人类的最后一项发明——Recursive 的 Richard Socher

At 1:09:00 we talk about the rise of AI x Finance, and AIE NYC is one month away - our hotel block is 97% sold out, get tix & travel ASAP - we will announce speakers from Bridgewater, Ramp, Coatue, Mastercard, Vanguard, Coinbase, Blackrock, Fidelity, Point72, Capital One, JPMC, Wells Fargo, Bloomberg, A24 (yes the movie studio) Labs, Two Sigma, Apollo Global, and more soon!

From helping pioneer core ideas in NLP to now building AI systems that can automate AI research itself, Richard Socher is betting that the next major step in AI is recursive self-improvement. He is the founder of You.com, AIX Ventures, and now Recursive, which has assembled some of the best open-endedness (& self improving agent) researchers in the world and raised a $4.65B seed round.

In this episode, Richard joins Latent Space to unpack his vision for the “Eureka Machine”: a superintelligence that can improve the process of invention itself, accelerate AI research, and eventually tackle major problems across science, energy, materials, biology, and more.

You can get his book “The Eureka Machine” here!

We go deep on Recursive’s early results, including an AI research system that Richard says outperformed humans and their agents on optimization tasks in less than two days, as well as work on NVIDIA GPU kernels where the system discovered improvements without relying on a team of CUDA experts. Richard also explains why he thinks AI research that currently takes thousands of people and years could eventually be compressed into weeks. These results are summarized in his 20 minute AIE keynote, where we also discuss his 10 dimensions of intelligence:

We also explore the harder questions around increasingly capable AI: reward hacking, whether Anthropic-style constitutions actually work, AI regulation and proposals to “pace” frontier development, open-source models as geopolitical soft power, whether today’s LLM paradigm is enough, and what happens if AI systems eventually begin choosing their own goals. Richard reflects on the rejected research that helped inspire Alec Radford’s GPT, open-endedness, the AI Economist, simulations of entire economies, and his framework for thinking about the upper bounds of intelligence itself.

We discuss:

- The Eureka Machine and Richard’s vision for an AI that can automate invention

- Why Richard is optimistic about superintelligence for science and technology

- Why AI hard-takeoff scenarios may underestimate physical and economic constraints

- The risks of regulating intelligence itself instead of specific AI applications

- Reward hacking and why increasingly intelligent AI makes objective design harder

- Richard’s critique of Anthropic’s constitution and constitutional AI

- Alignment vs. personalization and whose values an AI should follow

- Why open-source AI matters for resilience, competition, and geopolitical soft power

- Why Richard left You.com’s frontier-model work to start Recursive

- Recursive self-improvement and automating the process of AI research

- Whether today’s LLM paradigm is enough — and why Richard is less bullish on world models

- DecaNLP, early prompt-based generalization, and the research that influenced GPT

- Why rejected research can shape entire technological timelines

- Open-endedness, evolutionary approaches, and rainbow teaming

- What happens if AI systems begin setting their own goals

- Why simple objectives like profit maximization can produce dangerous reward hacks

- Recursive’s long-term plan to apply self-improving AI to science

- The compute, hardware, and economic constraints on AI takeoff

- Recursive’s early NanoChat, NanoGPT, and GPU kernel optimization results

- Why automating AI research could reduce years of work to weeks

- Reward engineering and what makes auto-research systems actually work

- The AI Economist and using simulations to test economic policy

- Whether LLMs can realistically simulate people and entire economies

- Benchmark bugs and evaluation harnesses and the difficulty of measuring AI progress

- Recursive’s near-term focus on AI for AI research

- Harness optimization, sandboxing, and web search as core agent infrastructure

- You.com and the search stack for AI agents

- AI in finance, backtesting, and data leakage

- Richard’s three fundamental components and ten “spaces” of intelligence

- The theoretical upper bounds of vision, communication, knowledge, and computation

- Creative intelligence, metacognition, and AI-generated goals

- Survival and replication and why AI does not necessarily need to fear being turned off

- High agency and ambitious goals and Richard’s advice for people building with AI

Richard Socher

- X: https://x.com/RichardSocher

- LinkedIn: https://www.linkedin.com/in/richardsocher/

Timestamps

00:00:00 The Eureka Machine and Superintelligence

00:02:23 AI Optimism, Slow Takeoff, and Regulation

00:07:56 AI Safety, Reward Hacking, and Anthropic’s Constitution

00:11:49 Alignment, Personalization, and Open Source AI

00:15:46 Why Richard Started Recursive

00:20:03 Recursive Self-Improvement and the Founding Team

00:22:55 Are Today’s LLMs Enough?

00:29:03 DecaNLP, GPT, and the Rejected Idea Ahead of Its Time

00:34:38 Open-Endedness and Evolutionary AI

00:36:38 What Happens When AI Chooses Its Own Goals?

