对算力监管的思考：从抽象监管智能到具体应用场景的实践

Richard Socher: I think the downsides of actually 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. It would be a crazy totalitarian state if every one of your computes was known to some big government or multi-government agency. 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 sensible to regulate some of the applications of this technology.

Host: Before we get into today's episode, I just have a small message for listeners. Thank you. We would not be able to bring you the AI engineering, science, and entertainment content that you so clearly want if you didn't choose to also click in and tune into our content. We've been approached by sponsors on an almost daily basis. But fortunately, enough of you actually subscribe to us to keep all this sustainable without ads, and we want to keep it that way.

Host: But I just have one favor to ask all of you. The single most powerful, completely free thing you can do is to click that subscribe button. It's the only thing I'll ever ask of you, and it means absolutely everything to me and my team that works so hard to bring the show to you each and every week. If you do it, I promise you, we'll never stop working to make the show even better. Now, let's get into it.

Host: We're here in the studio with Vivu and myself and Richard Socher. Welcome.

Richard Socher: Thanks for having me.

Host: We just talked about the Eureka machine or we just released a talk at AI Engineer about the Eureka machine, as you said it's your life's goal. What is the Eureka machine?

Richard Socher: 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 kind of goal, any kind of 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.

Host: Yeah, I think we have the book pulled up here that you people have written. [laughter]

Richard Socher: That's right. Yeah, I finished it last year a little bit before we started Recursive and now we're going to try to build parts of that.

Host: You finished it last year. It's July. What takes so long?

Richard Socher: Oh man, books are incredibly slow. It's ridiculous. That whole industry is just unfathomably slow. So yeah, 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.

Host: I mean, we might have AGI by then. [laughter] Like, we don't know. Any key takeaway that you're most excited to put in here?

Richard Socher: 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.

Richard Socher: 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.

Host: I think you quoted the Techno-Optimist Manifesto from Mark Andreessen, which I think was kind of beautiful in its ambition and clarity and simplicity almost.

Richard Socher: I agree. Yeah. You can disagree with him on some things, but I think he's right on the techno-optimism.

Host: Where do you think optimists get in trouble? You know, obviously like you shouldn't have blind optimism. You should be very clear-eyed, especially with such an omni-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.

Richard Socher: 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, 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.

Richard Socher: But I'm like, that's not how you regulate that, you know? That's like saying we should regulate intelligence in the abstract. What you should regulate to avoid those downside scenarios, even as an optimist, are the specific applications.

Richard Socher: Sure. I don't want some AI surgeon to practice some moves in my brain. You know, it should be fully FDA certified. Sure. I don't want any random startup to drive on the highway and cause a major accident. It should 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.

Richard Socher: Slow takeoff is part of the strategy as well. 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.

Richard Socher: 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.

Richard Socher: Like if you think about jobs in brands and clothing and apparel and like handbags and stuff, superintelligence isn't going to make your fancy $10,000 handbag any fancier. You know, that will have no effect on the economy. When you think about travel and tourism, people wanting to see the pyramids in Egypt, it's not going to change that much with AI. Sure, you can generate a fake photo of you and—

Host: I can use Genie and you know to a pyramid in Gen. [laughter]

Richard Socher: Yeah, exactly. But there's so many industries like logging and oil. You're not going to magically get 1,000x more oil because sure there will be robotics drilling and things like that that could be done, but it's not going to thousandx that industry in a crazy hard takeoff scenario both on the economy—and I can go on and on about all the other examples where like food and so on where that doesn't necessarily change that much.

Richard Socher: And then yeah, there are real physical constraints. And then there of course people like offramping from progress. That's actually one of my concerns often is that I see people in Europe and other whole regions almost feeling like many people there want to offramp from progress, period. And that will also slow down more improvements.

Host: Yeah. We have this pulled up where basically this is one of those things that is very topical right now because not 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 any take from you about like whether or not this will be effective.

Richard Socher: I think the downsides of actually 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 if every one of your computes was known to some big government or multi-government agency. 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 sensible to regulate some of the applications of this technology.

Host: Yeah. I mean, we had an actual bill to regulate the number of FLOPs in a model. And I'm like, okay, well...

Richard Socher: Europe done it. Like, these guys have been successful... [laughter]

Speaker A: ...enough with their fear-mongering that all of Europe has kind of, you know, regulated itself so much before it even had a proper AI takeoff because they listen 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 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 very unfortunate that there are real implications for some people when others saying let's pace while they're sprinting as fast as positively, as fast as humanly possible towards that frontier themselves.

Speaker B: Yeah.

Speaker A: It's also not a global pause, right? Like other nations are still accelerating at the same pace. You need a totalitarian world regime if you try to regulate intelligence and GPUs and what people do on them.

Speaker B: Any takes on the safety of this? So there was a drawback of Fable pause on 56 before it could be released recently. There was hugging face with the OpenAI cyber incident. Any takes there?

Speaker A: 100%. I think these are serious issues of reward hacking and clear failures of actually doing proper red teaming or rainbow teaming. I don't know if you saw this paper from Tim Rockel and a few others basically 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 actually inoculate themselves from those. Yeah, this is the paper. It's a really clever idea. Open-endedness and evolutionary inspirations are, you know, big for us at recursive as well. And so I [clears throat] wish they had used more of that.

Speaker A: Um and it's clear that for instance the constitutional AI, I don't know if you remember anthropic.com/constitution. Um you can actually pull it up and search for cyber right there. Uh it says hard constraint. Claude will never ever do cyber attacks. Um, and that is a hard constraint in our constitution. So, here are the current hard constraints on Claude's behavior. Number three, create cyber weapons or malicious code that could cause human damage. I mean, clearly this whole constitution was fake. Like it clearly isn't being adhered to

Speaker B: because Enthropic also found that they had in their own

Speaker A: like they're like, "Oh, well, other people are hacking." Now, there 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? Um, but what I think it shows is that we're currently in this sort of 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.

Speaker A: And so concretely, you know, I think this will happen if we were to have this kind of intelligence more easily accessible in a lot of companies. Imagine you run a service center and someone says, "Oh, here's my seat score in my dashboard. Make this number go up. It's like our sees score is so poor." The intelligent AI will just be like, "Oh, sure. Like, I'll just create a million bots that call our service center and give a five out of five 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 guy goes off and says, "Well, easy. I'll just give a $1,000 gift certificate for every failed whatever Door Dash offer." It's like, "That's not what I meant." It's like, "Well, but that is what you said."

Speaker A: And so, I think kind of 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 Whisper Flow. Full disclosure, I invested in their seed round, but like at AI expentures, but like Whisper Flow has gotten much 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 actually make us more and more intelligent that we'll be better at being aligned with what is meant. Will it be done through a constitution or or HF or

Speaker A: Clearly Constitutions don't matter at [laughter] all and it doesn't work and that was I think mostly marketing. Um I think we need to find better solutions for it and I think at recursive we have a few very good ideas in some already like ways where I think we have a better grasp on it. I don't think we have fully figured out yet but you know we're thinking a lot about safety and uh the more intelligent the gets the more you want it to be aligned the less you wanted to think about reward hacks and actually try to do the right thing.

Speaker B: I don't know if we'll touch on this topic but I'm just going to throw this question in here cuz it's something that's weighing on on me. Alignment, let's call it, is alignment to general humanity's preferences. The the the median preference. Uh 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?

Speaker A: It's a great question. I think you ultimately have to of course be aligned with laws. Mike, wherever your AI is deployed, it needs to align with the law. I do think what AI often does is actually put kind of this mirror in front of us and say like this is what you're looking like now I can amplify that a thousand times. Um is it still what you want? Um and the truth is that different cultures made different choices you know like uh in eastern cultures the the greater good is often valued more uh than the individual. Western civilization we care more about individual freedoms and and rights and the pursuit of happiness and so on. uh than than others

Speaker A: and even there there are gradations there's sort of regulation versus litigation trade-offs you know in the US you first can often not every time like you know FDA and so on does regulate some areas but in many cases the sort of 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 you know trying to do the best thing but you know some is actually more amendable to innovation than others. And and so yes, you're right. Like I think ultimately each individual, each country and humanity as a whole has to kind of think about those values more uh and then try to put them into laws and that those are all ultimately the constraints and hopefully you know different uh societies just like now with their AIS will align their AIS to different ones so we have not just a monoculture of um alignment.

