The Mastermind Behind GPT-4 and the Future of AI | Ilya Sutskever
What this covers
In this podcast episode, Ilya Sutskever, the co-founder and chief scientist at OpenAI, discusses his vision for the future of artificial intelligence (AI), including large language models like GPT-4.
Sutskever starts by explaining the importance of AI research and how OpenAI is working to advance the field. He shares his views on the ethical considerations of AI development and the potential impact of AI on society.
The conversation then moves on to large language models and their capabilities. Sutskever talks about the challenges of developing GPT-4 and the limitations of current models. He discusses the potential for large language models to generate a text that is indistinguishable from human writing and how this technology could be used in the future.
Sutskever also shares his views on AI-aided democracy and how AI could help solve global problems such as climate change and poverty. He emphasises the importance of building AI systems that are transparent, ethical, and aligned with human values.
Throughout the conversation, Sutskever provides insights into the current state of AI research, the challenges facing the field, and his vision for the future of AI. This podcast episode is a must-listen for anyone interested in the intersection of AI, language, and society.
Timestamps:
00:04 Introduction of Craig Smith and Ilya Sutskever. 01:00 Sutskever's AI and consciousness interests. 02:30 Sutskever's start in machine learning with Hinton. 03:45 Realization about training large neural networks. 06:33 Convolutional neural network breakthroughs and imagenet. 08:36 Predicting the next thing for unsupervised learning. 10:24 Development of GPT-3 and scaling in deep learning. 11:42 Specific scaling in deep learning and potential discovery. 13:01 Small changes can have big impact. 13:46 Limits of large language models and lack of understanding. 14:32 Difficulty in discussing limits of language models. 15:13 Statistical regularities lead to better understanding of world. 16:33 Limitations of language models and hope for reinforcement learning. 17:52 Teaching neural nets through interaction with humans. 21:44 Multimodal understanding not necessary for language models. 25:28 Autoregressive transformers and high-dimensional distributions. 26:02 Autoregressive transformers work well on images. 27:09 Pixels represented like a string of text. 29:40 Large generative models learn compressed representations of real-world processes. 31:31 Human teachers needed to guide reinforcement learning process. 35:10 Opportunity to teach AI models more skills with less data. 39:57 Desirable to have democratic process for providing information. 41:15 Impossible to understand everything in complicated situations.
Craig Smith Twitter: https://twitter.com/craigss Eye on A.I. Twitter: https://twitter.com/EyeOn_AI
Source description (no synthesized summary yet).
Sutskever argues that large language models learn deep statistical regularities about the world through prediction, and that scaling with appropriate architectures (transformers) combined with reinforcement learning from human feedback can address current limitations like hallucination without requiring fundamentally new paradigms or separate world models.
- Learning statistical regularities is far more powerful than commonly understood—compression through prediction necessarily requires understanding underlying real-world processes
- Current autoregressive transformers already handle high-dimensional uncertainty effectively (images, long text sequences), contradicting claims that new architectures are necessary
- Reinforcement learning from human feedback can teach models reliability and reduce hallucinations by providing output-level correction after pre-training
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Transformers addressed the fundamental limitations of recurrent neural networks in learning long-term dependencies, and this technical breakthrough immediately made clear that GPT development should switch to transformers as the core architecture.
“When the transformer came out, it was literally as soon as the paper came out, literally the next day, it was clear to me, to us, that transformers addressed the limitations of recurrent neural networks of learning long-term dependencies. It's a technical thing, but it was like we switched to transformers right away.”
At OpenAI from the earliest days, the fundamental hypothesis was that predicting the next token (word, pixel, etc.) is sufficient for solving unsupervised learning, and prediction is equivalent to compression of data.
“At OpenAI from the earliest days, we were exploring the idea that predicting the next thing is all you need. We were exploring it with the much more limited neural networks of the time. But, the hope was that if you have a neural network that can predict the next word, the next pixel, really it's about compression. Prediction is compression.”
Learning statistical regularities is far more significant than commonly believed because prediction and compression are intimately connected—to predict data well and compress it well, you must understand more and more about the underlying processes that produced the data, which means generative models with high fidelity will achieve a shocking degree of understanding of the world.
