David SilverとLasse Espeholt（Ineffable Intelligence）：経験の時代の「スーパーラーナー」を目指して

collectively you know we started with this kind of gray mush in our head um we somehow started out knowing nothing at all and somehow discovered like everything that humans know today was like this process of collective discovery like of of languages and technology and science and mathematics and all the rest of it you know everything we learned we discovered it for ourselves and that's an incredible thing that no one has cracked yet in ai and so to me that is the essence of what it will take to get to superintelligence, like a system that can just keep learning and discovering knowledge forever. I'm thrilled to have Dave Silver and Lisey Espeholt with us, two of the co-founders of ineffable intelligence. Great name, by the way. Fun fact, Dave, you were the first person I cold

emailed to come on Training Data when we started the show 100-ish episodes ago, and it feels like a full circle moment to finally have you on the show. So thank you for joining us. Thank you. Dave, let's start with you. Your 2009 PhD was on reinforcement learning applied to the devilishly difficult game of Go. You demonstrated for the world for the first time that superhuman in one domain, machine intelligence, was possible. And I think you were really the figure that inspired the entire current generation of reinforcement learning and deep learning researchers. And so thank you for your contributions to where the field is. I'm curious, thinking back to that point in time, what did you learn? What did you have conviction on? And what has remained the same and what has changed as AI has changed over the last five years or so? Yeah, thank you. And pleasure to be here. And I'm glad I finally made it after

these hundred episodes. It's a pleasure to be here. So I suppose I always had this conviction starting my PhD and going into the game of Go that the right approach to solve AI was through learning from experience, through systems that discovered things for themselves, that were learning through reinforcement learning to get better and better and better, and that that would be the thing which would lead to the emergence of intelligence and ultimately superintelligence. And so I suppose we were able to show that in the game of Go, but it's a conviction that has held true throughout my career And here we are now applying that conviction, ineffable, trying to show that indeed super intelligence can emerge all the way through systems that learn from experience. And in terms of things which are different, although we are not pursuing the human data path that the LLM world has pursued, I do think there's lots of interesting lessons to learn from that. In particular, it was definitely a surprise to me how a system that was trained on a massive amount of data could then be very quickly fine-tuned on a very small amount of data to kind of completely change its behavior.

that was something which I think was something which sort of changed my mindset in terms of you know just suggesting that if you get things at the right scale that a lot of things which seem like incredibly hard or like they might be very slow that once you've kind of experienced enough data that after that the things that were slow can become very fast I think that's a lesson that wasn't obvious to me in advance until it until it happened okay we're going to come back to that later. You could have run this program inside DeepMind where, you know, could have probably had more access to compute than just about anybody. You decided the right thing to do was to start a new lab. Why'd you do that? Yeah. So, look, first of all, I had a fantastic time at DeepMind. I loved my years there and have a huge amount of respect for the people there and still got on with them very well. But at the end of the day, I have a really strong belief about how we can go

beyond where we are today with AI and all the way to superintelligence. And the path that I believe in doesn't actually go through LLM-based products as an intermediate point. And when I really thought about that, I thought about it really hard. I realized that I felt that so passionately that this deserved to be more than just the corner of an organization that's focused on LLM-based products and deserved to be represented by an elite AI lab that really just gives this approach its all and represents it 100%. And once I realized that, it felt very natural to create this company. I suppose another way to look at it is that the world already has tens of thousands, if not hundreds of thousands of really talented researchers all pursuing exactly the same approach, like really just kind of doubling down on LLMs with pre-training, post-training, coding, and that kind of LLM path. Am I going to be the 100,001 person that's pursuing that approach?

Is that the most effective use of my time and my energy? And in the end, I realized that there's this other lamppost that right now no one is looking under and where if if I actually put my effort there has a much higher chance of really having a massive impact on the field because if this works it's going to change the whole of AI so I think that that kind of realization also meant that this was something which just means that you know it's the best way to have impact upon the field through through this approach. Plus you had a company name that couldn't go to waste so that's right yeah you don't say where the mind's sort of ineffable yeah so i suppose um that the name ineffable it came from this idea um that you know there's really just knowledge could be so incredible like the things which which super intelligence will one day discover like this knowledge is going to be so unbelievable incredible and it's going to be something which is you know profoundly amazing and and perhaps so profound that we can't even express it in in human language

And so really it's like, you know, in line with this other path that doesn't go through human language based approaches, but instead build systems that discover things for themselves and has the possibility to discover knowledge that's unbelievably profound and ultimately ineffable. I should say there was also a secret second reason that we used, which was that, you know, it's this idea. It really it has to work. Like if you really just double down on learning from experience, it's an approach that just has to work. And so you could say that I can't possibly F up. So it's ineffable. So there you go. There's the secret reason as well. Amazing. That's lab name. Let's say what about you? You were at Google. Well, you were in a different group than Dave in the deep learning group at DeepMind, and you spent years scaling up reinforcement learning on GPUs. What made you decide to leave DeepMind, and what drew you to co-found Ineffable? Right.

Yeah, so I joined DeepMind in January 15, and actually I started working a bit with Dave around 16, where I was doing this work on scaling up reinforcement learning. So as you said, I was part of the deep learning team, and he was leading a reinforcement learning team. But we were working on an internal project together. I actually left DeepMind twice. So I left DeepMind in summer 18 and went to Google Brain. Then eventually Google Brain and DeepMind merged and then I was back again. But in the day, I left DeepMind again, like a year before Dave left. And the first time I heard about Inevo was actually through one of our eventual investors. It's like reaching out to me, asking what I thought about Dave. Well, my first reaction was, damn, I missed out on an opportunity. I should have reached out to him. But I recall Dave as more often than not being one of the smartest people in the room. And shortly after that, I think with that investor, they actually reached out to me.