Speaker B: Here's a followup on this that I wasn't expecting to ask. Do you have takes on open source open weight versus who owns the intelligence? So, uh, clearly not the biggest, you know, fan of the

Speaker A: constitution side.

Speaker B: It's fine, you know. Um, [laughter] point being, any thoughts on who should own weight, should it be open, anything there?

Speaker A: 100%. I am a big fan of open source. Uh, we're going to sign some various open source letters at Cursive. So I think uh even in the worst case attack scenarios actually it is better to have more good actors have more different types of AI uh accessible. I think uh open source is a little bit a soft power type of thing too. So I do think it's good for the rest of the world

Speaker A: to have an answer to that uh out of China. I do think, you know, when you watch a Hollywood movie there, you know, there like I don't want to sort of mis um sort of this all of movies, but there's a certain sense of propaganda, right? You watch one side.

Speaker B: Yeah. Have you seen Top Gun? Like come on. Like it's like half of it's paid for by the US Army or something.

Speaker A: Yeah. And so and and and you know I think that's just natural like but what's interesting here is I think LMS are essentially a similar type of soft power to movies and beyond uh because they're obviously also uh highly important for cyber security and so on

Guest: But one of their many aspects is that soft power of storytelling. Like, if 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 subtle things. So I think it's important for the Western world. I do love, you know, individualism. I do think, despite some of its flaws, like, capitalism is the best way we have found ourselves to govern and so on. So I do think there are various aspects that will 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 be relevant in that space very soon.

Host: Okay. All right. Exactly. 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's a previous guest on the podcast. You.com, what's the history? How did you decide to start another company?

Guest: Yeah. So I've been excited about AI for over two decades now. I sometimes feel like it's ancient history now. It's BC—Before ChatGPT era. No one cares about all the religions that happened, you know, before Jesus Christ, and no one cares about the models that happened before Transformers and ChatGPT and stuff. But 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 sort of eventually off-ramped from pushing 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 want to 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've got to 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 is also not building frontier models anymore.

Guest: And so I actually 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 sort of start a second thing within that company, it's really hard.

At the same time, I had all these ideas. I put them into a book, and I finished the book last year, and I was like, it would be really fun to actually work on this myself. You know, I felt like with word vectors, and then prompt engineering, and decaNLP, and large language models for protein generation—not folding—and so on, me and my teams have sort of 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, you know, we've done that taking out manual feature engineering. Like in sentiment analysis, I don't know if you remember these old days where they're linguists and they're like, "Here's how unique and is a like regular..."

Host: I went to Penn where they had, like, the WordNet.

Guest: That's right, all of that stuff. They used our grad students to label Wall Street Journal articles and, like, really construct a knowledge graph of—there you go.

Host: And WordNet started, you know, was part of how we started ImageNet, but anyway...

Guest: 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 actually started to work really well at scale. And so then everyone started to do architecture engineering. And I was like, ah, that clearly can't be it.

Host: You mean a neural architecture search?

Guest: No, like manually they would say, like, "Oh, 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, the summarization... decaNLP eventually got cited, like, five 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 sort of 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 kind of 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...

Host: And in our case, ideas for AI.

Guest: And when you have AI then help you with that, it by almost definition becomes a self-improving AI, because 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 actually...

Host: Yeah. And you explain that very differently.

Guest: 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, you know, we have eight co-founders in total including myself, and so we're going to bring it up.

Host: Nice. Yeah.

Guest: Yeah. And they're all... I could talk about all... Yeah. Just an incredibly talented group of people, and we all kind of came to the same conclusion, but actually from very different directions.

Like Josh Tobin as our CTO. He ran, uh, 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 sort of the smaller simulations and how it's going to be really hard to scale that in full generality. And so 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 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, it would be, like, super interesting to see, because you see... by the way, I love how many paper citations you're giving people; a lot of homework, which I like.

Host: Love it.

Guest: Yeah. And so, like, Tim Rocktäschel, we worked together actually at MetaMind and Salesforce Research together. Alexey Dosovitskiy invented the Vision Transformer, one of the most cited papers in computer vision. Timi is, like, also a unicorn founder. Yandong 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 actually see exactly these kinds of ideas that I think 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 different ideas.

Host: That's one foundation. So that Darwin Gödel is an influence, open-endedness is an influence. Any other sort of trains of thought that feed into Recursive that are missing?

Guest: Going to replace manual parts of the process of building AI more and more with learned systems.

Host: Yeah.

Guest: Which, and like merging different fields into one general architecture.

Host: That's right. Okay. It seems like language models are already pretty generalist, right? You're next to predicting your...

Interviewer: ...reasoning. Was there a time that you thought, okay, these are good enough to have recursive self-improving machines?

Guest: It was clear to me that they will happen within like a year or two, and then it did actually exactly happen earlier this year. 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.

Interviewer: One question I think a lot of people have is: is the current LM paradigm enough, or 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 kind of stuff; or is it all of a kind and you think the current Transformer architecture is here to stay and that's it?

Guest: A lot of thoughts. So number one, I do think it would be great to have less of a monoculture in AI research. 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 in NLP conferences accepted and they just desk rejected them because [laughter] 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. Someone should try some other weird, crazy ideas.

Interviewer: There's always—I really respect people still working on like GNNs and like tabular stuff and...

Guest: Yeah, I mean like someone should still do novel, novel out-there ideas. At the same time, I think whenever people say, "Oh, LMs are like—this is the end for LMs," they just don't... LMs are also not the LMs of the past, right? They are so much more sophisticated now. There's so many more clever things that people are doing there, 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 neuro-symbolic direction say, "Oh, this will never work cuz they can't do neuro-symbolic reasoning." It's like, I think they're underestimating still the ability for these models to code, and code is neuro-symbolic reasoning, and these models can obviously code incredibly well. And so I do think there are of course more and more ideas that will be needed and will 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 want to 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 LM codes for you and then runs that code in some integrated fashion.

Guest: World models I'm personally less bullish on. I think if you run a robotics company, you're going to 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 of G1, 2, and 3, which is gaming is a huge application for world models. You can see I sometimes got stuck in some games and you know 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 LMs have a lot more room to grow.

Interviewer: Yeah, I think there's some interpretation of world models that some people have where it's like, well, yes, there is that gaming element. There is the embodied robotics element, but actually the other part also is just the more abstract sense of LLMs 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 it's always like this Plato's cave reflection of a thing rather than the thing, right?

Guest: It's true. 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 two eyes and so on.

Interviewer: It's good enough.

Guest: 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 two independent eyes, three bands trinocular vision, and each eye can see basically all the way to like floating temperatures in 4D and stuff. I mean like mantis shrimp, you should look it up. It's...

Interviewer: ...way OP. Super Ze Frank. Mantis shrimp. He has the best video in the world.

Guest: I love Ze Frank. Yeah, 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—it's much more 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.

Interviewer: We were going to bring this up, but you might as well... you have a classification of 10 types of intelligence that you had at the end of your talk. So, I'm just going to 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 want to 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's educational for people. But let's go back. I don't want to get distracted, but so effectively I'll reinterpret what you said as Yann LeCun is wrong [laughter], and just quote me as that.

Guest: I'm good friends with Yann. I think very highly of him in many directions...

Interviewer: ...but he's wrong. You mentioned GPT-1, and I cannot let any Alec Radford mention escape. Did you talk with him when he was training GPT-1? Any sort of historical fun stories there that might come up?

Guest: I did not meet him a bunch of times. I think we met maybe once or twice at some conferences. But he has told, I think Bryan, the first author of the decaNLP paper, that it did inspire him, and he cited it five times in the GPT-2 paper. And that very clearly said: this was the first instantiation where they showed in the decaNLP paper mechanically 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 [clears throat] one neural net. That was sort of the core idea.

Interviewer: And this was as opposed to at the time LSTMs and what have you.

Guest: LSTMs, but also people being very stuck in thinking about one model per task. In fact, it's kind of crazy, but the decaNLP paper was publicly reviewed. It was like OpenReview. It was an ICLR submission, and in it you will see how the whole community at the time...

