What’s Next for AI? OpenAI’s Łukasz Kaiser (Transformer Co-Author)

Thank you very much.

I think there is a lot to unpack there, so I want to go a little slower. There is this thing that's happening in AI, and in AI every week now, a lot is happening: new models, coding, doing slides, self-driving cars, images, videos. It's a nice field that doesn't make you bored for a long time. But through all of this, it's sometimes hard to see the fundamental things that are happening. And fundamentally, if you look at AI progress, it's been a very smooth exponential increase in capabilities.

This is the overarching trend. And there has never been much to make me, at least, and I think my colleagues in the labs, believe that this trend is not happening. It's a little bit like Moore's Law, right? Moore's Law happened through decades and decades, and arguably, you would say it's still very much going on, if not speeding up with the GPUs. But of course, it did not happen as, like, one technology was bringing you there for 40 years. There was one technology, and then another and another and another and another.

And this went on for decades, right? So from the outside, you see a smooth trend, but from the inside, of course, progress is made through new developments in addition to the increase of computer power and better engineering. And so all of these things come together. And in terms of language models, I think there was a big pivotal point. One point was, of course, the Transformers when it started, but the other point was reasoning models. And that happened, I think, o1-preview was about a year and a month ago or something like that.

So we started working on it maybe three years ago, but it's very recent. If you think of it as a paradigm, that's a very recent thing. So it's always like these S-curves, right? It starts, then it gives you amazing growth, and then it flatlines a little bit. We'll get to the pre-training, right? I feel pre-training, in some sense, is on the upper part of the S. Yes, but it's not like scaling laws for pre-training don't work. They totally work.

What scaling laws say is that your loss will log-linearly decrease with your compute. We totally see that. And clearly Google sees that, and all other labs. The problem is how much money do you need to put into that versus the gains you get? And it's just a lot of money, and people are putting it. But with the new paradigm of reasoning, you can get much more gains for the same amount of money because it's on this lower part, and there are just discoveries to be made, and these discoveries unlock insane capabilities.

So it's not like pre-training fizzled out. It's just we found out a new paradigm that, at the same price, gives us much more amazing development. And this paradigm is still very new. It happened so fast. I think if you blink, you may miss it. GPT-3.5, right? GPT-3.5 in ChatGPT, and it would give you answers, and it used no tools, no reasoning. It would answer you something. And now you have ChatGPT, and if you were not into it, you may have blinked and it also gives you answers, and you may say, okay, it's more or less the same, except ChatGPT now will go look on some websites, reason about it, and give you the right answer instead of something it memorized in its weights.

I very much used to like this example of, what time does the SF Zoo open tomorrow? Like, the old ChatGPT would tell you, right? Totally hallucinate from its memory an hour that it read, probably on the zoo's website from five years ago, and it didn't know what's today or tomorrow, so it would just assume it's a weekday. ChatGPT now knows what's today because it's in the system prompt. It goes to the zoo website, reads it, extracts the information.

If it's ambiguous, it probably checks three other websites just to confirm and then gives you the answer. But if you blink, you may think it's the same, but no, it's dramatically better. And as a consequence, since it can read all the websites in the world, it can give you answers in stuff that it wouldn't be able to even touch before. So there is tremendous progress, right? And it happened so fast that it may even be missed. I think one of the biggest things that I would say people kind of know on the inside and others don't is that already right now, it's not about the progress.

There are so many things ChatGPT or Gemini, any LLM, can do for you that people just don't realize. You can take a photo of something broken, ask how to repair it, it may tell you. You can give it college-level homework, and it will do it for you, probably. So that's absolutely amazing.

Well, it just happened. You said Codex, right? Programmers are conservative a little bit. I still use Emacs from time to time. All the coding tools, like, okay, it will complete one line for me, but people are very like, this is my editor, I write code here. Now people are like, no, this is Codex, I ask it to do stuff, I will fix it later, right? But I think it's the recent few months when the transition happened from people using it sometimes, but rarely, to now basically this being how a lot of people work in coding.

That's quite big. I'm not sure everyone's aware of it. But it's also like, if you don't do programming, why would you be aware of it? I do believe, though, that this will come to more and more domains.

Yes, there are a ton of extremely obvious things to fix. A large part of this is just hard to talk about on a podcast because it is in the engineering part. Every lab has their own infrastructure and their own bugs in the code. Machine learning is beautifully forgiving in some sense, in contrast to old software engineering, which would just yell at you when you made a mistake. Our Python code would generally probably run, except much slower and give you worse results if you run it wrong.

