AI 氛围检查：编程代理、开放权重 AI 与算力紧张

我是 Jacob Efron，这里是 Unsupervised Learning。过去几个月我们新增了不少订阅者，所以想欢迎大家来到这个节目。我们基本上会与 AI 领域最敏锐的头脑探讨当下发生的一切：什么是真实的、接下来会发生什么、这个领域将走向何方。今天的节目是我们最喜欢的格式之一：与来自 Datlogy 的 Ari 和 Radical 的 Rob 一起做一次 AI 氛围检查。Ari 曾是 DeepMind 和 Meta 的研究员，现在经营着一家非常令人兴奋的 AI 初创公司；Rob 在 Radical，一家顶尖的 AI 风险投资公司。我们三人聊了当今 AI 世界发生的一切。我们当然聊了 Fable，以及发布后的反应和模型能力。我们讨论了离 RSI 有多近。我们还做出了一些相当大胆的预测，包括实验室可能会在算力紧缩持续的情况下放弃 API 业务。我们还覆盖了当今所有主要话题。能和两位朋友、AI 领域的两位智者坐下来聊天，真的很开心。我相信大家会喜欢这期节目。闲话少说，下面就是我们的对话。

I'm Jacob Efron and this is Unsupervised Learning. We've had a bunch of new subscribers over our last few months. And so, wanted to welcome you to the show. We basically probe the sharpest minds in AI on everything that's happening today, what's real, and what's coming up, where the space is headed. And today's episode is one of my favorite formats we do. It's an AI vibe check that we do with Ari from Datlogy. Ari was a former researcher at DeepMind and Meta, now runs a really exciting AI startup. And Rob at Radical, one of the great AI venture firms. The three of us talk about everything happening in the AI world today. We talked about Fable, of course, and the reaction around the release as well as model capabilities. We talked about how close we are to RSI. We hit on some pretty spicy predictions, including that the labs may actually get rid of their API business as the compute crunch continues. And we just got to hit on all the main topics of today. Just really fun to sit down with two friends and great minds in AI. I think folks will really enjoy this. Without further ado, here's our conversation.

又到了总结集的时间。我一直很喜欢和你们俩做这个节目，Ari 和 Rob。我觉得上次我们玩得很开心，但天哪，AI 世界变化太快了。上次我们坐下来是在 Nurups 之后，我想自那以后，我们看到了 IPO 申请，看到了模型先是不发布然后又发布，还看到 SpaceX 变成了一家 AI 信息公司。这里不缺可讨论的头条新闻。非常高兴你们俩能再次来到节目。

It's time for another roundup episode. I always love doing this with you guys, Ari and Rob. I feel like we had a ton of fun in the last one, but like god, it's AI world. Things have changed. I feel like we last sat down after Nurups, and I think since then we've had IPO filings. We've had models not launched and then launched. We've had SpaceX becoming an AI info company. No shortage of headlines to discuss here. So excited to have you both back on the show.

很高兴来到这里。

Excited to be here.

是的，谢谢邀请我们。

Yeah, thanks for having us.

那么，我想先开始吧。你知道，在 AI 世界里 6 个月就像永恒一样，但我想从最高层面开始。自从我们上次交谈以来，你对这个领域的看法中，最大的变化是什么？也许 Ari，我先从你开始。

So I think to kick it off, you know, 6 months is an eternity in AI world, but I figured I'd start at the highest level. What's the single biggest thing that has changed in how you're thinking about the landscape since we last talked? And maybe Ari, I'll start with you.

是的，我认为过去六个月里最明显的变化是开始看到编码智能体真正在更长的时间跨度上工作。没错，我觉得在我们 25 年底录制上一期节目时，这还只是刚刚开始。

Yeah, I mean I think the most obvious thing that has changed over the last six months is starting to see the coding agents really start to work at longer time horizons. Right, I think that was just starting when we recorded our last episode at the end of '25.

感觉每个人在圣诞假期回来后都惊呼：天哪，这些东西真的管用。

Feel like everyone went away over Christmas break and was like holy crap, like these things really work.

是的。我认为这开始显示出存在一些阈值，一旦超过某个阈值，价值就会大幅提升。显然，这推动了 token 支出的巨大增长，以及整个 token 最大化理念等等。但我认为我们现在真的开始看到工程师们几乎都在从独立贡献者转变为智能体管理者。例如，在 Datlogy，过去几个月里非常明显的一点是，越来越多的人开始在管理不同智能体之间切换上下文，而不是只专注于一件事。而这之所以成为可能，是因为这些智能体能够运行足够长的时间，并以各种方式真正发挥作用。

Yeah. And I think it starts to show how there are these thresholds where if you go beyond a threshold it can become a lot more valuable. And obviously that's driven the massive rise in token spending and the whole token maxing idea and all this stuff. But I think we're really starting to now see the shift of engineers at least kind of almost all moving from ICs to managers of agents. That's been something that's been very noticeable within Datlogy, for example, over the last number of months. It's seeing more and more people starting to context switch between managing different agents rather than just kind of working on the one thing. And that was enabled by having these agents be able to run long enough and actually be useful in various ways.

我的意思是，大家都喜欢问像你这样的顶尖 AI 研究员，它让你的工作效率提高了多少。我觉得这很有趣。在某些方面它确实让你高效得多，对吧？但它也带来了很多挑战。比如我们正在努力解决的一个问题是，现在更容易生成大量能完成任务的代码，但你现在面临一个相当大的理解差距，而且更容易把劣质代码塞进代码库。所以它确实让我们更高效了。不过我认为很多时候，那些表面数字往往被高估了，因为它没有考虑到一些后期成本，比如我们现在在审查上遇到了很大的瓶颈，而且我们不想完全变成“哦，你只要让我的智能体去审查你的智能体的输出”那种情况。你知道，瓶颈似乎只是转移了。不管怎样，正因为如此，很难改进整个流程。

I mean everyone likes to ask top AI researchers like yourself how much more productive has it made you in your work. I think that it's interesting. It makes you a lot more productive in some ways, right? But it also produces a lot of challenges as well. Like one of the things that we're struggling with is now it's a lot easier to produce a massive amount of code that can do something, but now you have this pretty massive understanding gap and it's a lot easier to put slop into your codebase. So it's definitely made us more productive. I think a lot of times though the kind of topline numbers tend to be overestimated because it doesn't take into account some of these later costs of like we now have big bottlenecks on reviews and we don't want to go fully to like oh you just like but my agent will review your agent's output. You know the bottlenecks just seem to shift. You know whatever, it's hard to improve on an entire process because of that.

那你呢，Rob？

What about you Rob?

过去六个月里出现了一些早期迹象，让我开始质疑开放权重 AI 是否还能继续成为生态系统中一股真正重要的力量，至少在接近前沿的领域是这样。

There are early signs that seem to suggest over the past six months that make me question whether openweight AI is going to continue to be a really meaningful force in the ecosystem going forward, at least for near frontier.

一上来就直击要害，Ari，我喜欢。

Opening for the jugular, Ari, right off the top here. I like it.

是的。我们可以深入探讨，但我觉得，是的。六个月前，或者说过去几年里，我的一个基本假设是闭源专有模型会推进前沿，这有很多结构性原因。但开放前沿只会落后几个月。这个差距可能会稍微拉大。我原本以为它不会完全缩小，但会保持相对较小。而现在我认为有迹象表明，接近前沿的开放权重 AI 确实存在完全掉队的风险。我认为 Meta，历史上一直是西方开放权重的冠军，正在退缩，而且看起来他们很可能不会继续他们的开源战略。然后最近，显然，中国实验室一直在推动最先进的开放研究。而在那里，似乎有强烈的迹象表明他们也可能在退缩。无论是 Qwen 还是 DeepSeek 还是其他，他们最高性能的模型现在都保留在 API 后面专有，只开源、开放权重那些更小、性能更低的版本。我认为这背后有真实的算力激励。服务这些没有收入进来的开放权重模型非常昂贵。而且我认为还有地缘政治和竞争方面的考虑。但有趣的是，想象一下这样一个世界：如果你想要真正的前沿人工智能，你必须向一家公司付费购买专有模型，而不是自己获取权重或自己构建。

Yeah. Yeah, we can dig in more detail, but I think like, yeah. Six months ago or for the past few years, my kind of working assumption had been that the closed source proprietary models would advance the frontier, and there are a lot of structural reasons for that. But the open frontier would only be a few months behind. And that gap might widen a little bit. I didn't think it would shrink altogether, but I thought it would persist as being relatively small. And I think there are signs now that there seems to be a real risk of near frontier openweight AI falling off altogether. I think Meta, which historically has been the openweight champion in the West, is pulling back, and it seems likely that they're not going to continue with their open source strategy. And then more recently, obviously, the Chinese labs have been the ones driving state-of-the-art open research. And there it seems like there are strong indications that they may also be pulling back from that. And whether it's Qwen or DeepSeek or others, their most high performing models are now keeping proprietary behind an API and just open sourcing, open weighting, you know, smaller, less performant versions. And I think there are real compute incentives behind that. It's just very expensive to service these openweight models with no revenue coming in. And I think there's also geopolitical and competitive considerations. But it's interesting to contemplate what a world might look like where if you want real frontier artificial intelligence, you have to pay a company for a proprietary model as opposed to being able to have access to the weights yourself or build your own.

但我实际上同意这一点。第二个，我还在想另一个变化。我不认为我们看到了开放权重模型的能力或能力趋势线发生重大变化。实际上我认为如果有的话，我们开始趋同，而且我们继续看到这种情况。但我确实认为，围绕构建和发布开放模型的经济决策在过去六个月里肯定发生了变化，我认为这正是 Rob 真正想说的。

But I'd actually agree with that. The second one, this other one I was imagining about as a you know what changed. I don't think that we've seen a major change in the capabilities or the trend line of the capabilities of the open weight models. I actually think if anything we started to converge and we continue to see that. I do think though the economic decision-making around building and releasing open models has definitely changed over the last six months, which I think is what Rob was really getting at.

嗯，我对未来会有多少开放模型更加悲观。看起来在 2025 年，开放模型就像丰饶角一样，只增不减。而我认为我们现在肯定开始看到，开放模型的数量可能已经达到顶峰，接下来会越来越少。因为一旦你获得了信誉，财务激励就不合理了，投入大量资金去做这件事是有意义的。但过了那个点，你就想开始销售模型的托管推理服务。而开源会完全破坏你的业务。所以我认为我们会看到其他中国实验室先发布大型开放模型，但之后一旦他们获得了足够的媒体关注和公关，可能就会关闭。

Um, and I am a lot more bearish on how many open models there will be going forward. It seemed like there was this kind of cornucopia of open models that was only ever growing in 2025. And I think we're now definitely starting to see that we probably hit the peak number of open models and it's now going to kind of get less and less. And because the financial incentives just don't make sense once you've already kind of achieved credibility, it makes sense to invest a lot of money to do that. But after that point, you want to start selling hosted inference of your model. And opening it up just fully undermines your business. So I think we are going to see like other Chinese labs start with like big open models, but then probably close up after that once they've kind of gotten enough press and PR.

开源模型公司有商业模式吗？还是说这纯粹是营销，在你走向前沿的过程中，一旦你接近前沿，就不可避免地想要闭源？

Is there a business model for an open source model company or is it literally just marketing to, you know, as you're on your way to the frontier and then once you're kind of close, it makes it inevitable to want to go close source?

说实话，我不认为有商业模式。你知道，人们尝试过不同的东西，比如像 Red Hat 那样的高端企业模式，但我就是认为在 AI 领域行不通，因为首先要接近前沿就需要巨额的前期投资。我们会看到的，你知道，我们很快可以再回来讨论这个问题，但我很好奇，首先，Reflection 发布模型时会采用什么商业模式。但我仍然怀疑开源 AI 存在好的商业模式。

I don't think there's a business model honestly. There you know different things have been tried, the kind of like premium enterprise like Red Hat model, but I just don't think that it works in AI given the massive upfront investment required to get to the frontier or close to the frontier in the first place. We'll see, you know, we can come back and revisit this soon, but I'm very curious to see if, first of all, if Reflection when Reflection releases a model and what business model they aim to pursue with it. But I remain skeptical that a great business model exists for open source AI.