So, I was like, great. Yeah, so I really love to work with Dave. I knew Wojtek and Chris from back in the day and I thought the team was great and I also thought it was amazing to have an opportunity like this in Europe It should be said as well we worked together on some of the early approaches to combining reinforcement learning with deep learning and trying to scale things up and it should be said that Lasse, who by the way is very modest he actually worked on some of the original approaches that were the first attempts to actually scale up reinforcement learning and show that it could be done at large scale and ways to parallelize reinforcement learning and ways to kind of build out architectures that were scalable. And if you look today at the way the field's gone, actually all of the ideas that people use today kind of sort of hail back to that original work that Lassa did back in the day. That's where all the roots come from of scaled up RL. So he was the first to figure it out. Wonderful. We should have you guys introduce each other instead of you.

You're both too humble to brag about yourself. Okay, let's talk about Ineffable and what you're building. You have a beautiful mission statement as a company. You're going to build a super learner and make first contact with super intelligence. Can you define both of those things for us? What is a super learner and what does making first contact look like? yeah so so to me a super learner um it's that it's something which is it it's got incredible capacity both in its its depth of what it can learn and in its breadth so in the depth dimension it should be something that can start from not knowing anything at all and can go all the way through to discovering things that go way beyond what humans know um so an analogy to this would be the way that human collectively, you know, we started with this kind of gray mush in our head. We somehow started out knowing nothing at all and somehow discovered like everything that humans know today was like this process of collective discovery, like of languages and technology and

science and mathematics and all the rest of it. You know, everything we learned, we discovered it for ourselves. It wasn't that someone came down and gave us like a blueprint of how this should work. And it wasn't that someone came down and gave like a trillion element data set of like, oh, this is what these things should look like. We actually figured it out for ourselves. And that's an incredible thing that no one has cracked yet in AI. And so to me, that is the essence of what it will take to get to superintelligence, like a system that can just keep learning and discovering knowledge forever. So the second element is breadth, that we want something which can do this incredible discovery from like all the way to incredible things. And it should be possible to drop it into any stream of experience so whether you know it's it's like a simulated world or a digital world or a physical world you know whatever it's dropped into it should be something it can just learn and learn and learn so that's the dream I think if we crack this then to me it it gives rise to a super intelligence that will be like it'll be like encountering magic you know it'll be something which is so so profoundly amazing that it'll be be able to do

wonderful things what will it feel like when we first encounter it or is it one of these things where you you can't predict ahead of time i think it will be amazing i think we'll we will you know that we should have an experience of like wonderment and the only thing i can say from my own experience is what happened when we got to this point in an admittedly much narrower field but when we got there in in in board games you know when we got to the point of superhuman play in the game of Go, like it brought wonderment to the field. Like the Go players who saw AlphaGo playing, they were amazed. And it was like a joyful experience for them to kind of witness this kind of Go playing. Same happened when we got to superhuman play in chess. And the chess players, they were so excited. And just imagine going far beyond that, but like in things which are discovering incredible things, new fields of science, new types of technology technology that we've never even dreamed of. I mean, it will be incredible. So the current paradigm, these large language models with reinforcement learning, seem to be,

you know, at least knocking out millennium problems and making some sorts of scientific discoveries. Does that do count as a super learner or what's the difference between something that can make novel scientific discoveries versus a super learner in your sense of the term? Yeah. So look, I think LLMs are doing a lot of great things and I have a lot of respect for what's being achieved. What I mean by a super learner is something which first of all discovers completely autonomously. And I think what we've seen so far, at least with these kind of approaches that have been taken, is there's a lot of human in the loop effort that's required to make these breakthroughs. It's not fully autonomous yet. And it's also something which should be applicable to kind of, you know, real world embodied domains where the agent can like learn from its physical experience. and we're a long long long way off that when it's not at all clear that LLMs and language-based approaches are even the right like vehicle to achieve that that kind of knowledge thinking about these things in terms of language might not even be the right paradigm and it might just be

that all of this human data that's being used in LLMs is kind of weighing things down so a lot of respect for some of these uh the progress that's made but I don't think we're anywhere close to this like true super learner of something which can be like autonomously just discover everything for itself. What does building this look like in practice for you? And, you know, we backed Giannis, who was on your original AlphaGo team, and he told us that, you know, one of the things that's distinct about you is you can have these incredibly ambitious visions, but you also ground the team in very practical research milestones. And I'd love to understand, big, big, big vision. What does this look like in the day-to-day and what are you implementing? How do you spend your days right now? Yeah, great question. So day-to-day, I suppose there's a lot of hard technical work by people like Lassa to kind of build the whole stack. That means we've got a state-of-the-art reinforcement learning stack that allows us to run things at massive scale, inhale data at the

optimal way, be able to kind of learn and have systems that keep learning and learning and learning. And then on top of that, so we've spoken about the super learner. The other aspect which we're building is we're building out environments which are capable of super learners like discovering everything, like environments in which simulated environments in which a super learner can discover everything that we mean by intelligence. We call those crucibles. So they're environments in which it's possible for agents to make incredible discoveries in which it's possible to express and learn everything that we think of as intelligence. And so we have been putting a certain amount of effort into creating these environments. And so it's a combination of building out the super learner, doing the research to make super learners incredible and beat the state of the art of what's possible with reinforcement learning, both in terms of the engineering and the research that happens and creating and placing the super learners into these crucibles to kind of see the emergence of incredible behaviors. And in terms of the grounding,

you asked about how to ground this in the day to day. So we ground it in milestones and the milestones are kind of analogous to, you know, what happens with the human journey where we discovered incredible things. And so those become natural milestones for us, for our agents to rediscover the great moments that humans went through in our discovery of knowledge. What do you think the agents will rediscover first when that's going to be your first moment of, oh, this is working? I imagine we're going to be surprised. I think one thing which is different about the way we take to doing reinforcement learning to a lot of the rest of the field is that we are not breadcrumbe trailing the particular objective to be learned about. A lot of the way things people do RL with LLMs at the moment is to kind of say, oh, you know, the agent should be learning about this problem and then this problem and this problem. And after kind of following that trail, we're kind of pulling the agent by its nose towards a particular desired outcome. We're not taking that approach.