So you realize, oh no, it was wrong, and you improve it and the results get better. These are huge distributed computing systems. They're very complex to run. So there is a huge amount to improve and fix and understand in the process about just how to train your model and how to do RL, because RL is more finicky than pre-training. It's harder to do really right. So every day, this is our day-to-day work. On top of that, there is data.

We used to train on just, like, Common Crawl, basically. It's a big repository of the internet that people just scraped without regard for what, and some things came in, some didn't. It was a mess. So now, of course, every larger company has a team that tries to filter this and improve the quality. But it's a lot of work to really extract better data. Now synthetic data is becoming a thing, but when you generate synthetic data, it really matters how you do it, with what model, the whole engineering aspects of everything.

It's such a new domain that it was done somehow, it works, it's beautiful, but there is just so much to do better that people, I don't think, have any doubts that there is a lot there. And on top of that, there are the big things like multimodal. I mean, language models, they are now, as I'm sure you know, and most people realize, actually vision-language models, and they can also do audio. So they're multimodal models to some extent, but the multimodal part still lags behind the text part to a large extent.

So that's one big area where obviously you'd need to do better. And it's not a huge secret how you can do better. There are some methods that maybe will make it even amazingly better, but there are some very simple methods how you can do just better. But this maybe requires retraining your whole base model from scratch, and that takes a few months and it's a huge investment, so we need to organize it. So there is a lot of just work that will undoubtedly make things better.

I think the big question that people have in their mind is how much better will it make them?

So a reasoning model is like your base LLM, but before giving you the answer, it thinks what people call in the chain of thought, meaning it generates some tokens, some text that's meant not for you to read, but for the model to give you a better answer. And while it does this these days, it is also allowed to use tools. So it can, for example, in its so-called thinking process, go and browse the web to give you a better answer. So that's the superficial part of the thinking models.

Now, the deep part is that you start treating this thinking process as part of the model, basically. So it's not something the model generates and it's an output for you. It's something you want to train, right? You want to tell the model, you should think well, you should think so that the answer after this is good in whatever way. And this leads you to a very different way of training the model because models were usually trained with just gradient descent, the way deep neural networks are trained, meaning you say, predict the next word.

And you do a gradient, you differentiate your function from the model. They're not fully differentiable, but you approximate it and you train your weights to do that. And it was quite amazing that doing just that, you could make a chat. But with the reasoning model, you can't do that because there is this reasoning part that you can't differentiate through. So we train this with reinforcement learning. And reinforcement learning basically tells you, okay, there's just this reward and you need to do a bit of tries and reinforce, meaning push the model towards doing more of the things that lead to better answers.

And this kind of training has more restrictions than the training we used before. The training we used before, you took all of the internet, put it in, even if you didn't filter it very well, it would mostly work. Reinforcement learning, you need to be careful. You need to tune a lot of things, but you also need to prepare your data very carefully. So currently, and current for at least the most basic ways we use it currently, it needs to be fairly verifiable.

So, is your answer correct or not? You prepare data for that. You can do that in mathematics and coding very well. You can do this in science to some extent, right? You can have test questions, they're correct or not. But if it comes to, like, writing poems, is this poem good or not? For now, the reasoning models really shine in domains like science, and they've brought some improvements to non-science domains, but it's not quite as huge maybe yet as it could be.

But I mean, at least compared to mathematics and coding. Then there is the multimodal question. How do you do reasoning in multimodal? I think this is starting. I saw some Gemini creating images in the reasoning part. That's quite exciting, but it's very, very young.

At the beginning of ChatGPT, let's say there was pre-training. People did not do RL, right? But then you couldn't really chat with it. So chat was RLHF applied to a pre-trained model, but the RLHF was a different kind of RL of sorts, right? It was very small, and it was human preference that was telling you what is better. That's what the HF is, human feedback, right? You showed people pairs of stuff, you learned a model that says, well, people seem to prefer this as an answer, you trained with that, it would very quickly, what we today say, hack this model.

Like, if you trained the RLHF too long, it would start giving things that satisfy this model that seems supposed to model human preferences. So it was a bit of a brittle technique, but it was a bit of an RL that was extremely crucial to making the models chat. These days, I think most people move towards this big RL. It's still not as big as pre-training in scope, but it says you have a model that says whether this is correct or not, or if it's a preference, it's a very strong model that analyzes things and says you should prefer that.

And you have data that's restricted to a domain where you can do this well. And then you can also put some human preferences on top, but you make sure that you can run a little bit longer without making this whole grading fall apart. But again, this is our route today. I do believe the RL of tomorrow will be broader. It will work on general data. And maybe then it will expand to domains that go beyond where it shines today. Now, whether it will shine there is a different question.

