Ilya Sutskever (OpenAI Chief Scientist) — Why next-token prediction could surpass human intelligence

Dwarkesh Patel: But I would not underestimate the difficulty of alignment of models that are actually smarter than us, of models that are capable of misrepresenting their intentions. Are you worried about spies?

Ilya Sutskever: I'm really not worried about the way it's being leaked.

Dwarkesh Patel: We'll all be able to become more enlightened because we interact with an AGI that will help us see the world more correctly. Like, imagine talking to the best meditation teacher in history.

Ilya Sutskever: Microsoft has been a very, very good partner for us.

Dwarkesh Patel: So I challenge the claim that next-token prediction cannot surpass human performance. If your base neural net is smart enough, you just ask it like, what would a person with great insight and wisdom and capability do?

[Intro music fades]

Dwarkesh Patel: Okay, today I have the pleasure of interviewing Ilya Sutskever, who is the co-founder and chief scientist of OpenAI. Ilya, welcome to the Lunar Society.

Ilya Sutskever: Thank you, happy to be here.

Dwarkesh Patel: First question, and no humility allowed. There are many scientists—or maybe not that many scientists—who will make a big breakthrough in their field. There are far fewer scientists who will make multiple independent breakthroughs that define their field throughout their career. What is the difference? What distinguishes you from other researchers? Why have you been able to make multiple breakthroughs in your field?

Ilya Sutskever: Well, thank you for the kind words. It's hard to answer that question. I mean, I tried really hard. I gave it everything I got, and that worked so far. I think that's all there is to it.

Dwarkesh Patel: Got it. What's the explanation for why there aren't more illicit uses of GPT? Why aren't more foreign governments using it to spread propaganda or scam grandmothers or something?

Ilya Sutskever: I mean, maybe they haven't really gotten to do it a lot. But it also wouldn't surprise me if some of it was going on right now. Certainly, I imagine they'd be taking some of the open source models and try and use them for that purpose. I would expect this would be something they'd be interested in in the future. It's technically possible. They just haven't thought about it enough or haven't done it at scale using their technology. Or maybe it's happening, which is annoying.

Dwarkesh Patel: Would you be able to track it if it was happening?

Ilya Sutskever: I think large scale tracking is possible. Yes. I mean, it requires special operations. It's possible.

Dwarkesh Patel: Now, there's some window in which AI is very economically valuable—on the scale of airplanes, let's say—but we haven't reached AGI yet. How big is that window?

Ilya Sutskever: I think this window—it's hard to give a precise answer, but it's definitely going to be a good multi-year window. It's also a question of definition, because AI before it becomes AGI is going to be increasingly more valuable year after year, in an exponential way. So in some sense, especially in hindsight, it may feel like there was only one year or two years, because those two years were larger than the previous years. But I would say that already last year, there's been a fair amount of economic value produced by AI. Next year is going to be larger, and larger after that. So I think this is going to be a good multi-year chunk of time, from now until AGI, pretty much.

Dwarkesh Patel: Okay. Well, I'm curious if there's a startup that's using your models, right? At some point, if you have AGI, there's only one business in the world—it's OpenAI. How much window do they have, does any business have, where they're actually producing something that AGI can't produce?

Ilya Sutskever: Yeah, well, it's the same question as asking how long until AGI. I think it's a hard question to answer. I hesitate to give you a number, also because there is this thing where people who are optimistic, people who are working on the technology, tend to underestimate the time it takes to get there. But I think the way I ground myself is by thinking about a self-driving car. There is an analogy: if you look at the Tesla and the self-driving behavior of it, it looks like it does everything, but it's also clear that there is still a long way to go in terms of reliability. We might be in a similar place with respect to our models, where it also looks like we can do everything, and at the same time we'll need to do some more work until we really iron out all the issues and make it really good, really reliable, robust, and well-behaved.

Dwarkesh Patel: By 2030, what percent of GDP is AI?

Ilya Sutskever: Oh gosh, hard to answer that question. Very hard to answer.

Dwarkesh Patel: Give me an over-under.

Ilya Sutskever: The problem is that my error bars are in log scale. So I could imagine a huge percentage. I could imagine a disappointingly small percent at the same time.

Dwarkesh Patel: Okay, so let's take the counterfactual where it is a small percentage. Let's say it's 2030 and not that much economic value has been created by these LLMs. As unlikely as you think this might be, what would be your best explanation right now why something like this might happen?

Ilya Sutskever: My best explanation—and I really don't think that's a likely possibility, so that's the preface to the comment—but if I were to take the premise of your question, why were things disappointing in terms of the real world impact? My answer would be reliability. If somehow it ends up being the case that you really want them to be reliable and they ended up not being reliable, or if reliability turns out to be harder than we expect... I really don't think that will be the case, but if I had to pick one, if you told me, "Hey, why didn't things work out?" it would be reliability. That you still have to look over the answers and double check everything, and that just really puts a damper on the economic value that can be produced by those systems.