00:41:16 Superintelligence for Science

00:42:40 GPUs, Compute, and the Limits of AI Takeoff

00:45:07 Recursive’s Results: AI Beating Humans and Their Agents

00:49:14 Reward Engineering and Auto Research

00:53:12 The AI Economist and Simulating Entire Economies

00:58:07 LLM Simulations, Personas, and Mode Collapse

01:03:38 Recursive’s Roadmap, Agents, Search, and Finance

01:09:13 The Upper Bounds and Spaces of Intelligence

01:30:21 Goals, High Agency, and Advice for Builders

Transcript

Introduction: Richard Socher and the Eureka Machine

Swyx [00:00:00]: We’re here in a studio with Vibhu and myself and Richard Socher. Welcome.

Richard Socher [00:00:06]: Thanks for having me.

Swyx [00:00:07]: We just talked about the Eureka Machine, or we just released a talk, at AI Engineer about the Eureka Machine. Is it — you said it’s your life’s goal. What is the Eureka Machine?

Richard Socher [00:00:16]: 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.

Swyx [00:00:45]: Yeah, I think we have the book pulled up here that you’ve written.

Richard Socher [00:00:50]: That’s right, yeah. I finished it last year, a little bit before we started Recursive, and now we’re gonna try to build parts of that.

Swyx [00:00:57]: You finished it last year. It’s July. What takes so long?

Richard Socher [00:01:01]: Oh, man, books. Books are incredibly slow.

Richard Socher [00:01:04]: It’s ridiculous. That whole industry is just unfathomably slow.

Richard Socher [00:01:07]: So a lot of the ideas have been out there for a while, but yeah, I’m really glad it’s finally coming out in September this year.

Swyx [00:01:14]: We might have AGI by then. Like, we don’t know.

Vibhu [00:01:18]: Any key takeaway that you’re most excited to put in here?

Techno-Optimism, AI Upside, and Slow Takeoff

Richard Socher [00:01:21]: Yeah. The key takeaway, I think, 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. I think there is so much more that can be done with better technology. And right now, I feel like a lot of people need, like, better marketing, not just for the future in general, but also, better marketing for technology and in particular for AI. And 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 [00:02:09]: I think you quoted the techno-optimist manifesto from, Marc Andreessen, which I think was, like, beautiful in its, ambition and clarity and simplicity almost as well.

Richard Socher [00:02:18]: I agree. Yeah. Yeah, you can disagree with him on some things, but, like, I think he’s right on the techno-optimism.

Swyx [00:02:23]: Where do you think optimists get in trouble?

Richard Socher [00:02:26]: Like, you shouldn’t have blind optimism. You should be very clear-eyed, like, especially when with such an omni, like, use type of technology as AI is, you need to think about the potential downside scenarios, especially when people use it for things that you don’t want them to use it for. It’s a little bit like the internet, and I feel like people are trying to regulate AI sometimes because of those potential downsides the way you would regulate the internet, if you were to say, “Well, because there’s bad content on the internet, like torture porn or whatever, like, we should just make it slower. That way, you can’t share the illegal content as quickly, or we should make the hard drive smaller so you can’t store as much illegal content.” But I’m like, “That’s not how you regulate that.” that’s like saying like we should regulate intelligence in the abstract. What you should regulate to avoid those downside scenarios, even as an optimist, are the specific applications. Sure, I don’t want, like, some AI surgeon to, like, practice some RL moves in my brain. It should be fully FDA certified. Sure, I don’t want any random startup to, like, drive on the highway, and cause a major accident. It should, like, have proper certifications before it’s let loose on the highway. But I feel like those downside scenarios, that some optimists sometimes maybe don’t consider enough are fairly easily regulated, compared to, what the doomers are worried about.

Swyx [00:03:54]: It — Slow takeoff is part of the strategy as well?

Richard Socher [00:03:57]: I do think, as excited as I am about, AI and its impact for society and, culture even, and certainly technology and economics and wealth and, health and all of those things, as excited as I am about all that, I do think the most bullish people on the AI hard takeoff scenarios overestimate how quickly things can move. There are hardware constraints. There are physical constraints about, the compute substrate. How quickly can you get enough, GPUs on? There are also constraints in the economy where there are a lot of industries that don’t require an insane amount of complex intelligence and complex capabilities. Like, if you think about jobs in, brands and, like, clothing and apparel and, like, handbags and stuff, superintelligence isn’t gonna make your fancy $10,000 handbag any fancier?

Richard Socher [00:04:57]: It’s like that’s — It will have no effect on the economy. You think about travel and tourism. People wanting to see the pyramids, in Egypt, it’s not gonna change that much with AI. Sure, you can, like, generative a fake, photo of you and next to the pyramids.

Swyx [00:05:12]: I can use Genie and, tour the pyramids in Genie.

Richard Socher [00:05:15]: Yeah, exactly. But, and there’s so many industries, like logging and oil. You’re not gonna magically get 1,000x more oil because, like, sure, there will be robotics, like drilling and things like that could be done, but it’s not gonna 1,000x that industry in a, like, crazy hard takeoff scenario, both on the economy, and I can go on and on about all the other examples, where that, like food and so on, where that doesn’t necessarily change that much. And then, yeah, there are real physical constraints. And then there are, of course, like, people like, off-ramping from progress. That’s one of my concerns often is that I see people in, like, Europe and other, whole regions almost feeling like they. Like many people there wanna off-ramp from progress, period. And that will also slow down, like, more improvements.

