[0:00:00] Are you creating the strongest AI in the world? Our company is built on the belief that in the near future it will be equal or more intelligent than people. I believe that this development has no end. And where we are now is, in a sense, the beginning of a new paradigm that is also scaling. Wojtek Zaremba wrote to me and I went to OpenAI. It really shook my world. Will we get to the point where models will train even better models? For sure. But it will be a slightly different world than...
[0:00:29] Are you not afraid that AI will replace Kuba Pachocki one day? Yes. Since the discovery of atomic bombs, the world has not yet found itself in such an unstable place. What kind of revolution could it be? So this is something that stresses me out. The next step is AI, which is really able to build something in the real world. Jakub Pachocki. He won six times or took the leading places at the International Informatics Olympiad.
[0:00:57] In this field, we are the best in the world. After leaving OpenAI by Ilya Sutskever, he advanced to the position of head of scientists at OpenAI. Jakub, you're the chief researcher, chief scientist at OpenAI? Chief scientist, yeah. Today, he leads all scientists in the most important company worth $300 billion in AI development. In 2025, Time magazine put him on the list of the 100 most influential AI people on Earth.
[0:01:23] Who is Jakub Pachocki? Now I am Chief Scientist of OpenAI, which means I run a research program of our company.
[0:01:51] So I decide which directions we will take, how we will develop the AI we are working on. Are you creating the most powerful AI in the world? Our company is built on the belief that in the near future AI will be equally or more intelligent than people. I think that five or seven years ago it was more like
[0:02:23] Maybe there will be a moment when it will be. Now it seems to me that we see that it is more continuous. ShareGPT is in some way wiser than people, there are many more stupid ways and I think that it will just develop further. But yes, for sure a lot of our focus is to create the strongest AI in the world, but also to understand how it works, how people will cooperate with it, how it will have an impact on people.
[0:02:53] Do you sometimes think about how much the computers will be able to do? I mean, this is an object, actually, it's been an object for a long time, my work or research, right? It's just developing a little bit. When I was in high school, I was with professor Szwartowski, it was what I am able to program so that the computer is able to do. This is just, it was always for me such a huge, that's why I felt such a huge urge to do it.
[0:03:24] There is a clear entry to the program and it is supposed to write such an exit. I can't see how the computer is supposed to do it. And the fact that I'm able to come up with something and show that the computer is able to do it very quickly, it was fascinating to me. And that's why I went to theoretical informatics and I did a PhD in theoretical informatics, where it is proven what computers are theoretically able to do, and not.
[0:03:55] This is a very long-term science, it really touches the boundaries of mathematics, i.e. what functions can be calculated, what functions cannot be calculated. But at some point you abandoned mathematics in favor of deep learning. I didn't abandon mathematics, because deep learning is definitely mathematics, but it's a little different, it's more applied. How is it that Kuba Pachocki from such theoretical mathematics, when I look at your work on Harvard, they were about theoretical mathematics and suddenly ...
[0:04:22] You focused on practice, deep learning, applied mathematics. When I started my PhD in 2013-2014, my view of the field was that there are algorithms that we can prove that they work. And algorithms that we can prove that they work, even if they are not very effective at the moment, not very practical, they are worse than some heuristic methods that we can apply, they have the property that if the computer is bigger, they will scale.
[0:04:52] And if you look at the last decades of scaling and if you look at the future from that moment, you can conclude that it is not so important what computers are able to do. Now it is important what computers will be able to do in 50 years and we should build some theory to understand what are the actual limits of computer capabilities, how to think about it and really develop our understanding of the algorithm.
[0:05:22] And what also convinced me in this competition is that a large part of the most interesting things that computers were able to do were based on such theoretical algorithms. For example, such an optimization algorithm, which was used a lot in practice and is still used, linear programming, or in particular the maximum flow, which has always been one of the most magical algorithms in algorithmic competitions.