Dwarkesh Patel: That they'll be technologically mature. It's just a question of whether they'll be reliable enough.

Ilya Sutskever: Yeah, well, in some sense, not reliable means not technologically mature, if you see what I mean.

Dwarkesh Patel: Yeah, fair enough. What's after generative models? Before, you were working on reinforcement learning. Is this basically it? Is this a paradigm that gets us to AGI, or is there something after this?

Ilya Sutskever: I think this paradigm is going to go really, really far, and I would not underestimate it. I think it's quite likely that this exact paradigm is not going to be quite the AGI form factor. I hesitate to say precisely what the next paradigm will be, but I think it will probably involve integration of all the different ideas that came in the past.

Dwarkesh Patel: Is there some specific one you're referring to?

Ilya Sutskever: It's hard to be specific.

Dwarkesh Patel: So you could argue that the next token prediction can only help us match human performance and maybe not surpass it. What would it take to surpass human performance?

Ilya Sutskever: So I challenge the claim that next token prediction cannot surpass human performance. It looks like on the surface it cannot. It looks on the surface that if you just learn to imitate, to predict what people do, it means that you can only copy people. But here is a counter argument for why it might not be quite so. If your base neural net is smart enough, you just ask it like, what would a person with great insight and wisdom and capability do? Maybe such a person doesn't exist, but there's a pretty good chance that the neural net will be able to extrapolate how such a person would behave. Do you see what I mean?

Dwarkesh Patel: Yes, although where would it get that sort of insight about what that person would do, if not from the data of regular people?

Ilya Sutskever: Because if you think about it, what does it mean to predict the next token well enough? What does it mean, actually? It's a much deeper question than it seems. Predicting the next token well means that you understand the underlying reality that led to the creation of that token. It's not statistics—well, it is statistics, but what is statistics? In order to understand those statistics, to compress them, you need to understand what is it about the world that creates those statistics. And so then you say, okay, well, I have all those people. What is it about people that creates their behaviors? Well, they have thoughts and they have feelings and they have ideas and they do things in certain ways. All of those would be deduced from next token prediction. And I'd argue that this should make it possible—not indefinitely, but to a pretty decent degree—to say, can you guess what you'd do if you took a person with this characteristic and that characteristic? Such a person doesn't exist. But because you're so good at predicting the next token, you should still be able to guess what that person would do—this hypothetical, imaginary person with far greater mental ability than the rest of us.

Dwarkesh Patel: When we're doing reinforcement learning on these models, how long before most of the data for the reinforcement learning is coming from AIs and not humans?

Ilya Sutskever: I mean, already most of the data for reinforcement learning is coming from AIs. The humans are being used to train the reward function, but then the reward function in its interaction with the model is automatic, and all the data that's generated during the process of reinforcement learning is created by AI. So if you look at the current technique paradigm, which is getting significant attention because of ChatGPT—reinforcement learning from human feedback—there is human feedback. The human feedback is being used to train the reward function. And then the reward function is being used to create the data which trains the model.

Dwarkesh Patel: Got it. And is there any hope of just removing the human from the loop and having it improve itself in some sort of AlphaGo way?

Ilya Sutskever: Yeah, definitely. I feel like in some sense, our hopes—our plan—very much so. The thing you really want is for the human teachers that teach the AI to collaborate with an AI. You might want to think of it as being in a world where the human teachers do one percent of the work and the AI does 99% of the work. You don't want it to be 100% AI, but you do want it to be a human-machine collaboration which teaches the next machine.

Dwarkesh Patel: So currently, I haven't had a chance to play around with these models. They seem bad at multi-step reasoning, and they have been getting better. But what does it take to really surpass that barrier?

Ilya Sutskever: I think dedicated training will get us there. More improvements to the base models will get us there.

Dwarkesh Patel: Okay, but fundamentally, I also don't feel like they're that bad at multi-step reasoning. I actually think that they are bad at mental multi-step reasoning, but they're not allowed to think out loud. When they are allowed to think out loud, they're quite good. And I expect this to improve significantly, both with better models and with special training.

Dwarkesh Patel: Are you running out of reasoning tokens on the internet? Are there enough of them?

Ilya Sutskever: For context on this question, there are claims that indeed at some point we'll run out of tokens in general to train these models. And yeah, I think this will happen one day. And by the time that happens, we need to have other ways of training models, other ways of productively improving their capabilities and sharpening their behavior, making sure they're doing exactly what we want without more data.

Dwarkesh Patel: You haven't run out of data yet? There's more?