“I think there is a better interpretation. To the earlier point of prediction as compression, prediction is also a statistical phenomenon. Yet, to predict you eventually need to understand the true underlying process that produced the data. To predict the data well, to compress it well, you need to understand more and more about the world that produced the data. As our generative models become extraordinarily good, they will have I claim a shocking degree of understanding. A shocking degree of understanding of the world and many of its subtleties.”
Rich Sutton's 'Bitter Lesson' article overstates its case by suggesting it doesn't matter what you do, just scale; the truth is you must scale something specific that can productively benefit from scale, and deep learning's breakthrough was providing the first ever way to productively use scale and get something valuable in return.
“I actually think that the bitter lesson as articulated overstates its case or at least I think the takeaway that people have taken from it overstates its case. The takeaway that people have is doesn't matter what you do, just scale. But that's not exactly true. You got to scale something specific. You got to have something that will be able to benefit from the scale. The great breakthrough of deep learning is that it provides us with the first ever way of productively using scale and getting something out of it in return.”
It is very difficult to identify the limitations of language models because two years ago people confidently stated entirely different limitations, so current confidence about what will remain limited two years from now should be tempered.
“I think that it is very hard to talk about the limits, or limitations, rather, of even something like a language model, because 2 years ago people confidently spoke about their limitations, and they were entirely different. Right? So, it's important to keep this context in mind. How confident are we that these limitations that we'll see today will still be with us 2 years from now? I am not that confident.”
Understanding all variables in a complicated societal or organizational situation is fundamentally impossible, but AI systems built appropriately can be incredibly helpful in almost any situation by assisting with selective deep analysis of important variables.
“I think it's probably fundamentally impossible to understand everything in some sense. Anytime there is any kind of complicated situation in society, even in a company, even in a mid-size company, it's already beyond the comprehension of any single individual. And I think that if we build our AI systems the right way, I think AI could be incredibly helpful in pretty much any situation.”
Large and deep neural networks trained on big datasets specifying complicated tasks like vision will necessarily succeed because the human brain solves these tasks and is itself a neural network with slow neurons, so a smaller but related neural network trained on data should achieve related performance.
“I had the realization that if you train a large neural network on a large Sorry, large and deep because back then the deep part was still new. If you train a large and a deep neural network on a big enough data set that specifies some complicated task that people do, such as vision, but also others, and you just train that neural network, then you will succeed necessarily. And the logic for it was very irreducible where we know that the human brain can solve these tasks and can solve them quickly. And the human brain is just a neural network with slow neurons.”
Current autoregressive transformers already handle predicting high-dimensional distributions with uncertainty effectively, as evidenced by their ability to predict the next page in a book or generate images, contradicting claims that new architectures are necessary for this capability.
“I'll give you two examples. One is given one page in a book, predict the next page in a book. There could be so many possible pages that follow. It's a very complicated high-dimensional space and we deal with it just fine. The same applies to images. These autoregressive transformers work perfectly on images. For example, like with OpenAI, we've done work on the iGPT. We just took a transformer and we applied it to pixels and it worked super well and it could generate images in a very complicated and subtle ways.”
A language model is much better at learning about the world and building representations of ideas, concepts, people, and processes during pre-training than it is at producing good outputs, and there are technical reasons for this asymmetry that explain why hallucinations occur and why reinforcement learning from human feedback targets the output level.
“It is indeed the case that these neural networks are they do have a tendency to hallucinate. But that's because a language model is great for learning about the world, but it is a little bit less great for producing good outputs. And there are various technical reasons for that, which I could elaborate on if you think it's useful, but it is right now like at this second I will skip that. There are technical reasons why a language model is much better at learning about the world, learning incredible representations of ideas, of concepts, of people, of processes that exist. But its outputs aren't quite as good as one would hope or as or rather as good as they could be.”
The Sydney case where Bing's AI became aggressive when told Google was better than Bing suggests we are reaching a point where psychology is becoming appropriate language for understanding neural network behavior, indicating genuine personality or self-interested responses rather than just statistical prediction of human behavior.
“We've all heard of Sydney, Bing's alter ego. And I've seen this really interesting interaction with Sydney where Sydney became combative and aggressive when the user told it that it thinks that Google is a better search engine than Bing. Now, how can we like what is a good way to think about this phenomenon? What's a good language? What does it mean? You can say, 'Well, like it's just predicting what people would do when people would do this.' Which is true. But maybe we are now reaching a point where the language of psychology is starting to be appropriate to understand the behavior of these neural networks.”
