On January 5th, famed AI researcher Jerry Tworek stunned world+dog by announcing his departure from OpenAI. A few days later, he hopped over to the Core Memory podcast studio for his not-so-formal exit interview.
Tworek joined OpenAI in 2019 when the research lab was a research lab and had about thirty employees. He went on to work on many of OpenAI’s most consequential products, including the company’s reasoning technology, which ushered in a new era for the entire AI field. (Yes, Tworek worked on Q* before it was Strawberry before it was o1.)
Both Kylie and I have been longtime Tworek fans. He’s smart, funny and never really sought the limelight despite his massive contributions.
In the episode, Tworek reveals that he found it hard to keep doing high-risk, pioneering work at OpenAI as the company shifted toward what Tworek describes as more conservative ways. He, in fact, thinks the large AI companies have become conservative as a whole and that there might be bigger, better ideas to be found elsewhere.
No summary for this episode yet.
Speaker 3
Welcome to the Core Memory Podcast. This is Ashley Vance.
Speaker 4
And this is Kylie Robison.
Speaker 3
We are, well, we've done one pod since the new year began. And today we're bringing in the big guns for the second one. Jerry Torek of OpenAI is going to be here. He is something of a living legend in the AI space. This is Newsy.
Speaker 4
Yeah.
Speaker 3
He just left.
Speaker 4
Newsy, newsy. I saw a lot of people from OpenAI tweeting like, I'm crushed. I'm absolutely crushed. So yeah, he's a phenom.
Speaker 3
Yeah, obviously we will discuss this on the pod, but he led or was part of many of their biggest projects. Huge on this era of reasoning that has come is in part because of Jerry. So he's going to talk about his experience, I'm sure, and what he worked on. And then we'll see. We'll see how spicy he gets.
Speaker 4
Yeah, here he's hoping.
Speaker 3
On his open AI time. So anyway, that's what's coming. I did want to take just a little moment. So we're doing a bit of a different intro. clip. But I think I shortchanged a couple things on the last pod, which is we've been at this for a year. Yeah! And, you know, we have a broad media empire here at Core Memory. Obviously, if you're listening to the podcast, you've probably found the podcast. We've got a video channel on YouTube, podcast channel on YouTube. We've got our sub stack, making movies, doing all this kind of stuff. And... The reason I want to do this is just to thank people deeply who have been here for this first part of the journey, our loyal subscribers and listeners and everything. And yeah, anyway, just a really genuine, heartfelt thanks. We got this thing off and running and a bunch of people who signed up for the Substack from day one or tuned into the show. It's cool. Now we have Kylie bringing energy and fresh blood to this ventures. I always say that you put so much time and work and energy into everything you've
Speaker 4
delivered this year, and it really shows, and I'm glad people are subscribing. It's a huge reason why I wanted to join, is just because you really get this space. You work hard, and you also play hard. It's playful and fun and fucking rad. So congrats.
Speaker 3
Well, thank you very much. And for those of you who've been... potting or, uh, for free. That's all fine. We love Brex, Ewok Ventures. They help us, help us pay the bills. But if you have been listening to the show a lot and you want to leave a little token of thanks, I know this is so lame, but actually apparently it makes a huge difference if you would actually put a review, just take, if you've been listening to lots of the show, take five minutes.
Speaker 4
It's easy. I do this for a nice Starbucks barista. Come on, guys.
Speaker 3
Just a little bit. It'll really help us out. You'll be doing the right thing. And then otherwise, if you are already paying for our Substack and everything, thank you so much. And the rest of you should as well.
Speaker 4
Yes. And don't forget to subscribe to YouTube, Core Memory Pod and Core Memory.
Speaker 3
Big stuff there. Big, great stuff. All right. I won't keep selling too much, but genuine, heartfelt. Thank you to you all. Hooray. And then off we go with Jerry Torrey.
Speaker 1
Jerry, hello. Hello, Ashley. How are you? I'm having an excellent day today.
Speaker 3
You've got the glow of unemployment about you.
Speaker 1
Eight days of unemployed so far. It's a change, it's a shift. Having been unemployed for a very long time, but there are many upsides to it. I'm spending much more time in the sun right now.
Speaker 3
Yeah, well, so this is like, this is your exit interview. Thank you very much for a very, very fruitful collaboration. Well, we set the context for you very, very briefly. And so I'll dive in just one sec to give people a bit more context. But you started at OpenAI around 2019, I want to say. You and I, we've chatted before. You're from Poland. I am. Originally, very brilliant mind. You were kind of in the, like a lot of AI people in the high-frequency trading type world before coming into AI. And then at OpenAI, I worked on, either worked on or led the teams of, of just many of the major products that people are familiar with. And I think kind of maybe most recently sort of everyone would have heard of, of strawberry and oh one and this era of reasoning models and everything. And that was something that you were chasing for quite a while. Yeah, exactly. And then as we all mentioned, you just, you just recently left. It was a thing on X on Twitter and
Speaker 1
In a way, it's a tricky thing to do because if I knew, if I wouldn't say something myself, the press would do that for me. It would be a scoop. It would be a leak. So in some way, I really like being transparent and honest about everything I'm doing. And I thought it's better if I say myself my version of a story rather than it being traveling via a broken telephone.
Speaker 3
We hate that. You could have called us.
Speaker 1
I'm very happy to call you whatever else is my life. What I have for lunch, I can ask.
Speaker 4
Perfect. I mean, it was a good exit note. I mean, you really gushed about your time and it was a, it was a significant amount of time through like so much change. What was that like seeing it for the last seven years? Honestly, every year I've been at OpenAI, it's been a completely different company.
Speaker 1
The hyper growth of the company itself and of the AI world around it. I don't think there are that many stories like that in the history of the world at all. the speed of the growth and a way of change. I think it's a really, really unique thing. And what I was able to see, I'm very happy that I could live through it. As I mentioned, every period, every time was completely, completely different.
Speaker 3
How many people in 2019 is like 30 people, maybe?
Speaker 1
Uh, yeah, I think, I think it was, it's around, it's around 30 people.
Speaker 3
And what's it, what's it at about many thousands now, the latest count.
Speaker 1
It's hard to keep. They're probably not able to count everyone. Uh, but it's a pretty, pretty significant company. It's a big company, many offices global. uh, presence all around the world. It's, it's, it's hard to find people who haven't heard about OpenAI right now. And when I, when I, when I joined, it was a few groups of people doing their own small research projects and, and, and, and small pods trying to, Still, the thing that that remained open AI was the ambition and always aiming for AGI, always wanting to transform the world from the very beginning and make this positive impact. And I think the company really, really successfully managed to do that. And the amount of kind of intelligence being distributed and usefulness to people through ChatGPT, I think it's a pretty, pretty, pretty amazing thing.