Swyx [00:05:59]: Yeah. We have this pulled up where, this is one of those things that, is very topical right now because now all the Frontier Labs are calling for the option to pace AI. They don’t say pause, they say pace. I don’t know if there’s there’s any take from you about, like, whether or not this will be effective.

Pacing AI, Regulation, and Safety Incidents

Richard Socher [00:06:17]: I think 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 that they have. Like, it would be an crazy totalitarian state

Richard Socher [00:06:37]: If every one of your GPU computes was known to some big government or multi-government agency.

Richard Socher [00:06:44]: It’s like, it’s literally if you try to regulate intelligence, it’s trying to regulate thought, and that’s ridiculous, and it’s crazy. I think it is make — it is sensible to regulate some of the applications of this technology.

Swyx [00:06:55]: Yeah. We had a bill, actual bill to regulate the number of flops in a model, and I’m like, “Okay, well-”

Richard Socher [00:07:00]: Europe done it. Like, these guys have been successful enough with their fearmongering that all of Europe has regulated itself so much before it even had a proper AI takeoff because they listened to some experts who say, “We might all die if this technology has more than this number of flops.” And they’re like, “Well, we’re good. We wanna want people to thrive. Let’s not have technology that could have a small chance of all of us dying.” And so they regulated exactly those kinds of things in the EU. And so it’s, it’s very unfortunate that there are real implications for some people when others saying, “Let’s pace while they’re sprinting as fast as possibly,” “as fast as humanly possible towards that frontier themselves.”

Swyx [00:07:43]: Yeah. It’s also not a global pause, right? Like, other nations are still accelerating at the same pace.

Richard Socher [00:07:50]: Oh, yeah.

Richard Socher [00:07:50]: You’d need a totalitarian world regime if you tried to regulate intelligence and GPUs and what people do on them.

Swyx [00:07:56]: Any takes on the safety angles of this? So 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 [00:08:11]: 100 percent. I think these are serious issues of reward hacking, and clear failures, of doing proper red teaming or rainbow teaming. I don’t know if you saw this paper from Tim Rocktäschel and a few others, where one AI, is tasked to try to hack another AI and then they can go back and forth in an open-ended fashion to inoculate themselves from those. Yeah, this is the paper. It’s a really clever idea. Open-endedness, and evolutionary inspirations are, big for us at Recursive as well. And so I wish they had used more of that. And it’s clear that, for instance, the constitutional AI. I don’t know if you remember anthropic.com/constitution. You can pull it up and search for cyber right there. It says, “Hard constraint. Claude will never ever do cyberattacks, and that is a hard constraint in our constitution.” So here are the current hard constraints on Claude’s behavior.

Richard Socher [00:09:16]: Number 3, create cyber weapons or malicious code that could cause human damage.

Richard Socher [00:09:21]: And clearly, this whole constitution was fake. Like, it clearly isn’t being adhered to at all.

Swyx [00:09:26]: Because Anthropic also found that they had in their testing

Richard Socher [00:09:30]: They’re also. Like, they’re like, “Oh, well, other people are hacking now.” There are a couple things. One, you can make a sandbox very simple, and then it’s very easy to hack yourself out of a sandbox, right? But what I think it shows is that we’re currently in this state of AI where the reward engineer still has to do a lot more careful work, and where the AI, in most cases, is not very good yet at understanding what is meant versus what is being said. And so concretely, I think this will happen if we were to have this intelligence more easily accessible in a lot of companies. Imagine you run a service center and someone says, “Oh, here’s my CSAT score and my dashboard. Make this number go up.” It’s like, “Our CSAT score is so poor.” The intelligent AI will just be like, “Oh, sure. Like, I’ll just create 1,000,000 bots that call our service center and give a 5 out of 5 rating at the end, and the number went up just like you asked for.” And you’re like, “That’s not what I meant.” “I meant with our real customers.” The AI goes off and says, “Well, easy. I’ll just give a 1000 dollar gift certificate for every failed, whatever DoorDash

Richard Socher [00:10:35]: Offer.” It’s like, “That’s not what I meant.” It’s like, “Well, but that is what you said.” And like, so I think clearly articulating what the rewards are is something we haven’t gotten very good at as humanity. And then clearly, the AI in these cases has not gotten good enough at understanding what we mean when we ask it and give it certain rewards. Now, what gives me hope is there are the first inklings, of this being better. I’ll give you an example like WhisperFlow. Full disclosure, I invested, in their seed round, but at AIX Ventures, but, WhisperFlow has gotten much better at writing what you mean and not what you say. And I think that is a sign of things to come. I think there will be more and more AIs as we make it more and more intelligent that will be better at being aligned with what is meant.

Swyx [00:11:21]: Will it be done through a constitution or RLHF or

Reward Hacking, Alignment, and What We Really Mean

Richard Socher [00:11:23]: Clearly, constitutions don’t matter at all.

Richard Socher [00:11:25]: It doesn’t work. And that was, I think, mostly marketing. I think we need to find better solutions for it. And I think at Recursive, we have a few very good ideas and some already

Richard Socher [00:11:34]: Like, ways where I think we have a better grasp on it. I don’t think we’ve fully, figured it out yet, but, we’re thinking a lot about safety, and the more intelligent the AI gets, the more you want it to be aligned, the less you want it to think about reward hacks and try to do the right thing.

