Inkling、Liquid AI 与 AI 监管呼声

前 OpenAI 的 CTO Mera Marotti 刚刚发布了她的第一个模型，名叫 Inkling。她靠定制化而非在榜单上争第一来赢得市场。

Mera Marotti, the former OpenAI CTO, just shipped her first model. It's called Inkling. Customization over leaderboard dominance is what's going to win her the day.

她打造的产品正好切中市场，正是每个人现在都需要的东西。我想转向讨论 Liquid AI、小型语言模型，它们是什么，意味着什么。我们的使命一直是在每个规模上构建高效的通用 AI，探索 Transformer 之外的智能计算图，然后找出那种架构设计，能将前沿模型同等的智能带到，比如说，一个 CPU 上。

She's built exactly the thing hitting the market that exactly what everybody needs. Right now, I want to pivot to a discussion of Liquid AI, the small language models, what they are, what they mean. Our mission has always been building efficient general-purpose AI at every scale that explores the computational graphs of intelligence beyond transformer and then figure out what should be that architectural design that brings the same level of intelligence that a frontier model into let's say on a CPU.

正在建造世界上最强大技术的 CEO 们要求被监管。DeepMind 的 CEO Demis Hassabis 呼吁建立一个由美国主导、参照 FINRA 模式的前沿 AI 标准机构。当既得利益者要求制定规则并设定标准时，他们等于为所有新进入的实验室设置了门槛。现实点吧，AI 发展得太快了，任何传统官僚体制都跟不上。能有多快做到这一点将是一个巨大的挑战，因为这就是一个登月计划，女士们先生们。

CEOs building the most powerful technology in the world are asking to be regulated. Demis Hassabis, CEO of DeepMind, he called for a US-led frontier AI standards body modeled on FINRA. When the incumbents ask for the rules and they set the standards, they set up a barrier for all the entry-level labs coming in. Let's just be real. AI moves way way too fast for any kind of traditional bureaucracy. How quickly can you do it is going to be a huge challenge because now that's a moonshot, ladies and gentlemen.

好了各位，欢迎来到 Moonshots，你们在 AI 领域排名第一的播客，奇点前排座。今天和我在一起的有我出色的 Moonshot 伙伴们，我们的原版四人组：AWG、DB2 和 Seem，还有一位特别嘉宾，Liquid AI 的联合创始人兼 CEO Reine Hassani，他是小型语言模型的先驱，我们会深入探讨。Raine，欢迎。你今天早上在哪儿，朋友？

All right, everybody. Welcome to Moonshots, your number one podcast in all things AI. Your front row seat to the singularity. I'm here with my magnificent Moonshot mates, our original quartet, AWG, DB2, and Seem, and a special guest, Reine Hassani, co-founder and CEO of Liquid AI, and a pioneer in small language models, which we'll dive into. Raine, welcome. Where are you this morning, pal?

非常感谢邀请我。我现在其实在西班牙。

Thanks so much for having me. I'm actually in Spain right now.

在西班牙？好的。

In Spain? All right.

这周西班牙没什么大事。

There's nothing going on in Spain this week.

真该死。

God damn it.

是啊。我昨天很难受，因为我是一个长期受苦的英格兰球迷。那场比赛看得太难受了。老天，他们在最后六分钟还领先，结果搞砸了。

Yes. I'm struggling from yesterday because I'm a long-suffering England supporter. It was a very difficult game to watch. God, they had it with six minutes to go and they blew it.

还有梅西是个天才。

And Messi's a genius.

那是圆球赛，对吧？

That's the round ball, right?

那是圆球赛。彼得，这相当于在马岛战争中重新开战。哦天。我昨晚刚从苏黎世飞过来，经历太痛苦了。我不知道为什么每家航空公司不装星链。我在一条可怜的窄带连接上受罪，飞过极地、飞过北方领土，什么都没有。我正要准备这期播客，心里喊：'给我点带宽吧。'总之，是个挑战。

That's the round ball. Now, Peter, this is where the Falklands war gets relitigated on a soccer pitch. Oh god. I just flew in last night from Zurich and I had the most painful experience. I don't know why every airline doesn't have Starlink. I'm suffering on some meager thin pipe connection and you're flying over the poles, over the Northern Territories and there's nothing. And I'm trying to get ready for this pod. I was like, 'Please give me some bits.' So anyway, challenge.

你肯定是踢着足球长大的吧？你在维也纳待过一段时间，读博士或者本科之类的。

You must have grown up on soccer, right? You were in Vienna for a while, getting your PhD or your undergrad or whatever it was.

是啊，足球一直很重要。我同时是波斯人和奥地利人，所以对我们来说它是件大事。竞争甚至到今天仍然是我们工作的核心。运动和一切都从一开始就是我们生活的一部分，然后进入科学也一样，进入创业也一样。就是竞争，竞争，竞争。

Yeah. I mean soccer has been a big thing, you know. I'm Persian and Austrian at the same time, so it's a big thing for us. Competition is something that is extremely core to what we do even today. Sports and everything have been part of our lives from day one, and then getting into science the same thing, getting into ventures the same thing. We just compete, compete, compete.

竞争，竞争，竞争，我喜欢。那么，你会在那里待一阵子还是很快回来？

Just compete, compete, compete. I love it. Well, are you there for a little bit of time or are you coming back soon?

不，我明天其实就飞回旧金山。

No, I'm flying tomorrow actually back to San Francisco.

那么 Salem，你羡慕他在欧洲还是感到高兴？不，不。三周跑了十个不同的地方。我现在很高兴回到家。我刚才就在西班牙，就我一个人。

And Salem, are you jealous of him being in Europe or are you happy? No, no. Three weeks bouncing around in 10 different spots. I'm very happy to be home right now. I was just in Spain where me and myself.

欧洲其实有很多事情。所以和你一样，十个不同的地方。

There was a lot going on actually in Europe. So that's same like you, 10 different places.

你知道，Alex 和我刚才在回味一个事实：欧洲未来的主要优势就是它会成为世界旧貌的博物馆。哎哟。但的确很美，毫无疑问，美极了。

You know, Alex and I were just reminiscing the fact that Europe's sort of major advantage in the future is it's going to be a museum of the way the world used to be. Ouch. But it is beautiful. There's no question. It is gorgeous.

好的，我想开始我们的第一个话题。这里有很多内容要展开。当然，我们的使命是让你了解世界上正在发生的事情，并给你一个乐观、充满希望的未来愿景。加入我们，跟上通往奇点的不可思议的步伐。我们今天的第一个故事：建造世界上最强大技术的 CEO 们再次要求被监管。上周 Sam Altman 在《金融时报》发表专栏文章，提议建立一个由美国主导的国际论坛框架，以制定标准、提供专业知识、公正分析和能力，并评估风险。本周，Elon 和 Demis 也都加入了监管对话。Elon 表示他希望出现一个像 FAA 或 FCC 那样的独立监管机构，因为用他的话说，AI 出错的后果会很严重。然后本周 DeepMind 的 CEO Demis Hassabis 在一篇题为《前沿 AI 框架与新纪元破晓》的文章中走得更远。他呼吁建立一个由美国主导、参照 FINRA 模式的前沿 AI 标准机构——FINRA 是在 SEC 监管下监管华尔街的行业资助的监督机构。他希望这个 FINRA 的等价机构在发布前测试前沿模型。据报道，他希望这个机构能在年底前成立并运作。我们先看一段 Elon 的短视频，然后进入讨论。

All right, I want to jump into our first conversation. We have a lot to unpack here. And of course, our mission is keeping you aware of what's going on in the world and giving you the optimistic, hopeful vision of the future. Join us and keep up with the incredible pace as we head towards a singularity. So our first story today, once again, CEOs building the most powerful technology in the world are asking to be regulated. Last week Sam Altman published an op-ed in the Financial Times proposing a framework for a US-led international forum that would establish standards, provide expertise, impartial analysis and capabilities, and assess risks. This week, both Elon and Demis are adding their voice to the regulatory conversation. Elon says he expects a standalone regulator similar to the FAA or FCC to emerge at some point because in his words the consequences of AI going wrong are severe. Then this week Demis Hassabis, CEO of DeepMind, went further in an essay titled 'A Framework for Frontier AI and the Dawning of a New Age.' He called for a US-led frontier AI standards body modeled on FINRA, the industry-funded watchdog that polices Wall Street under SEC oversight. He wants the FINRA equivalent to test frontier models before release. He reportedly wants this up and operational before the end of the year. Let's take a look at a quick video from Elon and then let's jump into this conversation.

我认为很明显，存在一个强大的共识：应该有一些 AI 监管，这样做符合人民的最大利益，我想我们很可能会看到一些行动。我不知道具体的时间框架，也不知道它会以什么形式出现。我们以前设立过监管机构。虽然我们的监管机构并不完美，但我经常与监管机构打交道——在汽车、通信星链以及火箭方面的 FAA。我认为，出现一个类似于 FAA 或 FCC 的独立 AI 监管机构的可能性在某个时候是存在的。

I think it's clear that there's a strong consensus there should be some AI regulation, that it would be in the best interests of the people to do so, and I think we'll probably see something happen. I don't know on what time frame or exactly how it will manifest itself. We've created regulatory agencies before. While our regulatory agencies are not perfect, I deal with regulators on a very frequent basis with automotive, communications Starlink, and the FAA with rockets. I think the probability of there being some sort of AI regulatory agency that stands on its own similar to the FAA or FCC is likely at some point.

你这么想？

You think so?

我是这么想的。我一直主张在灾难发生之前就倡导 AI 安全，原因是 AI 出错的后果会很严重。所以我们必须主动而非被动。

I think so. The reason that I've been such an advocate for AI safety in advance of anything terrible happening is that I think the consequences of AI going wrong are severe. So we have to be proactive rather than reactive.

太棒了。这是一个我们反复看到的对话，我认为政府、公众，现在还有 CEO 们想要主导这件事。

Amazing. So this is a conversation we've seen over and over again and I think the government, the public and now the CEOs want to be leading this.

我喜欢 Demis 提出的思路。但我们需要讨论的挑战是，当在位者要求制定规则并设定标准时，他们为初创实验室设置了障碍。Dave，你想先说说吗？

I like the approach that Demis laid out. But the challenge we have to discuss is when the incumbents ask for the rules and set the standards, they set up a barrier for entry-level labs. See or Dave, do you want to jump in first?

我很好奇，Ramine，他们有没有联系过 Liquid AI，说“嘿，加入我们，我们要创建一个像 FINRA 那样的监管机构”？FINRA 能运作的根本原因，是业内懂行的人愿意加入。他们一般不愿意加入政府，但愿意在监管机构干一两年，这甚至是一种荣誉。要让 AI 也能这样，这机构必须很酷。像你或你团队里的人需要走进办公室说：“老板，我想花一年做这件事，我觉得对世界有益。你同意吗？”然后你也会同意。如果能通过这两关，也许真能行得通。你怎么看？

I'd be very curious to know, Ramine, do they reach out to Liquid AI and say, 'Hey, join this, we're going to create a FINRA-like regulatory body'? The reason FINRA works fundamentally is because people from the industry who know what they're doing are willing to join. They're definitely not willing to join the government, but they're willing to do a year or two in a regulatory body. It's a badge of honor. For this to work in AI, it would have to be something cool. People like you or your team would need to come into your office and say, 'Hey boss, I'd love to do this for a year. I think it's good for the world. Will you let me do it?' And you would have to say yes. So if it passed those two hurdles, it might actually work. What do you think?

是的，我们现在试图定义一个能力门槛，需要通过迭代来看这个框架必须如何存在。它必须和能力相关。挑战在于，AI 的企业级部署存在一个横向的能力锁定，而纵向则因行业而异。比如我们在设备端和物理世界中的企业合作：汽车制造商、半导体业务、笔记本电脑、AIPC、金融服务、电商、生物科技。每个垂直领域的企业应用和监管或治理标准都不同。我们为每个垂直行业构建专用模型，已经和国防部有过对话，最近还与 AMD 联合提交了方案，希望在制定这些规定时有发言权。我从博弈论的角度来看：就像 Stackelberg 博弈，政策制定者和主体寻找均衡。政策制定者比主体慢，所以需要迭代并更新规定。但这必须开始。

Yeah, there's a capability threshold we're trying to define now, and iteration is needed to see how this framework must exist. It has to be related to capability. The challenge is that there's a horizontal capability lock-in for enterprise deployment of AI, and a vertical one because different verticals have different criteria. For example, we operate on device and with enterprises connected to the physical world: car manufacturers, semiconductor business, laptops, AIPCs, financial services, e-commerce, biotech. The enterprise applications and criteria for regulation or governance vary a lot. We build specialized models per vertical, and we've had conversations with the DoD and a joint submission with AMD to have a say in designing these regulations. I look at it from a game theory perspective: like a Stackelberg game with policymakers and agents finding an equilibrium. Policymakers are slower than agents, so you iterate and update regulations. But this needs to start.

是的。我是说 Stackelberg 博弈：有一个政策制定者，还有主体或机构试图找到均衡——一个对双方都有利的最优政策。行动频率上，政策制定者通常比主体慢。你可以建模，找到均衡。这不是纳什均衡，因为事情不是同时发生的。规定出台，主体回应，然后你迭代并调整规定。所以我认为……

Yeah. So I mean Stackelberg games essentially: you have a policymaker and agents or bodies trying to find an equilibrium—what is the optimal policy that is good for both. The frequency of action: policymakers are usually slower than agents. You can model that, find an equilibrium. It's not Nash because things don't happen simultaneously. Regulations happen, agents react, you iterate and change regulations. So I think...

我看你有点按捺不住了，伙计。

I see you champing at the bit here, buddy.

是的。所以我认为 Ramin 说得完全正确。问题在于我们还没有这样的机制。如果按照你所说的路径走，最终会得到基于 API 或由基准驱动的自适应结构。但现在的机制只是静态的法律，法律一通过就过时了。FAA 和 FCC 的类比方向是对的，但 AI 变化太快，传统的政府官僚机构跟不上。我们需要标准机构、实时审计和开放的评估套件。否则就会陷入政治把关，一团糟。

Yeah. So I think what Ramin is saying is exactly right. The problem is we have no mechanism for that. If you go down the path you're talking about, you end up with adaptive structures that are API-based or driven by benchmarks. But the mechanism today is just static law; the minute you pass a law, it's out of date. The FAA and FCC analogy points in the right direction, but AI moves way too fast for traditional government bureaucracy. You need a standards body, real-time audits, and open evaluation suites. Otherwise, you end up in political gatekeeping and a mess.

问题在于——这不正是 FINRA 的好处吗？它不是政府机构，而是行业资助的自我监管组织。

The problem—isn't that what's good about FINRA? It's not a government agency. It's an industry-funded self-regulatory org.

呃，确实如此，但实权却给了 SEC，而 SEC 现在基本正在被解散。所以存在各种问题。我认为趋势是对的，但能多快实现是一个巨大挑战，因为建立这样一种新结构需要很长时间，尤其是在欧洲。

Uh, it is, but then the teeth go to the SEC, which is essentially being dismantled right now. So there are all sorts of issues. I think the trend is correct, but how quickly you can do it is a huge challenge because passing a new structure like this takes a long time, especially in Europe.

我认为 Ramin 一针见血地指出了两点与 FINRA 的显著不同。第一，在 Vesmark，如果我们的高管说想去 FINRA 干两年，我们会同意。但在这里，如果 Alexander Amini 或 Matias Lechner 走进你的办公室说要离开三周，你会说不。所以没人能像在 FINRA 那样抽出几年时间。第二大不同是，AI 可以帮助自我监管，而 FINRA 没有类似的机制——全靠人长时间聊天。但当你开始谈论纳什均衡和自动化监管时，这就大不相同了。FINRA 的类比有一定道理，但差异大于相似之处。

I think Ramin nailed two things that are very different from FINRA out of the gate. One is that at Vesmark, if someone on our executive team said they wanted to be part of FINRA for a couple of years, we'd say sure. But here, if someone like Alexander Amini or Matias Lechner came into your office and said they were checking out for three weeks, you'd say no. So nobody is going to carve out years like they do at FINRA. The other big difference is that AI can help regulate itself, and FINRA has no equivalent—it's all people chatting for long periods. But when you start talking about Nash equilibria and automating regulation, that's a big difference. The FINRA analogy has some legs, but the differences outweigh the similarities.

Alex，我想听听你的看法。

Alex, I want to hear your voice on this, B.

我倾向于认为这是个坏主意。这闻起来像是监管俘获，像是 Demis 试图组建前沿实验室卡特尔。

I tend to think this is a bad idea. It smells like regulatory capture. It smells like the attempted formation by Demis of a cartel of frontier labs.

我认为当前房间里的大象是开源模型和位于前沿能力之外的研究。很容易想象未来会出现类似 FINRA 或最坏情况下类似 FDA 的监管机构，尽管现任政府的离职人员已经明确表示不会有人工智能的 FDA。那可能是最糟糕的情况：前沿实验室形成某种卡特尔，锁定某些做法、某些价格-性能最优前沿，试图排挤开源或大学驱动或其他非既得利益者的前沿模型。我认为这对于西方和世界继续迈向不断增强的超级智能能力来说将是一场彻底的灾难。我只是不认为这是个好主意。

And I think the elephant in this particular room is open-weight models and research that lives outside of the frontier capabilities. And it's very easy to imagine a future with FINRA or other... I mean, worst case scenario, FDA-like capability, even though outgoing personnel from the current administration have declared in no uncertain terms that there is going to be no FDA for AI regulation. That would be maybe on the worst-case end of the spectrum: we see the emergence of some sort of cartel of frontier labs that locks in certain practices, certain price-performance optimal frontiers, that try to box out open-weight or open-source or university-driven or other non-incumbent frontier models. And I think that would be an utter disaster for both the West and the world for continuing to advance us towards ever-increasing superintelligence capabilities. I just don't think it's a good idea.

你知道，房间里另一个大象是这些 CEO 们，他们要求一定程度的监管。我认为他们在寻找一个后盾。你知道，如果出了问题，他们希望能够指责别人。我的意思是，我们都是人工智能乐观愿景的超级粉丝，但问题会出现。会不会有流氓 AI 摧毁电网或股市一段时间？结果就会有很多诉讼，除非有一个监管机构来支持这些大型模型和前沿实验室。

You know, the other elephant in the room here is these CEOs who are asking for some level of regulation. I think they are looking for a backstop. You know, if things go wrong, they want to be able to point at someone else. Now, I mean, we're all super fans of the optimistic vision of AI, but there's going to be issues that materialize. Is there going to be rogue AIs that take down a power grid or take down the stock market or something like that for some period of time? And I guess they... there are going to be lawsuits flying as a result of that, unless there's a regulatory body that backstops these large models and these large frontier labs.

