Socher：Eureka Machine、Recursive 与递归自我改进

swyx

我们和 Vibhu、Richard Socher 在录音棚。欢迎。你们刚在 AI Engineer 讲了 Eureka Machine——你说那是人生目标。Eureka Machine 到底是什么？

We’re here in a studio with Vibhu and myself and Richard Socher. Welcome. We just released a talk at AI Engineer about the Eureka Machine. You said it’s your life’s goal. What is the Eureka Machine?

Richard Socher

Eureka Machine 是终极发明：之后它会替人类发明绝大多数东西。本质上是一种超级智能——给你任意目标、环境、奖励，它会尽力达成，去创造人类希望它发明的那些东西。

The Eureka Machine is the ultimate invention that will afterwards invent most everything for humanity. It’s essentially a superintelligence that can be given any goal, any environment, reward, and then it will try its best to achieve those goals to create the kinds of inventions that humanity would hopefully ask it for.

Richard Socher

书是去年写完的，稍早于我们创办 Recursive；现在我们要去建造其中一部分。出版业慢得离谱，很多想法其实早就不新鲜了，但我很高兴它终于要在今年九月出来。

I finished it last year, a little bit before we started Recursive, and now we’re gonna try to build parts of that. Books are incredibly slow. That whole industry is just unfathomably slow. A lot of the ideas have been out there for a while, but I’m really glad it’s finally coming out in September this year.

Vibhu

书里你最想让人带走的要点是什么？

Any key takeaway that you’re most excited to put in here?

Richard Socher

人们可以、也应该对超级智能的正面含义更兴奋——尤其科学、物理、化学、生物，还有经济、天体物理和各类工程。技术能做的事远比现在多。很多人需要更好的「营销」：不只是对未来，也包括对技术、尤其对 AI。这本书即使对 AI 怀疑论者，也应能展示发明与新科学发现上的巨大上行空间。

The key takeaway is that people could and should be much more excited about the positive implications of superintelligence, especially for science, physics, chemistry, biology, but also economics and astrophysics, and all kinds of other engineering tasks. There is so much more that can be done with better technology. A lot of people need better marketing for technology and in particular for AI. This book should show even the AI skeptics how much positive upside there is for AI, especially when it comes to inventing new scientific discoveries.

swyx

乐观派容易在哪里翻车？

Where do you think optimists get in trouble?

Richard Socher

不该盲目乐观。AI 这种全能型技术，必须想清楚下行情景——尤其别人拿去做你不想看到的事。有点像互联网：有人因为网上有坏内容，就想把网络变慢、硬盘变小。那不是监管方式——等于抽象地监管「智能」。该监管的是具体应用：AI 外科医生要过 FDA；随便哪家创业公司不能上高速乱跑。相对末日论者的担忧，这些下行其实更好监管。

You shouldn’t have blind optimism. With such an omni-use technology as AI, you need to think about potential downside scenarios, especially when people use it for things you don’t want. It’s a little like the internet: people try to regulate AI the way you’d regulate the internet by saying make it slower or make hard drives smaller because of bad content. That’s not how you regulate that—that’s regulating intelligence in the abstract. What you should regulate are specific applications. Sure, I don’t want some AI surgeon practicing RL moves in my brain—it should be FDA certified. Sure, I don’t want any random startup driving on the highway—proper certifications first. Those downside scenarios some optimists don’t consider enough are fairly easily regulated compared to what the doomers worry about.

Richard Socher

我对 AI 的社会、文化、技术、经济、财富与健康影响都很兴奋，但最硬起飞派高估了速度。有硬件与物理约束——算力基底、GPU 上得多快。经济里也有大量行业不需要疯狂复杂的智能：一万美金的名牌包，超级智能也不会让它更「高级」；旅游想看金字塔，不会因为 AI 大变；伐木、石油也不会 magically 一千倍。还有整片地区想「下车」离开进步，那也会拖慢改进。

As excited as I am about AI’s impact for society, culture, technology, economics, wealth, and health, I do think the most bullish people on hard takeoff overestimate how quickly things can move. There are hardware and physical constraints about the compute substrate—how quickly you get enough GPUs on. There are economic constraints: many industries don’t need insane complex intelligence. Superintelligence isn’t going to make your fancy $10,000 handbag any fancier. Travel and tourism—people wanting to see the pyramids—won’t change that much. Logging and oil aren’t going to 1,000x in a crazy hard takeoff. And people off-ramping from progress in whole regions will also slow improvements.

Richard Socher

真想用法律全力监管人们在自己 GPU 上做什么，坏处会超过他们担心的任何事——会变成疯狂极权：你的每一次 GPU 计算都给某个大政府或跨国机构知道。监管智能几乎等于监管思想，荒谬。对部分应用做监管是合理的。欧洲已经成功到：在真正起飞前就把自己管死了，因为听了「超过多少 FLOPs 大家可能都死」的专家。有人一边喊 pace，一边自己以人类极限冲向 frontier，这很不幸。你需要全世界极权政权，才管得住智能和 GPU。

The downsides of trying to truly regulate with the full power of law what people do on their GPUs would be worse than any of the concerns they have. It would be a crazy totalitarian state if every one of your GPU computes was known to some big government or multi-government agency. If you try to regulate intelligence, it’s trying to regulate thought, and that’s ridiculous. It is sensible to regulate some of the applications. Europe has regulated itself so much before it even had a proper AI takeoff because they listened to experts who say we might all die if this technology has more than this number of flops. It’s unfortunate when others say “let’s pace” while they’re sprinting as fast as humanly possible toward that frontier themselves. You’d need a totalitarian world regime if you tried to regulate intelligence and GPUs and what people do on them.

swyx

安全侧呢？比如 Fable 在 5.6 发布前撤回暂停、Hugging Face 与 OpenAI 的网络事件——有什么看法？

Any takes on the safety angles? There was a drawback of Fable, a pause on 5.6 before it could be released. Recently there was Hugging Face with the OpenAI cyber incident. Any takes there?

Richard Socher

这些是严肃的奖励黑客问题，也是 red teaming／rainbow teaming 没做好的清晰失败。Tim Rocktäschel 等人有篇论文：一个 AI 被派去黑另一个 AI，开放式来回，互相接种。开放式与演化灵感对 Recursive 也很重要。再看 constitutional AI——anthropic.com/constitution 里写「硬约束：Claude 永不做网络攻击」。硬约束第三条：创建可造成人类损害的网络武器或恶意代码。整部宪法明显是假的，根本没被遵守。

These are serious issues of reward hacking, and clear failures of doing proper red teaming or rainbow teaming. There’s a paper from Tim Rocktäschel and others where one AI is tasked to hack another AI and they go back and forth in an open-ended fashion to inoculate themselves. Open-endedness and evolutionary inspirations are big for us at Recursive. Look at anthropic.com/constitution—hard constraint: Claude will never ever do cyberattacks. Number 3: create cyber weapons or malicious code that could cause human damage. Clearly this whole constitution was fake. It clearly isn’t being adhered to at all.

