On the recruitment wars. I mean, this got a lot of attention, clearly. And it looked like Meta was quite aggressive. What exactly does this tit-for-tat look like? What stage are we at?

Yeah, I mean, there is a pool of talent, right? And everyone kind of knows who they are. And, um, you know, I think many companies have realized that one of the key ingredients, not the only important ingredient, but one of the key ingredients to building a great AI lab is to get the best talent. And, um, I think it's not a surprise that, you know, Meta has been aggressively employing this strategy.

Um, you know, we haven't sat back idly. And I actually want to tell this story from OpenAI's point of view. Um, I think that a lot has been made in the media of, "Oh, you know, there's this unidirectional flow of talent over to Meta." Um, but the way that I've seen it is, you know, Meta, they've gone after a lot of people quite unsuccessfully. So, just to give you context, right, within my staff, within my direct reports, uh, before they hired anyone from OpenAI, I think they went after half of my direct reports, and they all declined. Um, and of course, you know, if they have something like $10 billion of capital per year to deploy towards talent, um, they're going to get someone. So, I actually feel like we've been fairly good about protecting our top talent. And, you know, it, it's been kind of interesting and fun to see it escalate over time.

Um, you know, uh, some interesting stories here are Zuck actually went and, yeah, hand-delivered soup to people that he was trying to recruit from us. >> Like a, a just a, just to show how far he would. >> Yeah, I think he, he hand-cooked the soup. And, and, you know, it was, it was shocking to me at the time. But, you know, over time, it, I've kind of updated towards these things can be effective in their own way, right? And, you know, I've also delivered soup to people that we've been recruiting from from Meta.

>> You're doing a soup, soup counting.

>> I, I've thought of, you know, if I had an offsite, the next offsite for my staff, I'm going to take them to a cooking class. Okay.

>> And, yeah, I mean, it's just been, um, yeah. But I, I do think, you know, um, there's something I've learned about recruiting.

>> Did you cook your soup?

>> Uh, it's better if you get like Michelin-star soup. [laughter] You know what I mean?

>> Yeah. No, no, no. I think Dejo is very, very good and, um, probably better than any soup I could cook. Um, but yeah, I, I do think there is something I've learned about, you know, just, um, how to go aggressively after top talent. And I think, uh, the, the thing I've been actually very inspired by is that, um, you know, at OpenAI, um, even among people who, who have for Meta, I haven't heard anyone say AGI is going to be developed at Meta first. Um, everyone is very confident in the research program at at OpenAI. And one thing that I've made very clear to, to my staff, to, to the whole research org, is we don't counter, uh, dollar for dollar with with Meta. Um, and the multiples that, uh, below what Meta is offering, that people are very happy to stay at OpenAI gives me so much conviction that, you know, people really believe in the upside and believe that we're going to do it.

>> Well, and you, and Alex, Alex Ting. He used to be one of the math compet, the the.

>> Yeah. Yeah.

>> I'm sure you guys hung out.

>> Yeah. I mean, I, I have hung out with Alex a handful of times, but we don't do much anymore. Yeah. I [laughter] mean, yeah.

>> Why did soup become the thing? It was just, it just.

>> I don't know. You know, it's, there's been soup, there's been flowers, there's been anything you can think of under the sun. But, um, I don't know. I think, you know, life's an adventure. I, I play into the meme.

>> Yeah. Yeah.

>> Is there any, any poker strategy to employ, like, as you're, you're thinking?

>> Well, again, I think it really goes back to what I've said about the media narrative. Um, the game is not to retain every single person in the org. It's to trust in this pipeline that we have for developing talent and to understand who the key people we [music] need to keep are and to keep those. And I think we've done a phenomenal job at that. [music]

[music]

>> We have a special treat today. I'm excited. Mark Chen is here from OpenAI. Um, he's the Chief Research Officer. He's somebody I've gotten to know over the last couple of years. Thank you so much for coming.

>> No, it's been great to know you for for so long.

>> I feel like, uh, you know, there's a handful of people in this world working on this very important project, and, and, you know, you're right at the top of it. So it's, it's so cool to, uh, to have a chance to chat.

>> Yeah. Thanks for having me on.

>> It's a, it's my pleasure. And, and, you know, there's a bunch of things that I want to talk to you about because I've gotten to know you, like we said, over those last couple of years. I want to get, I want, I want people to know a bit more about your biography. But I also know there's going to be AI enthusiasts who want us to go deep on a couple things there. So we'll, we'll try to do everything. Um, I wanted to start just by giving people a feel for your job, which in my head, and you, I mean, just correct me if I get any of this wrong, but you're, you're, you know, Sam has been, he, he's really into research. He's the boss. He's kind of at the top of the food chain. But then you and Yakob are working, >> together to shape OpenAI, OpenAI's research direction. And, and then you're in this additional part of this role is, is, is actually deciding which compute goes where onto these projects. So you kind of have to chart where OpenAI is heading and then the mechanics, yeah, of how you're going to get there.

>> Yeah. And this always strikes me as a horrible job because I picture people, um, doing everything in their power to get GPUs from YouTube.

>> It's true. People are very creative in the ways that they try to make backroom deals to to get the GPUs they need. But yeah, I mean, it is a big part of the job, right? Uh, figuring out the priorities for the research org, um, and also being accountable for execution. So, really, to that first point, um, you know, there's this exercise that Jakob and I do, um, every one to two months where we take stock of all the projects at OpenAI. And, uh, it's this big spreadsheet about 300 projects, and we go and try to deeply understand each one as best as we can. And, um, really rank them. And I think for a company of 500 people, it's important for people to understand what the core priorities are and for those to be communicated clearly, with explicitly verbally, and also through the way that we allocate compute.

>> 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 software runs right through those expenses, figures out where we're spending money, and takes care of so much stuff for you so you don't have to waste your time on it yourself. Go to brex.com/cormemory 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 you've got, when you're talking about the 500, these are the 500. This is the heart of the research team in an organization now that's thousands of people. Yeah. Okay. So, and then in that, when you're talking about this 300 projects, I, I imagine, I mean, obviously some of those are the giant frontier models, and then some are probably experiments that people are working on. And so, like, how do you possibly keep track of all that and then come to some sort of conclusion about what merits GPUs and what doesn't?

>> Absolutely. So I think, um, it is very important when doing this exercise to keep the core roadmap, um, in focus. And one thing that differentiates, I think, OpenAI, uh, with other, other big labs out there, is OpenAI has always had core exploratory research at its core. Um, we are not in the business of replicating the results of other labs, of kind of catching up to other labs in terms of benchmarks. That isn't really our bread and butter. We're always trying to figure out what that next paradigm is, and we're willing to invest the resources to make sure that, you know, we find that right. And, um, I think most people might be surprised at this, but more compute goes into that endeavor of doing exploration than it is to training the actual artifact.

>> It must be. It's still got to be, how do you stop yourself from being persuaded by someone? Because everybody's going to put, I, you know, like when I think about this sometimes, I picture when I was at the New York Times, you would have this Page One >> meeting where >> everybody wants to be on Page One.

>> Yeah.

>> Everybody thinks their story is the most important story. They're all doing their very best job to tell you why this thing is so important. Everybody in that room has worked >> weeks, months on on whatever they're pitching. And so it feels like life and death. And that, yeah, I mean, it just, it seems so difficult for me.

>> Yeah. No, it, it is also a difficult process. And I think the hardest calls you have to make are, you know, this is a project that we just can't fund right now. Um, but I also think that's good leadership. You need to clearly communicate that, hey, these are the priorities. This is what we're going to talk about. These are the types of results that we think move the research program. And, you know, there can be other things, but those have to be clearly number two.

>> And when you, like you were talking about, um, not being reactive to your competitors. Yeah. When I was looking through my notes, I, I don't know if I could go to the line >> quick enough, but I mean, this, this was like a point of pride that I saw that you feel like, um, some of the other companies are, well, you know, you guys were in this position where you were ahead, um, and, and setting the bar for others. And so, so they were reactive, right, to, to what you had coming out. We happen to be doing this interview a few days after Gemini 3 came out. And, you know, there is a degree to which your rivals >> um, >> at times, like, yeah, I mean, there's this back and forth going on. And, and I know the benchmarks are sort of controversial, how valuable they are. But you [clears throat] know, people go ahead on these things. So how do you also, um, as time has gone on, maintain that luxury or that intellectual position where you feel like, we're just going to do what we're going to do?

>> Yeah. I, I think AI research today, the landscape is just much more competitive than it's ever been. Um, and the important thing is to not get caught up in that competitive dynamic because you can always say, hey, you know, I'm going to ship an incremental update that puts me in front of my competitor for, you know, a couple weeks or a couple months. And I don't think that's the long-term sustainable way to do research because if you crack that next paradigm, that's just going to matter so much more, right? You're going to shape the evolution of it. You're going to understand, kind of, all the side research directions, um, around that sphere of ideas. And so when you think about, kind of, um, our research program as an example of this, right? We bet more than two years ago that we're really going to crack RL on language models. And this was a very unpopular bet at the time. Uh, you know, right now it seems obvious, but back then the environment was, hey, you know, there's this pre-training machine that's working great. There's this post-training machine that's working great. Why invest in something else? And I think today everyone would totally tell you, you know, thinking and language models, it's just a primitive you can't have, uh, can't live without. And, um, so we're, we're really there to make these bold bets and to figure out how we can scale and build the algorithms to really scale to orders of magnitude more compute than we have today.

>> It's just, it, you know, I mean, and I get that intellectually. In my, you know, it gets harder as you guys started as a, like, basically a pure research company. When you look at OpenAI today, I mean, you have product lines. There's parts of OpenAI that look much more familiar to a, >> mature Microsoft or a Google, where you have product lines, you've got all these different things that you have to serve. Typically, I feel like you guys are still young enough, so maybe you don't have these exact pressures yet, but, you know, as those companies go on, it always becomes, well, we're more focused on the things that are serving the bottom line than spending a ton of money on research always seems to get, like, dwindled down >> over time.