Do you really need to think very much doing some of the things? Maybe not, but maybe yes. Maybe we do more thinking and reasoning than we consciously call thinking.

I think this is a small part of it. I think that's one part. But if you think of economic tasks, making slides is important there, following instructions, doing calculations—it's not math. But it's still very verifiable, right? What I'm thinking about is when you do pre-training, you take the internet and you just ask, what's the next word? You could think before you ask what's the next word, obviously. Now, you don't want to think before every word, probably, but I don't know if you've ever looked at the training data for a real pre-training run, because I think people mostly don't realize how bad this is.

Booking.com is a great website compared with the average chunk of 2,000 words from the internet. It's a mess, right? And also a miracle that from this, the pre-training process gets you something reasonable. So you probably don't want—imagine you have a hotel website telling you it's a beautiful vacation—you don't necessarily want to have a very long chain of thought before that, right? If it was written by a person, there was probably some kind of thinking that went into it.

Maybe not as elaborate as the math and coding thinking, but maybe there was something going on. So maybe you want a little bit of thought before at least some of the text. And that our models can't do very well yet. I think they're starting. There is a lot of generalization in this reasoning. If you learn to think for math, you will sometimes do some—some strategies transfer very much: look up on the web and see what they say, and use that information.

So some of these things are very generic and they start to transfer. I feel like some are maybe not yet, especially thinking in the visual domains is very undertrained, I believe. But we work, so we will try to push for more of that.

So in the current ChatGPT, you will see a summary of the chain of thought on the side. So there is another model that takes the full chain of thought and shows you a summary because the full ones are usually not very nice to read. They more or less say the same thing, just in more messy words. So it's better to have a more readable summary. When you start with a chain of thought, the first paper about chains of thought, you basically just ask the model, please think step by step, and it would think.

So if you just pre-train a model on the internet and ask it to think step by step, it will give you some chain of thought. The interesting and most important point is you don't stop there. You say, okay, so you start with some way of thinking. And then you say, sometimes this leads to a correct answer and sometimes it leads to a wrong answer. So now I'm telling you, I have some training examples. You'll think 100 times, and say 30 lead you to the correct answer.

Then I'll train you on these 30 examples and say, this is the way you should be thinking. That's the reinforcement learning part of training. It changes dramatically how the models think. We see this for math and coding, but the big hope is it could also change how the models think for many other domains. Even for math and coding, you start seeing that the models start correcting their own mistakes. Earlier, if the model made a mistake, it'd generally just tell you what it did and insist that the mistake was right or something like that.

With the thinking, it's like, oh, I often make mistakes, but I need to verify and correct myself to give the correct answer. So this just emerges from this reinforcement learning. Which is beautiful, right? It's clearly a good thinking strategy to verify what you want to say. And if you think it may be an error, then think again. That's what the model learns on the most abstract level.

I was a mathematician and a computer scientist, but in theoretical computer science.

Yes, I was definitely very into math in high school and into computers also later in high school. Yes, I did my studies in Poland. I went for a PhD in Germany. It was a theoretical computer science and mathematics PhD. So I very much am a mathematician. Yeah, I was always fascinated by how is this thinking going? What is intelligence? As a child, I always wanted to emulate the brain. I thought, well, okay, maybe higher-level explanations are more interesting.

I did research in logic, but a little programming. But then there was this opportunity to join Google just as deep learning was starting off. I already had my tenured position in France, and the French system has this beautiful thing that you can take a leave of 10 years.

And you can still return anytime you want from the leave. So it's no risk.

Well, if you solve AGI, they may take you anyway. The nice part about the leave is that they will take you back even if you don't. But it's actually very important. I think there's a number of Nobel Prize winners who took this leave to just try something more risky. And sometimes it works. Sometimes it doesn't. There's a lot of luck in science and research. But it's very good to have this opportunity to take it. So I came to Google.

I came to Ray Kurzweil.

He was my first manager. He interviewed me and was very inspiring. My first interview was to join the YouTube UI team, and I was like, okay, I'm not going. And then I had an interview with Ray, and I knew him, of course, from his books, and he's a very inspiring person. So I was like, okay, let's go. The team was separate from Google Brain at that time. Then I moved to Google Brain, worked with Ilya Sutskever, another very inspiring person.

There's an amazing number of great people in AI, and in the Bay in general.

Well, we never got together.

I recently got a photo on Twitter of a photo session of all eight of us, and it was saying it was fake, but I knew it was fake because I don't think all eight of us were ever in the same physical room. These ideas developed from many sides before and after. Jakob and Illia worked on attention, like self-attention. Of course, attention was there from the encoder-decoder side.