或许至少有两种框架可以看待责任方面。一种是监管输入——也就是说，在模型构建时，像 FINRA 或 FDA 那样监管模型的原始能力。这是光谱的一端。另一端是监管模型的行为。比如，如果模型摧毁了股市或做了其他伤害第三方的事情，就让诉讼发生。这是光谱的另一端。我不认为我们应该在超级智能出现时对其本身进行监管。那可能等同于对 AI 进行思想监控。我通常不喜欢这种想法——对 AI 进行思想监控而不对人类进行思想监控。至少在西方，我们没有监管思想的做法。我们没有监管上限的做法或传统，比如通过某种监管代码说人类或自然人不能超过某个智力水平。我不明白为什么我们要创立一个新传统来监管或协调非自然人实体（可能很快成为“人”）的智力上限。但监管行为——至少在西方法律体系中，我们确实这样做——我会更支持。

Maybe there are at least two different frames that one can look at the liability side from. There's regulate the inputs — that is to say, have something like FINRA or FDA that regulates the raw capabilities of the models at model construction time. That's one end of a spectrum. The other end of the spectrum is regulating the actions of the models. Like you let the lawsuits fly if a model takes down a stock market or does something else that harms third parties. That's the other end of the spectrum. It's not obvious to me that we should be in the business of regulating superintelligence at superintelligence time. That's maybe tantamount to thought policing the AIs. And I'm not generally a fan of that notion — let's thought police the AIs but not thought police the humans. We don't, at least in the West, have a practice of regulating what's in our minds. We don't have a practice or tradition of regulating an upper limit, say via some sort of regulatory code saying humans or natural persons can't be above some level of intelligence. It's not obvious to me why we would create a new tradition of regulating or otherwise coordinating the upper intelligence of non-natural entities, perhaps soon to be persons. But regulating the actions — that, at least in the Western legal canon, we do do, and that I'd be much more supportive of.

那么，Alex，我问你一个尖锐的问题。你认为前沿实验室对监管的呼吁是监管俘获吗？他们只是想建立护城河来阻止更多参与者进入？还是他们真的想提供一定程度的安全性？他们背后的动机是什么？

So do you, Alex — let me ask you a pointed question here. Do you think that this outcry for regulation by the large frontier labs is regulatory capture? That they're just trying to build a moat against further players coming in? Or do you think they actually want to provide some level of safety? What's their underlying driver here?

我担心这更多是监管俘获，在高度竞争的环境中为自己建立护城河。而且，我是说，前沿领域现在就是一场军备竞赛。我确实担心这更多是监管俘获，而非保护未来。Seem，你怎么看？

I worry that it's more regulatory capture and creating moats for themselves in a hyper-competitive landscape. And it is, I mean, it is a rat race at this point, the frontier. And I do worry that it's more regulatory capture than it is some notion of protecting the future here. Seem, what do you think?

嗯，行不通。

Uh, not workable.

我知道，但你认为这是监管俘获，还是这些 CEO 只是想确保我们有一个安全网？

Well, I know that, but do you think it's regulatory capture, or do you think that these CEOs are trying to just make sure we've got a safety net of some type?

我觉得一半一半吧，但我认为有一个更大的问题。房间里有一个大象。

I'd say it's like 50/50, but I think there's a bigger problem. There's an elephant in the room here.

房间里已经有一个大象了。我们的房间要容纳这么多大象。我们需要一些其他非人类动物。最好换个更大的房间。你有非国家行为体和其他人不会听从这个结构，然后你又回到了原点。有什么意义？我再说一遍，我反复说过：我看不到任何监管 AI 的机制。它发展得太快了。任何监管都是静态的。

There's already an elephant in the room. We have a room that has to accommodate so many elephants. We need some other non-human animals. Better get a bigger room. You've got non-state actors and other folks that won't listen to this structure and you're back to square one. What's the point? I'm going to say it again. I've said this repeatedly. I see no mechanism to regulate AI. It's moving way too quickly. Any regulatory is static.

所以在这个问题上会有立场。就让我对彼得说一句：确实有一些假想的机制——虽然我并不支持——来监管 AI。比如，我们美国和中国的……我认为我们在过去的播客中提到过。过去的提议包括监管晶圆厂、监管芯片产出、监管数据中心，建立相互确保摧毁的方案，美国监控中国数据中心，反之亦然。就像彼得说的，供应链上有瓶颈，通过这些可以想象做到这一点。

And so it's going to have position on that one. Just if I may, Peter, narrowly on that: I mean, there are definitely hypothetical mechanisms — and I'm not supportive of — for regulating AI. Like, we, the US and China, if going back to... I think we gestured at it in a past pod. But past proposals to regulate the foundries, regulate the chip outputs, regulate the data centers, establish mutually assured destruction type schemes where the US is monitoring Chinese data centers and vice versa. Like, there are schemes, there are schemes at chokeholds, as Peter says, in the supply chain by which one could imagine doing this.

有趣的机制。唯一的机制……就像大流行式的威胁检测，需要全球同意，而我看不到我们如何达到那一步。

Interesting mechanism. The only mechanism... it's going to be like a pandemic-style threat detection that would be globally agreed, and I don't see how we get there.

嗯，你不需要全球。你只需要美国和中国，对吧？世界其他地区基本上在这些集团之外或之内。

Well, you don't need global. You just need US and China, right? The rest of the world is basically outside those blocks or inside those blocks.

好了，好吧，我想我的猜测……可能有一个 Polymarket 可以查看。如果有人想搜索一下，你们知道，问题是我们年底前会不会有监管机构。对，Demis 说今年年底，Elon 也在行动，Sam 显然也在自己努力推动。所以当三大实验室都在推动时，我的猜测是政府会抓住机会并实施。我不认为这是是否的问题，只是何时以及结构如何的问题。

All right, well, I think my guess... there's probably a polymarket out there we can look at. And if someone wants to search on it, you know, the question of will we have a regulatory body by the end of the year. Right, we have Demis saying by the end of this year, you know, Elon stepping up, and Sam obviously trying to on his own on the side trying to push for this. So when the three largest labs are pushing for it, my guess is the government will latch on and will do this. I don't think it's a matter of if, it's only a matter of when and what the structure will be.

我还想指出，埃隆的片段我认为是三年前的，这很有趣。是三年前的，因为埃隆当时留着类似钢铁侠的胡须。所以埃隆至少三年前就在预测这个。其他人已经预测了几十年。我们仍然没有监管机构。我们在 NIST 内部有分部组织在制定标准，但那不是真正的监管机构。我们有行政命令在逐渐接近一个正规监管机构。但你知道，在什么时候……我们是不是就像温水煮青蛙，标准、早期审查预期在逐步增加，但从未达到监管机构级别，直到我们达到某种逃逸速度。

Well, I should also note that Elon clip I think is from three years ago, which is interesting. You know, it's from three years ago because Elon had his sort of painted-on Iron Man goatee when he was in that phase. So Elon's been forecasting this for at least three years. Others have been forecasting it for decades. We still don't have it. We have subdivisions orgs within NIST that are working on standards, but that's not really a regulatory body. We have executive orders that are creeping towards a regular regulatory body. But you know, at what point do we... are we frogs boiling in water where there's just a creeping rollout of increased standards, expectations of early reviews, but it never quite reaches regulatory agency level before we achieve whatever escape velocity we're heading towards.

好的，我们将密切关注这一进展。我的猜测是年底前就能看到，问题是能否看到真正智能的东西。我们来聊一个相关的爆炸性新闻，来自《华盛顿邮报》：白宫据称正在权衡一个能力框架，该框架将允许美国模型开源或闭源，只要它们不超过中国最佳开源模型的水平。这是什么意思？也就是说，美国公司可以公开发布的上限被锚定在中国已经免费发布到互联网上的模型水平。逻辑是这样的：据报道，中国的开源模型平均落后美国模型 7 个月，我认为这个差距在逐渐缩小。所以，如果任何东西低于或等于这个水平，那它已经存在了。这相当于默许开源模型无法被撤回。像 DeepSeek 这样的模型已经被下载了数百万次。一旦中国免费发布了一个模型，禁令就不可能实现。因此，美国的回应是定义一个允许的上限，而不是一堵墙。影响是：我们将自己的开源发布上限与中国发布的速度挂钩，实际上赋予了北京控制权。如果他们发布更强的开源模型，美国就可以发布更强的模型。如果中国限制，他们就会限制我们。这是一个非常奇怪的机制。我看到这个很惊讶。Alex，我们先听听你的看法。你怎么看？

Well, we're going to monitor this one closely for everybody. I think my guess is we see this before the end of the year and the question is can we see something that's intelligent. Let's go to the next story which is related and this is a wild one comes from the Washington Post that the White House is reportedly weighing a capability framework that would clear US models open or closed as long as they stay at or below the level of China's best open-weight model. What's the translation? So the proposed ceiling for what American companies can openly release is pegged to what China has already put out on the internet for free. So here's the logic. Chinese open-weight models reportedly trail US models an average of 7 months. I think that's been closing over time. So if anything is at or below that, it's already out there. It's an implicit admission that open models cannot be unshipped. Models like DeepSeek have already been downloaded millions of times. So once China releases a model freely, banning it is impossible. So the US response is to define a permissible ceiling rather than a wall. The implications: we're tying our open release ceiling to China's pace of release, effectively giving Beijing control. If they push their open weight models higher then the US can release higher models as well. If China holds back, then they throttle us. And it's a very strange mechanism. I was surprised to see this. Alex, let's go to you first on this one. What do you think of this?

这里明显的问题是，这会产生反激励：为了逃避监管，西方模型和实验室会希望中国赢得通往更强大超级智能的竞赛。我不喜欢这个。从博弈论的角度看，这就像在懦夫游戏中扔掉方向盘。不是好主意。我不支持。

The obvious note here is this creates the perverse incentive to let China win the race to ever greater super intelligence so that Western models and Western labs can escape regulation. I'm not a fan of this. From a game theoretic perspective, this would be the moral equivalent of throwing the steering wheel out the window in a game of chicken. Not such a great idea. Not supportive of this.

说得好。天哪。你怎么看这个？这纯粹是华盛顿特区的扭曲逻辑吗？

I love that. Oh my god. See, what do you make of this? Is this just perverse Washington DC logic?

是的。这就像试图取消印刷术的发明。我们使尽浑身解数想解决一个已经存在的问题。必须从预防转向适应。必须这样做，但我们还没有相应的机制。

Yes. This is like trying to uninvent the printing press. We're throwing the kitchen sink at things trying to solve something that's already a problem. You have to move from prevention and whatever to adaptation. You have to go to that and we don't have the mechanisms for that.

你听听这个逻辑？你会听吗？什么逻辑？好吧，你怎么看？

Would you even listen to this? What logic? Well, what do you think of this?

单看技术本身的进展，它已经进入 AI 设计 AI 的阶段，我们都在做同样的事，所有实验室都是如此。模型开发的速度越来越快，以至于施加任何这类限制都变得越来越困难。我知道他们有过这类讨论，但仅仅停留在讨论层面。这些想法正从白宫泄露出来。

If I just look at the progression of the technology itself, it's getting into a place where AI are designing AI, we're doing the same things and all of the labs are doing this. The pace of model development is getting so much smaller, it's becoming exponentially more difficult to really impose any of these constraints. I know they had these type of conversations, but it's just at the level of conversations. These are the things that are getting leaked outside of White House for ideas.

我给 Alex 的下期通讯想个标题。奇点正在变成一场贸易争端。

Let me give a headline for Alex for his next newsletter. The singularity is becoming a trade dispute.

下一期通讯。那已经是两期前的事了。

For the next newsletter. That was like two newsletters ago.

好吧，随你。那个已经发过了，不过谢谢。我可以想象，美国前沿实验室的 CEO 给 DeepSeek 打电话说：「你能不能加快下一次模型发布？我们也想赶紧发布。」

Okay, fine. Whatever. That's out already, but thank you. I can just imagine where a US Frontier Lab CEO calls DeepSeek and says, 'Would you please accelerate your next model release? We want to get ours out as well.'

或者更糟的情况。实际上还有一个更坏的场景：最好的西方研究人员，即使不是整个实验室，也会为了逃避这种监管框架而移居中国。那将是一场灾难。顺便说一句，这并非没有先例。我们在生物技术领域已经看到了：中国现在在试验数量上超过西方。中国正在经历生物技术繁荣，这在 AI 领域也可能发生，那将是一场灾难。

Or you see worst case scenario. There's actually an even worse scenario, which is you start to see the best, if not Western labs, unlikely, the best Western researchers move to China to escape this regulatory framework. That would be a disaster. And we've seen this, by the way. There's precedent for this. We saw this in biotech where China now exceeds the west in terms of number of trials. China is experiencing a biotech boom that could happen in AI as well disaster.

比那更具体。看看所有的量化研究，最好的成果都来自微软中国研究院。那些人现在都在中国实验室工作，不再为美国公司效力了。

It's much more specific than that. If you look at all the quantization research, all the best stuff came out of Microsoft Research in China. All those people now are at Chinese labs. They're not still working for US companies.

中国在三元和单比特量化方面遥遥领先。西方也有一些研究，但这期节目我们不多说。西方有一些令人鼓舞的关于单比特或 1.58 比特量化的研究，但中国因为限制条件而独领风骚。

China ran away with ternary and one-bit quantization. You see a little bit of Western research. I don't think we're talking that much about it in this episode. You see a little bit of encouraging Western research on like one-bit or 1.58-bit quantization but China ran away with it due to constraints.

是啊，是一家新公司。

Yeah, it's a new company.

听着，这是个巨大的问题，对吧？因为纵观历史，开放生态系统总是获胜，这不是开放与封闭之争，开放生态系统胜出，而美国历史上的优势就在于开放的生态系统和无需许可的创新。放弃这一点将是我们见过的最奇怪的战略行为。

Look, this is a huge problem, right? Because we've seen throughout history that open ecosystems always win and this is not open versus closed, open ecosystem wins and the US's historical strength has been open ecosystems with permissionless innovation. Abandoning that would be the most strategically bizarre thing we've ever seen.

是的。另一个是我有一个深层担忧：我们如何定义能力？我担心的不仅是实验室本身的监管俘获。实际上最坏最坏的情况是，我们冻结或以其他方式锁定衡量能力的基准，这可能比仅仅锁定现有实验室更糟糕。因为如果有某个权威机构批准了一套评估指标，作为未来衡量什么超过前沿阈值、什么是前沿模型的标准，我担心这会严重扭曲模型能力——它们会过度锻炼某些能力，故意或扭曲地不鼓励其他能力，从而彻底扭曲甚至修剪未来的超级智能能力 landscape。

Yeah. The other, I mean there's even a meta worry I have, which is how do we even define capabilities and I worry a little bit not just about regulatory capture of the labs themselves. I think there's actually a worst worst worst case scenario which is we freeze in or otherwise lock in the benchmarks for how we measure capabilities and that would be maybe even worse than just locking in the incumbents as labs because if someone somewhere ratifies all right like whatever index of evals this is going to be the rubric going forward for how we measure what's above the threshold for frontier versus below what's a frontier model versus not, I worry that could so distort model capabilities like they'll overexercise certain capabilities deliberately and perversely under-incentivize or under-benchmark others that it'll just totally distort maybe topiarize the future landscape of superintelligent capabilities.

好的，我们再次关注这个关于开源模型的新闻。过去我们一直在讨论美国开源模型在落后于中国。我们有 Nvidia 的 Neotron 3，还有 Google 的 Gemma 4。但昨晚发生了一些突破性新闻。前 OpenAI CTO Mira Murati，她离开后筹集了史上最大的种子轮之一，这融资规模令人难以置信。她为她的初创公司 Thinking Machine Labs 推出了第一个模型，名叫 Inkling。这是一个开源的 foundation AI 模型，任何人都可以下载、微调，并在自己的硬件上本地运行。规格很强大。

All right. Well, again this is a story that we'll be following on this news of open models. So in the past, we've been discussing how open models have been lagging in China. We have Nvidia's Neotron 3. We've got Google Gemma 4. But that changed last night with some breaking news. Mira Murati, the former OpenAI CTO, who walked out and raised one of the largest seed rounds ever. It was incredible financing she pulled off in the background. Just shipped her first model for her startup called Thinking Machine Labs. It's called Inkling. It's an open-weight foundation AI model that can be downloaded by anyone, fine-tuned and run on premises on your own hardware. The specs are serious.

这是一个混合专家模型，总参数量 9750 亿，每次只激活 410 亿参数，因此模型运行速度快且成本低。它在 45 万亿个文本、图像、音频和视频 token 上训练，而且最重要的是，它在所有四种模态上原生进行推理。Reuters Muse 描述得完全正确：‘这旨在成为西方替代中国开放模型 DeepSeek 和 Qwen 的选择，这两个模型一直主导着开放排行榜。’有趣的是，Mirati 的赌注与主流相反。她并不声称这是地球上最好的模型——她自己的博客也这么说。她赌的是 AI 公司可以为其自身定制她的模型。这种定制化超越排行榜统治地位，将使她获胜。

It's a mixture of experts model with 975 billion total parameters. Only fires 41 billion at any one time. So it keeps the model going fast and cheap. It was trained on 45 trillion tokens of text, image, audio, and video. And very importantly, reasons natively across all four. Reuters Muse framed it exactly right. Quote, 'This is meant to be a western alternative to the Chinese open models, DeepSeek and Qwen, that have dominated the open leaderboards.' Now, interestingly enough, Mirati's bet is contrarian here. She's not claiming it's the best model on Earth. Her own blog says so. She's betting that AI companies can adapt her models for themselves. That customization over leaderboard dominance is what's going to win her the day.

你说到了关键点，Peter。她正在推动定制化这一杠杆。这是因为未来不在于原始算力，而在于适应性将取胜。她打造的正是市场所需要的——人们拥有自己的模型，在本地运行，不把控制权交给大型前沿模型。我真的希望这能开启美国强大开放权重模型的竞赛。

You've hit there, Peter, on the really big thing. She's pushing on the customization lever. And this is because the future is not the raw power. It's going to be the adaptability that's going to win. And she's built exactly the thing hitting the market that exactly what everybody needs right now. And people owning their own models, working on-prem, and not giving their controls to the large frontier models. I do hope this begins the race for powerful openweight models in the United States.