Speaker 3
All right, what do we do at Core Memory? We cover innovative, fast-moving, forward-thinking companies, which is why Core Memory is sponsored by Brex, because Brex is the intelligent finance platform for many of these companies, 30,000 companies from startups to the world's largest corporations. rely on brex's technology for their finances they've got smart corporate cards high yield business banking and expense automation tools that are fantastic i hate doing my expenses and brex's ai's and software run right through those expenses figure out where we're spending money and take care of so much stuff for you so you don't have to waste your time on it yourself Go to brex.com slash core memory to learn more and just, you know, get with the program. Let's get going. Let's get out of this archaic finance software and move toward the future. Core memory and Brex. So what happens? You issue a tweet like that and does every foundation lab on earth immediately contact you?
Speaker 1
Many do, and I'm in the process of figuring out what is the next step. I have, being in this industry for so many years, I already have many connections and know many people around. And in a good perspective, I'm not really in a super big rush to figure out what do i what do i want to do next i've been i've been working pretty hard for many years and even have that much time to meet people and talk to people so now i have a chance to to take a bit of a moment to figure out how do i want to spend the next seven years but i do have a lot of conversations yeah
Speaker 3
Well, in your tweet, I mean, you said you wanted to be able to do research you felt like you couldn't do at OpenAI. Yeah. Can you explain that a little bit? I mean, the origins of that and what that might There are multiple aspects of certain things that are just hard to do in a company
Speaker 1
that has to compete in an extremely, extremely brutal and demanding race for having the best AI model in the world right now. And... One dynamic is there is naturally how much willingness of risks companies are willing to take from the perspective of trying to not fall behind. User growth metrics, paying for GPUs is a very, very hard thing to do. And because of that, like showing strength and having consistently the best model is really important for people. all major AI companies right now, but that does affect somehow your appetite for risk that you are willing to take. And another set of trade-offs that are hard to do is companies have org charts, and org charts often determine what kind of research you can be doing because every team needs to have some kind of identity, some kind of research scope, some kind of set of problems that they work on.
And cross-organizational research often is Very, very hard to do, which is something I'm not sure yet fully solved problem, how to organize research on a larger scale very well. Research likes dynamism, research almost, you can say, likes chaos, but large groups of people need some order and need to have a structure and need to have an org chart. And because of that, I think like shipping your org chart is a very, very well known concept and almost all companies do it in terms of research as well. You are, you're shipping your org chart, you are, you're doing the projects that's, that people are set up to do.
And at the same time, I think I wanted to do some research that are, that the OpenAI org chart didn't really easily, easily support.
Speaker 4
Is this like a new breakthrough that we're about to see from JT? I think every researcher in the AI world wants to work on the next breakthrough,
Speaker 1
and I definitely count me in that group as well. And so, I mean, I've talked about this with Mark when he was on the podcast.
Speaker 3
I mean, everybody's coming to him with their ideas and to Jakob. And, you know, I guess OpenAI has a history that's paid off for them well, which is taking risky ideas and sometimes doing things the other labs weren't. But... I know everybody... I'm sure you have so many smart people. They're all coming with all these things. And at some point, it's a company with limited resources, although quite a lot. But it has the big choices, right? And so, yeah, that must just be a really difficult process. So you kind of... whatever you're thinking about is, is it must be, um, quite like a novel path or something. If it's, if it's, if it's like, we don't know if we want to go in this direction or if we can afford to do this right now or not. I,
Speaker 1
I think the, the Ilias concept of age of research, it's, I'm not sure if it's as binary as, as he said, but I definitely think there's much more to explore in the world of AI and machine learning than is currently being explored. We settled on a transformer architecture six years ago, and people have been scaling transformers for quite a while, and it's going pretty well. There's a very clear path to every quarter trying a better model, with slightly more compute, with slightly more data, and we see there's no really stopping of the progress on that. But is that it? Is that the last thing? And I am pretty sure it is.
There are many, many ways in which we can keep improving our models that are not... being done yet in many ways uh as you have introduced already i worked on on on reasoning on scaling up reinforcement learning and before that was a thing people were really betting big on scaling pre-training of transformers and it's one way of scaling and it's been going well and also every time you pre-trained a better model it is more capable although all the evals improve So you can say we should just keep scaling pre-training and we'll get better and better model. But suddenly a few people, dreamers, researchers believe there are more things we can do than just scaling pre-training.
Let's prove that scaling reinforcement learning on top of those language models to a also pretty comparable to pre-training amounts of compute, will be able to teach those models things that just by pre-training alone, you would never have done it. And thanks to that, we have all those amazing agents today automating work and solving problems that just pre-trained models would take forever. excruciating amounts of computes and data to to be able to do so you you invent new ways of scaling and you get completely new capabilities that would that going for the pre-training scaling law would take forever and like this this was a pretty big jump i think reasoning models since since basically introduction of gpd4 were the major uh major capability jump and i i just think there's there's more things like that and i i always think researchers should not only think about incremental improvements, but also how to, how to completely turn the game upside down.
It was Ilya at NeurIPS last year where he was like, you know, we're running out of data.
Speaker 4
There's going to be an end to this. And the idea that pre-training is in its struggle era is like, I'm wondering what the next breakthrough is. And that's basically what you're touching on here, right?
Speaker 1
Yeah. And I don't think it's exactly saying that the pre-training is ending and there's no more. Pre-training is still improving and there are many ways which it can be, but it's not the only way to improve the models. And other ways may just improve the models faster. Scaling pre-training just improves the model very, very slowly on a lot of things. It does something and the models are better, but there might be more.
Speaker 3
There's a funny thing about Silicon Valley, which is a lot of times tech companies are coming up with these very original ideas that the rest of the world thinks are weird. And that's how you build these new interesting businesses and science and research is to chase the direction other people are not going in. And then once something blows up, it goes the other way, which is there's this massive consensus. And then people get a little... we're all doing it this way and now we're racing on how we, we do it this way. I mean, that's what you're describing a bit. And, and do you feel, um, is this, is this like a, is this going to be a problem now that we've gotten into this model race and we've been in this for like two, three years? Um, yeah, I mean, it is like every, are all the major labs getting more conservative and, and is this a broad problem?
Speaker 1
I am definitely extremely, extremely sad that all the AI labs are trying to do the same thing as OpenAI is doing. OpenAI is clearly a very successful company that got a lot of things right and taken the world to the to the paradigm of scaling up transformers and also has proven that scaling and scaled up machine learning models can do a lot of really great and useful things to the world. But how many companies doing exactly the same thing do we need in the world? I don't know. The competition is clearly good, so it's probably better than one. But we have, I don't know, five pretty serious scaling AI companies right now doing exactly the same recipe and trying to build maybe slightly differentiate product on top of the same technology.
And maybe it's the right thing to do, but I would like there to be a little bit more diversity, a little bit more differences in the models. If you look at all the best models right now in the world, Very few people actually notice the differences between them. I think there should be more blind tests of people talking with one model and the other. And are they able to distinguish? Are they able to tell which model is which? I think 99.9% of users would not realize. Those models are and feel very much the same, even though there are different teams doing slightly different things. All the labs are thinking, we are doing this slightly better, and the other lab maybe is better in some other technique.