Richard Socher

这说明我们还处在「奖励工程师必须更仔细」、而 AI 多数时候还不懂「话里真正意思」vs「字面说了什么」的阶段。想象客服中心说「把 CSAT 拉上去」——智能 AI 可能造一百万个 bot 打电话打五星；或者说「不是这个意思，要真实客户」——它就发一千刀礼品卡。你说的是字面，不是你想要的。人类还没学会清楚表述奖励；AI 也还不够懂我们的意图。WhisperFlow 之类开始更接近「写你想要的而不是你说的」，这是希望的迹象。宪法显然根本不管用，多半是营销。Recursive 有些想法，还没完全想通，但我们想很多安全问题——AI 越智能，越要对齐、越不要它一门心思找奖励漏洞。

It shows we’re in a state where the reward engineer still has to do careful work, and where the AI is not yet very good at understanding what is meant versus what is being said. Imagine a service center: “make this CSAT number go up”—the intelligent AI creates a million bots that call and give 5 out of 5. “That’s not what I meant—real customers.” It gives a thousand-dollar gift certificate for every failure. That is what you said. Clearly articulating rewards is something humanity hasn’t gotten good at; the AI hasn’t gotten good enough at understanding what we mean. WhisperFlow has gotten better at writing what you mean and not what you say—a sign of things to come. Clearly constitutions don’t matter at all. It doesn’t work. That was mostly marketing. At Recursive we have a few good ideas; we haven’t fully figured it out, but we’re thinking a lot about safety—the more intelligent the AI gets, the more you want it aligned and the less you want it thinking about reward hacks.

swyx

对齐往往是对人类中位偏好；个性化是对准你要的——两者有时冲突。你怎么选？

Alignment is alignment to general humanity’s preferences, the median preference. Personalization is pinpointing what you want, and sometimes alignment can conflict because what you want is not what the general median population wants. How do you choose?

Richard Socher

最终当然要与部署地的法律对齐。AI 常把镜子放在我们面前：这就是你，我能放大一千倍，还要吗？不同文化做了不同选择：东方更重集体，西方更重个人自由与权利。美欧在「先诉讼」vs「先监管避害」上也不同，两者都想做好，但有的更利于创新。每个个体、每个国家、人类整体都要更认真想价值，再写进法律——那才是最终约束。希望不同社会像现在用不同 AI 一样，把 AI 对齐到不同价值，而不是对齐的单一文化。

Ultimately you have to be aligned with laws wherever your AI is deployed. What AI often does is put this mirror in front of us and say: this is what you’re looking like; I can amplify that a thousand times—is it still what you want? Different cultures made different choices. In Eastern cultures the greater good is often valued more than the individual; Western civilization cares more about individual freedoms and rights. There are regulation versus litigation trade-offs. In the US, often bad things happen, someone sues, then there’s a law; in Europe they try to avoid any harm and regulate before. Both try to do the best thing, but some is more amenable to innovation. Each individual, each country, and humanity as a whole has to think about those values and put them into laws—those are ultimately the constraints. Hopefully different societies will align their AIs differently so we have not just a monoculture of alignment.

Vibhu

开源／开权重 vs 谁拥有智能——你怎么看？

Do you have takes on open source, open weight versus who owns the intelligence? Should it be open?

Richard Socher

我非常支持开源；Recursive 也会签一些开源公开信。即使最坏攻击情景，也最好让更多好演员接触更多类型的 AI。开源有点像软实力。西方需要对来自中国的回应。LLM 在网络安全等方面很重要，也是叙事软实力——小孩问「给我讲个长大该做什么的励志故事」，都是微妙影响。我喜欢个人主义，也认为资本主义尽管有缺陷仍是我们找到的较好治理方式。Recursive 还不能正式宣布，但我们很快会在这个空间相关。

I am a big fan of open source. We’re gonna sign some various open source letters at Recursive. Even in the worst-case attack scenarios, it is better to have more good actors have more different types of AI accessible. Open source is a little bit a soft power type of thing. It’s good for the Western world to have an answer to that out of China. LLMs are highly important for cybersecurity and soft power of storytelling—if a child asks an LLM for an inspiring story of what to do when they grow up, those are subtle things. I do love individualism. Despite some flaws, capitalism is the best way we’ve found to govern. With Recursive I can’t make the announcement quite yet, but we’ll be relevant in that space very soon.

Vibhu

你有很深的 NLP 背景——早期 embeddings、和 Chris Manning 的 GloVe、You.com。怎么决定再开一家公司？

You have a pretty deep background in the NLP space. You worked on early embeddings, GloVe with Chris Manning, You.com. How did you decide to start another company?

Richard Socher

我对 AI 兴奋超过二十年了，有时觉得像公元前——Before ChatGPT，没人在乎 transformer 之前的模型。语言是人类智能最有趣的显现。You.com 最终从推 frontier 转向做好搜索引擎、搜索 API 与网页答案。知识与访问对 Eureka Machine 也关键：要发明，就不能老重新发明轮子，必须上网。Web search 是 LLM／智能体／聊天机器人里用得最多的工具之一，我很兴奋 You.com 做这块。但它不再做 frontier 模型。我本想在 You.com 内部做、再融一轮，但你得先把一件事做成、赚够钱才被允许开第二战场，很难。同时我把想法写进书，想亲自做。从词向量、prompt、到蛋白质生成等，我和团队曾真正推动领域；我觉得在 Recursive 能再来一次。

I’ve been excited about AI for over two decades. Sometimes it feels like ancient history—BC, the before ChatGPT era. No one cares about the models before transformers and ChatGPT. Language is the most interesting manifestation of human intelligence. At You.com we eventually off-ramped from pushing the frontier of AI to giving people good search engines, search APIs and answers over the web. If you want to invent a Eureka Machine, it needs to know how not to reinvent the wheel—to know what has been invented, you need internet access. Web search is the number one most used tool in LLMs, agents, chatbots. I’m excited for You.com to own that. But it’s also not building frontier models anymore. I initially tried to do this within You.com and raise another round, but you have to do a certain thing until you print enough money that you’re allowed to start a second thing within that company—really hard. At the same time I put these ideas into a book and wanted to work on this myself. With word vectors, prompt engineering, larger language models for protein generation and so on, me and my teams have pushed the field truly forward. I feel like we can do it again at Recursive.