Ilya Sutskever: Yeah, I would say the data situation is still quite good. There are still lots to go, but at some point, data will run out.

Dwarkesh Patel: Okay. Where is the most valuable source of data? Is it Reddit, Twitter, books? What would you trade many other tokens of other varieties for?

Ilya Sutskever: Generally speaking, you'd like tokens which are speaking about smarter things, tokens which are more interesting. So all the sources which you mentioned are valuable. Okay, so maybe not Twitter.

Dwarkesh Patel: Do we need to go multimodal to get more tokens, or do we still have enough text tokens left?

Ilya Sutskever: I think that you can still go very far in text only, but going multimodal seems like a very fruitful direction.

Dwarkesh Patel: If you're comfortable talking about this, where is the place where we haven't scraped the tokens yet?

Ilya Sutskever: I can't answer that question for us, but I'm sure that for everyone, there's a different answer to that question.

Dwarkesh Patel: How many orders of magnitude improvement can we get just not from scale or not from data, but just from algorithmic improvements?

Ilya Sutskever: Hard to answer, but I'm sure there is some.

Dwarkesh Patel: Is "some" a lot or is "some" a little?

Ilya Sutskever: There's only one way to find out.

Dwarkesh Patel: Okay. Let me get your quickfire opinions about these different research directions. Retrieval transformers—just somehow storing the data outside of the model itself and retrieving it somehow. Seems promising. But do you see that as a path forward?

Ilya Sutskever: I think it seems promising.

Dwarkesh Patel: Robotics. Was it the right step for OpenAI to leave that behind?

Ilya Sutskever: Yeah, it was. Back then, it really wasn't possible to continue working in robotics because there was so little data. If you wanted to work in robotics, you needed to become a robotics company. You needed to have a giant group of people working on building robots and maintaining them. And even then, if you're going to have a hundred robots, it's a giant operation already, but you're not going to get that much data. In a world where most of the progress comes from the combination of compute and data—that's where we've been—there was no path to data from robotics. So back in the day, when you made a decision to stop working in robotics, there was no path forward.

Dwarkesh Patel: Is there one now?

Ilya Sutskever: I'd say that now it is possible to create a path forward, but one needs to really commit to the task of robotics. You really need to say, "I'm going to build many thousands, tens of thousands, hundreds of thousands of robots, and somehow collect data from them and find a gradual path where the robots are doing something slightly more useful." Then the data that they get from these robots is used to train the models, and they do something slightly more useful. So you could imagine this kind of gradual path of improvement, where you build more robots, they do more things, you collect more data, and so on. But you really need to be committed to this path. If you say, "I want to make robotics happen," that's what you need to do. I believe that there are companies who are thinking about doing exactly that. But you need to really love robots and need to be really willing to solve all the physical and logistical problems of dealing with them. It's not the same as software at all. So I think one could make progress in robotics today with enough motivation.

Dwarkesh Patel: What ideas are you excited to try but you can't because they don't work well on current hardware?

Ilya Sutskever: I don't think current hardware is a limitation. I think it's just not the case.

Dwarkesh Patel: Got it. So anything you want to try, you can just spin it up?

Ilya Sutskever: Of course. You might say, "I wish current hardware was cheaper," or maybe it would be better if it had higher memory processor bandwidth, let's say. But by and large, hardware is just not a limitation.

Dwarkesh Patel: Let's talk about alignment. Do you think we'll ever have a mathematical definition of alignment?

Ilya Sutskever: A mathematical definition? I think is unlikely.

Dwarkesh Patel: Uh-huh.

Ilya Sutskever: I do think that we will instead have multiple—rather than achieving one mathematical definition, I think we'll achieve multiple definitions that look at alignment from different aspects. I think that this is how we will get the assurance that we want. By which I mean, you can look at the behavior in various tests, in various adversarial stress situations. You can look at how the neural net operates from the inside. I think you have to look at several of these factors at the same time.

Dwarkesh Patel: And how sure do you have to be before you release a model in the wild? Is it 100%, 95%?

Ilya Sutskever: Well, it depends how capable the model is. The more capable the model is, the more confident you need to be.

Dwarkesh Patel: Okay, so just say it's something that's almost AGI. Where is AGI?

Ilya Sutskever: Well, it depends what your AGI can do. Keep in mind that AGI is an ambiguous term. Your average college undergrad is an AGI, right? It's a little—yeah. But you see what I mean. There is significant ambiguity in terms of what is meant by AGI. So depending on where you put this mark, you need to be more or less confident.

Dwarkesh Patel: You mentioned a few of the paths towards alignment earlier. What is the one you think is most promising at this point?