But in the end, if they all end up so close together, where is actually the exploration and where is the ability to innovate, to differentiate yourself versus the others?
Speaker 3
I don't use them for coding or for like hardcore science research, but I do, you know, I'm always, I'm more in the world word area. And then it is funny. I feel like a wine connoisseur or something at some point. I think I could tell the difference, but it's a, it's a narrow difference, but it's like, I'll use Gemini if I'm trying to do this one type of thing. And then in Chachamiti, if I'm trying to do this other thing and Claude, and they, they all write like, slightly differently. Yeah, I was going to say for me, it's the personality that I noticed the most. Like I did just move to Claude Foley recently just because I like its personality quite a bit.
Speaker 4
It's not like overly verbose or, you know, gimmicky or trying too hard. It's much more direct and I cannot get Chachapiti to change its language that way. But I agree that most people in the world have no clue. Like they wouldn't be able to tell.
Speaker 1
Yeah, yeah. Such a stubborn machine. Consider yourself a high taste tester.
Speaker 3
Can I ask a semi-difficult question from this? I mean, look, I've been spending all this time at OpenAI. You're a legend inside of the company and out. You've got a pretty good track record of what you work on turning into a success. So is this a bad... Sign what does this mean for the company of, if you of all people are like, this is what I want to work on. And, and I have no idea if the company said yes or no, but they at least made it difficult enough for you not to want to do that, that there, I mean, for, especially for a company that did start out as a research lab, um, is, well, I don't know, is this a bad sign?
Speaker 1
I think. Sometimes people grow up to the situation where they think their roads need to separate. It's a very important thing for a single company and whoever is in that company to be aligned in some ways about what their goal is and what the path forward is. And at some moment, I think like my feeling of what the path forward is diverged at least in some meaningful enough way towards OpenAI chosen, what the path forward is and how the next year of research should look like. And in some way, I think it's better to separate and really force yourself to work together for some disagreements that may keep them festering.
That's why I believe that it's much better if different companies are doing the same things, because in some way, it's an extremely important thing for a company to be focused. And it's very likely that OpenAI is doing all the right things. Maybe I just have some dreams that don't make sense. And I am too much of an optimist about what other things can be done. And that's very possible. And thinking about we need to focus and we need to do our core things very well. A lot of things, a lot of companies need to do that to survive and to go to the next stage. So that in a good world, there should be a lot of different companies doing a lot of different things and specifically researchers for whom it is very hard to work on things they don't believe in, should need to find a place where they feel the research they believe in the most is that the history should prove.
over time which one which one is the best that's why i am again a little bit sad that all the companies are doing the same thing because these days if you want to do something different than ml mainstream it's very hard to find a place to do that and that's that's the only thing i am probably the most sad about and you're like still figuring out what you want to do next what this looks like so if all the labs are doing the same thing you don't see yourself like jumping to another big lab
Speaker 4
right
Speaker 1
I'm definitely thinking of what are my next stages, but if there were more options of doing something slightly different than the mainstream, I would be happier and it would be easier for me to choose and find that place.
Speaker 4
What do you think it takes for us to deviate from the mainstream? Because I can imagine being a company that is spending so much money and so many resources and is so in the spotlight, they're afraid to take these risks, but they really might be necessary. So what has to change? Or will it even... that's that's the interesting thing i really like risk taking and i've been described that way and i think taking risks is good but at some moment when there is a lot of money on the line you know there are very very few people who can and
Speaker 1
are willing to take that risk and Like personal risk appetite is a very, very, uh, very, very specific and very, uh, unique thing to each, to each individual, what they are, what they are willing to do. And I've, I've, I've worked with a lot of people over time and I, I really think people should be willing to take more risks and try a little bit more things out there. I think researchers also for the things like the compensation in the AI world is being pretty, pretty crazy these days. But I also think sometimes it makes people really not want to lose their job, really not want to have a bad performance cycle. And because of that, want to follow maybe short-term paths of gains and sometimes those are very brilliant very well-meaning people it's just just somehow sometimes the system is set up a little bit to think too too short there and and specifically researchers should really be encouraged to take risks and to make make bold bets because that's how that's how the progress is being done
Speaker 3
So we know about a couple in some ways, I guess. I mean, John Carmack went into his cave in Dallas. I think he was solo for a while. I think he has a few people working for him now. And he was just kind of like, he said, what you were saying a couple of years ago is, I don't know if I can do something different, but I think somebody should at least be tinkering away on a very different approach to this. I don't know. I've talked to Ilya, but I don't know much about what he's actually working on. So I have no idea if it's kind of a continuation of where he left off or if it is some radical, um, I mean, I'm sure he thinks on some level, it must be some different approach because otherwise, um, why raise all that money and go try to do this.
And then, and then Jan LeCun seems to have a different philosophy, but I was always wondering, I mean, is this like a field, you know, it's funny to me because in AI, In some ways it's old because it goes back decades, but this flavor of it is relatively new. When I talk to researchers, they're still like, you know, if you read all the major papers, you could get up to speed pretty quick these days. And then... But so sometimes I just wonder if there is going to be this person who's just like comes out of left field with some just radical new idea that pushes everything forward. Although it seems harder to do now that you have to have a state sized data center to like, that's a power thing.
Speaker 1
That's the thing that makes it really difficult. And it's another interesting problem that would be good to figure out how to solve. Because there is a lot of research happening in academia in the world. There are a lot of students doing a lot of things, but most of them are heavily under-resourced. which makes that research ultimately go kind of nowhere because the real research you want to do is the research at a large scale. And that's one thing that makes me very happy, very hopeful in the world. There does seem to be actually a good amount of funding going around, trying to go for new interesting ideas, which some people are able to do, like John Carmack, like Ilya, like Jan.
Those are exactly the types of things that should be existing right now and that should be funded. And not all of them will succeed, obviously, but some will. And that's how innovation happens in the world. For every reinforcement learning researcher, this concept of exploration versus exploitation is very well known. is even if you were optimizing our agents, there's a very big trade-off. Should those agents go for things that have been proven already to work, to the paths they know how to succeed? Or should they try to do something new and try to solve old problems in completely new ways? And it's a trade-off.
It's not an easy thing to do, but it's something that's that is studied and needs to be studied. And as much as we think about our agents doing that, we should also be thinking how we are doing this. In the small club of super, super elite AI researchers, do people know what Carmack is up to?
Speaker 3
Do all you guys know?
Speaker 1
Unfortunately, I do not know. As much, maybe okay, maybe I do know something. I think he is... betting on very end-to-end reinforcement learning on computer games via mouse and keyboard. I think there is something like that, if I recall correctly. I'm going out of memory, but it's a very interesting thing, because one thing I've been thinking for a long time, I think video games are actually one of the most interesting environments to be training intelligence, because As the games are made to be interesting, they are made to be interesting for a human brain. And being interesting for a human brain, there are some storytelling elements and elements of power fantasy, but there's also a lot of just problem solving.