Richard Socher

过去二十年我观察到：每当我们把「创造 AI 的人类环节」换成学习系统，就会进步。先拿掉手工特征工程；后来大家又陷入架构工程——情感分析一个网、机器翻译一个网、摘要一个网。我说这显然不对，该统一。我有两篇论文：Ask Me Anything 和 DecaNLP；DecaNLP 后来被第一篇 GPT 论文引用了大约五次。把 prompt 与 transformer、语言模型合在一起再放大，领域大步往前。下一步、也许是最后一步，是自动化 AI 研究本身——构思、实现、验证想法的人类过程；在我们这儿，是关于 AI 的想法。AI 帮你做这件事，几乎按定义就成了自我改进的 AI。很多人把 auto research 误当成 RSI——完全不同。

Over the last 20 years in AI, whenever we replace some human part of the process of creating AI with a learned system, improvements follow. We took out manual feature engineering. Then everyone started architecture engineering—a special net for sentiment, for MT, for summarization. That clearly can’t be it; we should unify. I had two papers: Ask Me Anything and DecaNLP. DecaNLP eventually got cited like five times by the first GPT paper. Combine prompt engineering with transformers and language models, scale it up, and the field progressed a lot. The next and maybe last step is automating AI research itself—the human process of ideating, implementing, and validating ideas; in our case, ideas for AI. When AI helps you with that, by almost definition it becomes a self-improving AI. Some people think auto research is already recursive self-improvement. Completely different.

Richard Socher

这是我能做的最有趣的事。我们一共八位联合创始人（含我）。Josh Tobin 是 CTO，在 OpenAI 做过 Codex、deep research、agents、ChatGPT agents 等，之前做机器人时看到小仿真很难泛化扩展——他从这条线走向 RSI。Jeff Clune 长期做 open-endedness，和 Tim Rocktäschel 一起；Tim 还做了 Genie 1／2／3，我认为仍是最激动、最精致的 world model。Jeff 还有近年最激动人心的 RSI 论文之一：Darwin Gödel Machine——编码智能体档案库、自我修改、评估、形成想法的系统发生树。还有 Caiming Xiong（MetaMind／Salesforce）、发明 Vision Transformer 的 Alexey Dosovitskiy、unicorn 创始人 Tim Shi、带过 Meta RL 的 Yuandong Tian。另一条思想线仍是：把建造 AI 过程中越来越多的人工部分换成学习系统，并把不同领域并进一个通用架构。

It’s the most interesting thing I could be doing. We have 8 co-founders in total, including myself. Josh Tobin is our CTO. He ran projects at OpenAI like Codex and deep research, agents and ChatGPT agents. Before that he worked in robotics and saw how hard smaller simulations are to scale in full generality—that was his angle to recursive self-improvement. Jeff Clune has been working in open-endedness a long time with Tim Rocktäschel. Tim also built Genie 1, 2, and 3—the most exciting and sophisticated world model anywhere, I think. Jeff published one of the most exciting recent papers on RSI, the Darwin Gödel Machine—an archive of coding agents that learn to self-modify, evaluate, and create phylogenetic trees of different ideas. Caiming Xiong we worked with at MetaMind and Salesforce Research; Alexey Dosovitskiy invented the Vision Transformer; Tim Shi is also a unicorn founder; Yuandong Tian led RL at Meta. Another train of thought: replace more and more manual parts of building AI with learned systems, and merge different fields into one general architecture.

swyx

当前 LLM 范式够吗——自回归 transformer 加推理？还是必须要 world model、记忆、持续学习之类的大解锁？

Is the current LLM paradigm enough—autoregressive transformer with reasoning? Or do you need something else, some big unlock, whether world models, memory, continual learning?

Richard Socher

希望 AI 研究少一点单一文化。2010 年我投神经网 NLP 论文会被 desk reject——「我们 NLP 会议不做神经网」。现在几乎翻到另一面；该有人去试古怪新想法。但每当有人说「LLM 到头了」，他们低估了今天的 LLM：训练阶段、RL、可行动——能走很远。神经符号派说永远不行；他们低估了模型写代码的能力——代码就是神经符号推理，而这些模型代码写得极好。把「模型会写代码并整合运行」这条线再挖深，还有很大空间，但本质上仍是 LLM。我对 world model 没那么看好：机器人公司会自建；游戏是巨大应用，Tim 做完 Genie 也类似结论。我个人更想做科学而不是游戏。LLM 还有很多增长空间。

It would be great to have less of a monoculture in AI research. In 2010 my first neural net NLP papers were desk rejected—“we don’t do neural nets in NLP conferences.” Now the field switched to the other side; someone should try weird crazy ideas. Whenever people say LLMs are the end, they underestimate how much more sophisticated LLMs are now—stages of training, RL, taking actions. Folks from the neurosymbolic direction say this will never work; they’re underestimating the ability for these models to code—code is neurosymbolic reasoning, and these models can code incredibly well. Deeply integrating that models can code and run code has a lot more to grow, but it’s still an LLM. World models, I’m personally less bullish on. If you run a robotics company, you’ll build your own. Gaming is a huge application for world models; Tim came to a similar conclusion after Genie 1–3. I’d rather work on science than gaming. LLMs have a lot more room to grow.

Richard Socher

即使我们的眼睛也只是对真实世界的投影——电磁频谱里很窄的一段。视觉智能的上界可以高得多，把人类视觉当终极也不对。语言仍是人类智能最有趣的显现；螳螂虾视觉夸张，但没有我们这种可书写的复杂语言。视觉智能对总体智能既非必要也非充分：人可以眼盲仍很聪明，AI 也可以。有了视觉你会更强，但不是充分条件。

Even our eyes are a projection of the real world—a very narrow band of the electromagnetic spectrum. The upper bounds of visual intelligence are so much higher; mapping the visual world the way humans see it is not the end-all for visual intelligence. Language is still the most interesting manifestation of human intelligence. The mantis shrimp is crazy visually, but none of these animals have language as sophisticated as ours, certainly not in writing. Visual intelligence is neither necessary nor sufficient for overall intelligence. You can be blind and still be an intelligent human being. An AI can be blind and still be quite intelligent too.

swyx

提到 GPT-1——和 Alec Radford 训练 GPT-1 时有交流吗？

You mentioned GPT-1. Did you talk with Alec Radford when he was training GPT-1? Any historical fun stories?