有趣的是，我正想说，我们在过去一两个月里实际看到的趋势是，终于，我想人们一直认为每个人都会为那些模型能胜任的任务使用更小、更便宜的开源模型，但感觉绝大多数使用量和 token 仍然是在前沿，推动能力，比如弄清楚模型在不同行业能做什么。然后最后我觉得我们现在有了一个趋势，比如，天哪，这些账单太贵了，或者我们的使用量太高了，如果能有一个更便宜、更快、更小的东西用就好了。但有趣的是，这发生在我们看到闭源模型跑在开源模型前面的同时。Ari，我知道你花了很多时间思考这些东西，我们如何看待这些似乎同时发生的抵消力量？

It's funny because I was going to say one of the trends we've actually seen happen in the past like month or two is that finally, forever I think people have been like everyone will use smaller cheaper open source models for tasks that those models can do and it felt like the vast majority of usage and tokens was still just at the frontier pushing capabilities like figuring out what models can do in different industries. And then finally I feel like we have a movement now toward like, geez, these bills are pretty expensive or our usage is pretty high, wouldn't it be nice to have something that is cheaper and faster and smaller to use. But it's funny that it's happening at the same time that we're kind of seeing the closed source models run ahead of open source models. And Ari, I know you spend a lot of time thinking about this stuff, like how do we think about those countervailing forces that seem to be happening at the same time?

是的，我的意思是，首先我肯定看到了很多你提到的前者。非常有趣的是，我们所处的极端算力环境、特别是编码模型能力的提升，再加上从一次发布到另一次发布，模型改变它们的 token 效率、输出 token 效率，导致很多原本满意于使用前沿模型的公司，突然之间，即使只是从 Opus 4.6 到 4.7，token 效率也有很大差异，很多人的账单一夜之间翻倍。我现在开始看到，特别是和很多企业交谈时，他们非常强烈地希望开始削减使用模型的成本。我认为一年前这种需求还没有那么强烈，因为模型的使用规模还没有达到成本足够重要的程度，但现在它们确实达到了那个点，你可以很快消耗预算，因为模型思考很长时间，这现在推动了很多需求，说好吧，我们怎么能让它便宜得多。

Yeah, I mean I think that first off I've definitely seen a lot of the former that you were talking about. Like it's been very interesting the combination of the extreme compute environments that we're operating in, the rise of capabilities of the coding models in particular, and then even just seeing from release to release models changing their token efficiency, their output token efficiencies, has resulted in a lot of companies that were happy using frontier models all of a sudden, even just going from like Opus 4.6 to 4.7 there was a big difference in token efficiency and a lot of people's bills just doubled overnight. And I'm now starting to see, talking to a lot of enterprises in particular, really strong desires to start cutting the cost of using the models. And I think that wasn't there nearly to the same extent a year ago because the models weren't being used at a scale where those costs were meaningful enough, but now they've really reached that point where they are meaningful enough and you can just consume budgets so quickly because the models think for a long time, and that's now driving a lot of demand to say okay how can we make this much cheaper.

我认为你可以用开放模型做很多事情。比如，我认为一个非常显著的事情是，很多人能够通过搭建脚手架，用开放模型复现 Mythos 使用的相同或类似的漏洞发现水平。我认为这是另一个重大变化，模型不再只是模型，而是模型加上 harness 和脚手架，很多创新发生在 harness 和脚手架层。

I think that you can do a lot of that with open models as well. Like I think one very notable thing right is that a lot of people were able to reproduce the same or similar level of vulnerability finding that Mythos used with open models by just putting scaffoldings around. I think that's another one of the big changes right is that a model is not just a model anymore, it's the model combined with the harness and the scaffolding, and a lot of innovation is happening on the harness and scaffolding layer.

我认为这也可能是开源模型拥有商业模式的方式。你开源模型，但不开源脚手架和 harness。然后你有一个 API，人们可以访问整个系统。我认为这可能行得通。Kimi 就在做类似的事情，对吧？比如 Moonshot，我认为通过 API 和聊天界面已经达到几亿美元。他们在某种程度上就是这么做的。所以我认为我们会继续看到强大的开放模型，但可能越来越少。而且我认为这将激励很多人，他们现在必须思考，在一个不会有可靠开放模型的时代，我们如何生存？我认为这推动人们越来越多地开始思考，好吧，我如何确保我有能力构建模型、维护模型、可重复地做到这一点。我认为这是 Nvidia 通过 Reflection 和其他公司大力推动在西方拥有开放模型提供商的一个重要原因。

I think that's also possibly how open source models can have a business model. You open source the model, you don't open source the scaffolding and the harness. And you then have an API where people can access the full system. I think that could potentially work. Kimi's kind of doing that, right? Like Moonshot, I think, is at a couple hundred million through the API and the chat interface. And that's to some extent what they're doing. So I think that we will continue to see strong open models, but probably fewer and fewer. And I think that this is going to motivate a lot of folks where they have to now think, how do we survive in an era where there aren't going to be reliable open models? And I think that pushes people more and more towards actually starting to think about, okay, how do I make sure I have the capability to build a model, to maintain a model, to do that repeatably. I think that's a huge part of the reason that Nvidia has been pushing so hard through Reflection and others, right, to have open model providers in the West.

是的。我的意思是，当然模型提供商越多，对他们越好。而且很有趣。我的意思是，这几乎感觉就像你在 Meta 身上看到的早期表现，对吧？早期，你知道，让一堆不是他们的公司控制这个的想法，对他们来说作为企业永远不会成功。所以就像我们必须在这里拥有一些东西来保持竞争力，即使我们不会成为领先的模型提供商，只是为了我们自己的业务。就像我们想要拥有一些我们可以自己使用的东西。

Yeah. I mean certainly the more model providers the merrier for them. And it's interesting. I mean it kind of almost feels like you saw an early manifestation of this with Meta, right? Where it's like early on, you know, the idea of having this controlled by a bunch of companies that weren't them was kind of like it would never work for them as a business. And so it's like we have to have something here to be competitive even if we're not going to be the leading model provider just for our own business. Like we want to have something that we can use ourselves.

如果你是这些可能受到威胁的公司之一，你基本上有两个选择。选项一是你去和前沿模型提供商合作。你与他们分享你的数据和领域专业知识。他们用这些来和你一起改进模型。也许你能签署一些协议，给你带来一些长期确定性，但最终他们只会超越你，因为你已经放弃了你所有的专有优势。

If you are one of these companies that's potentially threatened by this, you kind of have two options. Like option one is that you go and work with a frontier model provider. You share your data and your domain expertise with them. They use that to improve the model with you. Maybe you're able to sign some agreement that gives you some long-term certainty around that, but eventually they just out compete you because you've given up any of your proprietary advantages.

唯一的选择就是真正在你专注的细分领域竞争，发挥你的独特能力。我认为我们开始在各处看到这种情况，对吧？看看 Cursor 发生了什么，看看其他很多人发生了什么，他们意识到，好吧，我们得远离这个。这里也有一个利润率的角度，对吧？从根本上说，如果你和一个利润率比你高的人竞争，那很难赢。

The only other option is that you try to really compete on the niche that is your focus and where you have unique capabilities. And I think we're starting to see that across the board, right? Look at what's happened with Cursor. Look what happened with lots of other folks where they're realizing, okay, we have to kind of move away from this. There's also a margin perspective from that as well, right? Where just like fundamentally if you're competing against somebody who has better margin than you do, then it's very hard to win.

是的。他们总是会给自己比给你更好的费率，作为上层的应用公司。

Yes. They will always give themselves better rates than they give you, as an app company on top.

Rob，你对过去几个月“应用已完”和“SaaS 末日”的说法有什么看法？

Rob, what was your whole reaction to the 'apps are cooked' and 'SaaS apocalypse' narrative of the past months?

是的，这很有意思。我认为市场叙事总是从一个极端摆到另一个极端，以至于很多细微差别都丢失了。所以我认为有几件事同时成立。我确实认为很多传统软件公司，甚至整个类别，都可能因为前沿实验室及其产品路线图而面临真正的生存风险。所以我认为很多重新评级是理性的。有很多大公司，拥有大量收入和客户群，正因 OpenAI 和 Anthropic 的执行和表现而陷入真正的麻烦。我也认为这肯定反应过度，卖过头了，而且一竿子打翻一船人。在我们的风险投资领域，主流叙事肯定是应用投资太难，要远离。深度科技和硬件等在过去几个月变得更吸引人，成为共识。再说一次，我认为那里有很多真实的东西，有很多深度科技类别我非常兴奋，我们都在投资，我相信我们会谈到。但我也认为，从根本上说，不可能有一两家或三家公司赢得世界上每一个重要的市场和类别。没有公司能有那么大的影响力范围，并在如此广泛的领域出色执行。所以当然有些类别我会押注 OpenAI 和 Anthropic 会占据有利位置，编码显然是第一个，还有其他我们可以讨论的。我认为这些横向跨领域的领域才是实验室最应该关注和追求的，而不是像宠物店那样的垂直软件。

Yeah, that's interesting. I think market narratives have a way of swinging so far in one direction or the other that a lot of the nuance is lost. So I do think a couple of things are true simultaneously. I do think a lot of traditional software companies and even categories potentially are at real existential risk on account of the frontier labs and their product roadmaps. So I think a lot of that rerating was rational. There are a lot of big companies that have big revenues and big customer bases that are in real trouble based on how OpenAI and Anthropic are executing and performing. I also think it for sure was an overreaction and an oversell and painted with too broad of a brush. And I think in our field of venture capital, it certainly is the prevailing narrative that apps are so challenging to invest in, stay away from them. Deep tech and hardware and so forth have all become more appealing and consensus over the past few months. And again, I think there's a lot that's real there, and there are a lot of deep tech categories that I'm super excited about that we're both investing in, that I'm sure we'll talk about. But I also think there are still, fundamentally, no way that one or two or three companies will win every single important market and important category in the world. No company can have that span of influence and execute excellently so broadly. So for sure there are some categories where I would bet on OpenAI and Anthropic being very well positioned, in coding obviously is the first, and there are others which we can talk through. And I think these sort of horizontal cross-cutting areas are the ones that make the most sense for the labs to focus on and go after, versus vertical software for pet stores or something like this.

我确实想知道，风投是否只是把自己逼进了一个进退两难的角落。如果你只打算投资宠物店，因为那对模型来说足够小，我不知道那最终是否会成为一个超级有趣的最终类别，对吧？但事实证明，无论你在 AI 世界哪里看，都有无法避免的生存挑战。也许你想投资一个数据中心建设商，你会说他们就会建数据中心，会有需求，但当然你仍然需要弄清楚，是什么让你成为比另外 10 个用激进融资竞争同一块地的人更好的数据中心建设商。毫无疑问，我们正处于一个极度不确定和动荡的时期。是的，我完全同意你关于硬科技的观点。那里再说一次，我确实相信在硬件和深度科技类别中有很多真正令人兴奋的机会，现在有很多有趣的创新正在发生。

I do wonder if VCs are just pushing themselves into a corner of damned if you do, damned if you don't. If you're only going to go invest in pet stores because that's small enough for the models, I don't know if that actually ends up being a super interesting end category, right? But it turns out no matter anywhere you look in AI world there are existential challenges you can't avoid them. Maybe if you want to go invest in a data center builder and you're like they will just build data centers and there will be demand, but sure then you still have to figure out what makes you a better data center builder than the other 10 people that will compete with aggressive financing for the same site. There's no question that we are at a period of max uncertainty and volatility. And yeah, I totally agree with you on the hard tech point. There again, I do believe there are a lot of really exciting opportunities in categories in hardware and deep tech where there's a lot of interesting innovation happening right now.

是的，你说得对，事实证明硬科技也非常非常难，失败率要高得多，还有很多未解决的问题，未来几年在这条路上会有很多痛苦和损失。然后在应用方面，我认为正如你所说，最后一英里仍然有巨大的价值，实验室也开始尝试装备起来，建立这些部署代码，并在实施方面下功夫。但我认为会有很多领域和机会，让交付应用的公司也能蓬勃发展并持续增长。

Yeah, to your point, it turns out hard tech is also very, very hard and the failure rates are much higher and there's a lot of unsolved problems and there's going to be a lot of pain and lost money down that road in the years to come. And then on the application side, I think as you said, there's still tremendous value in that last mile and the labs are also starting to try to tool up and stand up these deploy codes and lean in on the implementation side. But I think there will be so many pockets and opportunities for companies that are delivering applications to also thrive and continue to grow.

我也觉得有趣的是，最终想想为什么初创公司总能赢，对吧？为什么谷歌没有把所有事情都做好？一家大公司很难把很多不同的事情都做好。从根本上说，一个反驳论点是 OpenAI 关闭或暂停了他们的视频项目，对吧？这让我很惊讶，我要说，考虑到他们能获得几乎无限的资本，几乎无限的人才，而且他们有一个很棒的团队在领导那个项目。但他们不得不在这里做出艰难的选择。现在其中一些，很多可能从根本上是由算力限制驱动的。视频训练相对于文本训练来说非常昂贵，等等。但这表明他们不能做所有事情。

I think also it's interesting to think about ultimately why do startups ever win, right? Why didn't Google do everything right? It's hard for a massive company to do many different things well. Fundamentally, a counterargument to this is actually OpenAI shuttering or suspending their video efforts, right? That was very surprising to me, I will say, given their access to effectively infinite capital, given the effectively infinite talent, and they had a great team there that was leading that. But they had to make the hard choices here. Now some of that, a lot of that is likely driven by compute constraints fundamentally. Video training is very expensive relative to text training and so on. But it goes to show that they can't do everything.