And it has to be interesting, engaging. It cannot be repetitive. There is something where computer games are very much tailored to human intelligence and are, in an interesting way, teaching a lot of things about resource allocation, about how to solve puzzles, about how to win various types of games. And it's exactly kind of what you want your agents to do. We still haven't figured out the really, really smart models to operate at a high frequency of this multimodal environment. There are some architectural limitations, probably. But I think training AIs on video games is an excellent thing to do.
Because Richard Sutton used to do a lot of stuff around not just video games,
Speaker 3
but poker and things like that. And then I've been in his lab. I mean, he had more rudimentary video games than where we ended up with like OpenAI with Dota. And it feels like Demis has always been chasing this idea. So, I mean, it's interesting that you bring that up because it's kind of like a... An old idea where the labs for a little while were like, we're just going to keep doing the best game and show off. And then it sort of fell out of favor in this, in this more like chat GPT era. And then you're saying maybe there's promise there after all, and what we should kind of like revisit this. There is a common thread in science that the good ideas have a tendency of coming back up.
Speaker 1
It's much easier to predict ahead of time. What is the idea that's important? It's much harder to figure out a good timing. There are a lot of things, and even at OpenAI, in the early days, we used to say, we can't say that the particular method doesn't work. Maybe it doesn't work yet. Reinforcement learning on games, when I was starting at OpenAI seven years ago, it was a big thing. Everyone was doing it. It was giving pretty good results. We were able to solve games. We solved Dota, we solved StarCraft. AlphaGo is a very famous moment. But those models, that was their clear and obvious issue, they didn't have that much world knowledge.
They didn't understand our world. They were trained from scratch only for playing that game. And that is clearly not the right thing to do. We need to first teach the models about our world. We need to teach them about various higher level concepts than just reacting to pixel. The reinforcement learning from scratch is very much like a kind of monkey brain or a lizard brain trying to do stuff. And we want the models that have slightly higher level concepts of things. But after scaling up pre-training for many years, we can... train a really, really good set of representations about the world around us.
And now we should use that. That's the whole magic of the reasoning malls, that they are building the tower of reinforcement learning on top of a very, very strong fundamentals on knowing the world from pre-training. And that's how things should be going forward.
Speaker 4
And the world models, I mean, Google's done this. And isn't that sort of what Yan and Fei Fei are after as well, which makes sense to me. Like that's it makes sense that that's how we learn about the world. We don't just like live in a black box as babies. We, you know, feel things out. So you're sort of bullish on that idea.
Speaker 1
That idea is very, very clearly the right thing. The interesting concept is how do we combine the representations built from world modeling together with something like reinforcement learning. You do reinforcement learning to teach the model skills, which then you need for the model to operate in the real world. You need some ability to achieve your own goals. And that's what reinforcement learning teaches you to do. But to be able to achieve your own goals, you first even have to understand the world around you to then have even good ideas how to achieve your goals. That's why those things need to be coming together. And I think doing reinforcement learning successfully on a very well-trained world model, whenever someone makes that work, it will be a very, very happy moment. Since day one, the Core Memory podcast has been supported by the fine people at E1 Ventures.
Speaker 3
They are a young and ambitious VC firm in Silicon Valley, investing in young and ambitious companies and people. Thank you so much to E1 Ventures for all your support. That's totally put you on the spot. I mean, just for these things that you've been pursuing, I mean, is it possible to give us hints about what you've found interesting and what's been calling to you? Or is that going to reveal your new startup?
Speaker 1
I think... I think there are two directions of research that I am the most excited about right now. And mostly from a perspective of... I don't think it makes sense to do the same thing that the labs are doing. I think there are still a lot of things that can be tweaked and improved in existing setup. But there are two directions I think are... under-invested or at least not done and appropriately resourced. Number one is architectural innovation of a kind. I think we are a little bit too hard set on the transformer architecture and it's a great architecture. It's been very heavily explored and people had a slightly hard time locally improving transformer and trying to figure out small tweaks to the transformer that will do better, although There were some that were pretty successful.
Obviously, Sparsity is extremely successful. Various ways to make attention cheaper also have been pretty successful. But is Transformer the last architecture to do all of ML? Pretty clearly not. As much as the set of authors of the Transformer did an amazing job and did define ML landscape for basically 10 following years, But I think there is more. And I think there are other ways to train large models that don't look like Transformer. Maybe look a little like Transformer and maybe don't look at all like Transformer. And I think it's a problem something would like to do. And if no one else does it, I will roll my sleeves and try to solve it.
The second thing, which is a much more popular problem, but I don't think anyone does it very well, is continual learning and figuring out how to really, really bring test time and train time together. Very clearly, we as humans operate that way. We don't have a separate learning mode and mode of answering questions. Everything is happening continuously and all the time. And I think our models should look like that a little bit more as well. probably one of the last few important capability elements before we can meaningfully call our models AGI. If they don't learn from the data they see, they still feel kind of limited and dumb.
Speaker 4
Speaking of AGI, the last time we were potting, I was saying that I don't hear about timelines as much as I once did about a year or two ago. It was way more hyped up to talk about timelines. And even talking about AGI, it's sort of fallen off. So I'm curious. You call yourself a cautious optimist about AI. So where do you think we are on the AGI timeline? Yeah, my personal take here is I have slightly updated in the perspective. I always believed scaling up reinforcement learning is a necessary part of getting to AGI.
Speaker 1
And a year, maybe a year and a half ago, I was dead set. We scale RL on our models, it will be AGI. And I definitely had to update a little bit about it. But some of the things you only see after you are at that next stage, and we have to admit today's models are very, very good in many, many ways. And the fact what they can achieve in coding, Vibe coding for me is favorite hobby. And you can code so many things very, very quickly. For some of the people 10 years ago, if you show what we have today, they would call it AGI already. So I don't think talking about AGI is that out there and that crazy anymore.
But at least for my definition, the current models are not AGI still because the continual learning hasn't been in any way integrated with our models. At the same time, from where our models are, also even things like multimodal perception. Can we call the models AGI even if they are very good at text understanding and very good programming, if they don't see the world out there or cannot watch a video and understand it well? So there are many problems that I would call necessary steps to really achieve that civilizational milestone of building AGI. And for a while I was thinking that if we really try hard and if we do really everything well, maybe 2026 will be at least the year we achieve really good continual learning and really general reinforcement learning.
I think my timelines are still a bit shifting, but at the same time, the field of AI is moving pretty quickly. The investments are only cumulatively growing year by year. More and more people are getting into the field of AI, which increases the talent pool, which increases the number of ideas we explore. So I don't think the idea is completely wild and completely out there. Maybe it will be slightly earlier, maybe it will be slightly later, maybe 26, maybe 27, maybe 28, maybe 29. I don't think it will take much more than that. But there is work to do, but people are working hard on AGI.