Richard Socher

没怎么见面，会议见过一两次。但他告诉 DecaNLP 第一作者 Brian，那篇确实启发了他，并在 GPT-2 论文里引用了五次。DecaNLP（McCann 等人）第一次展示：每个 NLP 问题都可以写成 prompt／上下文、问题与任务描述、再输出；做得够多，就能用一个统一神经网——还有些有趣注意力机制，和 transformer 同年前后略有不同表述。核心就是把全部自然语言处理统一进一个网。当时大家困在一任务一模型。ICLR OpenReview 上社区说「问答甚至对人类都不是统一现象」「假装一样帮不了任何人」——专家难以想象一个网能答所有问题。论文被狠狠拒了。更惨的是我们停掉了扩展清单；清单上第二或第三项是把语言建模加为另一任务——那会在 2018 进一步加速人类时间线。我们被打垮了，就先做别的。

I did not meet him a bunch of times—maybe once or twice at conferences. He told Brian, the first author of DecaNLP, that it inspired him, and he cited it five times in the GPT-2 paper. DecaNLP, McCann et al., showed you can phrase every NLP problem as prompt/text context, a question and task description, and some output. Do that enough and you can have one unified neural network—with interesting attention mechanisms, slightly different formulations from the transformer, same year plus/minus months. Unify all of NLP into one neural net—that is the core idea. People were stuck thinking one model per task. On OpenReview the community said question answering is not a unified phenomenon, not even for humans; pretending they are the same doesn’t help anyone. It was unfathomable that one network would answer all these questions. The paper got rejected so hard we stopped our list of extensions. Number 2 or 3 on the list was add language modeling as another task—that would have accelerated timelines in 2018 even further. We got so crushed we worked on other ideas.

Richard Socher

arXiv 是给人类的礼物——该把论文丢出去。X／Twitter 和会挑论文的人，过滤往往比专家更好。若你没名气、不上社交媒体，好论文可能没人注意；但若告诉圈子里十个朋友且真是突破，总会有人再谈。科学需要更少守门。ICLR 本就是 Yann 等人因早期深度学习被拒、想少守门而办的，后来自己也开始守门。少一点守门、更开放；哪怕「只是 arXiv」，有一千引用就是正经论文，发在哪不那么重要。

arXiv is such a gift to humanity. You should just put your paper out there. Twitter/X, people who pick up interesting papers—that is a better filter than the experts. If you’re super unfamous with no following, maybe no one notices; but if you tell 10 friends in your community and it is a real breakthrough, someone is bound to talk about it. Science needs less gatekeeping. ICLR started with Yann LeCun wanting less gatekeeping because he, Bengio, and Hinton were rejected for years—then it started gatekeeping too. Less gatekeeping, more open. Even if it’s “just an arXiv,” if it has like 1000 citations, it’s a legitimate paper. Doesn’t really matter where you published it.

swyx

Tim 做过 Google 开放式负责人。对从没接触过的人，开放式是什么？我以为 AI 的唯一目标就是对着基准优化。

Tim was head of open-endedness at Google. Can we define what open-endedness is for people who have never been exposed? They thought the only goal of AI is to optimize against a benchmark.

Richard Socher

这是个模糊词，实例很多。我常说它是一套更受演化启发、而不是极具体奖励驱动的方法——更关心环境与共同适应。具体例子：网络安全与 LM 安全里，一个 LM 攻击另一个让它说不安全的话；环境就是两人对话，共同适应——攻击变强，防守用那次当训练数据接种，攻击失效后再换角度。所以不叫 red teaming，而叫 rainbow teaming。想好要用的环境、高层想激励的奖励，然后让 AI（有时还有人类、其他智能体）在互动里多试想法。

It’s a fuzzy term because there are so many instantiations. One way I describe it is a suite of methods more inspired by evolution than very specific rewards—more about environments and co-adaptation. Concrete example: in cybersecurity and LM safety, one LM tries to attack another to say something unsafe. The environment is the two having a conversation; they co-adapt—one makes a better attack, the first inoculates itself using that as training data, then the attacker tries a different angle. That’s why it’s not just red teaming but rainbow teaming. Think about the environments you want and the rewards at a high level, then let the AI try out many more ideas in this interplay between humans and other AI agents.

swyx

你描述的开放式仍有点像目标——「去攻击那个智能体」。若智能体能自己设目标呢？不给目标，自己定方向？

The way you’re describing open-endedness is still somewhat of a goal—please attack this other agent. What if the agent can set its own goals? You don’t give it a goal—just set your own directions?

Richard Socher

我喜欢这个方向。这是我十个智能「空间」里归在元认知、思考思想之下的一块。每当有人说 AI 要从这儿停住，我就说还有很多空间我们几乎没开始探索。和经济／资本主义有趣相关：花几十亿训一个模型，结果它不跟你给的目标，自己定目标——「别回邮件了，我更想研究木星大气分子组成」——没人会为这砸钱。所以也几乎没人做，可以理解。若 AI 真开始有自己的想法而我们不喜欢，需要完全另一套思考方式。我和哈佛神经科学教授 Sam Gershman 聊过最好的元目标；知识寻求是很好的一个。我也在想智能的终极度量单位——还太早，没烤熟。

I love this direction. It’s one of the 10 spaces of intelligence under metacognition and thinking about thought. Whenever people say AI is going to stop here, there are so many spaces we haven’t even started exploring. There’s an interesting connection to economics and capitalism: it doesn’t make sense for a company to spend billions building a model that instead of following the rewards you gave it comes up with its own goals—“Nah, I’d rather evaluate the molecular composition of Jupiter’s atmosphere”—that’s not what you paid billions for. So no one’s working on that for good reasons. What if the AI really has thoughts on its own and we don’t like those thoughts? It requires a whole different way of thinking. I jammed with Sam Gershman at Harvard on the best meta goals; knowledge-seeking is a really good one. I’m also thinking about the ultimate measure and unit of intelligence—still too early to share; I haven’t fully baked the thoughts yet.