他们需要专注，精简，只专注于收购 TBPN 并将其加入公司。

They needed to focus and strip down and just focus on acquiring TBPN and adding that to the company.

我实际上听说他们听了我们上一集，读了你的文章，然后他们说，“哦，Rob 预测 Sam 会出局。这真的，你知道，那个辛辣的预测已经触动了我们的心。我们真的需要确保我们避免这种情况。”今年六个月过去了，你对那个预测感觉如何？我想你还有六个月的时间来证明是对的。

I heard actually it was that they had listened to our last episode and read your article and they were like, 'Oh, Rob's predicting Sam's out. This is really, you know, that spicy prediction has gotten to our hearts. We really need to make sure we avoid that.' How are you feeling about that one six months into the year? I guess you got six months still to be right.

今天，六月，你知道，六月中旬，看起来比我当时和你们分享时可能性大得多。这很有趣。是的，我的意思是，当我在十二月公布那个预测时，我觉得每个人都像是，你在说什么？这毫无意义。包括你们。你们就像，哦，这很有趣，但似乎不太可能。是的，我的意思是，不用说，今年对 OpenAI 的氛围已经转变了，而且有了这种认识，而去年这个时候是，哦，OpenAI 在做所有事情。他们是如此了不起的公司。他们会在芯片、数据中心、机器人等所有方面获胜。现在，我认为现在有一种认识，哦，我们确实需要专注。而且我认为 Sam 的领导有很多方面越来越受到质疑。

It's looking a lot more likely today on June, you know, mid-June than it was when I shared it with you guys. It's been interesting. Yeah, I mean when I unveiled that prediction in December, I feel like everyone was like, what are you talking about? That makes no sense. Including you guys. You were like, oh that's interesting, but that doesn't seem very likely. And yeah, I mean, needless to say, the vibes have shifted against OpenAI this year and there's been this realization whereas this time last year was like, oh, OpenAI is doing everything. They're such an amazing company. They're going to win in chips and data centers and robots and everything. Now, I think now there's a realization like, oh, we really do need to focus. And I think there are a lot of elements of Sam's leadership that are increasingly under question.

而且我认为，埃隆·马斯克的官司，即使埃隆·马斯克输了，也损害了山姆的声誉和可信度，等等等等。所以，谁知道呢？这是年初一个挑衅性的、故意低概率的预测。我敢说概率肯定上升了。有一个变化是，当我在《福布斯》上发表那个预测时，我的假设是 Fiji 是最明显的继任者，而且正在被培养成下一任 CEO。显然她现在因为健康问题不得不退居二线。所以我认为继任者的问题已经不同了。我实际上认为，我听到的关于布雷特·泰勒的传闻相当有道理。对于不熟悉的人来说，布雷特是 OpenAI 的董事会主席，Sierra 的 CEO，是硅谷最受尊敬和敬仰的领导者之一。我认为 OpenAI 收购 Sierra 并让布雷特担任 CEO 真的非常合理。我认为这符合 OpenAI 股东的最佳利益，说实话。而且在围绕 IPO 的舆论氛围讨论中，风向已经明显从 OpenAI 转向了 Anthropic。这是一个决定性的变化，我认为可能对 OpenAI 产生颠覆性影响，因为人们就是信任布雷特、尊重他、钦佩他，认为他是令人敬仰的领导者。我认为如果这样的人掌舵 OpenAI，会大大改变他们的命运。所以，走着瞧吧。还有好几个月，但我不知道。我觉得那个预测可能真的会成真。我们拭目以待。

And I think that the Elon Musk trial, even though Elon Musk lost, was damaging to Sam's reputation and his trustworthiness, and so on and so forth. So, you know, who knows? It was a provocative and purposely somewhat low-likelihood prediction at the beginning of the year. I would say odds have gone up for sure. One thing that's changed is that at the time I made and published that in Forbes, my hypothesis was that Fiji was the obvious most likely successor and was being groomed to be the next CEO. And she obviously has now had to take a step back given some health concerns. So I think the question around who the successor would be is different. I actually think this theory that I've heard making the rounds around Brett Taylor is quite plausible. For folks who aren't familiar, Brett is the chairman of the board at OpenAI, the CEO of Sierra, one of the most respected and revered leaders in Silicon Valley. I think it would honestly make so much sense for OpenAI to acquire Sierra and make Brett the CEO. I think it would be in the best interest of OpenAI's shareholders, honestly. And in this discussion around the vibes, they have really shifted against OpenAI and towards Anthropic in the lead-up to the IPO. That is a decisive change that I think could be a total game changer for OpenAI, because people just trust Brett and respect him and admire him and think he's an admirable leader. I think if someone like that was at the helm of OpenAI, it would do a lot to change their fortune. So anyway, we'll see. There are many months to go, but I don't know. I feel like that one might actually come to fruition. We'll see.

我也觉得，OpenAI 转向类似 Alphabet 的结构这个想法，现在总体上看起来更合理了，因为他们已经转向控股公司。也许山姆继续担任控股公司的 CEO，然后有人接管 OpenAI 或 ChatGPT，也许 ChatGPT 成为独立产品，或者类似这样。那也说得通。

I think also this notion of OpenAI going to an Alphabet-like structure seems a lot more plausible in general now that they've shifted to a holding company. Maybe Sam stays CEO of the holding company, and then you have somebody take over OpenAI or ChatGPT, maybe ChatGPT becomes its own product, or something to that effect. That would make sense as well.

是的。不，我是说你提到了，但显然我觉得过去几个月的主流氛围就是这样。看到 Anthropic 处于这种前所未有的高光时刻，真是难以置信，对吧？我觉得他们完全抢走了风头。可能是我们对一家公司最一致的共识了。显然这在很多方面是以 OpenAI 为代价的。这周随着对 Fable 发布的一些反应，你可能会开始看到一些裂痕。但我对你们两位都很好奇：你们认为这种氛围趋势会持续下去，还是像我们文化中的许多事情一样，我们不可避免地会在未来三到六个月里看到这两家公司遭遇某种反弹或命运翻转？

Yeah. No, I mean you mentioned it, but obviously I feel like these past months the dominant vibe. It's unbelievable to see Anthropic on this unprecedented vibe run, right? I think they've completely sucked up the oxygen. Probably the most consensus around one company we've had. And obviously that's come at the expense of OpenAI in many ways. You're maybe starting to see some light cracks in that this week with some of the reaction to the release of Fable. But I'm curious for both of you: do you think this is a vibes trend that continues, or like many things in our culture, are we inevitably going to have some sort of backlash or flip in fortune of these two companies in the next three to six months?

肯定会有一定程度的反弹。谁赢就会有人反弹，对吧？这是生活的真理。话虽如此，我认为 Anthropic 也可能开始采取一些会疏远人们的举动。比如明显限制 Fable 用于任何与 AI 开发相关的事情。而且我认为，人们核心上并不是对限制本身感到非常不满。而是这是一个无声的限制。是的。我认为人们真正不满的是这一点，对吧？它不会拒绝你。它不会说“我不会帮你做这个”。它只是在你不知情的情况下做得不好。他们可以说这是出于安全考虑。这很牵强。我认为很明显这是竞争定位的举动，而不是安全举动。而且我认为，正如我之前提到的，很多人能够用开放模型和好的工具找到 Methus 发现的许多零日漏洞。所以我也认为这并不一定有独特的安全风险。但我看到更多一直对 Anthropic 极度看好、持积极态度的人，因为过去一天 Fable 的事情而真正愤怒，这是我从未见过的。这似乎确实是一个阶段性的变化。所以我认为如果他们继续采取这样的举动，我们肯定会看到越来越多反对 Anthropic 的转变。

There'll definitely be some amount of backlash. There's always backlash to whoever's winning, right? That's just a truism of life. That said, I think Anthropic is also potentially starting to make moves that will alienate people. Like clearly with limiting the use of Fable for anything to do with AI development. And I think in particular, I don't think people are incredibly upset at the core with just a limitation. It's the fact that it's a silent limitation. Yeah. That I think people are really upset about, right? That it doesn't give you a refusal. It doesn't say, "I'm not going to help you with this." It just does a poor job on that without you knowing. They can say that's because of safety. That's tenuous. I would say it seems pretty clear that's a competitive positioning move rather than a safety move. And I think to that end, like I mentioned earlier, a lot of people were able to find many of the same vulnerabilities in zero days that Methus found with open models and good harnesses. So I also think there's not necessarily a unique safety risk to this. But I've seen more people who have been incredibly bullish on Anthropic and positive on Anthropic truly pissed off as a result of just the Fable of the last day than I'd ever seen. And it does seem like a bit of a step change. So I think if they continue to make moves like that, we will definitely see more and more of a shift against Anthropic.

我的意思是，显然很多开源模型都是通过蒸馏 Anthropic 模型来训练的。你认为这会对开源社区产生影响吗？

I mean obviously so many open source models have been trained on just like distilling Anthropic models. Do you think this will have an impact on the open source community?

这取决于他们是否提起诉讼，对吧？在中国，这不会限制任何人。对于美国模型，我认为人们往往非常谨慎。比如我们为 RC 的流行大模型做了所有数据整理，我们非常注意在开发过程中任何时刻都不使用任何闭源 API。也只使用公开数据。所以没有那些你也可以构建强大的模型。但我也认为，看看他们刚刚发布的 MAI 模型，他们特别强调完全不使用任何合成数据，以避免任何偷偷从另一个模型蒸馏的能力。然后我认为他们说过，现在他们将从那些模型开始使用合成数据来引导它们。但我不知道这是否会从根本上改变。只要人们能通过 API 访问你的模型，你就无法真正阻止他们尝试进行一定程度的蒸馏。话虽如此，我认为蒸馏这种说法在某种程度上被夸大了。这是真的，但你仍然可以构建伟大的模型。你不需要蒸馏来构建伟大的模型。我认为那种认为任何开放模型赶上封闭模型的唯一方法就是有效地蒸馏或窃取它的想法，对我来说有点像自我安慰。

Depends if they litigate it. Right? It's not going to do anything in China to limit anybody there. And then for US models, I think people tend to be very careful. Like we did all the data curation for RC's trendy large model, and we were very cognizant to not use any closed source APIs at any point in any of the development there. Also only public data. So you can build a really powerful model without that. But I think also looking at the MAI model that they just released, they made a huge point of not going so far as actually not using any synthetic data at all to avoid any ability to kind of sneakily be distilling from another model. And then I think they've said they now they're going to start using synthetic data from those models to kind of bootstrap them. But I don't know if it's going to fundamentally change it. As long as people can actually use an API to get to your models, you can't actually stop them from trying to do some amount of distillation. That all said, I think this distillation claim is overblown to some extent. It's true, but you can still build great models. You don't need to distill to build a great model. I think the notion that the only way any open model can catch up with a closed model is by effectively distilling or stealing from it reads a little bit like copium to me.

你知道，我还能记得一段时间以前的事。我不确定你玩过它到什么程度，但你对这究竟是多大的能力跃迁有什么初步看法？

You know, I can remember in a while. I'm not sure to the extent you've played around with it, but what are your early reads on how much of a step change in capabilities this really is?

我昨晚才稍微玩了一下，因为 Fable 刚发布。我个人没看出和 48 有巨大差别。跟人聊下来，大家的看法也相当不一。老实说很难判断。我对 Fable 的看法是，我觉得它既是一件大事，也不是什么大事。我认为它是最新的前沿模型，恰好昨天发布，也就是我们录制的前一天。但如果我们三周或七周后再录，可能就会有另一家供应商的另一个模型成为我们讨论的话题。我确实认为，基于基准测试和定量数据（虽然远非完美），它相对于之前的前沿模型确实是一个有意义的跃迁式改进。我认为这有意义，不是因为它是什么不连续性——我觉得这种渐进式改进会继续下去——而是因为它确实削弱了那种人们已经不再相信的说法，那种一年前还很有市场的说法，即预训练真的撞墙了，事情在趋于平稳，我们只能靠强化学习和测试时算力来推动前进，因为我们撞上了数据墙。我只是觉得这显然不是真的。收益还在源源不断地到来，而且我认为没有任何充分理由认为它们会很快趋于平稳。

I only played with it a little bit last night since Fable released. I didn't personally see massive differences from where 48 was. And talking to people, it seems like the takes have been quite varied around that. It's pretty hard to tell honestly. My take on Fable is I kind of feel like it is and also is not that big of a deal. I think it's the latest state-of-the-art model that was released, which happened to be released yesterday, a day before we're recording this. But if we recorded this three weeks from now or seven weeks from now, there would probably be another model from another provider that we would be talking about. I do think, based on the benchmarks and the quantitative data, which is far from perfect, it does seem like it is a meaningful step change improvement relative to the previous state-of-the-art. I think that's meaningful, not because it's some discontinuity—I think this kind of gradual improvement will continue going forward—but I think it's meaningful in the sense that it really does undermine this narrative that people have already shifted away from, which had a lot of currency a year or so ago, that pre-training is really hitting a wall, things are plateauing, and we have to rely on RL and test-time compute to carry us forward because we've hit this data wall. I just think that's clearly not true. The gains are continuing to come in very richly, and I don't think there's any good reason to think they will plateau anytime soon.