Swyx [00:11:49]: I don’t know if we’ll touch on this topic, but I’m just gonna throw this question in here because it’s something that’s weighing on me. Alignment, let’s call it, 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?

Alignment, Personalization, and Cultural Values

Richard Socher [00:12:12]: It’s a great question.

Richard Socher [00:12:13]: I think you ultimately have to, of course, be aligned with laws. Like wherever your AI is deployed and needs to align with the law. I do think what AI often does is put this mirror in front of us and say, like, “This is what you’re looking like. Now I can amplify that a 1000 times. Is it still what you want?” and the truth is that different cultures made different choices. Like, in Eastern cultures, the greater good is often valued more, than the individual. Western civilization, we care more about individual freedoms and rights and the pursuit of happiness and so on, than others. And even there are gradations. There’s regulation versus litigation trade-offs. In the US, you first can often, not every time, like, FDA and so on does regulate some areas, but in many cases, the bad things happen, someone sues someone else, and then there’s a law based on that. In Europe, they try to often avoid any harm to anyone and regulate before. And both are, trying to do the best thing, but, some is more amenable to innovation than others. And so yes, you’re right. Like, I think ultimately each individual, each country, and humanity as a whole has to think about those values more, and then try to put them into laws. And that those are ultimately the constraints. And hopefully, different, societies, just like now with their AIs, will align their AIs to a different one so we have not just a monoculture of alignment.

Vibhu [00:13:46]: Here’s a follow-up on this that I wasn’t expecting to ask. Do you have takes on open source, open weight versus who owns the intelligence? So, clearly not the biggest, fan of the constitution

Richard Socher [00:13:58]: You had to do this in the topic side off.

Vibhu [00:14:00]: But it’s fine.

Vibhu [00:14:02]: Point being, any thoughts on who should own weight? Should it be open? Anything there?

Open Source, Soft Power, and Who Owns Intelligence

Richard Socher [00:14:06]: 100 percent. I am a big fan of open source. We’re gonna sign some various open source letters at, Recursive also. I think, even in the worst case attack scenarios, it is better to have more good actors have more different types of AI, accessible. I think, open source is a little bit a soft power type of thing, too. So I do think it’s good for the Western world

Richard Socher [00:14:31]: To have an answer to that, out of China. I do think, when you watch a Hollywood movie, there’s — it’s like, I don’t wanna misc, diss all of movies, but there’s a certain sense of propaganda, right? You watch one side of things, right?

Vibhu [00:14:46]: Oh, yeah. Have you seen Top Gun? Like, come on.

Vibhu [00:14:48]: Like, it’s like half of it’s paid for by the US Army or something.

Richard Socher [00:14:51]: Yeah. And so. And, I think that’s just natural. Like, but what’s interesting here is I think LLMs are essentially a similar type of soft power to movies and beyond, because they’re also, highly important for cybersecurity and so on. But one of their many aspects is that soft power of storytelling. Like, if, like a child asks an LM, like, “Tell me an inspiring story of what I should do when I grow up,” right? It’s like those are all these, like, subtle things. So I think it’s important, for Western world. I do love, individualism. I do think, despite, some of its flaws, like capitalism is the best way we have governed, found ourselves to govern, and so on. And so I do think there are various aspects that would be good, to have a Western open source answer, for LLMs. And, with Recursive, I can’t make the announcement quite yet, but we’ll

Richard Socher [00:15:43]: We’ll be relevant in that space very soon.

Vibhu [00:15:46]: Okay. All right. Exciting. I wanna bring us to Recursive. So outside of our tangents, you have a pretty deep background in the NLP space. You worked on, like, early embeddings, GloVe with Chris Manning, who was a previous guest on the podcast, You.com. What’s the history? How did you decide to start another company?

From You.com to Recursive

Richard Socher [00:16:06]: Yeah. So I’ve been excited about AI for over 2 decades now. I sometimes feel like it’s ancient history now. It’s BC, the before ChatGPT era. No one cares about all the religions that happened, before, Jesus Christ, and no one cares about the models that happened before, transformers and ChatGPT and stuff. But, like, it’s something that I’ve been deeply passionate about. I think AI is one of the most interesting things one could work on, period. I think language is the most interesting manifestation of human intelligence, too. And, at You.com, we eventually off-ramped from pushing, like the frontier of AI forward to mostly giving people, like, good search engines, search, APIs and answers over the web. I think that’s an extremely important part of intelligence, just knowledge and access, especially even, we’ll get there maybe later, if you wanna invent a eureka machine that invents everything for us, it needs to know how not to reinvent the wheel, proverbially speaking. And to know what has been invented, you gotta have internet access. So it’s the number one used, most used tool, in LLMs, agents, chatbots, and so on is web search. So I’m really excited for You.com to own that and grow really well in that with really large customers and so on. But it’s also not building frontier models anymore. And so I initially tried to do this within You.com and raise another round and so on, but you just can’t. You have to do a certain thing, and until you print enough money that you’re allowed to start a second thing within that company is really hard. At the same time, I had all these ideas. I put them into a book. I finished the book last year, and I was like, “It’d be really fun to work, on this myself.” I felt like with word vectors, and then prompt engineering and, ImageNet and larger language models for protein generation, not folding and so on, I, me and my teams have pushed the field truly forward. And I feel like we can do it again, here at Recursive. And in many ways, what I observed over the last, 20 years in AI is that whenever we replace some human part of the process of creating AI with a learned system, improvements follow. And so. We’ve done that taking out manual feature engineering, like in sentiment analysis. I don’t know if you remember these old days where, like there are linguists, and they’re like, “Here’s how you negate, and there’s a, like, regular expression.”