Ilya Sutskever: I think that it will be a combination. I really think that you will not want to have just one approach. I think we will want to have a combination of approaches where you spend a lot of compute to adversarially probe it to find any mismatch between the behavior that you wanted to teach and the behavior that it exhibits. We look inside into the neural net using another neural net to understand how it operates on the inside. I think all of them will be necessary. Every approach like this reduces the probability of misalignment, and you also want to be in a world where your degree of alignment keeps increasing faster than the capability of the models. I would say that right now our understanding of our models is still quite rudimentary. We made some progress, but much more progress is possible. So I would expect that ultimately the thing that will really succeed is when we will have a small neural net that is well understood, that's given the task to study the behavior of a large neural net that is not understood, to verify it.

Dwarkesh Patel: By what point is most of the AI research being done by AI?

Ilya Sutskever: I mean, so today when you use Copilot, right? What fraction? How do you divide it up? So I expect at some point you ask your descendant of ChatGPT, you say, "Hey, I'm thinking about this and this. Can you suggest fruitful ideas I should try?" And you would actually get fruitful ideas. I don't think that will make it possible for you to solve problems you couldn't solve before.

Dwarkesh Patel: Got it. But it's somehow just telling the humans, giving them ideas faster or something. It's not itself interacting with the—

Ilya Sutskever: One example. I mean, you could slice it in a variety of ways. But I think the bottleneck there is good ideas, good insights. And that's something which the neural net could help with.

Dwarkesh Patel: If you designed a billion-dollar prize for some sort of alignment research result or product, what is the concrete criterion you would set for that billion-dollar prize? Is there something that makes sense for such a prize?

Ilya Sutskever: It's funny that you ask this. I was actually thinking about this exact question. I haven't come up with an exact criterion yet. Maybe something with the benefit of hindsight—a prize where we could say that two years later, or three or five years later, we look back and say that was the main result. So rather than having a prize committee that decides right away, we wait for five years and then award it retroactively. But there's no concrete thing we can identify yet, like "you solve this particular problem and you made a lot of progress." I think a lot of progress, yes. I wouldn't say that this would be the full thing.

Dwarkesh Patel: Do you think end-to-end training is the right architecture for bigger and bigger models? Or do we need better ways of just connecting things together?

Ilya Sutskever: I think end-to-end training is very promising. I think connecting things together is very promising. Everything is promising.

Dwarkesh Patel: So OpenAI is projecting revenues of a billion dollars in 2024. That might very well be correct, but I'm just curious: when you're talking about a new general purpose technology, how do you estimate how big a windfall it'll be? Why that particular number?

Ilya Sutskever: You look at the current—we've already had a product for quite a while now, back from the GPT-3 days from two years ago through the API. And we've seen how it grew. We've seen how the response to DALL-E has grown as well. And so you see how the response to ChatGPT is. I think all of this gives us information that allows us to make a relatively sensible extrapolation to 2024. Maybe that would be one answer. You need to have data. You can't come up with those things out of thin air, because otherwise your error bars will be off by—your error bars are going to be like 100x in each direction.

Dwarkesh Patel: But most exponentials don't stay exponential, especially when they get into bigger and bigger quantities, right? So how do you determine in this case that—

Ilya Sutskever: I mean, would you bet against AI?

Dwarkesh Patel: Not after talking with you. Let's talk about what a post-AGI future looks like. Are people like you—I'm guessing you're working 80-hour weeks towards this grand goal that you're really obsessed with. Are you going to be satisfied in a world where you're basically living in an AI retirement home? What are you concretely doing after AGI comes?

Ilya Sutskever: I think the question of what I'll be doing or what people will be doing after AGI comes is a very tricky question. I think where will people find meaning? But I think that's something that AI could help us with. One thing I imagine is that we'll all be able to become more enlightened because we'd interact with an AGI that will help us see the world more correctly and become better on the inside as a result of interacting. Imagine talking to the best meditation teacher in history. I think that would be a helpful thing. But I also think that because the world will change a lot, it will be very hard for people to understand what is happening precisely and how to really contribute. One thing that I think some people will choose to do is to become part AI in order to really expand their minds and understanding and to really be able to solve the hardest problems that society will face then.

Dwarkesh Patel: Are you going to become part AI?

Ilya Sutskever: Very tempting. It is tempting, yeah.

Dwarkesh Patel: Do you think there'll be physically embodied humans in 3000?

Ilya Sutskever: 3000? How do I know? It's going to have any 3000? What does it look like?

Dwarkesh Patel: Are there still humans walking around on earth? Or have you guys thought concretely about what you actually want this world to look like?