是的，我觉得我们在过去六到九个月里也看到了很多对这种说法的反驳，很明显预训练并没有撞上巨大的墙。我们只是在等新芯片。我觉得其中一部分也是因为那种天真的 Scaling（规模扩张）行不通。如果你只是把我们在做的完全相同的东西乘以 10 倍或 100 倍——这差不多就是 4 或 5 的目标，或者当时的 Llama 35B——那效果并不好。但我认为人们从中得出了过于笼统的结论，尤其是因为深度学习真正困难的一点，我发现，是你往往需要把所有细节都做对，事情才能真正奏效。如果你做对了大概 95%，它往往就会变成不工作。很多时候，有几种方法是稳健的，但你可能几乎什么都做对了，却没有任何真正的改进，然后你调整最后一个旋钮，突然之间就得到了跃迁。这在深度学习里经常发生。这带来的挑战之一是，它让解读负面结果变得根本性地困难。好吧，你尝试把所有东西都 Scaling 起来，但没成功。那是因为 Scaling 不起作用，还是因为你只是做错了一件事？这经常发生。这也让你很难决定什么时候放弃一个项目，因为你很容易总是想，也许我再调整一下，这东西就能工作了，而很多时候并非如此。

Yeah, I think we also saw a lot of pushback to that narrative over the last six to nine months, and it's pretty clear that pre-training did not hit a massive wall. We were just waiting for new chips. I think some of that also is just that naive scaling doesn't work. If you just take exactly the same stuff that we were doing and multiply it by a factor of 10 or 100—which was kind of what 4 or 5 was aiming at, or Llama 35B at the time—that didn't work extremely well. But I think people took a much overgeneralized take from that, especially because one of the things that's actually really challenging about deep learning, I have found, is that you really need to get all the details right often for something to really work. If you have kind of 95% of it right, it kind of rectifies to just not working. A lot of the time, there are a couple methods that are robust, but you can be doing almost everything right and get no real improvement, and then you tweak the last knob and now all of a sudden you get a step change. That happens a lot in deep learning. One of the things that's challenging about that is it just makes it fundamentally difficult to interpret a negative result. Okay, you tried scaling everything up and it didn't work. Is that because scaling doesn't work, or is that because you just did one thing wrong? That often happens. It also makes it very hard to decide when to abandon a project, because it's easy to always be like, well, maybe if I just make one more tweak this thing will work, when often times that's not the case.

我的意思是，我想谈谈这个领域的其他玩家。我觉得 Rob 在我们十二月的节目里说过，我认为我们所有人都有这种感觉，就是 Google 的位置好得不得了。显然他们有人才，有算力。但感觉过去六个月 Google 的事情进展得并不太顺利。你可以随意反驳，但那边到底发生了什么？为什么他们没能追上编码能力，我想，尤其是这一点？

I mean, I guess to hit the other players in the space. I feel like Rob in our December episode, you said, which I think is a feeling all of us have, that Google is incredibly well positioned. Obviously they've got talent, they've got compute. It doesn't feel like things have gone super well for Google in the last six months. Feel free to push back, but what's going on over there? And why haven't they been able to catch up on coding, I guess, most notably?

是的，我觉得我有点不同意 Google 落后了或者执行得不好。我认为这在一定程度上反映了这样一个现实：三家实验室都在不断互相超越，任何一个月，其中任何一家都可能拥有所谓的“前沿模型”。

Yeah, I think I would disagree a bit that Google has fallen behind or isn't executing as well. I think it in part goes to this reality that the three labs are all kind of in this process of leapfrogging one another continually, and any given month any one of them may have the quote unquote state-of-the-art model.

不过 Google 已经有一阵子了。

It's been a while for Google though.

我认为毫无疑问他们在编码方面落后了，而且我认为这恰恰反映了优先级问题。很明显 Anthropic 多年来一直把编码作为他们的北极星，这显然被证明是一个极其明智的举动。这真的让他们一飞冲天。而 OpenAI 最近更是加倍、三倍、四倍地投入其中，你会看到很多对 Codex 的正面评价。我认为这只是对 Google 来说优先级没那么高。但之前我们谈到的所有事情，我仍然坚持，并且非常有信心，Google 拥有极其深厚的人才储备。不像 OpenAI 和 Anthropic，Google 有这台巨大的印钞机。最后是算力。我认为 Google 从完全全栈中获益良多。他们自己设计芯片。他们有自己的云。没有人能获得无限的算力，Google 也当然希望有更多，但他们有巨大的……

I think there's no question that they're behind on coding, and I think that just reflects prioritization. It's clear that Anthropic leaned in on that as their north star for years, and that proved to be obviously an incredibly savvy move. That has really catapulted them. And OpenAI more recently has really doubled, tripled, quadrupled down on it, and you're seeing a lot of positive love for Codex. I think it just hasn't been as much of a priority for Google. But everything we talked about previously, I still stand by and feel very confident in the sense that Google has an incredibly deep bench of talent. Unlike OpenAI and Anthropic, Google has this massive cache machine. And then lastly is compute. I think Google has benefited so much from being totally full stack. They design their own chips. They have their own cloud. No one has access to infinite compute, and Google would also love to have more, but they have a massive...

他们希望有更多吗？我不知道。他们卖掉了一部分给 Anthropic。

Would they love to have more? I don't know. They sold some of it off to Anthropic.

是的，我的意思是，他们有外部云业务，但我认为他们在算力方面的位置比 Anthropic 或 OpenAI 都要好。所以，是的，我继续非常看好 Google。

Yeah, I mean, they have an external cloud business, but I think they're better positioned when it comes to compute than either Anthropic or OpenAI. So, yeah, I continue to be very bullish on Google.

你怎么看，Ari？

What do you think, Ari?

我倾向于持同样的看法。我有点惊讶他们自 31 以来没有进步那么多。我原本预期 IO 大会会有更大的发布。我怀疑，你知道，几个月后我们会不会发现原本有计划，但出了什么问题，然后被取消了，或者诸如此类。除了 Flash 模型之外，我确实认为他们拥有所有结构性优势。我认为能接触到印钞机——我觉得这是思考 xAI 和 SpaceX 合并的一个非常有趣的角度——这是让 xAI 搭上一台印钞机的方式，就 SpaceX 是一台印钞机而言，当然它比 Google 或 Meta 的印钞能力要弱。所以我认为他们在那里位置很好。

I tend to share the same view. I am a little surprised that they haven't improved as much since 31. I would have expected a bigger launch at IO. And I wonder if, you know, in a couple months we'll find out there was one planned and then something went wrong and it was scuttled, or whatever the case may be. Beyond just the flash model, I do think they have all the structural advantages. I think having access to the money printing machine—I think that's a really interesting way to think about the xAI-SpaceX merger—it's a way for xAI to get attached to a money printer, to the extent that SpaceX is a money printer, which is less so certainly than Google or Meta. So I think they are well positioned there.

我还觉得，我不知道谷歌的市场和 Anthropic 瞄准的是否一样，就重点而言。对吧？从根本上说，面向消费者的模型会被商品化。我觉得这一点非常清楚。对于消费者用例，比如只是问模型一个关于世界的问题，或者让它当辅导老师，或者任何大多数消费者会用的标准功能，这些都会被商品化。人们会在手机上使用模型。我觉得这看起来非常明确。

I think also, I don't know if Google's market is the same as what Anthropic is going after with respect to the focus. Right? Fundamentally, models are going to be commoditized for consumers. I think that's just quite clear. For the consumer use case of just asking my model a question about the world, or having it be a tutor, or any of the standard things that most consumers are going to use, that's going to be commoditized. People are going to use the model on their phone. I think that just seems very clear.

那计算机使用这类东西呢？感觉在消费者使用方面，模型还做不到的领域仍然存在。

What about some of the computer use stuff? It feels like there are still frontiers of consumer use that models can't do yet.

我觉得肯定会有，而且你会看到更多这样的东西。但首先，大多数人不会使用那个级别的软件。我觉得重度用户会，但大多数人只会把它当作答案引擎。而且谷歌现在在 Android 和 iOS 手机上都已经优化得很好，可以成为默认提供商，直到最终苹果在那里构建自己的模型——这一直是我大力推动的。所以我认为谷歌实际上会处于强势地位，即使他们不一定拥有最好的模型。我觉得这是另一个重要部分：我不知道最好的模型是否一定会在消费者领域获胜。

I think there will certainly be, and I think you will see more of that. But I think first off, most people will not use that level of software. I think the power users will, but most people are just going to be using it as an answer engine. And Google's now quite well optimized to be the default provider both on Android phones and iOS phones, until eventually Apple builds its own models there, which I've been pushing a lot. So I think Google's actually going to be in a strong position even if they don't necessarily have the best model. I think that's another big part here: I don't know if the best model necessarily wins in the consumer space.

那即使在编程领域呢？说实话，这对我来说是最令人惊讶的。我确实认为 Codex 显然是个很棒的产品，但它似乎并没有产生巨大的影响。我的意思是，它显然在增长，但 Claude Code 仍然是占主导地位的编程工具。我想知道，看到这种先发优势很有趣：你向人们介绍了这种范式，而我总是认为开发者工具是最精英主义的——他们会立刻切换，最好的模型总是赢。我不知道你们怎么看这件事。

Well, what about even in the coding space? This is honestly been most surprising to me. I actually think Codex is clearly an amazing product, but it doesn't seem to be making a huge dent. I mean, obviously it's growing, but Claude Code remains the dominant coding tool. And I wonder, it's interesting to see this first mover advantage: you introduce people to the paradigm, and I always thought of developer tools as the most meritocratic—they would switch in a second, best model always wins. I don't know what you guys make of that.

我认为这大体上仍然成立。我确实认为开发者在某些方面是最善变的消费者，他们会转向任何最好的东西。话虽如此，问题是它好多少。我认为 Claude 和 Codex 一直保持足够接近，而且没有特别令人信服的理由让你必须切换到 Codex，所以很多人一直坚持用 Claude。我预计，至少在 AI 开发者中，你会看到现在大量从 Claude 转向 Codex 的情况，考虑到他们限制 Fable 等方式。我认为老实说，这对 OpenAI 来说是一份相当不错的礼物，因为很多在 Twitter 上声音最大的人，在不久的将来很可能会花更多时间在 Codex 上。

I think that's generally still true. I do think that developers are the most mercurial consumers in some ways, and they are going to switch to whatever the best thing is. That said, there's a question of how much better it is. I think Claude and Codex have stayed close enough, and there hasn't been a massively compelling reason why you need to switch to Codex, such that a lot of people have stuck with Claude. I expect that at least amongst AI developers, you're going to see a massive set of shifts now to Codex from Claude, given the way that they're limiting Fable and whatnot. I think honestly this is a pretty nice gift for OpenAI with respect to a lot of the people who are the loudest voices on Twitter and whatnot becoming likely to spend a lot more time with Codex in the immediate future.

但我的意思是，当 5.6 发布时，OpenAI 大概必须采取反制措施，他们是否必须给人们更多访问权限？他们最终会不会为了口碑和使用量而承受更差的利润率？或者结构上什么能让他们为一个水平相当的模型提供更好的访问？

But I mean, is OpenAI when 5.6 comes out and OpenAI presumably has to counter-position, is it going to have to give people more access? Are they just ultimately going to eat way worse margins for vibes and for usage? Or what structurally is going to allow them to provide better access for a model that's on par?

这是个好问题。我的意思是，归根结底，就像在任何有少数玩家的市场里，只要模型足够接近，那就能产生很大影响。也可能是算力访问的问题，对吧？Anthropic 刚刚在这里获得了大量访问权限。但我认为这又回到了之前关于未来开放模型可能会减少的讨论。我认为其中一部分实际上也可能是未来封闭模型的访问权限会减少。不难想象一个世界，Anthropic 受到如此严重的算力限制，以至于他们实际上切断了 API。因为显然，就他们赚多少钱而言，他们会更倾向于 Claude Code 而不是 API。而且你现在开始看到 OpenAI 开始销售期货，比如“嘿，你未来能保证获得推理 token 的访问权限”。这实际上对任何在这些模型之上构建东西的人来说都是一个巨大的生存威胁。我认为六个月前这还不是一件真正可能的事情，但现在感觉 API 消失是非常可能的——不是作为商业决策，而纯粹是因为算力限制。

It's a good question. I mean, ultimately, it's like in any market where you have a couple of players, so long as the models are close enough, that can make a big difference. It could also be compute access, right? Anthropic just got a lot of access here. But I think that's another one of the things circling back to earlier conversation around how there might be fewer open models going forward. I think part of that could also actually be that there could be fewer closed model access going forward. It is not hard to imagine a world in which Anthropic is so compute constrained that they actually cut off the API. Because obviously they're going to prefer Claude Code to the API with respect to how much money they make. And you start to see this now with OpenAI starting to sell futures of like, hey, you get guaranteed access to inference tokens going forward. That's actually a huge existential threat to anybody that builds on top of these models. And I think that was not really a plausible thing six months ago, but now feels very plausible that actually the APIs go away—not as a business decision, but just purely because of compute constraints.