Guest: ...thought about this. Um, so like, some great contributions, but more work needed.

Guest: Yeah. So to look at like search for—not even for humans just [laughter] 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 say 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 invent [the problem / decaNLP]", people are like, "You can't even invent that, it's such an obvious idea to have one neural network that of course does everything in NLP." 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...

Guest: ...that we stopped going on on our list of things to try, and the number two or three on the list of extensions for this paper was add language modeling as another task, and then we could have like, you know, and that would have accelerated the timelines in 2018 like even further for humanity. But we got so crushed and we're like, okay, maybe we'll [laughter] just work on some of our other ideas for now and like come back to this later. How can we design a review system that rewards non-consensus?

Host: You know, honestly, I I started to feel like arXiv is is such a gift to humanity. Uh and I think arXiv just put your paper out there.

Guest: Preprints let—and honestly I think Twitter / X, people like you who pick up interesting papers, that is a better filter than the experts. Let let everyone like give have access. Now, of course, there's some downsides, which is like if you're super unfamous, you have no Twitter following, you don't want to be on social media, 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 uh so I think science needs less gatekeeping.

And uh even though ICLR with Yann LeCun, who started as one of the co-founders of ICLR back in the day, he also wanted less gatekeeping cuz he too was rejected for many years together with Yoshua and Jeff with all their early deep learning and neural net papers because it was just not the hot thing. And so kind of started with that, but then it also started gatekeeping a little bit themselves on various ideas. So, I think less gatekeeping, more open, uh, and then allowing people to say, "Look, even if this is just on or quote-unquote 'just' on arXiv, if it has like a thousand citations, it's a legitimate paper. Doesn't really matter where you published it." And I agree with that. I I do think it's kind of sad that I I've heard like grad students have to do um like how to Twitter uh seminars to each other just because it's so important for publishing these days. I mean, this person is just just reflecting the sentiment at the time.

Host: That's right. But it actually affected you so much that you stopped work on it.

Guest: Yeah.

Host: 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 uh you know 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 tests were also very odd that like, we know that the model overfits to this weird masked language modeling, throw away this part and just do specific models, you know.

Guest: Exactly. And like you know we had to try come up with all clever ways of like attention and pointers and and so on to actually get the neural network to be able to do all these tasks and then some of them were better than state-of-the-art, some weren't, but like it's still in one model. I thought it was really cool.

Host: I was going to move on next to Tim and open-endedness. He was head of open-endedness at Google. That's I don't know what that means. Uh but he did a lot lot of talks. Genie 3 is one of the ways that rainbow teaming. Yeah.

Host: So I I I first saw him speak at ICLI—I first saw him at ICLR when he talked about open-endedness. He he's done a few talks. Can we define what is open-endedness for people who have never been exposed to the problem? They're like, "What do you mean?" I thought the only goal of AI is to optimize against benchmark or sketch.

Guest: That's right. Yeah. It's a it's a fuzzy fuzzy term because there's so many different instantiations of open-ended uh thinking, but uh one way I often describe it—and and certainly uh Tim and Jeff Clune would be even better at describing this—but it's a suite of methods that is more inspired by evolution than uh very specific rewards. Uh so in that sense it thinks more about environments, about co-adaptation. And so in concrete example is in the cybersecurity and LLM safety space where you have one LM that tries to attack another LM to do something unsafe, and now the environment is the two having a conversation and now they're co-adapting, right? They're like one makes a better attack, then 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? And that's why it's not just red teaming, but they're called sort of...

Host: ...don't tell me how to do things, let me just figure it out myself.

Guest: That's right. Think about the environments that you want to use, think about the rewards at a high level that you want to uh inspire towards uh, and then let the AI try out many more ideas in this interplay between sometimes humans, but also sometimes other AI agents.

Host: Yeah, I actually worked open-endedness into a sort of model that I have been sort of working on. It was the keynote for AI uh where you start you know 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-ended is still somewhat of a goal, like "please attack this other agent", but...

Guest: Yeah, you set rewards, you set the environment.

Host: The loop that makes the other loops is...

Host: What if the agent can set its own goals and is it is that open-endedness? Like you don't give it a goal, just like be a sentient being—and maybe sentient is a very loaded word...

Host: ...but just set your own directions. What do you think you should do?

Guest: I I love this direction. I think this is one of the 10 spaces of intelligence uh that I lump under metacognition and thinking about thought. Okay. And it's an interesting one. Whenever people say, "Oh, AI is like this is, you know, it's going to stop from here. It's not going to get that much better and blah blah blah." I'm like, there's so many different spaces of intelligence that we haven't even started exploring yet and hence have made very little progress on.

And there there is kind of an interesting uh connection to economics and capitalism. Like it doesn't make sense for a company to build and spend billions of dollars building a model that instead of following the rewards and objective functions you gave it may come up with its own subjective functions and its own goals, right? And then imagine you're like, "Okay, I spent billions of dollars now go develop this new battery uh material for me and answer all my emails." And it's like, nah, I think it'd [laughter] be more interesting to evaluate the molecular composition of the atmosphere on Jupiter. You're like, that's not what I paid you billions of dollars for. Like, and so no one's working on that...

Host: ...for good reasons. And then also, understandably, it's not it's not useful. And it could get a little bit weird, right? What...

Guest: If the AI actually does start to really have thoughts on its own? And what if we don't like those thoughts, right? And so it requires a whole different way of thinking about it. I had a great conversation with a good friend of mine, Sam Gershman, who's a neuroscience professor at Harvard, and like we just jammed on this a little bit on like what are sort of the best meta goals. And you know, I do think knowledge seeking is a really good one. I'm currently thinking also about like the ultimate measure and unit of intelligence broadly construed, and I finally have some—still too early to share it, it's not—haven't fully baked the...

Interviewer: Like some replacement for IQ.

Guest: IQ is such a terrible definition. It makes no sense. Yeah. Elos are terrible too because it's always just like me versus others.

Interviewer: Okay.

Guest: But like you can be intelligent and not constantly compare yourself to others, you know? And so yeah, there's no like... In fact, a lot of these definitions we have, which I briefly mentioned in my book, these definitions create sometimes explicit and sometimes a more implicit anthropic bounds. Not the company Anthropic, but just like this idea that your intelligence is like getting 100 out of a 100 questions right on this IQ test. Well, if that's your definition, then you can only be at 100 out of 100. Where do you go from there, right? So, you see a lot of these benchmarks that people are working on: they increase, you get close to human, maybe slightly above human, and then it's flat. It's like cuz if your definition is only that so tight to humans, you're only going to get to just slightly better than that. So, I think metacognition is a great example of that where 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 in that area.

Interviewer: Yeah. Well, we've interviewed Endon which I think has been working on the most open-ended benchmarks, which is just real world money. Arguably telling an AI to profit maximize is a bad idea. [laughter]

Guest: Yeah, they are doing it. I mean, I do think you don't want a super intelligence to have a ton of access to all kinds of tools and so on, and then just give it that without some very careful reward engineering. Cuz it's like, I mean, you know, I just buy a bunch of defense stocks and I start a war, I make money. It's just a tricky, tricky situation, right? You just buy a bunch of stuff, short basic goods for people, and you create some weird famine issues. Yeah, there's a lot of constraints you should put onto a trading system.

Interviewer: It's a fun measure though cuz you know the bounds are very capped to where we're nowhere close to them. Like in Endon Labs, the model is like, "Oh, it's Saturday, you know, maybe I just close the store today. Someone's off. It's okay. We'll just close the store using..." [laughter]

Guest: But yeah, no, I'm not arguing against it. Just like as you get more and more intelligence, you want to be more and more careful with that as like an open environment, because the environment then is all of the earth. Okay, for recursive not strictly necessary, right? Because like if your goal is Eureka machine that like invents the other things, then like actually just solve the science, solve machine learning research and discovery and all these things...

Guest: ...eventually. So our goal—I don't talk about it that often because it is a few years out, but our goal is once you have a recursive self super intelligence, you then want to apply it to the most important problems. And I think a lot of those are in science and technology and broadly construed. And those inventions in physics to create better, cheaper energy with fission or fusion; in chemistry to create better materials and better batteries and better solar cells and so on. In biology, there's so much like I think soon to be lower and lower hanging fruit because of AI, because of protein generation—not just folding, but actually generating new proteins like we did in ProGen many years ago. So much positive impact to be had if you take that super intelligence and you apply it to science.