>> Yeah. And I think that's really one of the most special things about OpenAI. At its core, we're a pure AI research company. And I don't think you can say that of many other companies out there. And, you know, we were founded as a nonprofit, and I joined during that era. And I think the spirit is, you know, build AGI, advance AGI research at at all costs, um, and do it in a safe way, of course. Um, but yeah, I actually do think that's the best head fake to really creating value, right? If you focus and you win at the research, the value is easy to create. So, um, I think there's a trap of getting too lost into, like, oh, you know, um, let's drive up the bottom line, uh, when in reality, if you do the best research, that part of the picture is very easy.

>> And you, you started in 2018, and so you feel like that soul, that that, um, >> yeah, that core culture and that core nucleus, it's, it's really persisted.

>> It's still there. What does Elon say? Uh, what does he, he says we shouldn't call any of you guys researchers. It's just engineering, right?

>> Yeah. No, I think, yeah, we, no, it's true because I feel like once you have this hierarchy, um, and you elevate, let's say, research science, um, as a thing beyond engineering, you've completely already lost the game. Because, you know, when you're building a big model, so much is in the practice of optimizing all of those, you know, little percentages of, you know, how do you make your kernels that much faster? Um, how do you make sure the numerics all work? Um, and that's a deep engineering practice. And if you don't have that part of the picture, you can't scale to to the number of GPUs we use today.

>> So, because I think there, well, okay, but there is like a mystique that surrounds a researcher versus an engineer, you know what I mean? So are you, were you, do you feel like, um, it is better to kind of stay level-headed on that? Is that, is that kind of what you're saying? Or, or.

>> Well, I, I just feel like researchers, they come in so many different shapes, you know? Uh, some of our best researchers, they're, uh, they're the type that, you know, they come up with a billion ideas, right? And many of them are not good. But, you [laughter] know, just when you're about to be like, ah, is this person really worth it? They come up with some, you know, phenomenal idea. Um, some of them are just, you know, so good at kind of executing on on the clear path ahead. And so there's just so many different shapes of researchers, and I think it's hard to just lump it into one stereotypical type that works.

>> That makes sense. Um,

>> Okay. I won't, I won't belabor you with too many competitive, like, rival questions. It's just, since Gemini 3 did come out, I did wonder, what happens with you personally or the team when one of your rivals puts it out? Like, does everybody go and look and see what it can do? Is there, like, a, is there a prompt or a question that you throw at these new models to see what they can do?

>> Yeah. Yeah. So, um, to speak to Gemini 3 specifically, you know, it's a pretty good model. Um, and I think one thing we do is try to build consensus. You know, um, the benchmarks only tell you so much. Um, and just looking purely at the benchmarks, you know, we actually felt quite confident. Um, you know, we have models internally that, uh, perform at the level of Gemini 3, and we're pretty confident that we will release them soon, and we can release successor models that are even better. Um, but yeah, again, kind of the benchmarks only tell you so much. And I, you know, I, I think everyone probes, uh, the models in their own way. There, there is this math problem I like to give the models, uh,

>> I, I think so far, none of them has quite cracked it, even the thinking models.

>> Um, so, yeah, I'll wait for that.

>> Is this, is this like a secret math problem?

>> Oh, no, no. Um, well, if I announce it here, maybe it gets trained on. But, um, yeah, I do think, uh, it's one of the nice puzzles of last year. It's this, the 42 problem. So you want to create this random number generator mod 42, and you have access to a bunch of primitives, which are random number generators modulo primes less than 42. You want to make as few calls on expectation to these sub-generators as possible. Um, so it's a very cute puzzle, but, um, the language models, they get pretty close to the optimal solution, but I haven't seen one quite crack it.

>> Okay. This is, we're heading down a direction I want to ask you about. But then, just before we get there, so I know you're, I've seen you, you're very competitive. You've also told me, I think I found, I love competition. I hate to lose. Somewhere.

>> I really hate losing. I [laughter] hate losing.

>> Yeah. So I'm picturing, I'm just curious if this is at all right. I mean, if, you know, if we know Gemini 3 or whatever is coming out on a Thursday, I mean, are you up at like midnight throwing that problem at it, or is it, is it not quite that drastic?

>> Um, no. I mean, I think it's in long arcs, right? Um, and any endeavor, like, I, you know, I'm kind of a person who has obsessions. I, I think any endeavor, you have to play the long game. Um, and, you know, we've actually been focusing on pre-training, specifically supercharging our pre-training efforts for the last half year. Um, and I think it's a result of some of those efforts, uh, together with Yakub, focusing and building that muscle of pre-training at at OpenAI. Uh, you know, crafting a really superstar team around it, making sure that all of the important areas and aspects of pre-training are emphasized. Um, that's what creates the artifacts today that feels like we can go head-to-head with Gemini 3 easily on on pre-training.

>> Okay. And I, I want to ask about the pre-training stuff because I've, I've been talking to all you guys about this a lot. But, but, um, okay. But so, but you're saying that, um, you're less obsessed about lobbing, um, problems at these new models just when they appear and more at this, this long journey.

>> Absolutely.

>> Yeah. Okay. Um, okay. Hey, the reason I want to talk about the puzzle that you were at, I mean, I, >> you know, I first met Yakob before OpenAI ever started when he was doing a coding competition. And I got, I got like super into coding competitions for a while. There's this guy, >> Kennedy, I don't know if he's still famous, but he was like the Michael Jordan of of these coding competitions. And so I went to watch one at, uh, Facebook used to, I don't know if they still do, but they had an annual >> Hacker Cup.

>> Yeah, Hacker Cup.

>> And, and that's where I saw Yakob for the first time. And then I know you, I think, did math competitions in high school.

>> Yeah. Like, probably grade school through high school.

>> And then you also, did you also do IOI?

>> So I got into coding really late in life. Um, it was a roommate in college that convinced me to take my first coding class. Um, and I had all the hubris of a mathematician at that time, whereas like, you know, math is the purest and hardest science, and that's where, you know, you really prove your worth. I mean, I think I was probably too into the competition back then. Um, but yeah, I mean, it became this super rewarding endeavor. And, um, and, you know, it, it started out as purely a way to keep in touch with my friends from college. Um,

>> You went to MIT.

>> Yeah, I went to MIT. Um, you know, I graduated, and every weekend we would just log on and do these contests just to to keep in touch with each other. And, um, you know, over time, I found myself having a talent for it. You know, I, um, started competing fairly well, and then writing problems for for contests like the USA Coding Olympiad. Eventually started coaching that team. And, yeah, it's been a great community where I've met people like Scott, that you know.

>> Yeah. Yeah. Okay. So, you, so I think lots of people might be familiar with like math competitions because they probably see kids going through that. The IOI and in these coding competitions are a little bit different. I mean, it's, I mean, you'll know it so much better, but when I saw them, I mean, it looks like a, it's almost like a word problem that's a puzzle, and you're trying to kind of find the most efficient and correct way to solve that, and you're in this race against everybody. And, and everybody's like writing code on their computer, and then, and then some people try to get there really fast, but then their thing kind of doesn't solve the problem, right? And then, you know, there's like this trade-off, right?

>> Absolutely. Right.

>> And so, so you, you actually were on the MIT team?

>> No, no, it's something I did after college.

>> After college. Okay. But today you are like the coach of the US national.

>> Yeah. One of the coaches.

>> One of the coaches. Okay. And, and, >> was it last year or the year before? Like the US, like we hadn't won one in a long, long time, right?

>> Yeah. Didn't we?

>> Yeah. Yeah. Yeah. So, um, yeah. I mean, the, the team, I mean, it's, you know, you can never predict what the makeup of top talent looks like every year. Uh, but we had a very spiky team, I think, two years ago. Okay. And, um, yeah, I believe they won the Olympiad.

>> Because I feel like usually it's like China or Russia or like, uh, >> Belarus and Poland. I mean, right?

>> Yeah. And, and so this compet, the big competition [clears throat] >> takes place in a different country every year.

>> What does it look like? How many people show up?

>> Yeah. Yeah. So they, they take the top four students from every single country. Um, it is as much of a competition as it is a social event, you know. Um, this is a tight-knit community. They all do go on to do phenomenal things. And, um, yeah, it's this intense two-day contest where each day you get just three problems, um, five hours to solve them, and you can really feel the adrenaline and, uh, all the pressure in the room. Um, but it's also great fun. I think, um, people settle down and they make, you know, lifetime friends through it.

>> What do you, and like as coach, I mean, you're so freaking busy, man. How, what do you, uh, how much time do you spend on this? What does that look like?

>> Honestly, um, the kids are so self-motivated. Sometimes it's really about just managing their performance and and strategy. Um, I think, you know, you're going to have good days, you're going to have bad days, you're going to have good hours within the contest, bad hours, and you can't let that get into your head. Um, there's a lot of similarities between managing contestants and managing researchers. Uh, it's like on a much longer time scale, but, you know, like researchers have good months, bad months. You know, you can't really let those strings of failures get into your head because that's just the nature of research, right? And, um, I think a lot of its morale management, um, at a certain point.

>> Um, yeah, I think one other interesting thing that contests have helped me realize lately is when you put the models and deploy them towards solving these these contest problems, which they're quite good at these days.

>> Yeah, I was going to ask you about that.

>> Um, they, they work in a very different way from humans. You know, we typically think of these machines as, you know, they're very good at pattern recognition. You can take any problem, if it maps to a previous problem, it's probably going to be able to solve it. But what I've noticed is in some of the previous IO, there's this problem like message is very ad hoc. Um, I didn't think the models would solve it at all, but actually one of the easier problems for the AI. So, >> yeah, I mean, this has given me the sense that AI plus humans in frontier research, it's going to do something amazing just because the AI has a different intuition for what's easy and what's not.