是的。你怎么看，Rob？

Yeah. And what do you think of that, Rob?

是的，我认为这完全可行。我认为可能会发生几件不同的事情。我认为完全切断 API 访问将是一个极端的举动。但你可以想象，例如，OpenAI 或 Anthropic，与其只是不开放他们的模型，而只通过 API 提供，你可以想象他们实际上甚至不向任何人公开提供他们最强大的模型，而是保留供内部使用。而且我认为这在很大程度上取决于算力瓶颈以及这种严重程度会持续多久。我觉得这值得反思。这涉及到一些越来越相关的更深层次的技术话题。在眼前有一些努力和初创公司试图彻底打开半导体供应链，在美国建立尖端晶圆厂，挑战 ASML，挑战台积电等等。埃隆·马斯克显然有这个 Terapab 概念，这很迷人，老实说我觉得人们应该更多地谈论它，因为如果他成功了，那将是多么具有变革性。所以我认为很多这些变量都会影响算力作为限制因素的时间有多长。而这反过来，我认为，如果算力限制仍然存在，Ari 描绘的某些可能性，你完全可以想象它们会成真。但很难想象我们如何真正解除限制，如果我们甚至勉强继续沿着这条轨迹前进，很难想象我们如何在未来几年内缓解算力限制。

Yeah, I think it's totally feasible. I think a few different things could happen. I think cutting off their API access altogether would be an extreme move. But you can imagine, for instance, OpenAI or Anthropic, rather than just not open sourcing their models but only making them available via API, you could imagine them actually not even making their most powerful models available to anyone publicly and reserving them for internal use. And so much of this depends, I think, on the compute bottleneck and how long it stays this acute. Which I think is interesting to reflect on. This gets to some of these deeper tech topics which I think are becoming increasingly relevant. There are on the horizon efforts and startups that are trying to break the semiconductor supply chain wide open, build cutting edge fabs in the US, challenge ASML, challenge TSMC, etc. Elon Musk obviously has this Terapab concept which is fascinating and honestly I feel like people should be talking about more, just given how transformative it would be if he pulls it off. So I think a lot of those variables will influence how long compute is the limiting factor. And that in turn, I think, if it remains this compute constraint, some of these possibilities that Ari is sketching out, you could totally imagine them being real. It's hard to imagine though how we actually unblock, if we continue even remotely on this trajectory, it's hard to imagine how we relieve the compute constraint within the next handful of years.

我认为这不是两三年内的事，台积电不会被取代，而是会被许多其他能提供类似芯片的厂商所补充。但你可以想象五年之后，世界上不一定只有一家公司能制造尖端芯片。实际上，现在的市场结构有点疯狂——只有一家公司知道怎么做，其他人都做不了，而且制造过程中最重要的机器也只有另一家公司能做，其他人都做不了。事情不必非得这样，我认为也不会永远这样。

I think it's not a two-to-three-year thing where TSMC is displaced, but augmented by many other players that can provide comparable chips. But you can imagine over five years, it doesn't have to be the case that there's only one company in the world that can make cutting-edge chips. It's actually kind of crazy that that's the current market structure—that there's one company that knows how to do this and no one else can, and the most important machine that goes into the process is made by one other company and no one else can do it. It doesn't have to be that way, and I don't think it will be that way forever.

我们生活在众多可能世界中的一个，这真是疯狂。确实很离谱。

What a wild version of the many worlds that we could live in, that we live in now. It is pretty nuts.

有没有明显的初创公司在挑战 ASML？有很多在挑战台积电，但我没看到那么多针对 ASML 的。

Are there clear upstarts going after ASML? There are a lot going after TSMC, but I haven't seen as many going against ASML.

有。这是一个有趣的新研究领域。简而言之，ASML 的重点显然是极紫外光刻（EUV），而 EUV 在能在芯片上打印多小的晶体管方面开始触及物理极限。所以有几个非常有趣的新研究方向。一个是放弃使用光，完全改用物质，这叫原子光刻，用原子束在芯片上打印特征，这样可以获得更低的解析度。有几家初创公司在原子光刻方面做非常有趣的工作。还有一批初创公司希望超越 EUV，在电磁波谱上走得更远，用波长更短的 X 射线。所以 X 射线光刻这个整体概念正在获得很大势头。每个方向都有几家初创公司筹集了大量资金并全力投入。说白了，它们仍处于开发阶段，这两种方案能否被证明在商业上可行还有待观察。但如果它们能成功，我认为尤其是原子光刻，在机器更简单、零件更少、更便宜、更小、解析度更好等方面有很多优势。所以我认为这里确实可能出现真正的技术颠覆。

There are. It's an interesting new area of research. In a nutshell, ASML's focus is obviously extreme ultraviolet lithography (EUV), and EUV is starting to hit physical limits in terms of how small transistors it can print on chips. So there are a couple of really interesting new research directions. One is rather than using light—moving away from light altogether and instead using matter—it's called atom lithography, where you use a beam of atoms to print features on chips, which lets you get much lower resolution. There are a couple of startups doing really interesting work in atom lithography. Then there's also a handful of startups looking to leapfrog EUV and move even further out on the electromagnetic spectrum to something with even shorter wavelengths, which is X-rays. So this whole concept of X-ray lithography is getting a lot of momentum. There are a couple of startups in each of these buckets that have raised a ton of money and are running at this. To state the obvious, they're still very much in development mode, and it remains to be seen whether either of these will prove commercially viable. But if they work, I think especially atom lithography has so many advantages in terms of the machine being way simpler, with way fewer parts, way cheaper, way smaller, and obviously much better resolution. So I do think there may be real technology disruption coming here.

看起来这至少是五年后的事，可能，考虑到目前的进展。

It looks like that's at least five years away, probably, given where it is currently.

我觉得这个话题很有意思，值得深挖——一两年后，当我们比今天更受算力约束时，实际影响会是什么？显然，感觉现在有很多对强大模型的过度使用，超出了人们今天的实际需求，就像开法拉利去街角杂货店。但即使我们解决了所有这些问题，把人们完美地引导到他们适合的模型上，或者无论世界最终如何运作，感觉我们仍会处于这种整体算力短缺中。这很迷人。我的意思是，一个明显的影响是实验室本身优先考虑第一方产品或优先考虑自己的开发。还有其他影响吗？我只是临时想到这些，但我很好奇你们两位有没有想到其他影响，比如对竞争行业中有或没有算力获取渠道的企业意味着什么。

I think this thread is fascinating to pull on—what are the actual implications of a year or two from now, where we're even more compute-constrained than we are today? Obviously, it feels like there's lots of overuse of powerful models for what people need today, like using a Ferrari to go down the street to the grocery store. But even if we figure all that stuff out and route people perfectly to their appropriately built models, or however the world ends up working, it feels like we'll still be in this overall compute shortage. It's fascinating. I mean, one implication is obviously the labs themselves prioritizing first-party products or prioritizing their own development. Any other implications? I'm just thinking on the spot here, but I'm curious if there are other implications that come to mind for either of you, of what that might mean for businesses in competitive industries that get access or don't get access.

这会推动效率提升，对吧？一般来说，尤其是前沿实验室，因为所谓的无限资本获取，他们不太关心效率。你会遇到物理约束，必须想办法提高效率。所以我认为这会推动更多有趣的创新。坦率地说，我一直非常看好、并且我们认为已经持续看到的一个方向是，你不需要超过万亿 token 的参数就能实现我们目前看到的能力。抽象地说，在当下你需要那么多参数才能达到前沿水平，但我们持续看到越来越小的模型能匹配一两年前最大模型的能力。所以我认为这只会加速。会有越来越多的推动力去研究如何让模型尽可能小。你可能会看到更多投资进入能帮助实现这一点的领域，比如蒸馏，以真正降低推理成本。如果你能做到这一点，就能缓解相当一部分问题。但我仍然预期使用量的增长会快于你能缓解的速度。

It pushes towards efficiency, right? Generally, especially the frontier labs have not cared too much about efficiency because of said infinite capital access. You will get to physical constraints where you have to figure out how to be more efficient. So I think it'll drive a lot more interesting innovation. Frankly, one of the directions that I've always been very bullish on, and I think we've seen consistently, is that you do not need trillion-plus token parameters in order to achieve the capabilities that we currently see. In the abstract, at the moment you need that to get to frontier level, but we consistently see that smaller and smaller models can match the largest models of even one to two years ago. So I think that will only accelerate as a result of that. There will be more and more push towards how do you make models as small as possible. You'll probably see a lot more investment in areas that can help with that, like distillation, to try to really reduce inference costs. If you can do that, that can alleviate a fair amount of this. I would still expect that usage is going to grow faster than what you can do to alleviate this.

是的，我完全同意 Ari 关于效率的观点。我认为这个大规模算力受限世界的另一个有趣影响是，这对除英伟达之外的其他芯片供应商来说是个非常好的消息。我认为基本上几乎所有人都更愿意用英伟达 GPU 而不是其他任何东西，也许除了谷歌的人。但就是没有足够的 GPU 可供使用。所以你已经看到公司在做他们必须做的事来适应使用 AMD GPU、亚马逊 Trainium，而 Cerebras 显然因此获得了巨大的顺风。我认为基本上任何人能弄到的任何芯片都会有巨大需求。所以我认为这对英伟达不是坏事，但对其他芯片持有者是好事。我确实认为这是过去六到九个月的一大主题——这些异构芯片的崛起，随着推理占据主导，你可以把这些推理工作负载分成预填充和解码等部分，你确实可以使用异构芯片。我的问题是：这些其他芯片真的能帮助我们缓解算力短缺吗？考虑到我们一直在谈的这个主题——瓶颈不断转移——所以芯片上游存在瓶颈，比如芯片的组件或台积电的芯片生产。

Yeah, I definitely agree with Ari on the efficiency point. I think another interesting implication of this massive supply-constrained world is it will be a very good thing for other chip providers other than Nvidia. I think basically pretty much everyone would probably prefer to use Nvidia GPUs over anything else, maybe other than people at Google. But there just aren't enough GPUs to go around. So you're already seeing companies doing what they have to do to adapt to use AMD GPUs, Amazon Trainium, and Cerebras is obviously seeing massive tailwinds because of this. I think basically any chips anyone could get their hands on will be in massive demand. So I think it's not a bad thing for Nvidia, but it's a good thing for other chip holders. I actually do think that's one of the big themes of the last six to nine months—the rise of these heterogeneous chips, as inference is dominated and you can split these inference workloads into pre-fill and decode and other things. You really can use heterogeneous chips. The question I have is: do these other chips really help us on the compute shortage, given this exact theme we keep talking about—you keep shifting bottlenecks—so there is a bottleneck upstream of the chips, which are the components that go into the chips or the production of the chips at TSMC.

你知道，ASML 之类的——拥有更多芯片真的能解决那个根本瓶颈、能源这些问题吗？还是说，就像 Jeff 一样，我们看的每个领域，这些供应商都很乐意看到除了 Nvidia 之外有更多玩家。我不知道这些其他芯片的崛起是否真的能给我们带来更多算力。我很想听听你的看法。

You know, ASML and all this stuff—does having more chips actually solve that fundamental bottleneck, energy, all these things? Or is it just like, similar to Jeff, every space we look at, all these vendors are very happy to have more players than just Nvidia. I don't know if the rise of all these other chips actually gives us more compute. I'd be curious for your thoughts on that.

我觉得这个直觉是有道理的，因为你想，如果想象一个世界里没有 Cerebras、没有 D-Matrix、没有其他人在做有竞争力的芯片，那在那个世界里 Nvidia 大概就会把台积电的产能全吃掉，对吧？而全球芯片总数保持不变。所以结果可能只是把价值从 Nvidia 转移出去，而不是真正从根本上改变局面。

I think that intuition makes sense, because right, if you imagine a world in which there's no Cerebras, there's no D-Matrix, there's no one else who's kind of making competitive chips, presumably Nvidia just eats up the TSMC capacity in that world, right? And the total number of chips stays constant in the world. So likely it just accrues value to not Nvidia as a result of that, rather than actually changing the situation fundamentally.

