# Inside Anthropic's $100 Billion Al Compute Commitment | CFO Krishna Rao

## Metadata
- Channel: Invest Like The Best
- Duration: 82 min
- YouTube: https://www.youtube.com/watch?v=wEEZPpx8qow

## Transcript

**[00:00] Speaker A:** Every time we have a new model, there's a set of capabilities that are different. People tend to think about model intelligence as IQ. We think of it kind of differently. Intelligence for us is multi-dimensional. It's not just a score. What is the real-world capability of this model? Each model generation gives you the chance to do more with it, to do it better, to do it more efficiently, because we think the returns to frontier intelligence are extremely high. And it's extremely high, especially in enterprise. That's a core thesis of our business.  
每次我们推出新模型时,都会带来一系列不同的能力。人们往往把模型智能理解为智商。但我们的看法有所不同。对我们来说,智能是多维度的,不只是一个分数。这个模型在现实世界中的实际能力是什么?每一代模型都让你能用它做更多事情,做得更好,做得更高效,因为我们认为前沿智能的回报极高。尤其在企业领域,回报极高。这是我们业务的核心理念。  
**[00:41] Speaker B:** Krishna, I have been so excited for this conversation because you get to see from the inside one of the most interesting businesses in world history at maybe the most interesting time in world history, at least if you're a technologist or care about technology. One of the things that fascinates me most, just to dive right into something that I think we're both quite  
Krishna,我一直非常期待这次对话,因为你能从内部视角看到世界历史上最有趣的企业之一,而且是在世界历史上可能最有趣的时刻——至少对技术专家或关心技术的人来说是这样。最让我着迷的事情之一,我想直接切入我们都非常  
**[01:00] Speaker A:** passionate about is this question of compute that you have to deal with all day every day. It's a key part of what you do. It's a key part of what these companies are doing and there's just this whole revolution happening.  
热衷的话题,就是算力问题,你每天都要处理这个问题。这是你工作的关键部分,也是这些公司正在做的关键部分,而且正在发生一场革命。  
**[01:11] Speaker A:** I'd love you to just start by explaining what it's like to have to deal with that.  
我很想听你解释一下处理这个问题是什么感觉。  
**[01:14] Speaker A:** Like I understand at one point you were having like a daily meeting about how to allocate the compute and to who and why. Just like bring us into that part of your life because I think it's like right at the cutting edge of what's going on.  
据我了解,你们曾经有一段时间每天开会讨论如何分配算力,分配给谁,为什么分配。带我们了解一下你生活中的这部分吧,因为我觉得这正处于当前发展的最前沿。  
**[01:26] Speaker B:** Look, the compute that we procure is the lifeblood of our business. It is the most important thing in the company. It is the thing on which, it's like the canvas on which everything else gets built.  
听着,我们采购的算力是我们业务的命脉。它是公司里最重要的东西。它就像画布,其他一切都在上面构建。  
**[01:39] Speaker B:** And so the decisions we make and how much compute to buy are some of the most consequential and hardest decisions to make in the entire company.  
所以我们关于购买多少算力的决策,是整个公司最重要、最难做的决策之一。  
**[01:46] Speaker B:** You know, think of it this way. If you buy too  
你可以这样想。如果你买太多  
**[01:48] Speaker A:** Much compute, you go out of business. If you buy too little compute, you can't serve your customers and you're not at the frontier is the same thing.  
算力,公司就会倒闭。如果买太少算力,你就无法服务客户,也无法保持在前沿,结果是一样的。  
**[01:56] Speaker A:** So, you know, we talk a lot about this cone of uncertainty, but the idea of just these purchases that have these real world implications, right?  
所以,我们经常谈论这个「不确定性锥」,但这些采购决策会产生实际影响,对吧?  
**[02:03] Speaker A:** You can't just go out and, you know, buy a gigawatt of compute and have it delivered next week.  
你不能直接出去购买一千兆瓦的算力,然后下周就能交付。  
**[02:06] Speaker A:** You have to really think ahead to plan for this. And so, we really take a very disciplined approach to how we think about it.  
你必须提前认真规划。所以,我们对如何思考这个问题采取了非常严谨的方法。  
**[02:13] Speaker A:** So, we look bottoms up. You know, we model what we think demand will be. Obviously, we sometimes get that wrong.  
我们自下而上地分析。我们建模预测需求会是什么样。显然,我们有时会预测错误。  
**[02:17] Speaker A:** We think about the compute we need to stay at the frontier and we really look ahead and try to estimate that. And then as we go out and actually do these deals to procure compute, you know, flexibility is really important to us and so we build that flexibility into the deals themselves.  
我们考虑保持在前沿所需的算力,并真正向前看,尝试估算。然后当我们实际去做这些采购交易时,灵活性对我们来说非常重要,所以我们把灵活性构建到交易本身中。  
**[02:36] Speaker A:** into how we use the compute as well, because the way in which we bridge from a position we are today to where we want to go when the business is growing exponentially is to use that compute as efficiently as possible.  
也构建到我们如何使用算力中,因为当业务呈指数级增长时,我们从今天的位置过渡到想要达到的目标的方式,就是尽可能高效地使用算力。  
**[02:45] Speaker A:** I would say I spend 30 or 40% of my time on compute even today.  
我可以说,即使到今天,我仍然把30%到40%的时间花在算力上。  
**[02:51] Speaker B:** What does flexibility mean in that example?  
在那个例子中,灵活性是什么意思?  
**[02:53] Speaker A:** So it means a couple of different things. Number one, you know, we use three different chip platforms.  
它有几层不同的含义。首先,我们使用三种不同的芯片平台。  
**[02:57] Speaker A:** So we are customers of Amazon's Trainium chip, Google's TPUs and Nvidia's GPUs.  
所以我们是Amazon的Trainium芯片、Google的TPU和Nvidia的GPU的客户。  
**[03:02] Speaker A:** You know, we use these chips fungibly.  
我们可互换地使用这些芯片。  
**[03:04] Speaker A:** So if you think about the compute we buy, we're using it for model development.  
所以如果你想想我们购买的算力,我们用它来开发模型。  
**[03:07] Speaker A:** We're using it internally to speed up our own product and model development.  
我们在内部使用它来加速我们自己的产品和模型开发。  
**[03:10] Speaker A:** And then we're also using it obviously to serve customers.  
然后我们显然也用它来服务客户。  
**[03:14] Speaker A:** Across those three chip platforms, we're using compute for all of those internal and external uses.  
在这三个芯片平台上,我们将算力用于所有这些内部和外部用途。  
**[03:20] Speaker A:** And that  
而这  
**[03:25] Speaker A:** Flexibility, it actually took us a long time to be able to do that.  
在灵活性方面,我们其实花了很长时间才做到这一点。  
**[03:27] Speaker A:** We've invested in that over multiple years to be what I believe the most efficient users of compute amongst any of the frontier labs.  
我们在这方面投入了好几年时间,我相信现在我们是所有前沿实验室中算力使用效率最高的。  
**[03:35] Speaker A:** And that's not something that just happened overnight.  
这不是一夜之间就能实现的。  
**[03:39] Speaker A:** You know, when we started using TPUs, I think it was maybe the third generation TPUs was the first one we used at scale, people thought, "Oh, well, you're crazy. Everyone's using GPUs. Why aren't you using GPUs?"  
我们开始使用 TPU 的时候,我记得第一次大规模使用的是第三代 TPU,当时人们觉得「你们疯了吧,大家都在用 GPU,你们为什么不用 GPU?」  
**[03:49] Speaker A:** And we've invested very heavily to be able to use that compute incredibly flexibly.  
我们在这方面投入了大量资源,就是为了能够极其灵活地使用这些算力。  
**[03:54] Speaker A:** And then we look across the different generations of those chip platforms and use each generation for the best workload internally.  
然后我们会研究这些芯片平台的不同代次,在内部为每一代选择最适合的工作负载。  
**[04:01] Speaker A:** And so we really built this orchestration layer that gives us that flexibility to use all different types of compute and in doing so we also are able to get the most value out of it.  
所以我们真正构建了这样一个编排层,让我们能够灵活使用各种不同类型的算力,这样做也让我们能从中获得最大价值。  
**[04:14] Speaker A:** Am I thinking about this in the right way that like something like CUDA  
我这样理解对不对,比如像 CUDA 这样的东西  
**[04:16] Speaker A:** That has been a part of NVIDIA's story for a long time now that allows you to do a lot with the underlying actual hardware that you want to sort of eke your way into being as close to the bare metal as possible, and that's part of this flexibility and being able to control as many of the variables as you can.  
它一直是 NVIDIA 故事的一部分,让你能够充分利用底层的实际硬件,你想尽可能接近裸机层面,这就是灵活性的一部分,能够控制尽可能多的变量。  
**[04:31] Speaker A:** Is that the journey that you've been on?  
这就是你们一直在走的路吗?  
**[04:33] Speaker B:** That's part of the journey for sure, but it's also been actually pretty collaborative.  
这确实是我们旅程的一部分,但实际上也是一个很协作的过程。  
**[04:37] Speaker B:** So we work really closely with the Annapurna Labs team at Amazon to help influence the roadmap of these chips, because we believe what we're doing is really stressing the limits of what these chips are capable of.  
我们与 Amazon 的 Annapurna Labs 团队紧密合作,帮助影响这些芯片的路线图,因为我们相信我们正在做的事情真正在挑战这些芯片能力的极限。  
**[04:50] Speaker B:** And that means that a dollar of compute inside our organization goes further than I think it does anywhere else.  
这意味着在我们组织内部,每一美元的算力能发挥的作用比其他任何地方都要大。  
**[04:55] Speaker B:** But importantly, we basically want to utilize each chip to its best purpose within the company.  
但重要的是,我们基本上想让每个芯片在公司内部发挥它最擅长的用途。  
**[05:01] Speaker B:** So that does mean that we're building our  
所以这确实意味着我们在构建自己的  
**[05:05] Speaker A:** own compilers. We're really building things from the chip level up in order to have that customization and that flexibility to use it internally the way we think will generate the most ROI.  
编译器。我们真的是从芯片层面开始构建一切,以便拥有这种定制化能力和灵活性,按照我们认为能产生最大投资回报的方式在内部使用。  
**[05:15] Speaker B:** Can you explain this cone of uncertainty thing? Like I want to ask about all the component parts of this, but that feels like a really key starting point or overall frame for how to think about both sourcing and then the uses of compute. Can you just explain what that concept is?  
你能解释一下这个「不确定性锥」的概念吗?我想问这个的所有组成部分,但这感觉是一个非常关键的起点或整体框架,用来思考算力的采购和使用。你能解释一下这个概念是什么吗?  
**[05:27] Speaker A:** Sure. When you're building and growing a business exponentially, you know, really small movements in monthly or weekly growth rates result in compounding very, very different outcomes. And so as we're thinking ahead, you know, even with our revenue growth, it's really hard to predict this business, right? And it's really hard. I think humans mostly think linearly and you think incrementally. And that's something, you know, I've been at the company for two years. That's something  
当然。当你在指数级地构建和发展一个业务时,月度或周度增长率的很小变动,经过复利效应会导致非常非常不同的结果。所以当我们展望未来时,即使是我们的收入增长,也真的很难预测这个业务,对吧?这真的很难。我觉得人类大多是线性思考的,你会增量式地思考。这是我在公司两年来  
**[05:51] Speaker A:** That's a paradigm I've had to break for myself, right? To stop just thinking linearly and think on this exponential.  
必须为自己打破的一个范式,对吧?停止线性思考,开始用指数思维思考。  
**[05:57] Speaker A:** When you're on this exponential again, the range of outcomes starts to be really, really wide.  
当你处在这个指数曲线上时,可能的结果范围会变得非常非常宽。  
**[06:00] Speaker A:** We look at a range of scenarios and we look at different points in that cone of uncertainty over, you know, a one to two-year period, and then we kind of work backwards from that.  
我们会看一系列场景,看这个不确定性锥在一到两年期间的不同点位,然后我们从那里反推。  
**[06:11] Speaker A:** And what we want to do is we want to be at a place where we can, you know, obviously still be at the frontier. That's the most important thing.  
我们想要达到的状态是,显然仍然处于前沿,这是最重要的。  
**[06:16] Speaker A:** To be able to serve customers and then to be able to have enough internal compute to accelerate our employees.  
能够服务客户,然后能够有足够的内部算力来加速我们员工的工作。  
**[06:23] Speaker A:** It's interesting. If we were to say to our employees, you can't use our models anymore, we could serve billions of dollars of revenue with that compute that we allocate to employees internally, but we want to take a long-term view and a long-term perspective on that cone of uncertainty.  
有意思的是,如果我们对员工说,你们不能再使用我们的模型了,我们可以用分配给员工内部使用的那些算力来服务数十亿美元的收入,但我们想对这个不确定性锥采取长期视角和长期观点。  
**[06:40] Speaker A:** Because we want to range towards the top end of these outcomes, but we have to plan for that.  
因为我们想朝着这些结果的上限发展,但我们必须为此做好规划。  
**[06:44] Speaker A:** And as we go, that's how we think about buying compute in a disciplined way.  
随着业务发展,我们就是这样有纪律地购买算力的。  
**[06:48] Speaker A:** The most important thing is what happens if you are at one point in the cone of uncertainty but you've only bought compute for a different point.  
最关键的问题是:如果实际情况处于不确定性锥体的某个点,但你购买的算力却是按另一个点规划的,会发生什么?  
**[06:54] Speaker A:** That's where this compute efficiency is something that has really helped us out.  
这就是算力效率真正帮到我们的地方。  
**[06:58] Speaker B:** Can you bring us into the room for the conversations around the trade-offs between those?  
能不能带我们了解一下你们内部是如何权衡这些取舍的?  
**[07:01] Speaker B:** I'm so interested by those three buckets of like training, research, internal use broadly speaking, and then serving customer demand.  
我对这三大类特别感兴趣——训练、研究、内部使用,以及服务客户需求。  
**[07:08] Speaker B:** You naively might think like, okay, it's a third, a third, a third allocation or something. Like how much does that range around? What are the trade-offs like? What is that discussion like?  
你可能会天真地以为是三三制分配,各占三分之一。但实际分配比例会有多大浮动?权衡考量是什么?这种讨论是怎么进行的?  
**[07:19] Speaker A:** On an ongoing basis, in addition to meeting about compute procurement, we meet a lot about compute allocation.  
在日常运营中,除了讨论算力采购,我们还经常开会讨论算力分配。  
**[07:22] Speaker A:** I think what's  
我觉得  
**[07:25] Speaker A:** Important is it starts with a place where our culture is one that's incredibly collaborative, and that informs how this conversation happens. So there's not like FTEs—it's done in a very collaborative, not a zero-sum way.  
重要的是,这一切始于我们高度协作的文化,这种文化决定了讨论的方式。不像按全职员工数量分配那样——整个过程非常协作,不是零和博弈。  
**[07:37] Speaker A:** But there's a level of compute for model development that we will not go below, right? So even if it means it's harder to serve customers or we have to do kind of unnatural things when it comes to that, we want to continue to make that long-term investment in developing the best models.  
但模型开发有一个算力底线,我们绝不会低于这个标准。即使这意味着服务客户会更困难,或者我们不得不采取一些不太自然的做法,我们也要继续对开发最佳模型进行长期投资。  
**[07:52] Speaker A:** Because we think the returns to frontier intelligence are extremely high, and it's extremely high especially in enterprise.  
因为我们认为前沿智能的回报极高,尤其是在企业领域。  
**[07:59] Speaker A:** But so that kind of puts a floor on the compute that's allocated to model development.  
所以这就为模型开发设定了算力分配的下限。  
**[08:03] Speaker A:** And then as we think about the internal use of compute, it really helps us to speed up that model development and to speed up finding those compute efficiency multipliers that really get us more from each dollar.  
至于算力的内部使用,它确实帮助我们加快模型开发速度,并加快找到那些算力效率倍增器,让每一美元发挥更大作用。  
**[08:15] Speaker A:** Of compute. So when we're talking about it, each team is kind of representing what they would do with that compute.  
所以在讨论时,各个团队会说明他们会如何使用这些算力。  
**[08:20] Speaker A:** And then we have a really open and frank discussion about how we think about ROI.  
然后我们会非常开放坦诚地讨论如何看待投资回报率。  
**[08:26] Speaker A:** And because we can allocate that compute so dynamically, we can make changes. We can make adjustments in that on a relatively short time horizon.  
而且因为我们可以非常灵活地分配算力,所以能够在相对较短的时间内做出调整和改变。  
**[08:34] Speaker B:** The efficiency thing is so interesting to me. I'm curious if you have a sense of like how much more efficient you are versus your own internal benchmarks from a year ago or something, or versus others that you have some sense of how efficient they are. How do you measure what efficiency means?  
效率这个话题我很感兴趣。我好奇你们是否知道,相比一年前的内部基准,或者相比你们了解到的其他公司,你们的效率提高了多少?你们如何衡量效率?  
**[08:48] Speaker A:** So, there's a couple different ways I would think about it. From a model perspective, I think the analogy people have when these new models come out is like they're kind of like cars. You had a sedan before and then you might have the higher-end version of that sedan and you're kind of moving, moving, moving up.  
有几种不同的角度来看这个问题。从模型角度来说,人们对新模型发布的类比就像汽车一样。你之前有一辆轿车,然后可能升级到那款轿车的高配版,就这样一步步往上走。  
**[09:01] Speaker A:** The chain, and I think that is true in terms of model intelligence. The way the place that analogy kind of breaks down a little bit is people think, okay, I'm going from the sedan to the sports car.  
这个类比在模型智能方面确实成立。但这个类比有点不太准确的地方是,人们会想:好,我从轿车换到跑车了。  
**[09:10] Speaker A:** I'm going to get much less fuel efficiency, right? I'm not going to buy the sports car for the gas mileage.  
油耗肯定会高很多,对吧?我买跑车可不是为了省油。  
**[09:14] Speaker A:** In our case, we actually see both improvements, huge improvements in capability, but also in model efficiency.  
但在我们这里,实际上两方面都有提升——能力有巨大提升,模型效率也同样大幅提升。  
**[09:21] Speaker A:** And so, if you look at going from Opus, you know, 4 to 4.5, 4.6, and now 4.7, you know, each one of those leaps, they're not equal, but each one of those leaps to a new model has a multiplier in terms of how much more efficient it is at processing tokens effectively.  
所以如果你看从 Opus 4 到 4.5、4.6,再到现在的 4.7,每一次跃升——虽然幅度不完全相同——但每次升级到新模型,在处理 token 的效率上都有倍数级的提升。  
**[09:38] Speaker A:** And that just doesn't serve customers. That also helps us internally as well, because if you think about if we're using the model for, if we're doing reinforcement learning on the model, it's basically inference within a sandbox with a reward function.  
这不仅服务于客户,对我们内部也有帮助。因为如果我们用模型来做强化学习,本质上就是在一个带有奖励函数的沙盒环境中进行推理。  
**[09:52] Speaker A:** Right? And so if the model's better at more efficient inference, that RL is more efficient as well.  
对吧?所以如果模型的推理效率更高,那强化学习的效率也会更高。  
**[09:56] Speaker A:** And so we're able to do this kind of win-win where the customer is getting more capability when we release a new model.  
所以我们实现了双赢——当我们发布新模型时,客户获得了更强的能力。  
**[10:02] Speaker A:** And then we're able to serve that model sometimes again a multiple more efficient than the prior generation.  
然后我们就能部署那个模型,有时候效率能比上一代提升好几倍。  
**[10:09] Speaker A:** And then when we're in between generations, we're dynamically deploying efficiency improvements kind of in between these like more step function model changes.  
在两代模型之间的时期,我们会动态部署效率改进,就是在这些更大的阶跃式模型变化之间持续优化。  
**[10:20] Speaker A:** And so it is always getting more efficient over time.  
所以效率一直在随时间提升。  
**[10:22] Speaker A:** And what fuels that is the research team.  
而推动这一切的是研究团队。  
**[10:25] Speaker A:** So if you think about it, all these things are very connected—these various tasks and workloads that we have internally all kind of fit together in this way of, you know, doing R&D for model capabilities, for compute efficiency, for serving customers, and then having internal workloads that can be sped up by using the best models, sometimes models that we  
如果你仔细想想,所有这些事情都是紧密相连的——我们内部的各种任务和工作负载都以这种方式结合在一起,包括模型能力的研发、计算效率的优化、为客户提供服务,还有内部工作流程可以通过使用最好的模型来加速,有时候是我们自己的模型。  
**[10:46] Speaker A:** Unlike most software companies that try to maximize your time on their app to juice engagement, Ramp does the exact opposite. It understands that no one wants to spend hours chasing receipts, reviewing expense reports, and checking for policy violations.  
不像大多数软件公司想方设法让你在他们的应用上花更多时间来提升参与度,Ramp 恰恰相反。它明白没人想花几个小时追踪收据、审核报销单、检查违规情况。  
**[10:58] Speaker A:** So they built their tools to give that time back, using AI to automate 85% of expense reviews with 99% accuracy.  
所以他们开发的工具就是为了把时间还给你,用 AI 自动化处理 85% 的费用审核,准确率达到 99%。  
**[11:05] Speaker A:** And since Ramp saves companies 5%, it's no wonder that Shopify runs on Ramp, Stripe runs on Ramp, and my business does too. To see what happens when you eliminate the busy work, check out ramp.com/invest.  
而且 Ramp 能为公司节省 5% 的开支,难怪 Shopify 在用 Ramp,Stripe 在用 Ramp,我的公司也在用。想看看消除繁琐工作后会发生什么,可以访问 ramp.com/invest。  
**[11:19] Speaker A:** Felix by Rogo is a personal finance agent that turns a single prompt into finished client-ready work using your firm's own templates, context, and standards.  
Rogo 推出的 Felix 是一个个人财务智能体,能把一句简单的指令转化为完成的、可交付客户的工作成果,使用的是你公司自己的模板、背景信息和标准。  
**[11:26] Speaker A:** Send Felix an email like, "Take these comments and turn them for me," or "Update my tracker with the context of these emails," or "Run the ability to pay math on this buyer," and Felix sends back finished PowerPoint decks.  
给 Felix 发个邮件,比如「把这些评论整理成文档」,或者「用这些邮件的内容更新我的跟踪表」,或者「对这个买家做支付能力分析」,Felix 就会返回完成的 PowerPoint 演示文稿。  
**[11:38] Speaker A:** Excel models, and sourced research." Felix works the way your team already does, delivering work quickly and accurately around the clock.  
还有 Excel 模型和有来源的研究报告。Felix 按照你团队已有的工作方式运作,全天候快速准确地交付工作成果。  
**[11:45] Speaker A:** Learn more at rogo.ai/felix.  
了解更多请访问 rogo.ai/felix。  
**[11:49] Speaker A:** OpenAI, Cursor, Anthropic, Perplexity, and Vercel all have something in common.  
OpenAI、Cursor、Anthropic、Perplexity 和 Vercel 都有一个共同点。  
**[11:53] Speaker A:** They all use WorkOS. And here's why.  
他们都在使用 WorkOS。原因如下。  
**[11:56] Speaker A:** To achieve enterprise adoption at scale, you have to deliver on core capabilities like SSO, SCIM, RBAC, and audit logs.  
要实现大规模的企业级采用,你必须提供核心能力,比如 SSO、SCIM、RBAC 和审计日志。  
**[12:02] Speaker A:** That's where WorkOS comes in.  
这就是 WorkOS 的用武之地。  
**[12:04] Speaker A:** Instead of spending months building these mission-critical capabilities yourself, you can just use WorkOS APIs to gain all of them on day zero.  
与其花几个月时间自己构建这些关键能力,不如直接使用 WorkOS 的 API,从第一天起就拥有所有这些功能。  
**[12:11] Speaker A:** That's why so many of the top AI teams you hear about already run on WorkOS.  
这就是为什么你听说过的那么多顶尖 AI 团队已经在使用 WorkOS。  
**[12:16] Speaker A:** WorkOS is the fastest way to become enterprise-ready and stay focused on what matters most, your product.  
WorkOS 是最快达到企业级就绪状态的方式,让你能专注于最重要的事情——你的产品。  
**[12:22] Speaker A:** Visit workos.com to get started.  
访问 workos.com 开始使用。  
**[12:22] Speaker B:** You said something really important before, which is the returns to being at the frontier are really high.  
你之前说了一个很重要的观点,就是处于前沿位置的回报非常高。  
**[12:27] Speaker B:** Can you just explain  
能不能尽可能详细地解释一下。  
**[12:29] Speaker A:** That in as much detail as you can.  
这个问题。  
**[12:32] Speaker B:** Sounds obvious when you say it, but there's certainly been some camps that like, oh, I'll just, you know, I can use the six-month-old model and it's a fraction of the cost and I'll just use that and that'll be catching up all the time.  
听起来很显而易见,但确实有些人认为,哦,我就用六个月前的旧模型,成本只是一小部分,我就用那个,它会一直追赶上来的。  
**[12:42] Speaker B:** And that just hasn't been the case. Like everyone, the second Opus 4.7 comes out, like even me as a consumer, the thing you do is you switch it on or GPT 5.5 comes out, you switch on the new one right away. Like I want the best.  
但事实并非如此。就像所有人,Opus 4.7 一发布,甚至我作为普通用户,你做的第一件事就是立刻切换过去,或者 GPT 5.5 一出来,你马上就换到新版本。因为我想要最好的。  
**[12:55] Speaker B:** So talk about the returns to being on the frontier and why it's so high.  
那么谈谈处于前沿的回报,以及为什么这个回报如此之高。  
**[12:58] Speaker A:** I think it's a couple things. It's every time we have a new model, there's a set of capabilities that are different. People tend to think about model intelligence as IQ. It's a single number. Okay, this model was at 110 and then it goes to 125. We think of it kind of differently. Intelligence for us is multi-dimensional. It's not just a score. In fact, we find that yes,  
我认为有几个原因。每次我们推出新模型时,都会有一系列不同的能力。人们倾向于把模型智能看作智商,认为它是一个单一的数字。好吧,这个模型原来是110,然后变成了125。我们的看法有些不同。对我们来说,智能是多维度的,不只是一个分数。事实上,我们发现确实如此,  
**[13:19] Speaker A:** Everyone publishes their model benchmark cards and finds that a lot of those benchmarks are saturated.  
每个人都会发布他们的模型基准测试卡,然后发现很多这些基准测试已经饱和了。  
**[13:22] Speaker A:** You know, we publish it too.  
你知道,我们也发布这些。  
**[13:25] Speaker A:** But what our measurement is is what the customers tell us, like what is the real world capability of this model.  
但我们的衡量标准是客户告诉我们的,比如这个模型在真实世界中的能力是什么。  
**[13:30] Speaker A:** And as we've released better and better models, what we've seen is it's not just, you know, the outright intelligence.  
随着我们发布越来越好的模型,我们看到的不仅仅是纯粹的智能水平。  
**[13:36] Speaker A:** It's also the ability to do long horizon tasks, the ability to use tools or computer use, the ability to do things for agentic tasks that have specific value even faster, right?  
还有执行长期任务的能力、使用工具或计算机的能力、更快地完成具有特定价值的智能体任务的能力,对吧?  
**[13:48] Speaker A:** Which means that in some sense, you know, if you have two employees and they're maybe both equally capable, someone takes a week to do, you know, an assignment, someone does it in a day.  
这意味着在某种意义上,你知道,如果你有两个员工,他们可能能力相当,一个人花一周时间完成一项任务,另一个人一天就完成了。  
**[13:57] Speaker A:** Well, that second person, if they're continuing to do that, can be seven times better, right?  
那么,如果第二个人持续这样做,他可以好七倍,对吧?  
**[14:00] Speaker A:** They might be equally capable at something, maybe just take longer times to do it.  
他们在某件事上可能能力相当,只是完成所需的时间更长。  
**[14:04] Speaker A:** So all of those factor in to then how customers  
所以所有这些因素都会影响客户的  
**[14:08] Speaker A:** Experience it. And what we found very consistently is by releasing new models, the TAM is unlocked in a unique way.  
体验。我们非常一致地发现,通过发布新模型,总体可达市场(TAM)以独特的方式被解锁了。  
**[14:17] Speaker A:** Like more TAM gets unlocked, more use cases are possible.  
更多的TAM被解锁,更多的用例成为可能。  
**[14:20] Speaker A:** And a good illustration of that is this last four months that we've had at the company, right?  
一个很好的例证就是我们公司过去四个月的情况,对吧?  
**[14:24] Speaker A:** We started the year with about $9 billion of run rate revenue and we ended the quarter with, you know, north of $30 billion of run rate revenue.  
我们年初的年化收入约为90亿美元,到季度末时,我们的年化收入超过了300亿美元。  
**[14:30] Speaker A:** I mean, that kind of a change is really enabled by these model intelligence leaps and then the products that we build around them.  
我的意思是,这种变化真正得益于这些模型智能的飞跃,以及我们围绕它们构建的产品。  
**[14:40] Speaker A:** And so that's what I mean by the returns to frontier intelligence are really high.  
所以这就是我所说的前沿智能的回报真的很高的意思。  
**[14:43] Speaker A:** I think that's unique to enterprise because in consumer sometimes you don't see that as readily, that consumers really are pushing the limits of what the models can do, whereas in enterprise like our customers are always—now you know it started with coding but it's really expanded beyond that very meaningfully—but each model  
我认为这对企业来说是独特的,因为在消费者领域,你有时不会那么容易看到这一点,消费者真的在推动模型能力的极限,而在企业领域,我们的客户总是——现在你知道它始于编码,但已经非常有意义地扩展到了其他领域——但每一代模型  
**[15:01] Speaker A:** Generation gives you the chance to do more with it, to do it better, to do it more efficiently, and customers see that and then they invest really heavily in more tokens with the newer models, and we've just seen that cycle play out again and again, and that's a core thesis of our business that especially in enterprise, the returns to frontier intelligence are not slowing down.  
都让你有机会用它做更多事情,做得更好,做得更高效,客户看到了这一点,然后他们在使用更新模型的更多token上进行大量投资,我们一次又一次地看到这个循环上演,这是我们业务的核心论点,特别是在企业领域,前沿智能的回报并没有放缓。  
**[15:20] Speaker A:** The things that push that frontier is like a sci-fi story or something from books I was reading when I was growing up.  
推动这个前沿的事情就像科幻故事或我小时候读的书里的东西。  
**[15:26] Speaker A:** It seems as though in the major labs we've reached this point—someone on your team said it recently—of like recursive self-improvement, where the models themselves are building and doing a lot of the research to do, you know, the next generation of improvement.  
似乎在主要的实验室中,我们已经达到了这样一个点——你们团队的某个人最近说过——就像递归自我改进,模型本身正在构建并进行大量研究工作,来实现下一代的改进。  
**[15:41] Speaker A:** And that there is some sort of—if I think about the frontier that you're pushing and OpenAI is pushing and compare that to the open source models—that maybe the gap will widen as a...  
而且存在某种——如果我想想你们正在推动的前沿和OpenAI正在推动的前沿,并将其与开源模型进行比较——差距可能会作为一个...  
**[15:51] Speaker A:** Result of you getting there first to this like recursive thing. How do you think about that?  
你们首先到达这种递归状态的结果而扩大。你怎么看待这个问题?  
**[15:56] Speaker A:** Like what, tell us how we should think about this idea of recursive self-improvement in the models themselves, because it seems like getting there first is incredibly important because then you just can continue to separate yourself versus those that haven't reached it yet.  
告诉我们应该如何理解模型本身的递归自我改进这个概念,因为似乎首先到达那里非常重要,因为这样你就可以继续将自己与那些尚未达到的人区分开来。  
**[16:09] Speaker B:** I would say we do see progress accelerating. We see, you know, I can't speak for other companies, but for us the scaling laws are, you know, alive and well and we're seeing that, you know, even with releases more recently like Mythos, right now within the company, you know, 90 plus percent of our code is actually written by Claude Code, right? A lot of Claude Code is written by Claude Code, and so you think of this as like, why do we allocate compute internally? Why would we forego revenue for it? It's because the models themselves are helping us to build that next generation of models.  
我会说我们确实看到进展在加速。我们看到,你知道,我不能代表其他公司发言,但对我们来说,规模定律是,你知道,活跃且运作良好的,我们看到,你知道,即使是最近像Mythos这样的发布,现在在公司内部,你知道,我们90%以上的代码实际上是由Claude Code编写的,对吧?很多Claude Code是由Claude Code编写的,所以你可以把这看作是,我们为什么要在内部分配算力?我们为什么要放弃收入来做这件事?这是因为模型本身正在帮助我们构建下一代模型。  
**[16:42] Speaker A:** And so in addition to this capability leap that you would have just from the scaling laws, talent is really important.  
所以除了扩展定律带来的能力飞跃之外,人才也非常重要。  
**[16:48] Speaker A:** And that talent with the best models can really accelerate the development of the capabilities.  
拥有最佳模型的人才能够真正加速能力的发展。  
**[16:52] Speaker A:** And we're really seeing that.  
我们确实看到了这一点。  
**[16:53] Speaker A:** We don't really think about models as like closed or open.  
我们并不真的把模型看作是闭源或开源的。  
**[16:57] Speaker A:** We think of them as frontier or not.  
我们把它们看作是前沿的或非前沿的。  
**[17:00] Speaker A:** And the ones that are at the frontier, you know, clearly are capturing this economic value, driving meaningful ROI for customers.  
那些处于前沿的模型,显然正在捕获经济价值,为客户带来有意义的投资回报。  
**[17:06] Speaker A:** And we are just investing behind that thesis.  
我们就是在围绕这个论点进行投资。  
**[17:08] Speaker A:** And that means, you know, both compute but it also means talent to use that compute and use our own models to really accelerate the development.  
这意味着,既要投资算力,也要投资人才来使用这些算力,并使用我们自己的模型来真正加速开发。  
**[17:18] Speaker A:** The other piece of it is it's not just the models, it's the products that get built on top of them.  
另一个方面是,不仅仅是模型本身,还有基于模型构建的产品。  
**[17:21] Speaker A:** Right.  
对。  
**[17:23] Speaker A:** So we had 30 different product and feature releases in January.  
我们在一月份发布了30个不同的产品和功能。  
**[17:26] Speaker A:** The pace of that has accelerated as well and that's enabled in part by utilizing the models with  
这个节奏也在加快,部分原因是利用模型配合  
**[17:35] Speaker A:** The talent that we have to really accelerate ways to access this underlying intelligence.  
我们拥有的人才,真正加速了访问这种底层智能的方式。  
**[17:42] Speaker A:** That's kind of our theory of the case on the product side.  
这就是我们在产品方面的理论。  
**[17:43] Speaker B:** How do you think about this weird world where you mentioned the talent and the leverage and they're not writing code themselves and Claude's writing its own code?  
你怎么看待这个奇怪的世界?你提到了人才和杠杆作用,他们自己不写代码,而是Claude在写自己的代码。  
**[17:49] Speaker B:** It seems like the last step of that would be you don't even need the talent to tell the thing what to do, it just figures out what to do on its own and that's like the ultimate, you know, the thing then just runs and is only constrained by compute or something.  
似乎最后一步就是你甚至不需要人才来告诉它做什么,它自己就能弄清楚该做什么,那就是终极状态,你知道,然后它就自己运行,只受算力限制之类的。  
**[18:02] Speaker B:** Is that, am I being too crazy about that or is that future possible do you think?  
我这样想是不是太疯狂了,还是说这个未来是可能的?  
**[18:06] Speaker A:** I think that the core of our company is still a research lab.  
我认为我们公司的核心仍然是一个研究实验室。  
**[18:11] Speaker A:** I think it's maybe not as well understood, maybe it's getting more understood from the outside, but we're doing experiments.  
我觉得外界可能不太了解这一点,也许现在越来越了解了,但我们在做实验。  
**[18:17] Speaker A:** We are doing things that push the limits of what our models can do.  
我们在做的事情是推动我们模型能力的极限。  
**[18:22] Speaker A:** And that research and that  
而这个研究和这个  
**[18:24] Speaker A:** Engine is upstream of everything else that we've talked about, and so that is enabled by the models today.  
引擎是我们讨论的其他一切的上游,所以它是由今天的模型赋能的。  
**[18:31] Speaker A:** It's not entirely done by the models. Over time, we think that the models will get better.  
它不是完全由模型完成的。随着时间推移,我们认为模型会变得更好。  
**[18:35] Speaker A:** They'll be more helpful in that process. But having the best talent to set the direction, not just the priorities, but some of the new areas of discovery, it just actually makes that research talent even better, right?  
它们在这个过程中会更有帮助。但拥有最好的人才来设定方向,不仅仅是优先级,还有一些新的发现领域,这实际上让研究人才变得更好,对吧?  
**[18:47] Speaker A:** And so I think of it as accentuating and accelerating the talent that we already have.  
所以我认为这是在增强和加速我们已有的人才。  
**[18:52] Speaker A:** We talk a lot about how talent density beats talent mass, and I think that's true here.  
我们经常讨论人才密度比人才总量更重要,我认为这一点在这里也适用。  
**[18:58] Speaker A:** Like we want the densest collection of AI research talent and inference engineering talent, and that enabled with the best models we think is a really winning combination.  
我们想要的是最密集的 AI 研究人才和推理工程人才的集合,再配合我们认为最好的模型,这是一个非常有竞争力的组合。  
**[19:10] Speaker B:** How are scaling laws talked about internally? Like the sort of consensus has been you've got different components of them.  
内部是如何讨论扩展定律的?目前的共识是扩展定律有不同的组成部分。  
**[19:16] Speaker A:** You've got like pre-training, you've got post-training, you've got reasoning, and that all of these are kind of moving at different paces and to hit a true wall, they would all need to fall down.  
包括预训练、后训练、推理能力,这些都在以不同的速度发展,要真正遇到瓶颈,需要所有这些方面都停滞不前。  
**[19:24] Speaker A:** Like that's sort of like how the world is now conceptualizing scaling laws.  
这就是目前业界对扩展定律的理解方式。  
**[19:27] Speaker A:** How are they talked about internally? How do you think about them?  
你们内部是怎么讨论的?你是怎么看待这个问题的?  
**[19:31] Speaker B:** Yeah, I mean, we look at models at various points in their development.  
我们会在模型开发的不同阶段对其进行评估。  
**[19:35] Speaker B:** We can see, you know, during a pre-training run, how does this model compare to a prior model that we did on these kind of loss curves?  
在预训练过程中,我们可以看到这个模型在损失曲线上与之前的模型相比表现如何。  
**[19:44] Speaker B:** And that gives us a sense for model capability.  
这让我们对模型能力有了一个判断。  
**[19:46] Speaker B:** You can do the same thing as you think about RL.  
在强化学习阶段也可以用同样的方法。  
**[19:48] Speaker B:** And then probably as importantly is when customers get their hands on it, like what are they seeing?  
同样重要的是,当客户真正使用模型时,他们看到了什么。  
**[19:55] Speaker B:** Where are they identifying pain points?  
他们发现了哪些痛点?  
**[19:57] Speaker B:** And those pain points then become like training targets for us, right? We don't train on customer data on the enterprise side.  
这些痛点就成为我们的训练目标。不过在企业端,我们不会用客户数据来训练模型。  
**[20:04] Speaker B:** On the prosumer—  
在个人消费者端——  
**[20:06] Speaker A:** Side, it's only if you opt in. But customers tell us things like, "Hey, I wish the model were better at this or I had this particular place where it got stuck and I could build this other product, but the capability needs to be further than that."  
只有在用户选择加入的情况下才会使用。但客户会告诉我们:「我希望模型在这方面表现更好」或者「我在某个地方卡住了,我本可以开发另一个产品,但模型能力还需要进一步提升」。  
**[20:19] Speaker A:** What we usually tell them is, okay, build your product for that because we're going to on the R&D side improve that over time.  
我们通常会告诉他们,可以先为那个场景开发产品,因为我们会在研发端持续改进模型能力。  
**[20:26] Speaker A:** And so there is this connected loop, but internally we're always looking at different models that are being trained, different snapshots that we have, and comparing them internally and to a lesser extent externally against our own measure and then ultimately how our customers view them as well.  
所以这是一个闭环反馈机制。在内部,我们会持续观察正在训练的不同模型、不同的快照版本,在内部进行对比,也会在一定程度上与外部模型对比,既用我们自己的标准衡量,最终也看客户如何评价。  
**[20:41] Speaker B:** And it feels like there's just no slowdown in the scaling laws themselves. Is that a fair characterization?  
感觉扩展定律本身并没有放缓的迹象。这样理解对吗?  
**[20:47] Speaker A:** For us that's a fair characterization. Yeah, we are extremely—I mean—  
对我们来说确实如此。我们非常——我的意思是——  
**[20:50] Speaker A:** Obviously a bunch of the authors of the scaling laws papers are amongst our founders, but you know, notwithstanding that, we can be a bit of a skeptical bunch. Like, you know, we hold ourselves to a really high standard. Again, it's this kind of idea of a research lab that's very kind of scientific method, and people are constantly challenging previously held assumptions. But from what we see, the scaling laws are not slowing down.  
显然,很多扩展定律论文的作者都是我们的创始人,但即便如此,我们也是一群相当skeptical的人。我们对自己的标准很高。这又回到了研究实验室的理念,非常注重科学方法,大家会不断挑战之前的假设。但从我们看到的情况来看,扩展定律并没有放缓。  
**[21:08] Speaker B:** So if that's true, you said before it's hard for humans to be exponential in their thinking and not linear. Like, if that continues to be true for however many more, you know, turns of the crank here, how do you do that thing of not thinking linear and thinking exponential yourself in your job and for the business? Like, the implications are really hard to reason through. Exponential growth rate is one thing, but exponential growth of capability—like, I don't even know how to get my head around it.  
如果这是真的,你之前说过人类很难用指数思维而不是线性思维来思考。如果这种趋势继续下去,不管还要经历多少轮迭代,你如何在工作和业务中避免线性思维而采用指数思维?这真的很难推理。指数增长率是一回事,但能力的指数增长——我甚至不知道该如何理解。  
**[21:33] Speaker A:** So how do you get your head around it?  
那你是如何理解的?  
**[21:35] Speaker B:** We think about the world as scenarios. It's very hard to have a point estimate in this business.  
我们用场景化的方式来思考。在这个行业很难做出单点预测。  
**[21:39] Speaker B:** And then having a very low bar for updating your current prior or your current perspective, because it could be the case that something a month ago was true that's just not true today and that breaks your model.  
然后要对更新当前认知或观点保持非常低的门槛,因为可能一个月前还成立的事情今天就不成立了,这会打破你的模型。  
**[21:50] Speaker B:** And so this old, like, well, we'll forecast, you know, once a quarter and we'll, you know, we'll revisit this in three months at the next board meeting. That doesn't work for our business.  
所以那种「我们每季度做一次预测,三个月后的下次董事会再讨论」的老方法不适合我们的业务。  
**[22:00] Speaker B:** It's so dynamic that we have to always think about, oh, you know, our models couldn't do this before and they could do this now. What does that mean for the TAM? We've seen this in coding first, right, where, you know, starting with around Sonnet 3.5/3.6, we started to see this really remarkable jump in capability, which was then followed by adoption and usage and  
市场变化非常快,我们必须时刻思考:我们的模型以前做不到的事情,现在能做到了,这对潜在市场规模意味着什么?我们首先在编程领域看到了这一点——从 Sonnet 3.5/3.6 开始,我们看到能力出现了非常显著的跃升,随之而来的是采用率、使用量的增长,以及  
**[22:21] Speaker A:** Revenue, and you know, it was a little hard to predict that, but now we can use coding as an analog for a lot of what's happening elsewhere in the economy and elsewhere in our business. And so we kind of look at pattern recognition in our own business to try to predict what's going to happen in the future.  
收入的增长。虽然这有点难以预测,但现在我们可以把编程领域的情况作为类比,来理解经济其他领域和我们业务其他部分正在发生的事情。所以我们通过识别自己业务中的模式,来尝试预测未来会发生什么。  
**[22:37] Speaker B:** Literally 15 minutes before you got here, the news came out about your partnership with xAI in the Tennessee facility. Makes me curious about how you are canvassing the world for—like, that is an opportunity you decided to do. Like, I'm sure there's a universe of things that you've explored. What is the strategy for trying to get more in creative ways? Like, bring us a little bit more into that.  
就在你到这里前 15 分钟,你们与 xAI 在田纳西州设施合作的消息刚刚发布。这让我很好奇你们是如何在全球范围内寻找机会的——比如,这是你们决定做的一个机会,我相信你们肯定探索过很多其他可能性。那么,用更有创意的方式获取更多算力的策略是什么?能不能多讲讲这方面?  
**[22:58] Speaker A:** We announced a partnership with xAI for their Colossus facility in Memphis. We're really excited about that. It's going to allow us to continue to expand, especially on the consumer and prosumer side. But that's just one example of us just, as you said, looking  
我们宣布了与 xAI 在孟菲斯 Colossus 设施的合作。我们对此非常兴奋,这将让我们能够继续扩展,尤其是在消费者和专业消费者这一侧。但这只是一个例子,正如你所说,我们一直在寻找  
**[23:13] Speaker A:** For near-term compute, wherever we can get it, as the compute base grows, that near-term compute becomes a smaller and smaller fraction of what's available and what's out there.  
近期可用的算力,无论在哪里能找到。随着算力基础的增长,这些近期算力在可用总量中所占的比例会越来越小。  
**[23:22] Speaker A:** But we look at it as can we deploy that compute that's available productively.  
但我们关注的是:我们能否高效地利用这些可用的算力。  
**[23:29] Speaker A:** Sometimes the answer is yes and sometimes it's no, but if we can, then we look at the economic return on it based on what it's priced, what duration we have it for, where it's located, what type of compute it is, and how efficiently we can run it.  
有时答案是可以,有时是不行。但如果可以的话,我们会根据价格、使用期限、所在位置、算力类型以及我们能以多高的效率运行它,来评估经济回报。  
**[23:42] Speaker A:** So we have a process to assess, and that same process, by the way, we use to assess longer-term deals as well.  
所以我们有一套评估流程,顺便说一句,我们用同样的流程来评估长期合约。  
**[23:48] Speaker A:** So last month, we signed a 5 gigawatt deal with Google and with Broadcom for TPUs starting in 2027.  
上个月,我们与 Google 和 Broadcom 签署了一份 5 吉瓦的协议,用于从 2027 年开始使用 TPU。  
**[23:57] Speaker A:** We also signed a deal with Amazon for Trainium for up to 5 gigawatts as well. It was an over hundred billion dollar commitment.  
我们还与 Amazon 签署了一份最高 5 吉瓦的 Trainium 协议,这是一笔超过千亿美元的承诺。  
**[24:04] Speaker A:** And a lot of that compute is actually already  
而且这些算力中的很大一部分实际上已经在  
**[24:08] Speaker A:** Landing and will land in the rest of this year into next year. And so if you think about it, it's a bit of this layer cake of compute that's starting at different times with different capabilities, and we're very dynamically comparing that compute.  
交付中,并将在今年剩余时间到明年陆续到位。所以你可以把它想象成一个分层的算力结构,不同层在不同时间开始,具有不同的能力,我们会非常动态地比较这些算力。  
**[24:21] Speaker A:** It's price performance over time that's really, really important to us when it lands and what we think we can do with it internally in the business.  
对我们来说,真正重要的是随时间变化的性价比——算力何时到位以及我们认为能在业务内部用它做什么。  
**[24:29] Speaker A:** And so there's so many different variables you have to optimize for around, you know, what compute it is, at what cost, and over what time horizon.  
所以有很多不同的变量需要优化,包括是什么算力、成本多少、以及时间跨度有多长。  
**[24:37] Speaker A:** But we have a pretty dynamic way of looking at kind of near-term compute and then medium to long-term compute.  
但我们有一套相当动态的方法来看待近期算力和中长期算力。  
**[24:42] Speaker A:** But the things we're assessing are largely the same. What is different is just the time horizon.  
我们评估的内容基本相同,不同的只是时间跨度。  
**[24:46] Speaker B:** What about the trade-off? You said price per performance. The trade-off between like cost per token or something, throughput and speed.  
那权衡取舍呢?你提到了性价比。比如每个 token 的成本、吞吐量和速度之间的权衡。  
**[24:51] Speaker B:** From the customer perspective, they care about both speed  
从客户角度来看,他们两者都在意  
**[24:56] Speaker A:** Probably unlocks some capability and use cases that are really interesting that we don't know about yet as these things get faster.  
速度可能会解锁一些我们还不知道的、非常有趣的能力和用例,随着这些东西变得越来越快。  
**[25:00] Speaker A:** Can you talk a little bit about that trade-off in compute as you're assessing it?  
你能谈谈在评估算力时这种权衡吗?  
**[25:05] Speaker B:** As we look across three different chip platforms, we also have multiple generations of chips within it, right?  
当我们审视三个不同的芯片平台时,每个平台内部还有多代芯片,对吧?  
**[25:12] Speaker B:** So it could be TPU, V5e and V6 and V7 and Trainium 2, Trainium 3, all of them are at different places on the price performance curve.  
可能是 TPU V5e、V6、V7,以及 Trainium 2、Trainium 3,它们都处于性价比曲线上的不同位置。  
**[25:20] Speaker B:** And then we importantly look at how we will utilize it, right?  
然后我们还会重点关注如何利用它,对吧?  
**[25:24] Speaker B:** Price performance is important because of efficiency.  
性价比很重要,因为关系到效率。  
**[25:26] Speaker B:** Speed is also important for certain use cases as well.  
对于某些特定用例来说,速度也同样重要。  
**[25:29] Speaker B:** So we look at the compute down to a very granular level in terms of what it can deliver for us and when, and that's something that we do. Again, our compute team leads that, but we closely collaborate across the business to say where do we...  
所以我们会非常细致地审视算力,看它能在什么时候为我们提供什么样的能力。这是我们一直在做的事情。我们的算力团队主导这项工作,但我们会与整个公司密切协作,确定我们需要在哪些地方...  
**[25:43] Speaker A:** We need this compute and for what, right? Okay, we might need, you know, CPUs for RL, we might need this, you know, more leading edge compute, and we're going to deploy it for, you know, our best and fastest models or for training them. And so from our perspective, it's customer demand, but it's also really down quite granular in terms of what is each chip best for and then what will we have.  
我们需要这些算力来做什么,对吧?比如,我们可能需要 CPU 来做强化学习,可能需要更先进的算力,然后我们会把它部署到最好最快的模型上,或者用来训练这些模型。所以从我们的角度来看,这既取决于客户需求,也要非常细致地考虑每种芯片最适合做什么,以及我们将拥有什么样的算力。  