So attention is the mechanism that tells the model, as you're doing the next thing, look into your past and find the most similar things that you see in the past to what you are seeing right now. It came from the machine translation times where people wanted to align words in one language with words in another. They were like, okay, so this word, where in this previous sentence would it be? It's an analog of alignment for deep learning. It's now called attention in AI.

It just says, think of what comes to your mind as you are here now in this environment. What things from the past are similar to it? And this mechanism was already used in deep learning translation before, but there was one encoder model and the decoder would be looking at the encoder, but never at its own states. The main novelty of Transformer was self-attention, but Transformer is more than just this idea. I think that's important. I think that's the beauty of these eight weird people that somehow came together.

Together, even though not physically, to do it, is that we all approached it from different sides. So there were people working on the attention idea. You need to put this in a network that needs to have a lot of knowledge. So there is the feed-forward layer that expands and then contracts. So Noam was working on this, and nowadays we use mixtures of experts, which actually came before the Transformer. So how do you store knowledge in neural networks is another important question.

And it's part of this model too. And in deep learning, people laugh that ideas are cheap; making them work is the hard part. So how do you write the systems and the code and the baselines to actually make this train? And this is funny to say now, because nowadays you can take any deep learning framework and say, "x equals Transformer, x train," and it will basically work. But back then, it totally did not.

So you need things like learning rate warmup or tweaks to the optimizer that just work. And I did a lot of coding at that time. I was working on TensorFlow and parts of the framework. And I remember distinctly that people were like, "So you want to use the same model for a few different tasks? Like, why do you even do that? Like, if you have a different task, like if you do translation, you train one model. If you do parsing, you train another."

If you do image recognition, you train a third. Like, you never train the same model for three different tasks. Why do you even write APIs to do multiple tasks on one model? And I was like, "No, no, we're gonna do all tasks in one model." And people were like, "No, no."

Not against the idea. Google was such an amazing place at that time that they would very happily let you work on whatever you wanted. But I don't think there was widespread belief in doing multiple tasks with the same model, not to mention this idea that you—I still find this idea that you take basically the same model as Transformer—now there is a bunch of changes to it—but you could in principle take the same architecture as the decoder from the paper, train it on all of the internet, and it will basically start chatting with you.

That would have back then definitely sounded as a worthy dream, maybe as a dream, but not a reality that you'd expect five years later. It's very lucky that it actually works so well, right?

So Ilya Sutskever was my manager at Brain, and he went on to found OpenAI. He asked me a number of times whether I would like to join over the years. I found it a little bit too edgy at the time. Then Transformers came, so we had a lot of work with that. And then COVID came, and COVID was a tough time for the whole world, right? But Google totally closed. Google was reopening extremely slowly. So one part of me was, I find it very hard to do remote work.

I much prefer to work with people directly. That was one reason, but the other was also Google Brain, when I joined it, was a few dozen people, maybe 40, something like that. When I left, it was 4,000, 3,000 people.

Spread across multiple offices. It's very different to work in a small group and to work in a huge company. So with all this, Ilya was like, OpenAI, though, is in a much stabler state. We're doing language models. You know something about this. That may look like a good match. And I was like, okay, let me try. I've never worked in any company other than Google before, other than the university. So it was quite a change to the small startup group, but I like working in smaller groups.

It has its pleasures, right? It has a little bit of a different intensity sometimes. In general, I found it very nice. On the other hand, Google has merged, made Gemini, and I hear it's also a very nice place. I think, in general, the tech labs are more similar to each other than people think. There are some differences, but I think if I look at it from the world, from the university in France, the difference between this university and any of the tech labs is much larger than between one lab or the other.

They're organized. They're not very organized. I mean, we do organize them, but some have managers. Sometimes talk to them. No, but mostly people find projects. There are things to do, right? Like improve your multimodal models, improve your reasoning, improve your pre-training, improve whatever part of the infrastructure. People work on it. As we go through these parts, there is infrastructure, pre-training, reasoning. I think the parts are the same for most of the labs.

So there will be teams doing these things, and then sometimes people change teams. Sometimes new things emerge. There's always some smaller teams doing more adventurous stuff, like diffusion models at times. Then some of the more adventurous stuff, like video models, gets big, and then maybe they need to grow.

I don't think it's so much people that compete. I think it's more projects that compete for GPU access. There's definitely some of that. On the other hand, on the big picture of GPU access, a lot of this is just determined by how the technology works, right? Currently, pre-training just uses the most GPUs of all the parts. So it needs the most GPUs, right? RL is growing in use. Now video models, of course, use a lot of GPUs too.