>> So, okay. Is it vaguely, you know, when, when DeepMind did the whole AlphaGo thing, >> you know, there was that moment where it was doing things human, it was playing in ways humans hadn't played before. So, kind of like vaguely similar to that? Or,

>> I, I think so. I think so. Um, I think really with GPT-5 Pro, right? Um, there's been an inflection point in frontier research. Um, and one of the best anecdotes I have for this is, you know, I think three days after the launch, um, I met up with a friend who was a physicist. And, you know, he had been playing around with the models, felt like, you know, they, they were cute but not super useful. And I challenged him with the Pro model, just try something ambitious. And, you know, he put in his latest paper, it thought for 30 minutes and just got it. And I would say that that reaction in that moment, um, it was kind of like seeing Lisa Doll during that, you know, move 37, move 38. Um, and I just think that is just going to keep happening more and more for frontier mathematics, for science, for biology, material science. Um, the models have really gotten to that point.

>> I was gonna ask you this question, which is not very original, because I think we've been doing this ever since, kind of, Big Blue and and all the chess stuff, but, yeah, just as somebody who had followed all these competitions, if, >> I don't know, there's a sadness when you start seeing >> these models solving these things that were like the >> this height of these achievement for these very unique human minds.

>> Well, um, yes and no. I mean, I was good at competitive programming. I was never at the absolute top. And, um, maybe this is a way to get revenge. No, [laughter] but I, I do think, um, no, there's certainly a moment for myself, right? Um, you, we tracked coding contest performance while we were developing reasoning models for a while. And, you know, at at the start of the program, you know, they were not super great, you know, uh, at the level of any average competitor going, uh, going, going into the contest. And, yeah, over time, they just started creeping up and up in in terms of capability. And, and you still remember that moment when you walk into the meeting and they have where your performance is, and then the models exceeded that.

>> Um, man, that was also a shock to me. It's just like, wow, we've automated to this level of capability so fast.

>> And of course, you know, Yakob was there, still a bit smug, but within like one or two months, it was also surpassing him. So, >> um, yeah, no, the models are at the frontier today, right? Um, it's so clear by even through the results we've done this summer, um, at Codeforces, right? Top optimization competitive programmers in the world, um, I think it achieved second place there. Um, and so really, it's jumped from, you know, hundredth place last year to top five this year.

>> And, like, do you think we'll still be doing these competitions in 10 years?

>> I think so. I mean, they're just fun. Um, I mean, certainly a bunch of people who use it to, you know, had their resume are going to drop off from doing it, but I think the people who've always excelled at it the most are people who just do it for the fun of it. And I don't think that'll go away.

>> When I was doing this story, I mean, they were telling me that like if you're from >> Russia, or I don't know which countries, that you basically get like an automatic free ride to any university that you want. I mean, I see the guys on the US team go to like Harvard and MIT, so they seem to be doing okay, but it was, it doesn't seem like the US has a [laughter]

>> Yeah. I mean, don't you think it's going to, yeah. I mean, interviews, right? They're going to be kind of broken going forward. And I, everyone's seeing this a little bit. And, you know, even college exams or college homework, it's, it's all broken at this point, right? And I, I do think we're going to need new ways of assessing and gauging, you know, who's performing well, who's learned the, >> where somebody's actually at.

>> Yeah. Yeah. So I mean, I, I've had this idea here where, um, maybe for our interviews, we should just have candidates talk to ChatGPT. And, you know, it's a special kind of ChatGPT where the model is trying to gauge whether you know the material or or whether you're at the capability level to work at OpenAI. Um, and, you know, you have to have this conversation with it that convinces it deeply you belong at OpenAI. And of course, you know, you can't be allowed to jailbreak it, but, and we look at the transcript after. But maybe like tests like this will more accurately reflect in the future whether you know.

>> So you don't do that yet, but you're thinking about.

>> Yeah. Yeah. Just creative ways to revamp the interviews.

>> Yeah. Yeah. Well, I mean, Silicon Valley is famous for doing the like, >> brain teasers during the interviews and everything. Yeah.

>> Um, so you, I mean, we tal, you were very good at math [laughter] growing up, and, and I think you were, you were born on the East Coast?

>> Uh, yeah, born on the East Coast.

>> And then you lived on the West Coast too?

>> On the West Coast, and then you lived in Taiwan for like, for like, um, grade school to high school.

>> Four years.

>> Okay. Your parents worked at Bell Labs.

>> Yep.

>> So you come from like engineering stock. [laughter] Uh, I mean, it's a really interesting background, um, just because you kind of got like a flavor for all these innovation hubs. And, and especially with your parents being at at Bell Labs, and most of, >> I mean, yeah, I just grew up in a very scientific environment, you know, dinner table talk was puzzles and things like that. And, um, I also got kind of the more traditional, you know, Bell Labs East Coast experience, um, on the West Coast, my dad came to do a startup, so a little bit of that kind of new company got got exposed to that when I was young as well. And of course, the big jump to Taiwan, right? And I think it's a huge culture shock. You, you wear uniforms, you're in a school, it has barbed wire around the school, right? And, um, and also getting exposed to kind of that level of rigor. Um, I think it was just a number of really great experiences growing.

>> So, like the schools were harder?

>> Um, or,

>> Well, I would say it was just much more kind of, uh, you know, it's just there's a little bit less flexibility and freedom in the school system, but I think it also teaches you something.

>> Yeah. Okay. Since day one, the Core Memory podcast has been supported by the fine people at E1 Ventures. 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.

>> Yeah.

>> And you knew you wanted to come back to the US for college.

>> Absolutely. Yeah.

>> Okay.

>> And then, Okay. So, you're, you're at MIT. Um, you're kind of like in this interesting group. I guess MIT probably always has an interesting, >> Oh, man. Yeah. 2012 was such a great group.

>> Yeah. Like, who is there, sort of like an all-star list?

>> Oh, I mean, it was a great year. Like, um, I don't know if you knew, like Jacob Steinhart, you know, he's doing Trans Loom now. Um, he, uh, and I used to do projects together in in computer science class. Um, there was Paul Cristiano, who's, um, uh, a bunch of really >> phenomenal. He >> worked at OpenAI. Yeah. Um, a bunch of kind of big names in in AI came from that year.

>> And then, and then we were talking about the, the competitive, um, coding, like Scott Woo, who's at Cognition. I mean, he's kind of like famous now as like a meme on X for his math abilities, but, and you just got, you got to know him through the coding competition.

>> Oh, yeah. Through the coding community.

>> Okay. Okay. And then now I see the competitive end of you guys. The, the output of this to me looks like poker these days. I think I, I was, we were at an event which I think I have to, we have to keep secret or something like the, the specifics on this event. But, um, I think I'm okay to talk about this part, which is like, I >> late at night, >> I'm walking by this table, there's you, Scott, I think Sham from Palantir, and then, and like a handful of other people in a, like a fairly intense, looked intense to be, maybe it's, it's not for you guys, but like a fairly intense poker game. So you guys, this is where you've applied your your math and competitive skills now.

>> Yeah, I mean, poker is a really fun game. And, you know, I've talked about my life in terms of a series of obsessions. Poker was definitely one of these obsessions in in the past. And, um, I think the big revelation for me in poker was, um, you know, it's so much more a mathematical game than a game of reading people and bluffing. And I think the more you learn about poker, the more you update in that direction, right? Um, and I think that's, you know, I used to be a terrible bluffer. And when you know it's mathematically correct to bluff, then it's so easy, right? It's like, you, you don't feel any nervousness around it. Um, and, yeah, it's just so interesting that you have a game that I think is perceived as so human, um, but the underlying mechanics and how to win are so deeply mathematical. And, yeah, I kind of thought about this the other day where, you know, there's something about that in language modeling too, right? You have, um, this deeply human process of generating language, but there's this mathematical machine that can really do it as well as we can.

>> I think about that part all the time as a writer. And then I, >> did all this philosophy back in college with Wittgenstein and all these guys thinking about these things. Yeah. Um, well, how do you find like an edge? If you, and you and Scott both strike me as like supernaturally good at math, I don't understand how one of you is out calculating the other person.

>> Um, well, no, I mean, it, it is mostly, um, a forum for us to just kind of, um, just hang out and catch up with each other. Um, and, you know, today we don't take it as ser, I think there is an element to taking something like poker very seriously that takes the fun out of it. And, and so, you know, I, my obsession with poker, I think has ended more than a decade ago, and now it's just fun.

>> You're just saying this 'cause I saw Scott win both [laughter] days. I think.

>> You might be right about that.

>> He was taking it, he was taking it quite seriously. Um, and, and so like coming out of college, I mean, you had, in some sense, I was, >> Oh, I beat him on the plane though.

>> You beat him on the plane right home. Okay. All right. All right. So you did, you did, was it just you versus him, or it was like a group thing?

>> And maybe three or four people.

>> Okay. [laughter]

>> Um, I feel like a lot of, feel like there's three, I don't think I'm overgeneralizing too far, especially among, like, if you cut back to the 2018, >> right, >> sort of time frame. Um, as far as like people who were in AI at a high level, I mean, a lot had academic backgrounds. A lot were math prodigies or had gone on to sort of >> take their math background and and get into robotics or something, physics, like that. And then, and then there's this other bucket, which is people who had gone to Wall Street and done, um, high-frequency trading and and quants and and things like that. So that, that was the first path that you took. Was you went straight from MIT to to Wall Street.