**[26:08] Speaker B:** I'm always so curious by the metabolism of, in this case, Anthropic for new compute. Like, how fast could you take it? If I airdropped on you twice the compute that you have tomorrow, like, would you consume that? How fast would you consume that? If I airdropped 10 times the compute on top of you, how fast would you consume it? Can you calibrate us on that sort of thing? Like, is demand—actually, it feels like demand's unlimited between these three, you know, uses: training, internal, customer demand. Everyone's saying the same thing. Shortages everywhere, memory...  
我一直很好奇 Anthropic 对新算力的「代谢速度」。比如,你们能多快消化它?如果我明天空投给你们两倍于现有的算力,你们会用掉吗?多快能用完?如果我空投 10 倍的算力给你们,你们多快能消化掉?能不能给我们一个概念?感觉需求是无限的——训练、内部使用、客户需求这三个方面,大家都在说同样的话:到处都缺算力,内存...  
**[26:37] Speaker A:** Stocks, you know, mooning. Is it that extreme that like if you 2x'd or 5x'd or 10x'd the amount available to you tomorrow, you would just like more or less instantly consume it?  
股价飙升。真的有那么极端吗?就是说如果明天你们可用的算力翻 2 倍、5 倍或 10 倍,你们会或多或少立刻就用完?  
**[26:45] Speaker B:** This goes back to like how we use it and the fungibility of it. So the answer is, you know, we're constrained kind of across those use cases internally today.  
这要回到我们如何使用算力以及算力的可替代性。答案是,目前我们在内部的这些使用场景中确实都受到算力限制。  
**[26:55] Speaker B:** And you know I would say that you know a year or two ago it would have been harder to consume especially like a heterogeneous kind of compute drop in your example really quickly because these chip platforms are different and they are different, some are harder to operate, some of them have you know idiosyncrasies in terms of how we use it.  
我想说,一两年前,要快速消化算力会更困难,特别是像你举例中那种异构算力的突然增加,因为这些芯片平台是不同的,它们确实不同——有些更难操作,有些在使用方式上有特殊性。  
**[27:13] Speaker B:** I would say today that you know getting a bunch more compute I think it would be deployed you know very rapidly across those different use cases.  
但我想说,在今天,如果突然获得大量额外算力,我认为它会非常快速地部署到那些不同的使用场景中。  
**[27:19] Speaker B:** We probably have the same kind of allocation or calibration that we do with compute today but it's  
我们可能会采用与今天类似的算力分配或调配方式,但是  
**[27:28] Speaker A:** It's become a lot easier for us to spin up very quickly and deploy like almost any type of compute, and that's something again we think is a real advantage.  
对我们来说,快速启动并部署几乎任何类型的算力已经变得容易多了,这也是我们认为的一个真正优势。  
**[27:37] Speaker B:** Back to the ways the customers are using Anthropic, one of the interesting tensions and trade-offs that I'm fascinated to hear how you think through is between sort of the platform approach where I build my business on top of Claude and it powers my thing versus you doing the thing that I wanted to build.  
回到客户使用 Anthropic 的方式,有一个有趣的张力和权衡让我很想听听你们是怎么思考的:就是平台方式——我在 Claude 之上构建我的业务,它驱动我的产品——与你们直接做我想做的事情之间的权衡。  
**[27:54] Speaker B:** So this is like the classic, you know, Canva versus Figma or something like this.  
这就像经典的 Canva 对比 Figma 的例子。  
**[27:58] Speaker B:** How do you think about the right balance of how deep into the application layer you should go versus just being a pure enabling layer of like, we're going to provide the reasoning engine and the intelligence and, you know, world go forth and build whatever you want, you know, pay us through the API or whatever? That seems like a fascinating internal discussion and tension to some extent.  
你们如何看待应该深入到应用层多深,还是只做一个纯粹的赋能层——就是说,我们提供推理引擎和智能,然后大家去构建任何想要的东西,通过 API 付费之类的?这似乎是一个很有意思的内部讨论,在某种程度上也是一种张力。  
**[28:19] Speaker A:** Yeah, the way I would think about it is most of what we're building is platform, and we think that there's so many examples of where a platform can accrue a lot of value, but the customers who are building on that platform actually accrue even more value.  
是的,我的想法是,我们构建的大部分是平台,我们认为有很多例子表明平台可以积累大量价值,但在该平台上构建的客户实际上会积累更多价值。  
**[28:32] Speaker A:** We think that's what we're setting up for today.  
我们认为这就是我们今天正在建立的模式。  
**[28:34] Speaker A:** It's maybe akin to the early days of AWS, right?  
这可能类似于 AWS 的早期阶段,对吧?  
**[28:36] Speaker A:** If you think about the cloud platform and all the tools and services that are now built in—because it's not just the raw model access, it is, you know, prompt caching and the ability to use virtual machines and, you know, cloud code being called within there or dispatch or the cloud agents SDK, managed agents.  
如果你想想云平台以及现在内置的所有工具和服务——因为这不仅仅是原始的模型访问,还包括提示词缓存、使用虚拟机的能力、在其中调用 cloud code、dispatch、cloud agents SDK、托管代理等等。  
**[28:52] Speaker A:** All of these are effectively, I think of as vectors to access that model intelligence for other companies to build into their own products.  
所有这些实际上,我认为都是访问模型智能的向量,让其他公司能够将其构建到自己的产品中。  
**[29:01] Speaker A:** That's most of what we're focused on and really most of where we think the business is going from where we are today.  
这是我们关注的重点,也是我们认为从现在开始业务发展的主要方向。  
**[29:07] Speaker A:** That said,  
话虽如此,  
**[29:10] Speaker A:** We will also build our own applications on that same platform where a couple things are true.  
我们也会在同一个平台上构建我们自己的应用,前提是满足几个条件。  
**[29:15] Speaker A:** Number one, if we feel like we have a vision into where the models are going and we can kind of demonstrate that and create customer value in that, that might be something like Claude Code, right?  
第一,如果我们觉得我们对模型的发展方向有洞察,并且能够展示这一点并创造客户价值,那可能就是像 Claude Code 这样的产品,对吧?  
**[29:28] Speaker A:** We were able to say actually a lot of what's out there in the market was developer-led. Claude Code is a platform that's Claude-led and we think the models couldn't quite do that today when it was launched a little over a year ago, but we think they'll get there and they have.  
我们能够说,实际上市场上的很多产品都是开发者主导的。Claude Code 是一个由 Claude 主导的平台,我们认为在一年多前刚推出时,模型还做不到这一点,但我们认为它们会达到那个水平,而且它们确实做到了。  
**[29:41] Speaker A:** And so one is kind of building ahead to model capabilities.  
所以第一点是提前构建以适应模型能力的发展。  
**[29:43] Speaker A:** The second is thinking about ways to demonstrate value for the ecosystem that others might emulate.  
第二点是思考如何为生态系统展示价值,让其他人可以效仿。  
**[29:49] Speaker A:** Right? If you think about Claude for financial services or Claude for life sciences or even something like Claude security, these are ways in which we've kind of composed the platform.  
对吧?如果你想想 Claude for financial services、Claude for life sciences,甚至像 Claude security 这样的产品,这些都是我们组合平台能力的方式。  
**[29:59] Speaker A:** Again, we're building on the same platform as our customers and we think that creates like a level playing field.  
我们和客户使用同一个平台来构建产品,这创造了一个公平的竞争环境。  
**[30:04] Speaker A:** We also think that there's so much value that's going to accrue in some of these areas that, you know, our customers can win and we can win as well, which is why you've seen as we've launched some of these products, we've done them in a collaborative kind of partnership-oriented way, whether that be on the security side or, you know, design or financial services—we've partnered with the ecosystem.  
我们认为这些领域会产生巨大的价值,客户能赢,我们也能赢。所以你会看到我们推出这些产品时,都采用了合作伙伴导向的方式——无论是安全、设计还是金融服务领域,我们都与生态系统合作。  
**[30:25] Speaker A:** So I think of our strategy as mostly horizontal. We'll build vertical where, you know, we think we have some value to add or a perspective that's useful or a way to demonstrate to the market how we think about our platform adding value, and a lot of the value is going to accrue to the customers that are building on top of it.  
我们的策略主要是横向的。只有在我们认为能增加价值、提供有用视角,或者能向市场展示我们平台如何创造价值时,才会做垂直产品。大部分价值会流向在平台上构建应用的客户。  
**[30:44] Speaker A:** Our goal is to build the best models and then build the products and tools.  
我们的目标是构建最好的模型,然后开发相应的产品和工具。  
**[30:49] Speaker A:** And services that allow that intelligence to proliferate within customers.  
以及让这种智能能力在客户内部广泛应用的服务。  
**[30:53] Speaker B:** How much do you care that it's just a reality that people are scared of you, that there's a sense that because you control the most essential piece of these new applications, the underlying intelligence reasoning engine, that may totally be true and maybe already is true, that more of the value is occurring on top of the Anthropic platform than is being captured by it.  
你在意吗——人们确实害怕你们,因为你们控制着这些新应用最核心的部分,也就是底层的智能推理引擎。可能确实如此,甚至已经是现实:Anthropic 平台上产生的价值比平台本身捕获的更多。  
**[31:14] Speaker B:** But nonetheless, it's still scary to imagine, and I guess maybe you could say something similar about cloud and AWS or something like that.  
但不管怎样,这种想象还是让人害怕。也许你可以说云计算和 AWS 也有类似情况。  
**[31:22] Speaker B:** But how much do you think and care about the fact that some of your would-be customers or existing customers are in fact scared of you as a competitor?  
你有多在意这个事实——你的潜在客户或现有客户实际上把你当作竞争对手而感到害怕?  
**[31:29] Speaker A:** Part of what is hard in this business is it's changing so quickly, so the model capabilities  
这个行业的难点之一是变化太快,模型的能力有时连我们自己都感到意外。  
**[31:36] Speaker A:** sometimes even surprise us, and so when we release models or products on top of  
所以当我们发布模型或基于模型的产品时——  
**[31:41] Speaker A:** That like there is an element of what's happened, you know, in prior waves over the course of 5 years, 10 years, 20 years—it's happening in months now.  
以前需要 5 年、10 年、20 年才发生的变化,现在几个月就发生了。  
**[31:51] Speaker A:** And when we release things, people are also surprised by it in some ways, in the same way that we were surprised by it.  
我们发布新东西时,人们也会感到惊讶,就像我们自己当初惊讶一样。  
**[31:56] Speaker A:** But I think fundamentally what we are trying to do is be very partner-oriented towards the ecosystem.  
但从根本上说,我们努力以合作伙伴为导向来对待整个生态系统。  
**[32:02] Speaker A:** And that means that, you know, we have early access programs. We work very closely with customers. We listen to them about what capabilities they want.  
这意味着我们有早期访问计划,与客户紧密合作,倾听他们想要什么能力。  
**[32:12] Speaker A:** That doesn't mean that the things we release are sometimes not moments where you're like, "Wow, that's way more powerful than I thought it would be," or "I didn't realize the models would be able to do that this quickly."  
但这不代表我们发布的东西不会让人惊叹「哇,比我想象的强大太多了」或「没想到模型这么快就能做到这个」。  
**[32:20] Speaker A:** I think that part of that is a reality of where we are in this cycle and in this kind of development of intelligence.  
我认为这部分是我们所处的这个周期和智能发展阶段的现实。  
**[32:26] Speaker A:** But part of it is also we want to make those capabilities really accessible, and that—  
但另一部分原因是我们想让这些能力真正易于获取——  
**[32:33] Speaker A:** should accrue a lot of value to customers as well, and customers that are forward-footed on that and adopt, and frankly also ones that are building and using the tools that we offer on our platform.  
这会给客户带来很多价值,尤其是那些积极采用的客户,以及那些使用我们平台提供的工具来构建的客户。  
**[32:43] Speaker A:** We think we can actually accelerate them.  
我们认为可以真正加速他们的发展。  
**[32:45] Speaker A:** I think some of it's a reality of kind of frontier model development, but our approach to it is probably a little different and more partner-oriented.  
我认为这部分是前沿模型开发的现实,但我们的做法可能有点不同,更注重合作伙伴关系。  
**[32:52] Speaker B:** You said before going 9 to 30 in the first quarter.  
你之前说第一季度从 9 个客户增长到 30 个。  
**[32:53] Speaker B:** The pace is so insane, which makes me wonder about pricing—like the dynamic of how to price tokens or use of the system is so fascinating to me, because I think a lot of people a year ago would say price is going to just constantly fall.  
这个速度太疯狂了,这让我想到定价问题——如何给 token 或系统使用定价的动态变化让我很着迷,因为我觉得一年前很多人会说价格会持续下降。  
**[33:11] Speaker B:** But actually what's happening is it's going up in many cases, and this is true at different levels, whether it be the Mythos pricing that is quite high because it's so powerful, the cost of an H100, the rental price of  
但实际上很多情况下价格在上涨,这在不同层面都是如此,无论是因为功能强大而定价很高的 Mythos,还是 H100 的成本,或者——  
**[33:22] Speaker A:** The cost of an H100 is, well, you know, looks like a smile curve. I'm very curious why, if everyone is compute constrained, why everyone doesn't just raise prices a lot to try to find like what the right equilibrium is. And so I'd love you to just like riff on pricing, like how you think about it, what the trade-offs are, why not raise prices a lot.  
H100 的成本,嗯,你知道,看起来像一条微笑曲线。我很好奇,如果每个人都受算力限制,为什么不大幅提价来找到合适的平衡点。所以我想听你聊聊定价——你怎么考虑的,有什么权衡,为什么不大幅提价。  
**[33:41] Speaker B:** The company is only a little over 5 years old. This past March was the third anniversary of the first dollar of revenue into the business, and we only had a frontier model for real for the first time in March of 2024. So the time scale of these things is kind of, it's an important backdrop.  
公司成立才五年多一点。今年三月是公司获得第一笔收入的三周年,而我们真正拥有前沿模型是在 2024 年 3 月才第一次实现。所以这些事情的时间尺度是很重要的背景。  
**[33:56] Speaker B:** Our pricing has been relatively stable across, you know, Haiku, Sonnet, and Opus, and now Mythos is obviously newer, but you know, we made very few pricing changes. The biggest pricing change we made was to bring down the price of the Opus family when we launched Opus 4.5. And if we thought about why did we do that,  
我们的定价在 Haiku、Sonnet 和 Opus 这几个模型上一直相对稳定,现在 Mythos 显然是新推出的,但我们很少调整价格。我们做过的最大一次价格调整是在推出 Opus 4.5 时降低了 Opus 系列的价格。如果要说我们为什么这么做,  
**[34:18] Speaker A:** It's really because we found that Opus-class models were underutilized relative to their capability, right?  
实际上是因为我们发现 Opus 级别的模型相对于它的能力来说使用率偏低,对吧?  
**[34:23] Speaker A:** And so people were trying to often fit an Opus problem into a Sonnet workload.  
所以人们经常试图把本该用 Opus 解决的问题硬塞进 Sonnet 的工作负载里。  
**[34:30] Speaker A:** And because of the efficiency improvements that we were able to make, we were able to serve that very efficiently from our perspective, but actually bring down the price which made it more accessible to customers.  
由于我们在效率上取得了改进,从我们的角度来看可以非常高效地提供服务,同时还能降低价格,这让客户更容易使用。  
**[34:41] Speaker A:** And so it goes back a little bit to we want our customers to generate a lot of value from it and they're generating a ton of ROI from our models today.  
这又回到了一点,我们希望客户从中获得大量价值,而他们现在确实从我们的模型中获得了巨大的投资回报。  
**[34:52] Speaker A:** We want that to just continue because our goal is to proliferate this throughout the ecosystem.  
我们希望这种情况持续下去,因为我们的目标是让这项技术在整个生态系统中普及开来。  
**[34:56] Speaker A:** We think we're in the very, very early innings on all of these use cases.  
我们认为在所有这些应用场景上,我们还处于非常非常早期的阶段。  
**[35:00] Speaker A:** The best way to do that is to get this intelligence in the hands of as many businesses from startups to digitally native businesses to the largest companies in the world.  
最好的方式就是把这种智能能力交到尽可能多的企业手中,从初创公司到数字原生企业,再到世界上最大的公司。  
**[35:13] Speaker A:** That means that you have to make it in a price point that's accessible and that allows them to get a lot of value from it.  
这意味着你必须把价格定在一个可接受的点上,让他们能够从中获得大量价值。  
**[35:19] Speaker A:** The changing of the pricing for Opus actually, you know, you see this Jevons paradox, right?  
Opus 的价格调整实际上体现了杰文斯悖论,对吧?  
**[35:24] Speaker A:** Like we lowered the price of it, but the consumption went up way, way more than what you would have expected.  
我们降低了价格,但消费量的增长远远超出了你的预期。  
**[35:30] Speaker A:** And so because we kind of hit that sweet spot for customers, they were able to use it a lot more.  
因为我们找到了客户的甜蜜点,他们能够更多地使用它。  
**[35:35] Speaker A:** We had the efficiency to be able to serve it to customers at scale.  
我们有足够的效率能够大规模地为客户提供服务。  
**[35:39] Speaker A:** And then they were able to build that into their workload such that when we released Opus 4.6, it's a model improvement.  
然后他们就能把它整合到工作负载中,这样当我们发布 Opus 4.6 时,这是一次模型改进。  
**[35:44] Speaker A:** They can slot it in. We didn't change the price.  
他们可以直接替换进去。我们没有改变价格。  
**[35:46] Speaker A:** And so we think pricing stability is important.  
所以我们认为价格稳定性很重要。  
**[35:50] Speaker A:** And we also think that pricing to get that value and to see that kind of Jevons paradox happen is really important.  
我们也认为通过定价来获得那种价值,并看到杰文斯悖论发生,这真的很重要。  
**[35:57] Speaker A:** The other component of this is margins and how you think about margins as a business, again, because this...  
另一个要素是利润率,以及你作为一家企业如何看待利润率,因为这个...  
**[36:03] Speaker A:** It's so unbelievably capital intensive to build these frontier labs. You've got the leverage we talked about, which is efficiency, price—both those things relate to margin.  
建立这些前沿实验室需要极其庞大的资本投入。我们谈到的杠杆就是效率和价格——这两者都与利润率有关。  
**[36:14] Speaker A:** I apologize if there's a naive perspective, but given how much capital you need, why not just say we want to have a healthy margin and sort of set the price accordingly, and maybe that price can come down if efficiency is better or whatever? And so I'm curious how you think about margins as it relates to pricing in the business.  
如果这个问题太天真请见谅,但既然需要这么多资本,为什么不直接说我们想要一个健康的利润率,然后相应地设定价格,如果效率提高了价格也许可以降下来?所以我很好奇你如何看待利润率与定价和业务的关系。  
**[36:30] Speaker B:** Yeah, I would say we think about what is the return on our compute spend, right, writ large. So that is all of the different workloads that we've talked about, whether it's serving customers, model development. If you think of all of those, they're kind of in support of revenue over different time scales, right? If I serve inference, it's in support of revenue today. If I do model development, it might help for a  
是的,我会说我们考虑的是算力支出的整体回报率。这包括我们讨论过的所有不同工作负载,无论是为客户提供服务还是模型开发。如果你把所有这些都考虑进去,它们在不同的时间尺度上都是在支持收入,对吧?如果我提供推理服务,那是在支持今天的收入。如果我做模型开发,可能会帮助  
**[36:50] Speaker A:** Capability that unlocks TAM that drives revenue 6 months from now and everything in between.  
解锁某种能力,从而扩大潜在市场规模,推动 6 个月后的收入,以及介于两者之间的一切。  
**[36:55] Speaker A:** If I do internal acceleration to launch a new product, all of these things are in support of that.  
如果我做内部加速来推出新产品,所有这些都是在支持收入。  
**[37:00] Speaker A:** I will say our returns on that compute expense today are robust.  
我要说的是,我们今天在算力支出上的回报是稳健的。  
**[37:06] Speaker A:** They're robust and we think of it as what is the return on that full envelope of compute.  
它们很稳健,我们把它看作是整个算力投入的回报率。  
**[37:11] Speaker A:** And so we feel really good about where we are from that perspective.  
所以从这个角度来看,我们对目前的状况感到非常满意。  
**[37:16] Speaker A:** And we're balancing delivering value to customers with also seeing a really strong return on that compute ourselves.  
我们在为客户提供价值的同时,也在算力投入上获得了非常可观的回报,两者之间保持着平衡。  
**[37:22] Speaker A:** And if you think about when revenue grows, as we mentioned in Q1, it's not like we onboarded a bunch of new compute in that time period.  
如果你想想收入增长的情况,就像我们在第一季度提到的,并不是说我们在那段时间里突然增加了大量新算力。  
**[37:32] Speaker A:** We talked about compute comes based on a ramp that might have been determined 12 months ago.  
我们讲过,算力的部署是基于一个爬坡计划,这个计划可能在12个月前就已经确定了。  
**[37:36] Speaker A:** And so this idea of a variable cost that's like on the incremental to serve a customer is a little bit like it doesn't really fit our model.  
所以那种认为每服务一个客户就会产生增量可变成本的想法,其实并不太适用于我们的模式。  
**[37:47] Speaker A:** business, right? It tries to maybe fit our business into like a software paradigm, but that's not the case.  
对吧?这种想法试图把我们的业务套进软件行业的范式里,但实际情况并非如此。  
**[37:53] Speaker A:** And in actuality, compute is supporting all of these activities and we're really generating a robust return on that compute and that's our measuring stick.  
实际上,算力支撑着所有这些业务活动,而我们在这些算力上获得了非常稳健的回报,这才是我们的衡量标准。  
**[38:03] Speaker A:** And so, you know, I think it's something where we think of the compute envelope that we have as the thing that is able to govern how much we're able to drive revenue both over the short term and the long term.  
所以我认为,我们把现有的算力规模看作是决定我们能在短期和长期内推动多少收入增长的关键因素。  
**[38:16] Speaker B:** So if you're this great customer of the compute providers, what does that group need to do to be a great provider to you to help you drive that return?  
既然你们是算力供应商的优质客户,那么这些供应商需要做些什么才能成为你们的优质合作伙伴,帮助你们实现那样的回报呢?  
**[38:28] Speaker A:** So, we're fortunate in that we have really great partners in Amazon, in Google, in Microsoft, but also with Broadcom and Nvidia as well.  
我们很幸运,在Amazon、Google、Microsoft,以及Broadcom和Nvidia那里都有非常好的合作伙伴。  
**[38:37] Speaker A:** And so, our ecosystem, you know, we are the only model that's on all three clouds today. We're the only language  
在我们的生态系统中,我们是目前唯一一个在三大云平台上都部署的模型,也是唯一一个  
**[38:44] Speaker A:** Lab that's using all three of these chip platforms, and really these collaborations are much deeper than just like procurement.  
同时使用这三种芯片平台的语言实验室,而且这些合作的深度远不止采购那么简单。  
**[38:51] Speaker A:** I think that's something that's often lost.  
我觉得这一点经常被忽视。  
**[38:52] Speaker A:** If you think about our relationship with Amazon, you know, our teams are deeply embedded with the Annapurna Labs team.  
比如我们和Amazon的关系,我们的团队与Annapurna Labs团队深度合作。  
**[38:58] Speaker A:** You know, we are really good users of Trainium. We've spent a lot of time and energy and work closely with the team internally.  
我们是Trainium的优秀用户,投入了大量时间和精力,与他们的内部团队紧密协作。  
**[39:04] Speaker A:** We plan capacity together, right?  
我们一起规划算力容量,对吧?  
**[39:07] Speaker A:** If you think about the three clouds, they're great distribution engines for us too.  
如果你想想这三大云平台,它们对我们来说也是很好的分发渠道。  
**[39:11] Speaker A:** We have a really robust first-party business as well.  
我们自己的第一方业务也很强劲。  
**[39:15] Speaker A:** But these are multifaceted partnerships, whether it be on developing the chips themselves, landing that capacity, serving it, and then ultimately distributing it to the customers.  
但这些都是多方面的合作关系,涵盖芯片本身的开发、算力容量的落地、服务的提供,以及最终向客户的分发。  
**[39:26] Speaker B:** I'm thinking about your function, like the finance team and the ways that you might—I'm picturing this like ROI on  
我在想你的职能,比如财务团队以及你们可能采用的方式——我脑海中浮现的画面是关于算力投资回报率的  
**[39:31] Speaker A:** Compute things on different horizons with all these complex variables, which makes me wonder: how do you use these powerful tools yourself internally to run your group and the business? Like, what is the deployment of Claude Code and Claude in general on the finance team at Anthropic?  
在不同时间维度上有各种复杂变量的计算,这让我好奇:你们内部是如何使用这些强大的工具来运营团队和业务的?比如,Claude Code和Claude在Anthropic财务团队的部署情况是怎样的?  
**[39:47] Speaker B:** Yeah, so this is really interesting because we were using Claude Code, you know, about a year ago, and I started asking people like, "Is everyone just vibe coding?" And we started to use Claude Code as almost like an assistant, a digital co-worker, not just for coding tasks. And that actually was early in what eventually became CoWork, right? That was kind of an extension of Claude Code to say what it's done for agentic software development, it should do for all of knowledge work.  
这个问题很有意思,因为我们大约一年前就开始使用Claude Code了,当时我开始问大家「大家都在凭感觉写代码吗?」后来我们开始把Claude Code当作一个助手、一个数字同事来使用,不仅仅用于编码任务。这实际上是CoWork的早期形态,对吧?就是把Claude Code在智能体软件开发中的作用延伸到所有知识工作领域。  
**[40:18] Speaker B:** But then we started to productionize that, and I'm actually really proud. We spend a lot of time with our product team too. They kind of see how we use it and get input and feedback from that.  
然后我们开始将其产品化,我真的很自豪。我们也花了很多时间和产品团队合作,他们会观察我们如何使用,并从中获取意见和反馈。  
**[40:27] Speaker A:** But like today, you know, all of our legal entities, we can produce the statutory financial statements using Claude. And yes, a human checks it, but all of those financial statements are produced with Claude.  
但就像今天,我们所有法律实体的法定财务报表都可以用Claude生成。当然,会有人工审核,但所有这些财务报表都是由Claude制作的。  
**[40:38] Speaker A:** We also have a more real-time platform called Ant Stats. It used to take a lot of time to sift through all the data, get to conclusions, write a memo about it or publish a regular report on what's happening over the course of the day, what's driving it.  
我们还有一个更实时的平台叫Ant Stats。过去需要花很多时间筛选所有数据、得出结论、写备忘录或发布关于一天中发生了什么、是什么在驱动变化的定期报告。  
**[40:53] Speaker A:** We now have a library of skills for Claude that are specific to finance. I think last I checked there were over 70 of them that everyone can access through this common repository.  
我们现在为 Claude 建立了一个专门针对财务领域的技能库。我上次查看的时候,里面已经有超过 70 个技能,所有人都可以通过这个共享代码库访问使用。  
**[41:04] Speaker A:** And on top of that we built an MFR, a monthly financial review skill, and it can produce our monthly financial review. It's 90 to 95% ready and then all of our discussion becomes about what do we do, what are the implications, not what exactly happened because Claude is  
在此基础上,我们还构建了一个 MFR(月度财务回顾)技能,它可以生成我们的月度财务报告。这个报告能达到 90% 到 95% 的完成度,然后我们所有的讨论就可以集中在「我们该做什么、这意味着什么」,而不是「到底发生了什么」,因为 Claude 已经帮我们处理好了。  
**[41:21] Speaker A:** Not just reporting the weather, it's also helping to think about drivers and like why did the number change in the way it did, and that gives you tremendous insight into the business, both in terms of this like MFR that we do, but also on a daily basis. And so what used to take hours to produce, you know, a weekly report for, you know, what's driving revenue or what's driving our compute utilization, now comes down to 30 minutes, and then we can spend our time on the actual strategic implications of the business.  
Claude 不只是报告天气那样简单地呈现数据,它还会帮助思考驱动因素,比如为什么某个数字会以这种方式变化。这让你对业务有了深刻的洞察,不仅体现在我们做的这种 MFR 上,也体现在日常工作中。以前需要花几个小时才能完成的周报,比如分析收入驱动因素或计算资源利用率的报告,现在只需要 30 分钟,然后我们就可以把时间花在真正的战略层面思考上。  
**[41:47] Speaker A:** We can also get it in the hands of business, you know, leaders much more quickly, and so it's just meant that the insight engine is a lot faster within the company. We also have like, you know, I have a dashboard I look at token usage across...  
我们也能更快地把这些洞察交到业务负责人手中,所以公司内部的洞察引擎运转得快多了。我还有一个仪表板,可以查看整个公司的 token 使用情况……  
**[42:02] Speaker B:** The leaderboard.  
排行榜。  
**[42:03] Speaker A:** Yeah, we don't compensate people on it. No one's trying to token max for that, but it's really interesting because some of the most senior people...  
对,我们不会根据这个给人发奖金。没人会为了上榜而刻意刷 token 使用量,但有趣的是,财务团队里一些最资深的人……  
**[42:10] Speaker A:** Within the finance team are actually the biggest users of tokens, so it is not just, you know, the 22-year-old who joined and has a coding background and was doing that on the weekends and brought it to work.  
实际上是 token 使用量最大的用户。所以这不只是那种 22 岁刚入职、有编程背景、周末就在玩这些工具然后带到工作中的年轻人在用。  
**[42:20] Speaker A:** It's also people using the tools to change how they're working.  
也有很多人在用这些工具改变自己的工作方式。  
**[42:24] Speaker A:** Like, I think our number one user is our head of tax, and he's, you know, really focused on tax policy engines and automating large parts of the workloads that are happening within the team.  
比如我们使用量排第一的是我们的税务主管,他非常专注于税务政策引擎,并且在自动化团队内部的大量工作流程。  
**[42:35] Speaker A:** So I love seeing that, and I think I tell people, if we're not super users of this, if we're not pushing the limits of it, how can you expect customers to do that?  
我很喜欢看到这种情况。我跟大家说,如果我们自己都不是这个工具的超级用户,如果我们都不去突破它的极限,你怎么能指望客户去这么做呢?  
**[42:42] Speaker B:** As your business scales up, everything gets more complex, especially your compliance and security needs. With so many tools offering band-aids and patches, it's unfortunately far too easy for something to slip through the cracks. Fortunately, Vanta is a powerful tool designed to simplify and automate your security work and deliver a single  
随着你的业务规模扩大,一切都会变得更加复杂,尤其是合规和安全需求。市面上有太多工具只提供临时性的修补方案,很容易就会有东西从缝隙中漏掉。幸运的是,Vanta 是一个强大的工具,旨在简化和自动化你的安全工作,并为合规和风险提供单一的……  
**[42:58] Speaker A:** Source of truth for compliance and risk. There's a reason that Ramp, Cursor, and Snowflake all use Vanta. It frees them to focus on building amazing differentiated products knowing that compliance and security are under control.  
真实来源。Ramp、Cursor 和 Snowflake 都在使用 Vanta 是有原因的。它让这些公司能够专注于打造出色的差异化产品,因为他们知道合规和安全已经得到了妥善管理。  
**[43:09] Speaker A:** Invest Like the Best listeners get a special offer of $1,000 off Vanta when you go to vanta.com/invest.  
Invest Like the Best 的听众可以获得特别优惠,访问 vanta.com/invest 可以享受 1000 美元的折扣。  
**[43:17] Speaker A:** I know firsthand how complex the tech stack is for asset management firms. And seemingly every new tool and data source makes the problem even worse, adding more complexity, more headcount, and more risk.  
我深知资产管理公司的技术栈有多复杂。而且似乎每一个新工具和新数据源都会让问题变得更糟,增加更多复杂性、更多人力需求和更多风险。  
**[43:26] Speaker A:** Ridgeline offers a better way forward. One unified platform that automates away that complexity across portfolio accounting, reconciliation, reporting, trading, compliance, and more, all at scale.  
Ridgeline 提供了一条更好的前进道路。一个统一的平台,可以大规模自动化处理投资组合会计、对账、报告、交易、合规等各方面的复杂性。  
**[43:38] Speaker A:** Ridgeline is revolutionizing investment management, helping ambitious firms scale faster, operate smarter, and stay ahead of the curve.  
Ridgeline 正在革新投资管理行业,帮助有雄心的公司更快扩张、更智能运营,并保持领先地位。  
**[43:43] Speaker A:** See what Ridgeline can unlock for your firm. Schedule a demo at  
看看 Ridgeline 能为你的公司带来什么。在 ridgeline.ai 预约演示。  
**[43:46] Speaker A:** ridgeline.ai. Just as a human, does it freak you out at all that all of these—I've heard so many examples like this.  
作为一个人类,这些事情会不会让你感到有点不安?我听到过太多这样的例子。  
**[43:53] Speaker A:** It starts to feel like we just start doing the stuff that AI tells us to do, like in the sales example or the calendar or whatever.  
感觉我们开始只是做 AI 告诉我们去做的事情,比如销售的例子,或者日历管理之类的。  
**[43:59] Speaker A:** And maybe that's great. Maybe it's just such a better coordinator and, you know, wide thinker and optimizer than we ever could be that we should do what it tells us to do.  
也许这很好。也许它就是比我们更好的协调者、更广阔的思考者和优化器,所以我们应该按它说的做。  
**[44:08] Speaker A:** But it feels like ever so slightly dystopian to me that that reality is coming quickly.  
但对我来说,这种现实正在快速到来,感觉有那么一点点反乌托邦的味道。  
**[44:13] Speaker A:** I've had examples of it, too. And it feels kind of cool. Like it's helpful, but at the same time, if I really close my eyes, like, oh, I'm just like doing what it tells me versus me telling it what to do.  
我自己也有过这样的经历。感觉挺酷的,很有帮助,但同时,如果我真的闭上眼睛想一想,会觉得「哦,我只是在做它告诉我做的事」,而不是我告诉它该做什么。  
**[44:24] Speaker A:** It's a really interesting human dynamic that I'm curious for your take on.  
这是一个很有意思的人类动态变化,我很好奇你对此的看法。  
**[44:28] Speaker B:** I maybe have a slightly different view on it in that I think like it has made—you know, we've been able to hire great  
我的看法可能有点不同。我觉得它让我们能够在公司招到优秀的人才……  
**[44:36] Speaker A:** People at the company, but it has made even those incredibly talented people so much more productive, and there's a little bit of this—I think of it again like Jevons paradox, but for labor—which is that we have people who become incredibly more productive. We actually, we've hired a lot more people because of that, because there's no shortage of work to do. And now with the assistance of Claude, you know, people are spending less time in that MFR trying to reconcile some number, but they're actually thinking, oh, how do we reinvest this in the business?  
但它也让这些极具天赋的人变得更加高效。这有点像 Jevons 悖论在劳动力上的体现——人们的生产力大幅提升后,我们实际上雇佣了更多人,因为要做的工作永远不会短缺。现在有了 Claude 的帮助,人们在 MFR 上花更少时间去核对某个数字,而是真正在思考「我们如何把这些节省下来的时间再投资到业务中」。  
**[45:06] Speaker A:** How do we think about, you know, kind of dynamically allocating resources? Whereas before I'm working to tie out a number, or I'm, you know, in that accounting example, taking a long time to close the books.  
我们如何动态地分配资源?而以前我需要花时间去核对数字,或者像会计的例子那样,花很长时间才能结账。  
**[45:17] Speaker A:** So I actually think of it, you know, maybe even more optimistically, that it is an accelerant to our productivity, and that actually means that we can get a  
所以我其实更乐观地看待这件事,它是我们生产力的加速器,这意味着我们能完成更多工作。  
**[45:28] Speaker A:** A lot more done, and that even as we grow the team, those people are more productive as well as they come up the curve on how to use Cloud within our company, and I think that's starting to be true across many companies as well.  
即使团队在扩大,新成员在学习如何在公司内使用 Cloud 的过程中,他们的生产力也会更高。我认为这在很多公司都开始成为现实。  
**[45:38] Speaker B:** I'd love to talk about investors and capital formation. Of course, you've had to raise tons and tons of capital. At the same time, it seems as though, like, if I just squint my eyes and think about the multiple on current revenue, it's like not that crazy in terms of like where you're raising money.  
我很想聊聊投资人和资本形成的话题。你们当然筹集了大量资金。但同时,如果我眯着眼睛看当前收入的估值倍数,你们融资的估值其实也没那么夸张。  
**[45:52] Speaker B:** I'm so curious for you to teach us about what it's been like to interact with investors, like how you've seen their understanding of the company evolve and mature. Where do you think the investors as a group kind of like understand it now? Where are their misunderstandings about Anthropic? Tell us that side of your life.  
我特别好奇你能不能跟我们讲讲和投资人打交道是什么体验,他们对公司的理解是如何演变和成熟的。你觉得投资人群体现在对 Anthropic 的理解到了什么程度?他们对 Anthropic 还有哪些误解?跟我们聊聊你生活的这一面。  
**[46:07] Speaker A:** So I joined the company about two years ago. We were closing our Series D.  
我大约两年前加入公司,当时我们正在完成 D 轮融资。  
**[46:12] Speaker A:** At the time, you know, look, that was not a straightforward fundraising. The company really only had a frontier model in the middle of that fundraising. Towards the tail end of it, the FTX transaction was happening, which was liquidating a bunch of Anthropic shares, and so that was kind of the starting point. And at that point, the questions were around like, why do you need to have a frontier model? Like, what's the returns to this?  
那次融资并不顺利。公司当时只有一个前沿模型,而且在融资快结束时,FTX 的交易正在进行,清算了一批 Anthropic 的股份,这就是起点。那时候投资人会问,为什么你们需要有前沿模型?这能带来什么回报?  
**[46:36] Speaker A:** They were also around, you know, our mission and how we approach things. People said, well, hey, aren't AI safety and building a really big business—aren't those things at odds? And there were also a lot of other misconceptions: your sales force is really small, don't you need to scale it like all these enterprise software companies? And so there was just a bit of a paradigm around trying to fit us, you know, into a particular mold that had existed before.  
他们也会质疑我们的使命和做事方式。有人说,AI 安全和建立大型商业公司,这两件事不是矛盾的吗?还有很多其他误解:你们的销售团队这么小,难道不需要像其他企业软件公司那样扩大规模吗?所以当时有一种范式,就是试图把我们套进某个已有的模式里。  
**[47:01] Speaker A:** Over time, you know, it's—  
随着时间推移,情况发生了变化。  
**[47:04] Speaker A:** evolved. We at the end of 2024 we raised the Series E. You know the business had scaled to close to a billion dollars of run rate revenue, but the day of our first close was the day the DeepSeek news came out.  
2024 年底我们完成了 E 轮融资。业务已经扩展到接近 10 亿美元的年化收入,但我们首次交割的那天,正好是 DeepSeek 新闻爆出的那天。  
**[47:18] Speaker B:** Obviously we got the close done, but certainly a ton of volatility as people then said wait a minute, should I just totally rewrite how I think about AI in total? And so that was a Series E.  
我们显然完成了交割,但市场出现了巨大波动,人们开始说,等等,我是不是应该完全重新思考对整个 AI 行业的看法?这就是 E 轮的情况。  
**[47:29] Speaker A:** Obviously, we brought on great investors across all of these, but people still had some of those questions, but they looked at our forecast and they thought,  
我们在所有这些轮次中都引入了优秀的投资人,但人们仍然有一些疑问。不过他们看了我们的预测后会想:  
**[47:38] Speaker B:** okay, you know, I get it. You've grown to a billion dollars of run rate revenue so quickly, but there's no way you're going to be able to keep up. Yeah, that's just not possible, right? And there's laws of physics. You're in enterprise, which is great, but the  
好吧,我明白了。你们这么快就增长到 10 亿美元的年化收入,但你们不可能保持这个速度。这不可能,对吧?这违反物理定律。你们做的是企业市场,这很好,但  
**[47:52] Speaker A:** Adoption is going to take so much longer. I mean, look at how long it took with cloud and how many people are still on-prem.  
采用速度会慢得多。看看云计算花了多长时间,还有多少公司仍然在用本地部署。  
**[47:56] Speaker A:** The business continued to prove out the thesis that the return to frontier intelligence is really high that we are really focused on.  
但业务持续证明了我们的论点:前沿智能的回报确实很高,而我们真正专注于此。  
**[48:04] Speaker A:** What's really happened is model growth enabled by products and our go-to-market team and our distribution.  
实际发生的是,模型的增长是由产品、市场团队和分发渠道共同推动的。  
**[48:11] Speaker A:** And then I think what they also saw was that this thesis of like, hey, it's really important to build this transformative technology but to do it in the right way and do it responsibly, that that had this really interesting interlink with our business that most people didn't really understand or really believe.  
我认为他们还看到了一点,就是「以正确和负责任的方式构建这项变革性技术非常重要」这个理念,与我们的业务有着非常有趣的关联,而大多数人并不真正理解或相信这一点。  
**[48:31] Speaker A:** Which was that we invest in research not just in model development but also in AI safety research, right? Like we've pioneered interpretability, which is, think of it as like an MRI for the model to see inside the neural network how it works.  
我们不仅投资模型开发研究,还投资 AI 安全研究。我们在可解释性方面是先驱,可以把它想象成模型的核磁共振,能看到神经网络内部的运作方式。  
**[48:44] Speaker A:** We pioneered alignment science, which is  
我们在对齐科学方面也是先驱。  
**[48:47] Speaker A:** You know, you want the model to do what you tell it to do, and how often does it do that and how often does it stray from that. And those things are important for our mission and that's why we did them, but they had these downstream effects where it turns out if you can look inside the model, you're better at building them.  
你希望模型按照你的指令行事,它多久能做到这一点,多久会偏离。这些对我们的使命很重要,这也是我们做这些的原因,但它们产生了下游效应——如果你能看到模型内部,你就能更好地构建它们。  
**[49:02] Speaker A:** And then the last linkage, if you're selling to enterprises, like we now sell to nine of the Fortune 10, all of those enterprises are entrusting us with, you know, customer information, with their data.  
最后一个关联是,如果你要向企业销售,比如我们现在服务财富 10 强中的 9 家,所有这些企业都把客户信息和数据托付给我们。  
**[49:15] Speaker A:** They're interacting with their employees, sometimes even interacting with their customers as well. That is, those are the most sensitive workloads.  
我们与他们的员工互动,有时甚至与他们的客户互动。这些都是最敏感的工作负载。  
**[49:21] Speaker A:** The more and more of these businesses are running on cloud and our cloud platform. When you have this investment that we've made and will continue to make in safety, interpretability, alignment, like that actually inures to the benefit of the enterprise.  
越来越多的业务运行在云和我们的云平台上。当你在安全、可解释性、对齐方面进行了我们已经做的和将继续做的投资,这实际上对企业有利。  
**[49:36] Speaker A:** Customers as well, because all of our customers, if they're going to entrust us with all that access and all that data and the ability to work in the most sensitive workflows within their company, they want a company that they can trust.  
对客户也是如此,因为我们所有的客户,如果要把所有访问权限、所有数据以及在公司最敏感工作流程中工作的能力托付给我们,他们需要一家可以信任的公司。  
**[49:49] Speaker A:** And that's not why we invested in it, but it did have this kind of downstream effect that we've really seen prove out again and again to be a company that is both at the frontier, but one that is investing in safety and that you can trust.  
这不是我们投资的初衷,但它确实产生了一种下游效应,我们一次又一次地看到这一点得到验证——成为一家既处于前沿、又投资于安全、值得信赖的公司。  
**[50:00] Speaker A:** We've raised, you know, $75 billion since I joined the company.  
自从我加入公司以来,我们已经筹集了750亿美元。  
**[50:05] Speaker A:** We have another $50 billion that'll come in in the future from the Amazon and Google deals that we inked last month.  
未来还会有500亿美元陆续到账,这些资金来自我们上个月与Amazon和Google签订的协议。  
**[50:11] Speaker A:** And so that's a tremendous amount of capital, but it's a capital-intensive business and we need this capital to support that growth.  
这确实是一笔巨额资本,但我们这个行业本身就是资本密集型的,需要这些资金来支撑业务增长。  
**[50:16] Speaker A:** But you know, it all goes to the fact that, you know, the business is running very efficiently and so the reason we raise...  
但实际上,这一切都说明我们的业务运营效率很高,所以我们筹集资金的原因是……  
**[50:25] Speaker A:** This capital is more because of that cone of uncertainty than it is to fund, you know, actual losses in the business today.  
筹集这些资本更多是为了应对未来的不确定性,而不是为了弥补当前业务的实际亏损。  
**[50:30] Speaker B:** How is your own perception of this 10x perspective for a 10x growth of the business? Like the first time that happened, did you personally believe that it was possible? Did that seem absurd? And has it, like now that it's becoming consistent and so maybe it's becoming more commonplace to you or something, but what was your own view staring at this cone about the odds of hitting, you know, a 10x type of growth so many years in a row?  
你自己如何看待这种业务增长10倍的预期?第一次遇到这种情况时,你个人相信这是可能的吗?这听起来会不会很荒谬?现在这种情况变得越来越常见,可能对你来说也变得更加习以为常了,但你自己面对这个不确定性区间时,对于连续多年实现10倍增长的可能性是怎么看的?  
**[50:54] Speaker A:** Well, when I joined the business it had about 250 million of run rate revenue, and the plan was to get to a billion. And I said, great, in what year? And that was like linear thinking, right? And you know, consistently, Dario has been a much better predictor of the revenue than I have. I think we're going to close the gap over time as we get better.  
我刚加入公司时,年化收入大约是2.5亿美元,当时的计划是达到10亿。我就问,那是哪一年?这就是典型的线性思维,对吧?而事实证明,Dario对收入的预测一直比我准确得多。我想随着时间推移,我们会在预测方面做得更好,逐渐缩小这个差距。  
**[51:13] Speaker A:** At forecasting and understanding the business. But yeah, definitely the first time I saw it, you have all these arguments about the laws of physics and law of large numbers and this can't, you know, where is the revenue coming from and how can it be added this quickly and how can customers move this quickly and is this even possible in enterprise and all of those things start to get broken down over time as you see how the business works internally and you see how the adoption curves and the exponentials that are happening. Again, we have the exponential that's happening on revenue, but underlying that are these many other exponentials that support that.  
在预测和理解业务方面会越来越好。但确实,第一次看到这种增长时,你会有各种质疑——物理定律、大数定律、这不可能实现、收入从哪里来、怎么可能增长这么快、客户怎么可能转换这么快、在企业市场这真的可能吗——所有这些疑问随着时间推移都会被逐一打破,当你看到业务内部的运作方式,看到采用曲线和正在发生的指数级增长时。我们的收入呈指数级增长,但支撑这一增长的是许多其他的指数级因素。  
**[51:47] Speaker A:** You start to see and believe in that more. Now, that doesn't mean we're not disciplined and thoughtful about the forecast and how we think about the range of scenarios. But it does mean that my thinking has at least shifted a lot more from linear and incremental towards, you know,  
你会开始看到并相信这一点。当然,这并不意味着我们在预测和思考各种可能情景时不够严谨和周密。但这确实意味着我的思维方式已经从线性和渐进式转变了很多,转向了……  
**[52:03] Speaker A:** Leaning into this exponential and really, you know, believing in its potential and how this is just different than how other businesses have evolved.  
拥抱这种指数级增长,真正相信它的潜力,认识到这与其他业务的发展方式完全不同。  
**[52:12] Speaker B:** As you've talked to investors at every stage, I'm sure every stage, every round that you've raised, there's something that is like the most common or hardest thing to explain to investors or that they're struggling the most to understand and get their heads around. What is that today?  
在你与投资者交流的每个阶段——我相信每一轮融资都是如此——总会有一些最常见或最难向投资者解释的问题,或者他们最难理解和把握的东西。现在是什么?  
**[52:23] Speaker A:** I think it is this paradigm of how compute is used. Thinking of it as, you know, not just something that is like a variable cost over some time period, but really this resource that's so fully utilized, right? We run workloads on one day in the morning on a chip for inference and in the afternoon or evening we use it for model development. That is something that's, you know, that paradigm does not exist in a company like a software company or a factory, right? If you, you can't...  
我认为是算力使用的这种范式。不能把它仅仅看作某个时间段内的可变成本,而应该看作一种被充分利用的资源,对吧?我们在同一天早上用一块芯片做推理,下午或晚上就用它来做模型开发。这种范式在软件公司或工厂里是不存在的,对吧?你不能……  
**[52:59] Speaker A:** Repurpose if you have a bunch of people doing R&D and that's your R&D expense, they can't go and become cogs, right? And vice versa in most traditional companies.  
重新调配用途——如果你有一群人在做研发,那是你的研发支出,他们不能转而变成生产线上的工人,对吧?在大多数传统公司里反过来也不行。  
**[53:09] Speaker A:** Here, you really have that fungibility that's possible, and I think that's where the return on compute is so important. And I think people are beginning to understand that, but there's still a tendency towards treating it like, you know, oh I have to separate these two costs, when in actuality, you know, they're very self-reinforcing and that flexibility is actually what helps to drive revenue short-term and long-term.  
但在我们这里,这种可替换性是真实存在的,我认为这就是算力回报率如此重要的原因。我觉得人们开始理解这一点了,但仍然倾向于把它当作两种独立的成本来对待,而实际上,它们是相互促进的,这种灵活性正是推动短期和长期收入增长的关键。  
**[53:34] Speaker B:** If I was to force you out of your role and into an investor seat at a great big investing firm and then I said your job is to go grill these companies and invest in the best ones, like what questions would you be asking of the labs or companies that are building models to really get at the heart of, you know, the points of uncertainty, of skepticism, of like things that might not—  
如果我强行把你从现在的职位上拉下来,让你坐到一家大型投资公司的投资人席位上,然后说你的工作是去拷问这些公司,投资最好的那些,你会向那些构建模型的实验室或公司提出什么问题,来真正触及问题的核心——那些不确定性、质疑点,以及那些可能不会……  
**[53:58] Speaker A:** Make these the best businesses of all time. I'm curious maybe from that angle how you would approach it.  
让这些公司成为有史以来最好的企业的因素。我很好奇你会从这个角度如何看待。  
**[54:01] Speaker B:** So I would say a couple of things. First, what is the ROI on compute kind of all up? How are you utilizing it? And what return are you seeing today? And how is that coming over time? Right?  
我会问几个问题。首先,算力的整体投资回报率是多少?你们如何利用算力?今天看到了什么样的回报?随着时间推移回报如何变化?  