[0:05:50] When I was at the doctorate, new, fantastic methods were created based on linear algebra to solve such problems. Professor Aleksander Mądry, whom I didn't know at the time, worked at MIT and made incredible progress. Today working with OpenAI. Yes. I'll get back to him. So I was thinking that we are building understanding of what computers are able to do, which will be important in a few years. And then there were some first results.
[0:06:19] I mean, deep learning results have been there for a long time, but then they definitely started to be good for vision, right? To recognize images. But it was not clear to me then whether it was something that would be scaled, whether it was something that would last longer, because the theoretical foundations of deep learning are much worse, right? Because unlike such theoretical and classical informatics, in deep learning we are not able to prove that we actually find the optimal solution.
[0:06:49] So that we find the optimal parameters of the network that we optimize. We just make a network that looks reasonable, we optimize it for some time and we get a result. So it was easy to imagine that this is something that works a bit, but we don't know if it will continue. What made me think about my approach to the field was the results of DeepMind with AlphaGo.
[0:07:17] There were these fantastic Deep Blue results, right? Inspiring, that computers can be used to play chess. Well, okay, chess is a game that is relatively small, there are quite a few moves, and it was easy to believe that modern computers are able to search thoroughly all the possibilities here. Maybe not all of them, but a large part. Then there's Go. And in Go, there's a very large board, you can put stones anywhere, there's a lot of possible moves.
[0:07:47] And there was a consensus among the community that this game would take many, many years, that such algorithms as Deep Blue don't really have a chance. And what DeepMind showed then was that if instead of manually writing heuristics for how to search for all possible movements in the game, we use a neural network that learns on the basis of
[0:08:14] People's games or their own games, we are able to play better than people, in most cases. This really shook my world, because it is much easier to imagine how from playing in GOM you can get to solving, let's say, grammatical competitions or further writing scientific papers and really developing science than from recognizing pictures.
[0:08:45] And since then, I started to be more and more interested in it. Soon after, I finished my doctorate and I started talking to a few potential companies that wanted to conduct deep learning research there. At Harvard, I was more in the direction of machine learning. I actually have theoretical work there, but more related to the topic. And also, when I was at Harvard, I spent a lot of time there. I got to know it better then.
[0:09:16] We spent a lot of time together. As I said, he had amazing results on graphs in the same field as I worked before, but he came to similar conclusions as I did, so we talked a lot about how it will all change. We were both very dissatisfied with the fact that deep learning has such weak theoretical foundations and we don't quite understand how it will scale.
[0:09:46] How to think about it and we were wondering if we were able to build a better theory here and some better way to think about it all. And in particular, since some time ago, Wojtek Zaremba, who probably had such a plan to recruit people who were doing well in competition in economics, wrote to me from Facebook AI Research and then from OpenAI. And finally, after half a year of doctorate,
[0:10:17] After half a year of postdoc, I decided to jump into this field with both legs and I went to OpenAI. Do you have the feeling that you are one of the most influential scientists in the world today? Taking into account the scale of what you do? OpenAI? I'm asking you. Something that is... At OpenAI, we try to
[0:10:56] As long as we can promote AI development and the field in the right direction. But on the other hand, we don't publish a lot of scientific work, so I don't know if that statement would be fair. Aren't you afraid that AI will replace Kuba Pachocki one day? Do you have such a reflection sometimes? Yes, a little more than a reflection. This is something we actively think about.
[0:11:33] Consequences. The way I went to AI, from being interested in programming competitions, what computers are capable of, to AI is not very unique and there are a lot of fantastic people in our company who also do programming competitions. So when a few years ago someone asked me when I think we should really start seriously, when we will really be close,
[0:12:13] I said that when AI gets a gold medal or starts solving all the problems at the Math or Programming Olympiad, it's happening. It's happening now. We're in that moment now. So the next step is AI, which is really able to build something in the real world. For example, it's able to build useful programs or write books or conduct research and discover new things.