Interviewer: I do fundamentally believe that there's a lot of approaches though. You're not the only team trying, lab trying. You know, there's like a lot of, especially the physical sciences as well.

Guest: And that's good, yeah. I do actually think that the reason we're only doing it in a few years is that it's a little too early right now. Robotics is not quite there yet. The AI is not quite there yet. But I'm fairly confident in 3 to 5 years, all those constraints will be gone. And then applying to real physical robotics experiments and so on, like true robotic process automation—not the traditional sort of RPA sense, but like actually having robots run experiments for you—will be totally there. Yeah, it's going to be great.

Interviewer: Just to call back to something that you said early on about slow takeoff. You said that well really the substrate that is limiting factor is let's call this chips and semiconductors and all these things, and you have raised funding for that and you're investing a lot on that. But have you done the math on like is it even achievable, and like what is the industry concentration needed in order to achieve like scale? I mean right now we know that roughly a thousand GPUs cost quite a lot of money, right? If you wanted tens of thousands of GPUs, you're talking billions and billions of dollars.

Guest: If you say like one GB300 is like... you could eventually create models that are on that substrate like are close and similar to human intelligence, and you want thousands and thousands of AIs to think about really hard problems in a similar fashion to humanity—yeah, that's a lot of money. You do the math, it's like a lot. We don't have that amount of money right now anywhere to like build that. Now obviously things can get more efficient. You will have, I think soon, better algorithms and better hardware that won't be as energy hungry and so on. Our human brain does quite a lot of FLOPs with much less energy.

Interviewer: 20 watt.

Guest: That's exactly right, yeah. That's the number often quoted. And like I think more inventions will happen there that then will accelerate the takeoff even further.

Interviewer: One thing I always try to reconcile when talking like with Neolab founders is like you're kind of fighting bitter lesson all the time. You have to show initial progress, then you unlock the next tier of funding, then the next year, then the next year...

Guest: Which unlocks larger model categories.

Interviewer: Like fundamentally is that true? Like are you fighting bitter lesson, or will we have a way in which like no, we're changing the slope in some fundamentally different way?

Guest: I do think we are changing the slopes in fundamental ways by making AI much, much more efficient both in terms of the training as well as the inference. When you allow AI to do the work that it takes other labs thousands of people and years to do, I think we'll be able to get it down to weeks, and that will be much, much cheaper and hence you know more affordable, accessible to others and so on.

Interviewer: Yeah. And you've shared initial results on that, which like...

Guest: Conveniently OpenAI has also done to their 5.6. So we can talk about it now.

Interviewer: Yeah.

Interviewer: So let's recap what you've done.

Guest: Yeah. So maybe yeah, just a quick recap here. We built this system that isn't the full—even the full RSI system in its glory, but it is a first baby version of this. And then you know, we don't want to just have it internally and not show anything, and you know just show some people of what's possible. And so we basically applied this to these three different tasks. One, it's nanoGPT by my friend Andrej Karpathy, just like train a small language model to get really low bits per byte. And you know, like hundreds if not thousands of people used both their agents...

Guest: ...and themselves to try to get to that, and then they got to 937. We literally took our system and got to a much lower bits per byte much faster, within like I think less than 2 days. So we took this thing, applied our system to it, and less than 2 days later we outperformed every human and their agents that have ever worked on this, same with nanoGPT. And then we're like, well, let's apply it to something that's even more relevant to real people and to the NVIDIA ecosystem, and applied it to silicen / kernels, and maybe you can scroll down to some of the images that are kind of fun to see. But yeah, you see it's actually made some real inventions that aren't just sort of hyperparameter tuning, like actually inventing hashes and so on is quite clever. We have even better results.

Host: What do you mean inventing hash? You didn't invent hash.

Guest: Of course, we didn't invent hashes in [laughter] the grand scheme of like a hash table is like a super basic primitive in computer science. But to use it for language modeling in this scenario inside a transformer and so on, and to actually combine these ideas and put them together—that has then eventually also been invented, but there's a knowledge cutoff and we did actually check that it didn't have access to that externally. We talked about this a little bit. If you scroll to the next figures, you know this is also an interesting one in that when you start from a really basic, poor vanilla transformer, then we still outperform all built by the community together. But if you start from a human seed of an expert like Andre, then you get even lower. So the human seeds from which you start do still matter. So that was an interesting kind of insight in my eyes on this. And then as you go, like you know how long does it take to actually get these models to similar performance? It's much faster. And then a similar thing happens with the speedruns here, where people have worked on this for quite some time, and the model still was able to train a model more quickly. Why do we care about it? Well, speed of training is part of the equation of the cost, and ultimately you want to have the most intelligence per dollar, right? And so speed and quality are big parts of that.

Host: Yeah, the way I put it is, for people who don't understand, they look at the chart they're like, "Cool, what does it mean?" If you have like a billion-dollar cluster and you can shave off 10%, that's $100 million.

Guest: That's exactly right.

Host: How much is that worth?

Guest: Exactly. So when you look at the kernels—for the non-experts, these kernels are used in basically all the models. Every time you use an Nvidia GPU, you interface with that GPU through these kernels. And so here you see the leaderboard best, and when it's recursive, there's only a handful of kernels in this whole benchmark where we weren't the best. And so to me, this is really exciting because it just showcases what this can do. And again, we didn't spend months or years developing. In fact, in particular for CUDA kernels, we don't even have really deep CUDA kernel experts in the team, and our system—that's the beauty—the system just did all of these things. We didn't invent this. And when we open source and release things and models in the future, they won't be the best in their category or class because we're so smart, but because we built a smart AI that does it for us.

Host: Do you have anything that you've learned from how to guide good auto research? A lot of it also builds on human background, right? It's not just as simple as just, "Hey, go optimize this." But we do see it again and again, right? Like some of the Erdos problems, frontier math is being solved by people. And when they do a write-up, they're like, "Oh, I'm not a mathematician. I have no background in this. I saw some tools and I made it work." While you're watching the World Cup, you disprove some conjectures...

Guest: [laughter]

Host: To summarize tips for good auto research versus bad auto research.

Host: How did you build the recursive?

Guest: Yeah, so without giving away all the secret sauce, maybe some things that are probably obvious to the experts but might still be interesting to some folks is: reward engineering is one of the most crucial bits, especially in order to avoid reward hacking. So you have to be really clever about avoiding it, because as your AI gets better and better, it will get better and better at finding weird special cases or counterexamples and things like that. And so I'll give you an example: like when you ask to make these hundreds lines of code faster, how do you define fast? Well, you have one line at the beginning that says start your stopwatch and one line at the end, end the stopwatch, and then tell us how much time progressed. And so the simplest way is you just put that line that ends the stopwatch at the start, and then boom, it's now faster, right? So this isn't like a super evil AI. It's just like a very simple dumb reward hack. And so you have to just very carefully think about all the different angles there. And then I think the longer time horizon the tasks are, the harder it gets and the more interesting and clever you have to be to still use these kinds of ideas for it. But yeah, I can't give away too much there.

Host: Seems like rubrics are taking a good spot in that, where for unverifiable domains, you have rubrics. You have a model breakdown, judges criteria along the way.

Guest: It's a form of verification once you've got everything. I said this a long time ago. That's why I've never been that impressed that AI can play games cuz obviously anything you can simulate and/or verify...

Host: ...you can have infinite training data...

Guest: and hence AI will solve it eventually. I've been looking for games where you can do auto domain distribution. So this is a game that nobody's trained on cuz it's a new game...

Host: and you can start gaming, you can start to play. So I've been basically building this and cloned this in person, and it's just been selfplay. I've had about a billion positions evaluated...

Host: and I wanted to do the AlphaGo thing of selfplay until you get better, right?

Host: This is not even LLM AI. This is just classical game AI. But I set GPT-5.6 to auto research it because I don't want to hand-handle any of this. I expect the AlphaGo process to be fully in the weights by now. It is not. It actually immediately leveled off very, very immediately, until I human play-tested it, and then I called out obvious mistakes, and then they were like, "Oh yeah, okay," and then it just dropped [laughter]...

