What is Digital Life? with OpenAI Co-Founder & Chief Scientist Ilya Sutskever

0:06OpenAI, a company that we all know now, but only a year ago was 100 people, is changing the world. Their research is leading the charge to AGI. Since ChatGPT captured consumer attention last November, they show no signs of slowing down. This week, Ilan and I sit down with Ilya Sutskiver, co-founder and chief scientist at OpenAI, to discuss the state of AI research, where we'll hit limits, the future of AGI, and what's going to take to reach super alignment. Ilya, welcome to No Priors. Thank you. It's good to be here. Let's start at the beginning. pre AlexNet, nothing in deep learning was really working.

0:41And then given that environment, you guys took a very unique bet. What motivated you to go in this direction? Indeed. In those dark ages, AI was not an area where people had hope and people were not accustomed to any kind of success at all. And because there wasn't, there hasn't been any success, there was a lot of debate and there were different schools of thoughts that had different arguments about how machine learning and AI should be. And you had people who were into knowledge representation from a good old fashioned AI. You had people who were Bayesians and they liked Bayesian non-parametric methods.

1:23You had people who like graphical models and you had the people who like neural networks. Those people were marginalized because neural networks did not have the property that you can't prove math theorems about them. about something, it means that your research isn't good. That's how it has been. But the reason why I gravitated to neural networks from the beginning is because it felt like those are small little brains. And who cares if you can prove any theorems about them, because we are training small little brains, and maybe they'll become, maybe they'll do something one day. And the reason that we were able to do AlexNet when we did is because a combination of two factors, three factors.

2:04The first factor is that this was shortly after GPUs started to be used in machine learning. People kind of had an intuition that that's a good thing to do, but it wasn't like today where people exactly knew what they need GPUs for. It was like, oh, let's play with those cool, fast computers and see what we can do with them. It was an especially good fit for neural networks. So that was a very, that definitely helped us. I was very fortunate in that I was able to realize that the reason neural networks of the time weren't good is because they were too small. So like if you try to solve a vision task with a neural network, which has like a thousand neurons, what can it do?

2:44It can't do anything. It doesn't matter how good your learning is and everything else. But if you have a much larger neural network, you'll do something unprecedented. What gave you the intuition to think that that was the case? Because I think at the time, it was reasonably contrarian to think that despite to your point, you know, a lot of the human brain in some sense works that way or different, you know, biological neural circuits. But I'm just curious, like what gave you that intuition early on to think that this was a good direction? I think, yeah, looking at the brain and specifically the, if you, like all those things follow very easily.

3:18If you allow yourself, if you allow yourself to accept the idea, right now, this idea is reasonably well accepted. back then people still talked about it but they haven't really accepted it or internalized the idea that maybe an artificial neuron in some sense is not that different from a biological neuron so now whatever you imagine animals do with their brains you could perhaps assemble some artificial neural network of similar size maybe if you train it it will do something similar so there so that leads to that so that leads you to start to imagine okay like you almost imagine the computation being done by the neural network you can almost think like if you have a high resolution image and you have like one neuron for like a large group of pixels what can the neuron do it's just not much it can do but if you have a lot of neurons then they can actually do something and compute something so I think it was like it was considerations like this plus a technical realization The technical realization is that if you have a large training set that specifies the behavior of the neural network and the training set is large enough such that it can constrain the large neural network sufficiently.

4:37And furthermore, if you have the algorithm to find that neural network, because what we do is that we turn the training set into a neural network, which satisfies a training set. Neural network training can almost be seen as solving a neural equation. Solving a neural equation where every data point is an equation and every parameter is a variable. And so it was multiple things. The realization that a bigger neural network could do something unprecedented. The realization that if you have a large data set together with the compute to solve the neural equation, that's what gradient descent comes in.

5:23But it's not gradient descent. Gradient descent was around for a long time. It was certain technical insights about how to make it work. Because back then, the prevailing belief was, well, you can't train those neural nets anything. It's all hopeless. So it wasn't just about the size. It was about, even if someone did think, gosh, it would be cool to train a big neural net, they didn't have the technical ability to turn this idea into reality. You needed not only to code the neural net, you need to do a bunch of things right. And only then it will work. And then another fortunate thing is that the person whom I work with, Alex Krzyzewski, he just discovered that he really loves GPUs.