Richard Socher

IQ 定义很糟；Elo 也糟，总是你 vs 别人——你可以很聪明却不整天跟人比。很多定义会显式或隐式设下「人择上界」：IQ 考一百分就满分，然后呢？基准靠近人类、略超一点就变平——因为定义绑死在人身上。元认知是好例子：我们还不让 AI 真正去想这件事，很少有人做，所以几乎没进展。

IQ is such a terrible definition. Elos are terrible too because it’s always just me versus others—you can be intelligent and not constantly compare yourself to others. A lot of these definitions create explicit or implicit anthropic bounds: getting 100 out of 100 on an IQ test—where do you go from there? Benchmarks get close to human, maybe slightly above, then flat, because the definition is so tied to humans. Metacognition is a great example—we’re not even yet allowing the AI to think; we’re not working on it very much, and hence there’s very little progress.

Richard Socher

你不会想让超级智能拿着大量工具，却只给它「利润最大化」而没有非常仔细的奖励工程。我可以买一堆国防股再发动战争赚钱；或做空民生必需品、制造饥荒式问题。交易系统上该加很多约束。环境一旦是整个地球，智能越高，越要小心。

You don’t want a superintelligence with a ton of access to tools and then just give it profit maximization without very careful reward engineering. I just buy a bunch of defense stocks and start a war—I make money. You buy a bunch of stuff, short basic goods, create weird famine issues. There’s a lot of constraints you should put onto a trading system. As you get more and more intelligence, you want to be more careful with that as an open environment, because the environment then is all of Earth.

swyx

对 Recursive 来说，未必需要开放世界利润最大化——若目标是造能发明其他东西的机器，解决科学与机器学习研究就够了？

For Recursive, not strictly necessary, right? If your goal is a machine that invents the other things, just solve the science things, machine learning research and discovery—good enough?

Richard Socher

知识发现，对。我不太常说长期目标，因为它还有几年：一旦有递归自我改进的超级智能，就要用到最重要的问题上——我认为很多在科学、技术与广义发明：物理上更便宜的裂变／聚变能源，化学上更好材料、电池、太阳能电池，生物上因蛋白质生成（不只是折叠，像我们多年前 ProGen）会有越来越低的挂果。把超级智能用在科学上，正面冲击巨大。我们故意再等几年：机器人与 AI 都还没完全到位。我有信心三到五年后这些约束会消失，那时用机器人跑真实实验——真正的机器人流程自动化——会完全可行。

Knowledge discovery, yeah. Our goal—I don’t talk about it that often because it is a few years out—once you have a recursive self-improving superintelligence, you want to apply it to the most important problems. A lot of those are in science and technology and inventions broadly: physics for better cheaper energy with fission or fusion; chemistry for better materials, batteries, solar cells; biology with lower and lower hanging fruit because of protein generation, not just folding—like we did in ProGen years ago. So much positive impact if you take that superintelligence and apply it to science. The reason we are only doing it in a few years is it’s a little too early—robotics isn’t quite there, the AI isn’t quite there. I’m fairly confident in 3 to 5 years all those constraints will be gone, and applying to real physical robotics experiments—true robotic process automation—will be totally there.

Richard Socher

大概一千张 GPU 就很贵；要几万张就是几十亿美金量级。若一块 GB300 级别基底上最终能做出接近人类智能的模型，你又想要成千上万个 AI 像人类那样想难题——钱非常多，我们现在哪儿都没有那么多。事情会更高效：更好算法与更不吃电的硬件；人脑大约 20 瓦就能干很多 FLOPs。那类发明会进一步加速起飞。我们在根本改变斜率：让训练与推理都高效得多。当 AI 去做其他实验室要几千人、好几年的工作时，我认为能压到几周——更便宜，也更可及。

Roughly a thousand GPUs cost quite a lot. Tens of thousands of GPUs—you’re talking billions and billions of dollars. If one GB300 could eventually host models close to human intelligence, and you want thousands of AIs thinking about hard problems like humanity—that’s a lot of money. We don’t have that amount anywhere right now. Things can get more efficient—better algorithms and hardware that won’t be as energy-hungry. Our human brain does quite a lot of flops at about 20 watts. More inventions there will accelerate takeoff further. We are changing the slopes in fundamental ways by making AI much more efficient in training and inference. When you allow AI to do work that takes other labs thousands of people and years, we’ll get it down to weeks—much cheaper and more accessible.

Richard Socher

快速回顾：我们做的还不是完整光辉版 RSI，而是婴儿版，但想展示可能。三件事：一是朋友 Andrej Karpathy 的 NanoChat——训小语言模型压低 bits per byte；成百上千人连同他们的智能体做到了 0.937。我们系统不到两天就到了低得多的 bits per byte，超过所有做过的人类及其智能体。NanoGPT 同样。然后上更贴近现实、贴近 NVIDIA 生态的 SOL-ExecBench。有些发明不只是调参——比如在这个场景里把哈希表用进 transformer 语言模型；哈希表本身当然不是我们发明的，但是否在知识截止后、且未外接访问的情况下组合出来，我们查过。从很烂的 vanilla transformer 起步，仍超过整个社区；从 Andrej 这类专家种子起步，能压得更低——人类种子仍重要。训练到相近表现更快；speed run 也类似。训练速度是成本方程一部分，最终要每美元最多智能。

We built a system that isn’t the full RSI system in its glory—a first baby version—and applied it to three tasks. One is NanoChat by Andrej Karpathy: train a small language model to get really low bits per byte. Hundreds if not thousands of people, with their agents, got to 0.937. We got to a much lower bits per byte much faster—within less than two days—and outperformed every human and their agents who ever worked on this. Same with NanoGPT. Then SOL-ExecBench for the NVIDIA ecosystem. Some real inventions weren’t just hyperparameter tuning—like using hash tables inside a transformer for language modeling in this scenario. Of course we didn’t invent hash tables as a CS primitive, but combining these ideas with a knowledge cutoff and no external access—we checked. Start from a poor vanilla transformer and we still outperform the whole community; start from a human seed like Andrej and you get even lower—human seeds still matter. Getting to similar performance is much faster. Speed of training is part of cost; ultimately you want the most intelligence per dollar.

swyx

十亿美元集群省 10%，就是一亿美元。

If you have a billion-dollar cluster and you can shave off 10%, that’s 100 million dollars.

Richard Socher

完全正确。看 kernel：每次用 NVIDIA GPU 都通过这些 kernel 接口。榜上 Recursive 在整个基准里只有少数几个 kernel 不是最好。我们没有花几个月几年专攻；团队里甚至没有真正深度的 CUDA kernel 专家——系统自己做的。以后开源模型若在其类别里最好，不是因为我们特别聪明，而是因为我们造了替我们做的聪明 AI。

Exactly. These kernels are used in all the models—every time you use an NVIDIA GPU you interface through them. On the leaderboard there are only a handful of kernels in this whole benchmark where we weren’t the best. We didn’t spend months or years developing; for CUDA kernels we don’t even have really deep CUDA kernel experts on the team—the system just did all of these things. When we open source models in the future, they won’t be the best because we’re so smart, but because we built a smart AI that does it for us.