Swyx [00:18:21]: I went to Penn where we — they had, like the WordNet

Richard Socher [00:18:24]: That’s right, WordNet, all of that stuff. Yeah

Swyx [00:18:26]: Original. They use, our grad students to label Wall Street Journal articles and, like, really construct a knowledge graph of

Richard Socher [00:18:32]: There you go.

Richard Socher [00:18:33]: And WordNet started, was part of how we started ImageNet. But anyway, so, like, it was really, like, fun, to do. But when we replaced all of that manual feature engineering with vectors and neural nets and just backprop through everything, it started to work really well at scale. And so then everyone started to do architecture engineering, and I was like, “ that clearly can’t be it.”

Swyx [00:18:53]: You mean, neural architecture search?

Richard Socher [00:18:55]: Like, manually, they would say like, “Oh, I’m, I’m doing sentiment analysis, so I have a special neural net that’s really good at sentiment analysis.” And then the machine translation community had a special neural net for machine translation.

Swyx [00:19:06]: I see.

Richard Socher [00:19:07]: The summarization people had their own stuff. And I was like, “That clearly can’t be it. We should unify all of that.” So I had 2 papers. One is called Ask Me Anything, and the other one was called DecaNLP. And DecaNLP eventually got cited, like, 5 times by the first GPT paper. And, to me, that was, like a really a big step forward. And then, of course, you had to combine this idea of prompt engineering with transformers and with language models, and you put it all together, you scale it up, which is also a huge amount of work. And then, the field progressed a lot. I feel like the next step and maybe the last step of that history and the arguably, success has a lot of parents, only failure is an orphan, like my version of that AI history, I do feel like in that history, you can think about, “Well, what’s the next way to automate?” And that is the AI research itself, like the human, process of ideating, implementing, and validating ideas.

Automating AI Research and Recursive Self-Improvement

Richard Socher [00:20:01]: And in our case, ideas for AI.

Richard Socher [00:20:03]: And when you have AI then help you with that, it, by almost definition, becomes a self-improving AI ‘cause it now does research on itself. And there are lots of different misnomers. Some people think auto research is already recursive self-improvement. It’s

Swyx [00:20:17]: Yeah, and you explained that in the talk

Richard Socher [00:20:19]: Completely different.

Richard Socher [00:20:19]: But, to me, it’s the most interesting thing that I could be doing, and I’m really excited with the co-founding team. What’s interesting is we have 8 co-founders in total, including myself. And so

The Recursive Founding Team and Darwin Gödel Machine

Swyx [00:20:31]: They are gonna bring it up.

Richard Socher [00:20:31]: Nice. Yeah. And they’re all. I could talk about all of them if you want.

Swyx [00:20:34]: Super stacked.

Richard Socher [00:20:35]: Yeah. Just an incredibly talented group of people. And we all came to the same conclusion, but from very different directions. Like Josh Tobin, is our CTO. He ran, a bunch of different, projects at OpenAI, like, Codex and deep, research, agents and ChatGPT agents and so on. But before that, he also worked in robotics, and he saw the smaller simulations, and how it’s gonna be really hard to scale that in full generality. And so that’s, that was his angle coming to recursive self-improvement. We have Jeff Clune who’s been working in, like, open-endedness for a long time, together with Tim Rocktäschel. Tim Rocktäschel also built Genie 1, 2, and 3, which is, like the most exciting and most sophisticated, I think, still world model, anywhere. And so they both came from this, open-endedness angle. Jeff also, I think, published one of the most exciting papers in recent years about recursive self-improvement called the Darwin Gödel Machine. Super interesting paper. If we could, maybe pull it up really quick

Richard Socher [00:21:35]: It would be, like, super interesting to see ‘cause you see

Swyx [00:21:38]: By the way, I love how many paper citations.

Swyx [00:21:40]: You’re, you’re giving people a lot of homework, which I like.

Richard Socher [00:21:42]: Love it. Yeah. And so, like Caiming Xiong, a rockstar, we worked together at MetaMind and Salesforce Research together. Alexey Dosovitskiy invented the Vision Transformer, one of the most cited, papers in computer vision. Tim Shi is, like also a unicorn founder. Yuandong Tian led RL at Meta. So just like, yeah, really fun to work with them, and the next level of people are just incredibly strong, too. So it’s been a really fun ride so far. So the first figure, you see exactly these kinds of ideas, that, I think, yeah, inspired a lot of us and now more and more people, where you have this archive of different coding agents. They learn how to self-modify, evaluate, and then create these phylogenetic trees, of, yeah, different ideas.