6:00And he was perhaps one of the first person who really mastered writing really, like really performing code for the GPUs. And that's why we were able to squeeze a lot of performance out of two GPUs and do something and produce something unprecedented. So to sum up, it was multiple things. The idea that a big neural network, in this case, a vision neural network, a convolutional neural network with many layers, one that's much, much bigger than anything that's ever been done before could do something very unprecedented because the brain can see and the brain is a large neural network and we can see quickly so our neurons don't have a lot of time.

6:38Then the compute needed, the technical know-how that in fact we can train such neural networks and it was not at all widely distributed. Most people in machine learning would not have been able to train such a neural network even if they wanted to. Did you guys have any particular goal from a size perspective? Or was it just as, you know, and if that's biologically inspired or where that number comes from or just as large as we can go? Definitely as large as we can go. Because keep in mind, I mean, we had a certain amount of compute which we could usefully consume. And then what can it do? Maybe if we think about just like the origin of OpenAI and the goals of the organization, like what was the original goal and how's that evolved over time?

7:24The goal did not evolve over time. The tactic evolved over time. So the goal of OpenAI from the very beginning has been to make sure that artificial general intelligence, by which we mean autonomous systems, AI that can actually do most of the jobs and activities and tasks that people do, benefits all of humanity. That was the goal from the beginning. The initial thinking has been that maybe the best way to do it is by just open sourcing a lot of technology. We later, and we also attempted to do it as a nonprofit, seemed very sensible. This is the goal. Nonprofit is the way to do it. What changed?

8:12At some point at OpenAI, we realized, and we were perhaps among the earliest to realize that to make progress in AI for real, you need a lot of compute. Now, what does a lot mean? The appetite for compute is truly endless as now clearly seen, but we realized that we will need a lot. and a non-profit wouldn't be the way to get there, wouldn't be able to build a large cluster with a non-profit. That's where we became, we converted into this unusual structure called cap profit. And to my knowledge, we are the only cap profit company in the world. But the idea is that investors put in some money, but even if the company does incredibly well, they don't get more than some multiplier on top of their original investment.

9:03And the reason to do this, the reason why that makes sense, you know, there are arguments, one could make arguments against it as well. But the argument for it is that if you believe that the technology that we are building, AGI, could potentially be so capable as to do every single task that people do, does it mean that it might unemploy everyone? Well, I don't know, but it's not impossible. And if that's the case, it makes sense. It will make a lot of sense if the company that built such a technology would not be able to make infinite, would not be incentivized rather to make infinite profits.

9:46I don't know if it will literally play out this way because of competition in AI. So there will be multiple companies. And I think that will have some unforeseen implications on the argument which I'm making. But that was the thinking. I remember visiting the offices back when you were, I think, housed at YC or something, or, you know, cohabited some space there. And at the time, there was a suite of different efforts. There was robotic arms that were being manipulated. And then there was, you know, some video game related work, which was really cutting edge. How did you think about how the research agenda evolved and what really drove it down this path of transformer-based models and other forms of learning?

10:24So our thinking has been evolving over the years from when we started OpenAI. In the first year, we indeed did some of the more conventional machine learning work. By conventional machine learning work, I mean, because the world has changed so much, a lot of things which were known to everyone in 2016 or 2017 are completely and utterly forgotten. It's like the Stone Age almost. So in that Stone Age, the world of machine learning looked very different. it was dramatically more academic. The goals, values, and objectives were much more academic. They were about discovering small bits of knowledge and sharing them with the other researchers and getting scientific recognition as a result.

11:09And it's a very valid goal and it's very understandable. I've been doing AI for 20 years now. More than half of my time that I spent in AI was in that framework. And so what do you do? You write papers, you share your small discoveries two realizations. The first realization is just at a high level, it doesn't seem like it's the way to go for a dramatic impact. And why is that? Because if you imagine how an AGI should look like, it has to be some kind of a big engineering project that's using a lot of compute, right? Even if you don't know how to build it, what that should look like, you know that this is the ideal you want to strive towards.