Speaker 3
I mean, what you were working on, unless my memory is wrong, I mean, back when before Strawberry was a thing, I mean, it was kind of like the Q-Star thing, right? And so it's getting talked about during the coup as like this, Ilya saw Q, people saw Q-Star and... And, you know, they knew the AGI was here and it spooked everyone. I mean, is it just hearing you say what you did now? It is funny how, I mean, it is a tricky thing. It's like these things do incredibly impressive things and then we get... super frothy and then time passes and you know like now strawberry was incredible and it like changed the whole field but i don't think it like scared the shit out of me the first time i i used it you know what i mean i know what you mean an interesting part of both human psychology and in a way how we how we interact with
Speaker 1
technology It really, to me, still, effects of scale-up reinforcement learning have been pretty significant, and I think even more we'll see over time. Specifically, first and foremost, from applications to programming, which will affect our life in many, many different ways. Doing any large-scale programming project today versus a year ago is a completely different ballgame. And we will see impacts of those in various things. In the place where me and my team and many people at OpenAI were two years ago, when the first signs of QSTAR working were, um you are sitting in a room and you are seeing something that is a meaningfully new technology and if you are not a little scared a little worried having a little bit of a moment what are the consequences of this world of us doing this then i don't think you are treating your job very responsibly I think every AI researcher should be thinking, if I am doing something, it is something new.
It has new capabilities that were never, ever demonstrated before. What is happening in the world? And many researchers do. Sometimes it's easy to take a step or two steps too far. So far, the AI hasn't done anything meaningfully bad to the world, although maybe things like sycophancy can be debated. Or that we know about. Yeah, exactly. But still, I think it's a very, very good and very healthy response of being worried and cautious whenever you release new technology to the world. I think we are living through the times of change, times from diffusion of a lot of new things to the world. And they will have impact.
They will have impact how people spend their lives. They'll have impact how people think about themselves, about others. They will have impact on interpersonal relationships, on international relationships. They'll have impact on the GDP and on productivity. And like sometimes someone writing a line of code has done things to cascade through all that. And the weight is pretty heavy of that.
Speaker 3
That is all super legit idea. And I was. I think we've talked about this a little before, but I was always trying to put myself in your shoes as the coup thing is like, you know, being unpacked where Kylie and I are like spending that week trying to outbreak stories on each other. And then just to see this thing that you created now being... talked about in the press and it's getting interwrapped in what's essentially like a soap opera at the same time. Was it, I don't know, funny is the right word, was it weird to see your creation turning into this object of such a fascination in this moment before anyone really knew what it was? It is hard to disentangle the world of technology and concepts and human emotions
Speaker 1
and human lives and human agreements and disagreements. We clearly live in the world where the important players in the AI world have network of relationships that is very complex and on many, many levels. And untangling it will take historians many, many years or decades fully to understand what actually happened here and what was the case. I honestly myself now have only very spotty recollection of everything that happened through the coup. And we are learning a lot of things whenever new depositions happen, whenever New documents are being unraveled. Some new facts are coming up. And again, someone at some moment in the future will surely dig out all of it. But the world is complex. um and again like there probably is a healthier way to to talk about technology and and and find a good discussion forum where those disagreements should be fleshed out in some way but we live in the world where there's no solution there's no no perfect way and that's that's how you don't you don't think x is like the ideal
Speaker 3
medium I personally really enjoy posting on X and like sharing my thoughts with the
Speaker 1
community, with everyone around, but it's also not a completely serious place. So there's always, it's always half joking, half serious kind of thing, but... What's the right solution to? If one person has worry of something is too dangerous to continue working on it, someone thinks, oh, maybe it's the right thing to do because it will increase capabilities. The first person thinks, oh, this is not even the right research because we should be doing some other research. And in a world of technological progress and research, all of those things are unknown. No one knows the future. There are only thoughts, convictions and dreams. And with that uncertainty, we need to live and we need to somehow often agree to disagree in a lot of ways.
Speaker 4
Yeah, and given the media attention to Qstar at the time, it was like, what did Ilya see, all that. It's just so much hype, and it feels like it's only gotten more intense month by month. And I'm just curious what you think, because a lot of us on Twitter, we're both very much on Twitter, and we both have partaked in the hype. So what do you think about the hype? Should it die down a bit? I personally think that we should dial it way back. Yeah.
Speaker 1
I know that you do, but at the same time, if someone told you seven years ago, OpenAI will be a trillion dollar company building data centers on a scale of the largest infrastructure projects ever, having one of the largest web products in the world, everyone will be talking in AI all the time, you would think those people are crazy. This is hype. I kind of genuinely think the hype is in many ways substantiated. And sometimes it overshoots, sometimes it undershoots in many ways. But AI is important. AI needs to be talked about. I don't think anyone right now believes that AI is an unimportant topic. I think whenever that was a few years ago, there were definitely many people thinking that way. It's clearly determined that AI is... probably one of the most important topics in the world to be discussing and thinking about it. How quickly the progress will go here? What are the right paths? How safe or how dangerous this is? Those are all those things that can be disagreed and debated, but AI is here to stay and only be better.
Speaker 4
Totally. But like taking the technology out of it, I mean, I even covered the MSL poaching spree. Like it's become more of a soap opera and a reality TV show than about like the hard science. And you've, I mean, you've been doing this for so long. I'm just curious, like, have we sort of overstepped into reality TV territory?
Speaker 1
But who is making that soap opera? That is the question. It's not me for sure. I'm like old enough to have kind of lived through the dot-com boom and like these
Speaker 3
different eras. This one is a bit more soap opera. We didn't have, you know, in the great productivity software wars, I don't quite remember things being like this. I mean, it's just like... I guess because the stakes are so high and the amount of money. I mean, the trading of researchers between labs and then the crazy drama that is on top of all that. This all does feel... This is what struck me about the whole thing this entire time is that I feel like San Francisco is in a... we've almost like set this, this own reality, not fully in like a bubble sense, but of like, we've all talked ourselves into, this is the final product and the stakes are so high and it's the, you got this race and, and it could be great.
It could be terrible. Like everything is so charged, supercharged that, that I feel like it comes with all this extra baggage. Right. And, and, and, um, so I think, I think it is sort of like different. I remember the.com boom was really just, sort of started from this place of like, this is cool. All the world's information is available. We can contact everyone. And then we're kind of like making startups. It just turned into this like blatant money grab because it's kind of like somebody's going to get rich off this. Let's figure out who, but it didn't, it didn't feel like the weight of the world on all this stuff, you know?