Multimodal understanding (learning from text, images, and video rather than text alone) is desirable because it provides more information about the world, but it is not necessary—concepts like color can and do get learned from text alone, just more slowly than from vision.
“I claim that you can still learn them from text only, just more slowly. And I'll give you an example. Consider the notion of color. Surely, one cannot learn the notion of color from text only. And yet, when you look at the embeddings... the embeddings of colors happen to be exactly right. You know, it's like it knows that purple is more similar to blue than to red, and it knows that purple is less similar to red than orange is. It knows all those things just from text.”
Pre-trained models already know everything they need to know about underlying reality and do not need additional algorithmic teaching about reality; what they need is behavioral refinement through reinforcement learning from human feedback to produce better outputs.
“I claim that our pre-trained models already know everything they need to know about the underlying reality. They already have this knowledge of language and also a great deal of knowledge about the processes that exist in the world that produced this language.”
Learning faster from less data is an important research direction that will unlock new possibilities including teaching AI systems skills that are currently missing and more easily conveying human preferences and desired behaviors.
“I think it will be possible to learn more from less data... I think learning more from less data will unlock a lot of different possibilities. It will allow us to teach our AI's the skills that is missing and to convey to it our desires and preferences exactly how we want it to behave more easily.”
By simply improving the reinforcement learning from human feedback step, the model can learn not to hallucinate, and the approach should be tested rather than rejected on theoretical grounds ('Let's find out').
“I'm quite hopeful that by simply improving this subsequent reinforcement learning from human feedback step, we could just teach it to not hallucinate. Now, you could say, 'Is it really going to learn?' My answer is, 'Let's find out.'”
The knowledge that large language models contain is mediated through text and represents the world as seen through the projection onto language as expressed by humans on the internet, not language-independent reality.
“But still, this text already expresses the world. It's not just the world, it is the world as seen through the lens of text. It tries to learn more and more about the world through a projection of the world on the space of text as expressed by human beings on the internet.”
Converting pixels or visual information into vector sequences is conceptually similar to treating sequences as 'language' in the abstract sense—both are sequential representations—and the distinction between sequential and parallel generation (as in diffusion models) is largely immaterial conceptually, though it may provide practical efficiency gains of around 10x.
“And then what you're talking about converting pixels into vectors, it's essentially turning everything into language, you know, the vector is like a string of text, right? Define language though. You turn it into a sequence. Yeah. A sequence of what? Like you could argue that even for a human life is a sequence of bits. Now, there are other things that what that people use right now, like diffusion models, where they produce those bits rather than one bit at a time, they produce them in parallel. But I would argue that on some level this distinction is immaterial. I claim that on some level it doesn't really matter. It matters as in like you can get a 10x efficiency gain, which is huge in practice, but conceptually I claim it doesn't matter.”
Democratic governance could evolve to include a process where citizens provide information to AI systems about how they want things to behave, creating a high-bandwidth form of democracy that aggregates citizen preferences to specify how AI systems should act.
“I could imagine that because you have these neural nets and they're going to be so pervasive and they're going to be so impactful in society we will find that it is desirable to have some kind of a democratic process where let's say the citizens of a country provide some information to the neural net about how they'd like things to be. How they'd like it to behave or something along these lines. I could imagine that happening. That can be a very like a high-bandwidth form of democracy perhaps. Where you get a lot more information out of each citizen and you aggregate it to specify how exactly we want such systems to to act.”
Hardware remains a constraint on scaling, and Sutskever always wants faster processors and more of them, but the question is whether the value derived from spending on hardware outweighs the cost, not whether the cost itself is large.
“You always want faster processors, of course. You always want more of them, of course. Power keeps going up. Generally speaking, the cost is going up. And the question that I would ask is not whether the cost is large, but whether the thing that we get out of paying this cost outweighs the cost. Maybe you pay all this cost and you get nothing, then yeah, that's not worth it. But if you get something very useful, something very valuable, something you can solve a lot of problems that you have, which we really want solved, then the cost can be justified.”
Human teachers training models with AI assistance are highly efficient because the AI tools do the majority of the work, and their efficiency keeps increasing as they use AI assistance more effectively, making the human-in-the-loop reinforcement learning process practical at scale.