**[54:15] Speaker B:** These are the massive kind of unprecedented investments that companies like us are making. What is the return that you're generating on that and when does it come and what is the shape of it?  
像我们这样的公司正在进行前所未有的大规模投资。你们从中获得了什么回报,何时获得回报,回报曲线是什么样的?  
**[54:26] Speaker B:** So I think that's one. A second one is, how do your customers see ROI in what you do? Are people just using this for testing? Are they actually deploying this at meaningful scale?  
这是第一个问题。第二个是,你们的客户如何看待你们产品的投资回报率?人们只是在测试,还是真的在大规模部署?  
**[54:39] Speaker B:** You know, I could say for our business, like we're seeing that in spades. Our net dollar retention rate is over 500% on an annualized basis. And so, you know, with nine out of the Fortune 10...  
对于我们的业务,我可以说我们在这方面表现突出。我们的净美元留存率年化超过500%。而且财富10强中有9家都是我们的客户……  
**[54:54] Speaker A:** These are real, real customers making significant buying decisions.  
这些都是真实的客户,在做重大的采购决策。  
**[54:59] Speaker B:** No pilots anymore.  
不再是试点项目了。  
**[55:00] Speaker A:** Exactly. Like on the way here I was in an Uber and I signed two double-digit million-dollar commits during the car ride, which was like 20 minutes.  
没错。来这里的路上我坐在Uber里,20分钟的车程里我就签了两个千万美元级别的承诺订单。  
**[55:10] Speaker A:** So from that perspective, we're seeing it and we're now being judged by some of the biggest companies in the world, the most sophisticated buyers and startups that also have choice in the market and they're choosing us.  
从这个角度来看,我们确实看到了成效,现在世界上一些最大的公司、最成熟的买家以及市场上有选择权的初创公司都在评判我们,而他们选择了我们。  
**[55:22] Speaker A:** But I think one question I get a lot, or I would ask from the investor seat, the skeptical investor seat, is like how are your customers getting return from this?  
但我认为我经常被问到的一个问题,或者说如果我坐在持怀疑态度的投资人位置上会问的问题是:你们的客户如何从中获得回报?  
**[55:30] Speaker A:** Maybe a third one is, you know, how do you think about compute in the future and like where does it come from?  
可能第三个问题是,你怎么看待未来的算力,以及这些算力从哪里来?  
**[55:38] Speaker A:** Because obviously some of the places that we buy compute from, you know, they sell the compute to others, they...  
因为很明显,我们购买算力的一些供应商,他们也会把算力卖给其他人,他们自己也会...  
**[55:44] Speaker A:** might use the compute internally like what is the balance of that over time again for us like one reason why we have multiple different  
内部使用这些算力,那么随着时间推移这种平衡会是怎样的?对我们来说,我们与多家不同供应商合作的一个原因就是  
**[55:53] Speaker B:** And so your philosophy there is just like be involved with great players and have flexibility?  
所以你们的理念就是与优秀的合作伙伴保持合作,同时保持灵活性?  
**[55:58] Speaker A:** That's right, that's right. There's this crazy stat about AI just the generic concept being less popular than like Congress amongst like the general populace and it's kind of funny when you first hear, but when you really think about it, you're like, "This is kind of scary. Like, we need to solve this problem."  
没错,就是这样。有个很疯狂的统计数据,AI 这个概念在普通大众中的受欢迎程度甚至不如国会,第一次听到时觉得挺好笑的,但仔细想想会觉得「这其实挺可怕的,我们需要解决这个问题」。  
**[56:12] Speaker B:** It doesn't seem like the general world that isn't in technology, doesn't live in the Bay Area or New York, does not yet feel or understand why this is good for them just as measured by their opinion of it. What do you think we need to do as an industry about that problem?  
看起来不在科技行业、不住在湾区或纽约的普通人,还没有感受到或理解 AI 对他们有什么好处,至少从他们对 AI 的看法来衡量是这样。你认为作为一个行业,我们需要为这个问题做些什么?  
**[56:27] Speaker A:** Look, I think that if we think about the transformation that's happening, there's  
我觉得,如果我们思考正在发生的这场变革,  
**[56:35] Speaker A:** There have been other transformative waves, right, before all the way back to the industrial revolution, the internet, cloud, etc.  
之前也有过其他变革性的浪潮,对吧,一直追溯到工业革命、互联网、云计算等等。  
**[56:42] Speaker A:** I think what's one of the things that's different about AI is it's all happening so quickly.  
我认为 AI 的不同之处之一在于,一切发生得太快了。  
**[56:44] Speaker A:** You can have, you know, years or decades of progress that are being compressed into months.  
原本需要数年甚至数十年的进步,现在被压缩到了几个月内。  
**[56:51] Speaker A:** And going back to, you know, humans thinking in terms of exponentials versus linear, that can be jarring, I think.  
回到之前说的,人类习惯线性思维而不是指数思维,这种变化速度可能会让人感到不适应。  
**[57:00] Speaker A:** We are very optimistic generally about the potential for this technology.  
总体来说,我们对这项技术的潜力非常乐观。  
**[57:02] Speaker A:** I think that we as an industry can continue to do a better job of articulating, you know, Dario wrote this essay, Machines of Loving Grace.  
我认为作为一个行业,我们可以继续更好地阐述这一点,Dario 写过一篇文章叫《Machines of Loving Grace》(充满爱的机器)。  
**[57:11] Speaker A:** It's all about the potential for this technology to transform the way that we live.  
文章讲的就是这项技术如何有潜力改变我们的生活方式。  
**[57:16] Speaker A:** Whether that be in drug development and curing diseases that are more mainstream, but also rare diseases.  
无论是在药物研发和治疗常见疾病方面,还是罕见病方面。  
**[57:23] Speaker A:** Number two, in healthcare and how healthcare is delivered to raise our  
第二,在医疗保健以及医疗服务的提供方式上,提高我们的  
**[57:27] Speaker A:** Standard of living, you know, in the developing world and in places where resources are not as plentiful.  
生活水平,在发展中国家以及资源不那么充足的地方。  
**[57:33] Speaker A:** I think that all of those things are part of the promise and the potential of AI, and so we could probably do a better job of painting that picture and we want to show more tangible results for that over time.  
我认为所有这些都是 AI 的承诺和潜力的一部分,所以我们可能可以在描绘这幅图景方面做得更好,我们也希望随着时间推移展示出更多切实的成果。  
**[57:44] Speaker A:** I think that is coming and that's one of the things I'm most optimistic about.  
我认为这些成果正在到来,这也是我最乐观的事情之一。  
**[57:47] Speaker A:** I think on the other side though, we do, and this is again cultural to us, like we do want to articulate the risks.  
但另一方面,我认为我们确实需要,这也是我们文化的一部分,我们需要阐明风险。  
**[57:55] Speaker A:** Like I don't think we should just tell everyone everything's going to be great, because, you know, there are likely to be bumps on the road.  
我不认为我们应该只告诉大家一切都会很好,因为在这条路上很可能会遇到一些坎坷。  
**[58:03] Speaker A:** And so I think people generally gravitate towards more honest and balanced assessments, right?  
所以我认为人们通常更倾向于更诚实、更平衡的评估,对吧?  
**[58:07] Speaker A:** If I feel like somebody's just telling me all the good news and none of the bad news, then I'm like, okay, do I really trust this perspective?  
如果我觉得有人只告诉我好消息而不告诉我坏消息,那我就会想,好吧,我真的能相信这个观点吗?  
**[58:13] Speaker A:** I think that's where there's a need for balance and to  
我认为这就是需要平衡的地方,需要  
**[58:16] Speaker A:** Say like, look, these are some of the things that happen when change is compressed over a short amount of time.  
说,看,当变化在短时间内被压缩时,就会发生这些事情。  
**[58:21] Speaker A:** How do we work across, you know, commercial and government to actually come up with some of the solutions to that?  
我们如何跨越商业和政府部门,真正找到一些解决方案呢?  
**[58:27] Speaker A:** So I think it's about a clear articulation of the opportunities.  
我认为关键在于清晰地阐明机遇所在。  
**[58:30] Speaker A:** It's about really thinking about what those solutions may be.  
还要真正思考这些解决方案可能是什么样的。  
**[58:34] Speaker A:** And that's not any one company that can, you know, come up with it.  
这不是任何一家公司能够独自想出来的。  
**[58:37] Speaker A:** We don't have the blueprint that's going to solve everything, but to at least have that dialogue about some of the risks and downsides and what we can do to address it.  
我们没有能解决一切问题的蓝图,但至少要就一些风险和负面影响展开对话,讨论我们能做些什么来应对。  
**[58:45] Speaker A:** And then I think it's being transparent about that, about both of those things when we talk about it.  
然后我认为在谈论这些问题时,要对这两方面都保持透明。  
**[58:49] Speaker A:** And so over the long term, the opportunity is going to be significantly higher and greater than, you know, some of the risks and the downsides that will happen.  
从长远来看,机遇将远远大于那些会出现的风险和负面影响。  
**[58:58] Speaker A:** But that doesn't mean it's going to be perfectly smooth on the curve.  
但这并不意味着发展曲线会完全平滑。  
**[59:01] Speaker A:** The release of Mythos was such an  
Mythos 的发布是一个非常  
**[59:03] Speaker A:** Interesting moment. It was the first time many people, friends of mine that are careful watchers of this stuff said something like, "I'm like, this one kind of makes me scared."  
有意思的时刻。这是第一次,我的很多密切关注这个领域的朋友说:「这个模型让我有点害怕。」  
**[59:10] Speaker A:** So it relates back to the safety question.  
所以这又回到了安全问题。  
**[59:13] Speaker A:** It's also the first example of you coming out and saying, "We want to make sure this isn't used for bad, and it's maybe the first one that we are worried could be used for bad."  
这也是你们第一次站出来说「我们要确保它不被用于坏事」,可能也是第一个你们担心可能被用于坏事的模型。  
**[59:23] Speaker A:** I'm curious what that discussion was like internally before the world heard about it, the decision-making process around it.  
我很好奇在向外界公布之前,内部的讨论是什么样的,围绕它的决策过程是怎样的。  
**[59:30] Speaker A:** And just using that as an example to talk about the things that do scare you as we continue to advance and the scaling laws continue to hold.  
就以此为例,谈谈随着我们不断进步、scaling laws 持续有效,有哪些事情确实让你们感到担忧。  
**[59:36] Speaker B:** Yeah. I think one of the things about Mythos is that people maybe misconstrue it as just a cyber model.  
是的。我觉得关于 Mythos,人们可能误解了,以为它只是一个网络安全模型。  
**[59:42] Speaker B:** It is an incredibly capable model across many different dimensions.  
它其实是一个在很多不同维度上都极其强大的模型。  
**[59:47] Speaker B:** What we found was that cyber in particular was a place where it spiked  
我们发现网络安全是它能力特别突出的一个领域。  
**[59:52] Speaker A:** And so this was the first model that we decided to release in a different way, and the way in which we did that again is consistent with our mission, our principles, like we wanted to do it in that way.  
所以这是我们第一个决定以不同方式发布的模型,而我们采用的方式再次符合我们的使命和原则,我们想要这样做。  
**[01:00:02] Speaker A:** And so we have this phased approach to it because we think that when a model is this capable, and again cyber is the thing that people focused on, but you know there are other things as well.  
所以我们采取了分阶段的方法,因为我们认为当一个模型如此强大时——网络安全是人们关注的焦点,但其实还有其他方面。  
**[01:00:11] Speaker A:** We think again it can be used in a positive way, right, to patch code bases.  
我们认为它可以被用于积极的方面,比如修补代码库。  
**[01:00:16] Speaker A:** You've seen these examples where, um, you know, we had an open source code base that, you know, a prior model found 22 security vulnerabilities in and Mythos then found 250.  
你看到过这些例子,我们有一个开源代码库,之前的模型在其中发现了 22 个安全漏洞,而 Mythos 发现了 250 个。  
**[01:00:25] Speaker A:** Uh, so that is kind of scary, right, but that informed the way in which we released it.  
这确实有点吓人,对吧,但这也影响了我们发布它的方式。  
**[01:00:33] Speaker A:** So we didn't say we're never going to release it. We said let's do it in a phased way. Let's do it to, you know, a group that will expand over time where we can, you know, focus on this one cyber capability and how it can actually be used.  
我们没有说永远不发布它。我们说的是分阶段发布,先向一个会随时间扩大的群体发布,这样我们可以专注于这一项网络安全能力以及它实际上可以如何被使用。  
**[01:00:46] Speaker A:** Used positively, you know, in a defensive way as opposed to in an offensive way, and we think that's a template that could be used for the future. But because of this one particular area, we wanted to be cognizant of that in how we released it.  
用于积极的、防御性的方式,而不是攻击性的方式,我们认为这可以作为未来的一个模板。但正是因为这一个特定领域,我们在发布时想要对此保持警觉。  
**[01:00:59] Speaker B:** You're so big now that you run into everything and everyone.  
你们现在规模太大了,会遇到所有的事情和所有的人。  
**[01:01:02] Speaker B:** And one example of this is the government just a couple days ago said maybe there'd be this new system where you have to sort of pre-approve the release of a new model with the government before it was released to the public.  
举个例子,就在几天前,政府表示可能会推出一个新系统,要求在向公众发布新模型之前,必须先获得政府的预先批准。  
**[01:01:11] Speaker B:** Obviously, you had the crazy experience with the Department of Defense, which I'm really curious what that was like as you went through it.  
显然,你在国防部那边有过一段很疯狂的经历,我真的很好奇你当时是怎么度过的。  
**[01:01:15] Speaker B:** Like now everyone cares about this company and this technology and the couple other companies that are building it.  
现在每个人都在关注这家公司、这项技术,以及其他几家正在开发它的公司。  
**[01:01:23] Speaker B:** How do you navigate that stuff? And some of it's just, I guess, beyond your control, but I'm sure you're trying to  
你是如何应对这些事情的?我想有些事情可能超出了你的控制范围,但我相信你在尽力  
**[01:01:30] Speaker A:** Work with people as best you can. Maybe talk about those two examples of like the government now as a very relevant partner, player, you know, overseer, etc.  
尽可能地与人们合作。也许可以谈谈这两个例子,比如政府现在作为一个非常相关的合作伙伴、参与者、监管者等等。  
**[01:01:40] Speaker B:** Yeah. So, I think first and foremost, like we prioritize having a strong relationship on this because we do think that, you know, regulation has a role to play in how these models are developed over time.  
好的。首先,我认为我们优先考虑在这方面建立牢固的关系,因为我们确实认为监管在这些模型随时间发展的过程中应该发挥作用。  
**[01:01:52] Speaker B:** We are very like America first in our approach. We want the technology to support the US as well as, you know, democratic countries around the world.  
我们的做法非常「美国优先」。我们希望这项技术能够支持美国以及世界各地的民主国家。  
**[01:02:00] Speaker B:** And that's one of the reasons why we've been working closely with the administration on something like Stargate.  
这也是我们一直与政府密切合作推进 Stargate 这类项目的原因之一。  
**[01:02:08] Speaker B:** I do think that there's a balance, right? You want to be able to have innovation happen really quickly and have that not be slowed down, but you also want to have this kind of responsibility framework for how these things are deployed because we've long said  
我确实认为需要平衡,对吧?你希望创新能够快速发生而不被拖慢,但同时你也希望有一个责任框架来规范这些技术的部署,因为我们一直在说  
**[01:02:23] Speaker A:** That, you know, this technology has implications and we should have an honest conversation about them, and that includes with the government. And so I think that's, you know, I think the Nitros process is a good example of that.  
这项技术是有影响的,我们应该坦诚地讨论这些影响,包括与政府讨论。所以我认为 Nitros 流程就是一个很好的例子。  
**[01:02:34] Speaker B:** Can you teach us a bit more about the culture, like how you would describe the cultural tenets to your parents or something like this?  
你能多讲讲公司文化吗,比如你会如何向你父母描述这些文化准则?  
**[01:02:41] Speaker B:** And what feels like it really drives most of the culture? I'm especially curious about the writing. You hear often that, you know, Dario publishes these long essays every so often externally. My understanding is he does that way more frequently and there's a lot of writing culture internally. I'm just trying to get a sense of what the culture is like to be in and what makes it the most distinctive from other companies, maybe that you've worked at, or from other companies that are trying to do the same thing. And what's your sense of  
什么真正驱动了大部分文化?我特别好奇写作文化这一块。经常听说 Dario 会不时对外发布长篇文章。据我了解,他在内部发布的频率要高得多,而且内部有很浓厚的写作文化。我只是想了解一下身处其中是什么感觉,以及与你工作过的其他公司,或者与其他试图做同样事情的公司相比,最显著的区别是什么。你对  
**[01:03:11] Speaker A:** Of the differences and the distinctiveness, the culture is a real unique aspect of Anthropic and it is something that, you know, we do talk about externally, but it's different when you're kind of in there living it, and maybe I can tell you a little bit about some of my observations.  
这些差异和独特性有什么看法?文化确实是 Anthropic 非常独特的一面,我们确实会对外谈论它,但当你真正身处其中时感受是不一样的,也许我可以跟你分享一些我的观察。  
**[01:03:26] Speaker A:** First of all, we have seven co-founders, right? That shouldn't work on paper, but it really does in practice.  
首先,我们有七位联合创始人,对吧?这在纸面上看起来不应该行得通,但在实践中确实很有效。  
**[01:03:32] Speaker A:** And I think they've really set the example for the culture and the things that really matter to the company.  
我认为他们真正为公司的文化和真正重要的事情树立了榜样。  
**[01:03:39] Speaker A:** We do a culture interview and it's not some pro forma, you know, thing we do just to kind of check a box.  
我们会进行文化面试,这不是那种走形式、只是为了打勾的事情。  
**[01:03:47] Speaker A:** It is a real part of the evaluation process.  
它是评估流程中真正重要的一部分。  
**[01:03:49] Speaker A:** So somebody could be flying colors on everything else and really, really the smartest person you've met in this role, we won't hire them if they don't pass the culture bar.  
所以即使某人在其他所有方面都表现优异,真的是你在这个岗位上见过最聪明的人,如果他们没有通过文化关,我们也不会录用。  
**[01:03:58] Speaker A:** And the way I would describe it, um, I like that frame. How would you describe it to your parents is  
我会这样描述,嗯,我喜欢这个框架——你会如何向你父母描述它  
**[01:04:02] Speaker A:** It's one incredibly collaborative, and this means that we don't really tolerate the fiefdoms or the sharp elbows or the like "I need to take credit for this."  
首先是极度协作,这意味着我们真的不容忍各自为政、争抢地盘或者「我需要为此邀功」这种行为。  
**[01:04:13] Speaker A:** It's incredibly humble. It's like, you know, our competitors are incredibly capable and success is far from guaranteed.  
极度谦逊。就像,你知道,我们的竞争对手非常有能力,成功远非板上钉钉。  
**[01:04:19] Speaker A:** And I think that's really part of how the company operates. If we reach a milestone and something good happens, there's not confetti on the floor, it's like "what's next?"  
我认为这真的是公司运作方式的一部分。如果我们达成了一个里程碑,发生了好事,地上不会撒彩纸,而是「接下来做什么?」  
**[01:04:30] Speaker A:** And I think it's just that focus on the mission and the alignment that kind of is imbued throughout the culture of the company.  
我认为正是这种对使命的专注和贯穿整个公司文化的一致性。  
**[01:04:36] Speaker A:** The other thing I would say is, you know, there's rigorous debate, right? There's an intellectual openness and intellectual honesty that happens where people question things.  
另一点我想说的是,你知道,有严谨的辩论,对吧?有一种智识上的开放和诚实,人们会质疑事情。  
**[01:04:46] Speaker A:** People will really express a point of view, but then there's dialogue around it that's productive, and then we'll decide on a path forward.  
人们会真正表达自己的观点,然后围绕它进行富有成效的对话,最后我们会决定前进的路径。  
**[01:04:57] Speaker A:** After that happens, there's real alignment. So in something like compute allocation we were talking about before, people might have different perspectives on how to allocate that compute, but they will engage in a thoughtful discussion about where the returns are the highest or the best. And when they do that, you know, and we come to a decision, then there's alignment on it. There's not second-guessing, there's not this kind of politics or fiefdom. The other piece of it is it's remarkably transparent, the culture, right? So  
这之后,就会形成真正的共识。比如我们之前提到的算力分配问题,大家可能对如何分配算力有不同看法,但他们会进行深思熟虑的讨论,探讨哪里能获得最高或最好的回报。当他们这样做并最终做出决定时,就会达成共识。不会有事后质疑,也不会有那种政治斗争或山头主义。另一个方面是,这种文化非常透明。  
**[01:05:23] Speaker A:** Dario gets up in front of the company every two weeks, usually writes a short document, and he talks about, you know, usually three or four topics, and then takes open questions from the company. And these are not softballs, they're not like planted questions, they're just real questions that are on people's minds, and he answers them the best that he can. And it's not a decision-making forum, but it is a way for the company to get a window into how  
Dario 每两周会在全公司面前做分享,通常会写一份简短的文档,讲三四个话题,然后接受公司的开放提问。这些问题不是那种软性问题,也不是事先安排好的,而是员工真正关心的问题,他会尽力回答。这不是一个决策论坛,但它为公司提供了一个窗口,让大家了解  
**[01:05:50] Speaker A:** Leadership is thinking how he's thinking, and there's debate and dialogue in that, and I think that is something that people really value. Like, it is a transparent culture.  
领导层在思考什么、他是如何思考的,其中也会有辩论和对话,我认为这是大家真正看重的。这确实是一种透明的文化。  
**[01:06:00] Speaker A:** It is one where, you know, all seven of the co-founders are still at the company.  
在这种文化下,七位联合创始人全部还在公司。  
**[01:06:04] Speaker A:** The vast majority of the first, you know, 20 to 30 employees are still at the company.  
最早的二三十名员工中,绝大多数也还在公司。  
**[01:06:08] Speaker A:** And I think the culture underpins the reason why we've been able to attract and retain some of the best talent in the industry, right?  
我认为这种文化是我们能够吸引和留住业内顶尖人才的根本原因。  
**[01:06:16] Speaker A:** Because we don't always pay people the most. We have, you know, very competitive compensation packages.  
因为我们并不总是给员工最高的薪酬。我们的薪酬待遇很有竞争力。  
**[01:06:22] Speaker A:** But when you know Meta and others were out, you know, with these huge packages for some of the technical talent across the large language labs, I think we lost two people and other labs lost dozens.  
但当 Meta 和其他公司为大语言模型实验室的技术人才开出巨额薪酬包时,我们只流失了两个人,而其他实验室流失了几十人。  
**[01:06:31] Speaker B:** What parts of the business and the culture—I mean, specifically for researchers—why do you think that stat is true?  
业务和文化的哪些方面——我是说,特别是对研究人员来说——你认为为什么会有这样的数据?  
**[01:06:40] Speaker A:** I think it really is  
我认为这真的是  
**[01:06:43] Speaker A:** Underpinned by the culture, and that's not just something we feel. It's like empirically when you talk to people, it's, you know, I want to have the most impact possible.  
由文化支撑的,这不只是我们的感觉。从实际情况来看,当你和员工交流时,他们会说,我想产生尽可能大的影响。  
**[01:06:50] Speaker A:** I want to work in a place where, um, again, this idea of talent density mattering more than talent mass.  
我想在一个地方工作,在那里,人才密度比人才总量更重要。  
**[01:06:57] Speaker A:** And I want to work in a place that is actually collaborative versus I have to like fight for this one thing and I feel like it wasn't discussed and debated in the right way, or there wasn't transparency around how a decision was made.  
我想在一个真正协作的地方工作,而不是我必须为某件事争斗,感觉它没有以正确的方式被讨论和辩论,或者决策过程缺乏透明度。  
**[01:07:10] Speaker A:** I think that actually really matters because most of our team just wants to do really, really good work and they're attracted to the company for the mission.  
我认为这真的很重要,因为我们团队的大多数人只是想做非常出色的工作,他们是被公司的使命吸引而来的。  
**[01:07:18] Speaker A:** The idea of having an impact on a company like ours that is trying to develop this transformative technology but to do it in a responsible way, I think that that really matters to the people, not just on the research team but across the company, and that we think is a real advantage for us and it's not  
能够在像我们这样的公司产生影响——一家试图开发这种变革性技术但以负责任的方式进行的公司——我认为这对员工来说真的很重要,不仅是研究团队,整个公司都是如此,我们认为这是我们的真正优势,而且这不是  
**[01:07:37] Speaker A:** Something that we take lightly. We have this concept of a race to the top. We want, you know, we don't always have all the right answers. We don't always do everything perfectly, but we want others to look at some of the things we do and maybe emulate some pieces of that and actually have the technology be developed in a better way across the industry.  
我们轻视的东西。我们有一个「向顶端竞赛」的概念。我们并不总是有所有正确答案,也不总是把每件事都做得完美,但我们希望其他人能看到我们做的一些事情,也许模仿其中的某些部分,让整个行业以更好的方式开发这项技术。  
**[01:07:54] Speaker A:** I think people are also really attracted to that as well. Again, not that we have all the answers, but that we can be a part of contributing and leading to how this can go well for humanity.  
我认为人们也非常被这一点吸引。再次强调,不是说我们有所有答案,而是我们可以参与并引领如何让这项技术造福人类。  
**[01:08:05] Speaker B:** As you're having conversations with people internally, what does the frontier feel like to you? I don't just mean the model frontier. I mean like the next couple of rolls of the dice here in building AI in general.  
在你与内部人员的对话中,前沿对你来说是什么感觉?我不只是指模型前沿,我是指在构建 AI 方面接下来的几次尝试。  
**[01:08:21] Speaker A:** Everyone is kind of wise to like these things are powerful. Everyone's using them. It's becoming, it's diffusing. People are becoming accepting of  
每个人都意识到这些东西很强大。每个人都在使用它们。它正在扩散。人们正在接受  
**[01:08:29] Speaker A:** What feels to you like the frontier from the inside?  
从内部来看,什么让你感觉像是前沿?  
**[01:08:31] Speaker B:** I think it's this idea, and again it's because we're focused on enterprise and because we're really trying to change the productivity of knowledge work that's done in the economy.  
我认为是这个想法,这也是因为我们专注于企业,因为我们真的在试图改变经济中知识工作的生产力。  
**[01:08:42] Speaker B:** I think it is towards this vision or this goal of like a virtual collaborator.  
我认为是朝着虚拟协作者这个愿景或目标前进。  
**[01:08:49] Speaker B:** And so think of this as something that has context within your organization that can use all of the tools that are specific to you, whether they be homegrown tools or tools that you purchase, that has memory and the ability to effectively learn from mistakes you've made, but also mistakes that it's made over time.  
可以把它想象成这样的东西:它了解你组织内的上下文,可以使用所有你特有的工具,无论是自研工具还是购买的工具,它有记忆力,能够有效地从你犯过的错误中学习,也能从它自己随时间犯的错误中学习。  
**[01:09:08] Speaker B:** And then the ability to work over a very long time horizon on not just a task but an actual idea.  
然后它能够在很长的时间跨度内工作,不仅是完成一个任务,而是推进一个真正的想法。  
**[01:09:18] Speaker B:** And so what that means for us is the model capability has to continue to grow to support that.  
所以对我们来说,这意味着模型能力必须持续增长来支持这一点。  
**[01:09:22] Speaker B:** And then the products we build on top of it can unlock this.  
然后我们在此基础上构建的产品就能释放这种能力。  
**[01:09:26] Speaker A:** Virtual collaborator that we think can really accelerate knowledge work.  
我们认为这种虚拟协作者能真正加速知识工作。  
**[01:09:31] Speaker A:** But you have to get it in the right form factor.  
但你必须找到合适的产品形态。  
**[01:09:33] Speaker A:** Right. This is where intelligence is not just a single dimension. It's multiple things, but the virtual collaborator kind of combines many of those things, right?  
对。智能不是单一维度的,而是多方面的,虚拟协作者就是把这些方面结合起来,对吧?  
**[01:09:42] Speaker A:** Which is something that's not just generically smart, but is smart for your use cases.  
它不只是泛泛的聪明,而是针对你的具体使用场景的智能。  
**[01:09:45] Speaker A:** And I think again, what we're seeing in coding is something that we expect to see elsewhere.  
我们在编程领域看到的情况,我认为在其他领域也会出现。  
**[01:09:52] Speaker A:** For us, Claude for code has led the way on that as well as much of the business that we have great customers in that are pushing the coding frontier as well.  
对我们来说,Claude for code 在这方面引领了方向,我们也有很多优秀的客户在推动编程前沿的发展。  
**[01:10:04] Speaker A:** But then you also see something like Cowork come along and start to unlock that co-working faster than Claude for code was if you index them to the same point in time.  
但你也看到像 Cowork 这样的产品出现,如果把它们放在同一时间点比较,Cowork 释放协作能力的速度比 Claude for code 还要快。  
**[01:10:14] Speaker A:** That's kind of remarkable because developers are really fast adopters of  
这很了不起,因为开发者对这项技术的接受速度本来就很快。  
**[01:10:17] Speaker A:** this technology. But I think it's because the model capabilities and the products are pushing towards this notion of a virtual collaborator where even our product development today is not done by like one product manager with two engineers shipping something over 3 months.  
但我认为这是因为模型能力和产品都在朝着虚拟协作者的方向发展,现在我们的产品开发已经不是一个产品经理带两个工程师花三个月交付一个东西了。  
**[01:10:32] Speaker A:** It's shipping daily and there's a fleet of agents that are working across the company on a specific task.  
而是每天都在交付,有一整队 agents 在公司内协作完成特定任务。  
**[01:10:40] Speaker A:** Everyone kind of becomes a manager and I think the implications of that and the productivity gain that can come from that when it's in the right form factor is we're very very early in that but the potential for it is incredible, crazy to imagine.  
每个人都变成了管理者,我觉得当产品形态合适时,这种模式的影响和能带来的生产力提升——我们现在还处于非常早期的阶段,但它的潜力是难以想象的,太疯狂了。  
**[01:10:55] Speaker B:** I'm curious how you've had to personally evolve to be able to stay doing this.  
我很好奇你个人是如何进化来持续做这件事的。  
**[01:11:01] Speaker B:** Like you hear a lot about these stories about how the executives have to scale with the company or else they'll get new executives.  
你经常听到这样的故事,说高管必须随着公司成长,否则就会被换掉。  
**[01:11:08] Speaker B:** You know the business that you were at prior to this was a great business but it was a tiny—Cedar was a  
你之前做的业务是个很好的业务,但规模很小——Cedar 是个  
**[01:11:13] Speaker A:** Tiny, tiny fraction of the scale, so like everyone is in this new unprecedented thing.  
非常非常小的规模,所以现在每个人都在经历这种前所未有的事情。  
**[01:11:17] Speaker A:** You talked about the example of like getting out of linear into, you know, into more exponential type thinking.  
你提到了从线性思维转向指数型思维的例子。  
**[01:11:22] Speaker A:** That's one example of what I mean, but how have you managed it personally?  
这就是我说的一个例子,但你个人是怎么应对的?  
**[01:11:27] Speaker A:** Like, what have you had to do? What's been the most painful?  
你必须做什么?什么最痛苦?  
**[01:11:31] Speaker A:** Like, how do you manage your own ability to scale with this thing that's scaling faster than what we've seen before?  
你如何管理自己的能力,让自己能跟上这个增长速度超过以往任何时候的东西?  
**[01:11:38] Speaker B:** Yeah, it's really hard, but I think the important thing is to think in first principles, right?  
这确实很难,但我认为重要的是用第一性原理思考,对吧?  
**[01:11:45] Speaker B:** So this is like everyone has priors when they come to something new.  
每个人面对新事物时都有自己的先验认知。  
**[01:11:50] Speaker B:** Thinking in first principles and having like intellectual openness.  
用第一性原理思考,保持思想上的开放。  
**[01:11:52] Speaker B:** You know, I spent a lot of time with Tom Brown, our chief compute officer.  
我花了很多时间和我们的首席计算官 Tom Brown 交流。  
**[01:11:58] Speaker B:** He was actually one of the first people to interview me at the company, and I remember we went on a walk like, it was a bit before I  
他其实是公司最早面试我的人之一,我记得我们一起散步,那是在我  
**[01:12:06] Speaker A:** It started and we walked around the Mission in San Francisco for two and a half hours, and he started to tell me about his vision for the future of the company, and this is in 2024, early 2024.  
那次见面开始后,我们在旧金山 Mission 区走了两个半小时,他开始跟我讲他对公司未来的愿景,那是在 2024 年,2024 年初的时候。  
**[01:12:16] Speaker A:** And I'll be honest, it sounded crazy.  
说实话,听起来简直疯狂。  
**[01:12:18] Speaker A:** You walked me all the way home, and I remember I came in, I told my wife, I was like, this is going to be wild, like if even 10% of that is true, this is going to bend all paradigms of what not just things I've seen but what most people have seen.  
你一路把我送到家,我记得我进门后跟我妻子说,这将会是疯狂的经历,如果他说的哪怕只有 10% 是真的,这都将打破所有范式——不仅是我见过的,而是大多数人见过的。  
**[01:12:36] Speaker A:** And it turns out that a lot of what Tom said during that walk has come to fruition, but I remember that as like an early formative thing, coming home and being like, holy shit, like this is going to be totally different and new, and you know, a really incredible experience but also a really challenging one.  
结果证明 Tom 在那次散步中说的很多事情都实现了,但我记得那是一个早期的关键时刻,回到家时我想,天哪,这将完全不同,是全新的东西,会是一段非常不可思议的经历,但同时也会非常具有挑战性。  
**[01:12:54] Speaker A:** And that's what it's been.  
事实也确实如此。  
**[01:12:56] Speaker A:** The other piece of this is just hiring great people. You know, I try to hire people and I tell people during the interview process, I'm like, I'm not...  
另一个重要方面就是招聘优秀的人才。我在面试过程中会告诉候选人,我不是……  
**[01:13:02] Speaker A:** Really hiring you as like a direct report of mine. I'm hiring you as a partner and I want you to treat it as a partnership which means that there might be things that you and I disagree on.  
真的把你当作我的直接下属来招聘。我是把你当作合作伙伴来招聘的,我希望你也把这当作一种合作关系,这意味着你我之间可能会有意见不一致的时候。  
**[01:13:12] Speaker A:** I want to hear that and I want to like whiteboard it. I want to understand like you know we've hired people from some of the best companies in the world. They come to this from a different perspective right. They might come to it from a hyperscaler or a large software company or from financial services. In another lifetime you know I worked at Blackstone in the private equity group. Like that training is really valuable and thinking about things at a granular level and not losing that. Like I'm not somebody who is comfortable at 50,000 feet. That's just like not me. But you can't be at 500 feet at everything in this business. There's too much surface area and so having people who can be partners in that is really really critical.  
我想听到那些不同意见,我想在白板上讨论。我想理解——我们从世界上一些最好的公司招来了人才,他们带来不同的视角,对吧。他们可能来自超大规模云服务商、大型软件公司或金融服务行业。在另一段人生中,我曾在 Blackstone 的私募股权部门工作过。那种训练非常宝贵,能让你在细节层面思考问题而不失焦。我不是那种习惯在五万英尺高空俯瞰的人,那不是我的风格。但在这个业务中,你也不可能在所有事情上都保持在 500 英尺的高度。涉及面太广了,所以有能够成为合作伙伴的人真的非常关键。  
**[01:13:50] Speaker A:** I think the last piece is to think about you  
我认为最后一点是要思考  
**[01:13:53] Speaker A:** Know how the business evolves over time and where there might be moments or analogues to things that have happened in the past.  
业务如何随时间演变,以及在哪些时刻可能会出现与过去类似的情况。  
**[01:14:01] Speaker A:** I helped lead the financing that Airbnb did in the middle of the pandemic.  
我曾帮助主导了 Airbnb 在疫情期间的融资。  
**[01:14:07] Speaker A:** Very different situation, right? The business lost 70% of its revenue in seven weeks.  
那是完全不同的情况,对吧?公司在七周内失去了 70% 的收入。  
**[01:14:10] Speaker A:** I know Brian was just—did a show with you.  
我知道 Brian 刚刚——在你的节目上做过访谈。  
**[01:14:14] Speaker A:** That was a harrowing time, but it was also a time kind of without precedent, right?  
那是一段艰难的时期,但也是一段几乎没有先例可循的时期,对吧?  
**[01:14:18] Speaker A:** Where you had to think about things with a clear perspective when it was rapidly changing and there was not a good template or pattern to match.  
在那种情况下,你必须在快速变化中保持清晰的视角来思考问题,而且没有好的模板或模式可以参照。  
**[01:14:28] Speaker A:** And then on a personal level, look, it is hard to balance everything—family and friends—and certainly this job takes a big bite out of all that.  
然后在个人层面,说实话,平衡所有事情——家庭和朋友——是很难的,这份工作确实占据了很大一部分时间。  
**[01:14:39] Speaker A:** But what I do try to do maybe once a week is in a quiet moment just think, wow, this is really cool. It's an incredible opportunity to—  
但我确实会尝试每周一次,在安静的时刻想一想,哇,这真的很酷。这是一个难得的机会——  
**[01:14:50] Speaker A:** Work, you know, with this group of people on this problem at this company at this moment in time. And I try to do that, you know, again, maybe it's in a car ride, maybe it's late at night or something like that.  
能够和这群人一起,在这个时刻,在这家公司,解决这个问题。我会尝试这样做,可能是在开车的时候,可能是深夜或类似的时候。  
**[01:15:01] Speaker A:** Just having that recognition and that appreciation is really important.  
保持这种认知和感激真的很重要。  
**[01:15:07] Speaker B:** What did Tom tell you on the walk that sounded most crazy?  
Tom 在那次散步中告诉你的哪些事情听起来最疯狂?  
**[01:15:11] Speaker A:** I mean, we talked a lot about the scale of the compute infrastructure, what models could do in a short amount of time.  
我们聊了很多关于计算基础设施的规模,以及模型在短时间内能做到什么。  
**[01:15:19] Speaker A:** I think, you know, he described a world that I would have said is kind of sci-fi. But a lot of what we're experiencing here and now have really roots in that conversation.  
我觉得,他描述的世界我当时会认为是科幻小说。但我们现在正在经历的很多事情,其实都源于那次对话。  
**[01:15:29] Speaker A:** And so there's even more things that he talked about that are probably beyond where we are today.  
所以他还谈到了更多可能超出我们今天所处阶段的事情。  
**[01:15:34] Speaker A:** But I think the commonality of it was that, you know, everything is going to happen much quicker than we think and that both the implications but also the  
但我认为共同点在于,一切都会比我们想象的发生得更快,而且这带来的影响以及  
**[01:15:45] Speaker A:** Capabilities of that can change, and then he also had like a really incredible optimism about the future that I think, you know, we talk about internally kind of holding light and shade.  
能力都会发生变化,然后他对未来也有一种非常不可思议的乐观态度,我觉得这就是我们内部所说的「同时持有光明与阴影」。  
**[01:15:57] Speaker A:** That's one of the things we say, and I think like I came from that conversation with just a bunch of questions but also just a sense of positivity about what could happen in the future.  
这是我们常说的一点。我觉得那次对话结束后,我带着一堆问题离开,但同时也对未来可能发生的事情充满了积极的期待。  
**[01:16:06] Speaker B:** It seems like we spent most of our time talking about because it's been the reality that we exist at the high end of that cone.  
看起来我们大部分时间都在讨论这个,因为现实情况是我们一直处于那个锥形区间的高端。  
**[01:16:11] Speaker B:** What can you imagine that would cause that to change to the low end of that cone? Like how, if we were to do like some sort of premortem on a year from now, we're like, "Wow, actually we didn't need nearly as much compute as we thought or something like that." What can you imagine that would shift us meaningfully in that cone?  
你能想象什么情况会让我们转向锥形区间的低端吗?比如说,如果我们现在对一年后做个事前复盘,发现「哇,实际上我们需要的算力远没有想象中那么多」之类的。你能想到什么会让我们在这个区间内发生显著变化?  
**[01:16:29] Speaker A:** I think the first thing would be the diffusion rate within our customers, the use cases are playing  
我认为第一个因素是我们客户内部的扩散速度,也就是实际应用场景的推进情况。  
**[01:16:37] Speaker A:** Catch up to the model capability, and I think, you know, look, these are—we are talking about humans in large organizations with a set of tools and practices and things that they've been doing for a really long time. Change is hard, right? And so to the extent that that diffusion, you know, hits a wall or slows down or something like that, that could affect the kind of rate of change in terms of revenue growth.  
能否跟上模型能力的发展。你要知道,我们面对的是大型组织中的人,他们有一套已经使用了很长时间的工具、实践和做事方式。改变是很难的,对吧?所以如果这种扩散遇到瓶颈、放缓或者类似的情况,就可能影响营收增长的速度。  
**[01:17:01] Speaker A:** Certainly the scaling laws slowing down or not holding—we don't see that. We can't say that with 100% certainty. I think that would be silly. We certainly believe in the trajectory, but the model capabilities leveling off would be another thing.  
当然还有scaling laws放缓或者失效的可能——我们目前没有看到这种迹象。我们不能百分之百确定,那样说太愚蠢了。我们确实相信这个发展轨迹,但如果模型能力增长趋于平缓,那也会是一个影响因素。  
**[01:17:18] Speaker A:** And then, you know, maybe third is just how we think about being at the frontier. You know, today we're at the frontier. I think we're defining the frontier of agentic AI. We need to stay there, right? And it's a competitive  
第三点可能是我们如何看待保持在前沿的问题。现在我们处于前沿,我认为我们正在定义agentic AI的前沿。我们需要保持在那里,对吧?这是一个竞争激烈的市场。  
**[01:17:32] Speaker A:** Market, and we're going to continue to invest in the technology and the compute and the go-to-market that's required to be there, but that's not guaranteed either.  
我们会继续在技术、算力和市场拓展方面投入必要的资源来保持领先,但这也不是板上钉钉的事。  
**[01:17:40] Speaker B:** What are you most excited about? Like, you get to—you have a privileged seat. You sort of get to literally see the future because it's happening inside the business before those outside the business see it.  
你最兴奋的是什么?你的位置很特殊,可以说你能真正看到未来,因为这些变化正在公司内部发生,而外界还看不到。  
**[01:17:49] Speaker B:** With that perspective and in that seat, what are you most excited about in the future?  
站在这个角度,处在这个位置上,你对未来最期待的是什么?  
**[01:17:54] Speaker A:** I really think that the biotechnology and healthcare outcomes that can come from this technology are the things that I'm most optimistic about.  
我真的认为,这项技术能在生物技术和医疗健康成果方面带来的改变,是我最乐观看待的事情。  
**[01:18:05] Speaker A:** We may live in a world where you're diagnosed with a disease that is not curable, but in your lifetime that cure can be found much more rapidly and you actually might not die of that disease.  
我们可能会生活在这样一个世界:你被诊断出患有某种无法治愈的疾病,但在你的有生之年,治愈方法可以被更快地找到,你实际上可能不会死于那种疾病。  
**[01:18:14] Speaker A:** And I think of this as like, you know, a lot of what we're doing today is helping to speed up the drug development process, right? A lot of the paperwork and—  
我觉得这就像,我们现在做的很多工作是帮助加速药物开发过程,对吧?大量的文书工作和——  
**[01:18:25] Speaker A:** Clinical studies, reports, and things like that that are needed to be done—AI and our solutions in particular are helping to rapidly accelerate that.  
临床研究、报告之类需要完成的工作——AI,特别是我们的解决方案,正在帮助快速推进这些工作。  
**[01:18:32] Speaker A:** I'm really most optimistic and excited about when it goes further back into drug development and drug discovery, because, you know, our humans are incredibly capable at research, but if you think about these molecules and proteins, like, they're so complex and such small changes have such big implications for the outcomes—like, AI is perfect for that.  
我真正最乐观和兴奋的是当它进一步深入到药物开发和药物发现阶段,因为人类在研究方面的能力确实很强,但如果你想想那些分子和蛋白质,它们是如此复杂,微小的变化就会对结果产生巨大影响——AI在这方面简直完美。  
**[01:18:54] Speaker A:** If you think about what can happen when the lab's throughput goes up 10x or 100x and we can run that many more experiments, probably get better results faster, and that can be something that helps, you know, people around the world, right? And it doesn't have to be limited to a small set of diseases or disorders—it can really go much further down the chain. And so I think that has the potential to, you know, greatly alter the...  
想想当实验室的产出提高10倍或100倍时会发生什么,我们可以进行更多实验,可能更快得到更好的结果,这能帮助到全世界的人,对吧?而且不必局限于少数几种疾病或病症——它可以延伸到更广泛的领域。所以我认为这有潜力极大地改变……  
**[01:19:19] Speaker A:** Way that we live and the way that we interact, and that's really exciting to me.  
我们的生活方式和互动方式,这真的让我很兴奋。  
**[01:19:24] Speaker B:** I sure hope you're right. It sure seems like we're on that trajectory and it's quite a future to imagine.  
我真心希望你是对的。看起来我们确实在朝那个方向发展,想象一下那样的未来真是令人激动。  
**[01:19:29] Speaker B:** This is so much fun. I feel like we covered so many interesting aspects of the business that, you know, you don't—I don't think you've done this before. So, you know, don't get this amazing perspective.  
这次聊天太有意思了。我觉得我们涵盖了业务的很多有趣方面,你之前应该没做过这种访谈。所以能得到这种精彩的视角真是难得。  
**[01:19:37] Speaker B:** When I do these, I ask the same traditional closing question. What is the kindest thing that anyone's ever done for you?  
每次做这种访谈,我都会问同一个传统的结束问题。别人为你做过的最善良的事情是什么?  
**[01:19:42] Speaker A:** I have a brother who's five and a half years older than me, and we lived in California when he went to college.  
我有个比我大五岁半的哥哥,他上大学的时候我们住在加州。  
**[01:19:48] Speaker A:** He got into everywhere he applied to, and he was going to go to medical school after that.  
他申请的每所学校都录取了他,而且他之后打算去读医学院。  
**[01:19:51] Speaker A:** And so, I didn't know any of this at the time.  
当时我完全不知道这些事。  
**[01:19:53] Speaker A:** So he ended up going to college in-state and he did exceptionally well.  
所以他最后去了州内的大学读书,而且表现非常出色。  
**[01:19:59] Speaker A:** It's kind of years later that I kind of had to pull this out of him.  
多年以后我才从他那里把这件事问出来。  
**[01:20:03] Speaker A:** But, you know, in deciding where to go to college, you know, we were solidly middle class as a family. And, you know, this was like 25-30 years ago.  
但你知道,在决定去哪所大学读书时,我们家当时是标准的中产阶级。而且这是大约25到30年前的事了。  
**[01:20:14] Speaker A:** You know, the financial aid packages weren't, you know, as robust as they are today. And a big factor in his decision, I found out, you know, many, many, many years later, was, you know, wanting to give me the opportunity to go wherever I wanted.  
那时候的助学金项目远没有现在这么完善。我是很多很多年后才知道,他当时做决定的一个重要考虑因素,是想让我将来能有机会去任何我想去的学校。  
**[01:20:28] Speaker A:** Even though, you know, that was 6 years out and who knows how it would turn out. I didn't know that.  
虽然那是6年以后的事,谁也不知道会怎么样。但我当时完全不知道这些。  
**[01:20:34] Speaker A:** And it was something that, you know, 12-year-old me or 13-year-old me would have never really understood. But now, you know, many years later, I think that's something that was incredibly kind and is something that, you know, I still kind of hold with me today.  
这种事情,12岁或13岁的我是永远不会真正理解的。但现在多年过去了,我觉得这是一件非常善良的事,也是我至今仍然铭记在心的。  
**[01:20:46] Speaker B:** Wow, I've done this like 600 times or something. I've never heard an answer like that type. That's awesome and amazing.  
哇,我做这个节目大概做了600次了吧。从来没听过这样的回答。太棒了,太感人了。  
**[01:20:53] Speaker B:** Christian, thanks so much for doing this.  
Christian,非常感谢你来做这期节目。  
**[01:20:54] Speaker A:** With me. Yeah, thanks for having me, Patrick. Really enjoyed it.  
是的,谢谢你邀请我,Patrick。我很享受这次对话。  
**[01:21:00] Speaker B:** You know how small advantages compound over time? That's true in investing and just as true in how you run your company.  
你知道微小的优势是如何随时间累积的吗?这在投资中是真理,在你经营公司的方式上同样如此。  
**[01:21:05] Speaker B:** Your spending system is your capital allocation strategy. Ramp makes it smarter by default. Better data, better decisions, better economics over time. See how at ramp.com/invest.  
你的支出系统就是你的资本配置策略。Ramp让它默认就更智能——更好的数据、更好的决策、随时间推移更好的经济效益。访问ramp.com/invest了解详情。  
**[01:21:15] Speaker B:** As your business grows, Vanta scales with you, automating compliance and giving you a single source of truth for security and risk. Learn more at vanta.com/invest.  
随着你的业务增长,Vanta与你一起扩展,自动化合规流程,为你提供安全和风险管理的单一可信来源。访问vanta.com/invest了解更多。  
**[01:21:24] Speaker B:** Every investment firm is unique and generic AI doesn't understand your process. Rogo does. It's an AI platform built specifically for Wall Street, connected to your data, understanding your process, and producing real outputs. Check them out at rogo.ai/invest.  
每家投资公司都是独特的,通用AI并不理解你的流程。但Rogo理解。它是专为华尔街打造的AI平台,连接你的数据,理解你的流程,并产出真实的成果。访问rogo.ai/invest查看。  
**[01:21:37] Speaker B:** The best AI and software companies from OpenAI to Cursor to Perplexity use Work  
从OpenAI到Cursor再到Perplexity,最优秀的AI和软件公司都在使用Work  
**[01:21:41] Speaker A:** OS to become enterprise ready overnight, not in months. Visit works.com to skip the unglamorous infrastructure work and focus on your product.  
OS在一夜之间做好企业级准备,而不是花费数月时间。访问works.com跳过那些不起眼的基础设施工作,专注于你的产品。  
**[01:21:48] Speaker A:** Ridgeline is redefining asset management technology as a true partner, not just a software vendor.  
Ridgeline正在重新定义资产管理技术,作为真正的合作伙伴,而不仅仅是软件供应商。  
**[01:21:53] Speaker A:** They've helped firms 5x and scale, enabling faster growth, smarter operations, and a competitive edge.  
他们帮助公司实现5倍增长和规模化,实现更快的增长、更智能的运营和竞争优势。  
**[01:21:59] Speaker A:** Visit ridgelineapps.com to see what they can unlock for your firm.  
访问ridgelineapps.com,看看他们能为你的公司释放什么潜力。  