11:47So you want to somehow move towards larger projects as opposed to small projects. So while we attempted a first large project where we trained a neural network to play a real-time strategy game as well as the best humans. It's the Dota 2 project and it was driven by two people, Jakob Pachocki and Greg Brockman. They really drove this project and made it a success. And this was our first attempt at a large project. But it wasn't quite the right formula for us because the neural networks were a little bit too small. It was just a narrow domain, just a game. I mean, it's cool to play a game. And we kept looking.

12:27And at some point, we realized that, hey, if you train a large neural network, a very, very large transformer to predict text better and better, something very surprising will happen. This realization also arrived a little bit gradually. We were exploring generative models. We were exploring ideas around next word prediction. Those are ideas also related to compression. We were exploring them. The transformer came out. We got really excited. We were like, this is the greatest thing. We're going to do transformers now. It's clearly superior than anything else before it. We started doing transformers.

13:01We did GPT-1. GPT-1 started to show very interesting signs of life. And that led us to doing GPT-2. And then ultimately GPT-3. GPT-3 really opened everyone else's eyes as well to, hey, this thing has a lot of traction. There is one specific formula right now that everyone is doing. And this formula is train a larger and larger transformer on more and more data. I mean, for me, the big wake up moment to your point was GPT-2 to GPT-3 transition, where you saw such a big step function and capabilities. and then obviously with four open AIs published some really interesting research around some of the different domains of knowledge or domains of expertise or chain of thought or other things that the models can suddenly do in an emergent form what was the most surprising thing for you in terms of emergent behavior in these models over time you know it's very hard to answer that question it's very hard to answer because I'm too close and I've seen it progress every step of the way so as much as I'd like I find it very hard to answer that question i think if i had to pick one i think maybe the most surprising thing for me is the whole thing works at all you know it's hard it's and i'm not sure i know how to convey this what i have in mind here because if you see a lot of neural networks do amazing things well obviously neural networks is the thing that works but i have witnessed personally what it's like to be in a world for many years where the neural networks don't work at all.

14:34And then to contrast that to where we are today, just the fact that they work and they do these amazing things, I think maybe the most surprising, if I had to pick one, it would be the fact that when I speak to it, I feel understood. Yeah, there's a really good saying from, I'm trying to remember, maybe it's Arthur Clarke or one of the sci-fi authors, which is effectively, it says advanced technology is sometimes indistinguishable from magic. Yeah, I'm fully in this camp. Yeah, it definitely feels like there's some magical moments with some of these models now. Is there a way that you guys decide internally, given all of the different capabilities you could pursue, how to continually choose the set of big projects?

15:19You've sort of described that centralization and committing to certain research directions at scale is really important to OpenAI success. given the breadth of opportunity now, like what's the process for deciding what's worth working on? I mean, I think there is some combination of bottom-up and top-down where we have some top-down ideas that we believe should work, but we're not 100 % sure. So we still, we need to have good top-down ideas. And there is a lot of bottom-up exploration guided by those top-down ideas as well. And their combination is what informs us as to what to do next. And if you think about those bottom, I mean, either direction, top-down or bottom-up ideas, like clearly we have this dominant continue-to-scale transformers direction.

16:06Do you explore additional architectural directions or is that just not relevant? It's certainly possible that various improvements can be found. I think improvements can be found in all kinds of places, both small improvements and large improvements. I think the way to think about it is that while the current thing that's being done keeps getting better as you keep on increasing the amount of compute and data that you put into it. So we have that property. The bigger you make it, the better it gets. it is also the property that different things get better by different amounts as you keep on improving as you keep on scaling them up so not only you want to of course scale up what we are doing we also want to keep scaling up the best thing possible what is uh a i mean you you probably don't need to predict because you can see internally what do you think is improving most from a capability perspective in the current generation of scale?

17:06The best way for me to answer this question would be to point to the models that are publicly available. And you can see how they compare from this year to last year. And the difference is quite significant. I'm not talking about the difference between, not only the difference between, let's say you can look at the difference between GPT-3 and GPT-3.5, and then chat GPT, chat GPT-4, chat GPT-4 with vision. And you can just see for yourself. It's easy to forget where things used to be, but certainly the big way in which things are changing is that these models become more and more reliable. Before, they were only very partly there.