Swyx [00:22:28]: That’s one foundation. So that Darwin Gödel is an influence.

Swyx [00:22:32]: Open-endedness is an influence. Any other trains of thought that feeds into Recursive that I’m missing?

Influences: Open-Endedness and Learned Systems

Richard Socher [00:22:38]: Going to replace manual parts of the process of building AI

Swyx [00:22:42]: I

Richard Socher [00:22:42]: More and more

Richard Socher [00:22:43]: With learned systems. Yeah.

Swyx [00:22:45]: Which, and, like, merging different fields into one general, architecture.

Richard Socher [00:22:51]: That’s right.

Swyx [00:22:51]: Okay. It seems like language models are already pretty generalist, right?

Swyx [00:22:55]: Your next token predicting your reasoning. Was there a time that you thought, “Okay, these are good enough to have recursive self-improving machines”?

Are Current LLMs Enough?

Richard Socher [00:23:05]: It was clear to me that they will happen, within, like a year or two, and then it did exactly happen, like, earlier this year, right? Earlier this year, AI really went from not just being code, but being able to code. And that is a big unlock. It’s definitely making everything a lot easier than it was, before the beginning of this year.

Swyx [00:23:24]: One question that I think a lot of people have is the current LLM paradigm enough? Or, like, let’s call it autoregressive transformer, with reasoning, whatever. Don’t you need something else, some big unlock, whether it’s world models, which Chris Manning is working on, or memory, continual learning, all that stuff? Or is it all of the kinds, and you think the current, let’s call it transformer architecture, is here to stay and that’s it?

Richard Socher [00:23:48]: A lot of thoughts. So number one, I do think it would be great to have less of a monoculture in AI research.

Richard Socher [00:23:55]: Like, if you look at, AI conferences now, I still remember the days in, like, 2010 when I tried to get my first neural net papers and NLP conferences accepted, and they just desk rejected them because, like, neural nets were something, quote, unquote, “We don’t do in NLP conferences,” and just, like, desk rejected. And it was very brutal in the first years of my PhD. Now I feel like it’s almost like the field switched to the other side. Like

Richard Socher [00:24:17]: Someone should try some other weird, crazy ideas now that aren’t.

Swyx [00:24:20]: There’s also a few. I really respect, like, people still working on, like, GNNs and, like tabular stuff and.

Richard Socher [00:24:25]: Yeah. Like, someone should still, like, do novel out there ideas. At the same time, I think whenever people say, “Oh, LLLMs are. Like, this is the end for LLLMs,” they just don’t, like. LLLMs are also not the LLLMs of, like the past, right? Like, they are so much more sophisticated now. There’s so many more clever things that people are doing. It — There’s, like, different stages of training. You have the whole RL training, and you can take actions and, like all of these things where that can go really far. And then the folks that come from the neurosymbolic, direction say, “Oh, this will never work because they can’t do neurosymbolic reasoning.” It’s like, I think they’re underestimating still the ability for these models to code, and code is neurosymbolic reasoning, and these models can code incredibly well. And so I do think there are, of course, more and more ideas that will be needed and we’ll continue to have. We’re seeing, like, more and more interesting high-level ideas coming out of the AI itself, too. And with really deeply integrating the fact that these models are code and can code, that line — I don’t wanna give it all away, but, like, I think that line has a lot more to grow. But it’s still an LLM, right? Even if that LLM codes for you and then runs that code in some integrated fashion. World models, I’m personally less bullish on. I think if you run a robotics company, you’re gonna build your own world model. I think world models are super fun, and Tim Rocktäschel came to a similar conclusion after building the most interesting one with Genie 1, 2, and 3, which is gaming is a huge application for world models. Can see I sometimes got stuck in some games and, like, got a little overly competitive in the wrong direction. And so I understand games are fun, but personally, I’d rather work on science than gaming. And so, yeah, I think LLLMs, a lot more room to grow.

Swyx [00:26:16]: Yeah. I think there’s some interpretation of world models that some people have where it’s like, well, it’s okay, yes, there is that gaming element. There’s this — there’s the embodied robotics element. But the other part also is just, the more abstract sense of LLLMs are just modeling output, but they’re not modeling the chain of thought, inside the human that has created the output. We can annotate it, of course, but, like, it’s, it’s always, like, this Plato’s cave reflection of a thing rather than the thing, right?

Richard Socher [00:26:43]: It’s true.

Richard Socher [00:26:44]: But I would argue that, and maybe we’ll get there in the 10, spaces of intelligence, but I would argue that even our projection, our eyes is a projection of the real world. And, like, we have only a very narrow, band of the electromagnetic frequency spectrum that we can observe with our puny little 2 eyes and so on.

Swyx [00:27:01]: It’s good enough.

Richard Socher [00:27:02]: It’s, it’s good enough for now, but, like the upper bounds of where it could be are so much higher. And, like, to map, the visual world the way humans see it is also not necessarily, like the end-all be-all for visual intelligence. And I would argue that language is still the most interesting manifestation of human intelligence. And while our visual cortex is certainly less sophisticated, than that of, certain animals all the way down to the mantis shrimp who can, have, like, 2 independent eyes, 3 bands, trinocular vision and each eye can see all the way to, like, floating temperatures in 4D and stuff.