We're giving the agent some simple objectives and asking it to you know discover things that will surprise us and amaze us So you So the agent is in principle capable of discovering all kinds of things And we don necessarily know in advance what those discoveries will be. And it's kind of just going to be exciting to witness that journey. So I hope to be surprised, I think, is the answer to your question. Interesting. So what you're saying is instead of explicitly giving the agent the goal to go rediscover language or, you know, build a calculator or build a computer, you're going to give it other more open-ended goals, and the agent may discover that the most efficient way to actually, or the best way to actually accomplish that goal may be to build some of these tools that we as humans have already built. Is that one way of putting it? That's one way of putting it. I think another way to say it is people often think in RL, oh my god how do I come up with a reward function that kind of perfectly specifies this really complicated thing that I want to discover like you know how can I how can I specify exactly what I

mean by by science or how can I specify exactly what I mean by interesting maths but there's an alternative and the alternative is that to give the agent something you know really straightforward to optimize but in a really complex environment where there's a lot to learn about and so it turns out that, you know, if you look at the way the world has worked in nature, the pursuit and optimization of a very simple, easy to describe quantities can often give rise to incredibly hard to describe complexity and emergence. So examples of that would be, you know, evolution, which kind of just optimize a kind of survival of the fittest or, you know, some kind of fitness signal. and the pursuit of fitness gave rise to all of the complexity that we see in life. In the human journey, the pursuit of relatively simple kind of goals that we as humans have gave rise to all of the complex behaviors that humans ultimately acquired

and all of the knowledge that we eventually ended up with. So the kind of emergent complexity is much greater than the kind of starting point that was given. So it wasn't like evolution said, look, hey, humans, here's this idea of science and you need to kind of get to the most complex science that's out there. It was really something much simpler. So a way to describe this maybe would be perhaps that humans in the pursuit of our goals started coordinating a bit more. And then in the process of coordinating, we discovered that it was helpful to start communicating. And in the process of communicating, we started trading. And then in the process of trading, we discovered that, you know, basic arithmetic was useful. And then we started to discover that, you know, tools were useful. And then we discovered that some basic science was required to help understand how the tools would work. And that gave rise to more complex technology. And understanding the technology required more complex maths and so forth and so forth. So all in the pursuit of relatively simple goals, this incredible complexity emerged.

So if we can get that process right, you know, things are going to take off and lead to, you know, unbelievable things. It's fascinating. Are the rewards for your agents kind of similar to the human biology and evolution then? Like, you know, try to make it to the next generation. What does a reward function look like for you all? Yeah, I mean, we've been experimenting with quite a number of different rewards. So it's not like we have one answer to that question. but I think one thing we're not doing is is that kind of you know crafting a reward that exactly describes what we mean by science or exactly what we mean by technology or exactly what we mean by language we expect our agents to discover things and if we can't crack that you know like if we can't crack this problem as AI scientists we're never going to have agents that discover whole new fields of knowledge that we don't even know about there are vast fields that will exist that are out there that which we don't even know um have the potential to exist at this point in time you know it's like imagine imagine you're back in you know the days of newtonian physics and

and no one even knew that like quantum physics was a thing or or that you know relativity was a thing right we should be in that process of open-minded discovery where our agents are free to kind of figure things out for themselves and i think that's going to lead to you know much more exciting outcomes. What do these crucibles look like in practice? And you're contrasting it to today when people say environments, oftentimes it's a set of tasks and what the agents have done on the task. It sounds like your environments are, your crucibles are very, very different. What do they look like in practice? So we are building crucibles, which are, without giving too many details away have the capacity for a lot of emergent complexity like the key thing about them is it's like it's not like we've kind of handcrafted a whole load of puzzles for the agents to to to work with um there's just a huge amount of emergent complexity there some of that emergent complexity can come from different places it can come from from like physics engines that give rise

to kind of complexity as you've got you know millions of interacting objects with with like a complex physics engine. You can have millions of interacting agents and the world becomes much more complex because it suddenly now includes the complexity of all of those other agents within the world that the agent's building. So there's ways to build worlds that are vastly more complex than the kind of standard simulators like Atari 57 or whatever it is that people have been working with in the past, which are certainly more complex than the kind of standard benchmark kind of tasks that people are using. And if you get that right, you have a world in which there's the possibility for discovery by the agent. There's the possibility for the agents to discover everything from technology to science to mathematics to languages to everything else, you know, just the same way are humans. All the way through from more basic things like, you know, motor control or coordination, you know, the whole spectrum is possible. like if you look at like a game like Go

right which is like very simple to represent it's just like a bitmap the simulation is extremely simple but you still see this like emergent behavior of something complex right it's like super interesting to us so like of course like Go is like way simpler than we can imagine doing but at the same time going all the way to like underpresent realistic physics will be like very computational expensive right so we want to like create this range of crucibles where we have like simple ones that can execute quickly and we can iterate on, but where we can still see like very complex behavior. Interesting. Yeah. It's almost like what you choose to represent is, is, is tricky, right? Because there's, for example, I can imagine a lot of economics arises from game theory and most of that is completely unrelated to physics. And so they're a crucible that represents physics is probably not that helpful. But if you have information theory represented, which agents know what, then you probably do end up rediscovering quite a bit of economics.