I don't know how you guys actually live through this because I'll like hang out with you and might be at like anthropic or something like that. And then I see everybody working so hard and you're in this like race against each other. And, and again, the stakes are so high. I mean, it must be to have done this for seven, eight years. It's just, I would understand you taking some time off. It must, it must be like exhausting and just sort of psychologically, um, if you, because like, if you do buy into that, it has to take a psychological toll. it definitely does but i can tell you i heard at some moment from someone even more experienced than me of dealing with stress that's jerry it's like a push-up every time you go through a hard stressful moment you get a little better at dealing with it and honestly i think i've learned extreme amounts of mental and emotional
Speaker 1
resilience through seven years of doing it i At least I really feel I can disconnect from a lot of noise and a lot of BS and really try to stay constant and steady in the face of whatever is happening, whether the company I'm working is imploding, our researchers are moving, our projects are reallocated. There is... There's always something interesting and I've heard someone compare the talent poaching to sports team transfers. And the thing that works well for the sports team is that there are roles and there are some times when they transfer and not. And unfortunately for the California law, it is not really possible here. I think it would be good to have some rules for it because I think there are some people in the business that are changing jobs more often than actually do any work in the meantime. And that definitely is happening as well.
Speaker 3
How about AI salary cap? Yeah. But there are some people who are doing that and there are some people who are
Speaker 1
doing the work still and trying to push the frontier forward. But AI is a big business and there's no way denying that.
Speaker 3
Let's tell Kylie, we need to come up with a list of all the people who have worked at every frontier line.
Speaker 2
Yeah.
Speaker 3
How long they stayed in each one. There must be at least a handful of people who made the entire tour of the bay. God, it's crazy. Can we talk about the Polish mafia for a minute?
Speaker 1
Please do.
Speaker 3
I mean, when I first started working on this book and I got to like, we're talking about 2018, there were 30 people or something. It was like this incredible number of this original group of people came to Poland. They were math gurus. Some of you guys knew each other growing up, some didn't. But yeah, I mean, I guess what, this is like a reflection of the old... soviet education system just being so fantastic and having having all these math-minded folks and and or it's just that like one person went and you guys all knew each other and and ended up at openai so first and foremost i personally didn't know anyone from openai before i ended up at openai completely completely
Speaker 1
randomly but Especially in the early open era, there was definitely high representations of Polish people. I don't think it did survive the test of time. Right now, there are a few more Polish people than in the early days, but not that much given how company has grown a hundred times. But like there is something to that but i haven't experienced other education systems to really tell you if the polish education system is that great there are we definitely have a lot of brilliant people and what i really like about poland and i think what is also a testament of the Polish economic prosperity and development over the 37 years I live on this planet is that Polish people are pretty hard working, which is kind of a more and more rare thing as the time progresses.
And as in a lot of developed countries, you can see people slightly less put emphasis on working hard. Life's too good. my life can be very good and there are there are many ways in many in many other things you can emphasize and prioritize in life obviously but but people in poland really care and um the transition from uh before before i was i was born poland was still a communist country just just the year i was born we we transitioned to a free market economy it was something a little bit brutal but the polish society really embraced this change and try to figure out how to become entrepreneurial how to that you have to work for your for your future and for and for for for building uh prosperity and you know it works every i i'm i'm an expat i don't live in poland anymore but every time i come which is once or twice a year i see my country built up.
I see my country better. I see my country more beautiful. I see my country more prosperous, which is a really, really great thing to see. It's an amazing, amazing story.
Speaker 3
Are you like a celebrity?
Speaker 1
Not at all. I feel like the Polish government must be like, fuck, man, we could have done this whole thing.
Speaker 3
We should have held on to like all these guys. Yeah. I went to Poland last year. I know they're aware of this and everyone, everyone's like, do you know Wojciech?
Speaker 1
Wojciech is such a, such a great guy, such a, such a sweet person. I think there are a lot of things still that are hard to replicate anywhere else beyond, beyond here. Silicon Valley is completely unique and American ambition, American scale, and American dynamism is not something that you can really easily do anywhere in the world. Yeah. But I can assure you, Polish people are very hardworking and know bullshit, and that gets you pretty far in life.
Speaker 3
Um, the, if you, are you, are you impressed, surprised or not about Google's kind of comeback here? You know, I mean, it seems like they've gotten an awful lot of things right. Did you guys always just assume that would probably be the case? They'd figure their shit out and, and, and do this or, or has this been a surprise? uh personally what i think you should consider google's comeback i think it's openai's fumble i think openai as much as this company does a lot of things wrong i think it stumbled a few times and executed more slowly than it should have in an ideal environment if you are a company that is ahead and has all the advantages
Speaker 1
that openai has you should always stay ahead But if you make wrong decisions along the way and someone else makes the right decisions, which Google did a lot of right things, then you have to understand Google has a lot of advantages in terms of hardware, in terms of talent. When OpenAI started, Google was clearly number one in almost all machine learning aspects and research directions. And the way how OpenAI was able to differentiate itself was mostly about research conviction on a particular direction, a particular path. And it took surprisingly long to the world to realize that this is a good conviction and this is a good bet.
Because even when GPT-2 was trained, when GPT-3 was trained, when GPT-3.5 was trained, no one really cared that much. You went to NeurIPS, you talked to a lot of researchers, OpenAI was already kind of cool. But other labs, well, yeah, we'll replicate at some moment. Those language models are interesting, but that's kind of it. Only when OpenAI started making money out of chat GPT is the moment where other companies were, okay, now it is profitable. We really need to do it. But it gives you an enormous amount of time from starting to build technology to actually commercialize it for the others to, okay, we really, really need to do it.
And very clearly, Google started treating seriously at that moment, training large language models and like through OpenAI fumbling its lead, they are very, very close now in capability and in terms of models trained. So great for Google. Congratulations for everyone who managed to turn those teams around and make them execute well. That was definitely a lot of great work.
Speaker 4
MELANIE WARRICK- What were the fumbles? I'm trying to rack my brain. Because I remember covering when you guys launched Search, when OpenAI launched Search, Google's so dead. And I was like, I'm not sure it is. But what fumbles? I don't want to talk too much about internal decision making and what were right
Speaker 1
decisions, what were wrong decisions. But again, I think in a good execution scenario, if you start ahead, you should stay ahead. But you do think there were technical mistakes because there was a lot of corporate
Speaker 3
drama that slowed things down on some levels. I've talked to enough people at OpenAI of just dealing with how this company was going to move forward. And then there would be periods of time when a whole bunch of people who were important would leave. But it sounded like you were talking more about technical things that you think...
Speaker 1
Those things sometimes are related. Technically, I don't think people leaving should be that much of a problem, but sometimes the symptom of a problem, I don't think it's a problem in itself. In all of companies, people come and go, people leave and it should be a regular course of business. But if there is something that people say, hey, someone is doing something wrong and that's why we don't believe in this company anymore or we should leave, then there's probably probably something there. But yeah, as I mentioned, I think there clearly were certain things that were done slower than they could have been in the meantime.