>> Yeah. I mean, I don't wear that badge with too much pride, and [laughter] to be honest, like, you know, it, it was a, a path that was fairly common for, you know, very quantitatively oriented kids at at MIT. Um, and I mean, it was certainly a very meritocratic system, right? You could apply intelligence, and there's a very concrete reward function, the the amount of kind of profit that you would make. Um, but I think culturally it was hard for me. I, it was a place where when you discover something, your first instinct is to just keep it away from as many people as possible because your knowledge is what gives you your worth. And, and so it felt like you would have out of this an outgrowth of just even internally at a company, like these competitive dynamics, and people weren't very trusting of each other. Um, I think it was also felt like such a closed ecosystem, right? I think we today don't feel too much like, you know, when when someone in HFT finds a breakthrough that makes their algorithm a little bit faster, no one else feels it, right? And I, over time, just kind of felt like, you know, I woke up after four or five years, we're competing against the same exact set of players. Everyone was a little bit faster, but had the world really changed that much for it? Um, and it felt like time to do something else, right? It just, a bunch of things lined up back then. You know, there's the AlphaGo, um, uh, match, which I think was a huge inspiration for a lot of people at at OpenAI. And,

>> Did you play Go?

>> Uh,

>> I did not, but,

>> I think the sense in which, you know, the, the model was able to do something creative, I really wanted to understand what was going on behind that. So you're watching that happen, and had you been, had you been at all reading AI research papers and things like that, or,

>> To be honest, no. And then I saw that event, um, it was really inspiring, and that's when I started doing my deep dives into AI. So I, um, one of my goals after seeing that was reproduce the DQN results. Um, this is a network that was able to, uh, play a lot of Atari games effectively at a superhuman level. Um, and, and going from there, you know, that's how I got my start in AI.

>> And you were doing that, like, on the side, as you, so you just work all day and then go back and try to.

>> Yeah. Yeah. Yeah.

>> Yeah.

>> Okay. I mean, it is weird. I remember I was interviewing George Hotz, like, >> it might have been roughly 2018, um, maybe a little before that, you know, and he had just done this thing, >> building a self-driving car on his own in his garage. And then, you know, I mean, it's George, so he says, >> large statements sometimes that may may or may not be like exact or spot-on or or apply to other people. But he's like, he's like, AI is still so young. You can, you can basically learn the whole field if you read, I don't know what the number was, 10, 20, 30 research papers. I mean, it is fascinating to me that it's like, >> old in many ways, stretching back decades, but this particular moment, >> it's very shallow. Um, I, I always give this advice to people who are intimidated by getting into AI, so shallow, like, just spend three to six months, um, picking some project, like maybe, you know, reproduce DQN, and you can get to the frontier very quickly. Um, the last couple years has added a little bit of depth, but it's not anything like, you know, theoretical math or physics.

>> Do you think, is this a field where I asked Jacob this the other day, I don't know why I'm obsessed with this, but, you know, like in mathematics, you see, you see people tend to do their best work in their 20s or to have the big breakthrough, and then it's very hard as they get older to like have that same, >> kind of moment. Um, like what you're saying, you know, are we, are we dependent on on young people reading these papers and then having some insight, or is this, is this something where it's more, you can, you can keep going throughout your whole career?

>> I mean, I think you can keep going. I mean, OpenAI itself does have a pretty young culture, but I don't think you need to be young to do good research. I think there is something about being young and having less priors about this is the way it's done. Um, I think over time you may develop your own vision, which is a good thing, right? Um, but it also locks you into a frame of mind of like, oh, you know, this is how research is done, this is how good results come out. And I think younger researchers tend to have a little bit more plasticity around that that concept.

>> Yeah.

>> Um, so, as your career is at OpenAI, it's funny because it seems like you walked in the door and had a very, um, important, large position from the get-go. But when you first got there in 2018, it must have been, what, like, uh, 50 people?

>> Oh, no, it was much closer to 20.

>> Much closer to 20. Okay.

>> And it, it really looked like two teams back then. I came in as a resident. So someone who, you know, clearly not a specialist, not a PhD. Um, I think I was, I was only a resident, you know, throughout his tenure at OpenAI. So I was very lucky in that regard, just kind of learning the way that he thinks high-level about research. Um,

>> And a, like a resident in this case is, you're just the right-hand person to.

>> Oh, so it's, um, someone who comes in, usually from another field, who OpenAI wanted to invest in and train up in AI. Yeah. And so I think the first part of

A residency is like a six-month compressed PhD. Um, and and then going from there to to just, you know, getting into deeper and deeper research projects.

So you're kind of like talking to IA every day. Is he kind of shaping that that?

Yeah, he was responsible for for my projects, for my curriculum, for my learning. And um, and you know, I would just go to him for, you know, "Hey, like, what's this about? Like, why did people pursue this?"

Okay, yeah. And you, I mean, I think if you go like on LinkedIn, I mean, it would say you were the head of Frontier Research as your first job at OpenAI.

Oh, no. I, I was an IC for three-ish years. Yeah. So, um, I was doing independent research projects. Um, I worked in generative modeling because that was really kind of where IA's focus was at the time. And then, um, only after a while did I start managing teams.

And because most, you're talking about generative. I mean, most people point to Dolly as maybe the kind of first big project that the public would?

Mostly re is that is that fair?

Yeah. Yeah. So, um, I think that also marked the transition between when I was an IC and and a manager. So, uh, one of my own kind of big projects and one I'm still pretty proud of today is, uh, Image GPT, a proof of concept that even outside of language, you could put things like images into a transformer. Um, and the model would just internalize very good representations and understand, um, you know, the content of images. And, um, it's kind of like a proof of concept that you can do language modeling outside of pure text and get really, really great representations and scale them to be state-of-the-art with other methods. Um, that was, I consider like a precursor work to Dolly, which I was on the opposite side of managing. Um, and, um, I think between those, uh, another project I'm really proud of, uh, uh, doing doing IC work on is, uh, Codex, where we, you know, set up a lot of the framework for evaluating, um, coding models and also, uh, did a lot of in-depth study on how you can take language models and make them very good on code.

So, and what made you pick OpenAI? Because I, I could see it two ways in my head. One is, is big fish in a small pond. There's interesting people here. I remember the OpenAI of 2018 with 20 people. In my head, it was like, this is probably not going to work. Um, you know, [snorts] Google seems like they've got this locked up. And and this is like a pretty small group of people trying to take on something that appears to require many billions of dollars of of capital. And I mean, this was even before the scaling stuff. It was just like Google had invested so much [laughter] in in AI already. Um, kind of like in a different form than we think, but, you know, you're already kind of doing translation on your phone and things like that. So, um, was that like a hard decision for you or or you just stumbled into the OpenAI gig so quickly that?

Well, yeah, I mean, I think there are two things, right? You need ambition of vision. That was certainly what OpenAI had at the time, but you also need the talent to back it up. And, you know, I did feel like OpenAI was one of the rare places where the ambition was very large, but the talent was also large enough to fulfill that that gap. And, you know, I I was lucky. I knew people like Greg, um, before from from college. Yeah. And.

Oh, so yeah, Greg was, you overlapped at MIT.

Um, I think we did some math contests together. Okay. Yeah. Back in high school. [laughter] And, um, yeah, I I shot him a message actually and I was like, "Oh, you know, I I don't know if I have the right skill set, but this sounds like a place that's doing great work." So.

It still seems nuts, though, to just come at this, you know, out of nowhere and and now you're like leading [laughter] research.

No, it's surreal to me, too. It's surreal to me, too. Um, you know, even that transition from IC to manager, I was very hesitant about taking it. Um, I didn't know if managing was a skill set that I would be good at, and I was really enjoying IC work. I think, uh, um, I was having a lot of fun doing it, excelling at it, building really great collaborations. Um, but yeah, I mean, it's [snorts] really been a wild ride.

Yeah. Yeah. Well, okay, on that point, I mean, you've always struck me as a very nice, like level-headed guy. I have to say, you know, there's parts of OpenAI's history that are are quite, um, dramatic, soap opera, like, um, a little Game of Thronesy, you know, power struggles and and like to me, um, to be a manager in that. I will say now, like, I feel like things are a bit calmer than they they were, but but when you look backwards, it just seems like, um, I don't know, you you're saying you had to learn these skills. Feels. But some of this feels like opposite to your personality. Um, to have to like deal with all of that.

Honestly, you know, I've been lucky at OpenAI. Um, I genuinely say that and um, in the sense that I've had managers that have really advocated for me. Um, you know, they they saw my talent and and advocated for me. I think when I was an IC, you know, W. Check, he was like, "Oh man, you should bet on him for for Codex." And, um, and then later on, kind of reporting to Bob, I I've never asked for promotion or up-level. And, you know, it's just organically happened. Um, and everyone kind of along the way has given me great advice. Um, I think part of growing as a manager is just getting the reps. I don't think there's any better place to get the reps than at at OpenAI. You know, there's always challenges to solve. Um, and, yeah, I I think kind of developing that confidence. I actually think management is something where it's really just about the experience and, um, you know, there's, I would say, less so talent involved in it.

Yeah. I don't want to like embarrass you, and I don't know if this will or won't, and I assume you probably don't want to get too much into the the coup or the blip or whatever. We want to talk about anything. [laughter] Yeah. Yeah. Well, I just, I've I've interviewed so many people about this now. And I'm also gonna save some of my gems for my book. So, I won't, uh, I won't up myself. But there there's a couple moments in there where you, [clears throat] um, you know, you help, uh, get the researchers aligned around that the petition to like bring Sam back. And then and then I think just that either a day or two after that, there's there's kind of like this speech, you know, that given, I think at Greg's house, maybe, or.

I think Chelsea's house.

Okay. And and, um, you know, both of those struck me as pretty profound moments in especially for, um, I guess like standing up with for what you believe in and rallying the troops. I mean, yeah, like, uh, you know, in a in a moment of crisis, I don't know. So, did do those?