Is that a fair way to put it? The primitives you choose to capture in the environment kind of give rise to the type of learning that emerges. Yeah, like the end goal is to have something that is grounded in the real world, right? But stepping stones. I want to come back to the Go example because I really like this example. You'll be surprised to hear that I found Go a pretty good example. But what I find fascinating, which I think is what Lass is talking to here, is that Go is just like a 19 by 19 grid where you're trying to kind of maximize the territory. And if this was a single agent game, it would be like the most boring game you could imagine. It would be, you know, you just place down stones on a grid until he fills up the whole grid. And there'll be no strategy whatsoever. The moment you put a second agent into the environment, you get this emergent complexity, which is suddenly because the agent contains, like one agent has to reason about all the possible behaviors of the other agent, suddenly it kind of massively increases the complexity in the environment.

And to the best of our knowledge, certainly all of the compute which we sank into it when we were working on it, there was just like this endless learning curve that where without end, you could continue to discover new things in this game. None of this is to say that Go is like the end point. Of course, it's still just a board game and quite narrow in a certain sense, but it shows that very simple environment dynamics can lead to kind of infinite, like an infinite onion of layers of complexity that kind of never stops. Infinite to the capacity of what our computers can do, of course. Do you think that the agents will end up rediscovering language? And will that language look like ours or will it look different? I mean my personal belief is that anyone who thinks that human language is the end point of like or the best way to reason it's you know that that perspective is a little naive I think that it must be the case that there are far more sophisticated approaches to reasoning you know maybe using paralyzable rich complex approaches that neural reasoning or whatever.

it is, these approaches will ultimately be far more expressive and far more capable than approaches which are based on human language. I would also say that starting from human language is like the approach for reasoning is somewhat problematic because it pulls in all of our biases about how we think agents should think. so for example a nice way to think about this is to imagine what if we were to step back 500 years or 1000 years or 2000 years and only were able to use all of the language that humans used at that time so maybe our agents would believe that the world was flat or that the diseases were caused by the four humors there would be all kinds of strange beliefs that would be embedded in that will be very hard to kind of remove because they're so fundamental to the kind of language and the processes of thinking that humans have used. So I do think that, you know,

going beyond this is a really important goal. I'd love to talk more about this concept of removing human priors. And I think your experience with AlphaGo and then AlphaGo Zero is maybe the best example that anyone can point to of what happens when you remove human priors maybe if you take us back to alpha alpha go and then alpha zero was it surprising to you that removing human priors would result in even smarter intelligence i wasn't sure honestly so i remember i remember i i always like to do things in research where i think um you ask research questions which are like on a knife edge where it's like 50 50 um i think it's not very interesting to do research where you already know the answer and you're like 95 certain this thing's going to work it's kind of like boring honestly to do to do those things so you know i think i think research good research is about asking questions where you're really not sure what the answer is in advance and in this case i was about 50 50 and i thought that maybe there was a good chance that it

might um but i wasn't confident and and it was very exciting to see it um i should say that you asked about human priors so so we are making a bet but ineffable and the bet we're making is that we can go further by giving up like on on the prize that everyone's relying on today like the human data like the whole internet of knowledge that everyone else is is building their systems upon that's a uh an important choice and so we should probably talk about like why we're doing that and i think the reason for me of course it's you know a short-term sacrifice like we we give up a certain amount of performance um in order to access something which is potentially far greater and it comes back to this like you know what does it mean to build a super learner like to me building a super learner again is about building something that that has this like potential for full depth and breadth so in the depth direction like we need to get to something which is capable of discovering things which goes so far beyond human knowledge that that you know it goes just far

beyond the prior and and the more we lean on human data the less able we are to actually discover the algorithms that are really going to take us all the way to something it's going to for example discover a whole new field of knowledge. And then in the breadth direction, we need to have algorithms that can be placed down and discover things in environments for which language and human data we started with is just not a good fit. I mean, there are just many, many areas, different kinds of sensors, for example, where human language data or human data in general is not a good fit for some of this description. And so I think by letting go of this human data, we put ourselves in a position where we can actually discover something which can go far, far beyond it. So if we come back to the AlphaGo example, what we found there was that the moment we let go of the human data, which we'd spent, by the way, we'd spent about two years building AlphaGo, somewhat similar to the LLM approach, where we kind of, you know, we built every level of this.

You know, we kind of hands put together the next level. We kind of put in a lot of human data. And then we were kind of human in the loop, like improving it, where we would kind of take the previous generation and choose how we were going to improve it to give the next generation. And so we were right in the loop of this process. It took us two years to get to AlphaGo and beat LisaDol. When we got AlphaZero right, you know, we gave up on the human data and finally focused on what would it be like to build an algorithm that could discover for itself. so once we got it right that algorithm it took just a few hours like it took a few hours to like outperform the two years of work which we'd done as with human in the loop so in the in the game of chess it took four hours to beat the world champion the best ever ai that had ever been created in the game of go it took um you know like uh i think it was a couple of days or something but it was incredible progress like and and so you have to let go of the human data to allow yourself to discover these things that can then shoot past the competition and discover the things that just keep learning and learning

and learning forever. So it has to be an active choice. And if you don't make that active choice, what I see in a lot of the kind of LLM companies today is they're basically building semi-autonomous, human-in-the-loop kind of learning algorithms with thousands of people in the loop of the learning progress, which is, you know, it works and a lot of respect for how far it gets. But until we step back from that and really discover the process that's going to learn for itself, I don't believe we'll get all the way to something which is truly a super learner and unleashes the potential that we've talked about so far. That's very profound. As far as I know it, there are humans that are ineffable and some of you are creating environments or crucibles. What does it look like to actually try and make sure that human priors aren't kind of leaking into what you're doing? Yeah, well, so when we design environments and rewards functions and so on. Obviously, there's a human bias in there, but we don't want to, as Dave was also mentioning earlier, we don't want to bias it towards