---

## Deep Dive Summary

### Topic 1: Multi-dimensional model intelligence and enterprise capabilities
多维度模型智能与企业能力
_[00:00]_

**Q:** How does the company think about model intelligence differently from traditional IQ metrics?
**问：** 公司如何以不同于传统IQ指标的方式思考模型智能？

**A:** The speaker rejects the conventional single-score IQ approach to evaluating model intelligence, arguing instead that "intelligence for us is multi-dimensional" and focused on "real-world capability." Each new model generation unlocks different capabilities that enable enterprises to accomplish more tasks, execute them better, and operate more efficiently. The company's core business thesis rests on the belief that "returns to frontier intelligence are extremely high" specifically in enterprise contexts, where practical capabilities matter more than benchmark scores.
**答：** 发言人拒绝用单一IQ分数来评估模型智能，而是强调"intelligence for us is multi-dimensional"，关注的是"real-world capability"实际能力。每一代新模型都会带来不同的能力维度，让企业能完成更多任务、做得更好、效率更高。公司的核心商业论点建立在一个信念上："returns to frontier intelligence are extremely high"，尤其是在企业场景中，实用能力比benchmark分数更重要。

### Topic 2: Compute as the lifeblood of the business
算力是公司的生命线
_[01:00]_

**Q:** Why is compute procurement the most consequential decision and what are the risks of getting it wrong?
**问：** 为什么算力采购是最关键的决策，决策失误会带来什么风险？

**A:** Compute procurement represents existential stakes for AI companies because it serves as "the canvas on which everything else gets built" and determines both survival and competitiveness. The decision creates a brutal double bind: buying too much compute leads to bankruptcy, while buying too little means inability to serve customers and falling behind the frontier. This forces companies to navigate a "cone of uncertainty" with highly disciplined planning, since compute cannot be acquired on demand—you cannot "buy a gigawatt of compute and have it delivered next week." The company approaches this through bottoms-up demand modeling, though they acknowledge sometimes getting forecasts wrong, highlighting the inherent difficulty of predicting needs in a rapidly evolving landscape.
**答：** 算力采购对AI公司来说是生死攸关的决策，因为它是"一切构建的画布"，直接决定公司的生存和竞争力。这个决策面临残酷的两难困境：买多了会导致公司倒闭，买少了则无法服务客户、无法保持技术前沿地位。公司必须在"不确定性锥体"中进行高度自律的规划，因为算力无法即时获取——你不可能"下周就买到一gigawatt的算力"。他们采用自下而上的需求建模方法来应对，但也坦承有时会预测失误，这凸显了在快速演进的环境中预判需求的固有难度。

### Topic 3: Disciplined approach to compute planning and procurement
算力规划与采购的严谨方法
_[02:13]_

**Q:** How does the company model demand and build flexibility into compute deals?
**问：** 公司如何建模需求并在算力交易中构建灵活性？

**A:** The company takes a "bottoms up" approach to modeling compute demand while acknowledging forecasting uncertainty, focusing on requirements to "stay at the frontier" of AI capabilities. Flexibility is embedded both in procurement contracts and operational usage patterns, which becomes critical for bridging current capacity to future needs during "exponential" business growth. The strategic importance is evident in the speaker dedicating "30 or 40%" of their time to compute planning, reflecting its centrality to scaling operations efficiently.
**答：** 公司采用自下而上的方式建模算力需求，同时承认预测的不确定性，重点关注保持在AI能力前沿所需的资源。灵活性被嵌入到采购合同和实际使用模式中，这对于在业务指数级增长期间从当前产能过渡到未来需求至关重要。说话者将30-40%的时间投入到算力规划上，体现了其对高效扩展运营的核心重要性。

### Topic 4: Multi-chip platform strategy: Trainium, TPUs, and GPUs
多芯片平台策略：Trainium、TPU 和 GPU
_[02:51]_

**Q:** How does the company use three different chip platforms fungibly across workloads?
**问：** 公司如何在不同工作负载中灵活使用三种芯片平台？

**A:** The company has built an orchestration layer that enables fungible use of Amazon Trainium, Google TPUs, and Nvidia GPUs across model development, internal product acceleration, and customer serving. This flexibility "took us a long time" and required multi-year investment to achieve, positioning them as what the speaker believes are "the most efficient users of compute amongst any of the frontier labs." They strategically match each chip generation to its optimal workload, a capability that initially drew skepticism when they adopted TPUs at scale while "everyone's using GPUs." The orchestration approach maximizes compute value by treating different hardware platforms as interchangeable resources rather than being locked into a single vendor's ecosystem.
**答：** 该公司构建了一个编排层，能够在模型开发、内部产品加速和客户服务等场景中灵活调用 Amazon Trainium、Google TPU 和 Nvidia GPU。这种灵活性"took us a long time"，需要多年投入才实现，使他们成为说话者认为的"frontier labs 中计算资源利用效率最高的"。他们会根据每代芯片的特点匹配最适合的工作负载，这种能力在早期采用 TPU 时曾引发质疑，因为当时"everyone's using GPUs"。这种编排方式将不同硬件平台视为可互换资源，而非锁定单一供应商生态，从而最大化计算价值。

### Topic 5: Building close to bare metal and chip collaboration
底层硬件优化与芯片协作
_[04:14]_

**Q:** How does the company work with chip manufacturers and build custom compilers for efficiency?
**问：** 公司如何与芯片制造商合作并构建定制编译器以提高效率？

**A:** The company takes a collaborative approach with Amazon's Annapurna Labs team to influence chip roadmaps, believing their workloads are "really stressing the limits of what these chips are capable of." This partnership enables them to maximize compute ROI by ensuring "a dollar of compute inside our organization goes further" than elsewhere. They build custom compilers and develop "from the chip level up" to achieve the customization and flexibility needed to utilize each chip for its optimal purpose within the organization.
**答：** 公司与Amazon的Annapurna Labs团队紧密合作来影响芯片路线图，因为他们的工作负载正在"真正挑战这些芯片的极限能力"。这种合作让他们能够最大化计算投资回报，使得"组织内每一美元的算力"比其他地方发挥更大作用。他们构建定制编译器，"从芯片层面开始"向上开发，以获得所需的定制化和灵活性，让每个芯片在组织内发挥最佳用途。

### Topic 6: The cone of uncertainty in exponential growth
指数增长中的不确定性锥
_[05:15]_

**Q:** What is the cone of uncertainty and how does exponential growth create wide outcome ranges?
**问：** 什么是不确定性锥，指数增长如何创造广泛的结果范围？

**A:** The cone of uncertainty describes how small variations in growth rates compound into vastly different outcomes when a business grows exponentially, making prediction extremely difficult because "humans mostly think linearly" rather than exponentially. The speaker explains they examine multiple scenarios across this widening cone over 1-2 years and work backwards to ensure they can "still be at the frontier" while serving customers and providing internal compute to employees. The critical risk is being caught at one point in the cone while having only provisioned compute for a different point, which is why compute efficiency has been essential to their planning discipline.
**答：** 不确定性锥描述了在指数增长中，增长率的微小变化会复合成截然不同的结果，预测变得极其困难，因为"人类大多是线性思考"而非指数思考。Speaker 解释说他们会研究1-2年内这个不断扩大的锥体中的多种场景，然后反向规划，以确保能够"保持在前沿"，同时服务客户并为员工提供内部算力。关键风险在于实际处于锥体的某一点，但只为另一点准备了算力，这就是为什么计算效率对他们的规划纪律至关重要。

### Topic 7: Compute allocation across training, research, and serving
训练、研究和服务之间的算力分配
_[06:58]_

**Q:** How does the company allocate compute between model development, internal use, and customer serving?
**问：** 公司如何在模型开发、内部使用和客户服务之间分配算力？

**A:** The company maintains a strict floor on compute allocated to model development that "we will not go below," even if it creates challenges for customer serving, because they believe "the returns to frontier intelligence are extremely high" particularly in enterprise contexts. Internal compute use serves as an accelerator for model development by helping teams discover "compute efficiency multipliers that really get us more from each dollar," creating a virtuous cycle of optimization. Allocation decisions happen through collaborative, "not zero-sum" discussions where each team presents their use case and ROI, with the flexibility to "make adjustments on a relatively short time horizon" due to dynamic reallocation capabilities.
**答：** 公司为模型开发设定了算力分配的底线，即使影响客户服务也"不会低于这个水平"，因为他们认为前沿智能的回报极高，尤其在企业市场。内部算力使用通过帮助团队发现"compute efficiency multipliers"来加速模型开发，形成优化的良性循环。分配决策通过协作式、非零和的讨论进行，各团队展示使用场景和ROI，并且由于动态分配能力，可以在相对短的时间内做出调整。

### Topic 8: Measuring and improving compute efficiency over time
衡量和提升算力效率
_[08:34]_

**Q:** How do new model generations achieve both capability improvements and efficiency gains?
**问：** 新一代模型如何同时实现能力提升和效率增益？

**A:** Unlike the car analogy where upgrading from a sedan to a sports car sacrifices fuel efficiency, AI model generations achieve simultaneous improvements in both capability and efficiency, with each leap from Opus 4 to 4.5, 4.6, and 4.7 delivering "a multiplier in terms of how much more efficient it is at processing tokens." This creates a compounding benefit where more efficient inference not only serves customers better but also accelerates internal processes like reinforcement learning, which is "basically inference within a sandbox with a reward function." Between major model releases, the team continuously deploys incremental efficiency improvements, creating a system where research advances in model capabilities, compute efficiency, and serving infrastructure all reinforce each other.
**答：** 与汽车升级会牺牲燃油效率不同，AI模型的迭代在能力和效率上同时提升，从Opus 4到4.5、4.6、4.7的每次跃升都带来了"处理token效率的倍数级提升"。这种双赢局面不仅让客户获得更强能力，还加速了内部流程，比如强化学习本质上是"在沙盒环境中带奖励函数的推理"，模型推理越高效，RL训练也越快。在大版本发布之间，团队持续部署渐进式效率优化，形成了模型能力研发、算力效率提升和服务基础设施改进相互促进的飞轮效应。

### Topic 9: Returns to Being at the Frontier
前沿模型的回报优势
_[12:22]_

**Q:** Why are the returns to being at the frontier so high, and how do new model releases unlock more TAM and use cases?
**问：** 为什么前沿模型能带来如此高的回报？新模型发布如何解锁更多市场空间和应用场景？

**A:** The speaker argues that frontier model intelligence is multi-dimensional rather than a single IQ score, encompassing capabilities like "long horizon tasks," tool use, and speed—comparing it to employees with equal capability where one completes work in a day versus a week, making them "seven times better." Each model generation unlocks new TAM as enterprise customers "are always pushing the limits," evidenced by their company's growth from "$9 billion of run rate revenue" to "north of $30 billion" in four months driven by model releases. The speaker emphasizes this dynamic is "unique to enterprise" where customers immediately invest in more tokens with newer models, creating a reinforcing cycle that shows "the returns to frontier intelligence are not slowing down."
**答：** 前沿模型的智能是多维度的，不只是单一的智商分数，而是包括长时程任务处理、工具使用和速度等能力——就像两个能力相当的员工，一个用一周完成的工作另一个一天就能完成，后者的效率是前者的七倍。每一代新模型都会解锁新的市场空间，因为企业客户总是在突破模型能力的极限，这体现在公司四个月内从90亿美元年化收入增长到超过300亿美元。这种动态在企业市场尤为明显，客户会立即在新模型上投入更多token使用量，形成持续增强的循环，而且这种前沿优势并没有放缓的迹象。

### Topic 10: Recursive Self-Improvement and the Frontier Gap
递归自我改进与前沿差距
_[15:20]_

**Q:** How does recursive self-improvement work in practice, and will it widen the gap between frontier and non-frontier models?
**问：** 递归自我改进在实践中如何运作，它会扩大前沿模型与非前沿模型之间的差距吗？

**A:** Recursive self-improvement is now operational at major labs, where models actively contribute to building the next generation—at Anthropic, "90 plus percent of our code is actually written by Claude," including Claude itself. The company views this as justification for allocating internal compute rather than maximizing revenue, because models "helping us to build that next generation" create compounding advantages when combined with top talent. The distinction that matters isn't "closed or open" but "frontier or not," as frontier models capture economic value and benefit from this recursive acceleration, while the gap naturally widens through the combination of scaling laws, self-improvement loops, and talent leveraging the best tools.
**答：** 递归自我改进已在主要实验室实现，模型正在积极参与构建下一代——在 Anthropic，"90% 以上的代码实际上是由 Claude 编写的"，包括 Claude 自身。公司将此视为分配内部算力而非追求短期收入最大化的理由，因为模型"帮助我们构建下一代"会在与顶尖人才结合时产生复利优势。真正重要的区分不是"闭源或开源"而是"前沿或非前沿"，因为前沿模型能够捕获经济价值并受益于这种递归加速，而差距会通过 scaling laws、自我改进循环和人才使用最佳工具的组合自然扩大。

### Topic 11: Product Development Velocity and Model Leverage
产品开发速度与模型杠杆
_[17:18]_

**Q:** How are models being used to accelerate product development, and what does it mean when Claude writes its own code?
**问：** 模型如何被用来加速产品开发，当Claude编写自己的代码时意味着什么？

**A:** Speaker A emphasizes that product velocity has dramatically increased, with "30 different product and feature releases in January" enabled by combining talented engineers with models to "accelerate ways to access this underlying intelligence." When asked about the logical endpoint where AI autonomously determines what to build without human direction, Speaker A deflects by grounding the discussion back to fundamentals, stating "the core of our company is still a research lab." This suggests they see current model leverage as augmenting human talent rather than replacing strategic product direction, though the question of full autonomy remains unresolved.
**答：** Speaker A 强调产品速度显著提升，仅一月份就有"30 different product and feature releases"，这得益于工程师人才与模型结合来"accelerate ways to access this underlying intelligence"。当被问及 AI 是否会自主决定构建什么产品而无需人类指导时，Speaker A 将话题拉回基本面，强调"the core of our company is still a research lab"。这表明他们将当前的模型杠杆视为增强人类能力而非替代战略产品方向，但完全自主化的问题仍未明确回答。

### Topic 12: The Role of Human Talent in AI Research
人类研究人才在AI时代的核心作用
_[18:06]_

**Q:** What is the role of human research talent when models can do so much, and how does talent density beat talent mass?
**问：** 当AI模型能力如此强大时，人类研究人才的作用是什么？为什么人才密度比人才规模更重要？

**A:** The speaker emphasizes that their company remains fundamentally "a research lab" conducting experiments that push model limits, with this research engine being "upstream of everything else." While models are increasingly helpful in the research process, human talent remains essential for "set[ting] the direction" and identifying "new areas of discovery," with AI tools serving to accentuate and accelerate rather than replace this talent. The philosophy of "talent density beats talent mass" drives their strategy to build "the densest collection of AI research talent and inference engineering talent" working with the best models, creating what they believe is "a really winning combination."
**答：** 讲者强调公司的核心仍然是一个研究实验室，在做突破模型极限的实验，这个研究引擎是所有其他工作的上游基础。虽然模型在研究过程中越来越有帮助，但人类研究人才在"设定方向"和发现新领域方面仍然不可替代，AI工具的作用是放大和加速人才的能力而非取代。他们信奉"人才密度胜过人才规模"的理念，追求打造最密集的AI研究人才和推理工程人才团队，配合最好的模型，形成制胜组合。

### Topic 13: Internal View of Scaling Laws
内部如何衡量Scaling Laws
_[19:31]_

**Q:** How are scaling laws discussed and measured internally across pre-training, post-training, and reasoning?
**问：** 团队内部如何在预训练、后训练和推理阶段讨论和衡量Scaling Laws？

**A:** The team evaluates scaling laws by comparing models at different development stages through their "loss curves" during pre-training runs, which serves as a proxy for "model capability." This comparative approach allows them to benchmark new models against prior ones at equivalent training points. The same methodology extends to reinforcement learning (RL) phases, suggesting a unified framework for tracking capability improvements across the entire training pipeline.
**答：** 团队通过对比不同开发阶段模型的loss curves来评估Scaling Laws，这些曲线能反映模型能力的变化。具体做法是在预训练过程中，将当前模型与之前的模型在相同训练节点进行对比。这套方法同样适用于RL阶段，说明他们用统一的框架来追踪整个训练流程中的能力提升。

### Topic 14: Customer feedback loop and training targets
客户反馈循环与训练目标
_[19:48]_

**Q:** How does customer feedback inform model training and improvement?
**问：** 客户反馈如何影响模型训练和改进？

**A:** Customer pain points directly become "training targets" for model improvement, though the company maintains a strict policy of not training on enterprise customer data (prosumer data is only used with opt-in consent). When customers identify capability gaps or request features beyond current model performance, the team encourages them to "build your product for that" with the promise that R&D will improve those capabilities over time. The feedback loop operates through continuous internal evaluation of different model snapshots against both internal benchmarks and customer-reported experiences, creating a connected cycle between customer needs and model development priorities.
**答：** 客户遇到的痛点会直接转化为模型的训练目标，但公司严格遵守不使用企业客户数据训练的原则（个人用户数据仅在用户主动选择的情况下使用）。当客户提出能力需求或反馈模型性能不足时，团队会鼓励他们先基于现有能力构建产品，同时承诺 R&D 团队会持续改进这些能力。这种反馈循环通过内部持续评估不同模型快照、对比内部基准和客户实际体验来运作，在客户需求和模型开发优先级之间形成闭环。

### Topic 15: Scaling laws continuation and exponential thinking
Scaling laws 持续性与指数思维
_[20:41]_

**Q:** Are scaling laws still holding, and how do you think exponentially about capability growth?
**问：** Scaling laws 是否仍然有效，如何以指数方式思考能力增长？

**A:** The speaker confirms that scaling laws show no signs of slowing down, despite their team's skeptical and scientifically rigorous culture that "constantly challenges previously held assumptions." To handle exponential capability growth rather than linear thinking, they approach planning through "scenarios" instead of point estimates and maintain "a very low bar for updating your current prior," recognizing that quarterly forecasting cycles are inadequate for their dynamic business. They use coding as a concrete pattern recognition case study, where the "remarkable jump in capability" with Sonnet 3.5/3.6 led to measurable adoption and revenue increases, which now serves as an analog for predicting impacts across other domains and understanding how capability improvements translate to market expansion.
**答：** 发言者确认 scaling laws 没有放缓迹象，尽管团队文化非常严谨且"不断挑战先前的假设"。为了应对能力的指数级增长而非线性思维，他们采用"情景规划"而非点估计，并保持"极低的更新门槛"来调整认知，因为传统的季度预测周期已不适用于这种动态业务。他们将 coding 作为具体的模式识别案例，Sonnet 3.5/3.6 带来的"能力显著跃升"转化为可衡量的采用率和收入增长，现在用这个案例类比预测其他领域的影响，理解能力提升如何转化为市场扩张。

### Topic 16: xAI partnership and compute acquisition strategy
xAI 合作与算力获取策略
_[22:37]_

**Q:** How does Anthropic approach partnerships and compute deals across different timeframes?
**问：** Anthropic 如何处理不同时间范围的合作伙伴关系和算力交易？

**A:** Anthropic employs a unified evaluation framework for compute acquisitions across all time horizons, assessing whether they can "deploy that compute productively" based on economic return, pricing, duration, location, compute type, and operational efficiency. The strategy operates as a "layer cake of compute" with staggered deployment timelines: near-term opportunistic deals like the xAI Colossus facility in Memphis for consumer expansion, and massive long-term commitments including 5 gigawatt deals with Google/Broadcom for TPUs starting 2027 and Amazon for Trainium totaling over $100 billion. While the assessment criteria remain consistent—focusing on "price performance over time" and business fit—the key differentiator is simply the time horizon, with near-term compute becoming proportionally smaller as their overall compute base expands.
**答：** Anthropic 对所有时间跨度的算力采购采用统一评估框架，核心是判断能否"高效部署这些算力"，依据经济回报、定价、使用期限、地理位置、算力类型和运营效率来决策。整体策略呈现"layer cake"式的分层结构：短期机会型合作如 xAI Memphis 的 Colossus 设施用于拓展消费者市场，长期大规模承诺包括与 Google/Broadcom 的 5 gigawatt TPU 协议（2027年起）和与 Amazon 的 Trainium 协议，总投入超千亿美元。虽然评估标准一致——聚焦"price performance over time"和业务契合度——但关键区别在于时间维度，且随着算力基数增长，短期算力占比会逐渐缩小。

### Topic 17: Compute trade-offs: cost, speed, and performance
算力权衡：成本、速度与性能
_[24:46]_

**Q:** How do you evaluate different chip platforms and balance price-performance with speed?
**问：** 如何评估不同芯片平台并平衡性价比与速度？

**A:** Anthropic evaluates compute across "three different chip platforms" with multiple generations (TPU V5e/V6/V7, Trainium 2/3), each positioned differently on the "price performance curve." The assessment is highly granular, matching specific chip capabilities to use cases—"CPUs for RL" or leading-edge compute for their "best and fastest models"—with the compute team collaborating across the business to determine "what is each chip best for." While "price performance is important because of efficiency," speed matters critically for certain applications, and the decision framework balances customer demand with technical fit at a detailed level of "what it can deliver for us and when."
**答：** Anthropic 在评估算力时会考察三大芯片平台的多代产品（TPU V5e/V6/V7、Trainium 2/3），每种芯片在性价比曲线上的位置各不相同。评估过程非常细致，会将特定芯片能力与具体用例匹配——比如用 CPU 做强化学习，用前沿算力训练最优最快的模型。虽然性价比对效率很重要，但速度对某些应用场景同样关键，最终决策框架会在客户需求和技术适配之间找平衡，精确到每种芯片能在何时提供什么能力。

### Topic 18: Compute consumption capacity and deployment speed
算力消耗能力与部署速度
_[26:08]_

**Q:** How quickly could Anthropic absorb and deploy significantly more compute?
**问：** Anthropic 能以多快的速度吸收和部署更多算力？

**A:** Anthropic is currently "constrained across those use cases" for compute—training, internal operations, and customer demand—and could now deploy additional compute "very rapidly" if it became available. The speaker notes that a year or two ago, absorbing heterogeneous compute would have been harder due to "idiosyncrasies" in different chip platforms, but today the company has developed the capability to "spin up very quickly and deploy almost any type of compute." This operational flexibility in handling diverse hardware is viewed as "a real advantage," allowing them to maintain their current allocation strategy across use cases while scaling quickly when new resources arrive.
**答：** Anthropic 目前在训练、内部运营和客户需求这几个方面都受到算力限制，如果有额外算力可用，现在能够"非常快速地"部署。发言人提到，一两年前吸收异构算力会比较困难，因为不同芯片平台存在"特殊性"，但如今公司已经具备了"快速启动并部署几乎任何类型算力"的能力。这种处理多样化硬件的运营灵活性被视为"真正的优势"，使他们能够在获得新资源时快速扩展，同时保持现有的跨用例分配策略。

### Topic 19: Platform vs application layer strategy
平台层与应用层策略
_[27:37]_

**Q:** How does Anthropic balance being a platform provider versus building its own applications?
**问：** Anthropic 如何平衡作为平台提供商与构建自有应用？

**A:** Anthropic's strategy is "mostly horizontal" platform-focused, analogous to "the early days of AWS," where they provide model intelligence through various access vectors like prompt caching, cloud code, and managed agents. They selectively build vertical applications in three scenarios: demonstrating future model capabilities (like Claude Code being "Claude-led" rather than developer-led), showing ecosystem value through domain-specific offerings (financial services, life sciences, security), and maintaining a "level playing field" by building on the same platform as customers. The approach assumes that while the platform accrues significant value, customers building on top will capture even more, with partnerships ensuring collaborative rather than competitive dynamics.
**答：** Anthropic 的策略是"以平台为主"的横向布局，类似"AWS 早期"，通过 prompt caching、cloud code、managed agents 等多种方式提供模型能力访问。他们在三种情况下会选择性地构建垂直应用：展示未来模型能力（如 Claude Code 是"Claude 主导"而非开发者主导）、通过领域特定产品（金融服务、生命科学、安全）为生态系统示范价值、以及通过与客户使用同一平台来维持"公平竞争环境"。这一策略假设平台本身会积累大量价值，但基于平台构建的客户会获得更多价值，并通过合作伙伴关系确保协作而非竞争关系。

### Topic 20: Customer concerns about competition with Anthropic
客户对与 Anthropic 竞争的担忧
_[30:53]_

**Q:** How does Anthropic address customer fears about competing with the platform they build on?
**问：** Anthropic 如何应对客户担心在其平台上构建产品时会与平台方竞争的问题？

**A:** The speaker acknowledges that the rapid pace of AI development creates inherent tension, noting that capabilities evolving in "months" rather than "years" means "people are also surprised" by new releases just as Anthropic is. Rather than denying competitive concerns, they emphasize a "partner-oriented" approach through early access programs and close customer collaboration to understand desired capabilities. The core challenge is that model capabilities "sometimes even surprise us," making it difficult to fully predict or prevent moments when customers realize "that's way more powerful than I thought it would be."
**答：** 发言人承认 AI 快速发展带来的固有张力，能力演进从"年"缩短到"月"，导致"人们也会感到惊讶"，就像 Anthropic 自己一样。他们没有回避竞争担忧，而是强调通过早期访问计划和密切客户合作来采取"partner-oriented"的方式，了解客户需求。核心挑战在于模型能力"有时甚至让我们自己都惊讶"，很难完全预测或避免客户意识到"比我想象的强大得多"的时刻。

### Topic 21: Making AI capabilities accessible to customers
让前沿AI能力更易用
_[32:20]_

**Q:** How does Anthropic approach making frontier AI capabilities accessible and valuable to customers?
**问：** Anthropic如何让前沿AI能力对客户更易获取和有价值?

**A:** Anthropic's strategy centers on making frontier capabilities "really accessible" to drive customer value, particularly for those who are "forward-footed" in adoption. The company emphasizes that customers building with their platform tools can be actively accelerated by Anthropic's support. While acknowledging that some accessibility challenges are inherent to "frontier model development," Anthropic distinguishes its approach as "more partner-oriented" than typical industry practices, suggesting a collaborative rather than transactional relationship with customers.
**答：** Anthropic的策略核心是让前沿能力"really accessible"，为客户创造价值，尤其是那些积极采用的"forward-footed"客户。公司强调使用其平台工具构建的客户可以得到Anthropic的加速支持。虽然承认一些可访问性挑战是"frontier model development"固有的，但Anthropic将自己的方法定位为"more partner-oriented"，暗示与客户建立协作而非交易关系。

### Topic 22: AI pricing dynamics and why prices aren't rising
AI定价动态及价格为何不上涨
_[32:52]_

**Q:** Why doesn't Anthropic raise prices significantly despite compute constraints, and how do they think about pricing strategy?
**问：** 尽管算力受限，Anthropic为何不大幅提价，他们如何思考定价策略？

**A:** Anthropic prioritizes pricing stability and accessibility over maximizing short-term margins because they view the market as being in "very, very early innings" and want to "proliferate this throughout the ecosystem." Rather than raising prices during compute constraints, they actually lowered Opus pricing when launching 4.5 because the model was "underutilized relative to its capability"—customers were forcing "Opus problems into Sonnet workloads." This strategy triggered a "Jevons paradox" where lower prices drove consumption "way, way more than expected," ultimately generating more value for both customers and Anthropic. The company maintains that customers are already "generating a ton of ROI" and that accessible pricing enables broader adoption across startups to enterprise, which serves their long-term goal better than price optimization in a supply-constrained environment.
**答：** Anthropic优先考虑价格稳定性和可及性，而非短期利润最大化，因为他们认为市场仍处于"非常早期阶段"，希望"在整个生态系统中推广"这项技术。在算力受限的情况下，他们不仅没有提价，反而在推出Opus 4.5时降低了价格，因为该模型"相对于其能力被低估使用"——客户在用Sonnet的工作负载硬套Opus级别的问题。这一策略触发了"Jevons悖论"：更低的价格驱动消费量"远超预期"，最终为客户和Anthropic创造了更多价值。公司认为客户已经"获得了大量ROI"，可及的定价能促进从初创企业到大型企业的广泛采用，这比在供应受限环境中优化价格更符合长期目标。

### Topic 23: Margins and capital intensity in frontier AI labs
前沿AI实验室的利润率和资本密集度
_[35:57]_

**Q:** How does Anthropic think about margins and return on compute spend given the capital-intensive nature of frontier AI?
**问：** 考虑到前沿AI的资本密集特性，Anthropic如何思考利润率和算力支出回报？

**A:** Anthropic measures success through "return on compute spend" across all workloads rather than traditional margin-based pricing, viewing compute as supporting revenue across different time horizons—from immediate inference serving to model development that "unlocks TAM" months later. The company emphasizes their returns are currently "robust" and notes that compute capacity is determined by "a ramp that might have been determined 12 months ago," making the traditional software paradigm of variable costs per customer inadequate. This means revenue growth (like their Q1 performance) doesn't require proportional compute additions, as the fixed compute envelope governs both short-term and long-term revenue generation across training, serving, and product development.
**答：** Anthropic不采用传统的利润率定价思路，而是用"算力支出回报"来衡量所有工作负载的效益，将算力视为支撑不同时间尺度收入的投资——从即时的推理服务到数月后"unlocks TAM"的模型开发。公司强调当前回报"robust"，并指出算力容量由"12个月前确定的ramp"决定，这使得传统软件按客户边际成本的范式不适用。这意味着收入增长（如Q1表现）不需要同比例增加算力，因为固定的算力envelope同时支撑训练、服务和产品开发，驱动短期和长期收入。

### Topic 24: Partnerships with cloud and chip providers
与云服务和芯片供应商的合作关系
_[38:03]_

**Q:** What does Anthropic need from compute providers, and how do their partnerships with Amazon, Google, Microsoft, Broadcom, and Nvidia work?
**问：** Anthropic需要算力供应商提供什么，他们与Amazon、Google、Microsoft、Broadcom和Nvidia的合作如何运作?

**A:** Anthropic views its compute partnerships as fundamentally deeper than "just procurement," emphasizing that they are "multifaceted" collaborations spanning chip development, capacity planning, serving infrastructure, and customer distribution. The company holds a unique position as "the only model that's on all three clouds" and "the only language lab that's using all three of these chip platforms," which enables diversified access to compute resources. A concrete example is their relationship with Amazon, where Anthropic's teams are "deeply embedded with the Annapurna Labs team" and work as "really good users of Trainium," jointly planning capacity and co-developing chip capabilities. The speaker frames the three major clouds as "great distribution engines" that complement Anthropic's "robust first-party business," suggesting these partnerships serve both infrastructure and go-to-market functions.
**答：** Anthropic认为与算力供应商的合作远不止是简单的采购关系，而是涵盖芯片开发、容量规划、服务部署和客户分发的"multifaceted"深度协作。公司在市场上处于独特地位，是唯一同时部署在三大云平台上的模型，也是唯一使用全部三种芯片平台的语言实验室，这让他们能够多元化获取算力资源。以Amazon为例，Anthropic团队与Annapurna Labs深度合作，作为Trainium的重要用户共同规划容量并参与芯片能力开发。三大云平台不仅提供基础设施，还充当"great distribution engines"，与Anthropic自有的第一方业务形成互补。

### Topic 25: Internal use of Claude at Anthropic's finance team
Anthropic财务团队内部使用Claude
_[39:26]_

**Q:** How does Anthropic's finance team use Claude and Claude Code internally to transform their workflows?
**问：** Anthropic的财务团队如何在内部使用Claude和Claude Code来改变工作流程?

**A:** Anthropic's finance team has evolved from early "vibe coding" with Claude Code into a productionized system that now generates statutory financial statements for all legal entities and produces monthly financial reviews that are "90 to 95% ready" with minimal human intervention. The team built a library of over 70 finance-specific skills accessible through a common repository, enabling them to shift focus from "what exactly happened" to strategic implications, reducing weekly report production from hours to 30 minutes. Notably, the heaviest users are senior leaders like the head of tax who automates policy engines, not just junior employees, demonstrating that the tools fundamentally change how experienced professionals work rather than simply serving as coding assistants for technical staff.
**答：** Anthropic财务团队从早期用Claude Code进行"vibe coding"发展到现在的生产化系统,能够为所有法律实体生成法定财务报表,并产出"90到95%就绪"的月度财务审查报告。团队建立了一个包含70多个财务专用技能的共享库,使他们能够将重心从"到底发生了什么"转向战略影响分析,将周报制作时间从数小时缩短到30分钟。值得注意的是,使用量最大的是税务主管这样的高级领导者,他们用来自动化税务政策引擎,而不仅仅是技术背景的年轻员工,这表明工具从根本上改变了资深专业人士的工作方式。

### Topic 26: Human concerns about AI-driven decision making
人类对AI驱动决策的担忧
_[43:46]_

**Q:** Does it concern us that humans are increasingly following AI recommendations across various domains?
**问：** 人类在各个领域越来越多地遵循AI建议，这是否令人担忧？

**A:** The speaker expresses unease about the growing trend where "we just start doing the stuff that AI tells us to do" across domains like sales and calendar management. However, they acknowledge a potential counterargument: AI might be "such a better coordinator" and "optimizer than we ever could be" that following its recommendations could actually be beneficial. The question remains open-ended, reflecting genuine uncertainty about whether increasing AI dependence is concerning or simply rational given AI's superior optimization capabilities.
**答：** 说话者对人类"开始做AI让我们做的事情"这一趋势表示担忧，涉及销售、日程管理等多个领域。但他们也承认另一种可能：AI作为"更好的协调者"和优化器，其能力可能远超人类，因此听从它的建议也许是合理的。这个问题保持开放性，反映了对AI依赖度增加究竟是令人担忧还是理性选择的真实不确定性。

### Topic 27: AI autonomy and human-AI dynamics
AI 自主性与人机关系的转变
_[44:08]_

**Q:** What are the implications of AI systems becoming more directive rather than just responsive?
**问：** 当 AI 系统从被动响应转向主动指导时，会带来什么影响？

**A:** Speaker A expresses ambivalence about the shift toward directive AI systems, describing the experience as "ever so slightly dystopian" despite acknowledging its practical utility. The concern centers on a fundamental reversal in agency—moving from "me telling it what to do" to "doing what it tells me"—which raises questions about human autonomy in AI-assisted workflows. While A finds these systems "kind of cool" and "helpful" in practice, the underlying unease suggests tension between efficiency gains and preserving human decision-making authority.
**答：** Speaker A 对 AI 系统变得更具指导性持矛盾态度，尽管承认其实用价值，但仍觉得"有点反乌托邦"。核心担忧在于主导权的根本性逆转——从"我告诉它做什么"变成"我按它说的做"——这引发了关于人类自主性的疑问。虽然 A 认为这些系统"挺酷的"且"有帮助"，但潜在的不安表明，效率提升与保持人类决策主导权之间存在张力。

### Topic 28: AI as productivity accelerant and Jevons paradox for labor
AI 作为生产力加速器与劳动力的 Jevons 悖论
_[44:28]_

**Q:** How does AI increase productivity and what is the paradox of hiring more people as productivity increases?
**问：** AI 如何提高生产力，以及随着生产力提高反而雇佣更多人的悖论是什么？

**A:** Speaker A describes AI as "an accelerant to our productivity" that paradoxically leads to hiring more people rather than fewer, drawing an analogy to Jevons paradox applied to labor. The productivity gains from Claude free employees from low-value tasks like "trying to reconcile some number" or lengthy book-closing processes, allowing them to focus on strategic work like reinvesting in the business and dynamic resource allocation. Rather than reducing headcount, this creates capacity to tackle more ambitious work, so "there's no shortage of work to do" and the company has "hired a lot more people" who themselves become more productive as they learn to use the AI tools.
**答：** Speaker A 认为 AI 是生产力加速器，但会产生类似 Jevons 悖论的效应——生产力提升反而导致公司雇佣更多人。Claude 把员工从低价值工作（如核对数字、关账）中解放出来，让他们能专注于战略性思考，比如业务再投资和动态资源配置。这种效率提升不是减少人力，而是创造了处理更多工作的能力，所以公司反而招了更多人，而新员工在学会使用 AI 工具后生产力也会提升。

### Topic 29: Investor relations and capital raising journey
投资者关系与融资历程
_[45:38]_

**Q:** What has the fundraising experience been like and how have investor perceptions evolved?
**问：** 融资经历是怎样的，投资者的看法如何演变？

**A:** Anthropic's fundraising journey has been marked by persistent investor skepticism that evolved alongside the company's growth. The Series D two years ago faced fundamental questions about "why do you need to have a frontier model" and whether AI safety conflicted with building a profitable business, with investors trying to force-fit Anthropic into traditional enterprise software paradigms. By the Series E in late 2024, the company had scaled to nearly a billion dollars in run rate revenue, but the DeepSeek news breaking on the day of first close introduced new volatility as investors questioned whether to "totally rewrite how I think about AI in total." The narrative shifted from doubting the frontier model strategy to grappling with competitive dynamics and market assumptions, though the underlying pattern of investor uncertainty during critical funding moments remained constant.
**答：** Anthropic的融资历程充满了投资者的持续质疑，这些质疑随着公司成长而演变。两年前的Series D融资面临根本性问题，投资者质疑"为什么需要frontier model"，以及AI安全是否与商业盈利相冲突，他们试图用传统企业软件的范式来套用Anthropic。到2024年底的Series E时，公司已经达到近10亿美元的年化收入，但DeepSeek的消息恰好在首轮交割当天爆出，引发新的波动，投资者开始重新思考"是否应该完全改写对AI的认知"。叙事从质疑frontier model策略转向应对竞争动态和市场假设，但在关键融资节点投资者不确定性的底层模式始终存在。

### Topic 30: Series E fundraising and DeepSeek timing
E 轮融资与 DeepSeek 冲击
_[47:18]_

**Q:** What challenges arose when the Series E close coincided with DeepSeek news?
**问：** E 轮融资完成时恰逢 DeepSeek 发布，遇到了什么挑战？

**A:** The Series E closing coincided with DeepSeek's release, creating "a ton of volatility" as investors questioned whether to "totally rewrite how I think about AI in total." Despite reaching "a billion dollars of run rate revenue so quickly," investors doubted the company could maintain growth velocity, citing enterprise adoption timelines and comparing it to the slow cloud migration where "many people are still on-prem." The company countered skepticism by demonstrating that "the return to frontier intelligence is really high" and that their responsible AI approach had an "interesting interlink with our business that most people didn't really understand," ultimately proving their thesis through continued execution driven by "model growth enabled by products and our go-to-market team."
**答：** E 轮融资收尾时正好赶上 DeepSeek 发布，投资人开始重新审视整个 AI 投资逻辑，市场出现剧烈波动。尽管公司已经快速达到 10 亿美元的年化收入，投资人仍然质疑这种增长速度能否持续，他们担心企业市场的采用周期会像云计算迁移那样漫长。公司通过持续的业务表现证明了前沿智能的高回报价值，并展示了负责任 AI 开发与商业成功之间的深层联系——这一点大多数人此前并未真正理解，最终靠产品、GTM 团队和分发能力驱动的模型增长说服了投资人。

### Topic 31: AI safety research as business advantage
AI 安全研究作为商业优势
_[48:11]_

**Q:** How does investment in AI safety and interpretability create business value for enterprise customers?
**问：** 对 AI 安全和可解释性的投资如何为企业客户创造商业价值？

**A:** The speaker argues that AI safety research creates an unexpected business advantage through two mechanisms: technical improvements and enterprise trust. Investments in "interpretability" (described as "an MRI for the model") and alignment science not only serve the mission but make engineers "better at building" models by understanding their internals. For enterprise customers—including nine of the Fortune 10—who entrust the company with sensitive data and workflows, this safety investment signals trustworthiness, which has become critical as businesses run "the most sensitive workloads" on their cloud platform. The speaker emphasizes this wasn't the original motivation, but the "downstream effect" has proven valuable in raising $75 billion and winning enterprise contracts.
**答：** 嘉宾认为 AI 安全研究通过两个机制创造了意外的商业优势：技术改进和企业信任。对 interpretability（被形容为"模型的 MRI"）和 alignment science 的投资不仅服务于使命，还通过理解模型内部运作让工程师"更擅长构建"模型。对于那些将敏感数据和工作流托付给公司的企业客户（包括 Fortune 10 中的九家），这种安全投资传递了可信赖性，这在企业将"最敏感的工作负载"运行在他们的云平台上时变得至关重要。嘉宾强调这不是最初的动机，但这种"下游效应"在融资 750 亿美元和赢得企业合同中被反复证明有价值。

### Topic 32: Capital intensity and funding scale
资本密集度与融资规模
_[50:00]_

**Q:** How much capital has been raised and why is it needed?
**问：** 公司筹集了多少资金，为什么需要这么多资金？

**A:** The company has raised $75 billion since the speaker joined, with another $50 billion committed from recent Amazon and Google deals, reflecting the capital-intensive nature of the AI infrastructure business. Despite this massive scale, the speaker emphasizes that "the business is running very efficiently" and current operations are not generating losses. The primary driver for raising capital is to address "that cone of uncertainty" about future growth trajectories rather than to cover operational deficits, positioning the fundraising as strategic preparation for exponential scaling rather than survival.
**答：** 公司已经筹集了750亿美元，加上最近与Amazon和Google达成的协议，未来还将获得500亿美元投资，这反映了AI基础设施业务的资本密集特性。尽管融资规模巨大，发言人强调业务本身运营高效且没有亏损。筹集资金的主要原因是应对未来增长的不确定性cone，而不是弥补当前运营亏损，这是为指数级扩张做战略准备。

### Topic 33: Shifting from linear to exponential thinking
从线性思维转向指数思维
_[50:30]_

**Q:** How has the CFO's perspective evolved from linear to exponential growth expectations?
**问：** CFO 的视角如何从线性增长预期演变为指数增长预期？

**A:** The CFO initially approached the company's billion-dollar target with "linear thinking," asking "in what year?" rather than considering exponential trajectories. Early skepticism centered on fundamental constraints—"laws of physics and law of large numbers"—and doubts about whether enterprise customers could adopt technology fast enough to support such rapid growth. Over time, observing how "many other exponentials" underlying the revenue exponential actually materialized broke down these mental barriers, shifting the CFO's mindset from incremental planning toward "leaning into this exponential" while maintaining disciplined scenario planning. Notably, CEO Dario has "consistently" been a better revenue predictor than the CFO, though the CFO expects this gap to narrow as forecasting capabilities improve.
**答：** CFO 最初以"线性思维"看待公司的十亿美元目标，会问"哪一年实现"而非考虑指数增长轨迹。早期的怀疑集中在基本约束上——"物理定律和大数定律"——以及企业客户能否足够快地采用技术来支撑如此快速的增长。随着时间推移，观察到支撑收入指数增长的"许多其他指数曲线"实际发生，打破了这些心理障碍，CFO 的思维从渐进式规划转向"拥抱这种指数增长"，同时保持严谨的情景规划。值得注意的是，CEO Dario 在收入预测上"一直"比 CFO 更准确，尽管 CFO 预期随着预测能力提升这一差距会缩小。

### Topic 34: Explaining compute utilization paradigm to investors
向投资者解释算力利用范式
_[52:12]_

**Q:** What is the hardest concept for investors to understand about how Anthropic uses compute?
**问：** 投资者最难理解 Anthropic 如何使用算力的哪个概念？

**A:** The most challenging concept for investors is understanding compute as a fully utilized, fungible resource rather than a traditional variable cost. Unlike conventional companies where R&D teams and production workers cannot swap roles, Anthropic's compute infrastructure exhibits unique "fungibility"—the same chips run inference workloads in the morning and model development in the afternoon. This paradigm shift means investors struggle to separate inference costs from R&D expenses when they are actually "very self-reinforcing," and this flexibility is what drives both short-term revenue and long-term returns on compute.
**答：** 投资者最难理解的是把算力看作一种充分利用的、可互换的资源，而不是传统的可变成本。与传统公司中研发团队和生产工人无法互换角色不同，Anthropic 的算力基础设施具有独特的"fungibility"（可互换性）——同一批芯片上午跑推理工作负载，下午就用于模型开发。这种范式转变导致投资者难以将推理成本和研发支出分开看待，但实际上它们是"相互强化的"，这种灵活性正是驱动短期收入和长期算力回报的关键。