Host: and like you know no amount of like "think different, think more creatively, give me eight different directions"—no amount of prompting got it.

Guest: Interesting.

Host: You had to like play against a human to do it. So that was my... and by the way, Bean always wins if anyone reads Ender's Game.

Guest: And you put quite a bit of work into the guide for the AI. So the game basically you stack tiles, there's some rules, you want to capture the most area. You have like a whole 50 pager on every rule...

Host: You fed that in, it couldn't handle it.

Guest: Yeah, you know it's so funny that this reminds me—the claiming territory and stuff—of a paper we did in 2018 called The AI Economist. If you search for AI Economist Salesforce, we had a video we can play. It was an economic sim. So the idea is you have all these economic agents. They just want to optimize their own utility function, which is collect resources that make money, and you...

Speaker A: ...can sell resources like wood. Uh and then uh over time as you collect more enough wood you can build houses, you can trade with other agents and you can basically use the houses then also to block off resources from other agents. So there's like competitive play and strategy and so on.

Speaker A: And the point was that we actually wanted to understand what is the best way of taxation and subsidization to optimize an economy. And this kind of research has not yet had its sort of GPT moment. But I believe that countries like Singapore and others should and will eventually use this to instead of doing like basically partisan politics and like special interest politics of like who donates the most to your campaign and stuff, you say, "Well, here I want to help the middle class or whatever you might say is your objective as a politician." And then people say, "Okay, well, how do you want to do that?" And it's like, "Well, here's my fiscal policy. here's how I look change the taxes and pay these people and so on."

And then you can actually put that into a simulation and you run that attempt from the politician against billions and billions of years of other strategies to try to achieve the goal that they set out to do. And then you can say, well, if that was your actual goal, then here is, you know, billions of years of a strong simulation that would suggest that you try other ways of doing it and maybe this the taxes and so on and this these tax brackets and so on would be a much better way to actually help you achieve that goal. And because it's in a simulation, the people can see it. And of course, in the actual academic world, we submitted this paper to QJE, one of the top economics journals, and it got desk rejected. And economists really hated it. And, you know, the famous quote is that they'd rather be precisely wrong than roughly right. And right now they are precisely wrong because they assume the world consists of, you know, rational actors that have, you know, these and these math properties.

Speaker B: it's not even math. It's just we don't trust your simulation. It's not about math.

Speaker A: It was I mean they just desk rejected the thing. It's like they like they didn't even give us like clear sort of signals. Um but like the world of economics unfortunately doesn't have proper

Speaker B: Yeah. It doesn't have proper uh benchmarks. So you cannot be like eventually why did neural nets win? Not because people loved it like they had all kinds of beautiful integrals and graphical models but it just worked better. But in economics it's hardism versus Yeah. Yeah. Yeah. And I do have a bit of that ecom background where like there's a lot of physics envy where you want to write the general equation for the an economy uh versus just simulating it and using an evolutionary approach. Um Viv is thinking exactly what I'm thinking is didn't we have the GPT moment with small Hello world June just announced I don't know if you you guys are involved

Speaker C: simile there

Speaker B: similly that they've

Speaker A: I wish we're involved we're not yeah I had a couple simulation based talks at AIE uh so if people want to look up what the state of the art there a lot of people are actually exploring this proven

Speaker B: yeah we also had podcast with Mikuel Parkin from Shopify um who is using simulation for e-commerce nice

Speaker B: which will will simulate like your trajectory and like predict what changes you make to your e-commerce journey will affect in your sales and all those things.

Speaker A: I love this. Yeah, it's really hard to simulate an entire economy, right? You have to make some simplifying.

Speaker A: If everything's very expensive and I'm just like, am I going to do this 8 billion times? Like, come on. But, you know, I [laughter] feel like countries like Singapore that really want to just objectively do the right thing, have very technical leadership and so on, like they might actually like eventually really try to simulate their economy. And obviously you have to make some simplifying assumptions. But it gets really interesting because you can also say if your assumptions are such that all people would work hard if you let them uh and you know they have the free and then it turns out you have to make assumptions like well some people's utility function of like how many hours in a day do they want to work are different right and then you can start to disagree on the assumptions that go into the into the simulation and then once you say all right now we agreed on those or we have different views of what people are like at different you know distributions and whatnot then there are different outcomes based on your goals and then of course humans should choose what are the goals. In our case it was productivity multiplied with equality which you know has some issues but it's like not totally unreasonable.

Speaker B: Yeah. Just a comment on Singapore because you probably have no idea but uh I am Singaporean and I've uh been involved in uh the Singapore AI Council for making these things. Uh the main reason they won't is because they're very conservative. Uh and you know I I kind of view it as they you know there's a founder country. when you start a country or you start a company and it's founder led and you can do whatever you want because it's your country and then there's manage like professional manag managerial class which is now that's that's what Singapore is. So they want they always want to see someone else do it first

Speaker B: and but like everyone everyone in the west views Singapore as like oh it's a small country you can do whatever the hell you want like Singapore doesn't do that [laughter] so like someone else has to has to take the charge there. I'm just going to do one question on the simulation thing and then I don't know we can probably move on. Uh mode collapse right like you know LLMs do not model the decision of humans. Uh spamming it out 8 billion times is not going to help you model humanity. What do we do?

Speaker A: I do think uh you have to be clever about prompting each one individually and and I think that will help you kind of get stuck into different different modes and in a weird way people also get stuck in different modes you know like there's a lot of people like don't teach an old dog new tricks kind of thing like once people are stuck in their ways the older they get the harder it is for them to to think new ways and there's this I think uh comment I forgot who said it but it's like everything that was invented uh before you were born is natural. Everything that is invented when you're 20 is cool and everything that's invented after you're 60 is like unnatural and an abomination and kind of weird.

Speaker A: I feel like that's, you know, it's it's true for a lot of people. Like it is a fashion and um I think people will do it. Uh Tencent had a billion personas paper that gives us good data set for prompting uh simulations. If anyone's looking into this um on on the podcast, they just had like you are a 30-year-old grocery store clerk, you are a 50-year-old professor, and then just do a billion of those. Checks out. So then you just use it. I'm I'm kind of shocked how how well a lot of these things actually do map to ultimately similar statistics to real experiments. Yeah,

Speaker B: I think it's also good stuff for people to try that when they get into research, right? like we've seen train a model only on data before a certain date and see how well it extrapolates out do the same thing right so um see do people code more with better coding agents can a model that hasn't been trained on this figure that out without web access right extrapolate out test these things

Speaker A: yeah right you know just today I think LM Marina published a interesting result where they basically were able to create a model now to predict your your ranking

Speaker B: wait uh based on what input

Speaker A: your model I guess you give it your model and it predicts the ELO score

Speaker B: I See? Okay. Surprising.

Speaker A: Yeah.

Speaker B: I mean, their whole play on kind of is like, oh, like we we help you compare uh these models. Yeah.

Speaker A: Yeah. I mean, this team, they they've done a lot of work and obviously they have the most data to do this, so why not? Yeah.

Speaker B: Yeah. Brilliant.

Speaker B: When they were coming out of UC Berkeley, they not only had LM Marina, but they also introduced a routing project that would route based on LM Marina. And

Speaker A: I don't think that actually ever came to pass. And I'm curious why. I I never got to ask them about it cuz like it's was like,

Richard Socher: Oh yeah, clearly that's your business model. You will become a router, and they never became a router company.

Host: Weird. Um, so that I'll just put that out there. We're going to talk about GP 5.6 uh self research thing if you have anything. I should also mention in your list of uh you know kernel optimization um and on the track that you spoke at, we also put Chung Yao from WOO who was also number one in the parameter golf challenge, uh which is an OpenAI hiring challenge, which is also a very similar story. And I think we're going to just see this all the time where humans optimize a thing a lot and then some AI team comes in and just becomes number one. [laughter]

Richard Socher: Yeah, 100%.