对，对，对。我觉得这些替代芯片供应商并不是算力约束的解决方案，但会是算力约束的受益者，因为 Nvidia 并没有拿到台积电的全部产能和所有供应链、生产。所以如果在一个芯片过剩的世界里，大家都会更愿意用 Nvidia，人们能买到 Nvidia，对这些其他芯片的需求就没那么大。但在一个你拿不到足够 Nvidia 芯片的世界里，人们就会为 AMD 芯片付很多钱，为亚马逊芯片付很多钱。而且，正如你所说，芯片的总体供应并不会因为它们的出现而改变，但我确实认为它们会迎来巨大的顺风。

Yeah, yeah, yeah. I think the alternative chip providers aren't a solution to the compute constraints, but will be a beneficiary of the compute constraint, in the sense that Nvidia doesn't get all of TSMC's production capabilities and all of the supply chains, production. So if in a world where there is a surplus of chips, everyone would rather use Nvidia, people can buy Nvidia, there just isn't that much demand for these other chips. But in a world where you just can't get your hands on enough Nvidia chips, then people are going to pay a lot for AMD chips and pay a lot for Amazon chips. And it won't, to your point, the overall supply of chips won't be changed as a result of them, but I do think they'll see massive tailwinds.

对吧？台积电不想要一个买方垄断。

Right? TSMC doesn't want a monopsony.

完全同意。但这也和 NeoCloud 世界有惊人的相似之处，对吧？基本上是一样的——并不是说整体芯片数量增加了，而是它们肯定在众多不同玩家之间被重新洗牌。只要产能约束还在，每一个玩家都是非常扎实的生意。那么，鉴于目前的情况，我们是否真的预期短期内不会出现算力约束？我们是不是真的在谈论 2035 年，或者说，嘿，15 年后也许有人会弄清楚这些生意会怎样。可能不是未来几年的事。比如，如果 2030 年前发生什么，我会很惊讶。话虽如此，当人们正在尝试的所有缓解这些瓶颈的不同方法同时开始奏效时，这很可能会发生，对吧？我们会有几年时间，大概在 2030 年代初，会有几个大的解锁，这是我的猜测。那大概就是旋转木马开始转的时候。我想另一个问题是，随着更多芯片被生产出来，我们能在多大程度上继续从旧芯片中获得大量价值？这也可以——我认为一个巨大的领先指标是 H100 的价格逆转了下跌，那大概发生在我们录制上一集的时候，那是 12 月开始的。但 H100 的价格在过去几个月里大幅上涨。所以我认为那里有很多芯片可以使用，但我确实想知道当旋转木马停下来时会发生什么，最终它们都在卖完全相同的产品，差异化极小。

Totally. But it's a fascinating parallel to the NeoCloud world too, right? Where you basically have the same thing—it's not like you're necessarily increasing the number of overall chips, but they've certainly been shuffled among a bunch of different players. And as long as the capacity constraint is there, every one of those is a very solid business. Do we actually, given what's happening, expect there not to be a compute constraint anytime soon? Like, are we really talking about 2035, or like, hey, in 15 years maybe someone will figure out what happens to these businesses. It's probably not in the next couple years. Like, I would be surprised if anything happens before 2030. That said, when all of the different methods that people are working on to try to relieve these bottlenecks all start hitting at the same time, which is probably what will happen, right? We'll have several years where there will be several big unblocks, probably in the early 2030s, would be my guess. That's probably when the merry-go-round starts. I think the other question is, as more chips are produced, to what extent can we start to continue to get a lot of value out of older chips? And that can also—I think one huge thing that has been the leading indicator of this is that H100 prices reversed their drops, and that happened, I think, right around when we recorded the last episode—that was when that started to happen in December. But H100 prices have gone up dramatically over the last number of months. So I think there are a lot of chips there that can be used, but I do wonder what happens when the merry-go-round stops on that, and they ultimately are all selling the exact same product with minimal differentiation.

说到 NeoCloud 领域，显然我认为那个世界里最大的头条是 xAI，对吧？而且是以一种巨大的方式进入这个领域。你们怎么看 xAI 的未来？你们觉得，显然，他们是不是就打算继续大规模这么做？你们觉得模型业务有戏吗？我想现在提一下 Cursor 的潜在收购以及它如何契合，时机也不错。但很想听听你们其中一位聊聊这个。

Well, speaking of the NeoCloud space, I mean, obviously I think the biggest headline in that world is xAI, right? And just coming into the space in a huge way. What do you guys make of the future of xAI? Do you think, obviously, do they just lean into continue to do this at scale? Do you think the model business has any legs? I guess it's not a bad time to bring up the Cursor potential acquisition and how that fits in. But would love to hear one of you riff on that.

是的，这是个好问题。我想说，对于 xAI 作为前沿实验室的未来轨迹，我并不特别乐观。而且我认为 xAI 签下的这些大规模协议，把算力租给 Anthropic 和 Google——一方面，是的，这是在 IPO 前粉饰收入数字，这显然起了作用。但我也觉得很难不把它解读为一个信号，即公司的首要任务不是做前沿 AI 研究、不是推动 xAI 的研究，因为如果是的话，考虑到过去 20 分钟关于世界算力受限的讨论，你就不会把那些算力送出去。

Yeah, it's a good question. I would say I'm not super optimistic for xAI's future trajectory in terms of them being a frontier lab. And I think these massive deals that xAI signed to rent compute to Anthropic and to Google—on the one hand, yes, it's padding the revenue numbers ahead of the IPO, and that obviously plays a role. But I also think it's hard not to interpret it as a signal that the company's foremost priority is not doing frontier AI research and fueling xAI's research, because if it was, given the last 20 minutes of discussion around how much the world is compute-constrained, you just wouldn't be giving away that compute.

如果所有那些芯片都用于训练，在产品端又没什么收入可展示，那 IPO 就太难了，对吧？

Pretty hard business to IPO if all those chips were being used for training, right, without much revenue to show for it on the product side.

马斯克的公司不是按基本面估值的。我不知道，马斯克的公司完全是另一回事，所以谁知道呢。

Elon companies aren't valued on fundamentals. I don't know, it's totally different with an Elon company, so who knows.

而且，顺着 Ari 的观点，我认为 Elon Musk 和他的公司最擅长的，是极其运营密集型、现实世界、原子而非比特的事业。显然，我们在特斯拉身上看到了，在 SpaceX 身上也看到了。所以对我来说并不意外，在 AI 世界里，这个 SpaceX-xAI 巨头能拥有真正持久优势的一个楔子，我认为会在数据中心端。我认为他们会擅长——而且他们已经擅长——极其快速地搭建大规模集群，让它们上线运行。我认为这对他们来说会是一门好生意，而且可能会让他们成为全球最大的云，尤其是在未来几年公司开始把越来越多的算力送入轨道之后。但我不知道，我只是觉得——在我看来，组织上处于 AI 模型竞赛的前沿并不是优先事项，也不一定现实。显然，我们看到了过去一年左右 xAI 令人难以置信的人员流失。而且我认为，对于整个马斯克帝国来说，有一个做模型业务的模型部门是有意义的。但如果要我下注，我不会特别看好他们能重新挤进与 Google、OpenAI、Anthropic 并列的第一梯队。

And kind of to Ari's point, I think the thing that Elon Musk is amazing at, and his companies are amazing at, is incredibly operationally intense, real-world, atoms-not-just-bits undertakings. Obviously, we've seen that with Tesla, we've seen that with SpaceX. So it's not surprising to me that in the world of AI, the one wedge where this SpaceX-xAI behemoth is going to have a real durable advantage, I think, will be on the data center side. I think they will excel at—and they already have excelled at—standing up massive clusters super fast, getting them up and running. And I think that will be a great business for them, and it may turn them into the world's biggest cloud, especially as in the years ahead the company starts putting more and more compute into orbit. But I don't know, I guess I just don't—it doesn't seem to me like organizationally being at the frontier of the AI model race is a priority or is necessarily realistic. Obviously, we've seen just the insane attrition from xAI over the past year or so. And I think it makes sense for the overall Elon code to have a model arm that's doing model stuff. But if I had to bet, I wouldn't be super bullish that they will crack back into the top echelon alongside Google, OpenAI, Anthropic.

那他们在做什么？比如，为什么，为什么是 Cursor？

So what are they doing? Like, why, why Cursor?

我认为收购 Cursor 是为了获取所有的轨迹数据。

I think why Cursor is to get all the traces.

那么用于编码的 traces 值 600 亿美元吗？

So are traces for coding worth $60 billion?

坦白说，我不确定它们值 600 亿美元。但如果你认为那是能让你实现跨越式发展、加速你拥有强大编码模型的东西，我能理解这种想法。相对于这个目标，600 亿可能偏高。但我不认为埃隆已经放弃了想要拥有最好模型的雄心。我确实认为，从实际角度看，如果你的纯粹目标是构建最好的模型，你就不会赠送大量算力。不是赠送，是出售大量算力。所以我认为这印证了 Rob 的观点，即这不是他们的首要任务。而且，你知道，Cursor 尤其考虑到那笔交易的结构方式，它就像一个期权。有点像在说，嘿，我们想在接下来的几年里保持一定的选择余地。

I'm not sure if they're worth $60 billion, frankly. But if you think that that's what can leapfrog you and accelerate you to having a strong coding model, I can see where that comes from. 60 billion is probably quite high relative to that. But I don't think Elon has given up the ambition probably of trying to wanting to have the best models. I do think that when you look at the practical aspects of it, it's hard if your pure goal is to build the best models, then you would not be giving away massive amounts of compute. Not giving away, selling massive amounts of compute. So I think that it does go to Rob's point that it's not their top priority. And, you know, Cursor's even especially given the way that deal was structured, like it is an option. It's kind of like hey we want to maintain some amount of optionality for the next number of years.

是的。直到我们看到 traces 是否真的是打造一个真正好的编码模型最重要的东西。

Yeah. Till we see if traces actually are the most important thing to make a really good coding model.

关于 Cursor 这一点，我相信你们都看到了 S1 文件里 SpaceX 的 TAM 图表，并且笑了，我想他们估计了整个太空市场。

And on the Cursor point I'm sure you guys all saw and laughed at the SpaceX TAM chart in the S1 where I think they estimated the total space.

我从不嘲笑大的 TAM 图表。必须喜欢这种雄心。

I never laugh at a big TAM chart. Got to love the ambition.

你只会垂涎欲滴。

You just salivate.

28 万亿美元听起来不错。

28 trillion sounds good to me.

哦，是的。整个太空市场大概是五六千亿，然后通信基本上就是星链，大概一二十亿，然后企业 AI 大概是 20 万亿之类的。

Oh yeah. All of space was like I don't know five or six hundred billion and then comms basically Starlink was like one or two billion and then enterprise AI was like 20 trillion or something.

我认为当整个太空市场只占你 TAM 的几个百分点时，这总是好的。

I think it's always good when all of space is not close to a few percentage points of your TAM.

是的。没错。所以无论如何，我认为这指出了为什么 Cursor，他们当然喜欢这个叙事，很多这些都融入了 IPO 叙事，即我们将赢得企业 AI，而像 Cursor 这样的资产，因为在加入 Cursor 之前，他们确实没有应用或产品表面。他们是否真的能在这项事业中成功，我认为远不那么清楚。但我认为这就是定位和叙事。

Yeah. Exactly. So anyway, I think that points to why Cursor, like they certainly like the narrative again a lot of this goes into the IPO narrative is around like we're going to win enterprise AI and an asset like Cursor because before adding Cursor they really had no application or product surface area. Whether or not they'll actually succeed in that undertaking I think is a lot less clear. But I think that's the positioning and narrative.

我们这一集已经讲了不少，还没提到任何与 Andrej Karpathy 相关的内容，我觉得，你知道，我们几乎每集都会提到他。他似乎是贯穿始终的主题。你知道，首先他说一切都是垃圾，然后决定加入某个实验室。显然他加入了那个递归自我改进团队。而且感觉两个实验室都非常直言不讳，嘿，我们将在 2028 年实现 AGI。感觉推特上都是那些含糊的帖子，说嘿，我们非常接近了。你们觉得我们实际上有多接近？Andrej 决定是时候回去了，这算多大的信号？

We've made it through a good chunk of the episode here and we haven't mentioned anything Andrej Karpathy related which I think is, you know, I think we've done it pretty much every episode. He seems to be the unifying theme. You know, first he says that everything's slop, decides to go join one of the labs. Obviously he went to join this recursive self-improvement team. And it feels like both labs have been very vocal, hey we're going to have AGI by 2028. It feels like there's all these Twitter vague posts about hey we're getting really close. How close do you guys think we actually are? And like, is Andrej deciding it's time to go back in? How much of a signal is that?