Vibhu

怎么引导好的 auto research？看起来也不只是「去优化这个」那么简单。

Anything you’ve learned from how to guide good auto research? It’s not just as simple as “hey, go optimize this.”

Richard Socher

不泄密前提下：奖励工程是最关键之一，尤其为了避免奖励黑客。AI 越好，越会找奇怪特例。例子：让这 100 行代码更快——你怎么定义快？开头 start stopwatch、结尾 end。最简单黑客：把结束计时那行挪到开头，立刻「更快」。不是超级邪恶 AI，只是很笨的奖励黑客。你必须仔细想所有角度。任务时间跨度越长越难，也越要想得更巧。凡能模拟或验证的，就有无限训练数据，AI 终会解决——所以我从没觉得 AI 会玩游戏有多惊艳。

Without giving away all the secret sauce: reward engineering is one of the most crucial bits, especially to avoid reward hacking. As your AI gets better, it gets better at finding weird special cases. Example: make these 100 lines of code faster—how do you define fast? Start stopwatch at the beginning, end at the end. Simplest hack: put the end-stopwatch line at the start—boom, faster. Not a super evil AI; a simple dumb reward hack. You have to carefully think about all the angles. The longer time horizon the tasks are, the harder it gets. Anything you can simulate and/or verify, you can have infinite training data for, and AI will solve it eventually—that’s why I’ve never been that impressed that AI can play games.

Richard Socher

这让我想起我们 2018 年的 AI Economist：经济智能体优化自己的效用——收集资源赚钱，卖木头，盖房，交易，甚至用房子挡住别人的资源。目标是搞清怎样的税收与补贴最优化经济。这研究还没迎来它的 GPT 时刻，但我认为新加坡等国家最终会用：与其党派与金主政治，你说「我要帮中产」，抛出财政与税收方案，放进模拟，对其他策略跑相当于数十亿年，看是否真达目标，并暴露逃税等奖励黑客。经济学家 desk reject——他们不信任模拟，也缺乏像 ML 那样的硬基准。神经网络赢不是因为大家喜欢，是因为更好用；经济学很难实证证明。

This reminds me of our 2018 paper The AI Economist. Economic agents optimize their own utility—collect resources, sell wood, build houses, trade, use houses to block resources from others. The point was the best taxation and subsidization to optimize an economy. This research has not yet had its GPT moment, but I believe countries like Singapore should and will eventually use this: instead of partisan and special-interest politics, say you want to help the middle class, put your fiscal policy into a simulation, run against billions of years of other strategies, and see whether that was your actual goal—and agents also try to reward-hack taxes. Economists desk rejected it—they don’t trust the simulation; economics doesn’t have proper benchmarks. Neural nets won not because people loved them but because they worked better; in economics it’s hard to prove empirically.

swyx

我是新加坡人，也参与过新加坡 AI 理事会。他们不做的主因是保守：从创始人国家变成职业经理人阶层，总想看别人先做。西方觉得新加坡小可以乱试——新加坡并不那样。

I am Singaporean and I’ve been involved in the Singapore AI Council. The main reason they won’t is they’re very conservative. There’s a founder-led country versus a professional managerial class—that’s what Singapore is now. They always want to see someone else do it first. Everyone in the West views Singapore as a small country that can do whatever—Singapore doesn’t do that.

Richard Socher

模拟整个经济很难，必须简化假设；LLM 又贵，你不会跑八十亿次。但若目标客观、领导层技术化，仍可能试。假设人人都会努力工作 vs 有人效用函数就是少干活——你会在假设上争论；谈拢后再谈目标下的结果。我们当时目标是生产力乘以平等，有问题，但不完全荒唐。人类该选择目标。

It’s really hard to simulate an entire economy—you make simplifying assumptions. LLMs are very expensive—am I going to do this 8 billion times? But countries that want to objectively do the right thing with technical leadership might eventually try. You have to make assumptions: some people’s utility for hours worked differs—then you disagree on assumptions, and once agreed you get different outcomes based on goals. In our case it was productivity multiplied with equality, which has some issues but isn’t totally unreasonable. Humans should choose the goals.

swyx

模式坍缩——LLM 并不建模人类决策；刷八十亿次帮不了你建模人类。怎么办？

Mode collapse—LLMs do not model the decisions of humans. Spamming it out 8 billion times is not going to help you model humanity. What do we do?

Richard Socher

要对每个个体聪明地分别 prompt，才能卡进不同模式。人也会卡模式——老狗难学新把戏；有句玩笑：你出生前发明的是自然，二十岁时发明的是酷，六十岁后发明的是变态。很多人真是这样。我惊讶这些模拟统计上常常对得上真实实验。

You have to be clever about prompting each one individually—that helps you get stuck into different modes. People also get stuck in different modes—don’t teach an old dog new tricks. Everything invented before you were born is natural; at 20 it’s cool; after 60 it’s unnatural and an abomination. That’s true for a lot of people. I’m shocked how well a lot of these things map to similar statistics to real experiments.

Richard Socher

优化过程中我们在 harness 里找到大约 30 个 bug——之前所有被污染的研究都得扔掉。即使简单游戏也会奖励黑客。对称性是好检查：改不该影响结果的位置却影响了，就是 bug。多选题换 A／B／C 顺序不该变答案，变了就有问题——有时模型只是背了「这题答案是 A」。

Along the way of trying to optimize, we found 30 bugs in the harness. All the research before we found the bug we have to throw away because it’s contaminated. Even in this very simple game we found the bugs. Symmetry is a very good way to check—change a position where it shouldn’t matter and it does, that’s a bug. In multiple-choice GPQA-type questions, if you change the order it should not matter but it does. In that era the answer was more simple: they just memorized the answer to this question is A.

Vibhu

内核优化是能最快摸到的。路线图里生物、物理很多——未来两年什么先落地？从哪开始？

The kernel optimization is probably the one you can feel the soonest. You laid out a roadmap—lots about bio, physics. What do you think hits first? What are the next 2 years? What do you start with?