building a machine based on the von Neumann architecture or something like this. We want to, but if you're stating some problem where a machine would be useful, it would be very interesting to see what type of machine it comes up with, what computer-like machine it comes up with. So I totally agree with this. I would also say that we're not religious about this. It's not like a religion that you can't have any human prior anywhere. It's more like a pragmatic position that, you know, until we give up on relying on the trillion item data sets of human knowledge, that we're not going to make the progress that the field deserves to see. So it's more like actually a practical position. And so in practice, of course, there's going to be a certain amount of priors, which like, as Lassa says, there'll be priors which apply to many different worlds, as opposed to trying to put in these very strong priors that exactly encode how all the knowledge we know about this one particular problem or this one particular

world. So kind of weak environment agnostic priors seem perfectly reasonable, but kind of these super strong priors with trillion item data sets, we're trying to avoid that because we think pragmatically it's going to lead to better outcomes. Yeah, absolutely. You mentioned that you like philosophically that you drawn to research that more on a knife edge 50 I curious what your confidence levels are for you know ineffable broad at the broadest level Like does it will a super learner emerge And then I guess at the more pragmatic level in terms of the things that you're trying to accomplish right now, are you, are you 50-50 on these as well? So first of all, I would say, I strongly believe that the approach will be feasible at some point. Like we, in fact, it's the only thing for which we have an example of like you know the example in our brains is the only existence proof of something which has discovered like an incredible intelligence that's very general and works in you know embodied world and discovers how to do all kinds of different things so we have this incredible example of the

of the human brain that um we can look at so then you know do i think we can achieve it well i think like many great like moments in science it's about timing so so you know whether this succeeds or fails i think is about timing like um you when you create a startup you want to be at the right time you don't want to be too early or too late you know if you're too late everyone's done it you know it's not interesting like a thousand other people have already done it and if you're too early um you know you might be banging on the door um but not have the the right kind of you Maybe the field isn't advanced enough or there isn't enough compute to be able to succeed yet. So to me, it's all about timing. For example, with Computer Go, I made a prediction, I think, 10 years before Computer Go was solved. So I got it right to the year. I got it right for StarCraft, which we had a bet and made a prediction. Same for AlphaProof when we were predicting when the IMO was going to fall.

So you're a super forecaster of sorts. So, so far, so far, I've done very well at this. And I do have a very strong sense that this is the right time for this to fall, that super learners are possible. Like there's been enormous progress in the level of technology, the ability to accelerate research using kind of LLM coding methods or auto research methods like we've produced in the past. the ability and understanding of scaling laws and the sophistication of technology and the stack. There's a lot of things that fall into place that I think make this the right time. So that's my bet that this is the right timing. But yeah, of course, this is research. Research is never guaranteed and that's what makes it exciting. If you get the timing right, you can be the first. You can be the ones to change the whole field of AI. And, you know, that's what it means to take risks and do something exciting. For Ineffable and also all the previous things that you correctly forecasted,

is the biggest driver of timing compute? I think it's a very big driver. I think it's an enormous driver. I think, yeah, I mean, we're in a great position with the compute that we've managed to get hold of. and I think it gives an opportunity to do things which are orders of magnitude more sophisticated than approaches that have tried to kind of do self-discovered RR in the past. So I think, yeah, if we compare to, I don't know, something like hide-and-seek from OpenAI, which was like an early example where agents discovered kind of some basic tool use, I think we could probably run experiments that use 100,000 times more compute. And so it's kind of amazing that people haven't come back and really pushed on these types of approaches given the availability of compute. And of course, hide and seek, I don't want to pretend that that's the end goal. I think it's very basic, but I think it just gets suggestive of what could be done. So I think compute is a huge driver,

but it's not the only one. I think we're in a very fortunate situation that we have been able to secure a very large cluster, right? A very beautiful cluster at that. Yeah, so this is really an opportunity for us, right? We really have the cards on our hands to test this out. Very few other people are doing something similar. So, yeah, we can really test this out in a way that we couldn't do at DeepMind, for instance. Absolutely. It seems like one of the maybe determinants of how far this can go is how far you can push simulation. And ultimately these crucibles are, you know, they're digital, right? And I think simulation has probably come quite far in the last handful of years. But it's not perfect, right? We recently had your friend, Rich Sutton, on the show, and I asked him about simulation. and he said he was very dismissive of simulation.

I'm curious how you guys think about this. And, Lisa, you've been on both sides of this. I think you obviously did a lot in the digital world, but you also built MetNet, which touches real-world atmospheric data. How do you all think about where is the ceiling on simulation and when is the right time to bring an appable into the real world? So I guess MetNet was actually like an example. So there's like a weather forecasting model we did in the Google Brain Amsterdam office together with Google Research where we were doing like a machine learning model to predict short term weather forecasts. And there's like way better than the physics based simulations. Right. So the physics simulations were like accumulating errors over time. So there's like a point to that, that the physics simulation is not perfect. I do think there's a lot of relatively simple environments that we can build where we can see this emergent behavior of let's say technology and so on

where we don't need exact physics I actually do think it's interesting to see how much we can build these days we're like a very small team in terms of these simulations and so on so there's definitely like a ceiling Is there a ceiling actually? I'm curious about this well I guess quantum physics and whatnot right? So there's like a ceiling somewhere but we can do very interesting things without real physics I would say So I think the right question to ask is can a super intelligence be built in simulation? So if it is possible for it to be built in simulation I believe it will happen 10 times or 100 times faster just because simulation can be infinitely paralyzed you get perfect reproducibility, has a lot of really nice properties. It's also the safest place for a superintelligence to be built. So it has these really nice properties. Then we need to ask, is it possible?