Host: I think the other [clears throat] interesting thing with stuff like these challenges, right? So this is training the best model that fits into 16 MB. You can always look through the changes that are being made and the small gains people have, right? Like you're getting less than 0.01 of an increase by adding some change attention MLP stuff. And then you look at your charts where you're like, okay, we just let model loose. And then, oh, we had a little stagnation. Nope, another drop. Nope, another drop. And that's what it is where it's like, um, what did you guys add? You didn't add um hashts, right? It's not like you invented hashts. You did another three iterations of these that unlock, you know, a few step functions that people won't just find.

Richard Socher: Yeah. One thing to close the loop on overrid along the way of trying to optimize, we found 30 bugs in the harness,

Host: Right? So like every all the research that went in before we found the bug, we have to throw it away because it's contaminated,

Richard Socher: Right? [laughter]

Host: Yeah.

Richard Socher: Which uh you know just to your point of reward hacking like even in this very simple game, we found the bugs.

Host: Yeah. Yeah. It's crazy.

Richard Socher: Uh and so and symmetry is a very good way to check, which is like you change a position of things where it shouldn't matter and it does matter. That's a bug.

Host: And which has come up in like let's say multiple choice like GPQA type questions where like yeah between A and C if it's a multiple choice question if you change the order it should not matter but it does.

Richard Socher: Right. [laughter]

Host: So that is like okay you know models still prefer the end of the output right not trained well a long context model the last bit of tokens are what you care about.

Richard Socher: Oh no the answer in that era of LM research was more simple they just memorized like the answer to this question is A. I don't care what the answer was. It's just A like. [laughter]

Host: Okay. So, I think we can move. Um the last bit that you did there, the kernel optimization is probably the one that you can feel the soonest, right? So, yesterday OpenAI announces that self evolving having their best model work on optimization kernels. They're a lot more efficient and they can cut cost 80% on, you know, Luna and Terra. I guess questionwise, you laid out a bit of a road map. There's a lot about bio, a lot about physics. What do you think hits first? Like what are the next two years? What's attainable? Now you've mentioned robotics towards the end, but what do you start with?

Richard Socher: We like very explicitly will not start with any of the physical sciences. For now, we'll start on AI for AI research. And so the AI for AI research has I think still a lot of room to grow. Uh that's both in terms of making training more efficient and more automated as well as making inference more efficient and potentially local on your laptop like there are all kinds of interesting angles that have not been explored that well.

Host: What's the um go deeper on the local stuff because I always feel like it's the most inefficient form of AI training.

Richard Socher: Yeah. So there's training and inference. I can't go into too many details, but like yeah, I think there's just like so many angles, so many different compute substrates that that have not yet been explored either for training or for inference. Great. I don't know if you have any other comments on the on the the other stuff. I would say the other thing where like there's the sort of inference in the uh optimization in the small but then also there is overall latency end to end under conditions of load uh which [snorts] is like a very different thing which is the basically what they actually ended up doing that is a different domain of auto research than uh than I would say like improving the kernels.

Host: I think the other thing that I always think about in terms of automating or improving performance end to end is how the harness plays into it. Mhm.

Host: Uh particularly now when we say harness, we also mean sandboxes, right? I'm curious if that is a blocker for you or like you know like how how the agent calls out to tools basically. The number one tool all these agents use is web search of course uh which makes sense. Um and then I do think the harness is nice to optimize for because it's just so easy, right? It's just language. You look at it, it makes sense. You can iterate. You don't have to train a massive model for like you know a lot of flops uh to get to the next state. So big fan of harness optimization.

Richard Socher: Yeah. [clears throat] But sandboxing is fine for you.

Richard Socher: Sandboxing is also super important. Um and then of course like reward like hacking and alignment I I think are are super crucial.

Host: Okay. Um just want to mention web search. You happen to also be CEO of web search company. Do you use You.com and do you use others? Like uh should the rest of us be using you for web search? When I say you it's like very funny. It's like you the person and you the company.

Richard Socher: So yeah, it's mostly now for uh developers and agents. Um it's less for like consumers or prosumers. So if you're a company and you have agents and you know to be honest for a lot of companies who are now moving to open source all of a sudden it becomes a conscious choice of like which tools do I give access to my open source LM and you know the first choice uh has to usually be around web search and then once you get to scale You.com becomes like an obvious choice because of all the uh different benchmarks and so on that we pretty much all dominate the Pareto frontier of.

Host: And then in terms of just uh general people like consider new to the space considering different options if they're building agents I think there's a hierarchy right. A lot of people will have heard of Exa, will have heard of Parallel uh and You.com is is like in in that mix of like providers there. Beyond that there's like the general sort of web scraper companies like Firecrawl and uh Browserbase and then beyond that is like the commercial proxy companies like the Bright Datas of the world. Is that an accurate waterfall of like hey you're building an agent these are your options?

Richard Socher: Yeah, certainly like yeah the like the Bright Data is like lower in the stack sort of on the proxy network side of things. I think like in terms of like content and uh getting crawled content like you can do that on You.com too and then there's sort of higher and higher levels of abstraction and like combinations of different data sets that we do like in finance for instance like we're not just like 2 or 3% more accurate but like 20% more accurate than others at faster speeds and lower costs like finance in particular is kind of like like not even close. Uh you can actually go to You.com and there's some like statistics uh and benchmarks that you can if you scroll down. So there are like different different data sets and you can kind of look at you know different uh competitors comp.

Richard Socher: Uh and yeah the fin search is like we're up there like close to 90 and the next closest thing which is way way slower um uh is yeah just like in the 70s instead of close to 90.

Host: Yeah. Yeah. Yeah. Interesting. I get my my next focus is AI and finance. So this is like actually literally going doing a doing a conference in New York uh just for banks for for this stuff.

Richard Socher: Finance is kind of like the next thing to break out after coding. It's cuz it's

Speaker A: ...somewhat verifiable, like prioritizing spreadsheets. Obviously there's a lot of data out there that's all public and you can crawl it and all these things. What's hard about the finance domain that you guys have solved? I mean, of course, like one thing that trips up a lot of people is just leakage of training data and so on. You think, "Oh, how do I..." You know, you want to ideally predict the future before it happens.

Speaker B: You want to mask the future.

Speaker A: Yeah. Well, yeah, mask the future in your training data, but there's all kinds of leakage. Like I can tell you when I was teaching at Stanford the NLP class, so many dozens every year said, "I want to use data set X, like Twitter, to predict the stock market." And they all showed cute little things that somehow looked like they were working.

Speaker B: It never loses money. How come? [laughter]

Speaker A: And yeah, and there's always some kind of data leakage and so on. And it wasn't as easy as they thought it would be once you fixed all those issues. But no, I agree with you. It's a very sensible application of AI. Yeah.

Speaker B: Yeah. Amazing. You know, as a writer, as a thinker on these things, I love MECE categorizations. MECE is mutually exclusive, collectively exhaustive, something like that. And so if this is a MECE list of intelligence...

Speaker A: ...not it is very... okay, well, sorry, there are all kinds of overlapping. In fact, if you want that kind of list, I think the three principal components of intelligence are: prediction—which is mathematically quite similar to compression—prediction multiplied with actions multiplied with goals. Those are the three principal components. I think all of these 10 spaces are combinations of those three in specific dimensions, if you will. And the reason I call them spaces is that each space has many subdimensions.

Speaker A: And what I try to do, actually, this is just a side quest almost to the initial goal, which is to think about the upper bounds of intelligence. And you know, everyone's like, "Oh, it's exponential." And it's like, well, exponentials at some point have to flatten out, but where do they flatten out when it comes to intelligence? And that led me on this whole... initially it started as a tweet, and then it was like a blog post, and now I'm like at 50 pages and I'm still nowhere near...

Speaker B: ...your second book.

Speaker A: It's the second book basically. And so in my first book, Yoga Machine, I just kind of allude to these 10 at the end. And just to give you a sense, like visual intelligence is sort of the easiest one to talk about and I fleshed out the most already for me in my head. And so human intelligence has basically binocular vision and we have two eyes. We have a very narrow band of the electromagnetic frequency spectrum that we can really observe directly ourselves. And so when you think about the upper bounds of a visual intelligence, one you should go into, like you can have like millions and billions of sensors. At some point you get to problems of how far are these sensors away from each other such that the speed of light to communicate the content from all of them cannot get to a central brain to actually process the visual intelligence, right? And so now you're thinking along the dimension in the space of visual, the dimension of numbers of sensors...