So you need to split them like this. Then, of course, people will be, oh, but my thing would be so much better if I had more GPUs. And I've certainly said that a number of times too. So then you kind of push it: I really need more. And then some people may say, well, but there's only so much. There are never enough GPUs for everyone. So there is some part of the competition, but the big part is just decided by the fact of how the technology works currently.

Pre-training, as I said, I think it has reached this upper level of the S-curve in terms of science, but it can scale smoothly, meaning if you put more compute, you will get better losses if you do things right, which is extremely hard. And that's valuable. You don't get the same payoff as pushing RL, but it generally just makes the model more capable. And that's certainly something you want to do. I think what people underestimate a little bit in the big narrative is OpenAI, three, four years ago—I joined even before that—was a small research lab.

With a product called API, but it was not such a big—there was no GPU constraint on the product side, for example. All GPUs were just used for training. So it was very easy as a decision for the people to say, we're going to train GPT-4. This will be the smartest and largest model ever. And what do we care about small models? I mean, we care about them to debug the training of the big model, but that's it.

So GPT-4 was the smartest model and it was great, right? But then it turned out, oh, there's this chat, and now we have a billion users, and people want to chat with it a lot every day, and you need GPUs. So you train the next huge model, and it turns out you cannot satisfy this. People will not want to pay you enough to chat with the bigger model. So you just economically need the smaller model. And this happened, of course, to all the labs because the moment the economy arrived and it became a product, you had to start thinking about price much more carefully than before.

So I think this caused the fact that instead of just training the largest and largest thing you can for the money you have, we said, well, no, we're going to train the same thing, same quality, but smaller, cheaper. The pressure to give the same quality for less money is very large. In some sense, as a researcher, it almost makes me a little sad. I have a big love for these huge models that people say the human brain has 100 trillion synapses, or orders of magnitude.

Of course, they're not that exactly calculated, but our models don't have 100 trillion parameters yet. So maybe we should reach it. I would certainly love it. And you need to pay for it. And so I think this may be why people kind of think that pre-training has paused, because a lot of effort went into training smaller models. Now, on the side, people kind of rediscovered how amazing distillation is. Distillation means you can train a big model and then put the same knowledge from the big model—the big model is a teacher—to the little model.

People knew about distillation. It's a paper from a long time ago. But somehow, at least for OpenAI, I think maybe it was more in Google's DNA when Ariel's there, but people kind of rediscovered how important that is for the economics. But now it also means that, oh, training this huge model is actually good because you distill all the little ones from it. So now maybe there is a bit more of a return to large. It's also a matter: once you realize you have the billion users and you need the GPUs, you need to invest into them.

And of course, everyone sees this. There's a huge investment, but the GPUs are not online yet. So when they come online, and I think this may play into this, what people call resurgence of pre-training, we both understand that you can distill this amazing big model and there are now enough GPUs to actually train it. So it's resurging. But all of this fundamentally happens on the same scaling curve, right? It's not like we didn't know that you could do this.

It's more like the different requirements of different months have sometimes changed the priorities. But I think it's good to step back from it and think of the big picture, which is that pre-training has always worked. And the beautiful thing is it even stacks with RL. So if you run this thinking RL process on top of a better model, it works even better than if you run it on top of a smaller model.

I would say both. There is huge progress in understanding models. Fundamentally, think of the model that is ChatGPT. It talks to a billion people about all kinds of topics. It gets this knowledge from reading all of the internet. Obviously, you cannot identify—I cannot understand what's going on in there. I don't know the whole internet. What we can identify is—there was a beautiful paper, I think, just last week from OpenAI about if you tell the model that lots of its weights should be zeros, it should be very sparse.

Then you can really trace when it's thinking about one particular thing. Then you can trace what it's actually doing. So if you say, let's limit ourselves to this and really study this inside the model, then you can get a lot of understanding. And there are circuits in the models. Anthropic had great papers on that. So the understanding of what the models are doing on a higher level has progressed a lot, but then it's still an understanding of what smaller models do, not the biggest ones.

But it's not so much that these patterns don't apply to bigger models. They do. It's just that the bigger models do so many things at the same time that there is some limit to what you can understand. But I think this limit is a bit more fundamental than people think. It's like every very complex system. You can only understand so many things, and then you don't, right?