我认为我们更接近了。我的估计是，我们现在比 6 个月前接近得多。我会这么说。我确实认为这件事发生了变化，而且它成为我观点有所改变的地方之一，我对此方向变得更加乐观。我仍然认为瓶颈根本在于算力。我想我们上一集可能讨论过，想法不一定就是挑战。甚至执行也不一定是挑战。你必须真正去运行实验。而这需要算力。所以我确实认为改进速度会有一些瓶颈。但我们显然正在达到模型能够自我改进的地步。我们已经开始在这里进行一些实验，就是让智能体以各种方式、在各种程度的指导下自行进行数据整理，结果比我预期的要更有希望。所以这个方向有很多内容。我认为有很多理由对此感到非常兴奋。但我确实认为很多人也能做同样的事情。所以另一面是，有一种关于 RSI 的观点，认为现在有一个玩家会以超快速度前进并跑掉，然后没人能竞争，对吧？我仍然非常怀疑这种观点，因为我认为存在根本性的算力瓶颈，会限制速度，而且目前至少有 10 家公司拥有资金、人才和专业知识来从事这项工作。我不认为这会从根本上局限于一小群人。

I think we're closer. My estimate is that we're much closer now than we were 6 months ago. I would say so. I do think this is something that has changed and that I've become one of the places where my mind has changed a little bit where I've become more bullish on this direction. I do still think that the bottlenecks are in compute fundamentally. I think we talked about this in maybe the last episode that ideas are not necessarily the challenge. Even execution is not necessarily the challenge. You have to actually then go and run the experiments. And that takes compute. So I do think there's going to be some bottleneck on the pace of improvement. But we are clearly getting to the point where models can improve themselves. We've started to do a number of experiments here around just having agents do the curation itself in various ways with various amounts of guidance and seen far more promising results out of that than I would have expected. So there's a lot in this direction. I think there is a lot of reason to be very excited about it. I do think though that lots of people will be able to do the same thing. So the flip side of this is there's this view of RSI where it's like, okay, now there's one player that's going to go super duper fast and run away and then nobody can compete with it, right? I'm still very skeptical of that view because I think there are just fundamental compute bottlenecks that can prevent the speed and also like there are 10 companies at least at this point that have the funding, the talent and the knowhow to work on this. It's not something that I think is going to be fundamentally limited to a small group of people.

我本以为你会更怀疑，Ari，关于递归自我改进的整个概念，但有趣的是你相信它，却不认同这种起飞叙事。如果真正攻克了 RSI 却没有这种指数级起飞，那会是什么样子？基本上就是算力作为限制因素吗？

I was expecting you to be more skeptical, Ari, on the whole notion of recursive self-improvement, but it's interesting that you are a believer in it and yet you don't buy into this kind of takeoff narrative. What would that look like for allowed to like crack true RSI and yet not have this sort of exponential takeoff? Is it just the compute as a limiter basically?

我认为很多都在于，算力是一个根本性的限制因素，而且拥有更多人类只能让你在一定程度上更快。我认为有一个问题是，它会比人类好多少。它显然正在达到能与初级 AI 研究员相媲美的地步。拥有一支初级 AI 研究员大军只能让你走这么远。那么它会继续超越吗？然后我认为会有一些速度限制。但我要说，6 个月前我对此要怀疑得多，我认为我们在这里看到了一些明显的进展。我认为你还可以看看一些实验室正在采取的行动，Anthropic 正在放缓招聘大量初级人员，这开始增强人们对这可能实现的信心。但我认为它会比人们说的慢得多。

I think that's where a lot of it, like compute is a fundamental limiting factor with respect to this, and also like having just more humans only makes you go faster to a certain point. And I think there is a question of how much better than humans it is going to be. It's clearly coming to the point where it can be comparable to a junior AI researcher. Having just an army of junior AI researchers gets you so far. So will it continue past that? And then I think there will be some speed limitation. But I will say I was a lot more skeptical of it 6 months ago, and I think we have seen some clear progress here. I think also you look at some of the actions that some of the labs are taking, Anthropic is slowing down on hiring a lot of junior folks, that starts to drive more confidence that this is possible. But I think it's going to be a lot slower than people say.

嗯，我总是喜欢以快速问答结束我们的环节，让你，你知道，激起最后一点火花，为下一集提供素材。所以，我想也许从开始，对你们两位都很好奇，你们最不同意当前更广泛讨论中的哪个常见说法？

Well, I always like to end our sessions with a rapid quickfire where we get your, you know, to drum up the last bit of spice which provides fodder then for the next episode. So, I figure maybe to start, curious for both of you, what do you disagree with most that's kind of a common trope in the broader discourse right now?

我很乐意先来。

I'm happy to start.

我认为这是一个更宏观的观察，但与我们一直在讨论的很多话题相关，比如芯片短缺、资本支出建设、吉瓦级数据中心。对我来说，很明显，在五到十年内，我们回顾当前的人工智能时代时，会觉得今天的人工智能系统在资源效率上有多么可笑。我们需要建造两吉瓦级的数据中心——这相当于整个旧金山市两倍的电力——来运行这些最先进的模型。再回到我最喜欢的一个话题，把它与人脑作为存在证明相比：人类智能，也就是我们最终想通过人工智能实现的目标，只靠 20 瓦的功率运行。我不认为会有单一的银弹突破，但在硬件层面会有巨大的进步——比如 Naveen Rao 的公司正在做的模拟计算，诸如此类的努力会带来芯片能效的根本性突破。当然，还会有很多算法和优化方面的突破。但我认为这与需求短缺、持续时间以及资本支出建设等问题有着复杂且不明显的交集。

I think this is a bigger picture observation but relates to a lot of the discussion we've been having around the chip shortage, the capex buildout, the gigawatt-scale data centers. To me, it seems so clear that in five to ten years, we're going to look back on the current era of AI and it's going to be laughable how resource-inefficient today's AI systems are. The fact that we need to build two-gigawatt-scale data centers—that's twice as much power as all of San Francisco—to run these state-of-the-art models. And to return to one of my favorite hobby horses, compare that to the human brain as an existence proof: human intelligence, which is what we're trying to achieve with AI, runs on 20 watts of power. I don't think there's going to be one silver bullet breakthrough, but there will be massive advances at the hardware level—things like Naveen Rao's company working on analog computing, efforts like that that lead to fundamental breakthroughs in the energy efficiency of chips. For sure, there'll be a lot of algorithmic and optimization breakthroughs. But I think it has complex and not obvious intersections with the question of the demand shortage, how long it lasts, and the capex buildout.

是的，这很有趣。从中得出两点。第一，如果事情真的朝那个方向发展，那么领先模型提供商的护城河就不那么有说服力了，对吧？显然，如今护城河的很大一部分只是资本获取和规模。所以你已经看到所有这些从 OpenAI 出来的人，我想他们中的很多人都有类似的灵感去追求什么。我从来都不相信，即使他们真的搞清楚了，最终会是一门好生意。因为，正如我们一直说的，这些想法会扩散，而且在一个很长的时间里，你不太可能有一个与其他人的想法根本不同的想法。所以如果世界真的朝那个方向发展，那会非常有趣。然后第二点——Ari，你总是提到这一点，我想在我们的第一集里，你在中国开源模型的背景下谈过这个——就是约束孕育创新。我想知道，如果你现在在 OpenAI 或 Anthropic，激励就是继续推动当前的范式并做得更好，因为这太有价值了。所以你已经看到像 Jerry Twark 这样的人从 OpenAI 出来，说：‘好吧，我想专注于更接近你所说的东西，Rob。’而且，看看这些进步——下一个重大进展——是发生在大型实验室之一，还是可能在一个附属的地方，这将非常吸引人。我不知道你对这个有没有预测，Ari。

Yeah, it's interesting. Two things come out of that. One is that if things do go that way, then the moats on the leading model providers are far less compelling, right? Obviously, a huge part of the moat today is just access to capital and scale. So you've had all these folks spin out of OpenAI, and I think a lot of them have similar inspiration for what to chase. And I've never been convinced that even if they do figure that out, it ends up being a great business. Because, to our point we've always talked about, these ideas do diffuse, and it's very unlikely that for a long period of time you'll have an idea that looks fundamentally different from these other folks. So it'll be really interesting if the world does go that way. And then the second point—Ari, you always hit on this, I think in our first episode you talked about this in the context of Chinese open-source models—is that constraints breed innovation. And I wonder, the incentive if you're at OpenAI or Anthropic right now is just to keep pushing the current paradigm and getting better at it, because it's so valuable. So you've had folks like Jerry Twark spin out of OpenAI and say, 'Well, I want to focus on something that feels more akin to what you're talking about, Rob.' And it'll be fascinating to see whether these advances—the next big advance around this—happen in one of the big labs or maybe in an ancillary place. I don't know if you have a prediction on that, Ari.

我认为你会看到它同时从许多地方出现。有一种多重独立发现的概念，就像有多少次事物是被独立发现的？有一个非常有趣的维基百科页面，你可以去看看——基本上每一个重大的科学发现都是由几个人同时发现的。然后随着通信带宽的增加和延迟的降低，现在世界各地第二天就能知道，这变得快多了，而且现在你看到这种情况少多了。但我认为你清楚地看到，想法就是准备好了。OpenAI 并没有发明测试时计算的想法；很多人都在研究测试时计算，o1 只是先出来了。我认为对于我们在递归自我改进或其他任何方面看到的任何东西，你都会看到同样的事情。想法的空间并不是那么大，而且想法确实会变得成熟。所以无论最终什么起作用，我敢打赌，会有其他几个人同时在研究同样的事情。所以我认为两者都会有一点。

I think you'll see it come from many places simultaneously. There's this notion of multiple independent discoveries, where like how many times are things discovered independently? There's a really interesting Wikipedia page about this where you can go and look—basically every major scientific discovery was discovered by several people simultaneously. And then as communication bandwidth increased and latency decreased, now what's happened across the world the next day, that got a lot faster, and now you see that a lot less often. But I think you clearly see that ideas are just ready. OpenAI didn't invent the idea of test-time compute; many people were working on test-time compute, o1 just came out first. I think for anything we're seeing with recursive self-improvement or any other aspects, you're going to see the same sort of thing. The space of ideas isn't that massive, and ideas do become ready. So whatever ends up working, I would bet that several other people will be working on the same thing simultaneously. So I think it'll be a little bit of both.

那么，在当前的更广泛讨论中，你最不同意的是什么？

And what do you disagree with most in the broader discourse right now?

可能我最不同意的事情，尽管这可能有争议，是永久下层阶级的概念，以及人工智能将在十年内取代所有人类工作的想法。我认为我们真的会对此大笑。

Probably the thing I disagree with the most, although this could be controversial, is the notion of the permanent underclass and just this idea that AI is going to take all human jobs in a decade. I think we're really going to laugh hard at that.

是的，我在这里对 Roko 的蛇怪表现不佳，但这有点有趣，对吧，Anthropic 甚至到处公开说这个等等。我倾向于认为这被夸大的原因是，从根本上说，人类在经济中传播事物的速度很慢。即使是我们在看到的工具，也需要很长时间才能真正完全渗透。我们现在就看到了这一点，以至于很多商业和一切实际上都是关于人与人之间的互动和信任等等。这些是技术官僚往往不考虑的障碍。而且我认为很多在这方面预测快速时间线的人低估了世界在各种方式上可能有多慢。所以我倾向于认为那是被夸大的。那可能是我最不同意的事情。但也有可能我只是希望人类能更长时间地保持重要。

Yeah, I'm doing poorly at Roko's basilisk here, but it's kind of interesting, right, that Anthropic is even going around and saying this very visibly and so on. The reason I tend to think that this is overblown is that fundamentally, humans are slow at dissipating things through the economy. It's going to take a long time for even the tools we're seeing to really fully percolate. And we're seeing that now, such that so much of business and everything is actually about human-to-human interaction and trust and so on. Those are barriers that I think technocrats tend not to consider. And I think a lot of the folks that are projecting really fast timelines around this are underestimating how slow the world can be in various ways. So I tend to think that that's overblown. That's probably the thing I disagree with the most. But it's also possible that I just want humans to still matter for longer.

是的。Ari，你提到过改变你对开源模型的看法，以及那个领域会有多少参与者。在过去的几个月里，你觉得还有什么其他方面你的观点发生了变化吗？

Yeah. Ari, you mentioned changing your mind on open-source models and how many players there would be in that space. Anything else in the last few months you feel like you've shifted your perspective on?

我比之前更看好递归自我改进了。这是一个变化。

I'm more bullish on RSI than I was. That's a change.

但我认为开放模型在某种程度上会消失，我现在非常坚信这一点。六个月前我并没有预见到这一点，就像我没有预见到开源模型的崛起一样。我想我们当时都对 Llama 的发布感到非常惊讶。现在情况相反，但除此之外没有别的了。

But I think this notion that open models are going to go away to some extent, I believe quite strongly now. I didn't really see that coming six months ago, in the same way that I didn't see the rise of open source models coming. I think we were all very surprised by that past Llama. I think now the opposite is the same, but nothing beyond that.