[0:12:46] And I expect to see it before the end of next year. In three months, you managed to improve OpenAI results by 25% on the most difficult frontier math test. It is worth mentioning that this test consists of two parts. One of them is much more difficult than the other, so we are probably close
[0:13:16] These models are already doing great on the part that is a bit more similar to mathematical competitions, not so much involved in science, but I think it will take a while before they will actually solve the whole exam. But yes, it's happening. It's not like we see a direct way to it. And we certainly think a lot about what it means that we will have models that
[0:13:46] With whom not only is it nice to talk and maybe they will help to remember something. But they are better than us. Yes, and we can, let's say, take some kind of graphic card cluster. I mean, we will be able to take some kind of graphic card cluster, run a number of copy-text models on it and we will simply have a research laboratory. They will just think for themselves and produce scientific results. I come back to research work and scientific results, because I think it's ...
[0:14:15] In a sense, the most drastic consequence, because the scientific results lead to new technologies, especially to accelerate the development of AI, but also to changes in medicine. And it can start to accelerate at a pace to which we are not used to. I would also like to say that models will be able to not only look for evidence, because it is happening today, but to look for claims, to make scientific discoveries,
[0:14:45] If we want to solve really difficult problems, many claims are discovered in order to prove some other, more difficult claims. So it's not like these are completely different problems. If models are able to solve difficult tasks that require many months of work, then it's mostly creative work.
[0:15:14] You have to formulate new hypotheses and build some understanding. We wouldn't be able to train the models we train today if not for many years of work and foundation building and trying to understand what the models from years ago did.
[0:15:35] This time is short, because first we had Transformer 2017, then you gave the world to ChataGPT, it's been five years, but it's been less than two years since ChataGPT to the O1 model. When you look at the pace,
[0:16:00] Looking for answers to this question, do you take advantage of the background of values that the high school gave you, that was given to you by Szubartowski, that was given to you by Gdynia and the three in Gdynia? As much as possible, yes. Especially now, when so much is happening. We are building a very important technology, which is already of great importance to people, it will probably be even greater. Of course, there is also a whole active market here.
[0:16:32] Competition, some business. And to remember in all this what the mission is, that we want to build AI that will really change the world for the better. The assumption is that AI will be created, so we want AI to work for good people, not against us. Remember this and remember that we want to build further, we want to build understanding of how these models work,
[0:17:03] How to make it work? It doesn't differ so much from the fact that you have to find time in the noise of competitions and schools and give yourself a little space and focus on what is really important. I was talking to Wojtek Zerembał, I'm talking to you now, I was talking to Szymon Sidor once. You are very sensitive people.
[0:17:33] Kuba, how do you deal with these sensitivity issues in this race where you compete with Google, Meto, Anthropic, Mistral, now Chinese Deep Seek? In a moment, another Deep Seek will be born. How do you deal with it? I try to focus on what actually
[0:18:13] We want to achieve fundamentally. Our company is not built on the assumption that we have to build a big business here, which will be bigger than Google. Of course, now we have achieved a lot of success, the GPT is developing, there is competition, we have to think about how we will finance future research and we have to talk to each other a lot, but I have fantastic co-workers who also think a lot about these
[0:18:42] I am able to focus mainly on research work and on how what we are working on now leads to artificial intelligence that is able to conduct research work and how this artificial intelligence affects the world and what we can do to make this whole journey go well. The biggest and last breakthrough in AI development is O1 for you. Free reasoning, abstractional thinking. Yes.
[0:19:14] How do you achieve such model efficiency when you don't use TreeSearch, when you don't use decision trees? Observation, right? That language models like GPT-2 or GPT-3 do better than others. If you even ask to think for a moment in context. It was a long time ago, right? It was one of the first observations that really inspired people.
[0:19:42] So it's not a new idea that we want the model to think for a long time in the context. What took a bit of time was to develop the technology to be able to actually learn. So the basis of this technology, which is also not a big surprise, is reinforcement learning.