### Topic 35: Key investor due diligence questions
投资者尽职调查的核心问题
_[53:34]_

**Q:** What questions would you ask AI labs if you were an investor evaluating them?
**问：** 如果你是投资者，会如何评估 AI 实验室？

**A:** The speaker identifies three critical investor questions for AI labs: first, "what is the ROI on compute" and how returns evolve over time given the unprecedented capital investments being made. Second, whether customers are achieving real ROI beyond pilots, pointing to their own "over 500% net dollar retention rate" and Fortune 10 deployments as evidence of genuine enterprise adoption with "double-digit million-dollar commits." Third, the strategic question of compute sourcing and supply chain resilience, where their philosophy is to "be involved with great players and have flexibility" across multiple providers rather than depending on a single source.
**答：** 演讲者提出三个关键投资评估问题：首先是 compute 投资回报率及其演变趋势，这是前所未有的大规模资本投入；其次是客户是否获得真实回报，而不只是试点项目——他们自己的 net dollar retention rate "超过 500%"，Fortune 10 企业都在大规模部署，签约金额达到 "double-digit million-dollar" 级别；第三是 compute 供应链策略，他们的做法是与多家优秀供应商合作以保持灵活性，而非依赖单一来源。

### Topic 36: Public perception of AI and industry responsibility
公众对 AI 的看法与行业责任
_[55:58]_

**Q:** Why is AI less popular than Congress and what should the industry do about it?
**问：** 为什么 AI 的受欢迎程度低于国会，行业应该如何应对？

**A:** Speaker A attributes AI's poor public perception to the unprecedented speed of transformation, where "years or decades of progress" are "compressed into months," making it jarring for people who think linearly rather than exponentially. The industry needs to better articulate AI's potential through concrete examples like "drug development and curing diseases" and improved healthcare delivery in resource-scarce regions, as exemplified by Dario's essay "Machines of Loving Grace." However, A emphasizes the importance of "honest and balanced assessments" that acknowledge risks alongside benefits, arguing that people distrust perspectives that present "all the good news and none of the bad news." The key is demonstrating tangible results over time while maintaining transparency about the "bumps on the road" that come with rapid technological change.
**答：** 嘉宾 A 认为 AI 公众形象差的根本原因是变革速度前所未有，"数年甚至数十年的进步" 被 "压缩到几个月" 内完成，这对习惯线性思维的人来说很难适应。行业需要通过具体案例更好地阐述 AI 的潜力，比如 "药物研发和治愈疾病"、改善资源匮乏地区的医疗服务，Dario 的文章 Machines of Loving Grace 就是这方面的尝试。但 A 强调必须保持 "诚实和平衡的评估"，在展示收益的同时承认风险，因为人们不信任只报喜不报忧的说法。关键是随时间展示切实成果，同时对快速变革带来的 "颠簸" 保持透明。

### Topic 37: Cross-sector collaboration on AI solutions
跨部门合作应对AI挑战
_[58:21]_

**Q:** How should commercial and government sectors work together to address AI opportunities and risks?
**问：** 商业和政府部门应如何合作应对AI的机遇和风险？

**A:** The speaker advocates for collaborative dialogue between commercial and government sectors, emphasizing that no single company has "the blueprint that's going to solve everything." The approach requires "clear articulation of the opportunities" alongside transparent discussion of risks and potential solutions. While acknowledging the path won't be "perfectly smooth on the curve," the speaker maintains confidence that long-term opportunities will significantly outweigh the downsides, provided stakeholders engage in honest, ongoing conversation about both benefits and challenges.
**答：** 嘉宾强调商业和政府部门需要通过对话协作，因为没有任何一家公司能拿出解决所有问题的蓝图。关键是要清晰阐述机遇，同时坦诚讨论风险和可能的解决方案。虽然发展过程不会完全平滑，但嘉宾相信只要各方保持透明沟通，长期来看机遇会远大于风险。

### Topic 38: Mythos release and safety concerns
Mythos发布与安全担忧
_[59:01]_

**Q:** What made Mythos the first model that scared people and how did Anthropic approach its release?
**问：** 是什么让Mythos成为第一个让人担忧的模型，Anthropic如何处理其发布？

**A:** Mythos was "incredibly capable across many different dimensions" but particularly "spiked" in cybersecurity capabilities, making it the first model where Anthropic publicly acknowledged concerns about misuse. Rather than withholding release entirely, they adopted a "phased approach" that expanded access gradually to groups who could use it defensively, such as patching codebases—exemplified by finding 250 security vulnerabilities in an open source project where a prior model found only 22. This dual-use tension between offensive and defensive applications informed their release strategy, which they view as "a template that could be used for the future" when models exhibit concentrated capability spikes in sensitive domains.
**答：** Mythos在多个维度上都非常强大，但在网络安全能力上特别突出，这让它成为Anthropic首次公开承认存在滥用风险的模型。他们没有选择完全不发布，而是采用了"分阶段发布"策略，逐步向能够将其用于防御性用途的群体开放访问，比如用来修补代码库——一个典型例子是在开源项目中发现了250个安全漏洞，而之前的模型只找到22个。这种攻击性和防御性应用之间的双重性质决定了发布策略，他们认为这可以作为未来处理在敏感领域出现能力突增的模型的模板。

### Topic 39: Government relations and regulation
政府关系与监管
_[01:00:59]_

**Q:** How does Anthropic navigate government oversight, pre-approval requirements, and partnerships like Stargate?
**问：** Anthropic如何应对政府监管、预批准要求以及Stargate等合作项目？

**A:** Anthropic prioritizes building strong government relationships because they believe "regulation has a role to play in how these models are developed over time," taking an explicitly "America first" approach to support the US and democratic allies. The company is actively working with the administration on initiatives like Stargate while acknowledging the need to balance rapid innovation with a "responsibility framework" for deployment. They view government engagement as essential for having "an honest conversation" about the technology's implications, citing the Nitros process as an example of this collaborative regulatory approach.
**答：** Anthropic优先建立稳固的政府关系，因为他们认为"regulation has a role to play"在模型开发中，并明确采取"America first"策略支持美国及民主国家。公司正积极与政府合作推进Stargate等项目，同时承认需要在快速创新与部署的"responsibility framework"之间取得平衡。他们认为与政府接触对于就技术影响进行"honest conversation"至关重要，并以Nitros流程为例说明这种协作式监管方式。

### Topic 40: Company culture and seven co-founders
公司文化与七位联合创始人
_[01:02:34]_

**Q:** What makes Anthropic's culture distinctive, especially with seven co-founders and rigorous culture interviews?
**问：** Anthropic的文化有何独特之处，特别是七位联合创始人和严格的文化面试？

**A:** Anthropic's culture is anchored by seven co-founders who "shouldn't work on paper, but really does in practice," setting a tone of collaboration without "fiefdoms or sharp elbows" where even brilliant candidates are rejected if they fail the culture interview. The company operates with "intellectual openness and intellectual honesty" where rigorous debate leads to alignment, exemplified by compute allocation discussions that result in decisions without "second-guessing" or politics. Transparency is embedded through biweekly all-hands where Dario addresses "real questions that are on people's minds" rather than "softballs," and the culture maintains humility where reaching milestones prompts "what's next?" instead of celebration, reflecting a relentless focus on mission over individual credit.
**答：** Anthropic的文化由七位联合创始人奠定，这种配置"理论上不该成功，但实践中确实有效"，营造了一种拒绝"地盘主义和尖锐竞争"的协作氛围，即使是最聪明的候选人如果无法通过文化面试也会被拒绝。公司内部保持"智识开放和诚实"，激烈辩论后能达成一致，比如在计算资源分配讨论中做出决策后不会有"事后质疑"或政治斗争。透明度体现在每两周的全员会议上，Dario回答"员工真正关心的问题"而非"软球问题"，文化保持谦逊，达成里程碑时会问"下一步是什么"而非庆祝，体现了对使命的专注而非个人功劳。

### Topic 41: Transparency and leadership communication
透明度与领导层沟通
_[01:05:23]_

**Q:** How does Dario's biweekly all-hands and open Q&A contribute to company transparency?
**问：** Dario的双周全员会议和开放问答如何促进公司透明度？

**A:** Dario conducts biweekly all-hands meetings where he presents a short document covering "three or four topics" and then takes "real questions" from employees—not "softballs" or "planted questions"—creating a transparent window into leadership thinking. This open communication culture has contributed to exceptional retention, with "all seven of the co-founders" and the "vast majority of the first 20 to 30 employees" still at the company. The transparency advantage proved particularly valuable during competitive talent wars: when Meta and others offered "huge packages," the company "lost two people" while "other labs lost dozens," demonstrating that culture can compete effectively against pure compensation.
**答：** Dario每两周举行一次全员会议，准备简短文档讲解"三四个话题"，然后接受员工的"真实问题"——不是"软球"或"预设问题"——让全公司透明地了解领导层思路。这种开放沟通文化带来了极高的留存率："七位联合创始人全部"和"前20到30名员工中的绝大多数"都还在公司。透明度优势在人才争夺战中尤为明显：当Meta等公司开出"巨额薪酬包"时，公司"只流失了两人"，而"其他实验室流失了几十人"，证明文化可以有效对抗纯粹的薪酬竞争。

### Topic 42: Talent retention and researcher motivation
人才保留与研究员动机
_[01:06:31]_

**Q:** Why does Anthropic retain talent better than competitors despite not always paying the most?
**问：** 为什么Anthropic在薪酬并非最高的情况下仍能更好地留住人才？

**A:** Anthropic's retention advantage stems from a culture that empirically attracts researchers who prioritize impact and mission over compensation, with employees citing "talent density mattering more than talent mass" and collaborative decision-making with transparency. The company's commitment to developing transformative AI "in a responsible way" resonates across teams, creating what they call a "race to the top" where they aim to set industry standards even while acknowledging they don't have all the answers. This cultural foundation—rooted in meaningful work, collaborative environment, and contributing to humanity's benefit—proves more compelling to their talent pool than purely financial incentives.
**答：** Anthropic的人才保留优势源于一种文化：研究员们更看重影响力和使命感而非薪酬，员工们强调"人才密度比人才总量更重要"以及透明的协作决策机制。公司致力于"以负责任的方式"开发变革性AI技术，这一理念在各团队中产生共鸣，形成了所谓的"race to the top"——即便承认自己并非全知全能，也要为行业树立标杆。这种文化基础——有意义的工作、协作环境、为人类福祉做贡献——对他们的人才群体而言，比纯粹的经济激励更具吸引力。

### Topic 43: The frontier: virtual collaborators
前沿方向：虚拟协作者
_[01:08:05]_

**Q:** What does the AI frontier look like from inside Anthropic, particularly the vision of virtual collaborators?
**问：** 从Anthropic内部看，AI前沿是什么样的，特别是虚拟协作者的愿景？

**A:** Anthropic's frontier vision centers on building "virtual collaborators" designed to transform enterprise knowledge work productivity. These systems would possess organizational context, use company-specific tools, maintain memory to learn from past mistakes, and work over extended time horizons on complex ideas rather than isolated tasks. The approach requires both continued model capability growth and thoughtful product design, recognizing that intelligence isn't "just a single dimension" but must be contextualized for specific use cases. This represents a shift from generic AI assistants to deeply integrated collaborators that understand and adapt to organizational workflows.
**答：** Anthropic的前沿愿景聚焦于构建"虚拟协作者"，旨在变革企业知识工作的生产力。这类系统需要具备组织上下文理解能力，能使用公司特定的工具（无论是自研还是采购的），拥有记忆功能以从过往错误中学习，并能在较长时间跨度内处理复杂想法而非单一任务。实现这一愿景需要模型能力的持续提升和产品设计的精心打磨，因为智能不是"单一维度"的，而必须针对具体使用场景进行定制。这标志着从通用AI助手向深度融入组织工作流的协作者的转变。

### Topic 44: AI coding tools adoption and virtual collaboration
AI 编程工具的采用与虚拟协作模式
_[01:09:45]_

**Q:** How are AI coding tools like Claude and Cowork changing developer workflows and team structures?
**问：** Claude 和 Cowork 等 AI 编程工具如何改变开发者的工作流程和团队结构？

**A:** The speaker observes that AI coding tools are fundamentally transforming software development from traditional team structures into a model of "virtual collaborators" working alongside humans. While Claude for code pioneered this shift, newer tools like Cowork are achieving adoption "faster than Claude for code" when indexed to comparable stages, which is remarkable given that developers are already "really fast adopters." The speaker describes their own company's evolution where product development now involves "a fleet of agents" shipping daily rather than small teams working over months, creating a paradigm where "everyone kind of becomes a manager" overseeing AI collaborators. This represents an early but potentially transformative shift in how software gets built, with "incredible" productivity implications once the right form factors emerge.
**答：** 讲者观察到 AI 编程工具正在从根本上改变软件开发模式，从传统团队结构转向人类与"虚拟协作者"并肩工作的模式。虽然 Claude for code 率先开创了这一转变，但像 Cowork 这样的新工具在相同发展阶段的采用速度更快，这在开发者本就是"非常快速的采用者"的背景下尤为显著。讲者描述了自己公司的演变：产品开发现在依靠"一群 agents"每天发布，而不是小团队花几个月时间，形成了"每个人都成为管理者"监督 AI 协作者的新范式。这代表了软件构建方式的早期但具有变革性的转变，一旦找到合适的产品形态，生产力提升的潜力"令人难以置信"。

### Topic 45: Personal scaling as an executive
高管的个人成长与规模化
_[01:10:55]_

**Q:** How does an executive personally evolve to scale with a rapidly growing company?
**问：** 高管如何实现个人成长以适应快速增长的公司？

**A:** The speaker emphasizes that scaling as an executive requires "thinking in first principles" and maintaining "intellectual openness" rather than relying on prior experience. He describes a pivotal moment when Tom Brown, the chief compute officer, shared his vision during a two-and-a-half-hour walk in early 2024, which "sounded crazy" at the time but made him realize the company would "bend all paradigms" beyond what most people have seen. The key is acknowledging that everyone brings priors to new situations, but success at unprecedented scale demands setting those aside and being open to possibilities that seem implausible based on past experience.
**答：** 这位高管强调，要实现个人成长必须"从第一性原理思考"并保持"思想开放性"，而不是依赖过往经验。他描述了一个关键时刻：2024年初，首席计算官Tom Brown在两个半小时的散步中分享了他的愿景，当时听起来"很疯狂"，但让他意识到公司将"打破所有范式"，超越大多数人见过的规模。核心在于认识到每个人都会带着先验知识进入新环境，但在前所未有的规模下取得成功，需要放下这些经验，对那些基于过去看似不可能的事情保持开放。

### Topic 46: Tom Brown's vision and first principles thinking
Tom Brown 的愿景与第一性原理思维
_[01:12:18]_

**Q:** What was Tom Brown's early vision for the company and how did it shape expectations?
**问：** Tom Brown 对公司的早期愿景是什么，如何塑造了预期？

**A:** Tom Brown articulated a vision during an early walk that was so transformative the speaker told his wife it would "bend all paradigms" if even 10% proved true. Much of what Tom predicted "has come to fruition," validating the initial assessment that this would be "totally different and new." The speaker frames the experience as both "incredible" and "really challenging," acknowledging the dual nature of working on paradigm-shifting technology. This foundational conversation set expectations for a fundamentally different kind of company, one built on partnership rather than hierarchy, where the speaker hires people "as a partner" rather than direct reports.
**答：** Tom Brown 在早期的一次散步中描绘了一个极具变革性的愿景，以至于说话者回家告诉妻子，即使只有 10% 成真，这也会「颠覆所有范式」。Tom 当时预测的很多内容「已经实现了」，验证了最初的判断——这将是「完全不同和全新的」。说话者将这段经历描述为既「令人难以置信」又「极具挑战性」，承认了从事范式转变技术的双重性质。这次奠基性的对话为一种根本不同的公司设定了预期，一个建立在伙伴关系而非层级制度上的公司，说话者招人时把他们视为「合作伙伴」而非直接下属。

### Topic 47: Hiring partners and managing surface area
招聘合伙人与管理业务复杂度
_[01:13:02]_

**Q:** How does hiring great people as partners help manage the complexity of a fast-growing business?
**问：** 如何通过招聘优秀人才作为合伙人来管理快速增长业务的复杂性？

**A:** The speaker emphasizes hiring senior people as true "partners" rather than just direct reports, creating relationships where disagreement and debate are expected and valued. He acknowledges that while he prefers operating at a granular level rather than "at 50,000 feet," the business has "too much surface area" for one person to maintain that depth everywhere, making partner-level hires essential for distributed ownership. These partners bring diverse perspectives from "some of the best companies in the world" including hyperscalers, large software companies, and financial services, with the speaker citing his own Blackstone private equity background as an example of valuable training that informs decision-making.
**答：** 讲者强调要把高级人才当作真正的"合伙人"而非普通下属来招聘，建立一种鼓励分歧和辩论的合作关系。他承认自己更喜欢在细节层面工作而非"在50,000英尺高空俯瞰"，但业务的"表面积太大"，一个人无法在所有领域都保持这种深度，因此合伙人级别的招聘对于分布式所有权至关重要。这些合伙人来自"世界上最优秀的公司"，带来了hyperscaler、大型软件公司和金融服务等不同视角，讲者以自己在Blackstone私募股权的背景为例说明这种训练对决策的价值。

### Topic 48: Learning from Airbnb and unprecedented situations
从 Airbnb 危机中学习应对未知局面
_[01:14:01]_

**Q:** What lessons from past experiences like Airbnb's pandemic crisis apply to navigating unprecedented growth?
**问：** 从 Airbnb 疫情危机等过往经验中学到了什么，可以用于应对前所未有的增长？

**A:** The speaker draws on leading Airbnb's pandemic financing when the business "lost 70% of its revenue in seven weeks" to emphasize the importance of maintaining "clear perspective" during rapid change without established templates. Beyond crisis management tactics, he highlights a personal practice of pausing "maybe once a week" to recognize the unique opportunity of working "with this group of people on this problem at this company at this moment in time." This deliberate appreciation—whether "in a car ride" or "late at night"—serves as a grounding mechanism when navigating both unprecedented challenges and the demands of balancing work with personal life.
**答：** 讲者以主导 Airbnb 疫情期间融资的经历为例，当时业务"七周内损失了 70% 的收入"，强调在快速变化且没有既定模板的情况下保持"清晰视角"的重要性。除了危机管理策略，他分享了一个个人习惯："每周一次"在安静时刻停下来，意识到能够"在这个时刻与这群人在这家公司解决这个问题"是多么难得的机会。这种刻意的感恩——无论是"在车里"还是"深夜"——成为他在应对未知挑战和平衡工作生活时的心理支点。

### Topic 49: Tom's predictions about compute scale and model capabilities
Tom 对算力规模和模型能力的预测
_[01:15:07]_

**Q:** What specific predictions did Tom make about compute infrastructure and AI capabilities?
**问：** Tom 对算力基础设施和 AI 能力做出了哪些具体预测？

**A:** Tom painted a vision that initially seemed "sci-fi" but has proven prescient, centered on the idea that "everything is going to happen much quicker than we think" in terms of both compute infrastructure scale and model capabilities. The conversation left a lasting impression not just through technical predictions but through Tom's "incredible optimism about the future," which influenced the internal philosophy of "holding light and shade" - balancing awareness of risks with positive possibilities. While many of Tom's predictions have already materialized in current AI systems, he described capabilities "beyond where we are today" that continue to inform expectations about rapid advancement timelines.
**答：** Tom 描绘的愿景最初听起来像"科幻小说"，但后来被证明很有预见性，核心观点是算力基础设施规模和模型能力的发展速度会"比我们想象的快得多"。这次对话的影响不仅在于技术预测，更在于 Tom 对未来"令人难以置信的乐观态度"，这种态度影响了团队内部"holding light and shade"的理念——在认识风险的同时保持对积极可能性的关注。虽然 Tom 的许多预测已经在当前 AI 系统中实现，但他还描述了一些"超越今天水平"的能力，这些预测继续影响着人们对快速发展时间线的预期。

### Topic 50: Risks to the growth trajectory
增长轨迹的风险因素
_[01:16:11]_

**Q:** What factors could cause the company's growth to slow down or shift to the lower end of projections?
**问：** 哪些因素可能导致公司增长放缓或转向预测的低端？

**A:** The speaker identifies three primary risk factors that could dampen growth expectations. First, organizational inertia poses a significant challenge—"change is hard" within large enterprises with established practices, and if "diffusion rate within customers" slows as use cases struggle to keep pace with model capabilities, revenue growth could stall. Second, while they don't currently observe it, the possibility that "scaling laws slowing down or not holding" and model capabilities leveling off represents an existential risk to the growth thesis. Third, maintaining frontier position in "agentic AI" is critical but not guaranteed given competitive market dynamics, requiring sustained investment in technology, compute, and go-to-market execution.
**答：** 发言人指出了三个可能拖累增长预期的主要风险。首先是组织惯性问题——大型企业内部"change is hard"，如果客户内部的"diffusion rate"放缓，用例跟不上模型能力的发展速度，营收增长就会受阻。其次，虽然目前没有观察到，但"scaling laws"可能放缓或失效、模型能力增长趋于平缓，这是增长逻辑的根本性风险。第三，在"agentic AI"领域保持前沿地位至关重要但并非必然，在竞争激烈的市场中需要持续投入技术、算力和市场资源。

### Topic 51: Healthcare and biotech applications of AI
AI 在医疗和生物技术领域的应用
_[01:17:40]_

**Q:** What are the most exciting potential applications of AI in healthcare and drug development?
**问：** AI 在医疗和药物开发领域最令人兴奋的潜在应用是什么？

**A:** The speaker sees AI's greatest potential in accelerating drug discovery and development, envisioning a future where diseases currently incurable could have cures found "much more rapidly" within a patient's lifetime. While current AI applications focus on speeding up administrative aspects like "clinical studies, reports, and things like that," the real breakthrough will come when AI moves upstream into drug discovery itself, where the complexity of molecules and proteins makes AI "perfect for that" kind of work. The speaker anticipates lab throughput increasing "10x or 100x," enabling exponentially more experiments with better results that could benefit people globally across a much broader range of diseases beyond just a "small set."
**答：** 嘉宾认为 AI 最大的潜力在于加速药物发现和开发，设想未来那些目前无法治愈的疾病可以在患者有生之年更快找到治疗方法。虽然目前 AI 主要用于加速临床研究报告等行政工作，但真正的突破将发生在 AI 深入到药物发现环节——分子和蛋白质的复杂性使得 AI 特别适合这类工作。嘉宾预期实验室产出可以提升 10 到 100 倍，能够进行指数级更多的实验并更快获得更好的结果，这将惠及全球更广泛的疾病类型。

### Topic 52: Kindest thing - brother's college sacrifice
最善良的事 - 哥哥的大学牺牲
_[01:19:37]_

**Q:** What is the kindest thing that anyone's ever done for you?
**问：** 别人为你做过的最善良的事是什么？

**A:** The speaker's brother, five and a half years older, got into every college he applied to with plans for medical school, but chose to attend an in-state school instead. Years later, the speaker discovered that his brother's decision was driven by wanting to "give me the opportunity to go wherever I wanted" despite the family being "solidly middle class" with less robust financial aid packages 25-30 years ago. What makes this particularly meaningful is that the speaker was only 12 or 13 at the time and "would have never really understood" the sacrifice, yet his brother made this choice for someone six years away from college with an uncertain future.
**答：** 嘉宾的哥哥比他大五岁半，当年申请的所有大学都录取了他，本打算读医学院，但最终选择了州内大学。多年后嘉宾才了解到，哥哥做这个决定是为了"给我机会去任何我想去的地方"，尽管当时家庭"稳定的中产阶级"背景和25-30年前不够完善的助学金体系让这成为重要考量。这件事特别有意义的地方在于，当时嘉宾只有12、13岁，"根本不会理解"这种牺牲，而哥哥却为一个六年后才上大学、前途未知的弟弟做出了这样的选择。

### Topic 53: Sponsor: Rogo AI for Investment Firms
赞助商：Rogo AI 投资公司平台
_[01:21:24]_

**Q:** How can investment firms use AI that understands their specific processes?
**问：** 投资公司如何使用理解其特定流程的AI？

**A:** The speaker positions Rogo as a specialized alternative to "generic AI" that fails to understand investment firm workflows. Unlike general-purpose tools, Rogo is "built specifically for Wall Street" and integrates directly with firms' proprietary data systems to understand their unique processes. The platform's value proposition centers on producing "real outputs" tailored to each firm's specific methodology rather than generic responses.
**答：** 演讲者将 Rogo 定位为专门针对投资公司的 AI 平台，解决了"通用 AI"无法理解投资公司特定工作流程的问题。与通用工具不同，Rogo 是"专为华尔街打造"的，能够连接公司的专有数据系统，理解每家公司独特的投资流程。这个平台的核心价值在于根据每家公司的具体方法论产生"真实的输出结果"，而不是泛泛的通用回答。

### Topic 54: Sponsor: WorkOS for Enterprise Readiness
赞助商：WorkOS 企业就绪解决方案
_[01:21:37]_

**Q:** How can AI and software companies become enterprise ready quickly?
**问：** AI和软件公司如何快速实现企业就绪？

**A:** WorkOS positions itself as an infrastructure solution that enables companies to "become enterprise ready overnight, not in months" by handling the unglamorous foundational work. The service is used by prominent AI companies including OpenAI, Cursor, and Perplexity, suggesting it addresses common enterprise requirements like authentication, authorization, and compliance. The value proposition centers on allowing engineering teams to "skip the unglamorous infrastructure work and focus on your product" rather than building enterprise features from scratch.
**答：** WorkOS 定位为一个基础设施解决方案，让公司能够"一夜之间而非数月内实现企业就绪"，通过处理那些不起眼但必需的底层工作。OpenAI、Cursor 和 Perplexity 等知名 AI 公司都在使用这项服务，说明它解决了企业级常见需求，比如身份认证、权限管理和合规性。核心价值主张是让工程团队"跳过枯燥的基础设施工作，专注于产品本身"，而不是从零开始构建企业级功能。

### Topic 55: Sponsor: Ridgeline Asset Management Technology
赞助商：Ridgeline 资产管理技术
_[01:21:48]_

**Q:** What technology solutions help asset management firms scale and gain competitive advantage?
**问：** 什么技术解决方案帮助资产管理公司扩展并获得竞争优势？

**A:** Ridgeline positions itself as "redefining asset management technology" by functioning as a "true partner, not just a software vendor," emphasizing a relationship-driven approach beyond traditional software licensing. The platform has demonstrated measurable impact by helping firms achieve "5x" scale, though the specific metric (AUM, headcount, or revenue) is not specified in this segment. The value proposition centers on three operational outcomes: "faster growth, smarter operations, and a competitive edge," suggesting the technology addresses both efficiency and strategic differentiation challenges that asset managers face.
**答：** Ridgeline 将自己定位为资产管理技术的重新定义者，强调作为"真正的合作伙伴而非仅仅是软件供应商"的关系驱动型服务模式。该平台展示了可量化的影响力，帮助公司实现了"5倍"的规模增长，尽管这段内容中未明确具体指标（资产管理规模、人员数量或收入）。其价值主张围绕三个运营成果："更快的增长、更智能的运营和竞争优势"，表明该技术同时解决了资产管理公司面临的效率和战略差异化挑战。

---

## Vocabulary (CEFR B2+)

### frontier  /frʌnˈtɪr/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 24

**EN:** the extreme limit of understanding or achievement in a particular area; the most advanced level  
**CN:** 前沿，尖端领域；某一领域中理解或成就的极限

**Original examples:**
- [00:00] Every time we have a new model, there's a set of capabilities that are different. People tend to think about model intelligence as IQ. We think of it kind of differently. Intelligence for us is multi-dimensional. It's not just a score. What is the real-world capability of this model? Each model generation gives you the chance to do more with it, to do it better, to do it more efficiently, because we think the returns to **frontier** intelligence are extremely high. And it's extremely high, especially in enterprise. That's a core thesis of our business.  
  每次我们推出新模型时,都会带来一系列不同的能力。人们往往把模型智能理解为智商。但我们的看法有所不同。对我们来说,智能是多维度的,不只是一个分数。这个模型在现实世界中的实际能力是什么?每一代模型都让你能用它做更多事情,做得更好,做得更高效,因为我们认为前沿智能的回报极高。尤其在企业领域,回报极高。这是我们业务的核心理念。
- [01:48] Much compute, you go out of business. If you buy too little compute, you can't serve your customers and you're not at the **frontier** is the same thing.  
  算力,公司就会倒闭。如果买太少算力,你就无法服务客户,也无法保持在前沿,结果是一样的。
- [02:17] We think about the compute we need to stay at the **frontier** and we really look ahead and try to estimate that. And then as we go out and actually do these deals to procure compute, you know, flexibility is really important to us and so we build that flexibility into the deals themselves.  
  我们考虑保持在前沿所需的算力,并真正向前看,尝试估算。然后当我们实际去做这些采购交易时,灵活性对我们来说非常重要,所以我们把灵活性构建到交易本身中。
- [03:27] We've invested in that over multiple years to be what I believe the most efficient users of compute amongst any of the **frontier** labs.  
  我们在这方面投入了好几年时间,我相信现在我们是所有前沿实验室中算力使用效率最高的。
- [06:11] And what we want to do is we want to be at a place where we can, you know, obviously still be at the **frontier**. That's the most important thing.  
  我们想要达到的状态是,显然仍然处于前沿,这是最重要的。
- [07:52] Because we think the returns to **frontier** intelligence are extremely high, and it's extremely high especially in enterprise.  
  因为我们认为前沿智能的回报极高，尤其是在企业领域。
- [12:22] You said something really important before, which is the returns to being at the **frontier** are really high.  
  你之前说了一个很重要的观点,就是处于前沿位置的回报非常高。
- [12:55] So talk about the returns to being on the **frontier** and why it's so high.  
  那么谈谈处于前沿的回报以及为什么回报如此之高。
- [14:40] And so that's what I mean by the returns to **frontier** intelligence are really high.  
  这就是我所说的前沿智能的回报真的很高的意思。
- [15:20] The things that push that **frontier** is like a sci-fi story or something from books I was reading when I was growing up.  
  推动这个前沿的事情就像科幻故事或我小时候读的书里的东西。
- [15:20] The things that push that **frontier** is like a sci-fi story or something from books I was reading when I was growing up.  
  推动这个前沿的事物就像科幻故事或我小时候读过的书中的情节。
- [15:41] And that there is some sort of—if I think about the **frontier** that you're pushing and OpenAI is pushing and compare that to the open source models—that maybe the gap will widen as a...  
  而且存在某种——如果我想想你们正在推动的前沿和OpenAI正在推动的前沿,并将其与开源模型进行比较——差距可能会作为一个...
- [16:57] We think of them as **frontier** or not.  
  我们把它们看作是前沿的或非前沿的。
- [17:00] And the ones that are at the **frontier**, you know, clearly are capturing this economic value, driving meaningful ROI for customers.  
  那些处于前沿的模型显然正在获取这种经济价值，为客户带来有意义的投资回报。
- [46:12] At the time, you know, look, that was not a straightforward fundraising. The company really only had a **frontier** model in the middle of that fundraising.  
  当时，你知道，那不是一次简单的融资。公司在融资过程中实际上只有一个前沿模型。
- [46:12] At the time, you know, look, that was not a straightforward fundraising. The company really only had a **frontier** model in the middle of that fundraising. Towards the tail end of it, the FTX transaction was happening, which was liquidating a bunch of Anthropic shares, and so that was kind of the starting point. And at that point, the questions were around like, why do you need to have a frontier model? Like, what's the returns to this?  
  那次融资并不顺利。公司当时只有一个前沿模型,而且在融资快结束时,FTX 的交易正在进行,清算了一批 Anthropic 的股份,这就是起点。那时候投资人会问,为什么你们需要有前沿模型?这能带来什么回报?
- [47:56] The business continued to prove out the thesis that the return to **frontier** intelligence is really high that we are really focused on.  
  但业务持续证明了我们的论点:前沿智能的回报确实很高,而我们真正专注于此。
- [49:49] And that's not why we invested in it, but it did have this kind of downstream effect that we've really seen prove out again and again to be a company that is both at the **frontier**, but one that is investing in safety and that you can trust.  
  这不是我们投资的初衷,但它确实产生了一种下游效应,我们一次又一次地看到这一点得到验证——成为一家既处于前沿、又投资于安全、值得信赖的公司。
- [01:08:05] As you're having conversations with people internally, what does the **frontier** feel like to you?  
  当你与内部人员交谈时，前沿对你来说感觉如何？
- [01:08:29] What feels to you like the **frontier** from the inside?  
  从内部来看,什么让你感觉像是前沿?
- [01:08:29] What feels to you like the **frontier** from the inside?  
  从内部来看,什么让你感觉像是前沿?
- [01:09:52] For us, Claude for code has led the way on that as well as much of the business that we have great customers in that are pushing the coding **frontier** as well.  
  对我们来说,Claude for code 在这方面引领了方向,我们也有很多优秀的客户在推动编程前沿的发展。
- [01:17:18] And then, you know, maybe third is just how we think about being at the **frontier**. You know, today we're at the **frontier**. I think we're defining the **frontier** of agentic AI.  
  然后，第三点可能就是我们如何看待处于前沿。今天我们处于前沿。我认为我们正在定义智能体AI的前沿。
- [01:17:18] And then, you know, maybe third is just how we think about being at the **frontier**. You know, today we're at the frontier. I think we're defining the frontier of agentic AI. We need to stay there, right? And it's a competitive  
  第三点可能是我们如何看待保持在前沿的问题。现在我们处于前沿,我认为我们正在定义agentic AI的前沿。我们需要保持在那里,对吧?这是一个竞争激烈的市场。

**Extra example:**
- Scientists are exploring the **frontier** of quantum computing.  
  科学家们正在探索量子计算的前沿领域。

### allocation  /ˌæləˈkeɪʃn/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 4

**EN:** the process of distributing resources or duties for a particular purpose  
**CN:** 分配，配置；为特定目的分配资源或任务的过程

**Original examples:**
- [07:08] You naively might think like, okay, it's a third, a third, a third **allocation** or something. Like how much does that range around? What are the trade-offs like? What is that discussion like?  
  你可能会天真地以为是三三制分配,各占三分之一。但实际分配比例会有多大浮动?权衡考量是什么?这种讨论是怎么进行的?
- [07:08] You naively might think like, okay, it's a third, a third, a third **allocation** or something. Like how much does that range around? What are the trade-offs like? What is that discussion like?  
  你可能会天真地以为是三三制分配,各占三分之一。但实际分配比例会有多大浮动?权衡考量是什么?这种讨论是怎么进行的?
- [07:19] On an ongoing basis, in addition to meeting about compute procurement, we meet a lot about compute **allocation**.  
  在持续的基础上，除了讨论算力采购的会议，我们还经常开会讨论算力分配。
- [07:19] On an ongoing basis, in addition to meeting about compute procurement, we meet a lot about compute **allocation**.  
  在日常运营中,除了讨论算力采购,我们还经常开会讨论算力分配。

**Extra example:**
- The budget **allocation** for education has increased this year.  
  今年教育预算的分配有所增加。

### consequential  /ˌkɑːnsɪˈkwenʃl/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** important; significant, especially in having important effects or results  
**CN:** 重要的，重大的；尤指具有重要影响或结果的

**Original examples:**
- [01:39] And so the decisions we make and how much compute to buy are some of the most **consequential** and hardest decisions to make in the entire company.  
  因此，我们做出的决定以及购买多少算力是整个公司最重大、最困难的决定之一。

**Extra example:**
- The CEO's resignation was a **consequential** event for the company's future.  
  CEO的辞职对公司的未来是一个重大事件。

### exponentially  /ˌekspəˈnenʃəli/
**CEFR:** B2 | **Part of speech:** adv. | **Occurrences:** 3

**EN:** at a rate that becomes faster and faster; increasing more and more rapidly  
**CN:** 以指数方式，呈指数级地；以越来越快的速度增长

**Original examples:**
- [02:36] into how we use the compute as well, because the way in which we bridge from a position we are today to where we want to go when the business is growing **exponentially** is to use that compute as efficiently as possible.  
  这也涉及到我们如何使用算力，因为当业务呈指数级增长时，我们从今天的位置到达我们想去的地方的方式就是尽可能高效地使用算力。
- [05:27] Sure. When you're building and growing a business **exponentially**, you know, really small movements in monthly or weekly growth rates result in compounding very, very different outcomes.  
  当然。当你以指数级方式建立和发展业务时，月度或周度增长率的微小变化会导致截然不同的复合结果。
- [05:27] Sure. When you're building and growing a business **exponentially**, you know, really small movements in monthly or weekly growth rates result in compounding very, very different outcomes. And so as we're thinking ahead, you know, even with our revenue growth, it's really hard to predict this business, right? And it's really hard. I think humans mostly think linearly and you think incrementally. And that's something, you know, I've been at the company for two years. That's something  
  当然。当你在指数级地构建和发展一个业务时,月度或周度增长率的很小变动,经过复利效应会导致非常非常不同的结果。所以当我们展望未来时,即使是我们的收入增长,也真的很难预测这个业务,对吧?这真的很难。我觉得人类大多是线性思考的,你会增量式地思考。这是我在公司两年来

**Extra example:**
- The cost of computing power has decreased **exponentially** over the past decades.  
  过去几十年里，计算能力的成本呈指数级下降。

### paradigm  /ˈpærədaɪm/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 4

**EN:** a typical example or pattern of something; a model or framework for thinking  
**CN:** 范式，典范；思维的模式或框架

**Original examples:**
- [05:51] That's a **paradigm** I've had to break for myself, right? To stop just thinking linearly and think on this exponential.  
  这是我必须为自己打破的范式，对吧？停止线性思考，开始进行指数级思考。
- [46:36] And so there was just a bit of a **paradigm** around trying to fit us, you know, into a particular mold that had existed before.  
  所以当时存在一种范式，试图把我们套入之前已经存在的某种特定模式中。
- [52:23] I think it is this **paradigm** of how compute is used. Thinking of it as, you know, not just something that is like a variable cost over some time period, but really this resource that's so fully utilized, right?  
  我认为这是关于如何使用算力的范式。把它看作不仅仅是某个时期的可变成本，而是一种被充分利用的资源，对吧？
- [52:23] I think it is this **paradigm** of how compute is used. Thinking of it as, you know, not just something that is like a variable cost over some time period, but really this resource that's so fully utilized, right? We run workloads on one day in the morning on a chip for inference and in the afternoon or evening we use it for model development. That is something that's, you know, that paradigm does not exist in a company like a software company or a factory, right? If you, you can't...  
  我认为是算力使用的这种范式。不能把它仅仅看作某个时间段内的可变成本,而应该看作一种被充分利用的资源,对吧?我们在同一天早上用一块芯片做推理,下午或晚上就用它来做模型开发。这种范式在软件公司或工厂里是不存在的,对吧?你不能……

**Extra example:**
- The discovery challenged the existing scientific **paradigm**.  
  这一发现挑战了现有的科学范式。

### cone  /koʊn/
**CEFR:** B1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a solid or hollow shape with a circular base and sides that slope up to a point; used metaphorically to describe expanding range of possibilities  
**CN:** 圆锥体；（比喻）不断扩大的可能性范围

**Original examples:**
- [06:00] We look at a range of scenarios and we look at different points in that **cone** of uncertainty over, you know, a one to two-year period, and then we kind of work backwards from that.  
  我们会看一系列场景,看这个不确定性锥在一到两年期间的不同点位,然后我们从那里反推。

**Extra example:**
- The **cone** of uncertainty widens as we project further into the future.  
  当我们预测更远的未来时，不确定性的范围会扩大。

### trade-off  /ˈtreɪd ɔːf/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a balance achieved between two desirable but incompatible features; a compromise  
**CN:** 权衡，取舍；在两个理想但不兼容的特征之间达成的平衡

**Original examples:**
- [24:46] What about the **trade-off**? You said price per performance. The trade-off between like cost per token or something, throughput and speed.  
  那权衡取舍呢?你提到了性价比。比如每个 token 的成本、吞吐量和速度之间的权衡。

**Extra example:**
- There's always a **trade-off** between speed and accuracy in software development.  
  在软件开发中，速度和准确性之间总是存在权衡。

### capability  /ˌkeɪpəˈbɪləti/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 8

**EN:** the power or ability to do something; a feature or faculty that makes something possible  
**CN:** 能力，才能；使某事成为可能的特征或功能

**Original examples:**
- [13:25] But what our measurement is is what the customers tell us, like what is the real world **capability** of this model.  
  但我们的衡量标准是客户告诉我们的,比如这个模型在真实世界中的能力是什么。
- [13:25] But what our measurement is is what the customers tell us, like what is the real world **capability** of this model.  
  但我们的衡量标准是客户告诉我们的,比如这个模型在真实世界中的能力是什么。
- [13:25] But what our measurement is is what the customers tell us, like what is the real world **capability** of this model.  
  但我们的衡量标准是客户告诉我们的,比如这个模型在真实世界中的能力是什么。
- [13:25] But what our measurement is is what the customers tell us, like what is the real world **capability** of this model.  
  但我们的衡量标准是客户告诉我们的,比如这个模型在真实世界中的能力是什么。
- [16:42] And so in addition to this **capability** leap that you would have just from the scaling laws, talent is really important.  
  因此，除了仅从规模定律就能获得的能力飞跃之外，人才真的很重要。
- [19:44] And that gives us a sense for model **capability**.  
  这让我们对模型能力有了一个判断。
- [21:08] So if that's true, you said before it's hard for humans to be exponential in their thinking and not linear. Like, if that continues to be true for however many more, you know, turns of the crank here, how do you do that thing of not thinking linear and thinking exponential yourself in your job and for the business? Like, the implications are really hard to reason through. Exponential growth rate is one thing, but exponential growth of **capability**—like, I don't even know how to get my head around it.  
  如果这是真的，你之前说过人类很难进行指数级思考而不是线性思考。如果这在未来继续成立，无论还要经历多少次迭代，你如何在工作和业务中做到不进行线性思考而是指数级思考？这些影响真的很难推理。指数级增长率是一回事，但能力的指数级增长——我甚至不知道如何理解它。
- [24:56] Probably unlocks some **capability** and use cases that are really interesting that we don't know about yet as these things get faster.  
  速度可能会解锁一些我们还不知道的、非常有趣的能力和用例,随着这些东西变得越来越快。

**Extra example:**
- The new software has enhanced **capabilities** for data analysis.  
  新软件具有增强的数据分析能力。

### agentic  /eɪˈdʒentɪk/
**CEFR:** C2 | **Part of speech:** adj. | **Occurrences:** 2

**EN:** relating to the capacity of an AI system to act autonomously and make decisions to achieve goals  
**CN:** 智能体的，自主的；与AI系统自主行动和做出决策以实现目标的能力相关的

**Original examples:**
- [13:36] It's also the ability to do long horizon tasks, the ability to use tools or computer use, the ability to do things for **agentic** tasks that have specific value even faster, right?  
  它还包括执行长期任务的能力，使用工具或计算机的能力，更快地完成具有特定价值的智能体任务的能力，对吧？
- [01:17:28] I think we're defining the frontier of **agentic** AI.  
  我认为我们正在定义智能体AI的前沿。

**Extra example:**
- The company is developing **agentic** systems that can handle complex workflows independently.  
  该公司正在开发能够独立处理复杂工作流程的智能体系统。

### unlock  /ʌnˈlɑːk/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to make something available or possible that was previously restricted or inaccessible  
**CN:** 释放，开启（新的可能性或机会）

**Original examples:**
- [43:43] See what Ridgeline can **unlock** for your firm. Schedule a demo at  
  看看 Ridgeline 能为你的公司带来什么。在 ridgeline.ai 预约演示。
- [43:43] See what Ridgeline can **unlock** for your firm. Schedule a demo at  
  看看 Ridgeline 能为你的公司带来什么。在 ridgeline.ai 预约演示。

**Extra example:**
- Better education can **unlock** human potential and create opportunities.  
  更好的教育能够**释放**人类潜能并创造机会。

### thesis  /ˈθiːsɪs/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** a statement or theory that is put forward as a premise to be maintained or proved; a core belief or proposition  
**CN:** 论点，核心观点；（投资或战略的）核心理念

**Original examples:**
- [15:15] Our **thesis** is that frontier intelligence will continue to improve exponentially.  
  我们的**核心观点**是前沿智能将继续呈指数级改进。
- [17:06] And we are just investing behind that **thesis**.  
  我们只是在这个**核心理念**的基础上进行投资。

**Extra example:**
- The investment **thesis** was based on the growing demand for renewable energy.  
  这个投资**理念**基于对可再生能源日益增长的需求。

### recursive  /rɪˈkɜːrsɪv/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 3

**EN:** characterized by a process that repeats itself in a self-referential way, where each iteration builds upon the previous one  
**CN:** 递归的，自我循环改进的

**Original examples:**
- [15:26] It seems as though in the major labs we've reached this point—someone on your team said it recently—of like **recursive** self-improvement, where the models themselves are building and doing a lot of the research to do, you know, the next generation of improvement.  
  似乎在主要实验室中，我们已经达到了这样一个点——你们团队最近有人说过——就像**递归式**自我改进，模型本身正在构建并进行大量研究，以实现下一代的改进。
- [15:51] Result of you getting there first to this like **recursive** thing.  
  这是你们首先达到这种**递归式**改进的结果。
- [15:56] Like what, tell us how we should think about this idea of **recursive** self-improvement in the models themselves, because it seems like getting there first is incredibly important because then you just can continue to separate yourself versus those that haven't reached it yet.  
  告诉我们应该如何理解模型本身的**递归式**自我改进这个概念，因为首先达到这一点似乎极其重要，这样你就可以继续与那些尚未达到的竞争者拉开差距。

**Extra example:**
- The algorithm uses a **recursive** approach to solve complex problems by breaking them into smaller sub-problems.  
  该算法使用**递归**方法，通过将复杂问题分解为更小的子问题来解决。

### leverage  /ˈlevərɪdʒ/
**CEFR:** B2 | **Part of speech:** n./v. | **Occurrences:** 2

**EN:** the ability to multiply effectiveness or impact; to use something to maximum advantage  
**CN:** 杠杆作用，影响力；充分利用

**Original examples:**
- [17:43] How do you think about this weird world where you mentioned the talent and the **leverage** and they're not writing code themselves and Claude's writing its own code?  
  你如何看待这个奇怪的世界，你提到了人才和**杠杆作用**，他们自己不写代码，而是 Claude 在写自己的代码？
- [36:03] It's so unbelievably capital intensive to build these frontier labs. You've got the **leverage** we talked about, which is efficiency, price—both those things relate to margin.  
  建立这些前沿实验室的资本密集度令人难以置信。你拥有我们谈到的**杠杆作用**，也就是效率和价格——这两者都与利润率相关。

**Extra example:**
- Companies can **leverage** social media to reach a wider audience at lower cost.  
  公司可以**利用**社交媒体以更低的成本触达更广泛的受众。

### skeptical  /ˈskeptɪkəl/
**CEFR:** B2 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** having doubts or reservations; not easily convinced  
**CN:** 怀疑的，持怀疑态度的

**Original examples:**
- [20:50] Obviously a bunch of the authors of the scaling laws papers are amongst our founders, but you know, notwithstanding that, we can be a bit of a **skeptical** bunch.  
  显然，很多scaling laws论文的作者都是我们的创始人，但尽管如此，我们还是会持怀疑态度。

**Extra example:**
- Investors remained **skeptical** about the company's ambitious growth projections.  
  投资者对公司雄心勃勃的增长预测仍持怀疑态度。

### implication  /ˌɪmplɪˈkeɪʃn/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 6

**EN:** a possible effect or result of an action or decision; a conclusion that can be drawn from something  
**CN:** 影响，后果；含义

**Original examples:**
- [21:08] Like, if that continues to be true for however many more, you know, turns of the crank here, how do you do that thing of not thinking linear and thinking exponential yourself in your job and for the business? Like, the **implications** are really hard to reason through.  
  如果这种情况继续下去，无论还要经历多少次迭代，你如何在工作和业务中做到不线性思考而是指数级思考？这些**影响**真的很难推理清楚。
- [41:04] And on top of that we built an MFR, a monthly financial review skill, and it can produce our monthly financial review. It's 90 to 95% ready and then all of our discussion becomes about what do we do, what are the **implications**, not what exactly happened because Claude is  
  除此之外，我们还构建了一个 MFR，即月度财务审查技能，它可以生成我们的月度财务审查。它已经完成了 90% 到 95%，然后我们所有的讨论都变成了我们该做什么，**影响**是什么，而不是到底发生了什么，因为 Claude
- [41:21] Not just reporting the weather, it's also helping to think about drivers and like why did the number change in the way it did, and that gives you tremendous insight into the business, both in terms of this like MFR that we do, but also on a daily basis. And so what used to take hours to produce, you know, a weekly report for, you know, what's driving revenue or what's driving our compute utilization, now comes down to 30 minutes, and then we can spend our time on the actual strategic **implications** of the business.  
  它不仅仅是报告天气，还帮助思考驱动因素，比如数字为什么会以这种方式变化，这为你提供了对业务的巨大洞察，无论是在我们做的 MFR 方面，还是在日常基础上。所以过去需要花费数小时才能生成的每周报告，比如什么在驱动收入或什么在驱动我们的计算利用率，现在只需 30 分钟，然后我们可以把时间花在业务的实际战略**影响**上。
- [01:02:08] I do think that there's a balance, right? You want to be able to have innovation happen really quickly and have that not be slowed down, but you also want to have this kind of responsibility framework for how these things are deployed because we've long said the **implications** are significant.  
  我确实认为需要平衡，对吧？你希望创新能够快速发生而不被拖慢，但你也希望有一种责任框架来规范这些东西如何部署，因为我们长期以来一直说**影响**是重大的。
- [01:10:40] Everyone kind of becomes a manager and I think the **implications** of that and the productivity gain that can come from that when it's in the right form factor is we're very very early in that but the potential for it is incredible, crazy to imagine.  
  每个人都在某种程度上成为了管理者，我认为这带来的**影响**以及当它以正确的形式出现时可能带来的生产力提升——我们在这方面还处于非常早期的阶段，但它的潜力是难以置信的，难以想象。
- [01:15:34] But I think the commonality of it was that, you know, everything is going to happen much quicker than we think and that both the **implications** but also the opportunities are significant.  
  但我认为共同点是，一切都会比我们想象的发生得更快，**影响**和机遇都是巨大的。

**Extra example:**
- The new policy has far-reaching **implications** for small businesses.  
  新政策对小企业有深远的**影响**。

### granular  /ˈɡrænjələr/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 3

**EN:** characterized by a high level of detail or fine-grained analysis; broken down into small, specific components  
**CN:** 细粒度的，详细的，精细的

**Original examples:**
- [25:29] So we look at the compute down to a very **granular** level in terms of what it can deliver for us and when, and that's something that we do.  
  所以我们以非常**细粒度**的层面来审视计算资源，看它能为我们提供什么以及何时提供，这是我们在做的事情。
- [26:08] I'm always so curious by the metabolism of, in this case, Anthropic for new compute. Like, how fast could you take it? If I airdropped on you twice the compute that you have tomorrow, like, would you consume that? How fast would you consume that? If I airdropped 10 times the compute on top of you, how fast would you consume it? Can you calibrate us on that sort of thing? Like, is demand—actually, it feels like demand's unlimited between these three, you know, uses: training, internal, customer demand. Everyone's saying the same thing. Shortages everywhere, memory... We need to be very **granular** in our allocation.  
  我总是对 Anthropic 对新计算资源的代谢速度非常好奇。比如，你能多快消化它？如果我明天空投给你两倍的计算资源，你会消耗掉吗？多快能消耗掉？如果我空投给你 10 倍的计算资源，你多快能消耗掉？你能在这方面给我们一个校准吗？需求是不是——实际上，感觉需求在这三个用途之间是无限的：训练、内部使用、客户需求。每个人都在说同样的话。到处都是短缺，内存... 我们需要在分配上非常**精细**。
- [01:13:32] We track performance at a very **granular** level to optimize resource allocation.  
  我们以非常**细粒度**的层面跟踪性能，以优化资源分配。

**Extra example:**
- The report provides **granular** data on customer behavior across different regions.  
  该报告提供了不同地区客户行为的**详细**数据。

### calibration  /ˌkælɪˈbreɪʃn/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** the process of adjusting or fine-tuning something to achieve accuracy or optimal performance  
**CN:** 校准，调整，精细调节

**Original examples:**
- [27:19] We probably have the same kind of allocation or **calibration** that we do with compute today but it's  
  我们可能会采用与今天类似的算力分配或调配方式,但是
- [27:19] We probably have the same kind of allocation or **calibration** that we do with compute today but it's evolving.  
  我们可能有与今天对计算资源进行的相同的分配或**校准**方式，但它在不断演变。

**Extra example:**
- Regular **calibration** of the equipment ensures accurate measurements.  
  定期**校准**设备可确保测量准确。

### metabolism  /məˈtæbəlɪzəm/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the chemical processes in living organisms that convert food to energy; metaphorically, the rate at which resources are consumed  
**CN:** 新陈代谢；（比喻）资源消耗速度

**Original examples:**
- [26:08] I'm always so curious by the **metabolism** of, in this case, Anthropic for new compute. Like, how fast could you take it? If I airdropped on you twice the compute that you have tomorrow, like, would you consume that? How fast would you consume that? If I airdropped 10 times the compute on top of you, how fast would you consume it? Can you calibrate us on that sort of thing? Like, is demand—actually, it feels like demand's unlimited between these three, you know, uses: training, internal, customer demand. Everyone's saying the same thing. Shortages everywhere, memory...  
  我一直很好奇 Anthropic 对新算力的「代谢速度」。比如,你们能多快消化它?如果我明天空投给你们两倍于现有的算力,你们会用掉吗?多快能用完?如果我空投 10 倍的算力给你们,你们多快能消化掉?能不能给我们一个概念?感觉需求是无限的——训练、内部使用、客户需求这三个方面,大家都在说同样的话:到处都缺算力,内存...