That's a very tough question. I think less than you think. No, I mean, from GPT-4 to GPT-5, I think the biggest thing that changed is reasoning, meaning RL and synthetic data. As I told you, the pre-training part in that time frame was mostly about making things cheaper, not making things better. So, of course, the price has changed dramatically too, right? A thousand times, I think, or some of these orders of magnitude. The main improvement from GPT-4 to GPT-5 is adding reasoning with reinforcement learning, and this allowed us to generate synthetic data, which also improves the model.

So that's the big picture. In addition to that, ChatGPT is now a product used by a lot of people. So the post-training team has learned a tremendous number of lessons. And it's added things clearly experimented—wanted the model to be very nice to you, then it turned out to be too nice. Then now, when a lot of people use it, you need to be really careful about safety, right? There may be people who are in distress using the model.

The model needs to do something reasonable in these cases. It was not trained for it before. Now it is, and it makes the model much better. But at the same time, you don't want to refuse to answer any question that has any sign of anything. So as you work on these things, you make the model much better in use, not just for the people in distress, but for everyone who wants questions answered, but the answers to be reasonable.

And there were these things called hallucinations. It's still with us to some extent, but dramatically less than two years ago. Some part of that is because reinforcement learning can now use tools and gather data, and it also encourages the model to verify what it's doing. So that's an emergent thing from this reinforcement learning of reasoning. But also, you just add data because you realize sometimes the model should say, "I don't know." So you add this to the post-training data.

You say, well, we really need to give it a thought how the model should answer people in various situations. This is mostly this kind of, like, post-training improvement.

I don't work on post-training, and it certainly has a lot of quirks. But I think the main part is indeed RL, where you say, okay, is this response cynical? Is this response like that? And you say, okay, if you were told to be cynical, this is how you should respond. If you were told to be funny, try this one. So I do think RL is a big part of it.

There used to be a time, not that long ago, half a year in the distant past, where the models did have an alignment with technical stuff, right? So they would align either with RL runs or pre-training runs. That's why you had a beautiful model called GPT-4o, which was aligned with a pre-training run, which was obviously worse than the o3, aligned with an RL run that was the follow-up to o1, naturally, because you couldn't use the name o2.

But it was slightly better than the o4-mini, because that one was mini. And we had this beautiful model picker, and people kind of thought this was not the best naming for some whatever reason. So no, I mean, it was fairly obvious that this was very confusing, right? So now the naming is by capability. GPT-5 is a capable model. GPT-5.1 is a more capable model. Mini is the smaller model that's slightly less capable, but faster and cheaper. And the thinking models are the ones that do more reasoning, right?

Two is the newly pre-trained model, or maybe not, but the naming is detached from the technology, which also gives some—as OpenAI has grown, there is a number of projects, right? There is RL and pre-training, and there may be something just to make slides better or whatnot. And with distillation, you have the ability to put a number of projects into one model. It's kind of nice that you don't need to wait on all of them to complete at the same time, and you can try to periodically put this together, actually make sure that as a product, it's nice to the users and good, and do this separately from waiting on the new full pre-training run that takes months and so on.

So I feel like even though a little tear in my eye goes for the times where it was that pre-trained model number that was the number, as it's a product serving a billion users, it's maybe inevitable that—

So the model sees the task. It will decide on its own a little bit how long it should think. But you can give it additional information that can tell it to think even harder, and then it will think longer. So you now have the ability to steer that. I still think it is important to realize this is the fundamental change that came with reasoning models: using more tokens to think increases your capability, and it increases it, given the compute, way faster than pre-training, right?

So if you give GPT-5 the ability to think for long, it can solve tasks that are—we had these gold medals at Mathematical Olympiad and Computer Science Olympiad. So, they're amazing abilities. At the same time, the fundamental training method of reasoning is very limited to science data. So it's not as broad as pre-training, which I think—pre-training models felt kind of almost uniformly good or bad at things. I mean, this was still not uniform because it's not like teaching humans, right?

But the reasoning models are even more—people call it jagged, right? They have amazing abilities somewhere and then, close by, not so much. And that can be very confusing. It's something I always love: it's weird because you can say, "The model's amazing at Mathematical Olympiad." At the same time, I have a math book for—I have a first-grade daughter in first grade. She's 5 years old. I took one exercise from this math book, and none of the frontier models is able to solve it.

And you would be able to solve it in 10 seconds. So that's something to keep in mind. The models are both amazing, and there are tasks that they cannot do very well. I can show you this as an example. I think it's quite interesting to keep in mind. Let me start with Gemini 3, just to blame the competitors.

So it has, you see, two groups of dots on both sides. And the question is, is the number of dots even or odd? And if you look at it, you see, oh, they're like two identical things. So that would be even. That's what the 5-year-old is supposed to learn. But there is one dot that's shared. So now that must be odd. For this simple one, which has, like, I don't know, 20 dots or so, Gemini 3 actually does it right.