[0:20:07] We train on the basis of what the model does. So instead of copying the way people think, we just look at how the model thinks and strengthen the parts of his thinking that lead to better results. This is the same method that AlphaGo Zero used to teach neural networks
[0:20:37] How to play Go without looking at how people play. This is also the same method that we used, it was my first big project in OpenAI, which we used to teach the neural network how to play a computer game DOTA2. OpenAI, from the beginning we were very interested in
[0:21:03] Reinforcement learning. We thought about how we can get to the point when these networks learn from their actions. The combination of these original ideas with these fantastic GPT solutions leads to such models as A1. Such a paradigm has been changed, right? The supervision not of the conclusions, but of the way the model makes decisions. In a sense.
[0:21:35] The supervision is still on the conclusions. The difference is that we train the model not only to produce conclusions, but to strengthen its own currents of thought that lead to correct conclusions. What else can we do in science? So we can stop for a moment at the fact that we have this new paradigm, because it is already in itself. I think it's easy to move on to the daily routine,
[0:22:08] Recently, when we had a new paradigm, it was maybe GPT-1 or GPT-2, right? And the difference between such models and such GPT-4, on which ChatGPT was based, is huge, right? And where we are now is, in a sense, the beginning of a new paradigm, which is also scaling. So, indeed, we can develop on the same axis. And that's definitely not all. And I'll tell you what else we can do.
[0:22:40] But there is still a lot of work to be done on this axis, so I think we will see much smarter models based on similar methodology. What these models are not doing at the moment is that they do not have long-term memory. Neither from O1 nor from FGPT.
[0:23:01] Yes, there may be some simple memory mechanisms, but it's not like with a person who can talk every day and he will really build the context of everything that is happening, he will learn about himself, he will learn about the world around him. At the moment, most of these models are not in this state. Especially if we want to conduct very long-term research, if we want these models to work in the real world, this is one of the things that is still to be solved.
[0:23:31] Another important thing is that at the moment, let's say, such a model as O1 is some kind of chat. You can talk to the text. But of course, it could do a lot more. First of all, if it can use a computer, which will probably be soon. And also in the future, if it is able to function in the real world, there is still robotics,
[0:24:04] Are you talking about spatial understanding? Also, but I think that these technologies that we have, I think you can see how spatial understanding comes out of them, that it is not such a difficult part. We already have video generators, we have ways to watch video, so I think these models will be able to understand space. However, I think that in a sense it is a more difficult problem, or at least one that will take a little more time.
[0:24:30] Give them a robotic body that works really well. The transfer of the rights of physics awaits us. The transfer of the rights of physics to models. I mean, first of all, if we look at such models as SORA, models that generate video, they somehow understand the rights of physics. They are able to generate video, which in most cases modifies some of the rights of physics. Yes, but if you want to integrate robotics with artificial intelligence, it's not enough. It's not enough to understand space, but you also have to understand the effects.
[0:24:58] What specific robotic action, robotic arm causes in the physical world? Well, yes, but that's why, if we look at such a room, it can generate a video in which the hand pushes the ball and the ball will then roll in some reasonable way. Of course, it's not perfect yet. There is also an interesting question, right? How much understanding...
[0:25:21] The law of physics builds such a model as O1, or even GPT-4, from the very Russian text. I wonder if he is not able to create his own physics. Or is he able to complete the lack of questions we have about reality? Because we don't know much about the world around us. If we look at...
[0:25:51] Both GPT-4 and O1, these models try to predict what is happening in the real world. And if we try to predict really well what someone would write about what was happening in the real world, we have to be able to model the law of physics. It is possible that what we are modeling is a question from physics, or some scientific work on physics, or a book in which someone describes an event and the consequences of that event.