Ilya Sutskever: 3000. Well, let me describe to you what I think is not quite right about the question. It implies that we get to decide how we want the world to look like. I don't think that picture is correct. I think change is the only constant. So of course, even after AGI is built, it doesn't mean that the world will be static. The world will continue to change. The world will continue to evolve and it will go through all kinds of transformations. I really have no—I don't think anyone has any idea of how the world will look like in 3000. But I do hope that there will be a lot of descendants of human beings who will live happy, fulfilled lives where they're free to do as they wish, as they see fit, where they are the ones who are solving their own problems. One of the things which I would not want—one world which I would find very unexciting—is one where we build this powerful tool and then the government said, "Okay, the AGI said that society should be run in such a way, and now we should run society in such a way." I'd much rather have a world where people are still free to make their own mistakes and suffer their consequences and gradually evolve morally and progress forward on their own through their own strength. See what I mean? With the AGI providing more like a base safety net.

Dwarkesh Patel: How much time do you spend thinking about these kinds of things versus just doing the research?

Ilya Sutskever: I do think about those things a fair bit, yeah. I think it's a very interesting question.

Dwarkesh Patel: So in what ways have the capabilities we have today surpassed where you expected them to be in 2015? And in what ways are they still not where you would have expected them to be by this point?

Ilya Sutskever: I mean, in fairness, they did surpass what I expected them to be in 2015. In 2015, my thinking was a lot more: I just don't want to bet against deep learning. I want to make the biggest possible bet on deep learning. Don't know how, but I will figure it out.

Dwarkesh Patel: But is there any specific way in which it's been more than you expected or less than you expected? Like some concrete prediction you had in 2015 that's been proven wrong?

Ilya Sutskever: Unfortunately, I don't remember concrete predictions I made in 2015. But I definitely think that overall, in 2015, I just wanted to make the biggest bet possible on deep learning. But I didn't have a specific idea of how far things will go in seven years. In 2015, I did have all these bets with people in 2016, maybe 2017, that things will go really far. But specifics? So it's both the case that it surprised me and I was making these aggressive predictions, but I think maybe I believed them only 50% on the inside.

Dwarkesh Patel: What do you believe now that even most people at OpenAI would find far-fetched?

Ilya Sutskever: I think that because we communicate a lot at OpenAI, people have a pretty good sense of what I think. So we've reached the point at OpenAI where I think we see eye to eye on all these questions.

Dwarkesh Patel: So Google has its custom TPU hardware. It has all this data from all its users—Gmail and so on. Does it give it an advantage in terms of training bigger models and better models than you?

Ilya Sutskever: So I think when the first TPU came out, I was really impressed and I thought, "Wow, this is amazing." But that's because I didn't quite understand hardware back then. What really turned out to be the case is that TPUs and GPUs are almost the same thing. They are very, very similar. A GPU chip is a little bit bigger. A TPU chip is a little bit smaller. It may be a little bit cheaper, but then they make more GPUs than TPUs. So I think the GPUs might be cheaper after all. But fundamentally, you have a big processor and you have a lot of memory. And there is a bottleneck between those two. The problem that both the TPU and the GPU are trying to solve is that by the amount of time it takes you to move one floating point from the memory to the processor, you can do several hundred floating point operations on the processor, which means that you have to do some kind of batch processing. In this sense, both of these architectures are the same. So I really feel like hardware—the only thing that matters about hardware is cost, cost per flop, overall systems cost. And there isn't much that's different. I actually don't know what the TPU costs are, but I would suspect that if anything, the GPUs are more expensive because there are fewer of them.

Dwarkesh Patel: When you're doing your work, how much of the time is spent configuring the right initializations, making sure the training run goes well, and getting the right hyperparameters? And how much is it just coming up with whole new ideas?

Ilya Sutskever: I would say it's a combination. But coming up with whole new ideas is actually a modest part of the work. Certainly, coming up with new ideas is important, but I think even more important is to understand the results, to understand the existing ideas, to understand what's going on. Normally you have this neural net—a very complicated system—and you run it and you get some behavior which is hard to understand. Understanding the results, figuring out what next experiment to run—a lot of the time is spent on that. Understanding what could be wrong, what could have caused the neural net to produce a result which was not expected. I'd say a lot of time is spent as well, of course, coming up with new ideas. But I don't like this framing as much. It's not that it's false, but I think the main activity is actually understanding.

Dwarkesh Patel: What do you see as the difference between the two? In my mind, when you say "come up with new ideas," I think, "Oh, what happened if we did such and such?" Whereas "understanding" is more like, "What is this whole thing? What are the real underlying phenomena that are going on? What are the underlying effects? Why are we doing things this way and not another way?" And of course, this is very adjacent to what can be described as coming up with ideas, but I think the understanding part is where the real action takes place.

Ilya Sutskever: Does that describe your entire career? If you think back on ImageNet or something, was that more new idea or was that more understanding?

Dwarkesh Patel: Oh, it was definitely understanding. It was a new understanding of very old things.

Ilya Sutskever: What is the experience of training on Azure been like? Using Azure?