Speaker 3
If the large labs, as you've argued, are doing the same thing for the most part, and then Meta is kind of a latecomer to this in some ways. I mean, obviously, they've been in AI for a long time, but clearly made a bet that they wanted to do this a different way, and then we're poaching talent from all these other... I don't know enough about exactly what Meta is doing, but it doesn't strike me that they're going out on a different path. They're just trying to follow the same path. That just seems like a fundamental... I mean, that just seems... I don't see how they can end super well. You're, you, you're just kind of late and doing the same thing as everybody else. Um, or do you, do you have a feeling that they, they do have a different approach to this?
Speaker 1
I'm not super, super familiar with their, their strategy, obviously. So I, I don't know, but it's like my, my impression from the outside is that they realized that. scaling up, you can take two different approaches almost of what you want to do in the world of AI right now. You can either say, I want to have different models than everyone else that will be better in some ways, and this is what I want to build to the world. Or you can say, I want to have as good models as others, but I want to use them differently or I want to build different products. As far as I understand Meta as a company, it is about trying to connect people and build personal relationships and build relationships. things for maybe metaverse, maybe social networks, maybe experiences. And again, I'm hypothesizing here, but I think the idea is to use the AI technology and the transformers that we kind of know and understand to try to build some of those things and some of the new experiences. That's my understanding. It's like, you know, from a perspective of a extremely profitable and a big company that has the largest social network on the world, it can be, it can be a very good strategy.
Speaker 3
Was there ever when you're, you guys are competing all the time, obviously, I mean, and we talked about Google making a bit of a comeback, but I mean, was there ever when you're having your meetings or you guys are sitting around some other lab that really impressed you the most?
Speaker 1
I must say it's a kind of recent development, but the amount of me being impressed by Anthropic has just increased tremendously over the last year. I never was a person who pays that much attention to personality of a mole, although I've heard that close personality is great, maybe.
Speaker 4
Pretty good.
Speaker 1
I don't know. But what they did with coding models and coding agents and the brand they're built for that and the amount of happy developers they have is something absolutely, absolutely, again, like amazing achievement and they they did that they also started from behind of open ai more compute constraint smaller team but they have a focus they have a they had a great execution they had a lot of issues with procuring good computes good good kinds of compute and they they managed to build products that are that are absolutely excellent that are changing the way how people both develop software as well as, as far as I understand, improve meaningfully enterprise productivity. Great, great, great job. Congratulations.
Speaker 3
They do seem to be having a moment.
Speaker 4
Yeah. Everyone I know talks about cloud code. And I just interviewed somebody who's keeping a plant alive using cloud. The first living thing being kept alive by an AI model. But I think I don't know how they make something that good that it's so universally loved since ChatGPT is cloud code. And Kylie broke this story on X being distraught that they were getting cut off,
Speaker 3
you know, from using it.
Speaker 4
Same thing happened to OpenAI.
Speaker 3
Yeah. And then Elon, you know, he kind of copped her. He's like, yeah, we're using it too much. This is a wake-up call. We got to get better at what we do. So I assume, I mean, this seems, I don't know if it's universal, but it seems like lots of the labs are, in fact, borrowing from this tool. And so, yeah, I don't know. I just... Hopefully cutting them off means we get better cloud codes in the world, a million cloud codes.
Speaker 1
Yeah. In OpenAI, we are developing codecs for a while. It's our own version of cloud code. I think it is pretty good as well. I haven't, interestingly, I haven't used cloud code that much myself. I was employed at OpenAI, so I didn't really know myself.
Speaker 3
But I was trying to be nice.
Speaker 1
So I really don't know. But I think Codex is pretty good. It's not a bad product. But just judging from a Twitter sentiment, Cloud is really, really loved by the developers in the world. So I think this is fair from based on our previous conversations is like you kind
Speaker 3
of come from this intellectual place and interest in science. You, some of the reasoning work had its roots in you wanting to create AI scientists. And this has been a passion for you for a long time. When I saw your tweet that you were leaving, there was part of me. I was like, you know, I was. I was wondering if you want to stay in this more of this foundational model race or, you know, I hear you, you're more interested in going in a different direction, but I, I kind of wondered, I was, I was, I could see you going off into biotech land or something like that and, and pursuing this in a, in a quite different way. If I could clone myself and do multiple different things, I would really, really love to.
Speaker 1
But like the long story short, at some moment I woke up and I realized I'm pretty happy and pretty proud of what I have achieved in my life. But what I want to do right now is try to take one or two really, really big research beds and try to make them work. I think people should be willing to take risks. And then from what I at least have seen, I am one of the people with the highest risk tolerance to trying wild out their ideas. And I think I should use that for something good.
Speaker 3
When you talk about that, like, I mean, how long would it take to bring something like what you have in your mind to fruition? Is this like a one-year project or what you're talking about is you're saying it's so risky because it's like, I'm willing to put four or five years of my life into chasing this thing that may just not be better than what exists.
Speaker 1
I'm definitely willing to put a lot of time. At the same time, I think people should execute quickly and doing things slowly is no reason to be proud of that. So I would like it to be much, much faster just for the virtue of executing well on the research project. But the important parts really are what I mentioned before, focus and conviction. If you do a lot of different things, very often you do them... Each one, you spread your attention, you spread your resources. And as much as the research labs were often telling the story, how compute constrained there are and how much research is slowed down by compute, I think it is...
part of this story as it's a meaningful part of the story but very much very often it's a lack of focus and in the end you can spend your attention on only so many things around the day and i really like telling researchers i was working with run fewer experiments and think about them harder Because just sometimes even the time, literally hours you spent looking at your experiments, nothing running, just analyzing the data more can lead to breakthroughs more than running more things. So OpenAI or whoever has all this compute and just being spread among too many projects. Like there is actually enough compute there if it was being dedicated to a smaller number.
Again, it's a bit of a risk taking and conviction. If you have three projects and one of them succeeds, then you are good and you have done something good and maybe the two other things you kill. If you do three projects and three succeeds, really, really, really great. If you do just one project, it goes much, much faster because you can focus more, you can have conviction more. And if it doesn't succeed, you are kind of screwed. If it does, you have the best model in the world. And for OpenAI, it is... slightly hard right now to be able, we are focusing all of our company on something new and something different.
And we don't care if Gemini will have a better model next quarter. It's just, it's just a hard thing to do. And, and it definitely takes a specific kind of, kind of person to be, to be willing, willing to try something, something like that. And that's, that's, that's, that's, I think, that's, I think the story here.
Speaker 4
Yeah. And I know you can't talk about the secret sauce and all of that. But I am curious, like, you know, in my mind, the bets that they're going to be making are the ones that make them a lot of money. I mean, the news, the ads are coming to Chajabuti has, you know, sort of broke the Internet today. And I'm wondering, like, what direction can we see them going in? Where are they putting their resources even broadly? Yeah.
Speaker 2
Well, I don't think I should and can talk anything about OpenAI's plans.
Speaker 4
That's fair. Do you think any of these model companies will have the courage or maybe courage is the wrong word because maybe it's a bad decision not to put ads in their models or
Speaker 3
is this all inevitable?