[0:26:19] So even without looking at the pictures and without having a body, you can learn quite a lot about physics. Do you believe that this development has no end? This is a very interesting question. Generally speaking, yes, I believe that this development has no end. And in particular, one thing that is a bit unintelligent.
[0:26:52] We always talk about AGI, Artificial General Intelligence. In a sense, it is a holy grail for the whole field. That is, artificial intelligence that is really as smart as people. And what is a bit unclear and unintuitive is actually how smart people are. If you think about it that way, you can imagine that...
[0:27:22] Real AI is when we understood 100% how deep learning works, we built big data centers in space, we invented, fixed all the shortcomings of our methods and then we built AI, because we solved all the problems in the world. But the truth is probably more that the intelligence is somehow developing and people are somewhere, in some degree, in some place.
[0:27:50] And you don't have to understand all these problems. At some point, these models will be smarter than us on practically every axis. And it doesn't mean that we will understand how they work so well. And it doesn't mean that it will be in 50 years, it may be very soon. And it's a bit hard to internalize, I think. Have you ever thought that we are developing together with language models and that talking about AGI is doomed to failure? Because it treats some level...
[0:28:21] As a point of constant development, to which we strive, but we are not able to achieve it, because we are also developing together with models. So if we treat a person as a benchmark, as a reference point, then this reference point is developing. Of course, it can develop slower than a model, but it is developing. Today we will consider something else as a strong artificial intelligence than in 5 years, 10 years, 15 years, because we will also develop.
[0:28:51] Have you ever thought about it? Yes, for sure. What I said before. I think it's clear that it's not a point where models are smarter than people now. There are some things that people do better. There are some things that these models do better. And it will develop. People will be working more and more with such models. But on the other hand, at some point we will reach a point where we will be
[0:29:22] We want to use big, holographic clusters to conduct really open-ended scientific research in medicine. And I think that when we are at this point, we will probably still argue whether this is already AGI or not. Maybe these are just some parrots, but it will be a slightly different world than now. Aren't you afraid of lack of data? A lot has been said lately about the lack of data. To train new models, we can create synthetic data, of course.
[0:29:53] O1 already works based on synthetic data in a large part. Yes, there is this paradigm of GPT, where the language model can be learned from a huge amount of text written by people. And indeed, people wrote an infinite amount of text and at some point it doesn't scale anymore. However, things like AlphaZero,
[0:30:24] Dota. They didn't have such a problem. And I think we see a lot of ways in which the current methods are much less limited by data. Will we get to the point where models will train even better models? For sure. I think this is the direct consequence of models that are able to... Such recurrent loops? I mean, of course it will be gradual. I think that already this year some part...
[0:30:55] This is a simple part of the understanding. For example, some internal implementations in the infrastructure will be supported by thinking models. However, it will take some time before we get to the models that run research programs. However, I think this is expected.
[0:31:24] Of course, if we are in such a place, we must have very good ways to supervise such models, which is not trivial, because they will be much better at thinking about these problems than we are. A few years ago, this problem was not at all clear in what way we could approach it, this problem of supervising smart models than we are. One thing that I think is very promising in such models as O1 is that they have
[0:31:51] They have their own inner monologue, their own inner chains of thought, which, contrary to what such a GPT chat writes, they do not try to please the user. If such a model O1, if we ask him a question and he thinks that we are not right at all and it makes no sense, maybe he will not tell us directly, because he is somehow trained so that the user will at least tell us in a nice way.
[0:32:17] On the other hand, internally, it will think what it wants and we are able to see it. An internal monologue of thoughts. Yes. It's a very interesting problem. Especially now in the context of the DeepSeq release, there is definitely such a conversation. They give quite direct access for the user to how this model thinks. In our first release, we didn't give such access. What we also think about is that
[0:32:47] If we make such products in which part of this product is full access to what the model thinks, then of course, ultimately, it will be subjected to the same pressure as what ChatGPT writes. Users will want this model to think clearly and not think about how to deceive us or something like that. The problem is that if we supervise these models so that they don't think about how to deceive us,
[0:33:16] Then they can think about how to look for us in a way that is not obvious for us, which would be much worse. I wonder when and why young people today do not cope with the sense of humor, telling jokes. Where is the reason? Sometimes I wonder if it will not be the last thing that remains for us. The ability to tell jokes. I think they were not trained for it.