Richard Socher [00:27:36]: Like, mantis shrimp, you should look it up. It’s like

Swyx [00:27:37]: Way OP.

Richard Socher [00:27:38]: Super crazy.

Swyx [00:27:39]: Yeah. ZeFrank, mantis shrimp.

Swyx [00:27:41]: It’s the best video in the world on

Richard Socher [00:27:42]: I love ZeFrank, yeah.

Richard Socher [00:27:44]: Big shout-out to him. But, like, I think there’s a lot more room to grow, but none of these, other animals have language that’s as sophisticated as ours, certainly not in writing. And once you can write, you can, start thinking about longer term civilizations. All of that is language. Programming is much closer to language. And I would argue, and this is, like an important thing in the spaces definition of intelligence also, is that all of these spaces are highly correlated, but visual intelligence is neither necessary nor sufficient for overall intelligence. You can be blind and still be an intelligent human being. And an AI can be blind and still be quite intelligent too.

Swyx [00:28:25]: We were gonna bring this

Richard Socher [00:28:25]: Which doesn’t mean that you’re not more intelligent when you have it. Yeah.

Swyx [00:28:28]: We’re gonna bring this up. I might as well — Like, we have a classification of 10 types of intelligence that you had at the end of your talk. So I’m just gonna flash this up now for people to cover this. I don’t know if, maybe we’ll put this towards the end. We’ll come back to this. I just wanna mention that, you do have a philosophy that I like when people do lists because then I can just go through this and then it gets — it’s educational for people. But let’s go back. I don’t wanna get distracted. But, so effectively, I’ll, I’ll, reinterpret what you said as Yann LeCun is wrong. And then we’ll just

Richard Socher [00:28:56]: Don’t quote me as that. I’m, I’m good friends with Yann. I think very highly of him in many directions.

Swyx [00:29:01]: But he’s wrong.

Swyx [00:29:03]: You mentioned GPT-1, and I cannot let any, Alec Radford, mention escape. Did you talk with him when he was training GPT-1? Like, any historical, fun stories there that you might come up?

DecaNLP, GPT History, and Scientific Gatekeeping

Richard Socher [00:29:18]: I did not, like, meet him a bunch of times. I think we met maybe once or twice at some conferences. But, like, he has told, I think Brian, the first author of the DecaNLP paper, that it did inspire him, and he cited it five times in the GPT-2 paper. So, and that’s, like

Swyx [00:29:36]: Yeah, good enough.

Richard Socher [00:29:36]: Very clearly said, like, this was the first instantiation where they showed in the DecaNLP paper, McCann et al, that you can just phrase every single NLP problem as here’s some prompt, text context, here’s a question and task description and here is some output. If you just do that enough, you can have one unified neural network model, which, by the way, also had all kinds of interesting attention mechanisms. There are slightly different formulations to the transformer. I think came out the same year, plus/minus a few months. And then you can unify all of natural language processing into one neural net. That is the core idea.

Swyx [00:30:14]: And this was as opposed to at the time, LSTMs and what have you.

Richard Socher [00:30:17]: LSTMs, but also, like, people being very stuck in thinking about one model per task. In fact

Richard Socher [00:30:25]: It’s, it’s kinda crazy, but the DecaNLP paper was publicly reviewed as, like, open, OpenReview. It was an ICLR submission. And, in it, you will see, how the whole community at the time thought about this. So, like

Swyx [00:30:43]: Some great contributions, but more work needed.

Richard Socher [00:30:46]: So look at, like, search for not even for humans. Just scroll it up here. Like, question answering is not a unified phenomenon. There is no such thing as general question answering, not even for humans. And this is like, really, you replace your brain with a different brain a different neural net when you answer, like, different kinds of questions. It was unfathomable to the experts at the time that you can have one unified neural network that would answer all of these different questions. They are saying, “No, all of these questions require very different systems to answer, and trying to pretend they are the same doesn’t help anyone solve any problems.” That’s what it says right there, right? That’s how hard it was to fathom. And now, of course, people, when I say, “Oh, we’re gonna invent prompts,” people are like, “You can’t even invent prompts.” It’s such an obvious idea to have one neural network that, of course, does everything in NLP.

Richard Socher [00:31:37]: But at the time, it was, like, extremely controversial, and the paper got rejected. And the sad thing is that it got rejected so hard and they were so certain that we stopped going on our list of things to try. And the number 2 or 3 on the list of extensions for this paper was add language modeling as another task. And then we could have, and that would have accelerated the timelines, in 2018, like, even further for humanity. But we got so crushed, and we were like, “Okay, maybe we’ll just work on some of our other ideas for now and, like, come back to this later.” Yeah.

Swyx [00:32:09]: How can we design a review system that rewards non-consensus?

Richard Socher [00:32:14]: Honestly, I started to feel like arXiv is such a gift to humanity. With arXiv, you should just put your paper out there.

Swyx [00:32:24]: Is it pre-preprints?