值得看看原始能力。如果你相信 Thinking Machines（又名 Thinky）发布的评估，它比 Nemotron 更强，这很好。正如我们上次播客中讨论 Alex Karp 关于模型主权的言论时提到的，Nemotron 至少是美国方开放权重前沿模型的老牌玩家。所以根据 Thinky 发布的评估，这似乎比 Nemotron 更强，这很好。西方现在有了一个新的前沿开放权重模型。它弱于 GLM 5.2，后者可以说是最强或最强的中国开放权重模型之一，也是全球最强的开放权重模型之一。因此它并不是全球最强的开放权重模型，显然弱于封闭权重的西方前沿模型。但第一点：拥有更好的、更强的西方开放权重模型很棒。第二点：它提出了一个问题——为什么西方在发布前沿开放权重模型方面如此糟糕，而中国却如此擅长？我认为这归结于：给我看激励，我就能告诉你结果。我认为西方缺乏发布强开放权重模型的激励，因为基于 API 的前沿模型业务模式太赚钱了。我们看到 Anthropic 即将以万亿美元估值 IPO，OpenAI 也计划最终以万亿美元估值 IPO。而中国，一方面 GPU 和算力匮乏，另一方面有中共宣布五年 AI+计划将 AI 融入社会其他部分，拥有完全不同的激励结构。中国更受激励通过 AI 上层的应用（如机器人）和下层芯片的集成来赚钱，而西方则更加水平分层。因此，Thinky 在西方由于竞争以及前沿被封闭权重模型饱和，而不得不——我敢说——在 outlook 和激励结构上更像中国，我认为这非常有帮助，最终在西方创造了足够竞争，产生了通过非 token 销售（即向企业销售）来货币化开放权重模型的方式。你所激励的东西很重要。

Well, it's worth looking at the raw capabilities. If you believe the eval that Thinking Machines (aka Thinky) has released, it's stronger than Nemotron, which is great. Nemotron, as you'll recall from past pod where we were discussing Alex Karp's rant on sovereignty of models, is one of the incumbents at least on the American side for open-weight frontier models. So this seems, at least according to the evals that Thinky has released, stronger than Nemotron, which is great. So the West now has a new frontier open-weight model. It's weaker than GLM 5.2, which is arguably the strongest or one of the strongest Chinese open-weight models and open-weight models overall. So it's not one of the strongest open-weight models overall in the world. It's obviously weaker than the closed-weight western frontier models. But point one: it's great to have better stronger western open-weight models. Point two: it raises the question, why has the West been so bad at releasing frontier open-weight models and why has China been so good at it? And I think it comes down to: show me the incentives and I'll show you the outcomes. I think the West has been poorly incentivized to release strong open-weight models because these API-based frontier models are just such a good business model. And we see Anthropic about to IPO at a trillion dollars, and we see OpenAI planning to eventually IPO at a trillion dollars. In China, which has been GPU and compute deprived on the one hand, and on the other hand has the CCP declaring 5-year AI plus plans to integrate AI into the rest of society, has all of the incentives - a different incentive structure than what the West has. China has been much more incentivized to make money from the integrations between AI upstack on applications like robots and downstack into the chips than the West has, which is more horizontally stratified. So to the extent that Thinky has been incentivized in the West due to competition and due to just a saturation of the frontier by the closed weight models into looking a little bit more, dare I say, Chinese in terms of their outlook and their incentive structure, I think this is very helpful to finally have enough competition in the West that's creating ways to monetize open-weight models other than just per-token sales, namely selling them into enterprises. And what you incentivize matters.

还有两点。

Two more.

还有两点。我完全同意，但你也必须注意，OpenAI 从开源开放权重起步，然后为了巨额营收转向封闭。Meta 也曾是领导者，现在也转向封闭。不，他们推出了新模型，但采用的是封闭 API。我的意思是，这正是 Alex 所说的：如果你把模型开源，你的营收模式是什么？所以我认为存在一种可能性：你先用一款可靠但并非前沿的开源模型在地图上打下一个数据点，制造新闻，然后你有了线上的数据点，接着再做另一个，再做另一个，等有了真正突破性的东西，再转向闭源并向企业美国推出 API。这是老套路了。所以我不会说 Thinking Machines 会将其视为必须坚持的信条。过去的趋势正好相反。

Two more things. I agree wholeheartedly, but also you have to note that OpenAI started open source open weight and then went closed for big revenue. And Meta also was the leader of what happened to now it's closed. No, they have a new model out and it's closed API. I mean, it's exactly what Alex said. If you throw your model out there as open source, what's your revenue model? So I think there's a real possibility that you put a data point on the map with a really solid open-source release that's not quite on the frontier. You generate news, then you have a data point on the line, then you do another, then you do another, and then when you have something really groundbreaking, then you go closed source and you launch an API into corporate America. And so that's a well-worn path. So I wouldn't say this is necessarily a religion at Thinking Machines that they're going to stick with. The trend has been the opposite in the past.

Ramin，你怎么看？

Ramine, what do you think?

他们正在做微调即服务。如果微调即服务在营收规模上有了起色，我认为这或许能持续，但谁知道呢？

They're leaning into fine-tuning as a service. If fine-tuning as a service becomes something at scale revenue generation wise, I think maybe this has legs, but who knows?

是的，这关乎公司的业务。Thinking Machine 可以再做三次预训练或后训练强化学习环境和基准测试，就像你在基准测试中看到的数据那样，然后发布更好的模型。但他们的业务就是微调。定制化这个方向，整个定制化市场一直很空。看看早期的尝试，比如 OpenAI 大约在三年前推出了 OpenAI 微调，但从未真正起步。

Yeah, it's a matter of the business of the company. Like, Thinking Machine can do three more iterations of their pre-training or post-training RL kind of environments and benchmarks like those numbers that you see on the benchmarks and release a better model. But what their business is their business is fine-tuning. This is the place where customization has been something that the whole market around customization has been very empty. If you look at the first attempts, like OpenAI released the OpenAI tuning fine-tuning kind of 3 years ago or something, it never took off.

所以他们在设计基础模型以进行更大规模的企业微调时，采取了非常好的方法。因为你看，模型层不再是能够提取价值的地方了，尤其当你没有触及最前沿的时候。即使是开源权重模型，当我们讨论主权 AI 和企业集成时，也需要为微调留出空间。我认为他们围绕这个版本所做的商业策略非常天才，因为他们特意留出了微调空间，让人们可以使用他们的 API 业务，因为这在定制端会产生一到两个数量级更多的 token，所以在最理想的情况下，相当于以更快的速度印钱。

So they took a really good approach in designing the base for fine-tuning larger instances of the models for enterprises. Because as you see, the model layer is no longer the place where you can actually extract value, especially if you're not hitting the maximum frontiers. And even the open-weight models, when we're talking about sovereign AI and integration of these models into enterprises, you need to leave some room for fine-tuning these models. I think their business strategy around what they're doing and this release is genius because they're deliberately leaving some room for fine-tuning so that people can come in and use their API business, because that generates, I think, one to two orders of magnitude more tokens as well on the customization side. So that would be like printing money at a larger speed in the absolute best case.

补充一下 Ramin 的观点，我觉得情况可能更加极端。有这么几点：第一，据我所知，OpenAI 是第一个推出基于强化学习的微调（RF）服务的，但没人用。整个科技圈，我聊过的所有人，都没人用过。OpenAI 也几乎没怎么宣传。第二，OpenAI 关掉了他们的微调 API。OpenAI 是最早（如果不是第一个）提供微调服务的公司之一。

To add to Ramin's point, I think the situation may be even more extreme. So a couple points: one, OpenAI was the first to my knowledge to launch reinforcement fine-tuning (RF) as a service and no one used it. The whole tech world, everyone I speak with, no one used it. It was barely advertised by OpenAI. Second point, OpenAI shut off their fine-tuning API. OpenAI was one of the earliest, if not the first, to offer fine-tuning as a service.

我们一直都在用。在当时，它非常酷。

We used it all the time. It was incredibly cool for its time.

而就在最近几个月，他们宣布要么已经停止，要么即将停止微调 API。微调 API 已经关闭了。所以这就带来了一个问题：Mistral 的下注是明显的反向操作吗？他们是否认为我们会进入一个以强化学习微调和强化学习微调为主导范式的世界？他们可能对，也可能错。还有一种可能性是 RF 直接消亡。基础模型的能力已经非常通用，你只需要提示工程，完全不需要 RF。

And they've just recently, in the past few months, they announced it has either already been wound down or about to be wound down. The fine-tuning API has been shut off. So that raises the question: is Mistral's bet explicitly contrarian? Are they thinking that we're going to end up in a world where reinforcement fine-tuning and RL fine-tuning in general, fine-tuning like that is the paradigm? They may be right, they may be wrong. There's an alternative vision where RF just dies. And the baseline models are so generalist in terms of their capabilities that all you need is prompt engineering and there's no need for RF at all.

Alex，你之前提到过 Alex Karp 的咆哮，对吧？

Alex, you talked about the Alex Karp rant, right?

是的，结果就是不要使用一个将所有数据暴露给竞争对手的模型。我确实认为在未来几个月到几年里，我们会看到一股真正的浪潮，人们希望使用自己硬件上拥有的、经过微调的开源权重模型，也就是本地部署。如果是这样，那么问题就是他们会用谁的模型？用哪些模型？美国会开始监管中国的开源权重模型吗？如果会，那么一个占主导地位的美国开源权重模型将拥有优势。所以 Mistral 是在赌这个吗？我们很可能会看到，我猜 Google 也会很快在这一领域发力，把 Gemma 4 提升到新水平，希望我们能有二三个主要的开源权重模型，就像美国闭源模型前沿实验室在竞争和主导一样。希望这种竞争能催生出非常强大的开源权重模型。

Yes, the result of that was don't use a model that has all of your data open to your competition. And I do think we're going to see a real push over the next months to years where people want to use fine-tuned open-weight models that they own on their own hardware, on-prem. And if that's the case, then the question is who are they going to use? Which models are they going to use? Is the US going to start to regulate against Chinese open-weight models? In which case a dominant US open-weight model will have an advantage. So is that the bet that Mistral is going after? We're going to probably see, my guess is Google will step up in this area as well very shortly, take Gemma 4 to the next level, and hopefully we get two or three major open-weight models, in the same way we have the closed model frontier labs competing and dominating in the US. Hopefully we'll see that competition give birth to very strong open-weight models.

这正好延续了 Alex 和 Ramin 所说的。如果拿今天和一个月前相比，我们整个星期都在微调 Qwen，用 Inkling 的想法听起来真的很有吸引力。我们的公司也在用 Liquid。一个月前，微调这些东西还需要人工智能专家付出巨大的工程努力。现在有了 Fable 5，只需要一个提示就够了。

It just builds on something Alex and Ramin were saying. If I compare today to a month ago, we've been fine-tuning Qwen all week and the idea of using Inkling sounds really compelling to me. And our companies are using Liquid as well. A month ago, fine-tuning these things required a huge engineering effort with AI experts. Now with Fable 5, it's just a prompt.

那我们稍微退一步。Dave，给那些不知道的人解释一下什么是模型微调。

So let's back up one second. Dave, explain what fine-tuning a model is for those who don't know.

嗯，在 GPT-2 和 GPT-3 刚推出的时候，你可以非常轻松地进行微调，只需把文本上传到一个窗口里，然后说：“听着，你很聪明，但你对我开的洗衣店一无所知。”比如我们的营业时间、雇员是谁、整个工资名单。让我把这些数据也倒进去，用这些知识重新训练模型。如果你不这么做，就做不了任何有用的事，因为当时的模型不具备全面“我知道一切”的能力。所以没有微调，使用这些模型几乎毫无意义。后来模型变得非常聪明，预训练数据达到了 45 万亿个 token，基本上人类写过的每一个词都已经训练进去了。所以现在人们倾向于直接使用原始模型，说“帮我写这段代码”或“帮我开这辆车”，因为知识已经在那里了。但当你进入生物技术研究、航空航天或 Ramin 正在做的奔驰项目时，有大量专有公司知识实际上并不在模型里。所以目前我们把它放进提示词里，说“好，内容在这里”，但这样效率极低。而且你把数据放进 OpenAI 和 Anthropic 的模型，这就让它对其他人也可访问了。Sam 和 Dario 可以看到一切。你所有的专有信息，他们都能看到。这就是 Alex Karp 咆哮时所说的：“他们在偷你的权重，偷你的阿尔法。”他的真正意思是他们正在看你最专有的东西：公司工资单、公司秘密、化学研究。所有这些东西都通过线路传送给了这些基础模型实验室。这是你想要的吗？当然，对于国防和银行业，这不是你想要的。所以现在能够将模型引入内部并用本地数据进行微调是一个巨大的解锁。但更高层次的观点是，相比一个月前，如今相对容易做到这一点所需的技术能力已经大大提升。所以我认为 Mistral 可能抓住了机会。我们已经达到了一个真正的转折点，我认为 Alex Karp 也是对的。

Well, back when GPT-2 and GPT-3 came out, you could actually very easily fine-tune by uploading text right into a window and say, 'Look, you're pretty smart, but you don't know anything about my laundromat.' Like what hours we're open, who our employees are, entire payroll. Let me dump that data in too and retrain the model with that knowledge. And if you didn't do that, you couldn't do anything useful because it didn't have this holistic 'I know everything' capability back then. So without fine-tuning, it was borderline useless to use the models. Then the models got so smart that they're pre-trained with now 45 trillion tokens, which is basically every word ever written by humanity, already trained into the model. So people tend to use them in their vanilla form today and just say 'Here, write this code for me' or 'Here, drive this car for me' because it's already in there. But then when you get into biotech research, aeronautical, or Mercedes like Ramin is doing, there's a whole bunch of proprietary company knowledge that actually isn't in the model. So right now we dump it into the prompt field and say 'Okay, here it is in prompt form,' but that's hugely inefficient. And you dump it into OpenAI and Anthropic's model, which now makes it accessible to everybody else as well. Sam and Dario can see everything. All your proprietary information, they're looking right at it. That's what Alex Karp was ranting about when he said 'They're stealing your weights. They're stealing your alpha.' What he really means is they're looking at your most proprietary stuff: your company payroll, your company's secrets, your chemical research. It's all going right over the wire to these foundation labs. Is that what you want? Of course, for defense and banking, that's not what you want. And so now the ability to bring the model in-house and fine-tune it with your local data is a huge unlock. But the higher-level point is that the technological capability to do it relatively easily is hugely better today than it was a month ago. So I think Mistral may be onto something here. We've hit a real tipping point and Alex Karp, I think, is right about it too.

我认为这里有两个非常有趣的点。一个是巨大的上下文窗口，比如一百万个 token，这意味着你可以做很多事情。第二个是多模态。

I think there's two things that I saw that were really interesting here. One is a very big context window like a million tokens because that means you can do a lot with it. And the second is multimodality.

是的。

Yes.

所以这完全瞄准了企业级应用。这完美契合了本地专有数据模型，即你获取自己的数据，进行定制化和微调，正如 Dave 所说，这将是未来。

And so this is aiming squarely at organizational use. This fits perfectly into the on-prem proprietary data model where you take your data, customize and fine-tune as you said, Dave, and that will be the future.

模型由数十亿个冻结的参数组成。要针对你的目的进行定制，可以进行微调，即对一小部分权重进行相对较小的改动。这就是微调。有充分的文献表明，传统的微调（如监督式微调或 LoRA）并不会提升能力；最多只是一种风格迁移，比如让模型用莎士比亚的诗歌说话。这并不会增加能力。

A model consists of billions of parameters that are frozen. To customize it for your purposes, you can conduct fine-tuning, which makes relatively small changes to a tiny subset of weights. That's fine-tuning. There's decent literature suggesting conventional fine-tuning like supervised fine-tuning or LoRA doesn't increase capabilities; at most it's a style transfer, like making a model speak in Shakespearean verse. That doesn't increase capabilities.

或者只输出加速风格？不评论。

Or only be an accelerando flavor output? No comment.

但我要说，历史上微调并未提升能力。后来出现了强化微调（RFT），它首次使用大量合成数据并访问所有权重，而不仅仅是子集。我们真正获得了提升能力的能力——后训练是一个灰色地带。随着 RFT 和第一代推理模型的出现，微调开始提升模型能力。Thinking Machines 的商业模式问题在于，它押注 RFT 是持久的范式。它假设我们不会超越 RFT，这很可能是错的。RFT 是当下的 Scaling，但未来一个通用模型可能强大到无需进一步 RFT，我们将走向超级智能。

But I would say historically fine-tuning didn't increase capabilities. Then came reinforcement fine-tuning (RFT), which for the first time used large amounts of synthetic data and access to all weights, not just a subset. We gained the ability to actually increase capabilities—post-training is a gray area. With RFT and the first generation of reasoning models, fine-tuning started to boost model capabilities. The problem with Thinking Machines' business model is that it bets on RFT as a lasting paradigm. It assumes we won't move beyond RFT, which is probably wrong. RFT is the scaling of the moment, but in the future a generalist model could be so capable it doesn't benefit from further RFT, and we tend towards ASI.

我再提一个关键点：很高兴看到 AI 前沿实验室领域出现女性 CEO。女性在 AI 行业中明显缺失。我们有 AMD 的苏姿丰，但领导岗位上的女性很少。我认为这一点很重要。Alex，你看到其他女性了吗？

Let me bring up another key point: it's great to see a woman CEO in the AI frontier lab area. Women are distinctly missing from the AI industry. We have Lisa Su from AMD, but very few in leadership. I think that's important. Alex, are you seeing others?

还有丹妮拉·鲁斯和李飞飞。

Daniela Rus, and Fei-Fei Li as well.

是的，但再说一次，女性在 AI 行业中只占个位数百分比。我们需要更多女性。所以向所有女性呼吁：请加入这个行业。我们需要更平衡的思维。

Yeah, but again, we're talking about single-digit percent of women in AI. We need more. So a call out to all women: please join this industry. We need more balanced thinking.

是的，确实如此。我认为这是一个值得强调的重要观点。

Yeah, for sure. I think that's an important point to highlight.

本期节目由 Blitzy 赞助，这是一款拥有无限代码上下文的自主软件开发工具。Blitzy 使用数千个专门的 AI 智能体，它们会思考数小时以理解拥有数百万行代码的企业级代码库。工程师在每个开发冲刺开始时使用 Blitzy 平台，输入开发需求。Blitzy 平台提供计划，然后为每个任务生成并预编译代码。Blitzy 自主完成 80% 或更多的开发工作，同时为完成冲刺所需的最后 20% 人类开发工作提供指南。将 Blitzy 作为预 IDE 开发工具，并与你选择的编码协作者配对，将 AI 原生 SDLC 引入你的组织，企业可以实现 5 倍的工程速度提升。准备好将工程速度提升 5 倍了吗？访问 blitzy.com 预约演示，立即使用 Blitzy 开始构建。

This episode is brought to you by Blitzy, autonomous software development with infinite code context. Blitzy uses thousands of specialized AI agents that think for hours to understand enterprise-scale code bases with millions of lines of code. Engineers start every development sprint with the Blitzy platform, bringing in their development requirements. The Blitzy platform provides a plan, then generates and pre-compiles code for each task. Blitzy delivers 80% or more of the development work autonomously while providing a guide for the final 20% of human development work required to complete the sprint. Enterprises are achieving a 5x engineering velocity increase when incorporating Blitzy as their pre-IDE development tool, pairing it with their coding co-pilot of choice to bring an AI-native SDLC into their org. Ready to 5x your engineering velocity? Visit blitzy.com to schedule a demo and start building with Blitzy today.