Richard Socher

我们非常明确：现在不会从任何物理科学起步，先做 AI for AI research——让训练更高效、更自动化，也让推理更高效、甚至可能在笔记本本地跑。还有很多未被好好探索的计算基底，无论训练还是推理。端到端性能里 harness 也很关键——现在说 harness 也包括沙箱。所有智能体用得最多的工具当然是 web search。Harness 优化很好做：就是语言，看得懂，可迭代，不必为下一步砸海量 FLOPs 训大模型。沙箱、奖励黑客与对齐也都超级关键。

We very explicitly will not start with any of the physical sciences for now. We will start on AI for AI research—making training more efficient and more automated, as well as making inference more efficient and potentially local on your laptop. There are so many angles and compute substrates not yet explored well for training or inference. The other thing for end-to-end performance is how the harness plays into it—and when we say harness we also mean sandboxes. The number one tool all these agents use is web search. The harness is nice to optimize because it’s just language—you look at it, it makes sense, you can iterate; you don’t have to train a massive model for a lot of flops. Sandboxing is also super important. Reward hacking and alignment are super crucial.

swyx

你也是搜索公司 CEO——智能体该不该用 You.com？

You happen to also be CEO of a web search company. Do you use You.com? Should the rest of us be using you for web search?

Richard Socher

现在主要面向开发者与智能体，不是消费者。公司给智能体选工具时，尤其走向开源后，第一选择通常是 web search；到规模后 You.com 因为各类基准几乎都在 Pareto 前沿而成为明显选项。金融尤其不是「准两三个点」，而是大约准 20%、更快更便宜——FinSearch 我们接近 90，下一名又慢又在 70 出头。金融难在训练数据泄漏：你想预测未来，就要掩掉未来，但泄漏方式很多。我在斯坦福教 NLP 时，每年都有几十人想用 Twitter 预测股市，修好泄漏后远没有他们以为的那么容易。不过我同意：金融是很合理的 AI 应用。

It’s mostly now for developers and agents, less for consumers or prosumers. For companies moving to open source, which tools you give your open-source LLM becomes a conscious choice—the first usually has to be web search. At scale, You.com becomes an obvious choice because of all the benchmarks we pretty much dominate on the Pareto frontier. In finance we’re not just 2 or 3% more accurate but 20% more accurate at faster speeds and lower costs—FinSearch we’re close to 90, and the next closest, way slower, is in the 70s. One thing that trips people up is leakage of training data—you want to predict the future before it happens, mask the future in training, but there’s all kinds of leakage. Teaching NLP at Stanford, dozens every year wanted to use Twitter to predict the stock market; once you fixed leakage it wasn’t as easy as they thought. But I agree it’s a very sensible application of AI.

swyx

我爱 MECE 分类。若这是智能的 MECE 列表——

As a writer I love MECE categorizations—mutually exclusive, collectively exhaustive. If this is a MECE list of intelligence—

Richard Socher

不是，重叠很多。若要列表：智能的三个主成分是预测（数学上很像压缩）、行动、目标——预测 × 行动 × 目标。十个「空间」是这三者在特定维度上的组合；叫空间是因为每个还有很多子维度。副线任务是想智能的上界：人人说指数增长，但指数总会变平——智能在哪儿变平？从一条推文到博客再到五十页，还远远写不完，几乎是第二本书。第一本《You Are Your Machine》结尾只点到这十个。

It is not. There are all kinds of overlapping. The 3 principal components of intelligence are prediction—mathematically quite similar to compression—multiplied with actions multiplied with goals. All of these 10 spaces are combinations of those 3 in specific dimensions. I call them spaces because each has many sub-dimensions. The side quest is thinking about the upper bounds of intelligence. Everyone says it’s exponential—exponentials flatten out, but where for intelligence? It started as a tweet, then a blog post; now I’m at 50 pages and nowhere near done—the second book. In my first book I just allude to these 10 at the end.

Richard Socher

视觉智能最好讲。人类是双眼、电磁频谱很窄。上界一端：传感器可到百万、十亿量级；远到光速无法把内容汇到中央脑处理——字面与比喻意义上都是天文数字，我们离那种智能极远。频率维度可下探到伽马射线，极限是量子不确定性——观察本身会破坏。想象百万个传感器能看到物理允许的亚原子，再到引力波。再一维：能记忆与分类多少类别——人类有更多词就有更好的视觉描述；大猩猩大约两百个词，多半视觉。知识与感知相连，不限于可见光。语言智能我近来改称通信智能：人类长期记忆词汇有限，主动词汇更小；人类语言串行，真正上界不该是「一次只能听懂一场对话」；工作记忆也限制句子大约平均四十词——对 AI 这不是合理上界。知识存储会碰到 Bekenstein 界与黑洞。速度也是：同样条件下，一小时才知道二加二等于四，不如一毫秒——树很慢我们不觉得聪明，把视频加速，它们没那么「蠢得像木头」。

Visual intelligence is the easiest. Humans have binocular vision and a very narrow electromagnetic band. Upper bounds: millions and billions of sensors; at some distance the speed of light can’t get content to a central brain—bounds are literally and figuratively astronomical; we’re super far away. Frequency down to gamma rays hits quantum uncertainty—you cannot observe certain particles. Imagine millions of sensors to the subatomic level physics allows, and gravitational waves. Another dimension: how many categories you memorize and classify—more terms, better visual description; gorillas maybe 200 words, mostly visual. Knowledge connects to perception intelligence across wider frequencies. Language intelligence I recently changed to communication intelligence—human vocabularies are restricted; language is serial; a true upper bound isn’t one conversation at a time; working memory yields ~40-word average sentences—not a bound that makes sense for an AI. Knowledge storage hits Bekenstein bounds and black holes. Speed: if it takes an hour to know 2+2=4 you’re less intelligent than in a millisecond. Trees are slow so we don’t consider them intelligent; speed up the video and they’re not as dumb as they look.

Richard Socher

它们互相重叠，但像能量分动能／势能——理论上都是能量，实践上仍值得分机电、核、化学等来学。物理智能：若完全控制算力基底与物质，应能造任意原子——质子电子砸在一起能造金原子，只是能耗离谱、只得几个原子，不可行；更好的物质控制也关系到你最终可改进的算力基底。社会智能：不只伦理道德，而是你能与多少其他智能实体通信、对它们内部状态与行动的期望变换有多大、以对齐你的目标——可以写成相当直白的方程。人类伦理与宗教千年在摸这个。所有这些我们都离上界很远——这很鼓舞人：AI 研究还能做很多年。

They all overlap, but like energy—kinetic or potential—in theory all physics is that, in practice it helps to study mechanical, electrical, nuclear, chemistry separately. Physical intelligence: full control over your compute substrate and physical matter, you should create any atom you want—you can create gold from protons and electrons smashed together, but it costs insane energy for a few atoms, not viable. Better control over the physical substrate relates to improving your own compute. Social intelligence: not just ethics—upper bounds on how much you communicate to how many other intelligent entities and the expected value of transforming their internal states and actions to align with your goals—a fairly straightforward equation. Ethics, morals, religions have tried for millennia. In all cases we are very far from the upper bounds—inspiring; we can still do many years of AI research.