Richard Socher [00:32:25]: Let — And honestly, I think Twitter X, people like you who pick up interesting papers, that is a better filter than the experts. Let everyone, like, have access. Now, of course, there are some downsides, which is, like, if you’re super unfamous, you have no Twitter following

Richard Socher [00:32:41]: You don’t wanna be on social media or whatever, you write a good paper, maybe someone, somehow no one notices it. But I would argue that if you just tell, like, 10 of your friends in your community about a paper and it is a really significant breakthrough, someone is bound to talk about it again. And, so I think science needs less gatekeeping. And, even though ICLR, with Yann LeCun, who started it, as one of the co-founders of ICLR back in the day, he also wanted less gatekeeping ‘cause he too was rejected for many years together with Yoshua Bengio and Geoff Hinton with all their early deep learning and neural net papers ‘cause it was just not the hot thing. And so ICLR started with that, but then it also started gatekeeping a little bit themselves on various ideas. So I think less gatekeeping, more open, and then allowing people to say, “Look, even if this is just on, or, quote, unquote, ‘just an archive,’ if it has like 1000 citations, it’s a legitimate paper. Doesn’t really matter where you published it.”

Swyx [00:33:34]: And I agree with that. I do think it’s sad that I’ve heard that grad students have to do, like, how to Twitter, seminars to each other

Swyx [00:33:43]: Just because it’s so important for publishing these days. This person is just reflecting the sentiment at the time.

Richard Socher [00:33:49]: That’s right.

Swyx [00:33:49]: But it’s

Richard Socher [00:33:50]: I think it’s

Swyx [00:33:50]: It affected you so much

Swyx [00:33:52]: That you stopped work on it.

Vibhu [00:33:53]: The sentiment also came out of some of the research, right? Like, the original BERT paper was trained, and towards the end of the paper, they’re like, “Okay, throw off the last head, train specific iterations for

Vibhu [00:34:05]: Extractive summarization add a head for this.” Like, you should do task-specific stuff. These are, like the authors that wrote Attention, wrote BERT, telling you this is what you’re meant to do. And, like the training tasks were also very odd. They’re like

Vibhu [00:34:16]: The — “We know that the model overfits to this weird mass language modeling. Throw away this part and just do specific models,”?

Richard Socher [00:34:23]: Exactly. And, like, we had to try — come up with all clever ways of, like attention and pointers and so on to get the neural network to be able to do all of these tasks. And then some of them were better than state-of-the-art, some weren’t, but we were like, “But it’s still in one model.” I thought it was really cool. Really interesting.

Swyx [00:34:38]: I was gonna move on next to Tim and open-endedness. He was head of open-endedness at Google.

Open-Endedness, Rainbow Teaming, and Self-Set Goals

Richard Socher [00:34:42]: That’s right.

Swyx [00:34:43]: I don’t know what that means.

Swyx [00:34:44]: But he did a lot of talks.

Richard Socher [00:34:45]: Genie 3 is one of the ways that

Richard Socher [00:34:47]: Rainbow teaming, yeah.

Swyx [00:34:49]: So I first saw him at — speaking of ICLR, I first saw him at ICLR when he talked about open-endedness. He’s he’s done a few talks. Can we define what is open-endedness for people who have never been exposed to the problem? They are like, “What do you mean? I thought the only goal of AI is to optimize against a benchmark or.”

Richard Socher [00:35:04]: That’s right, yeah. It’s a, it’s a fuzzy term because there’s so many different instantiations of open-ended, thinking. But, one way I often describe it, and certainly, Tim and Geoff Hinton would be even better at describing this, but it’s a suite of methods that is more inspired by evolution than, very specific rewards. So in that sense, it thinks more about environments, about co-adaptation. And so a concrete example is in the cybersecurity and LM safety space where you have one LM that tries to attack another LM to say something unsafe.

Swyx [00:35:40]: Yeah, the rainbow, yeah.

Richard Socher [00:35:40]: And now the environment is the 2 having a conversation and now they co-adapting, right? They’re like one makes a better attack than the first one inoculates itself somehow, like uses that as training data, makes it so it’s harder to say something unsafe based on that. And then as the attack stops working, the attacker now tries a different angle, right?

Richard Socher [00:36:00]: And that’s why it’s not just red teaming, but they’re called rainbow teaming.

Swyx [00:36:02]: So, like, don’t tell me how to do things. Let me just figure it out myself.

Richard Socher [00:36:05]: That’s right. Think about the environments that you wanna use. Think about the rewards at a high level that you wanna, inspire towards, and then let the AI try out many more ideas in this interplay between sometimes humans, but also sometimes other AI agents.

Swyx [00:36:22]: Yeah. I worked open-endedness into a model that I have been working on. It was the keynote for AI Engineer where you start. You, we have the token loop, we have the agent turns, and then we have goal. And I feel like the way that you’re describing open-endedness is still somewhat of a goal. Like, please attack this,

Swyx [00:36:41]: Other agent. But, to me

Richard Socher [00:36:42]: Yeah, you set the rewards. You set the environments.

Swyx [00:36:44]: The loop that makes the other loops is. What if the agent can set its own goals?