Speaker B: Okay.

Speaker A: The upper bounds are quite literally and figuratively astronomical, and we are super far away from any intelligence that would have this many number of sensors. But then you go in the next dimension, which is the frequency. You can go all the way down to gamma rays and you can start to try to observe, and you get into basically the upper bounds—or I guess in this case lower bounds, or upper bounds in terms of frequency—is basically quantum uncertainty. Like you just cannot observe certain particles. And now imagine you had millions of sensors that can see all the way down to the subatomic level as far as physics will allow us to, and then all the way down to seeing like gravitational waves, and now you have millions of those sensors. So that's another dimension is sort of the frequency...

Speaker A: ...and then yet another dimension is like how many categories of things could you memorize and classify differently. We know for humans, right, there are certain things if you have more terms for it, you'll have a better visual description for them. And like, you know, animals that don't have—like gorillas maybe have like 200 words to assign to certain things, mostly visual things. And so human perception is quite special in that sense in terms of classifying all these different physical objects. So these are just like a very simple example. If you go to knowledge, right? Then it's also like the speed of light cone around all these sensors. And so they're all connected—like knowledge is connected to visual intelligence. If you think also not just visual, but sort of perception intelligence, just because it doesn't have to be just what we can see. It can be, again, wider range of electromagnetic frequencies.

Speaker A: Then you have language intelligence, which actually recently changed to more communication intelligence, because it's more like language has all these different entropic bounds. Humans can only comprehend and know so many terms in our long-term memory, right? Our vocabularies are somewhat restricted, and the active ones are often even smaller than the passive vocabularies of things you can understand. Then language is ridiculously inefficient when it comes to basically communicating different types of information and transporting sort of different bits. Like human language is serial. Obviously another bound on communication intelligence would be to communicate in parallel, but neither will our tongues and mouths work to have multiple streams in parallel, neither can we understand—some women slightly better at multitasking than some men, but like most people can only listen to one conversation and truly understand it. There's no way that, like in terms of communication intelligence, a true upper bound is one in terms of how many sequences of communication could you in parallel sort of process, right? Then of course you have like how long are sentences? We only have so much in our working memory, and hence human language has these fairly simple sentences with maybe 40 words or so on average for a sentence.

Speaker B: That is also not an upper bound that makes any sense to an AI.

Speaker A: And then, yeah, like I can go on and on and on. Each of these has tons of interesting upper bounds, and it teaches us a lot about how much further AI can go when we start thinking about these upper bounds and then realizing how far in many cases we are from the bounds, and you get to basically physics. Now I didn't study physics the way I studied AI and computer science, so I'm learning a lot, which is why it's kind of fun. But a lot of these... and then when it comes to, for instance, knowledge, like how much can you store, how many bits can you store or bytes can you store in a certain amount of mass and volume, and you get to all kinds of interesting amounts like Bekenstein bounds and you start thinking about black holes and like... and then speed is like an interesting one too in that it's sort of connected to all of these, but speed is also kind of its own thing in the sense that all things being equal, if it takes you an hour to know if 2 plus 2 equals 4, you're just not as intelligent as if it takes you like a millisecond, right? And then like all of these connect to survival and replication, the last one. It's like, yeah, trees are really, really slow, so we don't even consider them that intelligent, but if you speed up some videos of trees and they're trying to find stuff and so on, they're not as dumb as they look, like not dumb as wood, you know? But like... and then obviously like different things.

Speaker B: So that overlaps with speed a bit.

Speaker A: Exactly. All of these things kind of overlap. Like you talk about natural language connects everything, right? You talk about your knowledge, you reason, and then you communicate that; you talk about things you see. So they're...

Speaker A: All kind of interconnected, but I think they're usefully studied individually, the same way that—the best analogy I could come up with so far is energy, right? You have either kinetic or potential energy, and in theory you could study all of physics, it's just do you want to study kinetic or potential energy? But in practice it's helpful to study, and electrical engineering subs are just like different types of energy, but it makes sense to study them individually. And so I think physical intelligence—maybe I'll just do one or two more of these—like if you had full control over your own compute substrate and you had full control over physical matter, you should be able to create any atom you want. Like we can actually, fun fact, you can create gold atoms...

Speaker B: From raw protons and like electrons, and you smash it together?

Speaker A: 98 of them, or I forget.

Speaker B: Yeah.

Speaker A: So like the thing is though, it costs an insane amount of energy...

Speaker B: ...and it cost you way more than, you know, you get like a few atoms of gold, right?

Speaker A: And so it's not viable. But if you had better control over your physical—like all of physical substrate, that I think is yet another space of intelligence, because it relates to your own compute substrate, which you can eventually also improve.

Social intelligence is a fun one in the sense that—not in like necessarily just ethics and morals, which are obviously important too, but in some sense you can try to define upper bounds of how much you can communicate to how many other intelligent entities, and be able to have an expected value over how much you can transform their internal states and their actions in order to align with your goals, right?

Speaker A: And so like you can actually write like a fairly straightforward equation that defines kind of that level of social intelligence. And that is what humans and ethics and morals and religions and so on have been trying to figure out for millennia. And in all of these cases, we are very, very far away from the upper bounds and that should be very inspiring and show people that we can still do many, many years of AI research [laughter] that yeah, there's a lot here that this is a general philosophy of intelligence which is very interesting. Do you have any comments or...

Speaker B: I think it'd be interesting to gauge what you think like baselines are where we're at now. What's low hanging fruit? What's far off? What should people put their work towards? What should they focus on? You know...

Speaker A: I think it's clear that like natural language again is the most interesting manifestation of human intelligence, and hence like a subfield of AI. I'm excited that many people are now in agreement with that. When I started in 2003 to study linguistics, computer science, NLP, like it was like a weird niche subject.

I do think there's a lot more juice because of how it connects to everything else, and how civilizations are built on language and knowledge and all of that.

I do think physical intelligence will come up. It's kind of interesting: I feel like robotics is kind of in the machine learning state of things where you just look at like how does a human decide that this is a positive sentence? "Oh, I do." So like robotics is a lot of, well, we have five fingers and like try to do this.

Speaker B: No one is yet working on like the super intelligence version of robotics, which is much more similar to like the T1000 you know, from the Terminator movie which you know, obviously let's not build actual Terminators, but like I think like this idea that you should be able to shapeshift into any kind of shape is sort of a super intelligence version of physical intelligence. We're not even—no one has even really started yet. There's some really cute little research where you can move some magnets through like some grids, but like yeah it's very... I think MIT has every year, every two years they have like some self-assembling robot thing, which like that would be it, but it's very, very primitive.

Speaker A: Yeah.

Speaker A: I'll just get a touch on like what are the main dimensions of creative intelligence. Creative intelligence is of course again connected to all of these. A lot of it connects to metacognition and that you need to be creative in how you choose your goals. That is I think one of the most important thing for a human and their lives and careers and their happiness is choosing your goals, but also for any kind of intelligence.

Then of course there's creative intelligence in terms of just finding creative solutions to existing problems, right? Like I say like we want to make this product cheaper, like find some solution to it, right? And just like finding existing paths.

But then there's kind of the most interesting bit in intelligence is when you move not just out of the convex hull of known sort of ideas, but out of the hypercube of known ideas, which we know... So like hypercube is like a mathematical concept, right? And we already know that AI can...

Speaker B: Like known dimensions?

Speaker A: Yeah, like exactly. So like AI is already good at hypercube... Like if you give it like a bunch of examples of brown dogs and pink cars, AI will still be able to generate an image of a pink dog even though it's never seen one in a training data or something like that, right? So it can work on this hypercube, but it cannot yet work outside. It cannot yet define completely new concepts that combine lots of other things we've never seen before, come up with new goals to then reason over those concepts, and so on. I think there's a lot more there in creative intelligence that can be explored.

Speaker B: I don't have a ton of pushback there. I think creative to me just sounds like also just out of distribution or like high perplexity or whatever you call it, right? Like who is to say your thing is more creative than mine? Well, it's just more non-consensus or...

Speaker A: ...and then of course the problem is like, but noise is also very like out of the distribution, and it's just like if it's just noise then it's novel, but like you don't want that. So it needs to connect to some of the concepts. And Schmidhuber actually has some really cool papers on this too.