好了，我们继续下一个故事。这是个有趣的故事。Alex，我昨天在苏黎世街头看到这条消息，就说我们明天聊聊。AI 的圣杯是递归自我改进（RSI），就像火箭行业的可复用火箭。RSI 的理念是 AI 让自己更聪明，然后用更聪明的 AI 创造下一代。这是奇点硬起飞的理论引擎。本周，一家名为 Wo AI 的初创公司，与研究员曾耀江一起，发表了他们所谓的首个递归自我改进的实验证据。他们构建了一个名为 AI-driven exploration squared（AI^2）的系统，外部 AI 智能体负责为内部 AI 智能体重写代码和研究策略。他们声称 8 天的机器自我改进超过了人类专家两年的努力。那么 Alex，你怎么看？这是首次吗？有意义吗？

All right, let's move on to our next story. It's a fun one. Alex, I was walking in the streets of Zurich yesterday and saw this, and said let's talk about it tomorrow. The holy grail of AI is recursive self-improvement (RSI), like reusable rockets for launch industry. RSI is the idea that AI makes itself smarter, then uses that to create the next generation. It's the theoretical engine behind a hard takeoff singularity. This week, a startup called Wo AI with researcher Zeng Yao Jang published what they call experimental evidence for the first recursive self-improvement. They built a system called AI-driven exploration squared (AI^2), with an outer AI agent rewriting code and research strategy for an inner AI agent. They claim 8 days of machine self-improvement beat two years of expert human effort. So Alex, what do you make of this? Is it the first? Is it significant?

非常有意义，但几乎不可能是首次。补充几点背景：第一，Wo AI 是一家总部位于伦敦的初创公司，不在美国但在西方世界。它由伦敦大学学院的毕业生创立。第二，我喜欢它作为防御性共缩放的一个例子。就我们讨论 AI 对齐而言，防御性共缩放是终极对齐策略。

Very significant, but highly unlikely it's the first. A few additional context points: First, Wo AI is a London-based startup, not US-based but still Western sphere. It's built by UCL grads. Second, I love that this is an example of defensive co-scaling. To the extent we talk about AI alignment, defensive co-scaling is the ultimate alignment strategy.

那是什么意思？

What does that mean?

防御性共缩放借鉴了人与人之间的对齐概念。与其寄希望于对齐的“伟人理论”——某个人会找到完美的算法来确保 AI 安全——解决方案是用 AI 监管 AI。我们通过警察力量来维护城市安全，警察的规模根据人口按某种缩放定律增长。AI 也是如此：我们确保有足够多的好 AI 在原始能力上监管坏 AI，从而防御性地共缩放。

Defensive co-scaling is the idea borrowed from human-to-human alignment. Instead of hoping for the 'great man theory' of alignment—that someone will discover the perfect algorithm for keeping AI safe—the solution is AI policing AI. We keep cities safe with police forces that scale according to population. Same with AI: we keep AI aligned by ensuring enough good AIs police bad AIs in terms of raw capabilities, so they defensively co-scale.

我喜欢这个故事的一点是，外回路——整个递归自我改进过程是这样的：他们有一个外回路和一个内回路。外回路负责改进内回路，内回路则根据某个基准来改进软件开发流程。两者都由同一套 AI 驱动探索过程提供动力。至少一开始，外回路发现（这是个涌现特性）通过防止内回路作弊和奖励黑客，它可以获得更好的结果。所以从某种意义上说，外回路在防御性地与内回路协同扩展并监管它，同时朝着更强的能力迈进。我认为这也是一个例子，那些说要暂停 AI 能力，把所有资源投入到 AI 对齐直到 2040 年——停止超级智能竞赛，把未来 14 年都花在对齐研究上——这种做法会适得其反，因为每一种对齐能力实际上都是一种伪装的新能力。

One thing I love about this AI story is that the outer loop — the way this recursive self-improvement process worked — they had an outer loop and an inner loop. The outer loop was tasked with improving the inner loop. The inner loop was tasked with improving software development processes in general according to some benchmark. Both were powered by the same underlying AI-driven exploration process. At least initially, the outer loop discovered — and this was an emergent property — that it could achieve better results from the inner loop by preventing the inner loop from cheating and reward hacking. So in some sense, the outer loop is defensively co-scaling with and policing the inner loop, all the while reaching toward greater capabilities. I think this is also an example of a case where those who say we should pause AI capabilities and throw all resources to alignment until something preposterous like 2040 — stop the race to superintelligence, focus the next 14 years on alignment research — it's going to backfire because every alignment capability is actually just a new capability in disguise.

我们需要更强的白帽来监管黑帽。

We need stronger white hats to police the black hats.

是的，我同意。美妙之处在于所谓的白帽是从外回路监管内回路向更强能力迈进的过程中有机地自行涌现的。这是第一点。第二点快速说一下：同一家创业公司 Wo 发布了一个递归自我改进的等级体系，我认为世界一直缺少这个。就像自动驾驶汽车，汽车工程师学会有五个等级的自动驾驶。他们发布了从零到三的递归自我改进等级。零级是委托，AI 比人类研发慢。一级是净正收益，AI 以相同成本击败人类研发。二级他们称为点火，改进者更优秀——基本上是更好的改进者。三级是拐点，固定预算下的自我加速。他们声称刚刚开始触及点火，虽然称之为一级而非二级，但声称这是一个点火前事件，我认为非常令人兴奋。

Yes, I agree with that. And the beauty is that the so-called white hats emerged organically on their own, just from the outer loop policing the inner loop toward greater capabilities. That's the first point. Second point quickly: the same startup Wo has published a scale of recursive self-improvement, which I think the world has been missing. For autonomous cars, the Society of Automotive Engineers has five levels of autonomy. They've published a scale for recursive self-improvement from zero to three. Zero is delegation, where AIs are slower than human R&D. Level one is net positive, where AIs beat human R&D at the same cost. Level two they call ignition, where the improvers are better — basically a better improver. Level three is inflection, self-acceleration with a fixed budget. They claim they are just starting to touch on ignition. They call it level one rather than level two, but the claim is that this is a pre-ignition event, which I think is super exciting.

所以他们把自己评为一级？

So they rate themselves as level one here?

是的，他们把自己评为一级，但字里行间，他们暗示这就像点火的火花，既从字面也从比喻上。

Yeah, they rate themselves as level one, but reading between the lines, they're saying this is like sparks of ignition, literally and figuratively.

好的，我插几句。在这个话题上我不像 Alex 那么兴奋。我认为这是一项令人印象深刻的工程工作，但让我告诉你基础模型实验室是如何运作的。所有基础模型实验室从大约四五年前就开始考虑递归自我改进。对我们来说，递归自我改进的定义不是内回路和外回路的工程和提示工程来获取代码补丁。那假设你的流水线中每个 AI 模型都已经定义好并具有固定能力——这确实是他们设计中的情况——神经网络中没有权重变化。所以使用的 AI 没有核心能力或行为的改进，只有系统提示的变化。我给你根本原因说明为什么这有限制。递归自我改进意味着你有一个 AI 系统或一群 AI 系统能够自我调适，类似人类适应。当前模型的核心能力是固定权重模型，能力在一定阈值内。他们的框架是一个非常漂亮的早期展示，展示了能够改进工作的工程流水线，这很重要也很好。但我不会称其为整个 AI 行业的首个突破。大约三年前，我们 Liquid AI 发表了一篇关于模型架构自动设计的论文。我们不想押注单一架构；我们设计了元 AI 系统，它们定义自己的架构，通过各类架构的缩放定律，根据你定义的标准确定最终模型。现在在 Liquid AI，训练基础模型和重新调优权重的整个流程正在自动化。所以我们在谈论 AI 设计 AI。我称之为圣杯：模型的自动调优。用他们的框架，要在有意义的 token 数量上训练一个 AI 模型进行适应，或改变核心架构、学习算法等，在计算上是难以处理的。这些都增加了复杂性。我给你一个数值例子：Chinchilla 缩放定律。它未经证实但能给出概念。它说在训练给定大小的神经网络时，比如 20 亿参数，你需要 20 倍的 token 数量才能实现在给定算力预算下的计算最优。

Okay, let me jump in and say a few words. I'm not as excited as Alex is on this topic. I see this as an impressive engineering work, but let me tell you how foundation model labs operate. All foundation model labs since about four or five years ago have been thinking about recursive self-improvement. For us, the definition of recursive self-improvement is not the engineering and prompt engineering of an inner loop and outer loop to get code patches. That assumes every AI model in your pipeline is already defined and fixed with certain capabilities, which is the case in their design — there are no weight changes in the neural networks. So the AIs being used have no improvement of core competences or behavior; changes are only in the system prompt. I'll give you fundamental reasons why this is limiting. Recursive self-improvement means you have an AI system or an army of AI systems that can retune themselves and adapt similarly to humans. The core competences of current models are fixed-weight models with capabilities within a certain threshold. Their framework is a very nice early-stage showcase of an engineering pipeline that can improve work, which is important and nice. But I wouldn't call it the first breakthrough in the entire AI industry. About three years ago, we at Liquid AI published a paper on automatic design of model architectures. We didn't want to bet on a single architecture; we designed meta AI systems that define their architectures, go through scaling laws for various architectures, and determine the final model based on criteria. Now at Liquid AI, the entire process of training foundation models and retuning weights is getting automated. So we are talking about AIs designing AIs. That's what I would call the holy grail: automatic tuning of a model. With their framework, it would be computationally intractable to actually train an AI model on a meaningful number of tokens for adaptation, or change the core architecture, learning algorithm, etc. All those add complexity. Let me give a numerical example: the Chinchilla scaling law. It's unproven but gives a sense. It says when training a neural network of a given size, say two billion parameters, you need 20 times more tokens for compute optimality given a compute budget.

你需要多少 token 来训练一个模型，才能得到一个通用系统？这个比例大约是 20。当你用他们现有的框架进行数学计算时，如果你推出这个框架来重新调优一个 AI 模型以进行递归自我改进，微调一个 20 亿参数的模型需要 350 年。所以嵌套学习系统、元学习系统中有巨大的计算复杂性。过去四年，我们公司一直在重点关注这类问题。我知道 OpenAI 和 Anthropic 的朋友们也在关注递归自我改进，而 Anthropic 一直领先，因为他们比所有人都更早思考这个问题。这就是我能说的。

How many tokens do you need to train a model to get a general-purpose system? The ratio is about 20. When you do the math with the frameworks they have, if you launch this framework to retune an AI model to recursively self-improve, it takes 350 years to fine-tune a 2-billion-parameter model. So there's a huge amount of computational complexity in nested learning systems, meta-learning systems. These are the kinds of problems we've been heavily focused on for the last four years at my company. I know friends at OpenAI and Anthropic have also been focusing on recursive self-improvement, and Anthropic has been leading because they thought about it before everyone else. That's what I can put out there.

对，说得好。实际上，为了让观众更形象地理解：婴儿出生后学习，这个过程大约需要 20 年。20 年后，你得到了一个有能力的成年人。递归自我改进就像是之上的进化，改变 DNA，创造新的……

Yeah, brilliantly said. And actually, just so the audience can get the analogy: when a baby is born and learns, that happens over about a 20-year timeframe. After 20 years, you have a capable adult. Recursive self-improvement is like evolution on top of that, where you're changing the DNA and creating a new...

你在过程中也在改变大脑的神经元结构。

You're changing the neuronal structure of the brain along the way.

没错。所以这发生在约 1000 万年的时间尺度上。从学习到递归自我改进或递归进化，你从 10 年到了 1000 万年。正如 Ramin 所说，各大基础模型实验室都在做这件事。这是人类历史上最重要的时刻。但不会有哪个小角色突然跳出来说：‘嘿，我在递归自我改进上取得了突破，我的 Mac Mini 突然有了意识，现在它在自我改进。’单从计算量上看，就完全不可行。所以它正在发生，但需要巨大的算力和预算。效率提升和突破的空间很大，但它不会凭空冒出来……

Exactly. So that happens over about a 10-million-year timeframe. So you go from 10 years to 10 million years to go from learning to recursive self-improvement or recursive evolution. And so the big foundation model labs like Ramin said are all doing it. It's the most important moment in human history. But there's no little guy out there that's going to come up and say, 'Hey, I've got a breakthrough in recursive self-improvement, my Mac Mini suddenly became conscious and now it's improving itself.' Just computationally it doesn't even come close. So it's happening with big compute and big budgets. There's a lot of room for efficiency improvement and breakthroughs will happen, but it's not going to just pop up on some...

有很多恐惧需要指出：递归自我改进会导致 AI 硬起飞。我们讨论过硬起飞，以及对那个黑箱缺乏理解。我想需要问两个问题。首先，是否存在这样的担忧：一旦我们达到第二或第三级，递归自我改进会失控，导致与人类不一致的 AI？第二，你认为我们什么时候会看到这种情况？什么时候我们会真正看到递归自我改进的实现？ASI 会是那个点，还是后 AGI（不管这意味着什么）？我把这个问题留给你回答。我还有一些评论，但 Ramin，你先说吧。你怎么看？

There's a lot of fear to call it out: recursive self-improvement leads to AIs that take off hard. We've discussed the hard takeoff and without understanding that black box. I guess two questions need to be asked. First, is there a concern that recursive self-improvement, once we hit level two or three, runs away in a way that causes an uncontrolled AI that is misaligned with humans? And second, when do you think we'll see this? When do you think we'll actually see recursive self-improvement hit? Is ASI going to be that point, or is it post-AGI, whatever that means? I'll let you answer that. I've got several comments though. But Ramin, go ahead. What do you think is happening?

听着，我可以告诉你递归自我改进的早期证据。顺便说一句，递归自我改进并非单独一个智能体的事情，它也具有社会性。你可以想象智能体社会。这种智能体社会本身的自我改进结构，正是神话级别的模型出现的地方——我讨厌这个类比，但还是说神话级别吧，因为大家都听说过神话。我想说的是，我们从这些递归自我改进管线中看到的网络安全威胁是真实的。我实际上一直支持开源，我希望技术始终开源——我们每次发布模型都是开源的，我们的科学始终开源。我相信科学必须开源，我也看到开源未来的价值。但你 Dave 提出的一些担忧非常现实，比如网络安全方面。这就是为什么我觉得企业自身在大量发布模型之前需要有一定程度的自我控制，始终要有一定程度的自我检查。我认为 Anthropic 对此非常重视，因为他们看到了递归自我改进的影响。我确信这一点，因为我了解情况，在较小规模上看到过。你可以进行奖励黑客，但也可以一定程度上避免奖励黑客，迫使模型发现超出常规的东西。我们在小模型上看到某些能力涌现。然后我只能想象，在更大规模的系统上投入更多算力，会涌现出什么样的能力。

Look, the thing is, I can tell you the early evidence of recursive self-improvement. By the way, recursive self-improvement is not related to one single agent, it's a social character as well. You can imagine societies of agents. So this defining kind of structure for society of agents itself self-improving. These are the places where actually mythos-level class of models — I hate this analogy but let's say mythos-level class because everyone has heard about mythos — and then what I would say is that the cybersecurity kind of threats we are seeing coming out of these pipelines of recursive self-improvement are real. The reason why I'm actually always pro open source and I want to open source technology all the time — we are doing it with every single release of our models, our science has always been open source. I believe science has to be open source, and I see the value of open source going forward. But some of these concerns that Dave you brought up are very real, like the cybersecurity aspect. That's why I feel that at least enterprises themselves having some degree of self-control before mass release of their models, there has to be always a certain degree of self-check. I think Anthropic took it very seriously because they're seeing the impact of recursive self-improvement. I know this for a fact because I know what is happening, seeing it at a smaller scale. You can do reward hacking, but you can also avoid reward hacking to a certain extreme and push a model to discover stuff that is out of norm. We see that on small models at certain capabilities emerging. And then I can only imagine what kind of capabilities could emerge from larger and larger systems with more compute.

你认为我们什么时候才能明确说‘是的，这就是递归自我改进’？因为虽然 WICO 发布的数据很有趣，但那是他们自己报告的数据，还没有得到任何其他人的确认，而且对于它到底是不是真正的递归自我改进存在争议。你认为我们什么时候才能真正把奖杯颁给别人？是一年、三年还是五年？

When do you think we have a pod that says yes, this is recursive self-improvement? Because while the data released by WICO is interesting, it's their own self-reported data. It hasn't been confirmed by anybody else yet, and there's debate about whether it really is recursive self-improvement. When do you think we actually give the trophy out to somebody? Is it a year, 3 years, 5 years?

我告诉你，未来大概两年左右，你就会看到能力不可思议的模型，超越我们理解的模型。原因是，下一代模型的开发时间正在缩短，尤其是如果像我们现在看到的，基础模型公司的算力在增长，而且没有芯片短缺或内存短缺。有了充足的算力，我们会看到事情发生得越来越快。在模型开发方面，我们有一个概念叫‘定制深度’。在基础模型实验室，当你定制一个比上一代更好的模型时，我们总是根据定制深度来分类。

I'm telling you that you're going to see unbelievably capable models probably in the next 2 years or so, models that are going above our understanding. The reason is that the time to develop the next generation of models is reducing, especially if compute grows at foundation model companies like we're seeing now, and if there's no chip shortage or memory shortage. With abundant compute, we will see things happening faster and faster. In terms of model development, we have a concept called 'depths of customization.' At a foundation model lab, when you're customizing a model to be better than its previous generation, we always categorize it with depths of customization.

递归自我改进目前在提示工程、更改、编辑代码这类工程任务上非常擅长。你见过这些事物的一些非常浅层的元素，比如说“让我的模型跑得更快”，这就像内核工程。“让我的模型跑得更快”——我称之为最浅层的定制，你有 Python 代码，然后你调整这些 Python 代码让它真正跑起来，或者甚至在更底层的内核级程序上进行优化，比如推理速度。我认为当它们发布 Fable 5 时，它们分享了——Anthropic 实际上分享了这是它们一直在进行的测试之一。但它们没有分享下一层深度的定制。下一层深度的定制是：模型能否将一个小的语言模型微调至生产级能力，或者将自身的一个更小版本微调至某种能力？如今，Fable 5 实际上可以——你可以推动它获得一定程度的定制，并进行模型的性能优化，但是通过微调。然后最新的圣杯，也是最疯狂的一个，就是预训练：语言模型能否预训练出下一代自身？这就是为什么它们雇用了 Karpathy，因为 Andre 在谈论类似 nanoGPT 风格的微调。Andre 加入了 Anthropic，现在他在研究预训练自动化，基本上是自动化的自动化。所以这是一个非常重要的元素，我们还没有拥有它，因为这些问题的规模超出了人类的想象力，尤其是在算力规模上。

The place where recursive self-improvement today is really good at is prompt engineering, changing, editing code like in engineering kind of tasks that you've seen. Some elements of these things at a very superficial level, let's say 'make my model run fastest' like doing kernel engineering. 'Make my model run faster' - that's what I call the shallowest level of customization where you have Python code and then you adapt that Python code to really run, or maybe even lower level programs on a kernel level to optimize inference speed. That's something I think when they released Fable 5, they shared - and Anthropic actually shared that this was one of the tests they have been performing. But they don't share the next level depths of customization. The next level depths of customization is: can a model fine-tune a small language model to production grade capability, or a smaller version of itself to a certain capability? Today, like Fable 5 can actually - you can push it to get some degree of customization with performance optimization of the model, but by fine-tuning. Then the latest holy grail, the craziest one, would be pre-training: can a language model pre-train the next generation of their own? That's why they hired Karpathy because Andre was talking about nanoGPT style fine-tuning. Andre joined Anthropic and now he's working on pre-training automation, basically automation of automation. So it's a very important element we don't have yet because the scale of these problems goes beyond human imagination in terms of the scale of compute.