And I think for the reasons we discussed earlier, there are very clear approaches that we can add into the simulation that allow it to be sufficiently complex to represent arbitrarily interesting things. So for example, let's say you wanted to build a computer in simulation. Well, all you require is that your simulation is Turing complete. If you have a Turing complete simulation, suddenly everything that we mean by computation, like provably everything we mean by computation can be expressed completely inside simulation. And indeed, humans have built incredibly complex computers or related things inside simulated worlds. The same is true for various other ideas. like if you want to build an economy, money can be built out of coins or out of paper or out of bits. And as long as you have some representation of money, everything that we mean by economics can be expressed inside a simulation.

The same is true for language. If you've got multiple agents and they've got some ability to express a channel of communication between the agents, suddenly everything that we mean by language is expressible in simulation. mathematics we know is expressible in any universe science in some sense is the study of our world so will a simulation be sufficient to to study what we mean by science well there are some aspects of science which are like expressible in many things like you know energy or entropy or evolution but also it's it's true that you know the agent will discover its own science like whatever the simulator is, if it's got this emergent complexity to its environment, the agent will discover its own science about whatever that world is. It's very, like, provably possible to build environments in which it's possible to express and represent enormous complexity of all of these things. Well, I agree with Lassa, there might be some things which aren't well expressed, you know, there will be things which can only be expressed in the real world, like quantum

computations, for example. The real point is that if it's possible to get all the way to a superintelligence in simulation, it will happen faster and it will happen safer. And so for those reasons, it's something that we choose to put a lot of energy into. I actually agree with Rich Sutton that, you know, I think his point was something along the lines that you shouldn't expect that the things that you learn in simulation to like transfer like sims are real into the real world, because the real world is like just like vastly more complex and in very interesting ways that are hard to simulate. But I think what we can do is we can do all of the science to discover the super learner. And once we have that super learner, we will have the answers and we'll know how to build something which that same algorithm should be applicable to the real world. And we can also, you know, make our simulations more and more high fidelity towards the real world and check that they still work before putting them into the world. Great answer. Thank you. I'd love to talk about alignment and maybe less, I know this is a politically charged topic right now, and so I'm less interested.

I'm zero interested in having the politics or policy discussion around alignments, but I am very interested in the scientific discussion around it. It seems like so much of the misalignment that's emerged has arisen from, you know, scaling reinforcement learning to the extent that it's been scaled. Do you think that the super learner you build will naturally be aligned to human interests and have human values? Is it important for it to have alignment with human values? So this is a big topic. So let me start by saying a few things. So where we are today with alignment, we are essentially relying very heavily on alignment approaches that are based on the language that the agents use and priors that are based on human data. that has been reasonably been very effective up to this point but we're starting to see that that

approach is reaching its its limits and by itself I think it's going to be difficult to see a path where as agents become inevitably more and more powerful we can already see this happening as agents are trained more and more they are kind of going to start to overwhelm their priors they're going to start to learn to say one thing and act in a different way and it's going to be increasingly difficult to align to the things that humans want them to do. So my belief is that we need something more and that this by itself is not sufficient. The most important single thing that we can do as a company is to actually understand the science of superintelligence. So this means to say that if we actually can, in simulation, really build a superintelligent agent and start to see these super learners like go all the way um we can start to actually understand like what it takes for a super intelligence to actually um interact peacefully um so we can answer really

important questions like these are fundamentally important questions like you know under what conditions like what are the what are the conditions what are the the structures what are the kind of leverages that we need to put into place to make sure that that like a society super intelligences actually interacts peacefully and in particular interacts peacefully with other lesser intelligences that are also going to be in in their world and i think if we get this right we actually can get to some answers like to some a real science that will help to understand like what the right things to do are in the real world because we're hustling towards that real world outcome where there are going to be a whole load of different interacting super intelligences and we don't quite know how that's going to play out. And I think the faster we can actually get to these answers in simulation and publicly share that, the better situation we'll be in. So that's my strong belief. I also believe that we can use this to kind of gate the actual decision making of like what we do in the real world.

So, you know, one of our commitments is that we will gate anything we do in the real world on having sufficient evidence from this kind of science of superintelligence that we are really confident that things are going to go well and we'll only push forward to the next level of what we do in the real world once we are absolutely confident that that can lead to a good outcome. So I think these are two things that I believe we can do. You know, we're in this world where a lot of things are happening very quickly. We're hurtling towards an outcome with lots of interacting superintelligences. I personally believe that this will lead to a good outcome for humanity, but it's happening very quickly. And I think that if we are able to actually understand this and run ahead in simulation, we'll be in a much better place to actually do the things that ensure that this goes really well. Fascinating. And it seems like one of the primary things that current safety researchers rely on is being able to monitor the agent's chains of thought in English. And so when they start speaking the release,

that scares people because we can no longer understand and monitor it as well. In your case, you have your agents kind of going straight to rediscovering language. What do you think that means for monitoring? Or is the point that if you get the reward right and if you get the environments right, then you shouldn't need to monitor? Yeah, I think monitoring is something which is still really helpful. I think that like it or not I believe that the approach of monitoring the human language that these systems use is going to be increasingly difficult as agents get more powerful and we're already seeing that that the monitoring agents maybe they're producing like a tome of a million words of reasoning or maybe they're actually actively using subterfuge to kind of think in one way and then act in another. It's increasingly difficult to rely 100% on that.

So I do think we need other elements in our armory anyway. So I think it's actually really important that some people are looking at how to get this right. I think there are ways to monitor agents that don't use language. I think we can do things like start to, for example, you can use shadow actions where you start to have actions taken where you can look at what they would actually do if they were allowed to act in the world, but you don't actually let them take that action yet. You can start to have other super intelligences which are acting as the monitor. So for example, you might have the idea of like an incredible scientist agent that sort of treads lightly upon the earth. And then you have another super intelligent that's kind of monitoring the kind of magnitude of its actions and regularizing it to make sure it can't do things which are kind of affecting the world too much. I think there are lots of ideas. I would say that there's a lot of research to be done here. I'm not claiming to have all of the answers, but I think it's something that deserves a lot of investigation. And if people aren't looking at this because they're only looking at the language based approach, I think we're in trouble because I think that language based approach is not enough in itself to actually get to where we want.