17:51Right now, they are mostly there, but there are still gaps. And in the future, perhaps, these models will be there even more. We could trust their answers. They'll be more reliable. They'll be able to do more tasks in general across the board. And then another thing that they will do is that they'll have deeper insight. As we train them, they gain more and more insight into the true nature of the human world. And their insight will continue to deepen. I was just going to ask about how that relates to sort of model scale over time, because a lot of people are really stricken by the capabilities of the very large scale.

18:28models and the emergent behavior in terms of understanding of the world. And then in parallel, as people incorporate some of these things into products, which is a very different type of path, they often start worrying about inference costs going up with the scale of the model, and therefore they're looking for smaller models that are fine-tuned. But then, of course, you may lose some of the capabilities around some of the insights and ability to reason. And so I was curious in your thinking in terms of how all this evolves over the coming years. I would actually point out that the main thing that's lost when you switch to the smaller models is reliability.

18:59I would argue that at this point, it is reliability that's the biggest bottleneck to these models being truly useful. How are you defining reliability? So it's like when you ask it a question that's not much harder than other questions that the model succeeds at, then you'll have a very high degree of confidence that it will continue to succeed. So I'll give you an example. Let's suppose that I want to learn about some historical thing and I can ask, tell me what is the prevailing opinion about this and about that and I can keep asking questions. And let's suppose it answered 20 of my questions correctly.

19:36I really don't want the 21st question to have a gross mistake. That's what I mean by reliability. Or like, let's suppose I upload some documents, some financial documents. Suppose they say something, I want you to do some analysis and to make some conclusion and I want to take action on this basis and this conclusion. And it's like, it's not a super hard task and these models clearly succeed on this task most of the time. But because they don't succeed all the time and if it's a consequential decision, I actually can't trust the model any of those times and I have to verify the answer somehow.

20:07So that's how I define reliability. It's very similar to the self-driving situation, right? If you have a self-driving car and it's like, does things mostly well, that's not good enough. Situation is not as extreme as with a self-driving car, but that's what I mean by reliability. My perception of reliability is that, To your point, it goes up with model scale, but also it goes up if you fine-tune for specific use cases or instances or data sets. And so there is that trade-off in terms of size versus specialized fine-tuning versus reliability. So certainly people who care about some specific application have every incentive to get the smallest model working well enough.

20:49I think that's true. It's undeniable. I think anyone who cares about a specific application will want the smallest model for it. That's self-evident. I do think, though, that as models continue to get larger and better, then they will unlock new and unprecedentedly valuable applications. So, yeah, the small models will have their niche for the less interesting applications, which are still very useful. And then the bigger models will be delivering on applications. Okay, let's pick an example. So consider the task of producing good legal advice. It's really valuable if you can really trust the answer.

21:25Maybe you need a much bigger model for it, but it justifies the cost. There's been a lot of investment this year at the 7B in particular, but 7B, 13B, 34B sizes. Do you think continued research at those scales is wasted? No, of course not. I mean, I think that in the kind of medium term, medium term by high timescale anyway, there will be an ecosystem. There will be different uses for different model sizes. There will be plenty of people who are very excited for whom the best 7B model is good enough. They'll be very happy with it. And then there'll be plenty of very, very exciting and amazing applications for which it won't be enough.

22:13I think that's all. I mean, I think the big models will be better than the small models, but not all applications will justify the cost of a large model. What do you think the role of open source is in this ecosystem? Well, open source is complicated. I'll describe to you my mental picture. I think that in the near term, open source is just helping companies produce useful. Let's see. why would one want to have an open source to use an open source model instead of a closed source model that's hosted by some other company I mean I think it's very valid to want to be the final decider on the exact way in which you want your model to be used and for you to make the decision of exactly how you want the model to be used and which use case you wish to support and I think there's going to be a lot of demand for open source models And I think there will be quite a few companies that will use them.

23:12And I'd imagine that will be the case in the near term. I would say in the long run, I think the situation with open source models will become more complicated. And I'm not sure what the right answer is there. Right now, it's a little bit difficult to imagine. So we need to put our future hat, maybe futurist hat. It's not too hard to get into a sci-fi mode when you remember that we are talking to computers and they understand us. But so far, these computers, these models are actually not very competent. They can't do tasks at all. I do think that there will come a day where the level of capability of models will be very high.