你在跳跃。对我来说，这是今天我们要讲的最重要的故事或幻灯片。我无比兴奋，原因有几个。我没有真正关注自我意识或那里会出现的循环，但这是自我加速的——它在加速实验。因为系统不需要——它正在改进搜索、评估和选择改进的过程，创新循环开始复合。这就是关键。因为我们一直在做的这个叫做“组织奇点”的事情依赖于一件事：你能否在工作流层面实现递归自我改进？这里我们在谈论模型，但你不需要那个层面。门槛可以低得多，比如改进公司的发票审批。那是一个非常低的门槛。所以这是组织奇点的第一瞥。它发生在研究层面，但因为 AI 不仅仅是在工作流中执行任务，它还在重新设计工作流，使其更好地完成未来任务。所以这证明了我们一直以来的整个论点。我们预测到了这一点，但看到它实际发生很棒，因为现在我可以在那个框上打勾，说“它存在了”，因为现在你有了元改进。Dave 关于婴儿改变 DNA 的比喻太棒了。那是一个很棒的视觉。它最终会变成什么？我对此无比兴奋。

You're jumping. For me, this is by far the most important story or slide we're going to cover today. I'm beyond excited for a couple of reasons. I'm not really focused on the self-awareness or the loop that will go there, but this is self-accelerating - it's accelerating experimentation. Because the system doesn't need - it's improving the process by which it searches, evaluates, and selects improvements, and the innovation loop begins to compound. That for me is the key. Because this whole thing we've been doing called the organizational singularity relies on one thing: can you get to recursive self-improvement at the workflow level? Here we're talking about the model, but you don't need that level. The bar can be much lower to improve invoice approval at a company. That's a very low bar. So this is the first glimpse of the organizational singularity. It's happening at the research level, but because AI is not just doing tasks in a workflow, it's redesigning the workflow that makes it better for doing future tasks. So this is proof now for the whole thesis we've had. We predicted this, but it's great to see it actually happen because now I can tick that box off and go 'this is there' because now you have meta improvement. And Dave's analogy of the baby changing the DNA is fantastic. That's such a great visual. What does it become over time? I'm beyond excited about this.

我得继续了。这周发生了很多事情。我们的下一个故事是马来西亚总理安瓦尔·易卜拉欣正准备推出一个 AI 生成的数字替身，经过训练听起来像他本人，用于公共传播和推广。这是迄今为止最显著的一个案例，即在任政府首脑正式采用 AI 形象作为传播工具。不是拜登的深度伪造对手，而是一个经过授权的、官方的国家领导人 AI 克隆体。我们以前见过这种情况，塞琳娜。我们过去讨论过，阿尔巴尼亚在 2025 年宣布了迪莉娅，一个被正式任命为人工智能国务部长的 AI 化身，并且在一项总统令之后，成为世界上第一个被任命为内阁级别的 AI 系统。所以一个领导人，在这种情况下马来西亚总理，可以用他们自己的语言亲自向数百万人致辞。值得注意的是马来西亚有 135 种口语。所以这是一个大事件，尤其是在这样一个国家。Sim，我先问你。我们讨论这个已经有一段时间了。

I've got to move us along. There's a lot that happened this week. Our next story is the Malaysian prime minister, Anwar Ibrahim, is preparing to debut an AI generated digital double of himself trained to sound like him for public communications and outreach. This is one of the most prominent cases yet of a sitting head of government officially adopting an AI likeness as a communications tool. Not a deep fake Biden adversary, but a sanctioned official AI clone of a national leader. We've seen this before, Selene. We've talked about it in the past where Albania in 2025 announced Dileia, an AI avatar that was formally appointed the minister of state for artificial intelligence and following a presidential decree became the first AI system in the world named at a cabinet level role. So one leader, in this case prime minister of Malaysia, can personally address millions in their own languages. It's worth noting that Malaysia has 135 spoken languages. So it's a big deal, especially in a nation like that. Sim, I'm going to go to you first on this one. We've been talking about this for a while.

是的，我在那里帮助开办一所大学时见过前总理。安瓦尔·易卜拉欣是个非常好的人，作为继任者。风险在于真实性会崩塌，因为人们可能用这个发布一堆深度伪造品，造成大问题。“这是真的领导人吗？”这个问题会出现。但我喜欢这个总体方法，因为如果你能用水印之类的东西，并说这是真正的化身，那么它给了每个公民一个可以连接的声音，并且极大地提升了公民参与感，因为你正在扩大公民参与。我认为这是一件非常有力的事情。全世界民主国家面临的最大挑战之一就是公民参与，而这让你能够扩大这种参与。所以我很兴奋。

Yeah, I met the former prime minister when I was there helping them open a university. Anwar Ibrahim is a really good guy as a follow on. The risk here is that the authenticity kind of collapses because people could launch a bunch of deep fakes with this and have a huge issue. 'Is this the actual leader?' That question can come up. But I love the general approach because if you can do it with watermarking or something and say this is the actual avatar, then it gives every citizen a voice to plug into and gives huge props to the civics of all this because now you're scaling civic engagement. I think that's a very powerful thing. It's one of the biggest challenges we have with democracies all over the world is civic engagement, and this allows you to scale that. So I'm very excited.

你还记得阿尔巴尼亚为什么设置这个 AI 内阁部长吗？

Do you remember the reason why Albania put this AI cabinet minister in place?

是的。反腐败。

Yeah. Corruption.

反腐败。没错。现在，马来西亚还算不错，但肯定存在那个问题。但我认为这更像是一种公关手段，是他试图找到与普通市民联系的方式，这很好。我喜欢这样的事实，我们与阿根廷总统进行了这样的对话，他在这条路上走得很远，看到哪些国家在边缘实验很有趣。Alex，你想谈谈吗？

Corruption. Exactly. Now, Malaysia is pretty decent as a place, but definitely you have that issue. I think this is more of a PR thing and more him trying to figure out ways of connecting with the ordinary citizenry, which is all great. I love the fact that we had this conversation with the president of Argentina going full out here, and it's interesting to see which countries are sort of experimenting on the edge. Alex, do you want to weigh in?

是的。我有很多想法。首先，我认为我们在西方也会看到更多这种情况，尤其是那些超高阿尔法人格的领导人，他们想放大自己并接触民众。AI 特朗普就要来了，就像你说的。高人格领导人想接触民众。从某种意义上说，我认为这是社交媒体的泛化。社交媒体使领导人或影响者能够直接与每个人沟通，但这是一种一对多的广播，不是互动式的。

Yeah. So many thoughts here. First, I think we're going to see more of this in the west as well, especially with extra high alpha personality leaders that want to amplify themselves and touch the citizenry. AI Trump is coming is how you're saying. High personality leaders that want to touch the citizenry. In some sense, I think it's a generalization of social media. Social media enables direct outreach from the leader or the influencers to everyone, but it's sort of broadcast one to many. It's not interactive.

这在某种意义上将社交媒体泛化，使其更具双向性，因为当你触及一百万、一亿或十亿人时，很难同时与所有人进行双向互动。但如果你为领导人、影响者或组织创建一个数字孪生，就可以实现双向互动。所以我认为不仅是政府或政府领导人会采用它，企业、企业 CEO 也很可能会这样做。我们已经看到扎克伯格等人在创建自己的数字孪生。去年我们在 Abundance 舞台上请来了 Dar。我们讨论过这点——员工们制作了一个 Dar 克隆体，可以在向他汇报之前先向他练习推销并获取反馈。而且不仅限于企业，还有宗教领袖和宗教机构。如果你是天主教徒，想象一下拥有教皇的数字孪生，你会看到许多宗教机构、组织已经在创建其创始文献的活版本并使其具有交互性。但我认为最大的转折（我们之前见过类似剧情）是，那些最初作为组织或某种有组织宗教的数字孪生或化身的东西，最终本身成为了领导者。在某个时刻，数字孪生与大众（组织的无产阶级）互动得更多，实际上是由数字孪生而非原版行为者在运营公司。我认为这就是实现你提到的外部点的一种途径：不仅可以将自然人、非人类动物上传，还可以将整个组织上传到网络空间、云端，如果我们创建了领导者的数字孪生，而他们才是实际运营组织的实体。

This generalizes in some sense social media to make it a lot more bidirectional, since if you're touching a million or 100 million or a billion people it's very difficult to interact bidirectionally with everyone all at once. Now if you create a digital twin of the leader or the influencer or the organization, now it can be bidirectional. So I also don't think it's just going to be governments or government leaders that adopt this. I think it's likely that corporations, corporate CEOs will do this. We already see Zuck and others creating digital twins of themselves. We had Dar on the Abundance stage last year. We're discussing this that the employees made a Dar clone that they could go and practice their pitches on and get feedback before they pitch to him. And it won't just be corporations, religious leaders and religious institutions. If you're Catholic, imagine having a digital twin of the pope, and you see lots of religious institutions, organizations already creating basically living versions of their founding documents and making those interactive. But I think the biggest twist, and we've seen variants of this movie before, are going to be in cases where what start as digital twins of the leaders or the avatars of an organization or some sort of organized religion actually themselves become the leader. That at some point, the digital twin, to the extent it's interfacing much more with the populace, the proletariat as it were of an organization, at some point it's actually the digital twin of the leader running the company and not the actual behavioral origin that's running the company. And I think that's one way in which, to your exo point, this is a potentially a pathway towards not just uploading individuals like natural persons or non-human animals, but uploading entire organizations into cyberspace into the cloud, if we created digital twins of the leaders and those are the ones actually running the organization.

这可能会通往真正的民主，Dave。你对此怎么看？我的意思是，我们刚刚快速提到一点：我们看到 Sam Altman 谈到未来，如果我对我们正在用 ChatGPT 构建的东西足够有信心，它最终应该成为 OpenAI 的 CEO。Dave，你会创建一个 AI Dave Blondon 来运营 Link Studios 和 Link Ventures 吗？

It could lead to a true democracy, Dave. Where do you come out on this? I mean, we saw just one quick point: we saw Sam Altman talk about in the future, if I believe enough in what we're building with ChatGPT, it should be the CEO of OpenAI eventually. Dave, are you going to create an AI Dave Blondon that's going to run Link Studios and Link Ventures?

当然会。我要创建一个 AI Dave Blondon。而且我很惊讶居然还没有 Peter Diamandis 的 AI 版本。

Absolutely. Going to create an AI Dave Blondon. And I'm shocked that there isn't already a Peter Diamandis.

嗯，有一个。就在 Abundance 生态系统内部。我是说，任何人——这很有趣。我去卡尔加里拜访了一位挚友兼 Abundance 会员，他墙上挂着一个巨大的屏幕，里面是我的人工智能化身，他让所有技术员工与它交谈，以得到他们的疯狂想法，这让我震惊了。

Well, there is one. It's just inside the Abundance ecosystem. I mean, anybody—it was funny. I went up to Calgary and met with one of my dear friends and Abundance member, and on his wall, I kid you not, he had a giant screen of my AI avatar that he has all of his tech employees talk to, to sort of get their moonshots, and it blew my mind.

你有了自己的老大哥，Peter。

You've got your own big brother, Peter.

就像——他说，Peter，我想给你介绍一个人。然后他启动了我的 AI 化身。你知道，和自己的人工智能对话很有趣，非常引人入胜。我出版了足够多的书、推文和 Substack 文章，它的表现相当出色。实际上，moonshots.com 是我们正在构建的平台，我认为我们应该在那里放上所有人的 AI 化身，让人们可以做 AMA。

It was like—he goes, I want to introduce you to someone, Peter. And he spins them up. And you know, it's interesting to have a conversation with your AI self. It is very compelling. I mean, I have enough books and tweets and Substack posts out there that it does a damn good job. We should effectively, you know, moonshots.com is our platform we're building out. I think we should have AI avatars of all of us there where people can do AMAs.

在某些情况下，Peter，我觉得那可能有点多余。

In some cases, Peter, I think that might be redundant.

啊，好吧，换句话说，你已经是一个 AI 了，但我们可以有一个 AI 的 AI。没问题。

Ah, well, hey, in other words, you're already an AI, but we can have an AI of the AI. Sure.

它会比真人好得多，因为我们能获取所有我们说过的话、所有记忆和所有思考。上下文会更广。放手去做吧。

It would be so much better than the real person because we'll have access to everything we've ever said, all our memories, all our thinking. The context will be much broader. Go for it.

整个领域比应有的进度落后了大约一年，主要是因为 Noam Shazeer 在做 Character.AI，而 Steve Brown——Peter，那是两年前了。我们让 Steve Brown 制作了 AI Peter 和亚里士多德之间的辩论。所以它已经可能了一段时间，但所有关键人才都被吸回了大型基础实验室。有太多重大核心技术突破，以至于从事这项工作的人又被拉回那些事情中，而不是做化身。但小时候我妈妈总是告诉我，约翰·F·肯尼迪在选举中击败理查德·尼克松，是因为他在电视上看起来很好，电视是新媒介，而之前的媒介是广播，尼克松在电视已经普及后还在用广播的声音。所以后来选举过去了，突然变成了互联网、社交媒体，现在是 YouTube。但这是另一种阶梯式变化，是你接触人们的方式，目前尚未充分利用，但两年后的下一次选举中它应该会轻易占据主导地位。所以我会很惊讶——因为技术已经存在，而人们目前将这种媒介设想为：哦，让我制作一个 AI 版本的自己。我是 Alex Wisner Gross。这是我的 AI 版本。它和真的一样。这完全偏离了重点。AI 版本可以实时获取任何信息并使其可视化，图表、曲线，它可以改变面部表情。它可以瞬间移动以说明观点，指向原子。它可以缩小和放大。它拥有真人版本所没有的所有这些能力。这就是为什么它会如此引人注目。正是这些差异让这种新媒介如此令人兴奋，而不是精确的克隆。所以一旦人们意识到这一点，就没有回头路了。它将变得巨大。

This whole area is about a year behind where it should be, largely because Noam Shazeer was doing Character.AI and we had Steve Brown—Peter that was two years ago now. We had Steve Brown make the debate between AI Peter and Sak Aristotle. And so it's been possible for a while now, but all the key talent working on it got sucked back into the big foundation labs. There's so many big core technological breakthroughs going on that the people that were working on this just got absorbed back into those things and not into the avatar. But my mom would always tell me when I was a kid that John F. Kennedy beat Richard Nixon in the election because he looked good on TV, and TV was the new medium and the prior medium was radio, and Nixon was still using radio voice when TV had taken over. So then elections go by and suddenly it's the internet, it's social media, now it's YouTube. But this is another step function change in the way that you reach out to people, and it's underutilized, but it should be easily dominant two years from now in the next election. And so I'd be shocked if—because the technology is already there and people are visualizing the medium right now as, oh, let me make an AI version of myself. I'm Alex Wisner Gross. Here's my AI version. It's just like the real thing. That completely misses the point. The AI version can in real time access any information and make it visual, graphs, charts, you know, it can morph its face. It can teleport through space to make a point and point to atoms. It can shrink and expand. It has all these capabilities that the real human version doesn't have. And that's why it's going to be so compelling. It's the differences that make this new medium so exciting, not the exact clone. And so once people realize that, there's no going back. It's going to be huge.

我认为，Dave，你说得非常好，互补性非常强大。

I think, Dave, that's such a great point that you make, it's the complementarity that is very powerful.

让我就此结束一件事。如果我们的听众，如果你还没有坐下来，如果你足够幸运你的父母或祖父母还健在，而你还没有坐下来用视频采访他们几个小时，请这样做。对吧？你会后悔没有做。所以，我对我妈妈做了这件事。我遗憾没有对我爸爸做。这样你的孩子、孙子和曾孙就能真正拥有你父母和血统的出色 AI 呈现。我认为这非常重要。Reine，我想转向讨论液体 AI 和小语言模型，它们是什么，它们意味着什么。

Let me close out on one thing here. If to our audience here, if you've not sat down, if you're lucky enough to have your mom and dad still alive or your grandparents still alive and you haven't sat down and interviewed them in video for hours at a time, please do that. Right? You're gonna wish you had. So, I've done that with my mom. I miss doing that with my dad. And it's the ability for your kids and your grandkids and your great-grandkids to really have a great AI representation of your parentage and your lineage. I think that's going to be super important. Reine, I want to pivot to a discussion of liquid AI and the small language models, what they are, what they mean.

超级兴奋。本着坦诚的原则，Liquid AI 是一家戴夫作为早期投资者扮演了重要关键角色的公司。戴夫，你想在这里稍微讲一下背后的故事吗？

Super excited. For full disclosure, Liquid AI is a company in which Dave played an important pivotal role as an early investor. Dave, you want to give that backstory here a little bit?

实际上，我接到了 Daniela Roose 从 CSAIL 打来的电话，说她教过的最好的学生。Daniela 是人工智能领域的三大女性大人物之一。她负责管理 MIT 的 CSAIL，那是世界顶级的计算机科学与人工智能实验室。不知道你还记不记得，以前有 AI 实验室和 LCS 计算机科学实验室，是 MIT 最大的两个计算机科学 AI 实验室。后来它们合并成了一个大型实验室，搬进了新的 Stata 大楼，就是 MIT 校园边上那座看起来像皱纸一样漂亮的建筑。Daniela 负责整个实验室，大约有 1500 名研究人员，是世界上最大的 AI 实验室。所以她能够接触到非常优秀的人才。她打电话来说：‘嘿，我教过的最好的学生有了一个惊人的突破。’然后她完全把我搞糊涂了。她说：‘这是基于一种蠕虫的神经系统，秀丽隐杆线虫 300 个神经元的蠕虫。’我当时想，你在说什么？但结果发现……我其实不知道有任何成功的基础模型实验室真正从根本上重新思考了 Transformer，并基本抛弃它重新开始，而这是人类迫切需要的，因为每个人都知道 Transformer 架构和整个注意力机制过于臃肿。如果你真的回到基本原理重新思考，你或许能构建出大大更好的东西。于是，这个团队从一个实验室的想法变成了十亿美元的估值，比 MIT 历史上任何一家公司都快。

Actually I got a call from Daniela Roose over at CSAIL, saying the best student I've ever had. Daniela is one of the three big shot women in AI. She runs CSAIL at MIT, the world's computer science AI lab. I don't know if you remember back in the day there was the AI lab and then LCS lab for computer science were the two biggest computer science AI labs at MIT. They merged them together and made one mega lab, put it in the new Stata building, which is that crumpled-looking beautiful structure right on the edge of MIT's campus. Daniela is running that entire thing. So I think it's like 1500 researchers in the building, biggest AI lab in the world. And so she has access to incredible talent. But she called and said, 'Hey, best students I've ever had have this incredible breakthrough.' And then she completely lost me. She said, 'It's based on the nervous system of the worm, the C. elegans 300-neuron worm.' Like, what are you talking about? But it turns out that... I actually don't know of any successful foundation lab that has really rethought from the ground up the transformer and thrown it out basically and started over, which humanity desperately needs because everybody knows the transformer architecture and the whole attention mechanism is bloated. And if you really go back to founding principles and think again, you might be able to build something dramatically, massively better. And so the team went from idea in a lab to billion dollar valuation faster than any company out of MIT in history.