Dwarkesh Patel: Fantastic. Microsoft has been a very, very good partner for us. They've really helped take Azure and bring it to a point where it's really good for ML. We're super happy with it.

Ilya Sutskever: How vulnerable is the whole AI ecosystem to something that might happen in Taiwan? So let's say there's a tsunami in Taiwan or something. What happens to AI in general?

Ilya Sutskever: Like it's definitely going to be a significant setback. It's not going to be something equivalent to like, no one will be able to get more compute for a few years, but I expect compute will spring up. For example, I believe that Intel has fabs from just a few generations ago. So that means that if Intel wanted to, they could produce something GPU-like from like four years ago. So yeah, it's not the best. I'm actually not sure if my statement about Intel is correct, but I do know that there are fabs outside of Taiwan that are not as good, but you can still use them and still go very far with them. It's just a setback.

Dwarkesh Patel: Will inference get cost prohibitive as these models get bigger and bigger?

Ilya Sutskever: I have a different way of looking at this question. It's not that inference will become cost prohibitive. Inference of better models will indeed become more expensive. But is it prohibitive? Well, it depends on how useful it is. If it is more useful than it is expensive, then it is not prohibitive. To give you an analogy, suppose you want to talk to a lawyer. You have some case, you need some advice or something, you are perfectly happy to spend $500 an hour. So if your neural net could give you really reliable legal advice, you'd say, "I'm happy to spend $400 for that advice." And suddenly inference becomes very much non-prohibitive. The question is, can the neural net produce an answer good enough at this cost? Yes. And you will just have price discrimination, different models of different sizes. It's already the case today. On our product, the API, we serve multiple neural nets of different sizes and different customers use different neural nets depending on their use case. If someone can take a small model and fine tune it and get something that's satisfactory for them, they'll use that. But if someone wants to do something more complicated and more interesting, they'll use the biggest model.

Dwarkesh Patel: How do you prevent these models from just becoming commodities where these different companies just bid each other's prices down until it's basically the cost of the GPU run?

Ilya Sutskever: There is, without question, a force that's trying to create that. And the answer is you got to keep on making progress. You got to keep improving the models. You got to keep on coming up with new ideas and making our models better and more reliable, more trustworthy, so you can trust their answers. All those things.

Dwarkesh Patel: But let's say it's 2025 and the model from 2024 is somebody just offering it at cost and it's still pretty good. Why would people use a new one from 2025 if the one from just a year older is even better?

Ilya Sutskever: There are several answers there. For some use cases, that may be true. There will be a new model from 2025 which will be driving the more interesting use cases. There's also going to be a question of inference costs. You can do research to serve the same model at less cost. So the same model will be served at different costs for different companies. I can also imagine some degree of specialization too, where some companies may try to specialize in some area and be stronger in a narrower area compared to other companies. I think that too may be a response to commoditization to some degree.

Dwarkesh Patel: As over time, do these different companies' research directions converge or do they diverge? Are they doing similar and similar things over time or are they branching off into different areas?

Ilya Sutskever: In the near term, it looks like there is convergence. I expect there's going to be a divergence-convergence behavior where there is a lot of convergence on the near-term work. There's going to be some divergence on the longer-term work. But then once the longer-term work starts to yield fruit, I think there will be convergence again.

Dwarkesh Patel: Got it. When one of them finds the most promising area, everybody just...

Ilya Sutskever: That's right. Now, there is obviously less publishing now, so it will take longer before this promising direction gets rediscovered. But that's how I'd imagine it. Convergence, divergence, convergence.

Dwarkesh Patel: We talked about this a little bit at the beginning, but as foreign governments learn about how capable these models are, are you worried about spies or some sort of attack to get your weights or somehow abuse these models and learn about them?

Ilya Sutskever: It's definitely something that you absolutely can't discount. It's something that we try to guard against to the best of our ability, but it's going to be a problem for everyone who's building this.

Dwarkesh Patel: How do you prevent your weights from leaking?

Ilya Sutskever: Well, you have really good security people. The security people that we have, they've done a really good job so that I'm really not worried about the weights being leaked.

Dwarkesh Patel: What kinds of emergent properties are you expecting from these models at this scale? Is there something that just comes about de novo?

Ilya Sutskever: I'm sure things will come. I'm sure really new surprising properties will come up. I would not be surprised. The thing which I'm really excited about, or the thing which I'd like to see, is reliability and controllability. I think that this will be a very, very important class of emergent properties. If you have reliability and controllability, that helps you solve a lot of problems. Reliability means you can trust the model's output. Controllability means you can control it. It would be very cool if those emergent properties did exist.

Dwarkesh Patel: Is there some way you can predict it in advance? Like what will happen in this parameter count, what will happen in that parameter count?