**Extra example:**
- The startup's **metabolism** for burning through capital was unsustainable.  
  这家初创公司消耗资金的速度是不可持续的。

### fungibility  /ˌfʌndʒəˈbɪləti/
**CEFR:** C2 | **Part of speech:** n. | **Occurrences:** 2

**EN:** the property of being interchangeable or substitutable; the ability to replace one resource with another of the same type  
**CN:** 可替代性，可互换性

**Original examples:**
- [26:45] This goes back to like how we use it and the **fungibility** of it.  
  这又回到了我们如何使用它以及它的**可替代性**。
- [53:09] Here, you really have that **fungibility** that's possible, and I think that's where the return on compute is so important.  
  在这里，你真的拥有这种可能的**可替代性**，我认为这就是计算回报如此重要的地方。

**Extra example:**
- The **fungibility** of cloud computing resources allows companies to scale up or down quickly.  
  云计算资源的**可替代性**使公司能够快速扩展或缩减规模。

### constrained  /kənˈstreɪnd/
**CEFR:** B2 | **Part of speech:** adj. | **Occurrences:** 2

**EN:** restricted or limited by external factors or conditions  
**CN:** 受限的，受约束的

**Original examples:**
- [26:45] So the answer is, you know, we're **constrained** kind of across those use cases internally today.  
  所以答案是，我们目前在内部的这些使用场景中都受到限制。
- [33:22] I'm very curious why, if everyone is compute **constrained**, why everyone doesn't just raise prices a lot to try to find like what the right equilibrium is.  
  我很好奇，如果每个人都受到算力限制，为什么大家不直接大幅提价来找到合适的平衡点。

**Extra example:**
- The project timeline is **constrained** by budget limitations.  
  项目时间表受到预算限制的约束。

### tradeoff  /ˈtreɪdɔːf/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 2

**EN:** a balance achieved between two desirable but incompatible features; a compromise  
**CN:** 权衡，取舍

**Original examples:**
- [27:37] Back to the ways the customers are using Anthropic, one of the interesting tensions and **trade-offs** that I'm fascinated to hear how you think through is between sort of the platform approach where I build my business on top of Claude and it powers my thing versus you doing the thing that I wanted to build.  
  回到客户使用 Anthropic 的方式,有一个有趣的张力和权衡让我很想听听你们是怎么思考的:就是平台方式——我在 Claude 之上构建我的业务,它驱动我的产品——与你们直接做我想做的事情之间的权衡。
- [33:11] And so I'd love you to just like riff on pricing, like how you think about it, what the **trade-offs** are, why not raise prices a lot.  
  所以我很想听你谈谈定价，你是怎么考虑的，有哪些权衡，为什么不大幅提价。

**Extra example:**
- There's always a **tradeoff** between speed and quality.  
  速度和质量之间总是需要权衡。

### accrue  /əˈkruː/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to accumulate or receive over time, especially benefits or advantages  
**CN:** 积累，累积（尤指利益或优势）

**Original examples:**
- [28:19] Yeah, the way I would think about it is most of what we're building is platform, and we think that there's so many examples of where a platform can **accrue** a lot of value, but the customers who are building on that platform actually **accrue** even more value.  
  我的想法是，我们构建的大部分是平台，我们认为有很多例子表明平台可以积累大量价值，但在该平台上构建的客户实际上会积累更多价值。
- [30:25] So I think of our strategy as mostly horizontal. We'll build vertical where, you know, we think we have some value to add or a perspective that's useful or a way to demonstrate to the market how we think about our platform adding value, and a lot of the value is going to **accrue** to the customers that are building on top of it.  
  所以我认为我们的策略主要是横向的。我们会在认为有价值可以增加的地方构建纵向产品，或者有用的视角，或者向市场展示我们如何看待平台增加价值的方式，而大量价值将会积累到在其上构建的客户那里。

**Extra example:**
- Interest will **accrue** on your savings account monthly.  
  你的储蓄账户每月会累积利息。

### proliferate  /prəˈlɪfəreɪt/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to increase rapidly in number; to spread widely  
**CN:** 激增，扩散，迅速繁殖

**Original examples:**
- [30:49] And services that allow that intelligence to **proliferate** within customers.  
  以及让这种智能在客户内部扩散的服务。
- [34:52] We want that to just continue because our goal is to **proliferate** this throughout the ecosystem.  
  我们希望这种情况持续下去，因为我们的目标是让它在整个生态系统中扩散。

**Extra example:**
- Smartphones have **proliferated** rapidly over the past decade.  
  智能手机在过去十年中迅速普及。

### paradox  /ˈpærədɑːks/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 3

**EN:** a seemingly contradictory statement or situation that may nonetheless be true  
**CN:** 悖论，似非而是的论点

**Original examples:**
- [35:19] The changing of the pricing for Opus actually, you know, you see this Jevons **paradox**, right?  
  Opus 的定价变化实际上，你知道，你会看到这个 Jevons 悖论，对吧？
- [35:50] And we also think that pricing to get that value and to see that kind of Jevons **paradox** happen is really important.  
  我们也认为，通过定价来获得那种价值并看到 Jevons 悖论发生是非常重要的。
- [44:36] People at the company, but it has made even those incredibly talented people so much more productive, and there's a little bit of this—I think of it again like Jevons **paradox**, but for labor—which is that we have people who become incredibly more productive.  
  公司里的人，但它让那些极具才华的人变得更加高效，这有点像——我再次想到 Jevons 悖论，但是针对劳动力——也就是说我们的人变得极其高效。

**Extra example:**
- It's a **paradox** that the more choices we have, the harder it is to decide.  
  这是一个悖论：我们的选择越多，就越难做决定。

### robust  /roʊˈbʌst/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 4

**EN:** strong, healthy, and unlikely to fail or weaken  
**CN:** 强劲的，稳健的，健壮的

**Original examples:**
- [37:00] I will say our returns on that compute expense today are **robust**.  
  我要说的是，我们今天在算力支出上的回报是强劲的。
- [37:06] They're **robust** and we think of it as what is the return on that full envelope of compute.  
  它们是强劲的，我们把它看作是整个算力投入的回报。
- [39:11] We have a really **robust** first-party business as well.  
  我们也有一个非常稳健的第一方业务。
- [01:20:14] You know, the financial aid packages weren't, you know, as **robust** as they are today.  
  你知道，当时的助学金方案没有像今天这样完善。

**Extra example:**
- The company has a **robust** growth strategy for the next five years.  
  该公司有一个未来五年的强劲增长战略。

### envelope  /ˈenvəloʊp/
**CEFR:** B1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a flat paper container for a letter; in technical contexts, the outer boundary or limit of capability  
**CN:** 信封；（技术语境）外部边界，能力范围

**Original examples:**
- [37:06] They're robust and we think of it as what is the return on that full **envelope** of compute.  
  它们很稳健,我们把它看作是整个算力投入的回报率。

**Extra example:**
- The total **envelope** of compute resources determines what projects we can run simultaneously.  
  计算资源的总容量决定了我们可以同时运行哪些项目。

### insight  /ˈɪnsaɪt/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 2

**EN:** a deep understanding of a person, situation, or subject; a valuable piece of information or understanding  
**CN:** 洞察力，深刻见解

**Original examples:**
- [41:21] Not just reporting the weather, it's also helping to think about drivers and like why did the number change in the way it did, and that gives you tremendous **insight** into the business, both in terms of this like MFR that we do, but also on a daily basis.  
  不仅仅是报告天气，它还帮助思考驱动因素，比如数字为什么会这样变化，这让你对业务有了巨大的洞察，无论是在我们做的 MFR 方面，还是在日常基础上。
- [41:47] We can also get it in the hands of business, you know, leaders much more quickly, and so it's just meant that the **insight** engine is a lot faster within the company.  
  我们也可以更快地把它交到业务领导者手中，所以这意味着公司内部的洞察引擎快了很多。

**Extra example:**
- Her research provided valuable **insights** into consumer behavior.  
  她的研究为消费者行为提供了宝贵的洞察。

### dystopian  /dɪsˈtoʊpiən/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** relating to an imagined society that is undesirable or frightening, typically characterized by oppression or suffering  
**CN:** 反乌托邦的，令人恐惧的

**Original examples:**
- [44:08] But it feels like ever so slightly **dystopian** to me that that reality is coming quickly.  
  但对我来说，这个现实快速到来感觉有点反乌托邦。

**Extra example:**
- The novel paints a **dystopian** future where privacy no longer exists.  
  这部小说描绘了一个隐私不复存在的反乌托邦未来。

### interpretability  /ɪnˌtɜːrprɪtəˈbɪləti/
**CEFR:** C2 | **Part of speech:** n. | **Occurrences:** 2

**EN:** the quality of being able to be understood or explained, especially regarding how a system or model works internally  
**CN:** 可解释性；可理解性（尤指系统或模型内部运作方式）

**Original examples:**
- [48:31] We invest in research not just in model development but also in AI safety research, right? Like we've pioneered **interpretability**, which is, think of it as like an MRI for the model to see inside the neural network how it works.  
  我们不仅投资于模型开发研究，还投资于AI安全研究，对吧？比如我们在**可解释性**方面是先驱，可以把它想象成模型的核磁共振成像，能看到神经网络内部是如何运作的。
- [49:21] When you have this investment that we've made and will continue to make in safety, **interpretability**, alignment, like that actually inures to the benefit of the enterprise.  
  当你拥有我们已经并将继续在安全、**可解释性**、对齐方面所做的投资时，这实际上会让企业受益。

**Extra example:**
- The lack of **interpretability** in deep learning models remains a major challenge for regulatory approval.  
  深度学习模型缺乏**可解释性**仍然是获得监管批准的主要挑战。

### alignment  /əˈlaɪnmənt/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** in AI context, the process of ensuring that AI systems behave in accordance with human values and intentions  
**CN:** 对齐；一致性（在AI语境中指确保AI系统的行为符合人类价值观和意图）

**Original examples:**
- [48:44] We pioneered **alignment** science.  
  我们是**对齐**科学的先驱。
- [49:21] When you have this investment that we've made and will continue to make in safety, interpretability, **alignment**, like that actually inures to the benefit of the enterprise.  
  当你拥有我们已经并将继续在安全、可解释性、**对齐**方面所做的投资时，这实际上会让企业受益。

**Extra example:**
- Achieving **alignment** between AI objectives and human values is critical for safe deployment.  
  实现AI目标与人类价值观之间的**对齐**对于安全部署至关重要。

### inure  /ɪˈnjʊr/
**CEFR:** C2 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to come into effect or operation; to accrue benefit to someone (formal/legal)  
**CN:** 生效；使受益（正式/法律用语）

**Original examples:**
- [49:21] The more and more of these businesses are running on cloud and our cloud platform. When you have this investment that we've made and will continue to make in safety, interpretability, alignment, like that actually **inures** to the benefit of the enterprise.  
  越来越多的业务运行在云和我们的云平台上。当你在安全、可解释性、对齐方面进行了我们已经做的和将继续做的投资,这实际上对企业有利。

**Extra example:**
- The safety improvements will **inure** to the benefit of all enterprise users.  
  安全性改进将使所有企业用户受益。

### tangible  /ˈtændʒəbəl/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** clear and definite; able to be perceived or realized as real  
**CN:** 明确的，实实在在的；可感知的

**Original examples:**
- [57:33] I think that all of those things are part of the promise and the potential of AI, and so we could probably do a better job of painting that picture and we want to show more **tangible** results for that over time.  
  我认为所有这些都是AI的承诺和潜力的一部分，所以我们可能可以更好地描绘这幅图景，我们希望随着时间的推移展示更多实实在在的成果。

**Extra example:**
- The company needs to deliver **tangible** benefits to justify the investment.  
  公司需要提供实实在在的好处来证明这项投资的合理性。

### articulation  /ɑːrˌtɪkjuˈleɪʃn/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the clear and effective expression of ideas or feelings  
**CN:** 清晰表达；明确阐述

**Original examples:**
- [58:27] So I think it's about a clear **articulation** of the opportunities.  
  所以我认为这关乎对机遇的清晰**阐述**。

**Extra example:**
- The CEO's **articulation** of the company's vision inspired the entire team.  
  CEO对公司愿景的清晰**阐述**激励了整个团队。

### transparent  /trænsˈpærənt/
**CEFR:** B2 | **Part of speech:** adj. | **Occurrences:** 3

**EN:** open and honest, without secrets; easy to perceive or understand  
**CN:** 透明的；坦诚的；易于理解的

**Original examples:**
- [58:45] And then I think it's being **transparent** about that, about both of those things when we talk about it.  
  然后我认为要对此保持**透明**，当我们谈论这些事情时，对这两方面都要透明。
- [01:05:22] Being **transparent** about organizational culture and communication approach.  
  对组织文化和沟通方式保持**透明**。
- [01:07:05] Maintaining **transparent** practices in all operations.  
  在所有运营中保持**透明**的做法。

**Extra example:**
- The company maintains a **transparent** pricing policy with no hidden fees.  
  该公司保持**透明**的定价政策，没有隐藏费用。

### phased  /feɪzd/
**CEFR:** B2 | **Part of speech:** adj. | **Occurrences:** 2

**EN:** planned or carried out in stages over a period of time  
**CN:** 分阶段的；逐步的

**Original examples:**
- [01:00:02] And so we have this **phased** approach to it because we think that when a model is this capable, and again cyber is the thing that people focused on, but you know there are other things as well.  
  所以我们采取了这种**分阶段**的方法，因为我们认为当模型如此强大时，网络安全是人们关注的重点，但你知道还有其他方面。
- [01:00:33] We didn't say we're never going to release it. We said let's do it in a **phased** way.  
  我们没有说永远不会发布它。我们说让我们以**分阶段**的方式来做。

**Extra example:**
- The company implemented a **phased** rollout of the new software to minimize disruption.  
  公司实施了新软件的**分阶段**推出，以最大限度地减少干扰。

### navigate  /ˈnævɪɡeɪt/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to find a way through or manage a difficult or complicated situation  
**CN:** 驾驭；应对（复杂或困难的情况）

**Original examples:**
- [01:01:23] How do you **navigate** that stuff? And some of it's just, I guess, beyond your control, but I'm sure you're trying to.  
  你如何**应对**这些事情？我想有些事情超出了你的控制范围，但我相信你在努力。

**Extra example:**
- Leaders must **navigate** complex regulatory environments while maintaining business growth.  
  领导者必须在保持业务增长的同时**应对**复杂的监管环境。

### underpin  /ˌʌndərˈpɪn/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to support, justify, or form the basis for something  
**CN:** 支撑；构成……的基础；巩固

**Original examples:**
- [01:06:08] And I think the culture **underpins** the reason why we've been able to attract and retain some of the best talent in the industry, right?  
  我认为文化**支撑**着我们能够吸引和留住业内一些最优秀人才的原因，对吧？
- [01:06:43] **Underpinned** by the culture, and that's not just something we feel. It's like empirically when you talk to people, it's, you know, I want to have the most impact possible.  
  由文化**支撑**，这不仅仅是我们的感觉。从实证角度来看，当你和人们交谈时，他们会说，你知道，我想产生尽可能大的影响。

**Extra example:**
- Strong ethical principles **underpin** all of our business decisions.  
  强有力的道德原则**支撑**着我们所有的商业决策。

### precedent  /ˈpresɪdənt/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** an earlier event or action that is used as an example or guide for similar situations  
**CN:** 先例；前例；惯例

**Original examples:**
- [01:14:14] That was a harrowing time, but it was also a time kind of without **precedent**, right?  
  那是一段艰难的时期,但也是一段几乎没有先例可循的时期,对吧?

**Extra example:**
- This legal case could set a **precedent** for how AI companies handle data privacy.  
  这个法律案例可能为AI公司如何处理数据隐私树立**先例**。

### diffusion  /dɪˈfjuːʒən/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** the process by which something spreads or is spread widely  
**CN:** 扩散,传播(指某事物广泛传播的过程)

**Original examples:**
- [01:16:29] I think the first thing would be the **diffusion** rate within our customers, the use cases are playing.  
  我认为首先要看的是在我们客户中的**扩散**速度,以及用例的发展情况。
- [01:16:57] I think the first thing would be the **diffusion** rate within our customers, the use cases are playing.  
  我认为首先要看的是在我们客户中的**扩散**速度,以及用例的发展情况。

**Extra example:**
- The **diffusion** of new technology across industries takes time.  
  新技术在各行业的**扩散**需要时间。

### trajectory  /trəˈdʒektəri/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** the path or direction that something follows over time  
**CN:** 轨迹,发展路径(指某事物随时间发展的方向)

**Original examples:**
- [01:17:11] I sure hope you're right. It sure seems like we're on that **trajectory** and it's quite a future to imagine.  
  我真心希望你是对的。看起来我们确实在那条**轨迹**上,这是一个很值得想象的未来。
- [01:19:24] I sure hope you're right. It sure seems like we're on that **trajectory** and it's quite a future to imagine.  
  我真心希望你是对的。看起来我们确实在那条**轨迹**上,这是一个很值得想象的未来。

**Extra example:**
- The company's growth **trajectory** has exceeded all expectations.  
  公司的增长**轨迹**超出了所有预期。

### compound  /kəmˈpaʊnd/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to increase or accumulate by adding new elements over time  
**CN:** 复合增长,累积(指随时间不断叠加增长)

**Original examples:**
- [01:21:00] You know how small advantages **compound** over time? That's true in investing and just as true in how you run your company.  
  你知道小优势如何随时间**复合增长**吗?这在投资中是真理,在经营公司时同样如此。

**Extra example:**
- Interest will **compound** annually on your savings account.  
  你的储蓄账户利息会按年**复利计算**。

### procure  /prəˈkjʊr/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 3

**EN:** to obtain something, especially with effort or through a formal process  
**CN:** 采购,获取(尤指通过努力或正式流程获得某物)

**Original examples:**
- [01:26] Look, the compute that we **procure** is the lifeblood of our business.  
  你看,我们**采购**的算力是我们业务的命脉。
- [02:17] And then as we go out and actually do these deals to **procure** compute, you know, flexibility is really important to us and so we build that flexibility into the deals themselves.  
  然后当我们真正去做这些交易来**采购**算力时,灵活性对我们来说非常重要,所以我们把这种灵活性构建到交易本身中。
- [02:17] We think about the compute we need to stay at the frontier and we really look ahead and try to estimate that. And then as we go out and actually do these deals to **procure** compute, you know, flexibility is really important to us and so we build that flexibility into the deals themselves.  
  我们考虑保持在前沿所需的算力,并真正向前看,尝试估算。然后当我们实际去做这些采购交易时,灵活性对我们来说非常重要,所以我们把灵活性构建到交易本身中。

**Extra example:**
- The company needs to **procure** additional equipment for the new project.  
  公司需要为新项目**采购**额外的设备。

### lifeblood  /ˈlaɪfblʌd/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the most important thing that something needs in order to exist or function  
**CN:** 命脉,生命线(指某事物存在或运作所必需的最重要因素)

**Original examples:**
- [01:26] Look, the compute that we procure is the **lifeblood** of our business.  
  你看,我们采购的算力是我们业务的**命脉**。

**Extra example:**
- Innovation is the **lifeblood** of any technology company.  
  创新是任何科技公司的**命脉**。

### fungibly  /ˈfʌndʒəbli/
**CEFR:** C2 | **Part of speech:** adv. | **Occurrences:** 1

**EN:** in a way that allows items to be exchanged or substituted for one another  
**CN:** 可互换地,可替代地(指物品可以相互交换或替代的方式)

**Original examples:**
- [03:02] You know, we use these chips **fungibly**.  
  你知道,我们**可互换地**使用这些芯片。

**Extra example:**
- The system allows resources to be allocated **fungibly** across different projects.  
  该系统允许资源在不同项目之间**可互换地**分配。

### orchestration  /ˌɔːrkɪˈstreɪʃən/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the coordination and arrangement of complex activities or systems  
**CN:** 编排,协调(指对复杂活动或系统的协调和安排)

**Original examples:**
- [04:01] And so we really built this **orchestration** layer that gives us that flexibility to use all different types of compute and in doing so we also are able to get the most value out of it.  
  所以我们真正构建了这个**编排**层,让我们能够灵活使用所有不同类型的算力,这样做我们也能从中获得最大价值。

**Extra example:**
- The **orchestration** of multiple cloud services requires sophisticated automation.  
  多个云服务的**编排**需要复杂的自动化技术。

### multiplier  /ˈmʌltɪplaɪər/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 2

**EN:** a factor by which something increases or is increased in quantity or effect  
**CN:** 乘数，倍增因子

**Original examples:**
- [09:21] And so, if you look at going from Opus, you know, 4 to 4.5, 4.6, and now 4.7, you know, each one of those leaps, they're not equal, but each one of those leaps to a new model has a **multiplier** in terms of how much more efficient it is at processing tokens effectively.  
  所以如果你看从 Opus 4 到 4.5、4.6,再到现在的 4.7,每一次跃升——虽然幅度不完全相同——但每次升级到新模型,在处理 token 的效率上都有倍数级的提升。
- [09:21] And so, if you look at going from Opus, you know, 4 to 4.5, 4.6, and now 4.7, you know, each one of those leaps, they're not equal, but each one of those leaps to a new model has a **multiplier** in terms of how much more efficient it is at processing tokens effectively.  
  所以，如果你看从 Opus 4 到 4.5、4.6，现在到 4.7 的发展，你知道，这些跃升并不相同，但每一次跃升到新模型都有一个倍增效应，体现在处理 token 的效率提升上。

**Extra example:**
- Education is often seen as a **multiplier** for economic growth.  
  教育常被视为经济增长的倍增器。

### dynamically  /daɪˈnæmɪkli/
**CEFR:** C1 | **Part of speech:** adv. | **Occurrences:** 2

**EN:** in a way that is characterized by constant change, activity, or progress; in real-time or flexibly  
**CN:** 动态地，灵活地

**Original examples:**
- [08:26] And because we can allocate that compute so **dynamically**, we can make changes.  
  因为我们可以如此动态地分配算力，所以我们可以做出调整。
- [10:09] And then when we're in between generations, we're **dynamically** deploying efficiency improvements kind of in between these like more step function model changes.  
  然后在两代模型之间，我们会动态部署效率改进，就在这些阶跃式模型变化之间。

**Extra example:**
- The website **dynamically** adjusts its layout based on screen size.  
  这个网站会根据屏幕尺寸动态调整布局。

### inference  /ˈɪnfərəns/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** in AI/ML, the process of using a trained model to make predictions or generate outputs; generally, a conclusion reached based on evidence and reasoning  
**CN:** 推理；（AI领域）推断，模型推理

**Original examples:**
- [09:38] And that just doesn't serve customers. That also helps us internally as well, because if you think about if we're using the model for, if we're doing reinforcement learning on the model, it's basically **inference** within a sandbox with a reward function.  
  这不仅服务客户，对我们内部也有帮助，因为如果你想想，如果我们在模型上做强化学习，本质上就是在一个带有奖励函数的沙盒环境中进行推理。
- [09:52] Right? And so if the model's better at more efficient **inference**, that RL is more efficient as well.  
  对吧？所以如果模型的推理效率更高，那么强化学习也会更高效。

**Extra example:**
- The detective made an **inference** based on the evidence at the crime scene.  
  侦探根据犯罪现场的证据做出了推断。

### reinforcement  /ˌriːɪnˈfɔːrsmənt/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the action or process of strengthening or supporting something; in AI, a learning method where models improve through reward-based feedback  
**CN:** 强化，加强；（AI领域）强化学习

**Original examples:**
- [09:38] And that just doesn't serve customers. That also helps us internally as well, because if you think about if we're using the model for, if we're doing **reinforcement** learning on the model, it's basically inference within a sandbox with a reward function.  
  这不仅服务客户，对我们内部也有帮助，因为如果你想想，如果我们在模型上做强化学习，本质上就是在一个带有奖励函数的沙盒环境中进行推理。

**Extra example:**
- Positive **reinforcement** is an effective teaching strategy.  
  正向强化是一种有效的教学策略。

### saturate  /ˈsætʃəreɪt/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to reach a point where no more can be absorbed or accepted; to fill completely  
**CN:** 使饱和，使达到极限

**Original examples:**
- [13:19] Everyone publishes their model benchmark cards and finds that a lot of those benchmarks are **saturated**.  
  每个人都发布他们的模型基准测试卡，然后发现很多基准测试都已经饱和了。

**Extra example:**
- The market has become **saturated** with similar products.  
  市场上已经充斥着类似的产品。

### forego  /fɔːrˈɡoʊ/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to go without something; to give up or renounce  
**CN:** 放弃，舍弃

**Original examples:**
- [16:36] We **forego** potential revenue to allocate compute internally.  
  我们放弃潜在收入，将算力分配到内部使用。

**Extra example:**
- She decided to **forego** dessert to maintain her diet.  
  她决定放弃甜点以保持饮食控制。

### accentuate  /əkˈsentʃueɪt/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to make something more noticeable or prominent; to emphasize  
**CN:** 强调，突出

**Original examples:**
- [18:47] And so I think of it as **accentuating** and accelerating the talent that we already have.  
  所以我认为这是在强化和加速我们已有的人才优势。

**Extra example:**
- The lighting was designed to **accentuate** the architectural features.  
  灯光设计旨在突出建筑特色。

### consensus  /kənˈsensəs/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a general agreement or shared opinion among a group  
**CN:** 共识，一致意见

**Original examples:**
- [19:15] There's a growing **consensus** in the AI community about scaling laws.  
  AI 社区对于规模定律正在形成越来越多的共识。

**Extra example:**
- The team reached a **consensus** on the project timeline.  
  团队就项目时间表达成了共识。

### prior  /ˈpraɪər/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 2

**EN:** a belief or assumption held before new evidence is considered (often used in statistics and reasoning)  
**CN:** 先验信念，预设假设(常用于统计学和推理中)

**Original examples:**
- [21:39] And then having a very low bar for updating your current **prior** or your current perspective, because it could be the case that something a month ago was true that's just not true today and that breaks your model.  
  然后对更新你当前的**先验假设**或当前观点保持很低的门槛,因为可能一个月前正确的事情今天就不再正确了,这会打破你的模型。
- [21:39] And then having a very low bar for updating your current **prior** or your current perspective, because it could be the case that something a month ago was true that's just not true today and that breaks your model.  
  然后要对更新当前认知或观点保持非常低的门槛,因为可能一个月前还成立的事情今天就不成立了,这会打破你的模型。

**Extra example:**
- In Bayesian statistics, you start with a **prior** probability and update it as new data comes in.  
  在贝叶斯统计中,你从一个**先验概率**开始,然后随着新数据的到来不断更新它。

### analog  /ˈænəlɔːɡ/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** something that is comparable to something else in significant respects; a parallel or equivalent  
**CN:** 类比物,相似物;可比较的事物

**Original examples:**
- [22:21] Revenue, and you know, it was a little hard to predict that, but now we can use coding as an **analog** for a lot of what's happening elsewhere in the economy and elsewhere in our business.  
  收入方面,你知道,这有点难以预测,但现在我们可以用编程作为**类比**来理解经济中其他领域以及我们业务其他方面正在发生的事情。

**Extra example:**
- The company's expansion in Asia serves as an **analog** for how they might approach the European market.  
  该公司在亚洲的扩张可以作为他们如何进入欧洲市场的**类比**。

### assess  /əˈses/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 3

**EN:** to evaluate or estimate the nature, ability, or quality of something  
**CN:** 评估,评价,估定

**Original examples:**
- [23:23] So we have a process to **assess**, and that same process, by the way, we use to **assess** longer-term deals as well.  
  所以我们有一个**评估**流程,顺便说一下,我们用同样的流程来**评估**长期交易。
- [23:42] So we have a process to **assess**, and that same process, by the way, we use to **assess** longer-term deals as well.  
  所以我们有一个**评估**流程,顺便说一下,我们用同样的流程来**评估**长期交易。
- [23:42] So we have a process to **assess**, and that same process, by the way, we use to assess longer-term deals as well.  
  所以我们有一套评估流程,顺便说一句,我们用同样的流程来评估长期合约。

**Extra example:**
- The team needs to **assess** the risks before moving forward with the project.  
  团队需要在推进项目之前**评估**风险。

### heterogeneous  /ˌhetərəˈdʒiːniəs/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** diverse in character or content; consisting of dissimilar or diverse elements  
**CN:** 异质的,由不同成分组成的,多样化的

**Original examples:**
- [26:55] And you know I would say that you know a year or two ago it would have been harder to consume especially like a **heterogeneous** kind of compute drop in your example really quickly because these chip platforms are different and they are different, some are harder to operate, some of them have you know idiosyncrasies in terms of how we use it.  
  你知道我想说的是,一两年前,要快速消化特别是像你例子中那种**异构**计算资源会更困难,因为这些芯片平台是不同的,它们确实不同,有些更难操作,有些在使用方式上有特殊性。

**Extra example:**
- The data center manages a **heterogeneous** infrastructure with servers from multiple vendors.  
  该数据中心管理着一个包含多个供应商服务器的**异构**基础设施。

### idiosyncrasy  /ˌɪdiəˈsɪŋkrəsi/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a distinctive or peculiar feature or characteristic of something  
**CN:** 特性,特质;独特之处

**Original examples:**
- [26:55] And you know I would say that you know a year or two ago it would have been harder to consume especially like a heterogeneous kind of compute drop in your example really quickly because these chip platforms are different and they are different, some are harder to operate, some of them have you know **idiosyncrasies** in terms of how we use it.  
  你知道我想说的是,一两年前,要快速消化特别是像你例子中那种异构计算资源会更困难,因为这些芯片平台是不同的,它们确实不同,有些更难操作,有些在使用方式上有**特殊性**。

**Extra example:**
- Every programming language has its own **idiosyncrasies** that developers need to learn.  
  每种编程语言都有开发者需要学习的**特性**。

### emulate  /ˈemjuleɪt/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to match or surpass by imitation; to copy or reproduce the behavior or actions of  
**CN:** 效仿,模仿;力求赶上

**Original examples:**
- [29:43] The second is thinking about ways to demonstrate value for the ecosystem that others might **emulate**.  
  第二点是思考如何为生态系统展示价值,让其他人可以**效仿**。

**Extra example:**
- Many startups try to **emulate** the success of Silicon Valley giants.  
  许多初创公司试图**效仿** Silicon Valley 巨头的成功。

### incremental  /ˌɪŋkrəˈmentl/
**CEFR:** B2 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** relating to or denoting an increase or addition by regular degrees or in small amounts  
**CN:** 递增的,逐步增加的,增量的

**Original examples:**
- [37:36] And so this idea of a variable cost that's like on the **incremental** to serve a customer is a little bit like it doesn't really fit our model.  
  所以这种可变成本的概念,就像服务客户的**增量**成本,有点不太符合我们的模式。

**Extra example:**
- The company achieved success through **incremental** improvements rather than radical changes.  
  该公司通过**渐进式**改进而非激进变革取得了成功。

### multifaceted  /ˌmʌltiˈfæsɪtɪd/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** having many different aspects or features; complex and varied  
**CN:** 多方面的,多层面的,复杂的

**Original examples:**
- [39:15] But these are **multifaceted** partnerships, whether it be on developing the chips themselves, landing that capacity, serving it, and then ultimately distributing it to the customers.  
  但这些是**多方面的**合作关系,无论是开发芯片本身、获取产能、提供服务,还是最终分发给客户。

**Extra example:**
- Climate change is a **multifaceted** problem requiring solutions across technology, policy, and behavior.  
  气候变化是一个**多方面的**问题,需要在技术、政策和行为等方面寻求解决方案。

### repository  /rɪˈpɑːzətɔːri/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a place where things are stored and can be found; a central location for information or resources  
**CN:** 存储库，仓库；信息或资源的集中存放地

**Original examples:**
- [40:53] We now have a library of skills for Claude that are specific to finance. I think last I checked there were over 70 of them that everyone can access through this common **repository**.  
  我们现在有一个专门针对金融领域的Claude技能库。我记得上次查看时有超过70个技能，每个人都可以通过这个共享**存储库**访问它们。

**Extra example:**
- GitHub is the world's largest code **repository** for open-source projects.  
  GitHub是全球最大的开源项目代码**存储库**。

### automate  /ˈɔːtəmeɪt/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to make a process or system operate automatically without human intervention  
**CN:** 使自动化，使自动运行

**Original examples:**
- [42:42] As your business scales up, everything gets more complex, especially your compliance and security needs. With so many tools offering band-aids and patches, it's unfortunately far too easy for something to slip through the cracks. Fortunately, Vanta is a powerful tool designed to simplify and **automate** your security work and deliver a single  
  随着你的业务规模扩大,一切都会变得更加复杂,尤其是合规和安全需求。市面上有太多工具只提供临时性的修补方案,很容易就会有东西从缝隙中漏掉。幸运的是,Vanta 是一个强大的工具,旨在简化和自动化你的安全工作,并为合规和风险提供单一的……
- [42:42] As your business scales up, everything gets more complex, especially your compliance and security needs. With so many tools offering band-aids and patches, it's unfortunately far too easy for something to slip through the cracks. Fortunately, Vanta is a powerful tool designed to simplify and **automate** your security work and deliver a single  
  随着你的业务规模扩大,一切都会变得更加复杂,尤其是合规和安全需求。市面上有太多工具只提供临时性的修补方案,很容易就会有东西从缝隙中漏掉。幸运的是,Vanta 是一个强大的工具,旨在简化和自动化你的安全工作,并为合规和风险提供单一的……

**Extra example:**
- Many companies are trying to **automate** their customer service with AI chatbots.  
  许多公司正在尝试用AI聊天机器人**自动化**他们的客户服务。

### compliance  /kəmˈplaɪəns/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 3

**EN:** the act of obeying rules, regulations, or standards; conformity with requirements  
**CN:** 合规，遵守；符合规定或标准

**Original examples:**
- [42:42] As your business scales up, everything gets more complex, especially your **compliance** and security needs.  
  随着业务规模扩大，一切都变得更加复杂，尤其是你的**合规**和安全需求。
- [43:09] Vanta is a powerful tool designed to simplify and automate your security work and deliver a single source of truth for **compliance**.  
  Vanta是一个强大的工具，旨在简化和自动化你的安全工作，并为**合规**提供单一可信来源。
- [43:38] One unified platform that automates away that complexity across portfolio accounting, reconciliation, reporting, trading, **compliance**, and more, all at scale.  
  一个统一的平台可以自动化处理投资组合会计、对账、报告、交易、**合规**等各方面的复杂性，并且能够大规模运行。

**Extra example:**
- The company was fined for failing to maintain **compliance** with environmental regulations.  
  该公司因未能保持对环境法规的**合规**而被罚款。

### accelerant  /əkˈselərænt/
**CEFR:** C2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a substance or factor that speeds up a process or increases the rate of change  
**CN:** 促进剂，加速剂；加快进程的因素

**Original examples:**
- [45:17] So I actually think of it, you know, maybe even more optimistically, that it is an **accelerant** to our productivity, and that actually means that we can get a lot more done.  
  所以我实际上更乐观地认为，它是我们生产力的**加速剂**，这意味着我们可以完成更多的工作。

**Extra example:**
- The new funding acted as an **accelerant** for the startup's growth.  
  新的资金注入成为这家初创公司增长的**加速剂**。

### entrust  /ɪnˈtrʌst/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 2

**EN:** to give someone the responsibility of taking care of something or someone; to place confidence in  
**CN:** 委托，托付；信赖

**Original examples:**
- [49:15] They're interacting with their employees, sometimes even interacting with their customers as well. That is, those are the most sensitive workloads that they **entrust** to us.  
  他们与员工互动，有时甚至与客户互动。这些是他们**委托**给我们的最敏感的工作负载。
- [49:36] Customers as well, because all of our customers, if they're going to **entrust** us with all that access and all that data and the ability to work in the most sensitive workflows within their company, they want a company that they can trust.  
  客户也是如此，因为我们所有的客户，如果他们要**委托**我们访问所有数据，并在公司内部最敏感的工作流程中工作，他们需要一个可以信赖的公司。

**Extra example:**
- Parents **entrust** teachers with the education and safety of their children.  
  家长将孩子的教育和安全**托付**给老师。

### jarring  /ˈdʒɑːrɪŋ/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** shocking, disturbing, or creating an unpleasant or discordant effect  
**CN:** 刺耳的，令人不快的；令人震惊的

**Original examples:**
- [56:51] And going back to, you know, humans thinking in terms of exponentials versus linear, that can be **jarring**, I think.  
  回到人类以指数而非线性方式思考这一点，我认为这可能会让人感到**不适应**。

**Extra example:**
- The sudden shift in tone was **jarring** and caught everyone off guard.  
  语气的突然转变令人**不适**，让所有人措手不及。

### blueprint  /ˈbluːprɪnt/
**CEFR:** B2 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a detailed plan or model for achieving something; originally a technical drawing  
**CN:** 蓝图，详细计划；设计方案

**Original examples:**
- [58:37] We don't have the **blueprint** that's going to solve everything, but to at least have that dialogue about some of the risks and downsides and what we can do to address it.  
  我们没有能够解决一切问题的**蓝图**，但至少可以就一些风险和缺点以及我们能做些什么来应对进行对话。

**Extra example:**
- The architect presented a detailed **blueprint** for the new office building.  
  建筑师展示了新办公楼的详细**设计图**。

### misconstrue  /ˌmɪskənˈstruː/
**CEFR:** C1 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to interpret or understand something incorrectly; to misunderstand the meaning or intention  
**CN:** 误解，曲解；错误理解

**Original examples:**
- [59:36] Yeah. I think one of the things about Mythos is that people maybe **misconstrue** it as just a cyber model.  
  是的。我认为关于Mythos的一个问题是，人们可能会**误解**它只是一个网络安全模型。

**Extra example:**
- Please don't **misconstrue** my silence as agreement with your proposal.  
  请不要**误解**我的沉默为对你提案的同意。

### spike  /spaɪk/
**CEFR:** B2 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to increase suddenly and sharply  
**CN:** 急剧上升，激增

**Original examples:**
- [59:47] What we found was that cyber in particular was a place where it **spiked**.  
  我们发现网络安全领域尤其是一个能力激增的地方。

**Extra example:**
- Sales **spiked** dramatically after the product launch.  
  产品发布后销售额急剧上升。

### cognizant  /ˈkɑːɡnɪzənt/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** aware of and understanding something; having knowledge of  
**CN:** 意识到的，认识到的

**Original examples:**
- [01:00:46] But because of this one particular area, we wanted to be **cognizant** of that in how we released it.  
  但由于这个特定领域，我们希望在发布时对此保持充分认识。

**Extra example:**
- Leaders must be **cognizant** of the impact their decisions have on employees.  
  领导者必须意识到他们的决策对员工的影响。

### overseer  /ˈoʊvərsɪr/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a person who supervises or monitors something; someone in a position of authority  
**CN:** 监督者，监管者

**Original examples:**
- [01:01:30] Maybe talk about those two examples of like the government now as a very relevant partner, player, you know, **overseer**, etc.  
  也许可以谈谈这两个例子，比如政府现在作为一个非常相关的合作伙伴、参与者、监管者等等。

**Extra example:**
- The regulatory body acts as an **overseer** of financial institutions.  
  监管机构充当金融机构的监督者。

### fiefdom  /ˈfiːfdəm/
**CEFR:** C2 | **Part of speech:** n. | **Occurrences:** 2

**EN:** an organization or department over which someone exercises control as if it were their own private territory  
**CN:** 领地，势力范围（比喻某人控制的部门或组织）

**Original examples:**
- [01:04:57] After that happens, there's real alignment. So in something like compute allocation we were talking about before, people might have different perspectives on how to allocate that compute, but they will engage in a thoughtful discussion about where the returns are the highest or the best. And when they do that, you know, and we come to a decision, then there's alignment on it. There's not second-guessing, there's not this kind of politics or **fiefdom**. The other piece of it is it's remarkably transparent, the culture, right? So  
  这之后,就会形成真正的共识。比如我们之前提到的算力分配问题,大家可能对如何分配算力有不同看法,但他们会进行深思熟虑的讨论,探讨哪里能获得最高或最好的回报。当他们这样做并最终做出决定时,就会达成共识。不会有事后质疑,也不会有那种政治斗争或山头主义。另一个方面是,这种文化非常透明。
- [01:04:57] After that happens, there's real alignment. So in something like compute allocation we were talking about before, people might have different perspectives on how to allocate that compute, but they will engage in a thoughtful discussion about where the returns are the highest or the best. And when they do that, you know, and we come to a decision, then there's alignment on it. There's not second-guessing, there's not this kind of politics or **fiefdom**. The other piece of it is it's remarkably transparent, the culture, right? So  
  这之后,就会形成真正的共识。比如我们之前提到的算力分配问题,大家可能对如何分配算力有不同看法,但他们会进行深思熟虑的讨论,探讨哪里能获得最高或最好的回报。当他们这样做并最终做出决定时,就会达成共识。不会有事后质疑,也不会有那种政治斗争或山头主义。另一个方面是,这种文化非常透明。

**Extra example:**
- The CEO worked to break down departmental **fiefdoms** and encourage cross-team collaboration.  
  CEO致力于打破部门间的领地意识，鼓励跨团队协作。

### imbue  /ɪmˈbjuː/
**CEFR:** C2 | **Part of speech:** v. | **Occurrences:** 1

**EN:** to fill or inspire with a particular feeling, quality, or value  
**CN:** 灌输，使充满（某种感情、品质或价值观）

**Original examples:**
- [01:04:30] And I think it's just that focus on the mission and the alignment that kind of is **imbued** throughout the culture of the company.  
  我认为正是这种对使命的专注和一致性贯穿并渗透到了整个公司文化中。

**Extra example:**
- The teacher tried to **imbue** her students with a love of learning.  
  这位老师努力向学生灌输对学习的热爱。

### fruition  /fruˈɪʃn/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the realization or fulfillment of a plan or project  
**CN:** 实现，完成，取得成果

**Original examples:**
- [01:12:36] And it turns out that a lot of what Tom said during that walk has come to **fruition**.  
  事实证明，Tom在那次散步中说的很多事情都已经实现了。

**Extra example:**
- After years of planning, their vision finally came to **fruition**.  
  经过多年的规划，他们的愿景终于实现了。

### formative  /ˈfɔːrmətɪv/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** having an important and lasting influence on someone's development or character  
**CN:** 形成性的，有重大影响的

**Original examples:**
- [01:12:36] But I remember that as like an early **formative** thing, coming home and being like, holy shit, like this is going to be totally different and new.  
  但我记得那是一个早期的形成性经历，回家后我想，天哪，这将会完全不同和全新。

**Extra example:**
- Her **formative** years were spent traveling around the world with her parents.  
  她的成长关键期是和父母一起环游世界度过的。

### analogue  /ˈænəlɔːɡ/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** a person or thing seen as comparable to another; a parallel or equivalent  
**CN:** 类似物，相似情况

**Original examples:**
- [01:13:53] Know how the business evolves over time and where there might be moments or **analogues** to things that have happened in the past.  
  了解业务如何随时间演变，以及哪些时刻或情况可能与过去发生的事情相似。

**Extra example:**
- The current economic crisis has no clear historical **analogue**.  
  当前的经济危机没有明确的历史相似情况可供参考。

### harrowing  /ˈhærəʊɪŋ/
**CEFR:** C1 | **Part of speech:** adj. | **Occurrences:** 1

**EN:** extremely distressing or disturbing  
**CN:** 令人痛苦的，使人不安的

**Original examples:**
- [01:14:14] That was a **harrowing** time, but it was also a time kind of without precedent, right?  
  那是一段令人痛苦的时期,但也是一个前所未有的时期,对吧?