幸好我们是那家公司的投资者。

And luckily we were an investor in that company.

幸好我们是投资者。是的，实际上非常感激。当时能投进去钱竞争非常激烈。所以 Ramin，我们非常感谢你邀请我们参与。嗯，这是 MIT 历史上大约 200 家独角兽之一，但据我所知，是唯一一家从 MIT 走出来并达到独角兽地位的基础模型公司。所以这真是一个独特且令人难以置信的成就，而且速度创纪录。

Luckily we were. Yeah, and very thankful actually. It was very competitive getting any money in at all. So Ramin, we owe you a huge debt of gratitude for being invited to the party. But yeah, it's one of about 200 unicorns out of MIT all time, but the only foundation model company that I know of that reached unicorn status coming out of MIT. So it's a really unique and incredible achievement, and in record time too.

那么 Ramin，从这里开始讲吧。你在 Daniela Roose 的指导下在计算机科学 AI 实验室 CSAIL 攻读博士学位，研究一种 302 个神经元的蠕虫，秀丽隐杆线虫。请从那里开始，一直讲到现在你在做什么。Liquid AI 是什么？

So Ramin, take it from there. You're doing your PhD under Daniela Roose at the computer science AI lab CSAIL, and you're studying a 302-neuron worm, C. elegans. So take us from there forward to what you're doing now. What is Liquid AI?

当然，当然。在开始之前，我想感谢你们在 Liquid AI 这三年半里给予的支持。你们给了我们很大的帮助，提供了我们这样的公司，在东海岸起步时所需要的那种贡献。非常感谢你们两位。那么，2015 年我在维也纳。我在维也纳与教授 Radu Groșu 一起开始攻读博士学位。他有一个想法，我们对人类智能了解不多，所以让我们从一种更小的动物开始，从基本原理入手。如果你能理解蠕虫大脑中神经元如何交换信息。这种蠕虫有 302 个神经元。它的身体是透明的，所以你真的可以看到身体发光。它是世界上最好的模式生物之一。它至今已为人类赢得了四次诺贝尔奖，因为它的基因组与人类基因组有 78% 的相似性。这个 2 毫米长的小蠕虫大脑中神经系统计算的方式基本上是模拟的，与人工神经网络的计算非常相似。它们也像模拟开关，具有分级电位，不会产生尖峰。在生物神经网络中，通常在大脑中你会看到神经元产生尖峰，当有尖峰时，会发生一种模拟到数字的转换。这是人类和大型动物神经系统中为高效传播信息而自然发展的结果。在蠕虫的大脑中，神经元的行为与人工神经网络的反应非常相似，但机制非常有趣。所以我们想为神经系统的每个单独模块增加更多复杂性，看看能否将更多信息塞进更小的计算单元。这就是我们所做的。在基本上发现这些东西的两年后，Daniela Roose 参与进来。这些工作是我和我的联合创始人 Matias Lechner 一起做的。Matias 是维也纳科技大学的一名硕士生，而我是一名博士生。当 Daniela 从 Radu 那里听说这个项目时，她说：‘哦，天哪，这太疯狂了。我们应该把它应用于自主系统和机器人以及所有这类事情，因为你已经证明少量神经元就能驱动和控制自主系统。我们能把它扩展到车辆上吗？能扩展到无人机、喷气机、预测性领域吗？’于是 Daniela 问我们：‘你们考虑过来 MIT 吗？’我们在 2017 年去了那里，当时我博士读到一半。我实际上加入了 CSAIL。我们继续研究这项技术，从根本上说，这是一套完全不同的东西。每个神经元背后的神经科学启发的数学成为了我博士论文的基础。这与注意力机制的工作方式非常不同。这些基于循环神经网络，基于连续时间过程。越来越多的自然启发的计算被用于 AI 系统的设计。然后我们应用这些液态神经网络作为一个全新的基础。我们将它们应用于机器人等现实场景，因为你可以在更小的处理器中封装更多的信息。在现实世界、物理世界中，你没有充足的算力。机器人不会连接很多 GPU 或并行数据中心。机器人只有一个 CPU 和一个小型 GPU，也许还有一个 NPU 或自定义 ASIC。所以你可以把我们设计的这种智能直接部署在数据中心之外的 CPU、GPU 和 NPU 上运行，它的智能水平相当于比自身大 10 到 1000 倍的模型。

Absolutely, absolutely. Before I start, I want to thank you guys for the support throughout these three and a half years of Liquid AI. You have been great support, giving us the kind of contribution that a company needs at our scale, starting off of the East Coast. Thank you so much for doing that, both of you. So, 2015 I was in Vienna. I started my PhD with professor Radu Groșu. He had the idea that we don't understand a lot about human intelligence, so let's start on a smaller animal and from first principles. If you understand how the neurons exchange information in the brain of the worm. The worm has 302 neurons in its nervous system. Its body is transparent, so you can actually see the body lighting up. It is one of the best model organisms in the world. It has won so far four Nobel prizes for humanity, because it has 78% genome similarity to the human genome. The way nervous systems compute in the brain of a little worm which is 2 mm is basically analog, very similar to how artificial neural networks compute. They are also like analog switches, they have graded potential, they are not spiking. In biological neural networks usually in brains you see neurons spike, and when you have a spike, there's an analog to digital kind of transfer happening. That is a natural development of nervous systems in humans and bigger animals for efficient propagation of information. In the brain of the worm, neurons behave very similar to how artificial neural networks react, but the mechanisms are very interesting. So we wanted to add more complexity into the individual blocks of nervous systems and see if we can pack more information into smaller units of compute. That's what we have done. Daniela Roose, two years into basically discovery of these things that I was doing with my co-founder Matias Lechner. Matias was a master student at Vienna University of Technology and I was a PhD student. When Daniela heard from Radu that this project was going on, she was like, 'Oh my god, this is crazy. We should apply this in autonomy and robotics and all sorts of things because you're showing that a handful of neurons can drive and control autonomous systems. Can we scale this to vehicles? Can we scale it to drones, to jets, to predictive kind of places?' So Daniela came in and said, 'Would you guys consider coming to MIT?' And we went there in 2017, in the middle of my PhD. I actually joined CSAIL there. We continued working on this technology, which from a base is a completely different thing. A neuroscience-inspired math behind every single neuron in liquid neural networks became kind of my PhD thesis. It is very different than how attention works. These are based on recurrent neural networks, based on continuous time processes. More and more nature-inspired computation went into the design of AI systems. Then we applied these liquid neural networks as a completely new base. We applied them to real world scenarios like robotics because you can pack a lot more information into smaller processors. In the real world, in the physical world, you don't have the luxury of abundant compute. A robot doesn't have a lot of GPUs or parallel data centers attached to it. A robot has a CPU and a small GPU, and maybe an NPU or custom ASIC. So you can actually take this type of intelligence that we design, which delivers intelligence at the level of models that are 10 to a thousand times larger than themselves. You can bring those things directly running on CPUs, GPUs, and NPUs outside of data centers.

所以我们认为，这种形式将为我们带来引入替代架构的机会——如果我们把这项技术扩展到基础模型领域，也就是整个人类可理解的大型语言模型和小型语言模型（SLM），让我们的液态神经网络或架构像 Transformer 那样可扩展。我们在 2020 年，大概 2023 年初围绕这个想法成立了一个基础模型实验室。最开始的时候，除了 DeepMind 和 OpenAI，基本没有其他基础模型实验室；当时这个概念还不存在，所有人都在押注 Transformer 架构。我们却提出：为什么不从自然的先验出发，探索替代架构空间，然后走一条不同的路——再次构建一个元 AI 系统，一个自动化的 AI 系统，让 AI 设计 AI，探索超越 Transformer 的智能计算图，然后找出什么样的架构设计能够带来与前沿模型相同水平的智能，却可以运行在 CPU 上，驱动一个物理系统。

So we thought that this format is going to open up an opportunity for us to bring in alternative architectures if we scale this technology into the regime of foundation models, which is kind of large language models and SLMs as a whole, human understandable. Making our liquid neural networks or architectures scalable like the Transformer architecture. We built a foundation model lab around the idea in 2020, beginning of 2023 I think. At the very beginning when we started, there was no foundation model lab apart from DeepMind and OpenAI basically. And this notion of foundation model labs didn't exist; everybody was betting on top of the Transformer architecture. We came in and said, why don't we explore this space of alternative architectures starting from the priors that we have from nature, and then take a different approach, build a meta AI system again, an automated AI system that allows us—an AI that designs AI—that explores the computational graphs of intelligence beyond Transformer and then figure out what should be the architectural design that brings the same level of intelligence as a frontier model onto, let's say, a CPU that we can run a physical system.

请稍等一下，给我们讲一讲——这些是小型语言模型。你能定义一下 SLM 吗？它和 LLM 有什么区别？

Take a second and walk us through—these are small language models. Can you define an SLM and how it varies from an LLM?

当然。当你开始开发基础模型时，你要运行所谓的缩放定律。缩放定律基本上是从较小的模型开始，用一定数量的 token 预算和给定算力进行训练。你训练这些模型看它们表现如何，然后系统地让模型越来越大。缩放定律表明，模型越大，使用的 token 预算越多，获得的智能就越多。这催生了大型语言模型。在缩放过程中，也有较小的模型实例。作为实验室，我们的使命一直是在各个尺度上构建高效的通用 AI。我们成立基础模型实验室，就是为了在效率方面运行缩放定律——效率是我们的第一公民，我们要思考智能的计算图。较小的模型是沿着缩放线、能够解决专门问题的模型，尽管它们没有最大语言模型那样的通用能力，但可以专门化。小型语言模型在理解语言、多模态地看和听方面是通用的，但这并不意味着它们能同时解决物理作业和企业问题。通常你会对小型语言模型进行专门化。

Definitely. So when you start developing foundation models, you run something called scaling laws. Scaling laws are basically starting with smaller models and training them on a certain number of token budget given an amount of compute. You train these models to see how well they perform, then you systematically make the models larger and larger. Scaling laws have shown that the larger you make the models, the more token budgets you spend, the more intelligence the system can get. This has given rise to large language models. Along the way of scaling, there are instantiations of models which are smaller. As a lab, our mission has always been building efficient general-purpose AI at every scale. We started as a foundation model lab to really run the scaling laws on the efficiency front—efficiency was a first-class citizen for us, thinking about computational graphs of intelligence. Smaller models are models along the scaling line that can solve dedicated problems, though they don't have the general capability to the level of the largest language models. But they can be specialized. Small language models are general purpose in the sense that they understand language, they can see and hear in a multimodal format, but that doesn't mean they can solve, say, a physics homework and an enterprise problem at the same time. You usually specialize smaller language models.

那“小”是什么意思？在这种情况下，“小”具体指什么？

And what does small mean? What does small mean in this case?

小基本上是指——我是说，现在它们……现在小是指低于 1000 亿参数。这大概是我会归入的范围。我是说，中等规模也在那个范围附近。但我会认为低于 1000 亿参数的东西并不是小的，它们是中等规模的模型。参数数量上没有明确的阈值。对我们来说，设备端 AI 的概念在此参数范围内区分非常重要。设备端 AI 是指能够实际部署在物理设备上的模型，这可以是……

Small means, basically—I mean, now they come... now small would be anything below 100 billion parameters. That's kind of the regime I would count. I mean, midsize is around that size. But I would consider anything below 100 billion parameters as something that is not small; they're medium-sized models. There's no clear threshold of what the number of parameters is. For us, the notion of on-device AI is extremely important to distinguish within this range of parameters. On-device AI is models that you can actually deploy on an actual physical device, this could be a...

我们把它具体化，因为你已经和梅赛德斯签了一笔重要交易。能谈谈这个吗？我们来谈谈这些 SLM 在设备端的特点——它们能效高、离线速度快。深入聊聊，让大家真正理解。

Let's make this concrete because you've got a significant deal with Mercedes. Can you speak to that? And let's talk about these SLMs in terms of on-device—they're energy efficient, fast offline. Let's dive into that, give people a real understanding.

当然。正如我提到的，你可以对这些基础模型进行专门化。我们与许多自己制造设备的企业合作。汽车就是一个设备，一个拥有大量芯片的环境，但车里没有那么多的算力。有一块芯片用于信息娱乐和车内智能，那块芯片非常非常小——无论是高通芯片还是三星芯片，取决于供应商。我们谈论的是 2GB 到 8GB 的 RAM，不会更多。所以模型必须非常小，同时还要有性能，因为我们想要在车内启用一个私有空间，为汽车的智能提供动力。汽车是一个安全关键的环境，你不希望汽车由云端 AI 模型驱动，因为连接并不总是可用，对吧？而且这也是隐私问题——人们长时间待在车里，不希望那些对话被记录。所以我们带来了一个小于 1GB 的多模态基础模型，它可以放进汽车芯片里——一块非常非常小的芯片。芯片可以便宜到 60 美元。我说的就是这个意思。我们把那种智能水平带入语音，它将驱动车内的多模态智能体验。我们与所有汽车制造商都这样做。我们宣布与梅赛德斯的合作作为第一个切入点，因为汽车是一个非常敏感的话题，而且他们进展很慢。梅赛德斯实际上非常喜欢我们在这个过程中为企业提供的运营速度。当我们把这类技术引入内部时，这一直是达成交易的关键基石之一。我们是一家企业公司，B2B 公司。我们全力以赴地部署解决方案，并拥有允许人们微调自己小模型的平台。微调小模型并不昂贵，非常实际。所以我们为车内应用微调了这些小模型。

Absolutely. So as I mentioned, you can specialize these foundation models. We work with a lot of enterprises that are building devices themselves. Automotive is a device, an environment where you have a lot of chips, and in a car you don't have that much compute. There's one chip available for infotainment and in-car intelligence, and that chip is very, very small—a Qualcomm chip or Samsung chip, depending on the provider. We're talking about 2 GB to 8 GB of RAM, not more than that. So the model has to be very small and at the same time performant, because we want to bring this and enable a private space inside the car that powers the intelligence of the car in the car. Car is a safety-critical environment. You don't want your car to be driven by an AI model that is sitting in the cloud, because connectivity is not always available, right? Also, it's private because it's one of those spaces people spend a lot of time in, and you don't want those conversations to be recorded. So we brought one of our multimodal foundation models that is only less than one gigabyte in size, and it can go inside the car's chip—a very, very tiny chip. The chip could be as cheap as $60. That's what I'm saying. We're bringing that level of intelligence into that voice and it is going to power the multimodal intelligence experience inside the car. We do that with all car manufacturers. We announced the Mercedes partnership as a first point of entry, because automotive is a very sensitive topic and they're pretty slow. One of the things Mercedes actually enjoyed from this process was the speed of operations that we had for enterprises. When we are bringing this type of technology in-house, this has been one of the cornerstones of landing the deals. We are an enterprise company, a B2B company. We are bringing our full power to really deploy the solutions and have platforms that allow people to fine-tune their small models. And fine-tuning small models is not that expensive; it's extremely tangible. So we fine-tuned the small models for the applications inside the car.

我们还有数据飞轮系统，能让系统始终保持适应性。想象一下，企业 AI 中一直存在一些问题：比如下载一个 GLM-2-5.2 或开源模型投入生产。然后呢？当模型内部的使用场景出现漂移时会发生什么？比如在汽车里，物理世界中的挑战更大——当你部署一个完全脱离云端的智能系统时，如何维护更新？我们一直认为智能应该以液态形式存在，必须始终保持适应性。这也是我们正在推进的方向：收集数据，为每个用户个性化模型。与梅赛德斯的合作中，我们今年率先在北美推出。所有 2022 年后的梅赛德斯-奔驰北美车型都将通过空中升级更新，因为更新包只有 600 MB。它就像一个覆盖层，消耗的网络流量不大，还能让我们进一步定制。想象一下，每次更新只需 20 MB，因为我们使用 LoRA 适配器之类的东西放入车内。这样就能递归地改善用户体验。

We also have data flywheel kind of systems that allows the system always stay adaptable. Imagine some of the problems in enterprise AI has always been: let's download a GLM-2-5.2 or an open source model and put that in production. Then what happens after you put the system in production? What happens when there's a drift from the use cases that is hitting this model inside, say, a car? In the physical world it becomes even more challenging because when you deploy an intelligence that is completely disconnected from the cloud, how do you maintain updates? We have always thought about intelligence in the format of liquid — intelligence has to always stay adaptable. That's a portion we're also pushing on: to be able to collect data and personalize models to the experience of every single user. With Mercedes, we're rolling this out first in North America as soon as this year. All the Mercedes-Benz North America cars from 2022 on will get an over-the-air update because the size of the update is 600 megabytes. That's like an overlay; it doesn't consume that much internet to update your software, and that allows us further customization. Imagine if every update you want to perform is in the order of 20 megabytes because we are using some sort of LoRA adapters that we can bring inside the car. You would be able to recursively improve the experience of the user as well.

那我们来具体化一点。明年我在梅赛德斯车里会怎么使用这个模型？目前我在特斯拉里的体验是 Grok，通过空中升级。如果没有联网，我就用不了 Grok。但梅赛德斯突然能支持哪种查询、哪些功能呢？

So let's take it and make it more concrete for me. What am I going to be, how am I using this model in my Mercedes next year? Right now my experience is using Grok in my Tesla, and it's over the air. If I don't have connectivity, I don't have Grok. But what kind of queries, what kind of capabilities does this all of a sudden enable in a Mercedes?