Ilya Sutskever: I think it's possible to make some predictions about specific capabilities, though it's definitely not simple and you can't do it in a super fine-grained way, at least today. But I think getting better at that is really important, and anyone who has research ideas on how to do that, I think that can be a valuable contribution.

Dwarkesh Patel: How seriously do you take these scaling laws? If there's a paper that says you need this many orders of magnitude more to get all the reasoning out, do you take that seriously or do you think it breaks down at some point?

Ilya Sutskever: The scaling law tells you what happens to your next word prediction accuracy. There is a whole separate challenge of linking next word prediction accuracy to reasoning capability. I do believe that indeed there is a link, but this link is complicated. And we may find that there are other things that can give us more reasoning per unit effort. For example, some special things like reasoning tokens, I think they can be helpful. There can be probably some things.

Dwarkesh Patel: Is this something you're considering, just hiring humans to generate tokens for you, or is it all going to come from stuff that already exists out there?

Ilya Sutskever: I think that relying on people to teach our models to do things, especially to make sure that they are well-behaved and they don't produce false things, is an extremely sensible thing to do.

Dwarkesh Patel: Isn't it odd that we have the data we need at exactly the same time as we have the transformer, at the exact same time that we have these GPUs? Is it odd to you that all these things happen at the same time, or do you not see it that way?

Ilya Sutskever: It is definitely an interesting situation. I will say that it is odd, and it is less odd on some level. Here is why it's less odd. What is the driving force behind the fact that the data exists, that the GPUs exist, that the transformer exists? The data exists because computers became better and cheaper. We've got smaller and smaller transistors, and suddenly at some point it became economical for every person to have a personal computer. Once everyone has a personal computer, you really want to connect them with the network. You get the internet. Once you have the internet, you have suddenly data appearing in great quantities. The GPUs were improving concurrently because you have smaller and smaller transistors and you're looking for things to do with them. Gaming turned out to be a thing that you could do. And then at some point, NVIDIA said, "Wait a sec, maybe we'll turn it into a general purpose GPU computer. Maybe someone will find it useful." Turns out it's good for neural nets. It could have been the case that maybe the GPU would have arrived five years later or ten years later if gaming wasn't the thing. It's kind of hard to imagine what it means if gaming isn't a thing, but maybe there was a counterfactual world where GPUs arrived five years after the data or five years before the data. In which case maybe things wouldn't have been as ready to go as they are now. But that's the picture which I imagine. All this progress in all these dimensions is very intertwined. It's not a coincidence. You don't get to pick and choose which dimensions things improve in.

Dwarkesh Patel: How inevitable is this kind of progress? If you and Geoffrey and a few other pioneers were never born, does the deep learning revolution happen around the same time? How much does it delay?

Ilya Sutskever: I think maybe there would have been some delay, maybe like a year delay.

Dwarkesh Patel: Really? That's it?

Ilya Sutskever: It's really hard to tell. I hesitate to give a longer answer because, okay, well then you'd have GPUs keep on improving. At some point, I cannot see how someone would not have discovered it. Here's the other thing: if no one has done it, computers keep getting faster and better. It becomes easier and easier to train these neural nets because you have bigger GPUs. So it takes less engineering effort to train one. You don't need to optimize your code as much. When the ImageNet dataset came out, it was huge and it was very, very difficult to use. Now imagine you wait for a few years and it becomes very easy to download and people can just tinker. So I would imagine a modest number of years maximum. This would be my guess. I hesitate to give a longer answer, though you can't rerun the world. You don't know.

Dwarkesh Patel: Let's go back to alignment for a second. As somebody who deeply understands these models, what is your intuition of how hard alignment will be?

Ilya Sutskever: I think with the current level of capabilities, we have a pretty good set of ideas of how to align them. But I would not underestimate the difficulty of alignment of models that are actually smarter than us, of models that are capable of misrepresenting their intentions. I think it's something to think about a lot and to research. This is one area where oftentimes academic researchers ask me where the best place they can contribute is, and I think alignment research is one place where academic researchers can make very meaningful contributions.

Dwarkesh Patel: Other than that, do you think academia will come up with more insights about actual capabilities, or is that going to be just the companies at this point?

Ilya Sutskever: The companies will realize the capabilities. I think it's very possible for academic research to come up with those insights. It just doesn't seem to happen that much for some reason, but I don't think there's anything fundamental about academia. It's not like academia can't. I think maybe they're just not thinking about the right problems, or maybe it's just easier to see what needs to be done inside these companies.

Dwarkesh Patel: But there's a possibility that somebody could just realize?

Ilya Sutskever: Yeah, I totally think so. Why would I possibly rule this out?

Dwarkesh Patel: What are the concrete steps by which these language models start actually impacting the world of atoms and not just the world of bits?