23:51Like in the end of the day, intelligence is power, right? Right now, these models, their main impact, I would say, at least popular impact, is primarily around entertainment and simple questions. So you talk to a model, wow, this is so cool. you produce some images, you had a conversation, maybe you had some questions, good answer. But it's very different from completing some large and complicated task like, what about if you had a model which could autonomously start and build a large tech company? I think if these models were open source, they would have it difficult to predict consequence. Like we are quite far from these models right now.

24:31And by quite far, I mean by high timescale, but still, this is not what you're talking about but the day will come when you have models which can do science autonomously like deliver on big science projects and it becomes more complicated as to whether it is desirable that models of such power should be open sourced I think the argument there is a lot less clear cut, a lot less straightforward compared to the current level models which are very useful and I think it's fantastic that the current level models have been built. So that is maybe, maybe I answered a slightly bigger question rather than what is the role of open source models?

25:11What's the deal with open source? And the deal is up to a certain capability. It's great, but not difficult to imagine models sufficiently powerful, which will be built, where it becomes a lot less obvious as to the benefits of their open source. Is there a signal for you that we've reached that level or that we're approaching it? Like what's the boundary? So I think figuring out this boundary very well is an urgent research project. I think one of the things that help is that the closed source models are more capable than open source models. So the closed source models could be studied and so on.

25:54and so you'd have some experience with a generation of closed source model and then you know like, oh, these models capabilities, it's fine. There's no big deal there. Then in like a couple of years, the open source models catch up and maybe a day will come when we're going to say, well, like these closed source models, they're getting a little too drastic and then some other approaches needed. If we have our, you know, future hat on, maybe let's like think about like a several year timeline. What are the limits you see, if any, in the near term in scaling? Is it like data, token scarcity, cost of compute, architectural issues?

26:32So the most near term limit to scaling is obviously data. This is well known and some research is required to address it. Without going into the details, I'll just say that the data limit can be overcome. and progress will continue. One question I've heard people debate a little bit is the degree to which the transformer-based models can be applied to sort of the full set of areas that you'd need for AGI. And if you look at the human brain, for example, you do have reasonably specialized systems. They're all neural networks via specialized systems for the visual cortex versus areas of higher thought, areas for empathy or other sort of aspects of everything from personality to processing.

27:16do you think that the transformer architectures are the main thing that will just keep going and get us there or do you think we'll need other architectures over time so i have to i understand precisely what you're saying and i have two answers to this question the first is that in my opinion the best way to think about the question of architecture is not in terms of a binary is it enough but how much effort how much what will be the cost of using this particular architecture like at this point i don't think anyone doubts that the transformer architecture can do amazing things but maybe something else maybe some modification could have some compute efficiency benefits so you also better to think about it in terms of compute efficiency rather than in terms of can it get there at all i think at this point the answer is obviously yes to the question about well what about the human brain and with its brain regions i actually think that the situation there is subtle and deceptive for the following reasons so what i believe you alluded to is the fact that the human brain has known regions it has like it has a speech perception region it has a speech production region it has an image region it has a face region has all these regions and it looks like it's specialized.

28:36But you know what's interesting? Sometimes there are cases where very young children have severe cases of epilepsy at a young age and the only way they figured out how to treat such children is by removing half of their brain. Because it happens at such a young age, these children grow up to be pretty functional adults and they have all the same brain regions but they are somehow compressed onto one hemisphere so maybe some you know information processing efficiency is lost it's a very traumatic thing to experience but somehow all these brain regions rearrange themselves there is another experiment where that which was done maybe 30 or 40 years ago on ferrets so the ferret is a small animal it's a pretty mean experiment they took the optic nerve of the ferret which comes from its eye and attached it to its auditory cortex.

29:30So now the inputs from the eye starts to map to the speech processing area of the brain. And then they recorded different neurons after it had a few days of learning to see. And they found neurons in the auditory cortex, which were very similar to the visual cortex or vice versa. It was either they mapped the eye to the ear, to the auditory cortex, or the ear to the visual cortex. But something like this has happened. These are fairly well-known ideas in AI that the cortex of humans and animals are extremely uniform. And so that further supports the AI, like you just need one big uniform architecture.