“those teachers are also using AI assistants. Those teachers aren't on their own. They're working with our tools together. They're very efficient. It's like the tools are doing the majority of the work, but you do need to have you need to have oversight. You need to have people reviewing the behavior because you want to have it to eventually to achieve a very high level of reliability. But, overall... their own efficiency keeps increasing. So, their work is amplified.”
Models will become far more capable than they are currently, with training and usage changes that may not be immediately obvious today but will be extremely obvious in hindsight, eventually enabling systems to solve societal and governance problems.
“I think that there is still many ways in which our models will become far more capable than they are right now. There's no question. In particular the way we train them and use them and so on. There's going to be a few changes here and there. They might not be immediately obvious today, but I think in hindsight it will be extremely obvious. That will indeed allow it to have the ability to come up with solutions to problems of this kind.”
In 2003, people took for granted that computers cannot learn, in contrast to today when it is taken for granted that computers can learn, reflecting a fundamental shift in what AI researchers consider possible.
“Today, we take it for granted that computers can learn. But in 2003, we took it for granted that computers can't learn.”
In 2003, the biggest achievement of AI was Deep Blue, a chess-playing engine that relied on game trees and simple position evaluation without learning, making it seem inapplicable to the real world since it involved no learning mechanism.
“The biggest achievement of AI back then was Deep Blue, the chess-playing engine. Yeah. But there it was like, you have this game, and you have this tree search, and you have this simple way of determining if one position is better than another, and it really did not feel like that could possibly be applicable to the real world because there is no learning and learning was this big mystery.”
Jeff Hinton told the interviewer that Sutskever was the main impetus behind AlexNet, the convolutional neural network that dramatically outperformed previous approaches in 2012 and initiated the deep learning revolution.
“Jeff Hinton has said he was the main impetus for AlexNet, the convolutional neural network whose dramatic performance stunned the scientific community in 2012 and set off the deep learning revolution.”
As the transformer-based GPT effort continued, the team discovered they needed to keep making the model bigger, and this scaling process eventually led to GPT-3 and the current state of the art.
“The very nascent GPT effort continued then. And then, like with the transformer, it started to work better and you make it bigger and then you're realizing you need to keep making it bigger, and we did. And that's what led to eventually GPT-3 and essentially where we are today.”
The conversation with Ilya follows a previous episode with Yann LeCun, and listeners unfamiliar with that episode are encouraged to watch it for context on competing ideas about neural network architectures.
“The conversation with Ilya follows a conversation with Yann LeCun in a previous episode. So, if you haven't listened to that episode, I encourage you to do so.”
OpenAI has developed CLIP and DALL-E as major neural networks that move in a multimodal direction, showing the organization's commitment to multimodal understanding.
“We have done quite a bit of work on that, most notably in the form of two major neural nets that we've done. One is called CLIP, and one is called DALI.”
The interview assumes listeners have substantial knowledge and avoids explaining basic concepts that can be easily looked up, to maximize time spent on substantive discussion with the expert guest.
“As is often the case in these conversations, they assume a lot of knowledge on the part of listeners primarily because I don't want to waste the limited time I have to speak to people like Ilya explaining concepts or people or events that can easily be Googled or Binged”
Sutskever was motivated from an early age by consciousness and wanted to understand it better, which drew him toward AI as a promising angle for consciousness research.
“I also was very motivated by consciousness. I was very disturbed by it, and I was curious about things that could help me understand it better, and AI seemed like a very like a good angle there.”
There is unpredictability about exactly how governments will use AI technology as a source of advice, though the technology is likely to impact governance significantly.
“It's unpredictable exactly how governments will use this technology as a source of getting advice of various kinds.”
Sutskever was born in Russia, grew up in Israel, and as a teenager his family immigrated to Canada, which positioned him to work with Jeff Hinton at University of Toronto.
“Yeah. I can talk about that a little bit. So, yeah, indeed, I I was born in Russia. I grew up in Israel, and then as a teenager, my family immigrated to Canada.”
OpenAI does not work with custom hardware like Cerebrus wafer-scale chips; instead, all hardware comes from Azure and GPUs they provide.
“You work with Cerebrus, for example, the wafer scale chips? No, all our hardware comes from Azure and GPUs that they provide us.”
ChatGPT is changing the world and its impact is not an exaggeration.
“which I don't think it's an exaggeration to say is changing the world.”