Ilya Sutskever: I don't think that there is a clean distinction between the world of bits and the world of atoms. Suppose the neural net tells you that here is something that you should do and it's going to improve your life, but you need to rearrange your apartment in a certain way. And you go and you rearrange your apartment as a result. The neural net impacted the world of atoms.

Dwarkesh Patel: Fair enough. Do you think it'll take a couple of additional breakthroughs as important as the transformer to get to superhuman AI, or do you think we basically got the insights in the books somewhere and we just need to implement them and connect them?

Ilya Sutskever: I don't really see such a big distinction between those two cases. Let me explain why. One of the ways in which progress has taken place in the past is that we've understood that something had a desirable property all along which we didn't realize. Is that a breakthrough? You can say yes it is. Is that an implementation of something on the books? Also yes. My feeling is that a few of those are quite likely to happen, but that in hindsight it will not feel like a breakthrough. Everybody's going to say, "Oh well, of course, it's totally obvious that such and such thing can work." With the transformer, the reason it's being brought up as a big advance is because it was the kind of thing that was not obvious for almost anyone. But if an advance comes from something like the most fundamental advance of deep learning—that a big neural network trained with backpropagation can do a lot of things—where's the novelty? It's not in the neural network. It's not in the backpropagation. But it is most definitely a giant conceptual breakthrough because for the longest time, people just didn't see that. But now that everyone sees it, everyone's going to say, "Well, of course, it's totally obvious. Big neural network. Everyone knows that they can do it."

Dwarkesh Patel: What is your opinion of your former advisor's new forward-forward algorithm?

Ilya Sutskever: I think that it's an attempt to train a neural network without backpropagation. I think that this is especially interesting if you are motivated to try to understand how the brain might be learning its connections. The reason for that is that as far as I know, neuroscientists are really convinced that the brain cannot implement backpropagation because the signals in the synapses only move in one direction. So if you have a neuroscience motivation and you want to say, "Okay, how can I come up with something that tries to approximate the good properties of backpropagation without doing backpropagation?" That's what the forward-forward algorithm is trying to do. But if you are trying to just engineer a good system, there is no reason to not use backpropagation. It's the only algorithm.

Dwarkesh Patel: I've heard you in different contexts talk about the need for using humans as the existing example case that AGI exists. At what point do you take the metaphor less seriously and not feel the need to pursue it in terms of research? Because it is important to you as a sort of existence case. At what point do I stop caring about humans as an existence case of intelligence, or as the example in the model you want to follow in terms of pursuing intelligence in models?

Ilya Sutskever: I think it's good to be inspired by humans. I think it's good to be inspired by the brain. I think there is an art to being inspired by humans and the brain correctly, because it's very easy to latch on to a non-essential quality of humans or of the brain. Many people whose research is trying to be inspired by humans and by the brain often get a little bit specific. People get a little bit too focused on what cognitive science model should we follow. At the same time, consider the idea of the neural network itself, the idea of the artificial neuron. This too is inspired by the brain, but it turned out to be extremely fruitful. So how do you do this? What behaviors of human beings are essential? That you say, "This is something that proves to us that it's possible." What is inessential? That is some emergent phenomenon of something more basic, and we just need to focus on getting our own basics right. I would say that one can and should be inspired by human intelligence with care.

Dwarkesh Patel: Final question. Why is there in your case such a strong correlation between being first to the deep learning revolution and still being one of the top researchers? You would think that these two things wouldn't be that correlated, but why is there that correlation?

Ilya Sutskever: I don't think those things are super correlated. In my case, honestly, it's hard to answer the question. I just kept trying really hard, and it turned out to have sufficed thus far.

Dwarkesh Patel: So it's perseverance.

Ilya Sutskever: I think it's a necessary, but not a sufficient condition. Many things need to come together in order to really figure something out. You need to really go for it and also need to have the right way of looking at things. It's hard to give a really meaningful answer to this question.

Dwarkesh Patel: Ilya, it has been a true pleasure. Thank you so much for coming to the Lunar Society. I appreciate you bringing us to the offices.

Ilya Sutskever: Thank you. Yeah, I really enjoyed it. Thank you very much.

Dwarkesh Patel: Hey, everybody. I hope you enjoyed that episode. Just wanted to let you know that in order to help pay for the bills associated with this podcast, I'm turning on paid subscriptions on my Substack at dwarkeshpatel.com. No important content on this podcast will ever be paywalled, so please don't donate if you have to think twice before buying a cup of coffee. But if you have the means and you've enjoyed this podcast or gotten some kind of value out of it, I would really appreciate your support. As always, the most helpful thing you can do is to share the podcast. Send it to people you think might enjoy it, put it on Twitter, your group chats, etc. Just spreads the world. Appreciate you listening. I'll see you next time. Cheers. Thank you.