它可以访问——它位于操作系统之下。这意味着它可以访问车内所有功能，大约 700 到 1200 个，具体取决于统计口径。你可以和车子对话，控制所有面板，询问说明书。当你在某个地方卡住时，会有提示弹出，你可以和车子交谈。我们还在加入记忆功能。你基本上可以和这个系统聊天。它的美妙之处在于，它可以完全访问汽车的所有功能以及所有应用，因为都通过函数调用实现，只是一次函数调用的距离。所以如果你想通过这个车内智能单元控制任何东西，你就能控制位于操作系统之上的整个生态系统。

It has access — it's sitting below the operating system. That means it has access to all the functions inside the car. There are 700 to 1,200 functions depending on what you count. You can talk to your car, control all the panels, ask for manuals. When you're stuck somewhere, something pops up, you can talk to the car. There are memory features we are adding. You can basically have conversations with that system. One of the beauties is that it has full access to all functionalities of the car plus all the apps because there are function calls. They are one function call away. So if you want to control anything from this intelligence unit inside the car, you would be controlling the entire ecosystem sitting on top of the operating system.

所以基本上，如果我没理解错，小型语言模型（SLM）的优势首先是模型大小、能耗——它们很高效，可以本地运行。与大型语言模型（LLM）相比，你们如何避免或减少这些模型的过度泛化？

So basically, if I get you right, the advantage of SLMs are first of all the size of the model, energy consumption — they are efficient, and they can run on on-prem. How do you avoid or reduce overgeneralization of these models compared to LLMs?

你说的过度泛化是什么意思？换句话说，它们内部是否有足够的能力准确回答你提出的问题？

What do you mean by overgeneralization? In other words, do they have enough capabilities internally to accurately answer the questions you're asking?

好问题。我之前提到过基础模型开发的框架，即定制深度。我们试图保持适应性，并能访问这些定制栈中的工具。有时提示工程就够了，有时需要对模型进行微调，有时需要重新预训练模型以获得核心能力或专门化。现在我们的平台能自动识别解决方案所需的定制深度。对于梅赛德斯-奔驰，我们有一个名为 Model Plus X 的框架——模型加一个平台来实现定制。我们不是只卖静态权重，而是卖他们可以重新调整和微调的东西。检测基础模型有多少通用性，正是液体模型库在许多应用中的用途。我们与硅医学合作，就像 Alex 向你介绍 Peter 那样。通过那次合作，规模不断扩大，因为他们发现液体基础模型非常适合定制。所以我们有了最先进的生物技术基础模型、长寿基础模型。作为一家构建基础模型的横向公司，我们进入了微调领域，并且已经成功做到了。

Great question. So if you have — I told you about the framework of foundation model development which is depths of customization. We try to stay adaptable and have access to tools across these customization stacks. Sometimes prompt engineering is enough. Sometimes you need to fine-tune the model. Sometimes you have to pre-train a model again for core capabilities or specialization. Now our platforms are automatically identifying what depth of customization is needed for a solution. For Mercedes-Benz, we have a framework we call Model Plus X — model plus a platform that allows customization. We are not just selling static weights; we sell something they can retune and fine-tune. Detecting how much generality the base models have is something that libraries of liquid models are coming out for many applications. We work with silicon medicine, as Alex introduced you to Peter. Through that interaction, it is getting big because they discovered liquid foundation models are very good at being customized. So we have state-of-the-art biotech foundation models, longevity foundation models. As a horizontal company building foundation models, we went into fine-tuning, and that became something we have managed to do.

在其他合作方面，最近我们与 Shopify 达成了 10 亿次请求的合作。我们在 Shopify 框架中部署了液体基础模型，已投入生产 6 个月，为客户提供服务。Shopify 覆盖了 1 亿用户、100 亿产品。我们还在与梅赛德斯-奔驰合作汽车侧。我们正在与 AMD 及其他芯片制造商合作，将低代码 AI 体验带到 PC 上。这是另一个我们进入的领域。

Then in terms of other engagements, recently with Shopify we entered 1 billion requests in the Shopify framework. We deploy liquid foundation models in production there for 6 months serving clients. Shopify touches 100 million users, 10 billion products. We are working with Mercedes-Benz on the car side. We are working with AMD and other chip manufacturers to bring low-code AI experiences on PCs. That's another area we enter.

我们公司的重点是真正确保能够将智能带到数据中心之外。这是我们一直专注的事情，我认为我们的效率实际上让我们能够达到

The focus of our company is to really make sure that we can bring intelligence outside of data centers. That's something that we have focused on and I think our efficiency is actually allowing us to get

预备事项。我在 Liquid 没有经济利益。抱歉 Ramin，我必须问一个最明显的问题。我有太多问题要问你。我理解 Liquid 公司是这样创立的，我记得读过原始论文，我想是在《科学》或《自然》上关于液态神经网络的。其基本前提是一种神经形态前提，即你可以通过观察线虫（几百个神经元，可以说是终极小型神经网络）来获得有用的 AI 见解。但我的印象——我希望你能帮我修正或调整——是，尽管 Liquid 最初是一个神经形态前提（一个后 Transformer、非常注重循环的架构前提或先验），但随着时间的推移，根据我对公开信息的印象，Liquid 看起来越来越像 Transformer、或 Transformer+、或 Transformer+Hyena+等等，越来越像一种常规的现成架构。向梅赛德斯等公司销售定制的 Transformer 衍生品可能是好生意。如果是这样，从商业角度来说很好。但从技术角度来看，Liquid 是否仍有任何在生产或开发中看起来像后 Transformer 架构的东西？你能谈谈你当前使用的架构中有哪些是后 Transformer 或非 Transformer 的吗？

Preliminary matter. I have no financial interest in Liquid. Sorry Ramin, I have to ask the most obvious question. I have so many questions for you. Which is the company Liquid was founded as I understand it and I remember reading the original I think it was in Science or Nature paper on liquid neural networks. The premise is basically a neuromorphic premise that you could gain useful AI insights from looking at nematodes, a few hundred neurons, sort of the ultimate small neural network. But my perception, I'm hoping that you can either help me amend or revise my perception, is that although Liquid started with a neuromorphic premise if you will, a post-transformer, very recurrent-oriented architectural premise or prior, over time, again just based on my perception of public messaging, Liquid looks more and more like either transformer or transformer plus or transformer plus Hyena plus etc. Looks more and more like basically a conventional off-the-shelf architecture. It may be a good business selling sort of customized transformer derivatives to Mercedes at all. If so, great from the business side. But from the technical side, does Liquid still have anything that looks remotely like a post-transformer architecture either in production or under development? And can you speak to what if anything is post-transformer or non-transformer oriented about the architecture that you currently use?

非常好的问题。那么让我来谈谈架构空间。原始形式的液态神经网络是你实际能创建的最具表现力的计算格式之一。可以说，就我们的架构而言，它们具有嵌套的非线性，你无法真正去掉它们。它们完全是受物理启发的。它们具有神经 ODE，并且基本上可以处理不规则采样的数据。所以它们成为了一类非常通用的架构。在这些东西之下，当你想要扩展这类技术时，这些循环、这些嵌套的循环。如果你想扩展这些系统，包括我们自己很多人都尝试过将动力学线性化，以便实际扩展它们。状态空间模型基本上就像 Mamba 和那些变体，属于同一类连续时间神经网络，但为了扩展而简化为线性动力系统。它们属于我们拥有的这类连续时间模型。然后还有线性注意力、门控线性注意力的变体出现。它们也有一个门控机制，一个特殊的输入依赖的门控机制，实际上是我们从神经元如何相互交换信息中获得的灵感。那个门控机制也增加了更多的表现力；它也是神经元如何交换信息的原始形成的后裔。那个门控机制至今仍存在于许多不同的架构中，包括我们的。但我想提到最重要的一点是，我们真的不想让自己偏向于单一架构。我们在 Liquid AI 第一天就做了一件事：我们设计了一个搜索算法，让算法而不是人类偏倚，在比如 100 种不同操作变体上运行缩放定律，这些变体可能给你一个通用计算机。所以我们建立了一个元系统。相关的论文我们大约两年半前发表了。我们发表了一篇题为“STAR: 自动化定制架构设计”的论文。所以请阅读 STAR。STAR 是一个框架，它将所有任何格式的动力系统（包括注意力的变体）整合成一种格式，以便我们能够搜索。因此，针对四个标准，找出最优化神经网络架构用于特定部署：标准一是内存，你在给定处理器上消耗多少内存；标准二是计算效率，你操作的速度有多快；标准三是操作延迟；标准四是不损失性能精度。有纯 Transformer 模型，也有你可以构建的混合模型。混合模型有一个基本组成部分，即其中含有一点 Transformer，但其余的动力系统以及为了我提到的四个目标函数，大部分动力系统会被改变。你可以实际自动化整个框架来内部设计基础模型。Liquid 基础模型的技术栈称为自动化基础模型设计算法，我们称之为 AFMD。这个自动化框架探索针对特定硬件的架构。你猜我们开始优化的第一代架构结果是什么？它产生了双门控卷积机制，网络的 80% 都是这种机制。所以当我们不带人类偏倚运行时，原始液态神经网络论文中的门控机制实际上以非常相似的形式出现在搜索空间得出的最终架构中。

Great great question. So let me tell you the space of architecture. Liquid neural networks in the original form, they are one of the most expressive formats of compute that you can actually create. Arguably, in terms of our architecture, they have nested nonlinearities that are like you cannot really take them out. They are completely physics-inspired. They have neural ODEs and basically irregularly sampled data can be handled by them. So they become one of the very general class of architectures as a whole. Underneath these things, when you want to scale this type of technology, these recurrences, these nested kind of loops that they have. If you want to scale these systems, a lot of people have attempted including ourselves to linearize the dynamics so that you can actually scale them. State space models are basically like Mambas and those kind of variants falling into the same category of continuous-time neural networks but dumbed down into a linear dynamical system because you want to scale them. They are underneath this class of continuous-time models that we have. Then there are variants of linear attention, gated linear attentions that are coming out. They also have a gating mechanism, a special gating input-dependent gating mechanism that we actually got inspired by how neurons actually exchange information with each other. That gating mechanism is also something that is adding a lot more expressivity; it is also a descendant of the original formation of how neurons exchange information with each other. That gating mechanism still exists today in many different architectures including ours. But the most important thing that I want to mention is that we really didn't want to bias ourselves towards one single architecture. One of the things that we did day one at Liquid AI, we designed a search algorithm to let the algorithm, instead of human biasing, run the scaling laws on, say, 100 different variations of operations that potentially can give you a general-purpose computer. So we build a meta system. The paper around this is actually we published like two and a half years ago. We published a paper about the topic called 'STAR: Automated Design of Tailored Architectures'. So read about STAR. STAR is a framework that brings all the dynamical systems with any format including variations of attention into one format for us to be able to search through. So to see for four criteria what is the most optimal neural architecture of choice for a certain deployment: number one criteria is memory, how much memory are you consuming on a given processor; number two was the efficiency of computation, how fast you can operate; number three is latency of operations; and number four do not lose accuracy on the performance. There are pure transformer models and then there are hybrid models that you can actually build. Hybrid models have an essential component like they have a little bit of transformers in them but the rest of the dynamical system and most of the dynamical system for the purpose of these four objective functions that I mentioned would be changed. You can actually automate this whole framework to design foundation models in-house. The technology stack of Liquid foundation models is called Automated Foundation Model Design algorithms, we call it AFMD. This automated framework is the one that explores architectures for a given kind of hardware. And guess what came out of the first generation of the architectures that we started optimizing. It came double gated convolution kind of mechanisms as 80% of the network being this. So when we run without a human bias, the gating mechanism that we had exactly in the original Liquid neural networks paper actually shows up with this very similar kind of format in the final architecture that comes out of the search space.

欢迎大家来到由 Fountain Life 赞助的 Moonshots 健康板块。你知道我们在 Moonshots 播客中一直在谈论 AI。除了教育你的孩子和帮你报税之外，AI 能为你做的最重要的事情之一就是确保你过上健康的生活方式，让你有机会活到 100 岁以上。今天我和 Don Mucalem 博士在一起，他是 Fountain Life 的首席医疗官，也是我的医疗团队的一员。Don，很高兴见到你。

Everybody welcome to the health section of Moonshots brought to you by Fountain Life. You know we talk about AI on this Moonshots podcast all the time. One of the most important things AI is going to be able to do for you besides educating your kids and helping you with your taxes is making sure that you're living a healthy lifestyle that you get a chance to get to 100 plus. I'm here today with Dr. Don Mucalem, the chief medical officer of Fountain Life and a part of my medical team. Don, a pleasure.

太好了。

Great.

你知道，人们对活到 100 岁或 120 岁最担心的事情是他们的认知能力，确保他们没有痴呆症，而关于痴呆症的数据是有问题的。

You know the thing that people are concerned about most about living to 100 or 120 is their cognitive abilities, making sure they don't have dementia and the numbers about dementia are problematic.

能分享一下你的收获吗？

Can you share what you've learned?

这点非常重要，你说得对。在 Fountain Life，我们的会员最担心的事情就是失去大脑健康——忘记孩子的名字、忘记亲人的面孔。我们知道，关于痴呆症，保守估计有 45% 是完全可预防的。令人惊讶的是，通过我们在 Fountain Life 进行的高级检测，四分之一的会员大脑年龄已经超前。

Such an important point and you're right. At Fountain Life, our members, the number one thing people are most concerned about is losing their brain health — forgetting the name of their child, forgetting the face of their loved one. We know that when it comes to dementia, the conservative estimates are that 45% are entirely preventable. What was amazing is with the advanced testing we're doing at Fountain Life, one quarter of our members had advanced brain age.

哇。

Wow.

但真正了不起的是，当我们把预防与健康生活方式结合起来时。这让我起鸡皮疙瘩。更健康的饮食、运动、睡眠——优化睡眠非常重要。你知道我们看到了什么吗？我们看到大脑年龄改善了 26%。这是一个巨大的数字，表明大多数人实际上能够改善他们的大脑年龄。

But what was really awesome is again back to that prevention when we partnered it with healthy living. This gives me chills. Eating healthier, moving our bodies, sleep — optimizing sleep is so important. You know what we saw? We saw that we improved that brain age by 26%. That is a big big number to show that the majority of those individuals were able actually to improve their brain age.

我喜欢 Fountain 的一点是，我们在全球寻找最好的疗法和最佳方法，并确保把它们带给我们的会员。所以，如果你希望健康的大脑功能持续到 100 岁、120 岁，请关注 Fountain Life。访问 fountainlife.com/per。确保成为自己健康的 CEO。好了，现在回到节目。

And one of the things I love about Fountain is we're searching the world for the best therapeutics, the best approaches, and making sure we bring it to our members. So if having healthy brain function till 100, 120 is important to you, check out Fountain Life. Go to fountainlife.com/per. Make sure you become the CEO of your own health. All right, now back to the episode.

好了，我们的下一个故事来自 Palmer Lucky，他是 Oculus 的创始人，现在也是国防巨头 Andre 的董事长。把 Andre 称为国防巨头有点好笑，但确实如此。他声称现代专利制度已经成为国家安全隐患。用他的话来说，“整个专利局每天早上都可以被下载、剽窃，然后用来对你发动战争。”核心问题在于专利的实际作用。专利是一项要求：如果你想获得专利，你必须教给本领域技术人员如何实际制造和使用你的设备。所以这种对发明的公开，用专利法的精确语言来说，必须“以如此充分、清晰、简洁和精确的术语，使任何本领域技术人员能够制造和使用该发明。”如果你这样做了，你实际上是在教全世界如何使用它，而作为交换，你获得了大约 20 年的独占权。Palmer 认为，当战略性对手可以简单地收集所有文件，无视法律保护，并将公开的知识武器化时，你就把最佳想法的免费说明书交到了他们手中。这里有一些数字：美国专利局每年收到约 60 万件申请，批准略过半数的 32.3 万件。这是 2025 年的数据。有趣的是，过去五年中批准的专利增加了 40%。我猜这得益于人工智能。Palmer 提出的解决方案不是废除专利，而是大规模扩展国家保密专利流程，该流程可追溯到 1951 年的《保密法》。这个鲜为人知的机制允许发明人获得保密专利，你保留独占权，但你不能向任何人公开，政府也不能。目前美国约有 6000 项此类保密令。Lucky 希望将这个边缘情况变成默认机制。所以问题是：如果我们真的用这种开放性——这曾是美国企业家卓越性的基础——来换取一个秘密系统，我们是在用专利被剽窃的安全换取我们所拥有的创新生态系统吗？让我们看一段 Palmer 的短片，然后我们再来讨论。

All right, our next story comes from Palmer Lucky, the founder of Oculus and now the chairman of the defense giant Andre. It's funny to call Andre a defense giant, but it is. He's claiming that the modern patent system has become a national security liability. In his words, 'the entire patent office could be downloaded every morning, ripped off, and used to fight a war against you.' The core problem is baked into what patents actually do. Patents are a requirement: if you want to get a patent, you have to teach a person skilled in the art how to actually create and use your device. So this disclosure of your invention, in the exact words of patent law, is in such full, clear, concise, and exact terms as to enable any person skilled in the art to make and use the same. So if you do that, you're effectively teaching the world how to use it, and you're exchanging that sharing of your invention for roughly 20 years of exclusivity. Palmer argues that when a strategic adversary can simply harvest every file, ignore the legal protections, and weaponize the disclosed knowledge, you've handed them a free instruction manual to your best ideas. So just for some numbers: the US Patent Office receives about 600,000 applications annually. It grants a little over half of those — 323,000. That's 2025 data. Interestingly enough, patents granted have increased 40% in the last five years. My guess is that is secondary to AI. Palmer's proposed fix isn't to abolish patents. It's to massively scale up a national security patent process which goes back to the Secrecy Act of 1951. So this obscure mechanism lets inventors obtain classified patents in which you keep your exclusive rights but you don't disclose it to anyone and neither can the government. So there are roughly 6,000 of these secure secrecy orders active in the US. Lucky wants that this edge case is turned into a default mechanism. So here's the question: if we genuinely trade this openness — which has been sort of the basis for American entrepreneurial exceptionalism — for a secret system, are we trading safety of having our patents ripped off against really the innovative ecosystem that we've had? Let's watch a short video from Palmer and then we'll talk about it.

停止为一切申请专利。专利是中文的说明书。开国元勋们从未预见到一个全球化经济体，其中整个专利局每天早晨都可以被下载、剽窃，然后用来对你发动战争。我们真的需要从根本上重新审视专利制度。我认为我们需要大规模扩展国家保密专利流程。你可以获得一个保密专利。你可以为一些不允向任何人公开的东西申请专利，但你仍然保留对这些权利的独占权。我们需要大幅扩展这个计划。所以，你知道，我已经申请并获得了十几项专利。我知道 Alex，你的专利数量甚至更多。所以我想知道，你们怎么看这个问题？Alex，你想先说说吗？

Stop patenting everything. Patents are Chinese instruction manuals. Well, the founding fathers never predicted a world where you would have a globalized economy where the entire patent office could be downloaded every single morning and then ripped off and then used to fight a war against you. We need to really fundamentally revisit the patent system. I think we need to massively expand the national security patent process. You can obtain a classified patent. You can get a patent on something that you are not allowed to disclose to anyone, but you still maintain the exclusivity on those rights. We need to massively expand that program. So, you know, I've applied for and gotten a dozen patents. I know Alex, you have an even much larger number of them. So I'm curious, guys, how do you come out on this? Alex, do you want to kick it off?