**Extra example:**
- The survivors shared their **harrowing** experiences during the disaster.  
  幸存者们分享了他们在灾难中令人痛苦的经历。

### throughput  /ˈθruːpʊt/
**CEFR:** C1 | **Part of speech:** n. | **Occurrences:** 1

**EN:** the amount of material or items passing through a system or process  
**CN:** 吞吐量，处理能力

**Original examples:**
- [01:18:54] If you think about what can happen when the lab's **throughput** goes up 10x or 100x and we can run that many more experiments, probably get better results faster, and that can be something that helps, you know, people around the world, right?  
  如果你想想当实验室的吞吐量提高10倍或100倍时会发生什么,我们可以进行更多的实验,可能更快地获得更好的结果,这可以帮助世界各地的人们,对吧?

**Extra example:**
- The new system increased data **throughput** by 50%.  
  新系统将数据吞吐量提高了50%。

### pro forma  /ˌprəʊ ˈfɔːmə/
**CEFR:** C2 | **Part of speech:** adj./adv. | **Occurrences:** 1

**EN:** done as a formality or matter of form; perfunctory  
**CN:** 形式上的，例行公事的

**Original examples:**
- [01:03:43] The evaluation was merely **pro forma** rather than a genuine assessment.  
  这次评估只是例行公事,而不是真正的评估。

**Extra example:**
- The board meeting was a **pro forma** exercise since the decision had already been made.  
  董事会会议只是走个形式,因为决定已经做出了。

---

## Useful Phrases

### dive right into
**Type:** phrasal_verb

**EN:** to start doing something immediately without hesitation  
**CN:** 立即开始做某事，直接切入

**Original examples:**
- [00:41] One of the things that fascinates me most, just to **dive right into** something that I think we're both quite passionate about is this question of compute  
  最让我着迷的事情之一，让我们**直接切入**一个我们都很感兴趣的话题，就是算力问题

**Extra example:**
- Let's skip the small talk and **dive right into** the main issue.  
  我们跳过寒暄，**直接切入**主要问题吧。

### bring someone into
**Type:** phrasal_verb

**EN:** to include someone in a situation or give them insight into something  
**CN:** 让某人了解情况，带某人进入某个场景

**Original examples:**
- [01:14] Just like **bring us into** that part of your life because I think it's like right at the cutting edge of what's going on.  
  **让我们了解**你生活的那部分，因为我觉得这正是最前沿的东西。
- [06:58] Can you **bring us into** the room for the conversations around the trade-offs between those?  
  你能**让我们了解**关于这些权衡的讨论吗？

**Extra example:**
- The manager decided to **bring the team into** the decision-making process.  
  经理决定**让团队参与**决策过程。

### go out of business
**Type:** collocation

**EN:** to stop operating as a company, usually due to financial failure  
**CN:** 倒闭，破产

**Original examples:**
- [01:46] If you buy too much compute, you **go out of business**.  
  如果你买太多算力，你就会**倒闭**。

**Extra example:**
- Many small retailers **went out of business** during the pandemic.  
  许多小型零售商在疫情期间**倒闭**了。

### eke out
**Type:** phrasal_verb

**EN:** to obtain or achieve something with difficulty or effort  
**CN:** 努力获得，艰难取得

**Original examples:**
- [04:16] you want to sort of **eke your way into** being as close to the bare metal as possible  
  你想要**努力达到**尽可能接近底层硬件

**Extra example:**
- The team managed to **eke out** a narrow victory in the final minutes.  
  球队在最后几分钟**艰难取得**了险胜。

### work backwards from
**Type:** collocation

**EN:** to start with the end goal and plan the steps needed to reach it in reverse order  
**CN:** 从结果倒推，逆向规划

**Original examples:**
- [06:00] We look at a range of scenarios and we look at different points in that cone of uncertainty over, you know, a one to two-year period, and then we kind of **work backwards from** that.  
  我们看一系列场景，在一到两年的时间范围内看不确定性锥体中的不同点，然后我们**从那里倒推**。

**Extra example:**
- Amazon's approach is to **work backwards from** the customer experience.  
  亚马逊的方法是**从客户体验倒推**。

### break down
**Type:** phrasal_verb

**EN:** to fail or stop working properly; to analyze something into parts  
**CN:** 失效，崩溃；分解分析

**Original examples:**
- [09:01] The way the place that analogy kind of **breaks down** a little bit is people think, okay, I'm going from the sedan to the sports car.  
  这个类比**失效**的地方在于，人们认为，好吧，我从轿车换到跑车。

**Extra example:**
- The negotiation **broke down** when neither side would compromise.  
  当双方都不愿妥协时，谈判**破裂**了。

### juice engagement
**Type:** idiom

**EN:** to artificially increase user engagement or activity metrics  
**CN:** 人为提高用户参与度

**Literal:** 榨取参与度的汁液  
**Figurative EN:** to artificially boost or manipulate engagement metrics  
**Figurative CN:** 人为提升或操纵参与度指标

**Original examples:**
- [10:46] Unlike most software companies that try to maximize your time on their app to **juice engagement**, Ramp does the exact opposite.  
  与大多数试图最大化你在应用上的时间来**提高参与度**的软件公司不同，Ramp恰恰相反。

**Extra example:**
- Social media platforms use notifications to **juice engagement** and keep users scrolling.  
  社交媒体平台使用通知来**提高参与度**，让用户不停滚动。

### switch on
**Type:** phrasal_verb

**EN:** to activate or start using something  
**CN:** 打开，启用

**Original examples:**
- [12:42] Like everyone, the second Opus 4.7 comes out, like even me as a consumer, the thing you do is you **switch it on** or GPT 5.5 comes out, you **switch on** the new one right away.  
  就像每个人一样，Opus 4.7一发布，即使作为消费者的我，你做的事就是**启用它**，或者GPT 5.5发布了，你立即**启用**新版本。

**Extra example:**
- As soon as the new feature is available, I'll **switch it on** in the settings.  
  新功能一上线，我就会在设置里**启用它**。

### play out
**Type:** phrasal_verb

**EN:** to develop or happen in a particular way; to unfold over time  
**CN:** 发展，展开；以某种方式发生

**Original examples:**
- [15:01] customers see that and then they invest really heavily in more tokens with the newer models, and we've just seen that cycle **play out** again and again  
  客户看到这一点后就会大量投资购买更多新模型的tokens，我们一次又一次地看到这个循环不断上演。

**Extra example:**
- Let's wait and see how this situation **plays out** before making any decisions.  
  让我们先等等看这个情况如何发展，再做决定。

### get one's hands on
**Type:** idiom

**EN:** to obtain or access something; to start using something  
**CN:** 获得，得到；开始使用

**Literal:** 把手放在某物上  
**Figurative EN:** to obtain, acquire, or gain access to something  
**Figurative CN:** 获得、得到或接触到某物

**Original examples:**
- [19:55] And that gives us a sense for model capability. You can do the same thing as you think about RL. And then probably as importantly is when customers **get their hands on it**, like what are they seeing?  
  这让我们对模型能力有了感觉。在考虑强化学习时你也可以做同样的事情。然后同样重要的是，当客户真正使用它时，他们看到了什么？

**Extra example:**
- I can't wait to **get my hands on** the new iPhone and test its features.  
  我迫不及待想拿到新iPhone并测试它的功能。

### get stuck
**Type:** collocation

**EN:** to be unable to progress or continue; to encounter a problem that prevents advancement  
**CN:** 卡住，遇到障碍；无法继续进行

**Original examples:**
- [20:06] Hey, I wish the model were better at this or I had this particular place where it **got stuck** and I could build this other product, but the capability needs to be further than that.  
  嘿，我希望模型在这方面更好，或者我在某个特定地方遇到了瓶颈，我本可以构建另一个产品，但能力需要更进一步。

**Extra example:**
- I **got stuck** on the third question of the exam and couldn't move forward.  
  我在考试的第三题上卡住了，无法继续往下做。

### hold oneself to a high standard
**Type:** collocation

**EN:** to maintain strict expectations for one's own performance or behavior  
**CN:** 对自己要求严格，保持高标准

**Original examples:**
- [20:50] Like, you know, we **hold ourselves to a really high standard**. Again, it's this kind of idea of a research lab that's very kind of scientific method  
  你知道，我们对自己要求非常严格。这又回到了研究实验室的理念，非常注重科学方法。

**Extra example:**
- She **holds herself to a high standard** and never submits work that isn't her best.  
  她对自己要求很高，从不提交不是自己最佳水平的作品。

### turn of the crank
**Type:** idiom

**EN:** one iteration or cycle in a repetitive process; one step in a series  
**CN:** （重复过程中的）一次迭代，一个循环

**Literal:** 转动曲柄一次  
**Figurative EN:** one iteration, cycle, or step in an ongoing repetitive process  
**Figurative CN:** 持续重复过程中的一次迭代、循环或步骤

**Original examples:**
- [21:08] Like, if that continues to be true for however many more, you know, **turns of the crank** here, how do you do that thing of not thinking linear and thinking exponential yourself  
  如果这种情况在接下来的多次迭代中继续保持，你如何做到不用线性思维而是用指数思维来思考？

**Extra example:**
- With each **turn of the crank**, the machine learning model gets slightly better.  
  每经过一次迭代，机器学习模型就会变得稍微好一点。

### get one's head around
**Type:** idiom

**EN:** to understand or comprehend something difficult or complex  
**CN:** 理解，弄明白（复杂或困难的事情）

**Literal:** 把头绕过某物  
**Figurative EN:** to understand or mentally grasp something that is difficult or complex  
**Figurative CN:** 理解或在思维上掌握困难或复杂的事物

**Original examples:**
- [21:33] Like, I don't even know how to **get my head around** it. So how do you **get your head around** it?  
  我甚至不知道该如何理解它。那你是如何理解的呢？

**Extra example:**
- It took me weeks to **get my head around** quantum physics concepts.  
  我花了几周时间才弄明白量子物理的概念。

### canvas for
**Type:** phrasal_verb

**EN:** to search widely or systematically for something; to explore options thoroughly  
**CN:** 广泛搜寻，系统性地寻找

**Original examples:**
- [22:37] Makes me curious about how you are **canvassing the world for**—like, that is an opportunity you decided to do.  
  这让我很好奇你们是如何在全球范围内搜寻的——那是你们决定抓住的一个机会。

**Extra example:**
- The startup is **canvassing for** investors to fund their next round.  
  这家初创公司正在广泛寻找投资者为下一轮融资。

### layer cake
**Type:** idiom

**EN:** a structure with multiple distinct levels or layers stacked on top of each other  
**CN:** 分层结构；多层次的架构

**Literal:** 层层叠叠的蛋糕  
**Figurative EN:** a structure or system composed of multiple distinct layers or levels  
**Figurative CN:** 由多个不同层次或级别组成的结构或系统

**Original examples:**
- [24:08] And so if you think about it, it's a bit of this **layer cake** of compute that's starting at different times with different capabilities  
  所以如果你想一想，这有点像一个计算资源的分层结构，在不同时间以不同能力启动。

**Extra example:**
- The software architecture is a **layer cake** with the database at the bottom and UI at the top.  
  这个软件架构是一个分层结构，底层是数据库，顶层是用户界面。

### airdrop on
**Type:** phrasal_verb

**EN:** to suddenly provide or deliver something (especially resources) to someone  
**CN:** 突然提供或交付某物（尤指资源）给某人

**Original examples:**
- [26:08] If I **airdropped on** you twice the compute that you have tomorrow, like, would you consume that?  
  如果我明天突然给你提供两倍的算力，你能消耗掉吗？
- [26:08] If I **airdropped** 10 times the compute on top of you, how fast would you consume it?  
  如果我突然给你提供10倍的算力，你多快能消耗掉？

**Extra example:**
- The government **airdropped** emergency supplies on the disaster zone.  
  政府向灾区空投了应急物资。

### spin up
**Type:** phrasal_verb

**EN:** to start or launch something quickly, especially a system or service  
**CN:** 快速启动或开启某物，尤指系统或服务

**Original examples:**
- [27:28] It's become a lot easier for us to **spin up** very quickly and deploy like almost any type of compute.  
  我们现在能够非常快速地启动并部署几乎任何类型的算力。

**Extra example:**
- We can **spin up** a new server instance in less than five minutes.  
  我们可以在不到五分钟内启动一个新的服务器实例。

### accrue to
**Type:** phrasal_verb

**EN:** to accumulate or be gained by someone over time  
**CN:** 随时间积累或被某人获得

**Original examples:**
- [28:19] We think that there's so many examples of where a platform can **accrue** a lot of value, but the customers who are building on that platform actually **accrue** even more value.  
  我们认为有很多例子表明平台可以积累大量价值，但在该平台上构建的客户实际上会积累更多价值。
- [30:25] We also think that there's so much value that's going to **accrue** in some of these areas that, you know, our customers can win and we can win as well.  
  我们还认为在这些领域会积累大量价值，我们的客户可以赢，我们也可以赢。
- [32:33] That should **accrue** a lot of value to customers as well.  
  这也应该为客户积累大量价值。

**Extra example:**
- The benefits of this investment will **accrue to** shareholders over the next decade.  
  这项投资的收益将在未来十年内归股东所有。

### build ahead to
**Type:** phrasal_verb

**EN:** to develop or create something in anticipation of future capabilities or needs  
**CN:** 为未来的能力或需求提前开发或创建某物

**Original examples:**
- [29:41] One is kind of **building ahead to** model capabilities.  
  一个是为模型能力提前构建。

**Extra example:**
- We're **building ahead to** the expected demand surge next quarter.  
  我们正在为下季度预期的需求激增提前做准备。

### smile curve
**Type:** collocation

**EN:** a U-shaped curve showing value distribution, with high value at both ends and low value in the middle  
**CN:** 微笑曲线，显示价值分布的U形曲线，两端价值高，中间价值低

**Original examples:**
- [33:22] The cost of an H100 is, well, you know, looks like a **smile curve**.  
  H100的成本，嗯，你知道，看起来像一条微笑曲线。

**Extra example:**
- The manufacturing industry follows a **smile curve** where R&D and marketing capture most value.  
  制造业遵循微笑曲线，研发和营销获取了大部分价值。

### Jevons paradox
**Type:** collocation

**EN:** the phenomenon where increased efficiency leads to increased consumption rather than decreased usage  
**CN:** 杰文斯悖论，指效率提高导致消费增加而非使用减少的现象

**Original examples:**
- [35:19] The changing of the pricing for Opus actually, you know, you see this **Jevons paradox**, right?  
  Opus的定价变化实际上，你知道，你会看到这种杰文斯悖论，对吧？
- [35:50] And we also think that pricing to get that value and to see that kind of **Jevons paradox** happen is really important.  
  我们还认为，定价以获得该价值并看到那种杰文斯悖论发生是非常重要的。

**Extra example:**
- Fuel-efficient cars demonstrate **Jevons paradox** - people drive more because it's cheaper per mile.  
  省油汽车展示了杰文斯悖论——人们开得更多，因为每英里更便宜。

### slot in
**Type:** phrasal_verb

**EN:** to fit or insert something easily into an existing position or system  
**CN:** 轻松地将某物放入或插入现有位置或系统

**Original examples:**
- [35:44] They can **slot it in**. We didn't change the price.  
  他们可以直接插入使用。我们没有改变价格。

**Extra example:**
- The new module should **slot in** seamlessly with the existing architecture.  
  新模块应该能够无缝地融入现有架构。

### writ large
**Type:** idiom

**EN:** on a large scale; in a more obvious or exaggerated form  
**CN:** 大规模地；以更明显或夸张的形式

**Literal:** 大字书写  
**Figurative EN:** considered as a whole or in its entirety; on a broader scale  
**Figurative CN:** 从整体或全局来看；在更广泛的范围内

**Original examples:**
- [36:30] We think about what is the return on our compute spend, right, **writ large**.  
  我们考虑的是我们在算力支出上的整体回报。

**Extra example:**
- This problem is society's inequality **writ large**.  
  这个问题是社会不平等的放大版。

### tie out
**Type:** phrasal_verb

**EN:** to reconcile numbers or make financial figures match across different records  
**CN:** 核对（数字），使账目一致

**Original examples:**
- [45:06] Whereas before I'm working to **tie out** a number, or I'm, you know, in that accounting example, taking a long time to close the books.  
  而以前我要花时间去核对一个数字，或者在那个会计例子中，要花很长时间来结账。

**Extra example:**
- The accountant spent hours trying to **tie out** the discrepancies in the quarterly report.  
  会计师花了几个小时试图核对季度报告中的差异。

### come up the curve
**Type:** idiom

**EN:** to progress along a learning curve, to gain proficiency over time  
**CN:** 逐步掌握，沿着学习曲线进步

**Literal:** 沿着曲线向上走  
**Figurative EN:** to gradually improve skills or knowledge in a particular area  
**Figurative CN:** 在某个领域逐步提升技能或知识

**Original examples:**
- [45:28] That even as we grow the team, those people are more productive as well as they **come up the curve** on how to use Cloud within our company.  
  即使我们扩大团队，这些人在学习如何在公司内使用Claude的过程中也会变得更有生产力。

**Extra example:**
- New engineers typically **come up the curve** within three to six months of joining the team.  
  新工程师通常在加入团队后三到六个月内就能逐步掌握技能。

### at odds
**Type:** collocation

**EN:** in conflict or disagreement with something  
**CN:** 相矛盾，不一致

**Original examples:**
- [46:36] People said, well, hey, aren't AI safety and building a really big business—aren't those things **at odds**?  
  人们说，嘿，AI安全和建立一个大企业——这两件事不是相互矛盾的吗？

**Extra example:**
- His personal values were **at odds** with the company's aggressive sales tactics.  
  他的个人价值观与公司激进的销售策略相矛盾。

### prove out
**Type:** phrasal_verb

**EN:** to demonstrate or confirm something through evidence or results over time  
**CN:** 证实，验证

**Original examples:**
- [47:56] The business continued to **prove out** the thesis that the return to frontier intelligence is really high.  
  业务持续证实了前沿智能回报率非常高的论点。
- [49:36] But it did have this kind of downstream effect that we've really seen **prove out** again and again to be a company that is both at the frontier, but one that is investing in safety.  
  但它确实产生了这种下游效应，我们一次又一次地看到这被证实——成为一家既处于前沿又投资于安全的公司。

**Extra example:**
- The initial hypothesis took two years to fully **prove out** in production environments.  
  最初的假设在生产环境中花了两年时间才完全得到验证。

### inure to the benefit of
**Type:** collocation

**EN:** to result in an advantage or positive outcome for someone or something  
**CN:** 有利于，使受益

**Original examples:**
- [49:21] When you have this investment that we've made and will continue to make in safety, interpretability, alignment, like that actually **inures to the benefit of** the enterprise.  
  当你在安全性、可解释性、对齐方面进行了我们已经做出并将继续做出的投资时，这实际上有利于企业。

**Extra example:**
- The patent protections **inure to the benefit of** both the company and its shareholders.  
  专利保护对公司及其股东都有利。

### close the gap
**Type:** collocation

**EN:** to reduce the difference or distance between two things  
**CN:** 缩小差距

**Original examples:**
- [51:13] I think we're going to **close the gap** over time as we get better at forecasting and understanding the business.  
  我认为随着我们在预测和理解业务方面做得更好，我们会逐渐缩小差距。

**Extra example:**
- The company is working hard to **close the gap** between its current performance and industry leaders.  
  公司正在努力缩小其当前表现与行业领导者之间的差距。

### lean into
**Type:** phrasal_verb

**EN:** to embrace or commit more fully to something, often a challenge or opportunity  
**CN:** 全力投入，积极拥抱

**Original examples:**
- [52:03] But it does mean that my thinking has at least shifted a lot more from linear and incremental towards, you know, **leaning into** this exponential.  
  但这确实意味着我的思维至少已经从线性和渐进式转变为更多地拥抱这种指数级增长。

**Extra example:**
- Instead of resisting the change, the team decided to **lean into** the new technology.  
  团队决定积极拥抱新技术，而不是抵制变革。

### in spades
**Type:** idiom

**EN:** in large amounts, to a great degree  
**CN:** 大量地，非常充分地

**Literal:** 用黑桃（扑克牌中最大的花色）  
**Figurative EN:** in abundance, to an extreme or impressive degree  
**Figurative CN:** 大量地，极其充分地

**Original examples:**
- [54:39] You know, I could say for our business, like we're seeing that **in spades**.  
  你知道，对于我们的业务来说，我们正在大量地看到这一点。

**Extra example:**
- The new product delivered on its promises **in spades**, exceeding all sales projections.  
  新产品充分兑现了承诺，超出了所有销售预期。

### come up with
**Type:** phrasal_verb

**EN:** to think of or produce an idea, answer, or solution  
**CN:** 想出，提出（主意、答案或解决方案）

**Original examples:**
- [58:21] How do we work across, you know, commercial and government to actually **come up with** some of the solutions to that?  
  我们如何跨越商业和政府领域，真正想出一些解决方案呢？
- [58:34] And that's not any one company that can, you know, **come up with** it.  
  这不是任何一家公司能够单独想出来的。

**Extra example:**
- The team needs to **come up with** a new marketing strategy by next week.  
  团队需要在下周之前想出一个新的营销策略。

### on paper
**Type:** idiom

**EN:** in theory or when judged by appearance, but not necessarily in reality  
**CN:** 理论上，表面上看

**Literal:** 在纸上  
**Figurative EN:** theoretically or based on written plans, but not proven in practice  
**Figurative CN:** 从理论或书面计划来看，但未经实践验证

**Original examples:**
- [01:03:26] First of all, we have seven co-founders, right? That shouldn't work **on paper**, but it really does in practice.  
  首先，我们有七位联合创始人，对吧？这在理论上不应该行得通，但实际上确实可行。

**Extra example:**
- The plan looks great **on paper**, but we'll see how it works in the real world.  
  这个计划在理论上看起来很棒，但我们要看看在现实中效果如何。

### pro forma
**Type:** collocation

**EN:** done as a formality; perfunctory  
**CN:** 形式上的，敷衍了事的

**Original examples:**
- [01:03:43] We do a culture interview and it's not some **pro forma**, you know, thing we do just to kind of check a box.  
  我们会进行文化面试，这不是某种形式上的、只是为了打勾的事情。

**Extra example:**
- The board meeting was just a **pro forma** approval since the decision had already been made.  
  董事会会议只是形式上的批准，因为决定早已做出。

### flying colors
**Type:** idiom

**EN:** with great success; excellently  
**CN:** 大获成功，出色地

**Literal:** 飞扬的旗帜  
**Figurative EN:** with outstanding success or distinction  
**Figurative CN:** 以出色的成绩或表现

**Original examples:**
- [01:03:49] So somebody could be **flying colors** on everything else and really, really the smartest person you've met in this role, we won't hire them if they don't pass the culture bar.  
  所以即使某人在其他所有方面都表现出色，而且确实是你在这个职位上遇到的最聪明的人，如果他们没有通过文化标准，我们也不会雇用他们。

**Extra example:**
- She passed the exam with **flying colors**, scoring 98 out of 100.  
  她以优异的成绩通过了考试，得了98分（满分100）。

### sharp elbows
**Type:** idiom

**EN:** aggressive, competitive behavior; being pushy or ruthless to get ahead  
**CN:** 咄咄逼人的竞争行为，为了出头而强势或无情

**Literal:** 尖锐的肘部  
**Figurative EN:** aggressive tactics used to advance oneself at others' expense  
**Figurative CN:** 以牺牲他人为代价来提升自己的强硬手段

**Original examples:**
- [01:04:02] It's one incredibly collaborative, and this means that we don't really tolerate the fiefdoms or the **sharp elbows** or the like 'I need to take credit for this.'  
  首先是极其协作，这意味着我们不容忍小团体、咄咄逼人的竞争或者'我需要为此邀功'这样的行为。

**Extra example:**
- In that company, you need **sharp elbows** to climb the corporate ladder.  
  在那家公司，你需要强势竞争才能在职场上晋升。

### softballs
**Type:** idiom

**EN:** easy questions designed to make someone look good; questions lacking challenge  
**CN:** 容易回答的问题，旨在让某人表现良好的简单问题

**Literal:** 垒球（一种较软的球）  
**Figurative EN:** easy, unchallenging questions, often asked to help someone appear competent  
**Figurative CN:** 简单、不具挑战性的问题，通常是为了帮助某人显得有能力

**Original examples:**
- [01:05:23] And these are not **softballs**, they're not like planted questions, they're just real questions that are on people's minds, and he answers them the best that he can.  
  这些不是容易回答的问题，也不是事先安排好的问题，而是人们真正关心的问题，他会尽力回答。

**Extra example:**
- The journalist threw him **softballs** during the interview instead of asking tough questions.  
  记者在采访中向他提出了一些简单的问题，而不是问一些尖锐的问题。

### take lightly
**Type:** phrasal_verb

**EN:** to treat something as unimportant or not serious  
**CN:** 轻视，不重视

**Original examples:**
- [01:07:37] I think that that really matters to the people, not just on the research team but across the company, and that we think is a real advantage for us and it's not something that we **take lightly**.  
  我认为这对公司里的人来说真的很重要，不仅仅是研究团队，而是整个公司，我们认为这是我们的真正优势，我们不会轻视这一点。

**Extra example:**
- We don't **take** security threats **lightly** - every incident is thoroughly investigated.  
  我们不会轻视安全威胁——每一起事件都会被彻底调查。

### first principles
**Type:** collocation

**EN:** fundamental truths or basic assumptions that serve as the foundation for reasoning  
**CN:** 第一性原理，基本原理（作为推理基础的基本真理或假设）

**Original examples:**
- [01:11:38] Yeah, it's really hard, but I think the important thing is to think in **first principles**, right?  
  是的，这真的很难，但我认为重要的是从第一性原理思考，对吧？
- [01:11:50] Thinking in **first principles** and having like intellectual openness.  
  从第一性原理思考，并保持思想开放。

**Extra example:**
- Elon Musk is known for breaking down problems using **first principles** thinking.  
  埃隆·马斯克以使用第一性原理思维来分解问题而闻名。

### bend all paradigms
**Type:** collocation

**EN:** to fundamentally change or challenge established frameworks and ways of thinking  
**CN:** 彻底改变或挑战既定的框架和思维方式

**Original examples:**
- [01:12:18] this is going to **bend all paradigms** of what not just things I've seen but what most people have seen.  
  这将彻底改变不仅是我所见过的，而且是大多数人所见过的所有范式。

**Extra example:**
- This breakthrough technology will **bend all paradigms** in the healthcare industry.  
  这项突破性技术将彻底改变医疗行业的所有范式。

### come to fruition
**Type:** collocation

**EN:** to be realized or achieved; to become reality  
**CN:** 实现，成为现实

**Original examples:**
- [01:12:36] And it turns out that a lot of what Tom said during that walk has **come to fruition**  
  事实证明，汤姆在那次散步中说的很多事情都已经实现了。

**Extra example:**
- After years of planning, their vision finally **came to fruition**.  
  经过多年的规划，他们的愿景终于实现了。

### at a granular level
**Type:** collocation

**EN:** in great detail; examining the smallest components or aspects  
**CN:** 在细节层面，详细地

**Original examples:**
- [01:13:12] Like that training is really valuable and thinking about things **at a granular level** and not losing that.  
  那种训练非常有价值，能够在细节层面思考问题而不失去这种能力。

**Extra example:**
- We need to analyze the data **at a granular level** to identify the root cause.  
  我们需要在细节层面分析数据以找出根本原因。

### at 50,000 feet
**Type:** idiom

**EN:** at a very high, abstract level; viewing the big picture without details  
**CN:** 在非常高的抽象层面；只看大局而不关注细节

**Literal:** 在五万英尺高空  
**Figurative EN:** viewing something from a very high-level, strategic perspective without getting into details  
**Figurative CN:** 从非常高层次的战略角度看问题，不涉及具体细节

**Original examples:**
- [01:13:12] Like I'm not somebody who is comfortable **at 50,000 feet**.  
  我不是那种只在高层次思考就感到舒服的人。

**Extra example:**
- The CEO prefers to stay **at 50,000 feet** and let managers handle the details.  
  CEO更喜欢保持在战略高度，让经理们处理细节。

### take a big bite out of
**Type:** idiom

**EN:** to consume or use up a large portion of something  
**CN:** 占用或消耗很大一部分

**Literal:** 从某物上咬下一大口  
**Figurative EN:** to consume, use up, or reduce a significant portion of something (time, money, resources)  
**Figurative CN:** 占用、消耗或减少某物（时间、金钱、资源）的很大一部分

**Original examples:**
- [01:14:28] certainly this job **takes a big bite out of** all that.  
  这份工作确实占用了所有这些的很大一部分。

**Extra example:**
- The new mortgage will **take a big bite out of** our monthly budget.  
  新的房贷将占用我们月度预算的很大一部分。

### hold light and shade
**Type:** idiom

**EN:** to balance optimism with realism; to acknowledge both positive and negative aspects  
**CN:** 平衡乐观与现实；同时承认积极和消极的方面

**Literal:** 同时持有光明和阴影  
**Figurative EN:** to maintain balance between optimism and caution, or between positive and negative perspectives  
**Figurative CN:** 在乐观与谨慎之间保持平衡，或在积极与消极的视角之间保持平衡

**Original examples:**
- [01:15:45] we talk about internally kind of **holding light and shade**.  
  我们内部讨论时会说要同时看到光明和阴影。

**Extra example:**
- Good leaders **hold light and shade** - they're optimistic but also prepare for challenges.  
  优秀的领导者会平衡光明与阴影——他们乐观但也为挑战做好准备。

### hit a wall
**Type:** idiom

**EN:** to reach a point where progress stops or becomes very difficult  
**CN:** 遇到瓶颈，进展停滞

**Literal:** 撞到墙上  
**Figurative EN:** to reach a barrier or limit where further progress becomes impossible or extremely difficult  
**Figurative CN:** 达到一个障碍或极限，使得进一步的进展变得不可能或极其困难

**Original examples:**
- [01:16:37] to the extent that that diffusion, you know, **hits a wall** or slows down or something like that  
  如果这种扩散遇到瓶颈或放缓之类的情况

**Extra example:**
- After months of progress, the project **hit a wall** due to funding issues.  
  经过几个月的进展后，该项目因资金问题遇到了瓶颈。

### level off
**Type:** phrasal_verb

**EN:** to stop increasing or decreasing and remain steady  
**CN:** 趋于平稳，停止增长或下降

**Original examples:**
- [01:17:01] but the model capabilities **leveling off** would be another thing.  
  但模型能力趋于平稳将是另一回事。

**Extra example:**
- Sales growth is expected to **level off** next quarter after rapid expansion.  
  在快速扩张之后，销售增长预计将在下个季度趋于平稳。

---

## Complex Sentences

### [01:26]
**Original:** Look, the compute that we procure is the lifeblood of our business. It is the most important thing in the company. It is the thing on which, it's like the canvas on which everything else gets built.

**Translation:** 你看,我们采购的算力是我们业务的命脉。它是公司里最重要的东西。它就像是一块画布,所有其他东西都在上面构建。

**Core structure:**
- The compute is the thing on which everything gets built.  
  算力是所有东西在其上构建的基础。

**Structure tree:**
```
main clause: It is the thing on which...
relative clause: on which everything else gets built
metaphor: it's like the canvas
parallel relative clause: on which everything else gets built
```

**Grammar points:**
- **介词+which引导定语从句** - on which 引导定语从句,介词前置的正式用法
- **被动语态 get + 过去分词** - gets built 表示被构建,口语化的被动结构

### [02:36]
**Original:** And then as we go out and actually do these deals to procure compute, you know, flexibility is really important to us and so we build that flexibility into the deals themselves, into how we use the compute as well, because the way in which we bridge from a position we are today to where we want to go when the business is growing exponentially is to use that compute as efficiently as possible.

**Translation:** 然后当我们真正去做这些采购算力的交易时,灵活性对我们来说非常重要,所以我们把这种灵活性构建到交易本身中,也构建到我们如何使用算力中,因为当业务呈指数级增长时,我们从今天所处的位置过渡到我们想要达到的目标的方式,就是尽可能高效地使用算力。

**Core structure:**
- Flexibility is important, so we build flexibility into deals, because the way to bridge is to use compute efficiently.  
  灵活性很重要,所以我们把灵活性构建到交易中,因为过渡的方式就是高效使用算力。

**Structure tree:**
```
time clause: as we go out and do deals
main clause 1: flexibility is important
main clause 2: we build flexibility into...
reason clause: because the way...is to use
nested relative clause: in which we bridge from...to where...
time clause: when the business is growing
```

**Grammar points:**
- **多层嵌套从句** - 包含时间、原因、定语从句的复杂嵌套结构
- **the way in which结构** - 表示方式的定语从句,in which可替换为that或省略
- **from...to where...结构** - where引导名词性从句作介词to的宾语

### [03:39]
**Original:** You know, when we started using TPUs, I think it was maybe the third generation TPUs was the first one we used at scale, people thought, 'Oh, well, you're crazy. Everyone's using GPUs. Why aren't you using GPUs?'

**Translation:** 你知道,当我们开始使用TPU时,我想可能是第三代TPU是我们大规模使用的第一代,人们认为,'哦,你们疯了。每个人都在用GPU。你们为什么不用GPU?'

**Core structure:**
- When we started using TPUs, people thought we were crazy.  
  当我们开始使用TPU时,人们认为我们疯了。

**Structure tree:**
```
time clause: when we started using TPUs
parenthetical insertion: I think it was maybe the third generation
main clause: people thought
direct speech: 'Oh, well, you're crazy...'
rhetorical question: Why aren't you using GPUs?
```

**Grammar points:**
- **插入语打断句子结构** - I think...插入语使句子难以一次性理解完整
- **直接引语中的反问句** - Why aren't you...表达质疑和不理解

### [04:37]
**Original:** So we work really closely with the Annapurna Labs team at Amazon to help influence the roadmap of these chips, because we believe what we're doing is really stressing the limits of what these chips are capable of.

**Translation:** 所以我们与亚马逊的Annapurna Labs团队密切合作,以帮助影响这些芯片的路线图,因为我们相信我们正在做的事情确实在挑战这些芯片能力的极限。

**Core structure:**
- We work closely with the team to influence the roadmap, because we believe we're stressing the limits.  
  我们与团队密切合作以影响路线图,因为我们相信我们在挑战极限。

**Structure tree:**
```
main clause: we work closely with the team
purpose: to help influence the roadmap
reason clause: because we believe...
object clause: what we're doing is stressing...
nested object clause: what these chips are capable of
```

**Grammar points:**
- **双重what引导的名词性从句** - what we're doing作主语,what these chips are capable of作宾语
- **be capable of结构** - 表示能力范围,of后接名词性从句

### [04:50]
**Original:** And that means that a dollar of compute inside our organization goes further than I think it does anywhere else.

**Translation:** 这意味着在我们组织内部,一美元的算力比我认为的在其他任何地方都能发挥更大的作用。

**Core structure:**
- A dollar of compute goes further than it does anywhere else.  
  一美元的算力比在其他地方发挥更大作用。

**Structure tree:**
```
main clause: that means that...
object clause: a dollar goes further than...
comparison: than I think it does anywhere else
parenthetical: I think
ellipsis: it does (go) anywhere else
```

**Grammar points:**
- **比较级中的省略** - than后用it does代替goes further,避免重复
- **插入语打断比较结构** - I think插在比较结构中间,增加理解难度

### [05:27]
**Original:** When you're building and growing a business exponentially, you know, really small movements in monthly or weekly growth rates result in compounding very, very different outcomes.

**Translation:** 当你以指数级方式建立和发展一个企业时,月度或周度增长率的微小变化会导致截然不同的复合结果。

**Core structure:**
- Small movements in growth rates result in different outcomes.  
  增长率的微小变化导致不同的结果。

**Structure tree:**
```
temporal clause: When you're building...
main clause: movements result in outcomes
modifiers: really small / in monthly or weekly growth rates / compounding / very different
```

**Grammar points:**
- **When引导时间状语从句** - 从句描述背景条件,主句说明结果
- **result in复合动词** - 表示'导致',连接原因和结果
- **现在分词compounding作状语** - 修饰outcomes,表示'复合的'方式

### [06:23]
**Original:** If we were to say to our employees, you can't use our models anymore, we could serve billions of dollars of revenue with that compute that we allocate to employees internally, but we want to take a long-term view and a long-term perspective on that cone of uncertainty.

**Translation:** 如果我们对员工说,你们不能再使用我们的模型了,我们可以用内部分配给员工的那些算力来服务数十亿美元的收入,但我们想对这个不确定性锥体采取长期的观点和视角。

**Core structure:**
- We could serve billions of dollars of revenue, but we want to take a long-term view.  
  我们可以服务数十亿美元的收入,但我们想采取长期观点。

**Structure tree:**
```
conditional clause: If we were to say...
main clause 1: we could serve billions of dollars
main clause 2: but we want to take a long-term view
relative clause: that we allocate to employees
```

**Grammar points:**
- **虚拟条件句(were to)** - 表示与现在事实相反的假设
- **but连接转折** - 对比短期利益和长期战略
- **定语从句修饰compute** - that从句说明算力的用途

### [07:37]
**Original:** But there's a level of compute for model development that we will not go below, right? So even if it means it's harder to serve customers or we have to do kind of unnatural things when it comes to that, we want to continue to make that long-term investment in developing the best models.

**Translation:** 但是有一个用于模型开发的算力水平是我们不会低于的,对吧?所以即使这意味着服务客户会更困难,或者我们不得不在这方面做一些不自然的事情,我们也想继续对开发最佳模型进行长期投资。

**Core structure:**
- There's a level that we will not go below. We want to continue to make investment.  
  有一个我们不会低于的水平。我们想继续投资。

**Structure tree:**
```
main clause 1: there's a level of compute
relative clause: that we will not go below
concessive clause: even if it means...
main clause 2: we want to continue to make investment
```

**Grammar points:**
- **even if引导让步状语从句** - 表示'即使',强调不管困难如何都坚持
- **it means (that)省略** - it指代前面的情况,that从句作宾语
- **when it comes to** - 固定搭配,表示'当涉及到'

### [09:38]
**Original:** And that just doesn't serve customers. That also helps us internally as well, because if you think about if we're using the model for, if we're doing reinforcement learning on the model, it's basically inference within a sandbox with a reward function.

**Translation:** 这不仅服务客户。这也对我们内部有帮助,因为如果你想想,如果我们使用模型,如果我们在模型上进行强化学习,它基本上就是在一个带有奖励函数的沙盒内进行推理。

**Core structure:**
- That helps us internally because it's basically inference within a sandbox.  
  这对我们内部有帮助,因为它基本上是沙盒内的推理。

**Structure tree:**
```
main clause: That helps us internally
causal clause: because it's basically inference...
nested conditional: if you think about if we're using...
parenthetical: if we're doing reinforcement learning
```

**Grammar points:**
- **多重if条件嵌套** - 口语化表达,多个if从句层层递进说明情况
- **because引导原因状语从句** - 解释为什么对内部有帮助

### [10:25]
**Original:** So if you think about it, all these things are very connected—these various tasks and workloads that we have internally all kind of fit together in this way of, you know, doing R&D for model capabilities, for compute efficiency, for serving customers, and then having internal workloads that can be sped up by using the best models, sometimes models that we...

**Translation:** 所以如果你想想,所有这些事情都是紧密相连的——我们内部的这些各种任务和工作负载都以这种方式结合在一起,你知道,为模型能力做研发,为计算效率,为服务客户,然后拥有可以通过使用最佳模型来加速的内部工作负载,有时是我们...

**Core structure:**
- All these things are connected. Tasks fit together in this way.  
  所有这些事情都相连。任务以这种方式结合在一起。

**Structure tree:**
```
conditional: if you think about it
main clause: all these things are connected
appositive: these various tasks and workloads
relative clause: that we have internally
gerund phrases: doing R&D... / having internal workloads...
```

**Grammar points:**
- **破折号引出同位语** - these various tasks解释all these things
- **并列动名词短语** - doing R&D和having workloads并列说明方式
- **定语从句嵌套** - that从句修饰workloads和models,层层限定

### [10:46]
**Original:** Unlike most software companies that try to maximize your time on their app to juice engagement, Ramp does the exact opposite.

**Translation:** 与大多数试图最大化你在其应用上的时间以提升参与度的软件公司不同,Ramp做的恰恰相反。

**Core structure:**
- Ramp does the exact opposite.  
  Ramp做的恰恰相反。

**Structure tree:**
```
main clause: Ramp does the exact opposite
prepositional phrase: Unlike most software companies...
relative clause: that try to maximize...
infinitive of purpose: to juice engagement
```

**Grammar points:**
- **Unlike 介词短语前置** - 对比结构,修饰整个主句
- **不定式表目的** - to juice engagement 说明maximize的目的

### [13:48]
**Original:** Which means that in some sense, you know, if you have two employees and they're maybe both equally capable, someone takes a week to do, you know, an assignment, someone does it in a day.

**Translation:** 这意味着在某种意义上,如果你有两个员工,他们可能都同样有能力,有人需要一周时间来完成一项任务,有人一天就能完成。

**Core structure:**
- Someone takes a week, someone does it in a day.  
  有人需要一周,有人一天完成。

**Structure tree:**
```
main clause: Which means that...
that-clause: if you have two employees...
conditional clause: if you have two employees
parallel structure: someone takes... someone does...
parenthetical: you know (×2)
```

**Grammar points:**
- **Which 引导非限制性定语从句** - 指代前文整个句子的内容
- **条件从句嵌套** - if从句内含并列结构和对比
- **口语插入语** - you know 是口语填充词,可忽略

### [14:08]
**Original:** And what we found very consistently is by releasing new models, the TAM is unlocked in a unique way.

**Translation:** 我们非常一致地发现,通过发布新模型,总体可达市场以一种独特的方式被解锁了。

**Core structure:**
- What we found is the TAM is unlocked.  
  我们发现的是总体可达市场被解锁了。

**Structure tree:**
```
main clause: What we found is...
subject clause: What we found
predicative clause: the TAM is unlocked
prepositional phrase: by releasing new models
adverbial: in a unique way
```

**Grammar points:**
- **What 引导主语从句** - What we found 作主语,表示'我们发现的事情'
- **被动语态** - is unlocked 强调TAM被解锁的状态

### [14:24]
**Original:** We started the year with about $9 billion of run rate revenue and we ended the quarter with, you know, north of $30 billion of run rate revenue.

**Translation:** 我们以约90亿美元的年化收入开始这一年,并以超过300亿美元的年化收入结束这个季度。

**Core structure:**
- We started with $9 billion and ended with $30 billion.  
  我们从90亿开始,以300亿结束。

**Structure tree:**
```
parallel structure: We started... and we ended...
prepositional phrases: with $9 billion / with $30 billion
parenthetical: you know
modifier: of run rate revenue (×2)
```

**Grammar points:**
- **并列句** - 两个对称的动作形成对比
- **north of** - 口语表达,意为'超过,多于'

### [14:43]
**Original:** I think that's unique to enterprise because in consumer sometimes you don't see that as readily, that consumers really are pushing the limits of what the models can do, whereas in enterprise like our customers are always—now you know it started with coding but it's really expanded beyond that very meaningfully—but each model generation gives you the chance to do more with it, to do it better, to do it more efficiently, and customers see that and then they invest really heavily in more tokens with the newer models, and we've just seen that cycle play out again and again, and that's a core thesis of our business that especially in enterprise, the returns to frontier intelligence are not slowing down.

**Translation:** 我认为这对企业来说是独特的,因为在消费者领域,你有时不会那么容易看到这一点,即消费者真的在推动模型能力的极限,而在企业领域,我们的客户总是——现在你知道它始于编码,但它已经非常有意义地扩展到了那之外——但每一代模型都给你机会用它做更多事情,做得更好,做得更高效,客户看到这一点,然后他们在新模型的更多token上进行大量投资,我们一次又一次地看到这个循环上演,这是我们业务的核心论点,即特别是在企业领域,前沿智能的回报并没有放缓。

**Core structure:**
- That's unique to enterprise because consumers don't push limits, whereas enterprise customers invest heavily, and we've seen that cycle.  
  这对企业是独特的,因为消费者不推动极限,而企业客户大量投资,我们看到了这个循环。

**Structure tree:**
```
main: That's unique to enterprise
because-clause: consumers don't see that readily
whereas-clause: enterprise customers invest heavily
parenthetical: now you know it started...
parallel infinitives: to do more, to do it better, to do it more efficiently
and-chain: customers see... and invest... and we've seen...
```

**Grammar points:**
- **超长复合句** - 多个从句和并列结构连接,包含because/whereas/that等多层嵌套
- **破折号插入语** - —now you know...— 打断主句,补充背景信息
- **并列不定式** - to do more/better/efficiently 表示递进的三个目的

### [15:26]
**Original:** It seems as though in the major labs we've reached this point—someone on your team said it recently—of like recursive self-improvement, where the models themselves are building and doing a lot of the research to do, you know, the next generation of improvement.

**Translation:** 似乎在主要的实验室中，我们已经达到了这样一个点——你团队中的某个人最近说过——就像递归自我改进，模型本身正在构建并进行大量研究来做，你知道的，下一代的改进。

**Core structure:**
- We've reached this point of recursive self-improvement, where the models are building the next generation.  
  我们已经达到了递归自我改进的点，模型正在构建下一代。

**Structure tree:**
```
main clause: we've reached this point
parenthetical: someone on your team said it recently
appositive: of recursive self-improvement
relative clause: where the models are building...
infinitive phrase: to do the next generation
```

**Grammar points:**
- **破折号插入语** - 打断主句流，增加口语化信息
- **where引导定语从句** - 修饰point，说明到达的状态
- **不定式表目的** - to do表示研究的目的

### [15:56]
**Original:** Like what, tell us how we should think about this idea of recursive self-improvement in the models themselves, because it seems like getting there first is incredibly important because then you just can continue to separate yourself versus those that haven't reached it yet.

**Translation:** 比如什么，告诉我们应该如何思考模型本身递归自我改进的这个想法，因为似乎首先到达那里是极其重要的，因为那样你就可以继续将自己与那些尚未达到的人区分开来。

**Core structure:**
- Tell us how we should think about this idea, because getting there first is important.  
  告诉我们应该如何思考这个想法，因为首先到达很重要。

**Structure tree:**
```
imperative: tell us
indirect question: how we should think about...
causal clause 1: because getting there first is important
causal clause 2: because you can continue to separate yourself
relative clause: that haven't reached it yet
```

**Grammar points:**
- **间接疑问句** - how引导的从句作tell的宾语
- **动名词作主语** - getting there first作主语
- **嵌套because从句** - 两层原因状语从句，第二个解释第一个

### [16:09]
**Original:** I would say we do see progress accelerating. We see, you know, I can't speak for other companies, but for us the scaling laws are, you know, alive and well and we're seeing that, you know, even with releases more recently like Mythos, right now within the company, you know, 90 plus percent of our code is actually written by Claude Code, right?

**Translation:** 我想说我们确实看到进展在加速。我们看到，你知道，我不能代表其他公司发言，但对我们来说，缩放定律是，你知道，活跃且良好的，我们看到，你知道，即使是最近的发布，比如Mythos，现在在公司内部，你知道，我们90%以上的代码实际上是由Claude Code编写的，对吧？

**Core structure:**
- For us the scaling laws are alive and well, and 90 plus percent of our code is written by Claude Code.  
  对我们来说缩放定律良好，90%以上的代码由Claude Code编写。

**Structure tree:**
```
main clause 1: the scaling laws are alive and well
coordinate clause: we're seeing that...
object clause: 90 plus percent of code is written...
parenthetical insertions: you know (multiple)
prepositional phrases: with releases, within the company
```

**Grammar points:**
- **多重插入语** - you know反复出现，模拟口语停顿
- **被动语态** - is written强调代码的生成方式

### [17:43]
**Original:** How do you think about this weird world where you mentioned the talent and the leverage and they're not writing code themselves and Claude's writing its own code?

**Translation:** 你如何看待这个奇怪的世界，在这个世界里你提到了人才和杠杆作用，他们自己不写代码，而Claude在写自己的代码？

**Core structure:**
- How do you think about this world where they're not writing code and Claude's writing its own code?  
  你如何看待这个世界，他们不写代码而Claude在写自己的代码？

**Structure tree:**
```
main question: How do you think about this world
relative clause: where you mentioned...
coordinate clauses: they're not writing / Claude's writing
object series: the talent and the leverage
```

**Grammar points:**
- **where引导定语从句** - 修饰world，描述这个世界的特征
- **并列结构** - 多个and连接的并列成分

### [18:35]
**Original:** But having the best talent to set the direction, not just the priorities, but some of the new areas of discovery, it just actually makes that research talent even better, right?

**Translation:** 但是拥有最好的人才来设定方向，不仅仅是优先事项，还有一些新的发现领域，这实际上使研究人才变得更好，对吧？

**Core structure:**
- Having the best talent makes that research talent even better.  
  拥有最好的人才使研究人才变得更好。

**Structure tree:**
```
subject: having the best talent (gerund phrase)
infinitive: to set the direction
appositive expansion: not just priorities, but new areas
main verb: makes
object: that research talent
complement: even better
```

**Grammar points:**
- **动名词短语作主语** - having...整个短语作主语
- **not just...but结构** - 递进关系，强调范围扩大
- **make + 宾语 + 宾补** - 使役动词结构

### [20:06]
**Original:** But customers tell us things like, "Hey, I wish the model were better at this or I had this particular place where it got stuck and I could build this other product, but the capability needs to be further than that."