It finds out that it's an even number of dots, and it says that, and that's great. And then you have another puzzle, which is very similar, except now there are two mountains of dots, and there's also one dot shared at the bottom now. And right in context, right after that, you ask, okay, how about this one? And then it does some thinking, and it just totally misses that there is a shared dot, and it says the number is even. And it's like, in context, where you've seen this first example, how would you ever miss that?

GPT-5.1 Thinking. And it also solves the first. It sees the dot, it says it's odd, and then it sees the mountains and somehow it doesn't see the dot and it says it's even. The nice thing is, if you let it think longer, or if you just let it think again, it will see it. So if you use GPT-5.1 Pro, it takes 15 minutes. The human 5-year-old takes 15 seconds. GPT-5.1 Pro will run Python code to extract these dots from an image, and then it will count them in a loop.

So that's not quite—

I think this is mostly multimodal. The models are just starting. Like, you see, the first example they manage, so they've clearly made some progress, but they have not yet learned to do good reasoning in multimodal domains, and they have not yet learned to use one reasoning in context to do the next reasoning. Learning in context happens, but learning from reasoning in context is still not very strong. All of these, though, are things that are very well known, and the models are just not trained enough to do this.

It's just something we know we need to add into training. So I think these are things that will generally improve. I do think there is a deeper question whether—so multimodal will improve, this will improve. We keep finding these examples. So as the frontier will move, it will certainly move forward. Some things will smoothen. But the question is, will it still be just other things that you don't need to teach the human? Like every, okay, now you know how to use a spoon and a fork, but now if the fork has four instead of three ends, then you need to learn anew.

That would be a failure of machine learning. I am fascinated by generalization. I think that's the most important topic. I always thought this was the key topic in machine learning in general and in understanding intelligence. Pre-training is a little different because it increases the data together with your increase in model size. So it doesn't necessarily increase generalization; it just uses more knowledge. I do believe that reasoning actually increases generalization, but now we train it on such narrow domains that it may still be too soon.

But I think the big question in all of AI is: is reasoning enough to increase generalization, or do you need more general methods? I think the first step is to make reasoning more general, as we talked about before. That's my passion. That's also what I work on. There is still something there, right? Yeah. We push the models, they learn, they learn things that are around what we teach them. They still have limitations because they don't live in the physical world, because they're not very good at multimodal, because reasoning is very young, and there's a lot of bugs in how we do it yet.

But once we fix that, there will be this big question: is that enough, or is there something else big to make models generalize better? So we don't need to teach it every particular thing in the training data, that it just learns and generalizes. I think that's the most fascinating question. But I also think a good way to approach a question like that is to first solve everything that leads up to it. You cannot know whether there is a wall or not until you come close to it, because otherwise, AI is moving very fast.

Someone said it's like driving fast in a fog. You never know how far or close you are. So we're moving, we are learning a lot.

I think there is a lot of beautiful work that people are trying out. The ARC challenges inspired one set of people. There are models now that are very small and solve them, but with methods that I'm not sure are actually general. We'll need to see. Yann LeCun has been pushing for other methods. So I feel like his approach is more towards the multimodal part, but maybe if you solve multimodal right, maybe if you do JEPA, it also helps your other understanding.

There are a lot of people pushing fundamental science still. It's maybe not so much in the news as the things that push, but whatever you do, it will probably run on some GPU. If you get a trillion dollars of new GPUs, the old GPUs will be much easier to get, also for us. So I think this growth in LLM AI on the more traditional side is also helping people to have an easier time to run more experimental research projects on various things.

So I think there is a lot of exploration, a lot of ideas. It's still a little hard to implement them at a higher scale. The engineering part is the biggest bottleneck. GPUs are a bottleneck too when you scale really up, but implementing something that's larger than one machine, it's an experimental research project, so you don't have a team to do that. I think that's still harder than it should be, but Codex may get there, or CodeCoder.

This is the thing where AI researchers have great hope to help themselves and also other researchers: if you could just say, "Hey, Codex, this is the idea, and it's fairly clear what I'm saying. Please just implement it so it runs fast on this eight-machine setup or hundred-machine setup," that would be amazing. It's not quite capable of doing that yet, but it's capable of doing this more and more. I think that's what OpenAI says. They say, "We'd like an AI intern by the end of next year."

That's how I understand this. Can someone help us?