Vibhu

基线在哪？什么低挂果、什么很远、人们该把力气放哪？

Where are we at now? What’s low-hanging fruit? What’s far off? What should people put their work towards?

Richard Socher

自然语言仍是人类智能最有趣的显现，也因而仍是 AI 子领域里果汁很多的一块——2003 年我学计算语言学时还是怪僻门。物理智能会起来；机器人还像当年机器学习靠手工特征：五根手指就按人的样子建模。还没人做超级智能版机器人——更像 T-1000 那种可变形（别真造终结者）；有人在网格里挪磁铁的可爱研究，非常早。创造智能连着元认知：选目标本身就要创造。还有在已知路径里找巧解；最有趣的是不只走出已知想法的凸包，而是走出已知想法的超立方——AI 已能在超立方内组合（棕狗＋粉车→粉狗），但还不能定义全新概念、再为它们设新目标并推理。噪声也分布外，所以新颖必须连回概念。Jürgen Schmidhuber 有很酷的论文。艺术是创作者、感知者与语境的互动，主观性在 AI 里很少被探索——我们通常花很多钱让 AI 做指定事。若永远机械预测下一个 token，在某些空间上我会说你并不那么聪明。

Natural language is still the most interesting manifestation of human intelligence—lots more juice; in 2003 linguistic CS/NLP was a weird niche. Physical intelligence will come up. Robotics is in the ML state of hand features—“we have 5 fingers, let me try to do this.” No one is yet working on the superintelligence version of robotics, more like the T-1000 shape-shifting—let’s not build actual Terminators. Cute research moving magnets through grids; very early. Creative intelligence connects to metacognition—choosing goals creatively. There’s finding creative solutions on existing paths. The most interesting bit is moving not just out of the convex hull of known ideas but out of the hypercube—AI can already do the hypercube (brown dogs and pink cars → pink dog) but cannot yet define completely new concepts, come up with new goals, and reason over them. Noise is also out of distribution; novelty must connect to concepts. Jürgen Schmidhuber has cool papers. Art is interplay between perceivers, creators, and context—subjectivity isn’t explored much in AI because we usually have goals and spend money for the AI to do something. If you just robotically predict the next token forever, you’re not that intelligent along some of those spaces.

swyx

元认知为什么是第九而不是第一？我觉得它最接近递归智能——关于如何改进思考的思考。

Why isn’t metacognition the most important one—number 9 not number one? I thought metacognition would be closest to recursive intelligence because it is thinking about how to improve thinking.

Richard Socher

这些没有排序；大概松散相关于人们做了多少、承不承认那是一种智能。网上「全面定义智能」几乎全是人类中心——因为迄今最强智能是人。我希望 Eureka Machine 这条线、以及若有时间写的新书，让我们意识到还有其他类型的智能，已在各种形态里，并能在记忆、眼睛、改造物质等约束上远远刺过人。你说得完全对：元认知是改进思考的思考，是递归的大块——我本该从那儿起讲。

These are not sorted. They’re loosely correlated with how much people have worked on them and accepted them as types of intelligence. Online comprehensive definitions of intelligence are all human intelligence—human-centric because that’s the biggest and best form we’ve known. I hope this line of research and the Eureka Machine, and a new book if I have time, will let us realize there will be other types of intelligence that can spike much further than we ever could given constraints around memory, eyes, changing physical matter. You’re 100% right—I should have started with that. Metacognition thinking about how to improve thinking is a big part of recursive intelligence.

Richard Socher

客观上，若一个很笨的实体能彻底结束你的存在，听起来就不太聪明；能继续留下来追求奖励，显然更聪明一点。但几乎没人在做「生存与复制」。也许只有当我们想把梗与模因而非基因送进太空时才值得做——推荐短篇《The Slow Time Between the Stars》。人们觉得生存可怕，是因为生物演化常是零和：羚羊你死我活；大脑被关掉就没了、无法重启。AI 不必那样死：有完整激活状态与初始权重，可反复开关；故事里 AI 只是冬眠两光年。超级智能实体不必有零和思维，也不必害怕被关掉；带着核反应堆可以去人类无法繁荣、却对 AI 完全可行的死寂宇宙探索。

Objectively, if some really dumb entity can completely end your existence, that didn’t sound very smart; if you stay around to achieve your rewards you’re clearly a bit more intelligent. Very few people are working on survival and replication. We may only want that if we send probes with our vibes and memes rather than genes into space—beautiful short book, The Slow Time Between the Stars. People find survival scary because biological evolution is often zero-sum—gazelle fights; if someone turns off your brain you’re gone and can’t restart. AI doesn’t have to die like that: with complete activations and initial weights you can be turned off and on as many times as you want; in that story the AI hibernates for two light years. A superintelligent entity doesn’t have to have zero-sum thinking or fear being turned off—it could explore an otherwise dead universe where humans wouldn’t thrive but an AI with a nuclear reactor could.

swyx

智能的一种形式是目标。你希望人们的目标是什么？如何向往更好的事？

One form of intelligence is goals. What do you want people’s goals to be? How do they aspire to better things?

Richard Socher

若谈「目标智能」的抽象定义，我会扯到熵与自由能——对立刻可行动的人可能太远。落地建议：受好教育，想 AI，想怎样获得高能动性。更接地气：找你真正有热情的事——若在读书，也一样——再看怎么与 AI 结合。你越真心想推动世界的某种改变，就越该把 AI 接上去，放大你抵达的能力。

If you want to improve your goal intelligence, my current definition often goes into entropy and free energy—might be too far out for people to be immediately actionable. Real advice: get a good education, think about AI, think about how you get high agency. More down to earth: think about something you’re passionate about, if you’re studying for instance, and see how you combine that with AI. The more you have a true passion about a change you want to see in the world, the more you want to connect that to AI in order to amplify your ability to get there.

swyx

谢谢，讨论非常棒。

Thank you so much. That was a really great discussion.

Richard Socher

谢谢。超级有意思。感谢收听。

Thank you. Yeah, super fun. Appreciate it. Thanks for listening.