**Translation:** 但客户会告诉我们这样的事情:「嘿,我希望模型在这方面能更好,或者我遇到了一个特定的卡住的地方,我本可以构建另一个产品,但能力需要比那更进一步。」

**Core structure:**
- Customers tell us things like "I wish the model were better."  
  客户告诉我们「我希望模型能更好」这样的事情。

**Structure tree:**
```
main clause: customers tell us things
direct speech: I wish... / I had... / I could build... / but the capability needs...
wish clause: the model were better (subjunctive)
compound structure: or I had... and I could build...
contrast: but the capability needs to be further
```

**Grammar points:**
- **虚拟语气 (wish + were)** - wish 后用虚拟语气表达与现实相反的愿望,用 were 而非 was
- **复合并列结构** - 多个 or/and/but 连接的从句,表达复杂的客户反馈

### [20:26]
**Original:** And so there is this connected loop, but internally we're always looking at different models that are being trained, different snapshots that we have, and comparing them internally and to a lesser extent externally against our own measure and then ultimately how our customers view them as well.

**Translation:** 所以存在这样一个连接的循环,但在内部我们总是在查看正在训练的不同模型、我们拥有的不同快照,并在内部进行比较,在较小程度上也对外部进行比较,对照我们自己的标准,然后最终也看我们的客户如何看待它们。

**Core structure:**
- We're looking at different models and comparing them against our measure and how customers view them.  
  我们在查看不同模型并对照我们的标准和客户的看法进行比较。

**Structure tree:**
```
main clause: we're always looking at... and comparing them...
object 1: different models (that are being trained)
object 2: different snapshots (that we have)
comparing: internally and externally
against: our own measure / how customers view them
```

**Grammar points:**
- **并列宾语 + 定语从句** - 多个宾语(models, snapshots)各带定语从句,结构复杂
- **to a lesser extent** - 插入语表示程度递减,增加句子层次

### [20:50]
**Original:** Obviously a bunch of the authors of the scaling laws papers are amongst our founders, but you know, notwithstanding that, we can be a bit of a skeptical bunch.

**Translation:** 显然,许多扩展定律论文的作者都在我们的创始人当中,但你知道,尽管如此,我们可能是一群有点怀疑态度的人。

**Core structure:**
- Authors are amongst our founders, but we can be skeptical.  
  作者在我们创始人中,但我们可能持怀疑态度。

**Structure tree:**
```
clause 1: authors are amongst our founders
clause 2: but we can be a skeptical bunch
parenthetical: you know, notwithstanding that
modifier: of the scaling laws papers
```

**Grammar points:**
- **notwithstanding** - 正式用语,表示「尽管」,相当于 despite
- **插入语 (you know)** - 口语化插入语,打断句子流畅度,增加听力难度

### [21:39]
**Original:** And then having a very low bar for updating your current prior or your current perspective, because it could be the case that something a month ago was true that's just not true today and that breaks your model.

**Translation:** 然后对更新你当前的先验或当前的观点设置非常低的门槛,因为可能会出现这样的情况:一个月前正确的事情今天就不正确了,这会打破你的模型。

**Core structure:**
- Having a low bar for updating your perspective, because something true before is not true today.  
  对更新观点设置低门槛,因为以前正确的事情现在不正确了。

**Structure tree:**
```
main: having a low bar for updating...
reason: because it could be the case that...
nested clause: something (that) was true / (that) is not true / (that) breaks your model
time contrast: a month ago vs. today
```

**Grammar points:**
- **动名词短语作主语** - having a low bar... 整个动名词短语作主语
- **多重定语从句省略** - something 后连续三个省略 that 的定语从句,需仔细辨析

### [24:08]
**Original:** And so if you think about it, it's a bit of this layer cake of compute that's starting at different times with different capabilities, and we're very dynamically comparing that compute.

**Translation:** 所以如果你想一想,这有点像一个计算的层蛋糕,在不同时间以不同能力开始,我们非常动态地比较这些计算资源。

**Core structure:**
- It's a layer cake of compute, and we're comparing that compute.  
  这是一个计算的层蛋糕,我们在比较这些计算资源。

**Structure tree:**
```
conditional: if you think about it
main 1: it's a layer cake of compute
modifier: that's starting at different times with different capabilities
main 2: we're comparing that compute
adverb: very dynamically
```

**Grammar points:**
- **隐喻表达 (layer cake)** - 用「层蛋糕」比喻分层的计算资源结构,需理解隐喻含义
- **定语从句 + 介词短语** - that's starting... with... 修饰 compute,描述其特征

### [26:55]
**Original:** And you know I would say that you know a year or two ago it would have been harder to consume especially like a heterogeneous kind of compute drop in your example really quickly because these chip platforms are different and they are different, some are harder to operate, some of them have you know idiosyncrasies in terms of how we use it.

**Translation:** 我想说,在一两年前,要快速消耗掉你例子中那种异构计算资源会更困难,因为这些芯片平台是不同的,它们确实不同,有些更难操作,有些在我们使用方式上有特殊性。

**Core structure:**
- It would have been harder to consume compute because these platforms are different.  
  消耗计算资源会更困难,因为这些平台是不同的。

**Structure tree:**
```
main: it would have been harder to consume...
reason clause: because these chip platforms are different
parallel structure: some are harder..., some have idiosyncrasies...
time marker: a year or two ago
```

**Grammar points:**
- **虚拟语气(过去)** - would have been 表示对过去情况的假设
- **并列结构** - some...some... 列举不同情况

### [28:36]
**Original:** If you think about the cloud platform and all the tools and services that are now built in—because it's not just the raw model access, it is, you know, prompt caching and the ability to use virtual machines and, you know, cloud code being called within there or dispatch or the cloud agents SDK, managed agents.

**Translation:** 如果你想想云平台以及现在内置的所有工具和服务——因为这不仅仅是原始模型访问,还包括提示缓存、使用虚拟机的能力、在其中调用的云代码、调度、云代理SDK、托管代理等。

**Core structure:**
- Think about the platform and all the tools that are built in.  
  想想这个平台和所有内置的工具。

**Structure tree:**
```
conditional: If you think about...
object: the cloud platform and tools
relative clause: that are now built in
explanatory clause: because it's not just..., it is...
long enumeration: prompt caching and... and... or... or...
```

**Grammar points:**
- **破折号插入** - 用破折号插入解释性从句,打断主句
- **长列举** - 多个并列成分用 and/or 连接

### [29:15]
**Original:** Number one, if we feel like we have a vision into where the models are going and we can kind of demonstrate that and create customer value in that, that might be something like Claude Code, right?

**Translation:** 第一,如果我们觉得我们对模型的发展方向有远见,并且能够展示这一点并在其中创造客户价值,那可能就是像Claude Code这样的东西,对吧?

**Core structure:**
- If we have a vision and can demonstrate that, that might be something like Claude Code.  
  如果我们有远见并能展示它,那可能就是像Claude Code这样的东西。

**Structure tree:**
```
conditional: if we feel like...
nested clause: where the models are going
parallel verbs: have... and can demonstrate... and create...
main result: that might be something like...
```

**Grammar points:**
- **feel like + 从句** - 表示主观感觉或认为
- **指示代词 that** - 指代前面整个条件从句的内容

### [30:04]
**Original:** We also think that there's so much value that's going to accrue in some of these areas that, you know, our customers can win and we can win as well, which is why you've seen as we've launched some of these products, we've done them in a collaborative kind of partnership-oriented way, whether that be on the security side or, you know, design or financial services—we've partnered with the ecosystem.

**Translation:** 我们还认为,在这些领域中将会积累如此多的价值,以至于我们的客户可以赢,我们也可以赢,这就是为什么你会看到,当我们推出这些产品时,我们以一种协作的、面向合作伙伴的方式来做,无论是在安全方面、设计方面还是金融服务方面——我们都与生态系统合作。

**Core structure:**
- There's so much value that customers can win and we can win, which is why we've done them in a collaborative way.  
  有如此多的价值,客户和我们都能赢,这就是为什么我们以协作方式来做。

**Structure tree:**
```
main: there's so much value that...
result clause: that customers can win and we can win
relative clause: which is why you've seen...
temporal clause: as we've launched...
enumeration: whether that be... or... or...
```

**Grammar points:**
- **so...that 结构** - 表示程度和结果
- **which 指代前句** - which 指代整个前面的情况
- **whether...or 列举** - 列举多种可能情况

### [31:41]
**Original:** That like there is an element of what's happened, you know, in prior waves over the course of 5 years, 10 years, 20 years—it's happening in months now.

**Translation:** 也就是说,在之前的浪潮中,在5年、10年、20年的过程中发生的事情——现在在几个月内就发生了。

**Core structure:**
- What happened over years is happening in months now.  
  多年发生的事情现在几个月就发生了。

**Structure tree:**
```
subject: there is an element of what's happened
relative clause: what's happened in prior waves
time phrase: over the course of 5 years, 10 years, 20 years
contrast: it's happening in months now
```

**Grammar points:**
- **what 引导名词性从句** - what's happened 作介词 of 的宾语
- **破折号对比** - 用破折号引出对比信息,强调时间压缩

### [32:33]
**Original:** But part of it is also we want to make those capabilities really accessible, and that should accrue a lot of value to customers as well, and customers that are forward-footed on that and adopt, and frankly also ones that are building and using the tools that we offer on our platform.

**Translation:** 但其中一部分原因也是我们想让这些能力真正易于获取,这应该也会为客户带来很多价值,包括那些在这方面积极主动并采用的客户,坦率地说,还有那些正在构建和使用我们平台上提供的工具的客户。

**Core structure:**
- Part of it is we want to make capabilities accessible, and that should accrue value to customers.  
  部分原因是我们想让能力易于获取,这应该为客户带来价值。

**Structure tree:**
```
main clause 1: part of it is we want to make...
main clause 2: that should accrue value to customers
main clause 3: and customers that are forward-footed...
main clause 4: and ones that are building...
relative clauses: that are forward-footed / that are building and using
```

**Grammar points:**
- **多重并列结构** - 四个and连接的并列成分,层层递进,口语中容易迷失主线。
- **省略结构** - customers后省略了should accrue value,ones指代customers。
- **forward-footed** - 复合形容词,意为'积极主动的',非常用表达。

### [33:11]
**Original:** But actually what's happening is it's going up in many cases, and this is true at different levels, whether it be the Mythos pricing that is quite high because it's so powerful, the cost of an H100, the rental price of...

**Translation:** 但实际上正在发生的是,在许多情况下价格正在上涨,这在不同层面都是真实的,无论是因为非常强大而相当高的Mythos定价,还是H100的成本,或是租赁价格...

**Core structure:**
- What's happening is it's going up, and this is true at different levels.  
  正在发生的是价格在上涨,这在不同层面都是真实的。

**Structure tree:**
```
main clause: what's happening is it's going up
subject clause: what's happening
coordinate clause: this is true at different levels
whether clause: whether it be the Mythos pricing...
appositive examples: the cost of..., the rental price of...
```

**Grammar points:**
- **whether it be虚拟语气** - whether引导让步状语从句,用be原形表示虚拟,意为'无论是...还是...'。
- **同位语列举** - 多个名词短语并列举例,句子未完成(被打断)。

### [34:18]
**Original:** It's really because we found that Opus-class models were underutilized relative to their capability, right?

**Translation:** 这真的是因为我们发现Opus级别的模型相对于其能力而言未被充分利用,对吧?

**Core structure:**
- It's because we found that models were underutilized.  
  这是因为我们发现模型未被充分利用。

**Structure tree:**
```
main clause: It's because we found that...
object clause: that Opus-class models were underutilized
modifier: relative to their capability
```

**Grammar points:**
- **relative to** - 介词短语,'相对于',表示比较关系。
- **underutilized** - under-前缀表示'不足',utilized是过去分词作表语。

### [36:03]
**Original:** It's so unbelievably capital intensive to build these frontier labs. You've got the leverage we talked about, which is efficiency, price—both those things relate to margin.

**Translation:** 建立这些前沿实验室需要如此难以置信的资本密集投入。你拥有我们谈到的杠杆,也就是效率和价格——这两者都与利润率相关。

**Core structure:**
- It's capital intensive to build labs. You've got leverage, which is efficiency and price.  
  建立实验室需要资本密集投入。你拥有杠杆,即效率和价格。

**Structure tree:**
```
sentence 1: It's capital intensive to build...
sentence 2: You've got the leverage
relative clause: which is efficiency, price
appositive: both those things relate to margin
```

**Grammar points:**
- **It's + adj + to do结构** - it作形式主语,to build...是真正主语。
- **破折号同位语** - 破折号后补充说明leverage的具体内容。
- **capital intensive** - 复合形容词,'资本密集型的',商业术语。

### [36:50]
**Original:** If you think of all of those, they're kind of in support of revenue over different time scales, right? If I serve inference, it's in support of revenue today. If I do model development, it might help for a capability that unlocks TAM that drives revenue 6 months from now and everything in between.

**Translation:** 如果你考虑所有这些,它们在不同的时间尺度上都是在支持收入,对吧?如果我提供推理服务,它支持的是今天的收入。如果我做模型开发,它可能有助于一种能力,这种能力会解锁TAM,从而推动6个月后的收入以及介于两者之间的一切。

**Core structure:**
- They support revenue over different time scales. Inference supports revenue today. Model development helps for a capability that drives revenue later.  
  它们在不同时间尺度上支持收入。推理支持今天的收入。模型开发有助于推动未来收入的能力。

**Structure tree:**
```
conditional 1: If I serve inference, it supports revenue today
conditional 2: If I do model development, it might help...
relative clause 1: that unlocks TAM
relative clause 2: that drives revenue 6 months from now
modifier: and everything in between
```

**Grammar points:**
- **嵌套定语从句** - capability后跟两个that从句,层层修饰,逻辑链条长。
- **in support of** - 介词短语,'支持,为了支持',正式表达。
- **everything in between** - 省略结构,指'介于今天和6个月后之间的所有时间点'。

### [38:03]
**Original:** And so, you know, I think it's something where we think of the compute envelope that we have as the thing that is able to govern how much we're able to drive revenue both over the short term and the long term.

**Translation:** 所以，你知道，我认为这是一个我们将所拥有的计算资源包视为能够控制我们在短期和长期内能够驱动多少收入的东西的情况。

**Core structure:**
- We think of the compute envelope as the thing that governs how much we drive revenue.  
  我们将计算资源包视为控制我们驱动多少收入的东西。

**Structure tree:**
```
main: we think of X as Y
X: the compute envelope that we have
Y: the thing that is able to govern...
nested clause: how much we're able to drive revenue
time modifier: both over the short term and the long term
```

**Grammar points:**
- **think of A as B 结构** - 将A视为B，后接复杂的定语从句修饰B
- **嵌套从句** - 定语从句中包含宾语从句(how much...)

### [38:44]
**Original:** And so, our ecosystem, you know, we are the only model that's on all three clouds today. We're the only language Lab that's using all three of these chip platforms, and really these collaborations are much deeper than just like procurement.

**Translation:** 所以，我们的生态系统，你知道，我们是目前唯一在所有三个云平台上的模型。我们是唯一使用所有这三个芯片平台的语言实验室，而且这些合作实际上比单纯的采购要深入得多。

**Core structure:**
- We are the only model on all three clouds and the only lab using all three chip platforms.  
  我们是唯一在三个云平台上的模型，也是唯一使用三个芯片平台的实验室。

**Structure tree:**
```
compound sentence: three independent clauses
clause 1: we are the only model that's on all three clouds
clause 2: We're the only language Lab that's using...
clause 3: these collaborations are much deeper than...
```

**Grammar points:**
- **并列复合句** - 三个独立分句通过句号和连词连接
- **定语从句修饰** - that引导的定语从句修饰model和Lab

### [39:15]
**Original:** But these are multifaceted partnerships, whether it be on developing the chips themselves, landing that capacity, serving it, and then ultimately distributing it to the customers.

**Translation:** 但这些是多方面的合作伙伴关系，无论是在开发芯片本身、获取产能、提供服务，还是最终将其分发给客户方面。

**Core structure:**
- These are multifaceted partnerships.  
  这些是多方面的合作伙伴关系。

**Structure tree:**
```
main: these are multifaceted partnerships
whether clause: listing four parallel activities
- developing the chips
- landing that capacity
- serving it
- distributing it to customers
```

**Grammar points:**
- **whether...or 结构的省略形式** - whether it be表示列举多种情况，使用虚拟语气
- **并列动名词短语** - 四个动名词短语作whether从句的宾语

### [41:04]
**Original:** And on top of that we built an MFR, a monthly financial review skill, and it can produce our monthly financial review. It's 90 to 95% ready and then all of our discussion becomes about what do we do, what are the implications, not what exactly happened because Claude is Not just reporting the weather, it's also helping to think about drivers and like why did the number change in the way it did.

**Translation:** 除此之外，我们构建了一个MFR，即月度财务审查技能，它可以生成我们的月度财务审查。它已经完成了90%到95%，然后我们所有的讨论都变成了关于我们该做什么、有什么影响，而不是到底发生了什么，因为Claude不仅仅是在报告天气，它还帮助思考驱动因素，比如为什么数字会以这种方式变化。

**Core structure:**
- Our discussion becomes about what we do and implications, not what happened, because Claude helps think about drivers.  
  我们的讨论变成了关于我们做什么和影响，而不是发生了什么，因为Claude帮助思考驱动因素。

**Structure tree:**
```
main: our discussion becomes about X, not Y
X: what do we do, what are the implications (two questions)
Y: what exactly happened
reason clause: because Claude is not just...but also...
nested: why did the number change
```

**Grammar points:**
- **not...but also 并列结构** - 强调Claude的双重功能
- **间接疑问句作介词宾语** - what do we do等疑问句作about的宾语
- **嵌套原因状语从句** - because从句中包含另一个why引导的从句

### [42:10]
**Original:** Within the finance team are actually the biggest users of tokens, so it is not just, you know, the 22-year-old who joined and has a coding background and was doing that on the weekends and brought it to work.

**Translation:** 在财务团队中实际上是最大的代币使用者，所以这不仅仅是那个加入公司、有编程背景、周末在做这个并把它带到工作中的22岁年轻人。

**Core structure:**
- It is not just the 22-year-old who joined and brought it to work.  
  这不仅仅是那个加入并把它带到工作中的22岁年轻人。

**Structure tree:**
```
main: it is not just the 22-year-old
relative clause chain modifying '22-year-old':
- who joined
- (who) has a coding background
- (who) was doing that on weekends
- (who) brought it to work
```

**Grammar points:**
- **多重定语从句并列** - 四个并列的定语从句修饰同一先行词，省略了重复的who
- **倒装结构** - 句首'Within the finance team are'为地点状语提前的完全倒装

### [44:36]
**Original:** I maybe have a slightly different view on it in that I think like it has made—you know, we've been able to hire great people at the company, but it has made even those incredibly talented people so much more productive, and there's a little bit of this—I think of it again like Jevons paradox, but for labor—which is that we have people who become incredibly more productive.

**Translation:** 我可能对此有稍微不同的看法,因为我认为它让我们能够雇佣到优秀的人才,但它也让这些极具才华的人变得更加高效,这有点像杰文斯悖论,但是针对劳动力的——也就是说我们的员工变得更加高效。

**Core structure:**
- I have a different view in that it has made people more productive.  
  我有不同的看法,因为它让人们更高效。

**Structure tree:**
```
main: I have a view
  reason clause: in that I think...
    object clause 1: it has made people productive
    interruption: you know, we've been able to...
    coordination: but it has made...
    parenthetical: I think of it like Jevons paradox
    relative clause: which is that...
```

**Grammar points:**
- **in that 引导原因从句** - 表示'因为,在于',比 because 更正式
- **破折号插入语** - 多处插入打断主句流畅性
- **which 指代前文整体** - 指代整个 Jevons paradox 概念

### [46:36]
**Original:** They were also around, you know, our mission and how we approach things. People said, well, hey, aren't AI safety and building a really big business—aren't those things at odds?

**Translation:** 这些问题也围绕着我们的使命以及我们如何处理事情。人们说,嘿,AI安全和建立一个大企业——这两件事不是相互矛盾的吗?

**Core structure:**
- People said aren't AI safety and building a business at odds?  
  人们说AI安全和建立企业不是矛盾的吗?

**Structure tree:**
```
main: People said...
  quoted question: aren't AI safety and building business at odds?
    interruption: well, hey
    repetition: aren't those things at odds?
    dash break: between subjects and repetition
```

**Grammar points:**
- **否定疑问句** - aren't 开头表示预期肯定回答
- **at odds** - 固定搭配,表示'矛盾,不一致'
- **破折号重复强调** - 用 those things 重述前面内容加强语气

### [48:44]
**Original:** We pioneered alignment science, which is you want the model to do what you tell it to do, and how often does it do that and how often does it stray from that.

**Translation:** 我们开创了对齐科学,也就是你希望模型按照你的指令行事,它多久能做到这一点,又多久会偏离这一点。

**Core structure:**
- We pioneered alignment science, which is about model behavior.  
  我们开创了对齐科学,它关于模型行为。

**Structure tree:**
```
main: We pioneered alignment science
  relative clause: which is...
    definition: you want the model to do what you tell it
      embedded clause: what you tell it to do
    coordination: how often does it do that
    coordination: how often does it stray
```

**Grammar points:**
- **which is 引导定义性从句** - 解释前面的专业术语
- **what 引导宾语从句** - what you tell it = the thing that you tell it
- **并列疑问句** - 两个 how often 问句并列说明频率

### [49:02]
**Original:** And then the last linkage, if you're selling to enterprises, like we now sell to nine of the Fortune 10, all of those enterprises are entrusting us with, you know, customer information, with their data.

**Translation:** 然后最后一个关联是,如果你向企业销售,就像我们现在向财富10强中的9家销售一样,所有这些企业都在把客户信息、他们的数据托付给我们。

**Core structure:**
- If you're selling to enterprises, those enterprises are entrusting us with information.  
  如果你向企业销售,这些企业就在把信息托付给我们。

**Structure tree:**
```
fragment: the last linkage
  condition: if you're selling to enterprises
    example: like we sell to nine of Fortune 10
  main: all enterprises are entrusting us
    objects: customer information, their data
    filler: you know
```

**Grammar points:**
- **条件状语从句** - if 从句描述前提条件
- **entrust sb with sth** - 固定搭配'把某物托付给某人'
- **like 引导举例** - 口语化用法,相当于 such as

### [51:13]
**Original:** But yeah, definitely the first time I saw it, you have all these arguments about the laws of physics and law of large numbers and this can't, you know, where is the revenue coming from and how can it be added this quickly and how can customers move this quickly and is this even possible in enterprise and all of those things start to get broken down over time as you see how the business works internally and you see how the adoption curves and the exponentials that are happening.

**Translation:** 但是,确实第一次看到时,你会有所有这些关于物理定律和大数定律的论点,这不可能,你知道,收入从哪里来,怎么能增长这么快,客户怎么能行动这么快,这在企业中甚至可能吗,所有这些问题随着时间推移开始被打破,当你看到业务内部如何运作,看到采用曲线和正在发生的指数增长时。

**Core structure:**
- The first time I saw it, you have arguments, and those things get broken down as you see how the business works.  
  第一次看到时,你有这些论点,这些问题随着你看到业务如何运作而被打破。

**Structure tree:**
```
time clause: the first time I saw it
main: you have arguments
  list: laws of physics, law of large numbers
  questions: where is revenue from, how can it be added, how can customers move, is this possible
main: all things get broken down
  time clause: as you see...
    how clause 1: how business works
    how clause 2: how curves and exponentials are happening
```

**Grammar points:**
- **多重并列疑问句** - 连续5个疑问句表达各种质疑
- **get broken down** - 被动语态表示'被打破,被推翻'
- **as 引导时间状语从句** - 表示'随着...'

### [52:23]
**Original:** Thinking of it as, you know, not just something that is like a variable cost over some time period, but really this resource that's so fully utilized, right?

**Translation:** 把它看作，你知道的，不仅仅是某个时间段内的可变成本，而是真正被充分利用的资源，对吧?

**Core structure:**
- Thinking of it as not just a variable cost, but a fully utilized resource.  
  把它看作不仅是可变成本，而是被充分利用的资源。

**Structure tree:**
```
gerund phrase: Thinking of it as...
not just X but Y structure
X: something that is like a variable cost
Y: this resource that's so fully utilized
relative clauses modifying both X and Y
```

**Grammar points:**
- **not just...but (also)** - 表示递进关系，强调后者更重要
- **动名词短语作独立成分** - Thinking of it as 作为话语标记引出解释

### [53:09]
**Original:** Here, you really have that fungibility that's possible, and I think that's where the return on compute is so important.

**Translation:** 在这里，你真的拥有这种可能的可替代性，我认为这就是计算回报如此重要的原因。

**Core structure:**
- You have that fungibility, and that's where the return is important.  
  你拥有这种可替代性，这就是回报重要的原因。

**Structure tree:**
```
compound sentence: clause 1 + and + clause 2
clause 1: you have that fungibility that's possible
clause 2: I think that's where the return is important
where clause: 表语从句说明原因/位置
```

**Grammar points:**
- **where引导表语从句** - where在此表示抽象的'方面/原因'，非具体地点
- **指示代词that的多重指代** - 第一个that指fungibility，第二个that指前面整个情况

### [54:15]
**Original:** These are the massive kind of unprecedented investments that companies like us are making. What is the return that you're generating on that and when does it come and what is the shape of it?

**Translation:** 这些是像我们这样的公司正在进行的大规模前所未有的投资。你们从中产生的回报是什么，何时到来，以及它的形态是什么?

**Core structure:**
- What is the return you're generating, when does it come, and what is the shape?  
  回报是什么，何时到来，形态如何?

**Structure tree:**
```
three parallel questions connected by 'and'
Q1: What is the return (that you're generating on that)
Q2: when does it come
Q3: what is the shape of it
relative clause: that you're generating on that
```

**Grammar points:**
- **并列疑问句** - 三个wh-问句用and连接，共享同一话题
- **定语从句省略** - return后省略了关系代词that/which

### [56:42]
**Original:** I think what's one of the things that's different about AI is it's all happening so quickly. You can have, you know, years or decades of progress that are being compressed into months.

**Translation:** 我认为AI的不同之处之一是，这一切发生得如此之快。你可以看到，数年或数十年的进步被压缩到几个月内。

**Core structure:**
- What's different is it's happening quickly. Progress is being compressed into months.  
  不同之处是它发生得很快。进步被压缩到几个月内。

**Structure tree:**
```
sentence 1: I think [what's different] is [it's happening quickly]
what clause: 主语从句
that clause: 定语从句修饰things
sentence 2: You can have progress that are compressed
relative clause: that are being compressed into months
```

**Grammar points:**
- **what引导的主语从句** - what's...that's different作整体主语，表示'...的事情'
- **被动语态的进行时** - are being compressed强调持续进行的被动过程

### [58:03]
**Original:** And so I think people generally gravitate towards more honest and balanced assessments, right? If I feel like somebody's just telling me all the good news and none of the bad news, then I'm like, okay, do I really trust this perspective?

**Translation:** 所以我认为人们通常倾向于更诚实和平衡的评估，对吧?如果我觉得有人只告诉我所有好消息而没有坏消息，那么我会想，好吧，我真的相信这个观点吗?

**Core structure:**
- People gravitate towards honest assessments. If somebody tells me only good news, do I trust this?  
  人们倾向于诚实的评估。如果有人只告诉我好消息，我会相信吗?

**Structure tree:**
```
conditional structure: If...then...
if clause: I feel like somebody's telling me X and none of Y
then clause: I'm like, do I trust this
feel like: 后接宾语从句
embedded question: do I really trust this perspective
```

**Grammar points:**
- **feel like后接从句** - feel like可接that从句(省略that)表示'觉得/感觉'
- **I'm like口语用法** - 口语中like引出说话人的想法或反应，相当于'I think/say'

### [58:49]
**Original:** And so over the long term, the opportunity is going to be significantly higher and greater than, you know, some of the risks and the downsides that will happen.

**Translation:** 因此从长远来看，机会将会显著地高于并大于那些将会发生的风险和不利因素。

**Core structure:**
- The opportunity is going to be higher than the risks.  
  机会将会高于风险。

**Structure tree:**
```
main clause: the opportunity is going to be higher than...
time adverbial: over the long term
comparative structure: higher and greater than
object of comparison: the risks and downsides that will happen
```

**Grammar points:**
- **比较级结构** - higher and greater than 构成双重比较，强调程度差异
- **定语从句** - that will happen 修饰 risks and downsides

### [59:03]
**Original:** It was the first time many people, friends of mine that are careful watchers of this stuff said something like, "I'm like, this one kind of makes me scared."

**Translation:** 这是第一次许多人，我那些密切关注这些事情的朋友们说出类似这样的话：'我觉得，这个有点让我害怕。'

**Core structure:**
- It was the first time people said something.  
  这是第一次人们说了某些话。

**Structure tree:**
```
main clause: It was the first time...
time clause: (that) many people said something
appositive: friends of mine
relative clause: that are careful watchers
direct speech: this one kind of makes me scared
```

**Grammar points:**
- **It was the first time (that)...** - 固定句型，that 常省略，后接完整句子
- **同位语 + 定语从句** - friends of mine 是 many people 的同位语，that 从句修饰 friends

### [01:00:16]
**Original:** You've seen these examples where, um, you know, we had an open source code base that, you know, a prior model found 22 security vulnerabilities in and Mythos then found 250.

**Translation:** 你已经看到这些例子，我们有一个开源代码库，之前的模型在其中发现了22个安全漏洞，而Mythos随后发现了250个。

**Core structure:**
- You've seen examples where a prior model found 22 vulnerabilities and Mythos found 250.  
  你已经看到例子，之前的模型发现了22个漏洞，Mythos发现了250个。

**Structure tree:**
```
main clause: You've seen these examples
relative clause: where we had a code base
nested relative: that a prior model found 22 vulnerabilities in
coordinate clause: and Mythos found 250
```

**Grammar points:**
- **介词 + 关系代词** - found vulnerabilities in (the code base)，in 提前到定语从句末尾
- **并列结构的省略** - 第二个分句省略了 in the code base，避免重复

### [01:00:33]
**Original:** So we didn't say we're never going to release it. We said let's do it in a phased way. Let's do it to, you know, a group that will expand over time where we can, you know, focus on this one cyber capability and how it can actually be used.

**Translation:** 所以我们没有说永远不会发布它。我们说让我们分阶段进行。让我们向一个会随时间扩大的群体发布，在那里我们可以专注于这一个网络安全能力以及它实际上如何被使用。

**Core structure:**
- We said let's do it to a group where we can focus on the capability and how it can be used.  
  我们说让我们向一个群体发布，在那里我们可以专注于这个能力以及它如何被使用。

**Structure tree:**
```
main clause: We said let's do it to a group
relative clause: that will expand over time
adverbial clause: where we can focus on...
coordinate objects: this capability and how it can be used
```

**Grammar points:**
- **关系副词 where** - where 引导定语从句修饰 group，表示抽象地点（情境）
- **并列宾语** - focus on 后接两个并列宾语：名词短语和 how 引导的从句

### [01:02:08]
**Original:** I do think that there's a balance, right? You want to be able to have innovation happen really quickly and have that not be slowed down, but you also want to have this kind of responsibility framework for how these things are deployed because we've long said that, you know, this technology has implications and we should have an honest conversation about them, and that includes with the government.

**Translation:** 我确实认为需要有一个平衡，对吧？你希望能够让创新快速发生并且不被减缓，但你也希望有这种责任框架来规范这些东西如何被部署，因为我们长期以来一直说这项技术有影响，我们应该就此进行坦诚的对话，这包括与政府的对话。

**Core structure:**
- You want innovation to happen quickly but also want a responsibility framework because this technology has implications.  
  你希望创新快速发生，但也希望有责任框架，因为这项技术有影响。

**Structure tree:**
```
main clause: You want to have innovation happen and have that not be slowed down
coordinate clause: but you also want to have a framework
purpose clause: for how these things are deployed
causal clause: because we've long said that...
nested object clause: this technology has implications
```

**Grammar points:**
- **have + 宾语 + 动词原形** - 使役动词结构，have innovation happen 表示'让创新发生'
- **复杂因果关系** - because 引导原因状语从句，内含 that 引导的宾语从句，层层嵌套

### [01:03:49]
**Original:** So somebody could be flying colors on everything else and really, really the smartest person you've met in this role, we won't hire them if they don't pass the culture bar.

**Translation:** 所以即使某人在其他所有方面都表现优异,而且确实是你在这个职位上遇到的最聪明的人,如果他们没有通过文化标准,我们也不会雇用他们。

**Core structure:**
- We won't hire them if they don't pass the culture bar.  
  如果他们没有通过文化标准,我们不会雇用他们。

**Structure tree:**
```
concessive clause: somebody could be flying colors...
coordinate clause: and (somebody could be) the smartest person...
main clause: we won't hire them
conditional clause: if they don't pass the culture bar
```

**Grammar points:**
- **让步状语前置** - 让步内容在主句之前,强调对比。
- **省略结构** - and 后省略了 somebody could be。
- **条件状语从句**

### [01:04:02]
**Original:** It's one incredibly collaborative, and this means that we don't really tolerate the fiefdoms or the sharp elbows or the like "I need to take credit for this."

**Translation:** 首先是极其协作,这意味着我们真的不能容忍山头主义、争强好胜或者那种'我需要为此邀功'的态度。

**Core structure:**
- It's collaborative, and this means that we don't tolerate fiefdoms.  
  它是协作的,这意味着我们不容忍山头主义。

**Structure tree:**
```
main clause 1: It's incredibly collaborative
coordinate clause: and this means that...
object clause: that we don't tolerate...
parallel objects: fiefdoms / sharp elbows / the like "I need..."
```

**Grammar points:**
- **并列宾语** - 三个 or 连接的并列成分。
- **指示代词 this** - 指代前面整个句子的内容。

### [01:04:57]
**Original:** After that happens, there's real alignment. So in something like compute allocation we were talking about before, people might have different perspectives on how to allocate that compute, but they will engage in a thoughtful discussion about where the returns are the highest or the best.

**Translation:** 在那之后,就会有真正的一致性。所以在像我们之前谈到的计算资源分配这样的事情上,人们可能对如何分配计算资源有不同的看法,但他们会进行深思熟虑的讨论,讨论在哪里回报最高或最好。

**Core structure:**
- People might have different perspectives, but they will engage in discussion.  
  人们可能有不同看法,但他们会进行讨论。

**Structure tree:**
```
time clause: After that happens
main clause: there's real alignment
contrast: people might have perspectives, but they will engage...
embedded question: how to allocate / where the returns are highest
```

**Grammar points:**
- **插入语** - we were talking about before 插入修饰 compute allocation。
- **间接疑问句** - how to 和 where 引导的从句作介词宾语。
- **转折连词 but**

### [01:05:23]
**Original:** And these are not softballs, they're not like planted questions, they're just real questions that are on people's minds, and he answers them the best that he can.

**Translation:** 这些不是容易回答的问题,也不是那种事先安排好的问题,它们只是人们心中真实的问题,而他会尽其所能地回答这些问题。

**Core structure:**
- These are real questions, and he answers them.  
  这些是真实的问题,他会回答它们。

**Structure tree:**
```
parallel negative clauses: not softballs / not planted questions
positive clause: they're just real questions
relative clause: that are on people's minds
coordinate clause: and he answers them the best that he can
```

**Grammar points:**
- **并列否定句** - 两个否定句强调对比。
- **定语从句** - that 修饰 questions。
- **方式状语** - the best that he can 表示尽力程度。

### [01:09:08]
**Original:** And so think of this as something that has context within your organization that can use all of the tools that are specific to you, whether they be homegrown tools or tools that you purchase, that has memory and the ability to effectively learn from mistakes you've made, but also mistakes that it's made over time.

**Translation:** 所以把它想象成这样一个东西:它在你的组织内有上下文,可以使用所有专属于你的工具,无论是自主开发的工具还是你购买的工具,它有记忆力,能够有效地从你犯过的错误中学习,也能从它自己随着时间犯过的错误中学习。

**Core structure:**
- Think of this as something that has context and memory.  
  把它想象成有上下文和记忆的东西。

**Structure tree:**
```
imperative: think of this as something
relative clause 1: that has context
relative clause 2: that can use tools
relative clause 3: that are specific to you
concessive clause: whether they be...
relative clause 4: that has memory and ability
infinitive phrase: to learn from mistakes
```

**Grammar points:**
- **多层定语从句嵌套** - 多个 that 从句层层修饰,结构复杂。
- **虚拟语气 whether...be** - whether 引导让步从句,用虚拟语气。
- **并列结构** - mistakes you've made 和 mistakes that it's made 并列。

### [01:10:04]
**Original:** But then you also see something like Cowork come along and start to unlock that co-working faster than Claude for code was if you index them to the same point in time.

**Translation:** 但随后你也会看到像 Cowork 这样的东西出现，并开始解锁协同工作的速度比 Claude for code 更快——如果你把它们对标到同一时间点的话。

**Core structure:**
- You see Cowork come along and start to unlock co-working faster than Claude was.  
  你看到 Cowork 出现并开始更快地解锁协同工作。

**Structure tree:**
```
main: you see something come along and start to unlock
comparison: faster than Claude was
conditional: if you index them to the same point
```

**Grammar points:**
- **感官动词 + 宾语 + 不带to的不定式** - see something come/start 结构，宾语补足语用动词原形
- **比较结构的省略** - than Claude was 后省略了 unlocking co-working

### [01:10:17]
**Original:** But I think it's because the model capabilities and the products are pushing towards this notion of a virtual collaborator where even our product development today is not done by like one product manager with two engineers shipping something over 3 months.

**Translation:** 但我认为这是因为模型能力和产品正在推动虚拟协作者这一概念，在这种模式下，即使是我们今天的产品开发也不再是由一个产品经理带两个工程师用三个月时间交付某个东西。

**Core structure:**
- It's because the capabilities and products are pushing towards this notion where development is not done by one manager with two engineers.  
  这是因为能力和产品正在推动这一概念，在这种模式下开发不再由一个经理带两个工程师完成。

**Structure tree:**
```
main: it's because capabilities are pushing towards notion
relative clause: where development is not done by...
modifier: one manager with two engineers shipping something
```

**Grammar points:**
- **where 引导定语从句修饰抽象名词** - where 修饰 notion，表示在这个概念/情况下
- **现在分词作后置定语** - shipping something 修饰 engineers，描述他们的工作状态

### [01:12:18]
**Original:** You walked me all the way home, and I remember I came in, I told my wife, I was like, this is going to be wild, like if even 10% of that is true, this is going to bend all paradigms of what not just things I've seen but what most people have seen.

**Translation:** 你一路把我送回家，我记得我进门后告诉我妻子，我说，这将会很疯狂，如果其中哪怕只有 10% 是真的，这都将打破所有范式——不仅是我见过的，而是大多数人见过的。

**Core structure:**
- I told my wife this is going to bend all paradigms of what people have seen.  
  我告诉妻子这将打破人们见过的所有范式。

**Structure tree:**
```
main: I told my wife
direct speech: this is going to be wild
conditional: if even 10% is true
result: this is going to bend paradigms
relative: of what people have seen
```

**Grammar points:**
- **多层嵌套的直接引语** - I was like 引入口语化引语，内含条件句和结果句
- **not just...but 结构** - 强调递进关系，从个人经验扩展到普遍经验

### [01:13:02]
**Original:** I'm hiring you as a partner and I want you to treat it as a partnership which means that there might be things that you and I disagree on.

**Translation:** 我雇用你是作为合作伙伴，我希望你把这当作一种伙伴关系，这意味着可能会有你我意见不一致的事情。

**Core structure:**
- I'm hiring you as a partner which means there might be things we disagree on.  
  我雇用你作为合作伙伴，这意味着可能会有我们意见不一致的事情。

**Structure tree:**
```
main: I'm hiring you as a partner
relative: which means that...
noun clause: that there might be things
relative: that you and I disagree on
```

**Grammar points:**
- **which 引导非限制性定语从句** - which 指代前面整个句子的内容
- **介词 + 关系代词的省略** - disagree on 后省略了 which/that，先行词是 things

### [01:13:27]
**Original:** Like that training is really valuable and thinking about things at a granular level and not losing that. Like I'm not somebody who is comfortable at 50,000 feet. That's just like not me. But you can't be at 500 feet at everything in this business.

**Translation:** 那种训练非常有价值，在细节层面思考问题并且不失去这一点。我不是那种在五万英尺高度感到舒适的人。那不是我的风格。但在这个行业里，你不可能在所有事情上都停留在 500 英尺的高度。

**Core structure:**
- That training is valuable and thinking at a granular level is important. But you can't be at 500 feet at everything.  
  那种训练很有价值，细节思考很重要。但你不能在所有事情上都保持在 500 英尺高度。

**Structure tree:**
```
parallel structure: training is valuable and thinking...and not losing
contrast: I'm not somebody who is comfortable at 50,000 feet
adversative: But you can't be at 500 feet at everything
```

**Grammar points:**
- **动名词短语作主语的并列** - thinking...and not losing 与 training 并列，共同作为讨论对象
- **隐喻表达** - 50,000 feet 和 500 feet 比喻思考的抽象程度，需理解商业语境

### [01:15:45]
**Original:** But I think the commonality of it was that, you know, everything is going to happen much quicker than we think and that both the implications but also the capabilities of that can change, and then he also had like a really incredible optimism about the future that I think, you know, we talk about internally kind of holding light and shade.

**Translation:** 但我认为其共同点是，你知道，一切都会比我们想象的发生得更快，而且这既包括影响也包括能力都可能改变，然后他对未来也有一种非常令人难以置信的乐观态度，我认为，你知道，我们内部谈论的那种保持光明与阴影的平衡。

**Core structure:**
- The commonality was that everything will happen quicker and he had optimism about the future.  
  共同点是一切都会发生得更快，而且他对未来很乐观。

**Structure tree:**
```
main: the commonality was that...
predicative clause 1: everything is going to happen...
predicative clause 2: both implications and capabilities can change
coordinate clause: he had optimism
relative clause: that we talk about internally
```

**Grammar points:**
- **多重并列从句** - that 引导的两个表语从句并列，后接 and 连接的独立分句
- **嵌套定语从句** - optimism 后的 that 从句修饰 optimism，内部又包含宾语从句
- **插入语** - you know 作为口语插入语打断句子流畅性

### [01:16:37]
**Original:** I think the first thing would be the diffusion rate within our customers, the use cases are playing catch up to the model capability, and I think, you know, look, these are—we are talking about humans in large organizations with a set of tools and practices and things that they've been doing for a really long time.

**Translation:** 我认为第一件事是我们客户内部的扩散速度，用例正在追赶模型能力，而且我认为，你知道，看，这些是——我们谈论的是大型组织中的人类，他们拥有一套工具和实践以及他们已经做了很长时间的事情。

**Core structure:**
- The first thing would be the diffusion rate, and we are talking about humans in large organizations.  
  第一件事是扩散速度，我们谈论的是大型组织中的人类。

**Structure tree:**
```
main 1: the first thing would be the diffusion rate
appositive: the use cases are playing catch up
main 2: we are talking about humans
modifier: with a set of tools and practices
relative clause: that they've been doing
```

**Grammar points:**
- **同位语结构** - 逗号后的句子解释说明 diffusion rate 的具体含义
- **复杂介词短语** - with 引导的介词短语包含多个并列成分和定语从句

### [01:18:32]
**Original:** I'm really most optimistic and excited about when it goes further back into drug development and drug discovery, because, you know, our humans are incredibly capable at research, but if you think about these molecules and proteins, like, they're so complex and such small changes have such big implications for the outcomes—like, AI is perfect for that.

**Translation:** 我真的最乐观和兴奋的是当它进一步深入到药物开发和药物发现时，因为，你知道，我们的人类在研究方面非常有能力，但如果你想想这些分子和蛋白质，就像，它们是如此复杂，如此微小的变化对结果有如此大的影响——就像，人工智能非常适合这个。

**Core structure:**
- I'm optimistic about when it goes into drug discovery, because humans are capable but AI is perfect for complex molecules.  
  我对它进入药物发现感到乐观，因为人类有能力但人工智能更适合复杂分子。

**Structure tree:**
```
main: I'm optimistic about when...
temporal clause: when it goes into drug development
causal clause: because humans are capable
adversative clause: but if you think about molecules
conclusion: AI is perfect for that
```

**Grammar points:**
- **介词 + 时间状语从句** - about 后接 when 引导的从句作宾语，较少见的结构
- **转折与条件嵌套** - because 从句内部包含 but 和 if 引导的复杂逻辑关系
- **破折号插入** - 破折号引出总结性评论，打断句子连贯性

### [01:18:54]
**Original:** If you think about what can happen when the lab's throughput goes up 10x or 100x and we can run that many more experiments, probably get better results faster, and that can be something that helps, you know, people around the world, right?

**Translation:** 如果你想想当实验室的吞吐量增加10倍或100倍时会发生什么，我们可以进行更多的实验，可能更快地获得更好的结果，而这可能是能够帮助世界各地人们的东西，对吧？

**Core structure:**
- If you think about what can happen when throughput goes up, that can help people.  
  如果你想想吞吐量增加时会发生什么，那可以帮助人们。

**Structure tree:**
```
conditional: If you think about...
object clause: what can happen when...
temporal clause: when throughput goes up
consequence 1: we can run more experiments
consequence 2: that can be something that helps people
```

**Grammar points:**
- **三层嵌套从句** - if 从句内嵌 what 宾语从句，其内又嵌 when 时间从句
- **省略主语的并列结构** - probably get 省略主语 we，与前句并列但易造成理解困难

### [01:20:14]
**Original:** You know, the financial aid packages weren't, you know, as robust as they are today. And a big factor in his decision, I found out, you know, many, many, many years later, was, you know, wanting to give me the opportunity to go wherever I wanted.

**Translation:** 你知道，当时的经济援助方案不像今天这样完善。而且他决定中的一个重要因素，我发现，你知道，很多很多年后，是，你知道，想给我去任何我想去的地方的机会。

**Core structure:**
- A big factor in his decision was wanting to give me the opportunity to go wherever I wanted.  
  他决定中的一个重要因素是想给我去任何地方的机会。

**Structure tree:**
```
main: A big factor was wanting to give...
parenthetical: I found out many years later
infinitive phrase: to give me the opportunity
infinitive clause: to go wherever I wanted
relative clause: wherever I wanted
```

**Grammar points:**
- **插入语分割主谓** - I found out 插在主语和系动词之间，严重打断句子结构
- **动名词作表语** - wanting 引导的动名词短语作表语，包含复杂的不定式结构

### [01:21:24]
**Original:** It's an AI platform built specifically for Wall Street, connected to your data, understanding your process, and producing real outputs.

**Translation:** 这是一个专门为华尔街打造的AI平台，连接到你的数据，理解你的流程，并产生真实的输出。

**Core structure:**
- It's an AI platform.  
  这是一个AI平台。

**Structure tree:**
```
main clause: It's an AI platform
post-modifier 1: built specifically for Wall Street (past participle)
post-modifier 2: connected to your data (past participle)
post-modifier 3: understanding your process (present participle)
post-modifier 4: producing real outputs (present participle)
```

**Grammar points:**
- **过去分词作后置定语** - built/connected 修饰 platform，表被动完成
- **现在分词作后置定语** - understanding/producing 修饰 platform，表主动进行
- **并列结构** - 四个分词短语并列修饰同一名词

### [01:21:37]
**Original:** The best AI and software companies from OpenAI to Cursor to Perplexity use WorkOS to become enterprise ready overnight, not in months.

**Translation:** 从OpenAI到Cursor再到Perplexity，最好的AI和软件公司使用WorkOS在一夜之间而非数月内做好企业准备。

**Core structure:**
- Companies use WorkOS to become enterprise ready.  
  公司使用WorkOS来做好企业准备。

**Structure tree:**
```
subject: The best AI and software companies
modifier: from OpenAI to Cursor to Perplexity
verb: use
object: WorkOS
purpose: to become enterprise ready
time contrast: overnight, not in months
```

**Grammar points:**
- **不定式表目的** - to become 说明使用WorkOS的目的
- **from...to...to 结构** - 列举多个例子的固定搭配

### [01:21:53]
**Original:** They've helped firms 5x and scale, enabling faster growth, smarter operations, and a competitive edge.

**Translation:** 他们帮助公司实现5倍增长和扩展，使其能够更快增长、更智能运营并获得竞争优势。

**Core structure:**
- They've helped firms scale.  
  他们帮助公司扩展。

**Structure tree:**
```
main clause: They've helped firms 5x and scale
result clause: enabling faster growth, operations, and edge
parallel objects: faster growth / smarter operations / competitive edge
```

**Grammar points:**
- **现在分词作结果状语** - enabling 表示帮助带来的自然结果
- **help + 宾语 + 动词原形** - help 后省略 to 的不定式结构