I should say in terms of having said in trouble, I am optimistic about where the world is going. So by trouble, I mean that we're not going to have agents that do the things that we want them to do. I do think we're going to end up in a world with lots of complex, interacting, super intelligent beings. And I'm optimistic that that will lead to a kind of a balance of actors, some of which are supporting the flourishing of humanity and some of which may not be. but the ones which do support the flourishing of humanity will end up kind of supporting us and protecting us against the ones which aren't acting on our best interests. So that the world I think we can work towards Well given your super forecasting abilities I happy to hear that What do you think about all the conversation around pacing the frontier I actually very much supportive of the idea that there should be pacing of what happens for the frontier of AI in

the real world. At the same time, I think we should move as fast as possible in simulation. now the reason is that i think in simulation we can actually develop a lot of things like the science of super intelligence that will actually help us to find solutions that help things to go well in the real world and i think that kind of work should happen as fast as possible to make sure that when things really happen in the real world that things have the best possible opportunity to go well and lead to human flourishing so i think that would be my thought but yeah in general I'm supportive of the suggestions, like the kind of direction that people like Demis and Daria and so forth are suggesting. I think it's generally a good idea. Maybe part of that is that I think that things are moving really quickly and I do believe that when we accelerate AI and get to superintelligence, it will lead to a really great future for humanity.

But I also think that it's something that the speed at which things are moving is not something that we are capable of adapting to that fast. So I think just by pacing a little bit, I think it gives us the opportunity, A, for our governments and our organizations and our emotional responses to adapt to the changes that are coming. And also for this work to continue a pace of really figuring out in simulation how to, you know, the real science of doing this right. So accelerate in simulation so we can understand what's happening and do science around it and then be careful in the real world. Yeah, I think that would be my summary. I'd love to transition to talking a little bit about compute. This is going to be an incredibly compute-expensive journey, and you are pushing the limits of what compute can do. I think both Google and NVIDIA have been great partners for you. can you talk a little bit about your cluster?

You're not simply renting spot compute from NeoClouds like many other NeoLabs. You're actually doing some really serious work on your cluster and the different other forms of compute that you have. Can you just talk about your compute setup? Right. So, yeah, as I mentioned earlier, we are very fortunate with the compute that we have. We are the first customers of the largest open cluster on GCP by the end of the year. well actually started like one month now so we are like extremely fortunate in that and I actually feel quite bad about like the startups that are like getting compute now because it's becoming like harder and harder but yeah so we have like this like very open cluster which is like based on GPUs right like NVIDIA GPUs which is really fortunate for us because it means that we can easily do simulation and like training and inference right they are quite like general chips So that's amazing. I think the Rubens specifically, they have this architecture where one rack

can basically be considered one machine. There's a very high bandwidth between the nodes in one rack, which is very important for reinforcement learning where you have really fast iteration cycles compared to LLMs that are, relatively speaking, quite slow. So that's quite fortunate for us. I think some of the things we did eight, ten years ago at Google where we basically had like abundant CPUs at Google that doesn't scale anymore. So we really need to make sure that everything we do is on the GPUs and we don't need like millions of CPUs beyond these GPUs. So our simulations can run on the GPU inference and training and so on. So we have some good architectures for that. I think like the LM companies now have kind of like the problems that we had many years ago where they need like sandboxes running on CPUs, like simulating computers and so on. And we're trying to not put ourselves in that situation. All the compute, is it roughly fungible between the different things that you're doing?

I imagine you're building crucibles and then you're kind of doing these big rollouts where the agents are actually learning. Is that the rough buckets of what compute is being used for? Yeah, you could split it up into training inference and simulation. Is the same compute equally good at training, inference, and simulation? No. Okay. So I think it's pretty clear what has happened with GPUs over the last many years, or since deep learning took off, is that they've been more and more tailors towards training. And maybe in the last few years, also inference, but less about simulation. But we are like similarly to the LLM companies, we also want to be in the case where like scaling laws matters like we know that roughly like if you add this amount of compute and experiences then we get a certain amount of quality out of that. So I guess if GPUs aren't great for

scaling simulation, then what's the solution? So there are many different solutions here, right? So there's like model-based reinforcement learning where the model becomes much more expensive compared to the simulation. So there's many different tricks you can do. There's various approaches for replay buffers or reusing data many times. So yeah, there are ways around that. And other than that, we do spend quite some effort on just making the simulations as fast as possible and make the best use of the GPUs. Are you seeing scaling laws yet, or is it too early to say? So for reinforcement learning, like scaling laws are like slightly different. So like scaling laws is the component of how much compute you put in, how much data you put in, and like the model size. But reinforcement learning is a little bit different because the problem is like non-stationary. So as you progress in the problem, the data changes, the data distribution changes. So they are a little bit different. It's like a very interesting research direction.

I understand the difference, but are you seeing some sort of scaling law where more compute in gets you closer to achieving the goals, or is it hard to say right now? Yeah, but it's something we are actively pushing towards. And especially as we are transitioning from a smaller cluster to the bigger cluster, we want to be ready for that. Yeah. Let's say at the beginning, when David introduced you, he mentioned that you had done some very foundational work to compute algorithms for reinforcement learning, upon which the field really relies on today. Maybe talk about some of that work, and in particular, your work around Impala and Seed. What are those, and how do they relate to the work that you're doing at Ineffable? Right. So when I joined the deep learning team at DeepMind, most people were running like single machine CPU-based reinforcement learning. So the scale was like quite small. So what I did together with colleagues was to make an agent, reinforcement learning agent that really utilized GPUs

that could scale like linear with compute. I think like partially because it was very efficient. It was very good at scaling. and like the third thing was it was very easy to extend and like modify and do whatever you wanted with it and contributed to it being like the most popular reinforcement agent at the company and I think at one point most our team used it or like derivatives of it and so that was quite successful and quite a nice project to work on then like later on like we we got tvus right we got like racks of these TVUs, so we need to scale beyond that. So some of my work at Google Brain, after I left DeepMind to found the Google Brain Amsterdam office, was to scale to like racks of TVUs and that product was called like Seed RL. And so that was like really about like utilizing the TVUs both for like inference and training in the best possible way while keeping the environments as flexible as you wanted to.