Yeah. I mean, that did feel like a very pivotal moment for me. I think in the days following the blip, right? There was a lot of uncertainty. Um, and, you know, myself, Nick, uh, Barrett at the time, we felt this responsibility of, you know, the wolves are at the heels, right? Everyone's getting calls from from all these competing labs being like, "You should come work here instead." And I just set this goal of, I will not lose a single person. Um, and we didn't. Um, and it was just, you know, every day, um, opening up our houses, you know, people could come here, they could, you know, have a place where they let out their anxiety. Um, and then also just helping them keep in touch with the leadership team, um, having a way for them to feel like they could make a difference. And I think, you know, over time, people really felt this spirit of, "Hey, we're all in this together. Um, how do we make a difference? How do we signal to the world that we're all together?" And, um, you know, we, I've been kind of driving back and forth, uh, between a couple houses. And, uh, we had this idea of like, "Hey, you know, we need to show the world that we're all seriously aligned and we're going to work for Sam." And, um, that's when the petition came together. And I, you know, the idea, I think, got solidified at 2:00 AM. We got more than 90% of the whole research org signed, I think, by the morning. Um, and it was just everyone like calling their friends being like, "Hey, are you in or are you are you not?" And, yeah, I think in the end, you know, it was very close to 100 people signing that petition.

Is well, I mean, that must have put you in something of a tough spot, though, just because especially at the outset, it was kind of like Ilia and Sam were on opposite sides, and Ilia is your mentor. And then I know Ilia kind of, uh, comes back.

Um, yeah, I don't know. Was that awkward?

Um, no, no. It was hard. I mean, it's it's a low-information environment. Um, but fundamentally, it's just, uh, and and yeah, I mean, I think at the moment, you know, you could very reasonably conclude like, "Was like, did Sam do anything here?" You know, is there, um, but would Greg and Yakob, like, people of super high integrity, quit over that? Um, I just felt like, you know, there was some part of the story that that was being misrepresented here.

Yeah. Yeah. With the, you know, Yakob's been there for a very long time. Um, like, what should people know about Yakob that they don't?

It's interesting, 'cause he's a super funny guy.

He's hilarious. Oh my gosh. He has this like sarcastic humor. Um, and, yeah, it cracks me up so much, honestly. Like, um, yeah, that's one of my favorite things about OpenAI today. Just like the level of alignment I have with Yakob. Um, I feel like we go into a meeting, um, we can just bounce off ideas and quickly get to alignment and then, you know, deliver the same message and kind of like operate on different parts of a big roadmap together. Um, and, yeah, it's just one of the big privileges I have working at OpenAI.

Yeah, I mean, go going to that actually, that point about, um, you know, keeping people together. Like, I still feel that way about OpenAI research. Um, I think we're, uh, still under attack. [laughter]

Yeah. No, I, we are a family. Yeah. We're always under attack. Look, when any, and this is how I know we're in the lead, right? Any company starts, where do they try to recruit from? It's OpenAI. Um, and, you know, they want the expertise. They want our our vision, kind of our our philosophy of the world. And we've made so many star researchers, right? Um, I think OpenAI more than anywhere else has been a place that makes names in in AI today. Um, and I still feel that same level of protectiveness. Like, you come after, I'm going to do anything in my power to make sure, you know, they're happy, they're at OpenAI, and they understand, you know, how their role fits into the roadmap.

I, yeah, I this is something I battled with as I was doing the book or even just watching events unfold in real time is like, when I go back through the history, I mean, you've got, you've got Ilia in 2012 making sort of like a big breakthrough. And then, you know, you've got Noam McGoo in 2017 doing Transformers. And then you've got Alec Radford, you know, like sometimes the story is these individuals, really pushing the field forward. And it feels like a field that's still so young that you can have have this individual. And then and it seems like there's this group of, I don't know what the number is, but let's call it like 8 to 10 who seem to have, uh, an ability to do that repeatedly. And they're really shaping where this is all gone. And so when I started seeing like John Schulman leave or Alec leave, and then you know, there's kind of was like, "Wow, okay, well, if you've lost a chunk of this all-star team, how do you it seems like a kind of field where you you sort of can't just replace that." And yet, you know, it was it was kind of like after that that you guys, uh, pushed forward on reasoning and and some of these other spots. Yeah. So I don't know. I I've intellectually had trouble.

I, I do disagree with that as the the overarching way to do good research today. Um, I think there's certainly a lot of top-down steer, you know, we bet on directions. But people, it at OpenAI has this beautiful culture of being bottom-up in a very deep way too, where some of the best ideas just organically emerge from sometimes the most surprising of places. And, um, I think really the the great thing has been just like watching some of these bets unfold, take shape, get scaled, um, and reasoning being a a core example of that.

Yeah. And okay, so in this, but like this idea that, um, like how star dependent are we? Because you still see Google spend an ungodly amount of money to bring Noam back, you know what I mean? Yeah. And so this makes me think, okay, this is how this works.

Yeah. I mean, I think it's a mix, right? Like you have to you have to invest in your pipeline because I'm very confident in our ability to create stars. But yeah, there's certainly very good people out there and everyone knows that they're good. Um, I think if there's one thing that, you know, on on the flip side, I've learned from from Meta is, you know, OpenAI can also go very aggressively after star talent. And, um, you know, there's this very aggressive recruiting approach that, you know, I've taken a couple pages from as well. Um, [laughter] but yeah, I I I think it's, we should always just be trying to assemble the best team.

Yeah. Uh, in service of the mission that we want to accomplish. It's funny because it is like a relatively small world and like all you guys hang out even though you you're like rivals and then.

Um, yeah, it's it's a brutally competitive industry on all fronts, right? Um, but again, that's what I love. Um, I'm a deeply competitive person. I hate to lose. And yeah, on research, on recruiting, all of these fronts, um, I'll work very hard on them.

It reminds me because I'm like a semiconductor, well, I'm a history nerd, but just the early semiconductor days were not that far off. I mean, you had you had all these semiconductor startups come at once. They were all pushing the limits of physics and somebody would discover something at one.

They'd go to the bar and have a. It's like people, they're engineers. They can't like stop from like sort of, uh, sharing knowledge with each other. But and then they're also getting, uh, pulled. Like, you know, it's hard. Each company is is kind of quickly getting this breakthrough in one way or another.

Yeah. I mean, you raised an interesting point of, you know, there is going to be some base rate diffusion of ideas. And, um, I think there's two ways a company can respond to that. You can create these deep silos of like, "Hey, you know, we're going to protect information in all these ways." I don't think OpenAI operates that way. And we don't think that's the right way to operate. Um, we just will outrun other people as fast as we can. And, um, I love the culture of openness. People in research freely share ideas. And I think that's the way to make the fastest progress.

And how like how do you, Sam, and Yakob now work together? I I think people sometimes, um, if you read the the announcements and everything, you can tell that Sam is research-oriented over over like, uh, day-to-day running of the company, you know what I mean? You can tell research is more of his his passion. And and even just like in the the titles and in the way it's been organized, especially recently.

Um, you and Yakob are so deep on this stuff. And I know Sam is is technical, but, um, you guys are are like in it all the time. And then, you know, Sam, uh, is having conversations with everyone. Yeah. I'm just curious about this dynamic between the three of you and how, I mean, are you guys, I mean, I guess you're not always alignment on in alignment on what is going to get the resources, but, um, yeah, I was I was just curious about you guys dynamic.

Yeah. Yeah. So, I mean, um, it's a very tight cohort. You know, I talk to Sam and Yakob every day. Um, and, you know, with Sam, he loves research. He loves just learning about research. Um, he loves talking to researchers. I think in some ways, he's very effective at getting a pulse on the research org. I rely on him also to just, you know, "Are there any hidden latent problems here? Um, go and find them out, you know, surface them to me." Jakob and I,

personality or techno. It could be just small things like, "Oh, you know, um, like just even the way the office is laid out makes it harder for this team and this team to collaborate." Um, and the two of them is like need to collaborate to to help unlock this breakthrough that we want. Um, I mean, all these things are, you know, very, very important. And I think Jakob and I, we spend a lot of time figuring out how to design the work for success. You know, I think, um, pairing people with the right strengths together. Um, you know, also how to like incentivize people to work on directions that we find are important. Um, yeah, that that's a lot of the work that we do.

The and Sam, what he, um, like, is he reading papers? Is he chatting with you guys? Is.

Yeah. Yeah. I mean, I think he he does his fair share of reading papers. He talks to researchers and just understands how how they think about the world, the type of research that they're doing. Um, and of course, he's responsible for a huge umbrella of things outside of that.

All right, I'm going to ask some nerdy questions now, but I'm going to try to, um, I don't know if I can dwarf cash level, but [laughter] I'm going to do my best. And, you know, I'll ask. I don't know how top secret some of this stuff is, but but, um, anyway, well, maybe you'll slip up and we'll just, we'll get it out. Um, you know, in the the meetings I have been on, and I don't think I'm revealing because we talked about it a bit. I think I'm safe here, but, you know, pre-training seems like this area where it feel it seems that, um, my sense is you you guys feel like you've figured something out. You're excited about. You think this is really going to be like a major advance. It was also, I think, um, either a neglected spot or something of a sore spot. You know, previously, things weren't maybe working [clears throat] exactly, uh, how you guys had expected or hoped. Um, like, what can you tell us about what you figured out and and, you know, some sort of frame of reference on on we've seen these periodic big leaps forward.