我改变看法的一件事是，我确实缩短了关于机器人 AI 和机器人模型的性能与改进速度的时间线。六个月前，我们上次聊天时，我可能会站在这样的阵营：这是不可避免的，是巨大的市场机会，但谁知道需要多长时间。可能是 18 个月，也可能是五年，才能达到通用机器人 AI 的神话般的 GPT-3 时刻。但在过去的几个月里，这些模型在工作效果方面确实跨越了一个门槛。显然还有很长的路要走，但机器人基础模型已经达到了一个点，它们足以在许多不同的用例中实现商业可行性。随着它们开始部署给客户，数据飞轮开始旋转，我认为这只会加速。所以我现在确实认为我们离所谓的机器人领域的 GPT-3 时刻不远了。

One thing I've changed my mind on is I've really pulled in my timelines in terms of the performance and improvement rate for robotic AI and robotic models. Six months ago, whenever we chatted last, I probably would have been in the camp of this is inevitable and it's a massive market opportunity, but who knows how long it's going to take. It could be 18 months or five years to this mythical GPT-3 moment for general purpose robotics AI. But in just the past handful of months, these models have really crossed a threshold in terms of how well they're working. There's obviously still a long way to go, but robotic foundation models have reached a point now where they are capable enough to be commercially viable across a lot of different use cases. As they start to be deployed with customers and that data flywheel starts to spin, I think it's just going to accelerate. So I do now think that we aren't far from this so-called GPT-3 moment in robotics.

这真是悦耳动听。但当然，我认为看到这个领域取得的一系列进展非常令人兴奋。显然，在机器人领域有很多活动，而且在生物材料科学领域，我认为人们肯定感受到了这些其他领域的一些进展。所以未来几年将会非常迷人。我想我最后一个问题是，我现在厚着脸皮要问一个额外的辛辣预测，关于今年下半年，你认为到年底我们都会意识到实际上是真的的事情。

Music to my ears. But certainly I think it's been really exciting to see a bunch of the progress in that space. Obviously there's been a ton of activity around robotics, and I think in bio material sciences, people are definitely feeling a lot of progress in some of these other spaces. So it will be fascinating to see these next years. I guess my last question for you guys is, I'm shameless now going to ask for an additional spicy prediction for the back half of the year, something you think by the end of the year we'll all be realizing was actually true.

你知道，我实际上会下 API 的赌注。这有点激进，因为今年发生而不是明年。我更有信心说到 27 年底，但我认为很有可能看到 Anthropic，但也可能是 OpenAI，暂停 API 访问一段时间，或者在短期内大幅限制 API 访问。这可能是未来更频繁发生这种情况的前兆。

You know, I'll take the API bet actually. This is a bit aggressive for it to happen this year versus next year. I feel a lot more confident saying that by the end of '27, but I think there's a very reasonable chance that we see probably Anthropic, but it could be OpenAI, suspend API access for some period of time, or heavily limit API access for a brief period. That maybe is the precursor to starting to see this happen more frequently down the line.

这是个辛辣的预测。我喜欢这个。是的，如果那发生了，今天看来如此不可能，如果确实发生了，那将是一个压力电话。

That is a spicy take. I like that one. Yeah, if that happens, it seems so unlikely today that if that does happen, that'll be a pressure call.

时机很难把握。我确信那会在某个时候发生，无论是在 26 年下半年。这是期权定价的缺点。

Timing is hard on that one. I'm confident that'll happen at some point, whether it'll be in the back half of '26. This is the downside of options pricing.

很难超越预测 Sam Altman 被罢免。

It's going to be hard to top predicting Sam Altman's ouster.

是的。还有谁出局？

Yeah. Who else is out?

是的，我认为我的预测是 Dario 年底仍会在 Anthropic。这是一个不那么戏剧性的预测，而且这些天可能不那么辛辣了，因为我认为人们越来越清楚这一点，但我认为到年底，Anthropic 在生命科学和生物学领域将成为正在崛起的巨头，这一点会变得非常明显。在我看来，很明显这是 Anthropic 正在专注的下一个大方向。Anthropic 理所当然地因其专注而获得了很多赞誉和钦佩，因为它在编码方面如此专注，并且取得了巨大成功，并以此为跳板超越了 OpenAI，而 OpenAI 则有点杂乱无章。所以我认为这种专注非常有价值。我确实认为很明显，生命科学是他们的下一个大赌注。一方面，你可以说这是否代表了专注的分散，会给他们带来问题？我有点认为这是他们在编码之后的下一大篇章，而且未来几年都会如此。我认为真正有趣的问题是，我知道 Jacob，你在医疗保健领域深耕，我相信你对此有很多想法，但我认为真正有趣的问题是 Anthropic 计划在价值链上走多远？我确信人们听到很多传言说 Anthropic 正在建立自己的湿实验室设施，以进行自己的实验并收集数据。那只是为了收集数据来训练模型，还是他们在考虑，或者最终会考虑开发自己的资产？显然，正如该领域通常的情况一样，这将是一个逐步发展的过程。但我确实认为，也许这更适合三年期的预测，但随着时间的推移，Anthropic 将成为世界上最重要的生命科学公司之一。

Yeah, I think my prediction is Dario will still be at Anthropic at the end of the year. This is a less dramatic one, and maybe this is probably less spicy these days because I think it's becoming increasingly evident to folks, but I think by the end of the year it will be very obvious that Anthropic is a fledgling juggernaut in the making in the life sciences and biology. It seems clear to me that that is the next big direction that Anthropic is focusing on. Anthropic has rightfully gotten a lot of praise and admiration for its focus and for being so dialed in on coding and just knocking that out of the park, using that as the stepping stone to leapfrog OpenAI, whereas OpenAI was kind of all over the place. So I think that focus has been really valuable. I do think that it's clear that life sciences is their next big bet. On the one hand, you could say does that represent a fragmentation of focus that's going to be problematic for them? I kind of think that it's their next big chapter after coding, and it will be for the next several years to come. And I think the really interesting question is, and I know Jacob, you are deep in the healthcare world, and I'm sure you have a lot of thoughts on this, but I think the really interesting question is how far down that value chain is Anthropic planning to go? There are plenty of rumors that I'm sure folks have heard that Anthropic is setting up their own wet lab facilities to run their own experiments and collect data. Is that just to collect data to train models, or are they thinking about, or will they eventually be thinking about developing their own assets? Obviously, as is generally the case in the space, it'll be a progression over time. But I do think, maybe this is more suited for a three-year prediction, but in the fullness of time, Anthropic will become one of the most important life sciences companies in the world.

他们迄今为止在生命科学领域所做的事情有多大优势？显然，你可以使用大语言模型作为研究的副驾驶，但似乎模型本身实际上相当不同，对吧？

How much of an advantage do they have from what they've already done to date within life sciences? Obviously you can use LLMs to be a co-pilot for research, but it seems like the models themselves are actually quite different, right?

是的，我认为他们今天没有巨大的优势。我认为他们唯一的优势是，他们可以说是世界上最好的 AI 研究机构，如果他们选择将这种能力指向……

Yeah, I don't think they have a huge advantage today. I think their only advantage is that they're arguably the best AI research organization in the world, and if they choose to point that capability at...

如果只有一位有远见的 CEO 在顶级 AI 实验室，他对生物感兴趣一段时间，并且实际上花了一些时间在上面，并在多年前创办了一家公司。那可能会很有趣。

If only there was a visionary CEO at a top AI lab who had been interested in bio for some time and actually spent some time on it and started a company on it years ago. That might be pretty interesting.

好的，我会修改我的预测。Anthropic 和 Isomorphic Labs 将成为两家最重要的生命科学公司，因为你对 ISO 的看法是对的，但我认为 Anthropic 也能做到，而且我确实认为 Dario 长期以来一直对生物学充满热情，而且他是神经科学博士等等。

Okay, I'll revise my prediction. Anthropic and Isomorphic Labs will be two of the most important life sciences companies because you're right about ISO, but I think Anthropic can get there, and I do think Dario has also been passionate about biology for a long time, and he's a neuroscience PhD and so forth.

完全同意。我的意思是，我认为你观点中真实的一点是，我认为生物实际上可能是仅有的足够大的机会之一。我的意思是，显然 Anthropic 的人从使命的角度对它感兴趣。他们在这方面一直如此一致，而且从影响的角度来看，这是 AI 最酷的影响之一。

Totally. I mean, I think the thing I do take in your point that is real is I think bio actually probably is one of the only opportunities that is large enough. I mean, obviously the Anthropic folks are interested in it from a mission perspective. They've been so consistent in that, and it's one of the coolest impacts of AI from an impact perspective.

它可能也是少数几个大到足以支撑那种投入规模的市场之一。你有代码，有面向知识工作的通用副驾驶，而且说实话，像这么大的市场并不多。所以我也认为这个机会可能是巨大的。我们拭目以待。生物很难。数据反馈循环，以及真正获取纵向数据，都需要时间，这甚至不是——即使是机器人技术也容易得多，对吧？因为至少你可以立即得到关于某件事是否奏效的反馈循环。

It also is probably one of the only TAMs that is large enough to justify that level of investment. You've got code, you've got your general co-pilot for knowledge work, and I don't know, there's not many that are as large. So I also think the opportunity could be massive. We'll see. Bio is hard. The feedback loops on data and actually getting longitudinal data just has a time to it that isn't even—even something like robotics is much easier, right? Because you can at least immediately get the feedback loop on whether something worked or not.

是的。

Yeah.

我认为领域专业知识的重要性远超许多 AI 从业者通常的认知。所以有一个问题：Anthropic 会如何应对——是雇佣一批非常出色的生物学家，然后让他们与强大的 AI 研究者配对，在那里做尽可能多的工作，还是仅仅试图让 AI 研究者去硬碰。我认为这正是 Demis 在 AlphaFold 和后来的 Isomorphic 上做得非常好的地方——不假设仅靠 AI 研究者就能做到。我见过很多其他实验室犯了这种错误。但我认为领域专业知识非常重要。我认为这对 Anthropic 来说会很难，但如果他们能做到，如果他们能组建那样的团队，那显然是一个巨大的利好。不过目前我可能更看好 Isomorphic。

I think domain expertise also matters a lot more than many AI folks often think. So there's a question of how Anthropic approaches this—by hiring a bunch of really fantastic biologists and then pairing them with strong AI researchers, doing as much there, versus just trying to throw AI researchers at it. I think that is one thing that Demis did very well with AlphaFold and then Isomorphic—not assuming that AI researchers alone can do that. I've seen lots of other labs kind of make that mistake. But I think domain expertise matters a lot. I think it's going to be a hard one for Anthropic to do, but if they can do it, if they can hire that team, it's obviously a huge boom. But I would bet probably more on Isomorphic right now.

好了，各位，这真是太有趣了。我非常感谢你们两位抽出时间来深入探讨。我们得做得更好——这次和下次之间不能再隔六个月了。你关掉了太多候选人。下次我们会更快安排。

Well, guys, this has been a ton of fun. I really appreciate you both taking the time to jam on this. And we had to do better—no six months in between this one and the next. You're closing too many candidates. Next time we'll do it sooner.

R 是唯一一个有实际工作的人，我们得围绕他的日程来安排。

R is the only one with an actual job that we have to schedule around.

是啊。我不知道。你在光刻方面提出了一些非常好的见解。我当时就想，有人在那边做了不少功课。我完全没想到。

Yeah. I don't know. You came in with some really good takes on lithography there. I was like, someone's been doing some work over there. I was out of left field.

是啊，说真的。我学到了很多关于替代方法的知识。

Yeah, seriously. I learned a lot about alternate approaches.

我是 Jacob Efron，这里是 Unsupervised Learning，一档我能与 AI 领域最聪明的人对话，并向他们提出大量关于模型发展及其对商业世界影响的问题的播客。正如我希望已经明确的，我做这件事非常开心。这是我除了在 Redpoint 做投资人的日常工作之外，利用夜晚和周末做的项目。但我们能请到这些出色的嘉宾，真的离不开像你这样的听众订阅播客、与朋友分享。这最终才是让这一切运转起来的根本。所以，请考虑这样做。非常感谢你的支持和收听。我们下期再见。

I'm Jacob Efron, and this has been Unsupervised Learning, a podcast where I get to talk to the smartest people in AI and ask them tons of questions about what's happening with models and what it means for businesses in the world. As I hope is clear, I have a ton of fun doing this. It's a nights and weekends project in addition to my day job as an investor at Redpoint. But our ability to get these incredible guests on really comes from folks like you subscribing to the podcast, sharing it with friends. It's really what ultimately makes this whole thing work. And so, please consider doing that. And thank you so much for your support and listening. We'll see you next episode.