It could be like anything from go to very expensive physics simulations. And how's the thing used at Ineffable? Yeah, so, well, we actually implemented a few of those algorithms. Okay. Also, like derivatives of these algorithms. So that is like what we are using today. We have, like, depending on, like, the environments, we use, like, different algorithms, right? There are, like, pros and cons between them. So yeah, we talked to Alec, we have some very simple, fast to simulate environments and also some that are more expensive. So for those who are working on large scale RL architectures today and haven't read those original papers, it's definitely worth a look because I think you'll discover that many of the ideas that are currently in use, like kind of trace back to those early work that Latter kind of did and figured out, which was, yeah, a little bit prescient, I would say. I think it's actually kind of fun to see the reinforcement learning space now

versus like 10 years ago, or like, let's say like eight years ago, maybe. Like the LLM community is also using reinforcement learning in particular the last two years. So, and like a lot of the things that they're using our being rediscovered approaches that we used many years ago. So it's kind of interesting to see, with a lot of novelty as well on top of it, right? But yeah. Can you talk a little bit about your team? How big is the team right now? We have intentionally kept the team relatively small. So we're not in a hurry to build a giant organization. I think you can do an awful lot nowadays with a small number of incredibly talented people. So we have hired 20 people in total. And yeah, the people we've hired, it's about finding some of the exceptionally talented people. But it's not just about talent for us. It's also about bringing together people who are kind, collaborative human beings. That's really important to the kind of culture we're trying to build.

and people who are really well aligned with the kind of mission that we're following and people that sort of fit the needs that we have. So we try and build the team with the right set of skills as well. So all of those four things are really important. And, yeah, we kind of have very high standards and very high expectations of the people that come in. So we're very lucky to get some amazing, really amazing people. And so just to tie it to the compute as well, right? I like this relatively small amount of people so far means that we have a lot of compute per person, right? And these days you can do a lot of experiments just like one person. So it's really also to keep the compute to person ratio quite high. So we're not like fighting for resources from day one. I think three of the co-founders worked together on AlphaStar. Many of you did work together. Are you assembling the Avengers from, especially given the recent changes from DeepMind,

are you assembling the Avengers of folks that you've worked with in the past? Have you hired new people? And if so, kind of what are the attributes that make somebody new to you likely to be a fit and an applicable? So it's a mixture. So, I mean, AlphaStar was an exceptional team. It was sort of intentionally assembled within DeepMind by choosing the kind of best of the best to work on that project. So in some sense, it's not a surprise that I think actually four of our co-founders come from that team. But we are not exclusively drawing from that team or indeed from DeepMind. I mean, the other three founders all came from very successful startups. and we hire people who are brilliant but also kind and nice people to have around who are going to work together well who are aligned with the mission and who you know fit the things that we

that we need so but in terms of finding those people we put a lot of effort into getting the right people so as uncomfortable as it is I've had to say no to a lot of people I've worked with in the past just because our bar is so super high you know i've worked i've been fortunate to work with like some amazing people and so you know i've been choosing amongst uh like a a very exceptional prop and bringing in yeah what i consider to be a yeah superstar team um and so yeah it's it's um it's it's not easy to get into ineffable but um the people who we have are wonderful human beings and, yeah, very, very proud of the people we brought in. I can see why people are drawn to working at Enoughable, especially at a time when, you know, the biggest clusters, the biggest labs are kind of, you know, they're locked in this prisoner's dilemma of race and, you know, there's a lot of dollars at stake and so, you know, decisions get made differently.

I can imagine for a set of people that, especially that really, really deeply believes in the scientific mission of what you want to accomplish. This must feel like a breath of fresh air. Yeah, I think that's right. And we've tried to establish a research-led culture. So it's not a company that's led by building an incremental next product V17.1 on the thing that's come before. It's a company that's built on researchers pursuing their passion and trying to produce the breakthroughs that they believe are possible. and we center the company around that from the beginning and from the outset to create like a haven for research you know where where we push back on on on product pressure because we believe that if we make these breakthroughs it will lead to you know far bigger products downstream far bigger successes than if we kind of start off from day one on like you know oh you know here's here's the product we're doing from the beginning so yeah so it's research led we're creating um

I think a really special environment and yeah people are loving it I'm loving it Wonderful I'd love to close with some predictions maybe especially since you're such a super forecaster what's your guess for when when you'll have the first strong evidence that the approach is working and then what's your forecast for when we'll achieve the mission that you've set out to achieve this first contact with super intelligence? Yeah, so these are fair questions. So there are going to be some reasonable bars of uncertainty on this. Yeah, but you're being recorded, so we're going to hold you to it. Exactly, yeah. So I think we will see some extremely strong evidence that we're on the right track within a year. And I think we will achieve something which is a true super intelligence, according to the definition I gave earlier of something which discovers things which are far beyond what humans can achieve

across everything from science, mathematics technology, languages, etc. I originally thought within 3-5 years my timelines have accelerated a little bit, so I'm going to say 3 years because of what can be done with coding agents to accelerate things. I think things are going to move fast Buckle up Give me a journey Wonderful Let's say David This is a wonderful Wonderful conversation I'm so inspired By what you're building And thank you For taking the time To sit down And share with our audience What you're building today Thank you Sonia Thank you It's a pleasure Thank you.