[0:33:46] I expect it to improve significantly. I know that we don't know what intuition is, so it's hard to use such a term, but it requires a lot of knowledge, a lot of context, data analysis obtained with all the senses that AI doesn't have. Maybe not necessarily with all the senses. I admit that at the moment the models are quite bad at making jokes,
[0:34:18] These are our newest models. I sometimes pull out some funny hair. But I agree that it is clear, especially in comparison to how great they are at math and stuff. It's not something that they do well. But it's not something where they are well trained at the moment. But I really expect it to change pretty soon. So I'm definitely not saying that this will be AGI when there will be a sense of humor. I think the sense of humor will be pretty soon.
[0:34:47] It will definitely be something very interesting and it will be a slightly different feeling, especially if we were able to get to the models that are able to produce interesting jokes as well as they are now able to solve math problems. I think it would be pretty cool and amazing, but I think it will happen soon enough. Is there something that Kuba Pachocki is afraid of, looking at the development of technology? There are certainly a few things about which
[0:35:26] I think a lot. What we've just talked about is in the long term, when these models are already conducting research work, building their new versions, how it will work, how we will supervise it. This is a big problem we have to solve. There is also a problem, maybe a little more short-term, even if the models are not so smart yet, but they are able to do a large part of the work.
[0:35:54] This is a real technological revolution. Since the discovery of atomic bombs, the world has not yet been in such an unstable place as such a revolution could be. So this is something that stresses me out. Finishing our meeting, what does your day look like? What does Kuba Pachocki's day look like? You get up in the morning, or maybe not in the morning, although probably in the morning. Rather in the morning.
[0:36:31] I've been sharing my time every other week for some time now. It was an idea of mine. I sit in the office all week long, meet people and talk about how to run some projects. We talk about research directions and I try to help people create space and direct it all in some other place.
[0:37:02] I try to focus on one specific technical problem, maybe on the research direction I'm working on, or to cooperate with one team to really be able to delve deeply. Or to be with yourself. Different. I often work with Szymon during this time, with whom we are very close co-workers since the beginning of my career at OpenAI. Or with other co-workers from the company.
[0:37:32] The problem is that there is a lot of context. We are building on a few layers of some understanding, some new research directions, and if I would only jump from meeting to meeting all the time we are here, at some point I would no longer be able to...
[0:37:54] I really want to think deeply about it and that's how I'm trying to build some space for myself. And it works. As it should. What would you like to say to young people in Poland, programmers, who are interested in you, in the whole team of Poles in OpenAI? In general, being here, I also have the impression that you are doing a lot of good work for Poland, because our reputation here is very, very strong. What would you like to say to young people?
[0:38:25] Is there something like that? I think that the lesson that I learn many times in my life, maybe the first time from Prof. Szuwartowski, but also later, is that if something really interests you, and maybe it's not exactly what you're working on at the moment, or maybe you're not that good at it yet, but you're really passionate about it, you can do it. It's not like...
[0:39:07] Sometimes it's so hard to realize that you're building an image of what you're doing. I'm a student at school and I'm doing this, or I'm a programmer and I'm building a specific type of system.
[0:39:34] So of course I won't talk about it because it's something different and someone else is doing it better and so on. The understanding that you can actually do it and that nothing holds you back so fundamentally is incredibly strong. And of course, here in the Krzemowa Valley, it's such an atmosphere, it's a bit of a mecca for people who want to do something of their own. And here people don't really feel such barriers, but even here, at least I often have to remind myself of it.
[0:40:05] Thank you for this meeting. Thank you.