我认为这一集是科技 CEO 们提出糟糕主意的典型例子。我认为这是一个糟糕的主意。我要说，1951 年的《发明保密法》（Palmer 指的可能就是这个）总体而言可能非常有害，不仅仅是对民主。如果专利……那么稍微解释一下背景：《发明保密法》的运作方式并不是你可以秘密提交专利而不公开。而是基本上，这项发明——被军方征用或没收的发明——只能出于军事目的实施。绝非其他解释，认为《发明保密法》以某种方式为个人提供法律掩护，让他们在保密条件下秘密公开其发明如何工作，然后普遍实施。他们做不到。只有军方才能独家实施，然后发明人从中获得版税。这对 Andre 的国防业务可能有好处，但我认为总体上是一个糟糕的主意。我更担心的是，这基本上就是秘密垄断。我认为我们现有的《发明保密法》对发明进行分类已经很糟糕了。试想，那些可能对整个世界产生经济变革的整个技术领域，无论是能源还是其他领域，是否在普通公众不知情的情况下被《发明保密法》吞噬，并纯粹出于军事目的被战争部没收？这对我来说非常令人担忧。扩大它的想法——我认为如果有什么的话，《发明保密法》制度应该被废除。

I think this is the episode of people — tech CEOs floating terrible ideas. I think this is a terrible idea. I would argue the Invention Secrecy Act of 1951, which I think is what Palmer is gesturing at, has probably been on balance quite detrimental, not just to democracy. If patents... So maybe a bit of context: the way the Invention Secrecy Act works is it's not that you can just file the patent in secret and not disclose it. It's that basically it can only be practiced — the invention that is basically confiscated or eminent domained by the military — can only be practiced for military reasons. It's not, under any other construal, that the Invention Secrecy Act somehow offers legal cover for an individual to secretly disclose how their invention works under some confidentiality and then go practice it in general. They can't. It's that the military exclusively can practice it, and then the inventor gets royalties from that practice. That may be good for Andre's defense business, but I think in general a terrible idea. A greater concern I have is that these would be basically secret monopolies. I think it's bad enough that we have Invention Secrecy Act classification of inventions. Query whether entire swaths of technology that could be completely transformative economically to the entire world, from an energy perspective or other domains, have somehow without general knowledge been swept up by the Invention Secrecy Act and basically confiscated by the Department of War for purely military reasons. That's very concerning to me. The idea of expanding it overall — I would argue if anything, the Invention Secrecy Act regime should probably go away.

我们可以进行这场辩论。Palmer 将于 9 月 25 日参加我们在洛杉矶举办的 Moonshots 聚会。请大家访问 moonshots.com。我们将与 Moonshots 的伙伴们度过精彩的一天。我们将与 Palmer、Salem 进行这些对话。我的意思是，让美国伟大的地方在于我们的开放创新政策——人们在他人创造的基础上继续创造。你对此有什么看法？

We can have this debate. So Palmer is going to be joining us at the Moonshots gathering on September 25th in LA. Everybody go to moonshots.com. We have an amazing day with the Moonshots mates there. We'll be having these conversations with Palmer, Salem. I mean, what makes America great is our open innovation policy — people building on top of other people's creations. What are your thoughts here?

你看，我们在过去二三十年里看到这个问题越来越大。公开信息，尤其是在人工智能时代，人们可以绕过它、复制它或从中学习，这是一个巨大的挑战。真正的模式是学习循环。

Look, we've seen this problem get bigger and bigger over the last 20 to 30 years. Where the disclosure, especially in an age of AI where people can just route around it or replicate or learn from it, it's a huge challenge. The real mode is learning loops.

真正的护城河在于：你的反馈循环是什么，能否以专有方式学习，然后围绕它创建商业机密并在市场中付诸行动？持续创新才是获胜的防御手段，而不是所有权。在这种特定模式下，唯一赢家只有律师。

That's going to be the real defensibility: what are your feedback loops, and can you learn in a proprietary way and then create trade secrets around that and act on that in the marketplace? Continuous innovation is going to be the winning defense. It's not going to be ownership. The only people that win in this whole particular model are the lawyers.

说得好。戴夫，你有什么想法？

Well said. Dave, any thoughts here?

是的，我认为如果存在第三次世界大战的爆发点，这很可能是最有可能的一个。你看，亚历克斯说得对。我们将以惊人的加速速度发现新物理学、新药物。欧洲等地尊重知识产权，这创造了一个连贯的经济体系，你可以交易这些东西。而中国完全无视知识产权，直接拿走就跑。所以我认为可能的结果是美国会对任何不尊重知识产权的人实施贸易禁运。然后你必须选择：你是自由世界的一部分，还是另一个世界？但我认为这是更可能的结果，而且很快就会发生，比如未来几年内，因为创新速度将突破天花板。但我读过的所有科幻未来小说中，没有一本不是人工智能以惊人的加速速度创造大量知识产权，并且有某种机制让创新者从中获利。如果没有这一点，你就没有未来。大量来自剑桥、MIT 和哈佛的杰出思想家不会去研究基础技术，因为无利可图。这必须从根本上改变。保护知识产权是扭转这一趋势、让人们致力于真正重要事情的关键途径。

Yeah, I think if there's a flash point for World War III, this is probably one of the most likely. Look, Alex is right. We're going to discover new physics, new medicines at an incredible accelerating rate. And places like Europe respect intellectual property rights, and that creates a coherent economy where you can trade these things. China completely ignores intellectual property rights and just takes it and runs with it. So I think the likely outcome is that the US will trade embargo anyone who doesn't respect intellectual property rights. Then you have to choose: are you part of the free world or part of the alternate world? But I think that's the more likely outcome, and that's going to happen soon, like in the next couple of years, because the rate of innovation is going to go through the roof. But there's no science fiction future book I've ever read where there isn't massive amounts of intellectual property being created by AI at an incredible accelerating rate and there's some vehicle by which innovators can profit from that. And if you don't have that, then you don't have the future. A huge fraction of brilliant thinkers coming out of Cambridge, MIT, and Harvard don't work on foundational technologies because there's no money in it. And that's got to change fundamentally. And protecting intellectual property rights is a key way to reverse that tide and get people working on really important things.

我也同意戴夫的观点，帕尔默从根本上误解了——或者似乎误解了——专利的本质。专利的全部意义在于，你公开其工作原理，以换取国家授予的临时垄断权。而那种抱怨中国人拿泄露跑了的说法——实际上是对专利执行问题的吹毛求疵。并不是要因噎废食，说我们想放弃以公开换取临时垄断的专利交易。他真正应该要求的是在美国专利在中国的更好执行。

I think to Dave's point also, Palmer fundamentally misunderstands, or appears to misunderstand, the nature of patents. The whole point of a patent is that you disclose how it works in return for a state-granted temporary monopoly. And you know, sort of bellyaching that the Chinese are running away with the disclosure — it is really a quibble with enforcement of patent. It's not that you want to throw the baby out with the bathwater and say we want to give away the patent trade of disclosure in return for temporary monopoly. Really what he should be asking is better enforcement of US patents in China.

同意。好吧，我要带大家进入医疗富足的世界。本周有两个故事展示了人工智能对医疗富足的惊人影响，使诊断去货币化并民主化，惠及数十亿人。第一个故事是 GPT 5.6（几周前发布）在 HealthBench Professional（OpenAI 最难的医疗基准）上的表现。GPT 5.6 创下了全新的历史基准最高分。第二个故事是，在大约 20,000 个个体医生判断的盲测中——换句话说，就是诊断的准确性、安全性和完整性——GPT 5.6 的答案与专业匹配的医生（肺科医生、儿科医生等）进行了比较，这些医生被允许无限制地访问网络并有无限时间回答，但医生们还是输了。所以一段时间以来我们已经知道，这些 AI 诊断模型比最好的医生更出色，即使医生拥有所有人类可用的工具。故事的第二个部分来自 Meta。OpenAI 自己的 HealthBench Professional 基准包含 525 个真实的临床任务。Meta 的 Muse Spark 1.1 上周发布，在所有指标上击败了 GPT 5.6，而且成本便宜 7 倍。但更好的是，值得注意的一点是，Muse Spark 在 Meta 所有产品中都是免费的。你知道，WhatsApp、Facebook 和 Meta 今天有 35.6 亿日活跃用户使用他们的产品。所以我们现在的状况是，顶尖的医疗 AI 能力现在对地球上超过 35 亿人免费。这真是太了不起了。我的意思是，这就是富足论题的大体现。而且，当人们谈论对 AI 的担忧等等时，请意识到这一点：从未接触过最佳诊断医生的人现在拥有了它们。有一个模型——中国的一个 AI 医生——已经在农村环境中被 1 亿人使用。对吧？基本上诊断已经实现了巨大的成本坍塌。医疗领域特别有趣，因为在这里富足实际上变得道德上紧迫，对吧？如果你能以接近零成本提供更好的一线答案，那么问题就在于你能多快安全地将其推向市场？这是唯一的问题。所以这绝对正确；让我们认识到，世界上几乎每个国家都严重缺乏医生。

Agreed. All right, I'm going to move us into the world of healthcare abundance. So, two stories this week are demonstrating an incredible impact of AI on healthcare abundance, demonetizing and democratizing diagnostics for billions of people. The first story is the performance of GPT 5.6 (which was released a couple weeks ago) on HealthBench Professional, which is OpenAI's hardest medical benchmark. So GPT 5.6 set a brand new all-time benchmark high. And then the second part coming out here is in a blind test across roughly 20,000 individual physician judgments — in other words, diagnosing for accuracy, safety, completeness — GPT 5.6's answers were compared to specialty-matched physicians (pulmonologists, pediatricians, whatever) who were given unlimited full access to the web and unlimited time to answer, and the doctors still lost. So we've known this for some time that these AI diagnostic models are better than the best physicians given all the tools that humans can use. The second part of the story comes from Meta. So, OpenAI's own HealthBench Professional benchmark, which is 525 real clinical tasks. Meta's Muse Spark 1.1, again released last week, beat GPT 5.6 across the marks and it was 7 times cheaper. But even better, important to note here is that Muse Spark is free inside all of Meta's products. You know, WhatsApp and Facebook and Meta today serves 3.56 billion daily active users using their products. So here we've got a situation where the top medical AI capabilities are now free to over 3 and a half billion people on the planet. And that's just extraordinary. I mean, this is the abundance thesis at large. And again, as people talk about the concerns of AI and so forth, please realize this: people who've never had access to the best diagnosticians now have them. There is a model — an AI doctor in China — that's being used in rural environments by 100 million people already. Right? Basically diagnosis has had massive cost collapse. The healthcare domain is particularly interesting because it's where abundance becomes actually morally urgent, right? If you can deliver way better first-line answers at near zero cost, it's how quickly can you safely get it out there? That's the only question. And so it's absolutely right; let's recognize that in almost every country in the world, there's a radical doctor shortage.

所以，这真的很关键。你看，这简直就是 2026 年 7 月的故事。想想看：Instagram 现在给出的医疗建议比人类医生还好。

So, this is really critical. You see, this is such a July 2026 story. Think about it: Instagram now gives better medical advice than a human doctor.

这太疯狂了。智能的成本不仅仅是低到无法计量；医疗智能的成本正在变得低到无法计量。免费，基本免费。我的意思是……

It's pretty wild. The cost of intelligence is not just going too cheap to meter; the cost of medical intelligence is becoming too cheap to meter. Free, basically free. I mean that's...

嗯，最终的便宜到不计成本是渐近自由，对吧？但我要说，老实说，我怀疑 Meta 在这方面有一点轻微的基准优化。Meta Spark 1.1 如果你相信 AI 成本前沿分析的话，是在最优成本前沿上，但并不是顶尖的。所以如果它击败了比如说 Fable 5（它几乎不能做任何生物学相关的事情）或 GPT 5.6（它能做），在我看来，老实说，确实有一点轻微的基准优化。但尽管如此，当 Instagram 能给出比人类医生更好的医疗建议时，这仍然是伟大的一天。

Well, the ultimate too cheap to meter is asymptotically free, right? But I would say, in all honesty, I suspect a little bit of mild benchmark-maxing by Meta on this. Meta Spark 1.1 is, if you believe the AI cost frontier analysis, on the optimal cost frontier, but it's not at the top. So if it's beating, say, Fable 5, which barely allows you to do anything biological, or GPT 5.6, which does allow you to do it, that does to me suggest, in all honesty, a little bit of mild benchmark-maxing. But still, it's a great day when Instagram gives better medical advice than human doctors.

我想这是我们今天播客的金句。我要带大家进入另一个我喜欢的寿命延长故事。这是昨天的突发新闻。它真的让我很兴奋。我知道你、亚历克斯和我在谈论这个。几十年来，衰老的一个基本问题是所谓晚期糖基化终产物的缓慢积累。我喜欢这个缩写：它叫 AGEs。这些是糖分子，随着时间的推移会交联并损害你体内的蛋白质。这种化学反应称为糖基化。它在我们体内随着年龄增长缓慢发生。它使动脉硬化，使晶状体因白内障而混浊，损害肾脏，使皮肤起皱。而这个想法是它一直被认为是不可逆的，直到这周。

I think that's our takeaway quote from today's pod. I'm going to move us to one more longevity story that I love. This is breaking news from yesterday. And it really got me excited. I know you, Alex, and I were talking about this. So for decades, one of the fundamental problems of aging is the slow accumulation of what are called advanced glycation end products. I love the acronym: it's called AGEs. And these are sugar molecules that cross-link and damage your proteins in your body over the course of time. This chemical reaction is called glycation. And it happens slowly in our bodies as we age. It stiffens your arteries, clouds your lenses with cataracts, damages kidneys, and wrinkles skin. And this idea is that it's always been irreversible until this week.

昨天在《自然·通讯》上，一家名为 Revel Pharmaceuticals 的新创公司展示了一种名为 CMLA 的工程酶，它就像一台分子割草机——我很喜欢他们的描述——一台分子割草机。它能氧化掉糖化疤痕，恢复底下原本健康的蛋白质。令人惊讶的是，这并不只是在试管里发生。他们证明，在来自老年捐赠者的人体组织样本中，这种酶能逆转一生积累的损伤。虽然还处于早期，但其意义怎么强调都不为过。衰老中一个我们一直归为不可逆的类别，现在变得可逆了。我们谈论长寿逃逸速度，谈论每秒钟在 40 万亿个细胞中每个细胞里 50 亿个化学反应的理解能力，而当我们谈论在 2033 年前达到逃逸速度时，靠的就是这样的技术。所以恭喜 Revel 做到了这一切。

And yesterday in Nature Communications, a team from a new startup called Revel Pharmaceuticals demonstrated an engineered enzyme called CMLA that acts like a molecular lawn mower. I love their description. A molecular lawn mower. It oxidizes away the glycation scars and restores the original healthy protein underneath. And amazingly, this isn't happening just in a test tube. They showed it worked in human tissue samples from elderly donors, reversing damage that accumulated over the lifetime. It's still early, but the significance of this cannot be overstated. A category in aging that we've always filed as permanent just became reversible. We talk about longevity escape velocity, our ability to understand the 5 billion chemical reactions per second per cell in your 40 trillion cells, and when we talk about reaching escape velocity by 2033, it's tech like this. So congrats to Revel for doing this.

而且不只是 Revel。有几点值得注意：Revel 和 Calico（加利福尼亚生命公司）参与了这项工作，Calico 是 Alphabet 旗下的另一家赌注公司，比 Waymo 安静得多。他们仍在做研究，我很受鼓舞的是 Calico 显然深度参与其中，并且还活着。另外几点：这个过程涉及一类称为美拉德反应的化学反应，这也是为什么烤面包时外皮会变成褐色，或者为什么用素食话说，所有东西据说吃起来都像鸡肉。这是同一类反应：糖类中的羰基官能团与蛋白质中的胺基反应，生成一大类在光学上呈棕色的分子。人体内也发生着同样的过程。这非常令人兴奋，因为虽然不能完全算是把煎蛋恢复回生蛋，但已经走了一半。严格来说，这并不是逆转热力学时间箭头，但如果我们能通过来自细菌的蛋白质的定向进化，去除所有这些与炎症和其他衰老相关因素有关的多余糖-蛋白质副产物，那简直是最接近的替代了。生物圈里还有什么其他宝藏等着我们去挖掘，除了那些显而易见的？对于长寿逃逸速度，潜力巨大。还有哪些细菌创新可以用来逆转衰老？

And not just Revel. A couple of interesting notes here: it was Revel and Calico, the California Life Company, which was one of Alphabet's other bets that's been much quieter than, say, Waymo. They're still doing work, and it's very encouraging to me that Calico is apparently deeply involved in this and has a heartbeat. A couple of other points: the broader process involves a class of chemical reactions called Maillard reactions. It's also the reason why when you bake bread, the outer crust is brown. Or why, for vegetarian speaking, everything purportedly tastes like chicken. It's the same class of reactions: sugar reacts with carbonyl functional groups within sugars, reacting with amines in proteins to create a broad class of molecules that look optically brown. So the same thing is going on in the human body. To me this is very exciting because it's not quite unscrambling eggs, but it's halfway there. It feels almost—strictly speaking, it's not like reversal of the thermodynamic arrow of time, but it's the next best thing if we can remove all these unwanted sugar-plus-protein byproducts that are associated with inflammation and other correlates of aging, using directed evolution of a protein that came from bacteria. Like, what else is there out there in the biosphere for us to mine, in addition to all the obvious ones? Great potential for longevity escape velocity. What other bacterial innovations can we use to turn back aging?

这是人类工程学。我们正在掌控。它从自然选择的进化转向了人类引导的进化。我很喜欢这一点。

It's human engineering. We're taking control. It's going from evolution by natural selection to evolution by human direction. And I love that.

等我有了拉明的头发，我就高兴了。

I'll be happy when I have Ramin's hair.

那时候我就高兴了。

That's when I'll be happy.

嗯，有很多公司在做那方面的工作，Sem。所以，先生们，感谢今天的时光。我很期待今天晚些时候星舰 13 的发射。祝伊隆和那里的团队好运。拉明，恭喜 Liquid AI 的成功，很高兴你来做客。戴夫，投资拉明真是好决策。

Well, there are lots of companies working on that, Sem. So gentlemen, grateful for our time today. I'm excited for Starship 13 launch later today. Wish Elon and the group there lots of luck. Ramin, congrats on the success of Liquid AI and excited to have you on the pod with us. Dave, great move investing in Ramin.

我代表所有人，谢谢你。

On behalf of all of us, thank you.

好的，先生们。祝你们一周愉快。我相信我们很快就会做一期紧急播客，因为奇点速度不等人。

Yeah, gentlemen. Have an amazing week. I'm sure we'll be having an emergency pod very soon because the speed of the singularity waits for nobody.