Absolutely. So I think the way I would describe at a high level, um, the last two years is, you know, we've put so much resourcing into into reasoning, into understanding this primitive and making it work. And it really has worked. And I do think one byproduct of that is you lose a little bit of muscle on, uh, your other functions like pre-training and post-training. Um, in the last six months, Jakob and I have done a lot of work to build that muscle back up. Um, I think pre-training is really a muscle that you you exercise. You need to make sure, you know, all the info is fresh. You need to make sure, uh, people are working on optimization at the frontier, are working on numerics at the frontier. And, um, I think you also have to make sure the mind share is there. Um, that's kind of one one of the recent things I've been focusing a lot on, just kind of directing and shaping what people talk about at the company. And very much today, that that is pre-training. Um, we think there's a lot of room in pre-training. You know, a lot of people say scaling is dead. Um, we don't think so at all. Um, in some sense, um, you know, all the focus on RL, um, I I think, um, it's a little bit of an alpha for us because we think there's so much room left in in in pre-training. And, and I think as a result of these efforts, you know, we've been training much stronger models. And that also gives us a lot of confidence carrying into, you know, Gemini 3 and other releases coming this end of the year.

Like the way I picture it in my head sometimes is that you guys have been on this, you've just been running so fast. The whole field has been running so fast. And so we're at a moment where it's like, okay, we've, uh, gathered up this vast volume of information from the internet. We've we've thrown it onto this supercomputer and and that, you know, ChatGPT pops out. And then we're just on this this like this incredible race that's going on. And so like when I hear you guys, I'm just trying to think about this in like a to level set, maybe for people who don't follow this as closely. Um, so you, you know, in that initial moment, you just had so much data, you're throwing it at this machine. You try to shape that data a bit initially. And what now, we're just learning like more efficient ways to shape that. Just it's not always clear on what the mistakes were.

Um, so I do think, um, yeah, you touch on something I've been thinking about a lot, right? Um, when you think about pre-training, right, you're taking human written data and you're teaching the model how to essentially emulate it, right? Um, it understands human patterns of writing. Um, and in some sense, that also bottlenecks, um, and and puts a ceiling on the capability that you're able to achieve, right? You you can't really surpass what humans have written when you're imitating what humans have written. And so, you know, you you work on things like RL, um, there, you know, you can really provide steer towards the hardest tasks that humans, uh, can come up with and, um, have the model basically think outside of the box, outside of, um, what it's learned from from imitating humans and achieve higher levels of capability. But there is this kind of interesting problem now of how do you go beyond, um, what humans are able to do today? And I do find a serious measurement problem there too. Um, even in the sense of like, can humans judge superhuman performance in in the sciences, right? Um, would how how would we know that like this superhuman mathematician is better than that superhuman mathematician? And, uh, we really do need to kind of come up with better evaluations for, um, what it means to make progress in this world. Right? We've been lucky up to this point, right? There have been contests like the IMO, IOI, really just like gauging who's the top one mathematician in the world, right? Um, but when the model capabilities go beyond humans, there are no more tests.

Right? Okay. You just made me think of a question going back to the IOI stuff. I mean, and sorry, we're going to come back. I just, you just totally popped in my head. I mean, like often I would see the, uh, kids who were amazing at those competitions, they would get hired somewhere like a Google or Facebook or something, but they weren't always like the, you know, the top executive or the top the the most famous engineer afterwards. And maybe it was like by choice, but I don't think Gennady was like the Michael Jordan ended up working at any of these companies. And that totally could be by choice. I'm not trying to disparage him, but but it's not clear to me, um, like even, okay, so it's not clear to me that the human who excels at that is is necessarily like the greatest engineer you're ever going to have. And so like, yeah, may I mean, if an AI is particularly good, like what are we learning?

Yeah, that's a thing I quite like about working in AI. I think more so than, uh, in in standard engineering culture, it is a meritocracy. Um, in that, you know, I've tried this many times before and learned this lesson many times before, but it is hard to put in someone to lead a group who doesn't have the respect of the researchers that they're leading. Um, and I think this is more so the case in research than than anywhere else. You have to make very strong technical calls of like, you know, "This is the right right path when there's a disagreement. This is the right kind of project." Um, and if you make those calls wrong, you know, you lose the respect of of your researchers. So, um, yeah, one of the fun things in working with, uh, in AI and and creating a a strong AI org is, you know, all my ventures are very deeply technical. And it's fun to talk to them about the technical things.

Yeah. Yeah. Okay. And then Okay. On this on I'm pre-training again for a second. Okay. You know, like to me, in my head, it feels like Transformers helped kick off this massive, massive leap. I mean, reasoning to me feels very comparable, if not even, uh, sort of more amazing. I mean, are we, when I talked to you guys over the last few months, my, you know, and I can never tell if this is optimism,, uh, if you guys are just putting the best foot forward when I'm chatting to everyone. My my sense when I talk to you, to Greg, to Yakob, you know, to Sam is is that you guys kind of feel like you've been putting in hard engineering work for like three, four, five years that hasn't fully, uh, manifested itself. Um, and so then I can never tell how excited or not to be. I mean, like when you guys are hinting at some of the stuff you're seeing, do you feel like it is, you can already tell that it is a comparable leap forward, in terms, uh, to these big epochal kind of things?

I think so. You know, um, I I think when we launched GPT-5, you know, we, uh, we talked a lot about synthetic data as well. Um, you know, there are many other threads of this form that we think are holding quite a bit of promise and that we're scaling up pretty aggressively right now. And I think it's always about maintaining that portfolio of bets, uh, taking the ones that are providing more empirical promise and and scaling and supporting them at at an even greater degree.

But it was like two weeks ago, Andre Karpathy, who used to work at OpenAI, you know, he went on Dark Casual's podcast and seemed to like deflate, uh, some giant portion of the AI industry by saying, you know, I think he was saying what that AGI was like 10 years, 10 years off. And then when I hear, and then I I heard Dario talking about a week ago. I mean, he seemed to be holding on very much to like massive scientific dis his, um, what is he called? The the nation of geniuses. He seemed to be holding still on kind of like maybe a little slower, but like a two-year timeline on that. You know,

Yeah. When you heard what Andre said, what did you?

Yeah. I mean, I think Twitter, they love this like cycle of, you know, "It's so over. So back." And, you know, whatever plays into the narrative at the time, I think, you know, just becomes amplified. Yeah, I'm trying to make a clip here, but, you know, I the way I think about it, yeah, I mean, it's like AGI, I mean, everyone defines their own point for AGI. I I think even at at OpenAI, um, you can't get everyone in the same room and be like, "Hey, this is my clear definition of AGI," and it's it's it's consistent. And so, I kind of think about it as something like, you know, you're in the industrial revolution, right? Do you consider the, you know, having machines make textiles, is that the industrial revolution, or is it the steam engine? You know, everyone kind of has their different definition. And, um, I think we're in the middle of this process of producing AGI. For me, I think the thing I index most on is, are we producing novel scientific knowledge? And are we advancing the scientific frontier? And I feel since the summer, there's been a tremendous phase shift on that front.

Okay. Like from stuff that you're seeing in terms, the the first things that are jumping to my head are all these, uh, startups that are in the biotech space that are showing, you know, one-shot antibodies and and molecules. But I have no idea if that like, what are you?

Yeah. Yeah. So, I mean, I was so inspired by that encounter with the physicists that went back and thought, "Hey, well, we should just create OpenAI for science." And the goal being, I think, for the small set of scientists today who realize the potential of these models and feel like they want to lean in and accelerate, we should do the best that we can to accelerate them. And, you know, I know there are similar efforts that, you know, other companies, um, aim towards pushing the scientific frontier. But I think what we want to do, and, um, I would say a little bit of a framing in terms of how we differ from, let's say, Google's, uh, efforts to to to work on science is, we want to allow everyone, uh, the ability to, you know, win the Nobel Prize for themselves. Um, it's less so about us winning that at at OpenAI, which would be nice, but we want to build the tooling and the framework so that all scientists out there feel that accelerative impact. And we think we can push the field collectively.

Well, and when the discoveries that you're saying you're excited about? I mean, are there any others like specifically that that you've?

Yeah. Yeah. So, um, I think there's, you know, if you want a a huge list of of these, um, you can go on Seb's Twitter account. Um, so recently, you know, there's a GPT-5 paper on an open convex optimization problem that, you know, uh, is actually whose Twitter account?

Uh, Sebastian. Okay. Yeah. Yeah. And, you know, it's like, um, very related to some of the core ML problems that we're solving. Um, I know there was, um, I think people kind of dismiss these things as, "Oh, is it just fancy literature search or something like that?" Um, it's quite a bit more complicated than that. And, you know, there's some examples I could go into, but.

I might, I'm honestly overwhelmed at the moment, 'cause, you know, I'm sort of a generalist, but I cover biotech a lot. And it's like, uh, every two days, man, I'm walking in and it's, "Wow, we we're making an AI scientist. We we one-shotted an antibody." And then so like, part of me gets excited. And, you know, at least a handful of these companies, I know the people, and they're real scientists. And like, but then there's so much of it that I'm like, either, uh, something amazing is happening, or every, it's, it's like, it's, it's kind of too much for me to be able to discern where reality is.

Yeah. I mean, I wouldn't be surprised if it's happening in biology. Um, personally, I have the most expertise in, you know, computer science and mathematics. And, you know, we, we do have the experts there that can confirm that these are discoveries being made. So that's the thing that gives me the most confidence. But I'm not surprised at all it's happening in biology.

But like what you're saying is is kind of different than the I gr, the narrative changes every like three weeks, it seems like. But like what you're saying is is sort of different because the biggest knock, even before Andre said that, it seemed to me from the, you know, what I was listening to, um, I was listening to like a politics podcast, uh, Sager, I think it's Breaking Points, is their podcast. You know, he's, uh, pretty smart, guy who's knowledgeable, but, I mean, he's just been on AI and the lack of progress, and this is all like make-believe and all in. And so, you know, if these discoveries aren't happening, I mean, I feel like the public is aware of this.

Just to be clear, you know, um, while setting up OpenAI for science, we've talked to a lot of physicists, a lot of mathematicians. And actually, most of the people we've talked to aren't that bullish on AI. I think they still believe, "Hey, you know, um, this thing isn't something that can solve new theorems. There's no way it could do that. You know, there must be something else going on." And that's why I feel like empowering the set of people who really do believe and lean into it, like those people are going to just, you know, outrun everyone else. And we want to build the tools and convince people, like, this is the right way to do scientific research.