So it is both an engineering and model problem. You want to do some engineering task, like you have some machine learning idea, you want Codex to implement it for you, test it on some simple thing, find the bugs. So it needs to run this thing. This is not something you would do in an hour, right? That's something you'd spend a week on. So the model needs to spend a considerable amount of time because it needs to run things, wait for the results, then fix them.

The model is not like it's going to come up with the correct code out of thin air, right? It's just like us. It needs to go through the process. And oftentimes in the process, since it was not trained on anything very long in its training, or maybe very few, but certainly nothing that went on for a week, it can get lost. It can start doing loops or doing something weird. That's, of course, not something you want. So we try to train in a way that makes it not happen, but it does.

So how can you make the model actually run a process that requires this larger feedback loop without tripping up? And the other thing is transformers have this thing called context. So they remember all the things that they have seen in the current run, and that can just exceed the memory available for your run. And the attention matrices are n by n, where n is this length, so they can get huge. So instead of keeping everything, you say, well, I'm going to just ask the model on the side to summarize the most important thing from the past, put it in context anew, and forget some part of it, right?

So it's a very basic form of forgetting—that's compaction, right? And that allows you to run for much longer if you do this repeatedly. But again, you need to train the model to do that. And when you do it, it works to some extent. I don't think it works well enough to replace an AI researcher yet. It made a fair bit of progress. I think another part of progress that's a little understated on the research side, but is very important, is allowing the model to connect to all of these things.

So models now use tools like web search and Python routinely, but to run on a GPU, to have access to a cluster, it's hard to train models with that because then you need to dedicate the thing for the model to use. And that has security problems. And how do models connect with the external world? It's a fundamentally very hard problem because when you connect in an unlimited way, you can break things in the real world. We don't like models to break things for us.

So that's a part where people work a lot. It overlaps with security, right? You need to have very good security to allow models to go on and train on the things they need to train on.

I just showed you a 5-year-old exercise that the model doesn't do. I think we need to keep that in mind.

There is hope the next model will do it. Right.

Okay. Well, for me, yes. I think we have some way to go on the models. Progress has been rapid, so there is good hope there'll be less and less of this. But on the other hand, for now, you don't need to do a deep search to find things where you'd really want a human to do that task because the model's not super good. On the other hand, the Transformer paper started with translation. I recently went to a translation industry conference. The translation industry has grown considerably since then.

It has not shrunk. There's more translations to be done. Translators are paid more. The question is, why would you even want a translator if the model's so good in most of the cases? The answer is: sometimes. Imagine you do a listing for a newspaper, but in a language you don't know, and GPT-5 will almost certainly translate it correctly for you if it's into Spanish or French or any high-resource language. Would you still publish it without having a human who speaks that language look at it?

Would you publish it if it's a UI of ChatGPT that a billion people are going to see? It's a question of trust. Probably right, but if you have a million users, a billion users, maybe you will pay the $50 for someone to just take a look over it before you publish it. So this is in an industry that fundamentally is totally automated, right? There is still the question of trust, and I think that's a question we will grapple with for a long time.

There are also just things you want a person to do. I don't think we will have no things to do, but that doesn't mean that some things we do may not dramatically change. And that can be very painful for people who do these things. And so this is a serious topic that many people are engaging with, but I don't think there will be this global lack of anything for people to do.

What do you personally find really interesting as a researcher? Well, I find this general data reinforcement learning is my pet peeve and what I work on, luckily. For example, robotics is probably just an illustration that we are not doing that well in multimodal and that we're not doing that well in general RL, in general reasoning yet. The moment we do really well in multimodal and we manage to generalize reasoning to the physical domains that the robot needs, I think we'll see amazing progress.

When it does, I have a feeling, given that a lot of companies are launching hardware that's kind of teleoperated or glove-operated or something, so my suspicion is, by the moment we make this progress, which maybe it'll be next year, maybe it'll be in a few more years, the hardware will maybe be there by then, and having a robot in the home may be, like, a big visible change, more visible than chat. I mean, given how quickly we got used to the self-driving cars in San Francisco, maybe it will be only visible for, like, the first two days, and then be like, yeah, sure, the robot's there.

It's always been cleaning since I can remember, like the last three months. It's stunning to me how quickly we get used to these things, right? The self-driving cars in San Francisco are something people got used to so quickly. So maybe this will happen for robots too. Nevertheless, I do think it will be quite dramatic in our perception of the world when it happens. Hardware is hard, though, right? Robots may have accidents in the house. You need to be very careful.

So maybe it will take longer to deploy them and actually make it a scalable business. We'll see. It is amazing that we are at this point where we can start thinking, like, yes, maybe that will come soon.

Thank you very much, Matt. Thank you for the invitation. Great to talk to you.