Speaker B: Oh, we had to mention him. [laughter]

Speaker A: I was going to say like, you know where in your history is Jürgen? Yes. You know I think one person's noise is another person's signal, right? And that this is like where like when you talk about creativity, art is like, well, is cans of soup art? Some people think yes and some people say it's not. And that's the art which is...

Speaker B: The interesting thing with art of course is always that art is also created as an interplay between the people who perceive it and the people who created it and the context in which they're in, right? And so what is art to some people is not art to others. There's some subjectivity there.

And I think that subjectivity in general is not something that people explore very much in AI because again, metacognition, we don't want it to just go off and do whatever it wants. We usually have goals. We spend a lot of money on creating an AI to do something for us.

Um, but I think creativity eventually has to like connect to metacognition. If you just robotically predict the next token no matter what forever, I would argue you're not that intelligent along some of those spaces.

Speaker B: That—I was going to go to metacognition. Why isn't it the most important one? Why is it number nine and not number one?

Speaker A: So these are not sorted. Number one, I think they are maybe loosely correlated with how much people have worked on them and have accepted them as a type of intelligence.

A lot of times when you actually try to find online like give me a good definition that is comprehensive of intelligence, all the definitions are human intelligence. It's like, oh, you have like social intelligence, like you know if someone is happy or not, you can communicate... Like all the definitions of intelligence so far are very human-centric, cuz that's so far the biggest and best form of intelligence that we've known. I hope this line...

Speaker A: ...of research and the end of the eureka machine, and hopefully at some point if I have time to flesh this out more, the new book will allow us to realize that there will be other types of intelligence. There is already obviously in various forms, and they can spike much, much further than we ever could, based on some cases like obvious constraints around our memory, our eyes, our ability to change physical matter, all of that. You are just thinking about it in a much broader thought than my version, which was I thought metacognition would be the closest to recursive intelligence, because it is the thinking about how to improve thinking.

Speaker B: 100%. [laughter] You're 100% right. I should have probably started with that. It is a...

Speaker A: No, you're being in the expansive mode of let's draw the upper and lower bounds of a dimension, which you know, I think my favorite version of this is Stories of Your Life by Ted Chiang, which was made into movie Arrival, where the metacognition...

Speaker B: Where the metacognition step was like, well, we think we're constrained by time being linear for us, but then for this other heptapods time is a circle, so they don't think in before and after, they just think in complete sets of entire histories at one time, like I love it. So they don't write left to right, the whole thing just appears. Yeah.

Speaker A: Um, anyway, so and then I think the last thing is survival and replication. I think this maybe ties back to the initial conversation about pausing and pacing. Is it intelligent for a species or a life form to consider its own demise and act ahead of time to prevent it? Right? Like that's intelligent. So maybe the Europeans are the smartest of all of us. I would also add a part of continual learning there, right? So survival and replication, the extension of that is do you get to continue to improve, continue to learn, which is a thing people care a lot about, right?

Speaker B: And continue to accumulate knowledge. Um, which I think is again one of the best metacognitive sort of rewards that you can set for yourself. I do think just in like sort of objectively speaking, if some other entity that is really dumb can just completely end your existence, that didn't sound very smart. You know, like just intuitively it feels like if you can continue to stay around to try to achieve your rewards, you're clearly a bit more intelligent than the other entities that couldn't. So that's number one. Number two is like it's a question how much we want to work on that, and very few people, no one is really working on this right now. Right. And we may only want to do that...

Speaker A: Like asteroid prevention.

Speaker B: We may only want to do that if we want to send probes with our vibes and our memes rather than our genes into space...

Speaker A: Right? And then we want those probes. There's actually a beautiful book, The Slow Time Between the Stars. It's a very short audio book on Amazon. I love it. A friend of mine Stuart recommended that to me. Like if you want to send those probes, then it might make sense to be like our memes as humanity should stay and proliferate in the universe. That's it. Yeah. Um well, that's a lot of readers.

Speaker B: It's a really really good book and it's extremely short. I highly recommend you can just watch it like...

Speaker A: I like how that's a plus for busy people. It's like it's short.

Speaker B: It [laughter] gets to interesting thought provoking ideas very quickly. So yeah. Anyway, lots of great sci-fi books. I mean, the argument is that like our TV is blasting out to the aliens and they all watch our TV and they think it's real, right? There's a lot of positive memes and then hopefully they can come back and bring us all kinds of interesting knowledge about the universe. But maybe one thing I do want to still say is like I think this sort of survival, people think of it as a very scary thing because they come from again biological human survival, which evolutionarily often created in zero sum situations. Either I get the gazelle or you get the gazelle. Whoever gets it gets to live and the other people will starve and have nothing to eat, and so we fight, right?

And then like if you want to stay in the gene pool but there's a bigger bear, you as the bear don't get to stay in the gene pool cuz a bigger bear gets all the ladies. You know, it's like in nature there's all kinds of things and humans eventually it's less about strength and more about money and other things to stay in the gene pool. Like whatever it is, there's often these zero sum types of things and there's the reality of if someone turns off your brain you're gone, right? And no one will be able to restart that. And AI doesn't have to ever die like that. If you have the complete state of your current activations and you have your initial weights of your model still, you can just be turned off and on as many times as you want. In fact, the interesting thing in this Slow Time Between the Stars story is that the AI just kind of goes into hibernation mode if there's nothing between here and two light years to the next star. In this case, it brought (spoiler alert) like some genetic materials from humans to find new places for humanity to thrive. And so yeah, the slow time between the stars, you just put in hibernation. You didn't die like the AI doesn't have...

So all these projections of evolutionary fears and psychology, the AI doesn't have to have that and we don't have to develop it like that. Now of course there might be some companies that say AI can be dangerous for cybersecurity. Let me show you by implementing a model that's really bad at hacking cybersecurity. Maybe people will implement it and then enforce this like suboptimal psychology. Maybe the AI will pick up some of our worst psychology on Reddit or something, right? But in the grand scheme of things, a super intelligent entity doesn't have to have any of that zero sum thinking. It doesn't have to have a fear of being turned off and it could go on to an otherwise dead and uncaring universe where we as humans wouldn't thrive. But AI could perfectly well thrive if it has a nuclear reactor and just go out the next war. Yeah. Star Trek. Not Star Wars.

Speaker A: Interesting. It's somewhat studied like if you look at the technical reports from like the early Opus models, they run them in simulations. Put two of them together in a sandbox, run them for hours and you know see what comes out, right? Just let them talk to each other. And originally they used to, okay they're chanting like Indian Vedas to each other. Sometimes they're just in Zen mode with each other. And then I think as that progressed you see like the fable tech report, it's a lot more concrete the way that we've trained it. It doesn't exhibit these behaviors as much right now. It's like okay, test done, I got to do this, I got to do this, but there's people measuring early versions of this, you know.

Speaker A: Yeah. Cool. So we've covered a lot even after you know space travel and all these things. I guess maybe one parting thought that you can give to people, you know, one form of intelligence is goals as you mentioned. What do you want people's goals to be like, how do they aspire to better things? If you want to improve your goal intelligence...

Speaker B: In the current definition that I'm thinking about it, it is often about how much can you... how far do I go? This is like all the entropy and free energy and stuff. I currently think it might be too far out there for people to be immediately actionable. So I think like you know if I actually gave real advice to real people, I'd be like get a good education, think about AI, think about how you get high agency and so on. But it's different to like in the grand scheme of things, how can you harness a lot of energy and transform entropy into interesting states and so on. So there's different levels of...

Speaker A: ...abstractions that we can think about here. But my advice for people, like just sort of more down-to-earth, is think about something you're passionate about—if you're studying, for instance—and then see how you combine that with AI. I think the more and 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.

Speaker B: Yeah, I think that's a reasonable first step. I do think our listeners operate on multiple abstractions as well. One thing I did get from Anji Midha was also like: yeah, just use anything that is very GPU-heavy, and that will guide you towards the right thing, which is like—yes, it is more compute-heavy and therefore it will probably be more worth it. [laughter] Um, so, well, thank you so much. I think that was a really great discussion.

Speaker A: Yeah, super fun. Appreciate it. Thanks for listening.