Okay. And so I mean, so like on that point, I mean, I grant you that everybody's definition of AGI is different, but you're, like, at least what I'm hearing is, I mean, [clears throat] you, whatever you want to call it, um, you feel like in the next year or two, is we're just seeing dramatic things happen.

Yeah. I mean, it is a bit of a meme, right? It's like, you ask someone, "When is AGI?" "It's two years away, right?" Um, and I don't think we're in that world anymore. And it it's like these results in math and science that that are giving me this conviction. But at at OpenAI, within the research, we set two very concrete goals, right? Within a year, we want to change the nature of the way that we're doing research. And we want to be productively, um, relying on AI interns in in the research development process. And within two and a half years, we want AI to be doing end-to-end research. And I think it's very different, right? Like today, you know, you come up with an idea, you execute on it, you implement it, you debug it. It means within a year, we're quite confident we can get to a world where we control the outer loop, we come up with the ideas, but the model is in charge of the implementation, the debugging.

Yeah. Okay. Are there, I beyond pre-training, when I talk to you guys, um, sometimes I get the sense, similar sort, the thing it's like we all have in our heads, at least people where I sit, that there's been this massive infrastructure build out. That, um, the models seem to get better every time you 10x them. That, um, you know, there was a, there was a story, uh, for a while that as you guys were going from like four to five, you weren't seeing the results you wanted, even though, um, you were getting more compute. But then the more I talked to you guys, the more it sounds to me like you feel we haven't actually, that things were moving so fast back then that we haven't actually seen, um, the moment where we made the leap to the 10x computer. I don't know if I asked that question very eloquently, but.

Yeah, I mean, I I do have a thought to share here, which is, you know, when people ask me, um, like, "Do you guys really need all this compute?" It's such a shocking question because, you know, day-to-day, I'm dealing with so many compute requests. And, you know, the really my frame of mind is, you know, if we had 3x the compute today, I could immediately utilize that very effectively. If we had 10x the compute today, um, probably within a small number of weeks, fully utilize that productively. And so I think the demand for compute is really there. I don't, I don't see any slowdown. And yeah, I it almost baffles me when I hear people ask, like, "Oh, do you guys really need more compute?" Yeah. Doesn't doesn't make sense to me.

And you think we in the broad strokes of the question I asked badly. [laughter] Um, do like along the lines of where you guys seem very optimistic about what you've cracked on pre-training, are you equally, not just like this demand that people want more GPUs, but are you, do you see pretty clearly that that same scaling is about to kick things higher?

Yeah, we we absolutely want to keep scaling the models. And I think we have algorithmic breakthroughs that enable us to scale the models. And, um, you know, I think there's a lot impressive about Gemini 3. Um, one thing that, kind of reading into the details that I've noticed is, you know, when you look at stuff like their SWE Bench numbers, there's still a big thing around data efficiency that they haven't cracked, right? They haven't made that much movement on it. And I think we have very strong algorithms there.

Yeah. Well, and there was this leaked memo from, I mean, Sam was sounding quite somber about Gemini 3, man. In this memo, I'm trying to find the quote. Did you? Well, you obviously, I'm sure you got the memo. Um, it like, it it seemed like a bit of a moment. Um, yeah.

Well, I do think part of Sam's job is to inject urgency and pace. And and that's also part of my job as well. Um, I think it is important for us to be laser-focused on on scaling. And I do think, you know, Gemini 3 is exactly like the right kind of bet that that Google should be pursuing. Um, you know, at the same time, you know, I would calibrate that by saying, you know, a large part of our jobs is to inject as much urgency into the org as possible. Yeah.

And, um, it is a good model. Um, I think we have a response. Um, and I think we can execute even faster to the the follow-up.

How much do you, um, get involved with things like, and I'm sure you're going to tell me exactly what it looks like with Johnny Ives' device?

Cool. Cool. Yeah.

Yeah. Like, is that, is that an area that, um, research plays in?

Yeah. Yeah, it is. And actually, um, I was just having dinner yesterday.

You can describe it to me if you want.

Yeah. [laughter] Absolutely. So it looks like this. Well, um, yesterday I was, uh, [laughter]

Yeah, just having, uh, dinner with Johnny, uh, with with some researchers as well, um, our head of pre-training and also post-training. And, um, really the way I think about ChatGPT in the future, right? Um, today, when you look at how you interact with ChatGPT, um, it feels very dumb to me. It doesn't feel very thinking native, right? And you go to it with a prompt, right? You get a response. And then it's doing no productive work for you until you give it the next prompt. And if you give it a similar prompt, you know, it's going to think for the same amount of time. It hasn't gotten smarter because you asked the first prompt. And, you know, I think the the future is going to be a world where, you know, memory is going to be a a much kind of improved feature. Every time you go go to ChatGPT, it learns something deep about you. It reflects about why you would ask this question, related questions, you know, anything. Um, and then the next time you go to it, it's going to be that much smarter. And I think it really begs the question of how do you design a device that has this as the dominating thesis. Um, and, yeah, I I thought that's been a very productive experience.

Do you have one?

Do I have one? Um,

I may or may not have one. [laughter] Um, what I think about when I think about you guys talking to Johnny is that like at Apple, you had this company that was centered around hardware. It's something that Steve Jobs obsessed about all the time.

It's like, uh, you know, it's a craft. It's like an art form. Um, whether it's you, Sam, Greg, Yakob, whomever, as far as I'm aware, none of you guys have really done a hardware product before.

Um, Sam seems to take design very seriously. I could tell from his the buildings in his house and things like that. But, you know, there's no sort of track record to speak of of like, I always thought of Steve Jobs as having like taste, you know, and then and I've I've had a couple, uh, bosses over the years like Josh Taring used to run Business Week. He kind of, he just always struck me as this guy. He just had taste, you know, whether it was the way something looked, the way a story should be. There was like this innate thing that was on this really high level. It strikes me that's kind of like what's required here. I guess that's why you have someone like Johnny on some level, but but you have to have this like, uh, back and forth. How do we know that like any of you guys have taste and are, you know, can shape a hardware product?

Honestly, we don't need to have taste ourselves. And and that that is Johnny's job. He he's our discriminator on on taste. And I think actually one thing that's been really nice is just realizing that the way they work in design and the way we work in research is there's some deep parallels there, right? There's like so much exploration and ideation. And you explore a bunch of hypotheses. You take your time. Um, and then you create kind of the thing that you're happy, the artifact at the end that you're happy about. And,

[clears throat]

Um, yeah, it's, it's been really nice to kind of have them fold into the company. And there's just a lot more direct communication about, "Here's the capabilities that we're going to ship, and here's what the form factor looks like, and how to gel them."

Okay. And I, this is like a crass way to put this, but, um, because I spend my life adoring and talking to these people, but, you know, sometimes I'm just like, man, I just don't know if a bunch of math nerds are the ones that you want making like, uh, the AI computer, you know? But I guess it is this blend that you're talking about.

Uh, yeah, I mean, honestly, um, there, you're you're right in that the people who are the best at building AI capabilities are slightly different from the people who have the best taste. And we do have teams built of people who have really great taste for model behavior. And I think there's like a, a very different kind of philosophy and a very different kind of set of questions you you need to keep asking yourself. Um, one one example of like a a good taste question, right? Like, um, you can imagine this being like in the model behavior interview is like, "What should ChatGPT's favorite number be?"

Um,

What should it, what number be?

ChatGPT's favorite number be?

Oh, okay. Okay. Okay. I'm curious what you would ask. You would answer.

What I think its favorite number should be.

Well, I have a stupid answer, which is that I went to Pomona College and 47 is this like number of lore there. So.

Okay. Okay. Okay. Yeah. I mean, that's a good answer. Yeah. [laughter]

Um, the, okay. Um, I'm going to let you go in a second. You've been really generous. I appreciate it. Um, is there Well, I'm going to ask you a question that ChatGPT told me to ask you, which is, is, uh, you know, it.

Says, "Like, if you look back in five years, um, are there any kind of like small, fragile, nascent ideas that you're seeing right now that you, your instinct is telling you might be at the heart of a big breakthrough?"

Yeah, there's a couple. I would say a handful of ideas. Um, I can't go into too much detail on them, but yeah, I'm really, really excited to scale them up. Yeah.

[laughter]

"Are there any hints, any, uh, buckets of areas where these fall?"

Yeah, I mean, I've been concentrating a lot on pre-training. So, uh, you know, some pre-training adjacent, um, a small number of ideas in RL as well, and a small number of ideas of how to put it all together.

"Yeah. Okay. All right. I tried. I tried. So, and you may or may not have a device. And no, no hints."

Yeah. Uh, no. No hints.

[laughter]

Um, okay. Well, we covered tons of ground. I really appreciate it. Is there, I feel like I'm letting the nerds down a little bit, um, as far as like the AI obsessives.

"Are there any, any, um, technical, anything you see people like kind of getting wrong about you guys at the moment that, um, you would set the record straight on?"

Yeah, I, I mean, I think the most important thing is, um, just, I think anyone at OpenAI in research would tell you that it is just a research-centric company. It's a pure AI bet. Um, at the core of the company, the ambition is to build AGI. It's to build it without distractions. And I think the, you know, anything when it comes to building products, it all flows very easily from that. Um, yeah, when it comes to what we want to do in research, it's, you know, we want to automate AI research. I think selfishly, like we want to accelerate our, our own progress. And then we want to automate scientific discovery. And of course, we want to automate the ability to do economically useful work. And, um, I think all these pillars are falling. And, and you see that kind of the big update in the last year has just been like in that second pillar of automating scientific research. It's happening.