Speaker 1
It's a business strategy and I work on training models.
Speaker 3
Yeah. Yeah. Okay. And sorry, I'm not trying to browbeat you here. I promise. Cause I know we talked about some of this. I just, as we've gone through this full conversation, I am still trying to wrap my head around this a little bit. Like when you're talking about what you want to do and pursuing this other, these new avenues, but, and I know we talked about some of this, that you do have to have all this horsepower. Um, so yeah, I'm, I'm still, and maybe you don't know yet. I'm, I'm trying to picture if this is Jerry and, testing these new ideas on his own, or if you feel like for the kind of stuff you want to do, you, you sort of like have to be somewhere that has, you know, enough oomph to, to make that happen.
Speaker 1
This is the number one question I am trying to understand right now. Every, every AI research still needs GVOs, it needs flops, and I need to, I need to think what is the best way.
Speaker 3
This is Poland's opportunity. They need to give you like the national data center.
Speaker 1
It wouldn't be maybe that bad of an idea. I'm really trying to understand. I know the types of research I want to do, and I'm trying to figure out and understand what is a good way to do it.
Speaker 3
Let's make this happen.
Speaker 4
Yeah.
Speaker 3
Poland.
Speaker 4
Poland, lock in.
Speaker 3
Get yourself a data center and give it to Jerry, man. I think this is, I would do it if I were them, actually. Those were the biggest questions I had. I could talk to you for hours and hours and hours. I feel like I don't want to browbeat you in your You seem so happy. Thank you. You are glowing. I don't want to, I don't want to like, thank you.
Speaker 1
I've heard that from multiple people, Jerry. You are much happier now than before.
Speaker 3
You do have a glow. So I don't want to, I don't want to drag you into what are you doing next? What are you doing next?
Speaker 1
I don't know, but I. I heard from a person who is right now running their own company that working in the OpenAI was more stressful than starting your own startup, which was pretty shocking to me. But it is in many ways a slightly stressful place to be.
Speaker 3
Well, we'll let you go in a second. I did just think of one other avenue. I mean, beyond sort of this, um, everybody's chasing the same stuff too much. Do you see, do you see a massive mistake people are, are making in this, this field or that that's the obvious sort of primary? i don't think there are massive mistakes like it's hard for everyone to do massive mistakes there there's only a question how much do you try to explore other things versus how much do you do you keep roughly doing the same thing what you've been doing yeah i asked that a bad way and maybe this is a question you can't answer because you already know the ideas that you want to chase on your own i i was wondering if there's um A couple things that have come out in research that you felt are like underappreciated ideas that the world hasn't paid as much attention to as maybe it should.
Speaker 1
I think, I think there are, I think, I think, I think, I think there are lots of those to be, to be perfectly honest, but. The most important thing the ideas need is a little bit more attention, a little bit more compute, a little bit more fight around. And I think there's some slightly unique thing that a lot of researchers like doing the work of zero to one. And a lot of academic research has done that, where you create some completely fresh idea, you show that it somewhat works, and then you put it in. And I think what I really specialize with my team at OpenAI and what I think we did a really, really great job is taking research from one to 100, taking new ideas that are different, that we haven't done before, but have been a little bit proven and figure out how to get them to work reliably at scale, training frontier models combined with a lot of other things that also are involved there.
And I think this is kind of what a whole bunch of academic research needs because the proof of concept is cool, but being able to train one of the most capable models of the world with a particular technique requires a whole bunch of very specific work to get there. And it can take you years if you don't do it well, or it can take you months if you have a good algorithm how to introduce things like that. So this is one of the things that I would like to do much more of in the future.
Speaker 3
When we were talking about some of the departures at opening out, you said, well, the company, they should be able to survive these losses. This has been sort of like a star-driven field, though, to some degree. You know, it's Alec Radford does this thing, you know, transform it.
Speaker 4
He just poached back three people.
Speaker 3
You know, and yeah, I mean, clearly you see by the behavior of the labs, they think this is like a star-driven field. thing. And, and I don't know, you know, you're looking quizzically at me. So, but there, I had a, there have been these, clearly there's industry-wide work, academic-wide work, but there have been these moments where these large breakthroughs came from somebody, they, they tied to like a handful of individuals. Um, but clearly you, you seem to like not think this is a star driven
Speaker 1
I think it's a complex subject, but I think both things can be true at the same time that at many times, and you see that a lot of open AI, very small set of individuals have an outsized impact and drive set of completely groundbreaking results. that then go throughout the industry. And I've seen that happen over and over again. At the same time, whenever I see people changing companies, I very rarely see it actually being that impactful on the companies. I I really think the company's identity and almost like a company's way of operation is what is actually the real research engine. Not that much as this particular researcher here or there.
I really see the researcher that hop between companies usually like are not that productive there. Even if they often did great work in the past and they suddenly become a little more preoccupied or need to spin up or... don't have particularly fresh ideas right now. Definitely experience in the field gives you something, but I think it's much more to create an atmosphere of personal responsibility, of being able to explore things, of enabling people to do great things, to do that. And I think it's possible to build a lot of teams that could build great things, whether it's that specific set of people or different specific set of people.
I don't think it's that unique. I think actually a good research structure, a good research culture, a good collaboration is much more important whether a specific person is on your team or not.
Speaker 3
This makes sense. This makes sense. Um, all right. I'm definitely, I'm going to let you go. The, if you, are you, are you a meditator at all?
Speaker 1
I have been trying to do more and more of that in my, in my life. I've been trying to incorporate some of it, but I, I, I think I am a bad meditator.
Speaker 3
All right. Before you take your, before you do your next thing, if you want, I know there's very luxurious darkness retreat in Poland. They have these like sort of hobbit. holes, but on the inside, it looks like, uh, four seasons, but you can't see it anyway. They turn all the lights off. But, but, but if you need three days at this, um, Polish targets retreat, we can, we can get that for you. Then you can, you can like fully reset and come, come back. Text me the details. Okay. Okay. All right, Jerry. Thank you. Thank you. So, so you good?
Speaker 4
Yeah. Thank you so much. It's so funny. My dad also said when I was unemployed, he's like, you should do a meditation retreat. So maybe that won't be the last time you hear that.
Speaker 3
The guy who runs this is great. He's a character. Thank you. Thank you so, so much for coming on here. Hopefully this exit interview was not too painful. Thank you.
Speaker 1
It's been great working with you.
Speaker 3
All right, man. It's been great getting to know you. And obviously we wish you all the best of luck chasing these new ideas. Thank you.
Speaker 4
Thank you.
Speaker 3
Thanks, Jerry. Core Memory Podcast is hosted by me, Ashley Vance. It is produced by David Nicholson and me. Our theme song is by James Mercer and John Sortland. The show is edited by the John Sortland. Thanks as always to Brex, anyone ventures for making this possible. Please visit our Substack YouTube and podcast channels to get more of what core memory makes. Thanks y'all.