"Yeah."

"How old are you now?"

Um, 34, about to turn 35.

"About to turn 35. Okay. Are you able to have like a social life or are you are you..."

No, honestly, not. Um, I, yeah, I think every day the last two weeks, you know, it's been work calls till 1, 2 a.m. Uh, but I, but I love doing it. It's just, um, you know, there's, there's a lot of work to get done. There's a lot of people I want to recruit. There's, um, a lot of steering that needs to be done. And like, why waste this golden moment? It's like, if we're in the middle of something like an industrial revolution, um, you got to take as much advantage of it as possible.

"Yeah. I hear stories about you sleeping at the office and..."

Oh, yeah, that, that was a fun one too. Um, no, honestly, it's just, um, yeah, I think there are times in the company, um, I think that was right after, um, Mira left and went to found their own company. Um, it just the job demands it. And, um, like, I think when I peel it all back and examine that deep emotion, it's just this protectiveness of of the research.

"That was after Mira, Mira left."

Yeah. Yeah. I spent, spent a month kind of sleeping in, in the office. And it's just like, I, I need to protect the research. They, they feel like, you know, it feels like my baby. Yeah.

"So, you guys have gone through these waves. There's the coup. Everybody's trying to steal your people."

[snorts]

"I guess everybody's trying to steal your people all the time, but you have this inflection point. Mira leaves. Meta decides they're going to fire up this massive lab. Do you think, are we, like, are we past, has everybody fired their their shot at this point?"

You know, um, you know, I have a staff meeting, right? I talk to, talk to my reports and I'm like, okay, well, here, here's the thing that I'm working on and, you know, once I get back to, once I'm done with this thread, you know, I'm, I'm going to zoom out and, you know, there's no more fires. Um, no, no. I, I've fully internalized at this point, you know, the stakes are high enough for, for building AGI that there's always going to be something. And, um, I think the important thing is just being able to understand what the important things are in the midst of all of all of these these things going on.

"Do you like months have passed since there was sort of the DeepSeek moment or whatever? I guess it was like December 2024, I think."

Yeah. Yeah. Earlier, earlier this.

"Yeah. Yeah. Or January. Yeah. I mean, is there anything, um, now, you know, it felt like people lost their minds for a second. Um, just like reflecting on it now, um, and seeing what they've done since, just, just like thoughts, I guess, on open source models and Chinese open source models."

Yeah. I mean, I think that was one of the first points in time when, um, I just realized how important it is that we just stay true to our research format. I think when, when that came out, you know, it, it went viral, right? Like everyone was like, "Oh man, like, has OpenAI lost its way? Are these models catching up?" And, um, "What's the response? What's the response? What's the response?" And I think rightfully, the thing that we did was, um, we just stumbled down our own, on our own research program. And I don't think it was the right, it was the wrong call at all. Like, um, I haven't seen the DeepSeek follow-up model. Um, you know, I, I think they're, they're a very strong lab, but fundamentally, like, let's just keep focusing on innovating. I think, um, you know, DeepSeek was a great kind of replication of the ideas in, in our O series of models, but let's just focus on innovating.

"Do you think 500 people is the, does that number grow as the company grows or this is like the optimum number for kind of like big ideas you can chase at one time?"

Um, no, honestly, uh, I feel like it can be done with even less. And, um, again, you know, as we get AI researchers or AI interns, uh, there's a real question of how do you design an around that. But, um, I'm certainly a person who cares a lot about heavy talent density. Um, when, like, I, I like to run a lot of experiments, um, of this vein. For, for instance, in quarter two of this year, um, I thought, hey, you know, I'm just not going to open up any headcount for anyone research. And, you know, if you want to hire people, uh, you got to figure out who's not [snorts] on the boat. And, um, I think these kind of exercises are quite important. You know, you don't want a, or to diffuse into something that's not manageable. And you want to keep the talent bar very high.

"I, okay, I promise this is the last question. So, yeah, sorry, I have to set you free. Um, uh, the..."

"I remember being in a meeting and..."

"I think you and Yakob were kind of on the same page here, but I remember you for sure."

"Um, sort of like this idea of of who gets attribution for a project and..."

"You seem to be of the stance that like people are obsessing over that a bit too much and and clearly AI has its roots in academia where you are very proud when you have a paper and it's a big deal and and attribution is a huge thing. Um, I think I'm remembering that meeting right and yours. Yeah. And and so..."

"What we've reached a new stage where that is less of a big deal or it's just, you, this is a company and who did what is less important."

I, I actually really love this topic. And, um, I think overfixation on credit is a very bad thing. Right. I, I think, you know, but on the other hand, I actually feel like it's important as a company for us to recognize credit, both internally and externally. And a lot of companies have actually shied away from this. You know, uh, we've moved away from publishing papers, credit lists, I think, broadly throughout the industry. But Yakob and I ended up making the call that we're going to do it at OpenAI. And of course, the counterargument is always like, man, you're like handing your top performers on a platter, you know, everyone else is going to be recruiting these guys aggressively. But I don't think that's important, right? Like, we should just recognize the people who are doing great work. We should continue to be this pipeline for creating AI superstars. And, um, yeah, honestly, it's important for us to make names for the people who are doing the best work at the company. So...

"But you seem to also be saying people, the individual researchers should maybe obsess about this less. Where or am I am I totally misremembering?"

No, I, I think there was a sentiment in the room of of that form. Um, actually Yakob and I held more of a dissenting view on that.

"Okay. Okay. Okay. It's been a while. It's in my notes. Perfect."

Yeah. Yeah. Yeah.

[laughter]

Yeah. But, um, I, I think we got to give credit where it's due, even at the risk of everyone knowing who our top talent is.

"Okay. Okay. Um..."

"I will make an even stronger statement that I think OpenAI is the place where we allow for the most external credit per capita."

"Okay."

"By a large margin."

"Okay. All right. Well, I'm check my notes on. Well, now I've got more."

Absolutely. Absolutely.

[laughter]

"Yeah."

"I just, I, I just remember it being a topic of discussion and and and I, there were numerous opinions. So, um, that, that's funny. Um, in that, okay, I lied. Last question, I swear. So, you know, you got there in 2018. I mean, it was a research company. It was a nonprofit. Um, the company started among the the founders, you know, being this, this counter to, um, to Google and and with sort of, you know, making sure AGI arrived safely was kind of the goal. Um, you came at this from high-frequency trading and and and saw these interesting things happening."

"You know, like how much in your..."

"I'm sure you're going to say you want this to happen safely. I get that. But like if you look at your career path, you're a smart, curious human who saw this interesting thing happening. It's not like a requirement that you like really give a philosophically about this or or want to see, you know, a super intelligence. Yeah. I mean, but anyway, like let's hear from you on on like why are you doing this in the first place?"

Yeah. So I think, um, really on the safety and alignment piece, um, I managed the alignment team at OpenAI as well. And, um, I honestly feel like some of the grand challenges over the next one or two years are alignment. And I think for people paying attention to this slice of research broadly in the field, OpenAI, I think has probably done the best work in the last year. And why I say that is like, there's been so much work on things like scheming, right? The more RL compute that you pump into the model, the more you can measure things like self-awareness, self-preservation, uh, potentially even situations where the model can scheme. And it's scary because the model can come to you with the right answer at the end, the answer that you expect, but arrive at it from a very kind of twisted way, right? And, um, I think as the models do more complex tasks for us, uh, having a handle on what its thought process is, um, is going to be super, super important. And, okay, Chat told me to ask you a question along these very lines, which is, I mean, you're talking about a field, mechanistic interpretability, where we're trying to..."

"Is a a term that captures trying to understand this black box and how it operates and and I guess the heart of the question was, do our skills at doing that keep up with the complexity of the AI systems or do we just get to this runaway point where it's like we're never going to learn how this thing works?"

Yeah. Um, so I think one of the decisions that went all the way back to OAI's release, which I'm very proud of, is we decided that we weren't going to supervise the the model thinking process. Um, and I think when you put incentives into the model to, uh, you know, give you a thinking process that is appealing to a human, um, it won't necessarily be honest with you, right? It won't say, kind of tell you it's, it's its true intentions. And, and so we've actually through that channel been able to maintain observing the thinking process of the model as a tool towards understanding alignment. And, um, you know, there was a paper that was published just a couple months ago, uh, with DeepMind, with Anthropic, really exploring kind of how this will evolve as a tool over time. And, and so, you know, I think we've made a lot of fairly good choices in in design here. Um, yeah, I really do worry about this world in the future where the model will tell us something super convincing, but we can't be sure, um, whether the model is aligned with us, right? Aligned with our values. And, and so I think there are a lot of interesting directions here, like, um, can you set up games, right? Or can you set up frameworks or environments where, you know, like models supervise each other, or they co-evolve together in a certain way where, like, the only, like, stable equilibrium is one where, you know, the models are honest. Um, and, yeah, I think there's a lot of very exciting work to do there.

"Okay. All right. Okay. I'll behave myself now. Thank you so much for joining us. I, I am glad I'm old enough now that I don't have to take a job interview from like a super intelligent [laughter] uh, chatbot that like, I feel like, uh, that you can't sort of try to charm your way past and and..."

"Great, Ashley. You would do great at."

"I don't know, man. I don't know. I'm feeling okay, but I've, I've, I've old enough not to have to probably do that. U Thank you, Mark, so much. You're, I know you're super busy, so thank you for your time."

Thank you so much for your time, too. All right, man. It was fun. Really a pleasure.

"Okay."

[music]

The 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. And the show is edited by John Sortland. Thanks as always to Brex and Elone Ventures for making this possible. Please visit our Substack, YouTube, and podcast channels to get more of what Core Memory makes. Thanks, y'all.

Get Our Mobile App

OpenAI's Research Chief Is Ready For War - EP 46 Mark Chen

Transcription