18 Jun 2025 · 31m · Greg Brockman, John Collison · 22 quoted facts
Greg Brockman argues that OpenAI did not set out to prove the scaling hypothesis, the idea that more computing power keeps improving AI. It observed this while building Dota-playing bots and then chased it. He says progress comes from controlling inputs such as experiments and measurements, not from promising outcomes, and that AI keeps finding new growth curves when old ones flatten.
The main tension is between rapid capability gains and real-world limits. Brockman admits the API project felt doomed, and that products are held back by missing phone and software hooks, by energy supply, and by the difficulty of making models reliable. He stays optimistic: when a capability is good enough, people tolerate inconvenience, and markets and national competition will eventually supply the power.
Brockman says the Dota team doubled compute from 16 to 32 cores and kept seeing about 2x gains each week. That convinced OpenAI to chase scale rather than start from a product problem.
His outcome-based Dota milestones failed, while milestones for experiments and features worked. Good measurement of where you stand was, he says, almost as important as progress itself.
OpenAI launched an API because it needed revenue and had no single problem to solve, and the first paying user was a text game. Reliable business uses only arrived with GPT-4 and better post-training.
If a model is capable enough, users will accept clumsy workflows like screenshot-and-upload. Brockman expects phone makers and others to then build the convenient interfaces.
Brockman calls code refactors perhaps the killer enterprise feature because they need heavy mechanical work rather than deep insight. He says Codex already finds functionality in big code bases better than he can.
It is the idea that giving AI models more computing power keeps improving them. Brockman says OpenAI did not set out to prove it but saw it working in the Dota 2 project, and he avoids claiming OpenAI was first.
After GPT-3, OpenAI needed a product to keep raising funds but had no specific problem to solve, so it offered an API and let others find uses. Brockman says people were interested but nobody was committing to build a company on it.
He says two to five years for something like a Millennium Prize problem. He imagines a whole data center, possibly with robotic labs, working on a single problem.
Brockman says it is very possible, and that OpenAI is pushing for far more power. He believes markets and national competition with other countries will eventually force the building.
Brockman expects AI to take over more drudgery and become a full coworker, with people acting as managers of agents. He singles out large refactors as a strong business use.
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Greg Brockman dropped out of MIT in 2010 to become Stripe's first engineer, later became its CTO, and co-founded OpenAI in 2015.
“Greg, in 2010, dropped out of MIT to become our first engineer and went on to become Stripe's CTO. In 2015, after he left, he co-founded OpenAI.”
In 2012 neural networks solved image recognition much better than any other method.
“in 2012, basically, you know, image recognition for the first time, you could solve with the neural net much better than anything else.”
By 2014 neural network approaches were producing great results in machine translation.
“And, you know, 2014, suddenly you're getting great results in machine translation.”
OpenAI's Dota 2 project began with 16 cores on a desktop and scaled to 32 cores, with Jakob and Shimon leading it from the machine learning side.
“We started out with 16 cores to train a little agent on Jakob and Shimon who are leading the project from an ML perspective on their desktop, and then they scaled to 32 cores”
At The International Dota tournament, OpenAI's bot went 3-0, 3-0 and then 2-1 against its first-day opponents.
“So we showed up at the international, this tournament blind. First day, we had three players that we played against. We went 3-0, 3-0, and then 2-1.”
A professional player beat the bot in one game by using a Dota item the bot had never trained against.
“it turned out that this pro that we're playing against, that he had used an item we'd never trained against.”
The OpenAI Dota bot had learned a baiting strategy that looked like a bug to its human tester.
“He'd realized that what had happened was it had learned a baiting strategy.”
OpenAI combined two Dota bots into one, which beat the world's number one player.
“So what if we just stitch the two bots we have together? And then that bot was just undefeatable. And we played against this number one player and won.”
In 2018 an OpenAI board member told Brockman that near-term AGI would imply massive economic value from AI already, which Brockman called a fair criticism then.
“In a world with the near-term AGI, you would expect massive economic value to be delivered by AI already.”
After GPT-3, someone at OpenAI proposed releasing an API and letting customers find the uses, rather than building a specific product.
“And so someone had the idea of saying, well, why don't we just make a API and let people figure it out?”
AI Dungeon, a text-based adventure game, was according to Brockman OpenAI's first paying API user.
“in fact, I believe they were our first paying user”
Brockman describes GPT-3 as an impressive demo machine whose outputs were too unreliable for real businesses until GPT-4 and better post-training.
“I think in many ways, GPT-3 was like the world's best demo machine.”
Bill Gates said GPT-4 was the best demo he had seen since Xerox PARC, and Brockman says Gates told him so the night he saw it.
“Bill Gates was saying recently that GPT-4 was the best demo he'd ever seen since Xerox PARC.”
Sal Khan, founder of Khan Academy, reacted to a GPT-4 demo by saying it was the thing for personalized tutoring.
“And we showed him GPT-4. He's like, this is the thing.”
OpenAI's ChatGPT plugins, released about two years before the interview, were limited to three at a time because the models were not reliable enough.
“Then we limited it to like three plugins at a time.”
OpenAI describes five levels of AGI: chatbots, reasoners, agents, innovators and organizations, and Brockman says the field is somewhere in level three.
“starting from chatbots to reasoners to agents to innovators to organizations”
OpenAI is building large data centers in Abilene, Texas, together with Oracle.
“I recently posted some pictures of our visit to Abilene, Texas, where we're building these big data centers together with our partner, Oracle.”
Brockman describes reinforcement learning as a data production mechanism, in which the AI trains on data it generates itself.
“If you think about the RL paradigm, fundamentally that's a data production mechanism.”
Brockman says OpenAI's Codex coding tool is good at working in large code bases, and finds where functionality lives better than he can.
“So if you look at something like Codex, it's actually great at operating in a big code base.”
Brockman compares OpenAI's strategy to Disney's: build one core asset, here the base model, and turn it into many products.
“maybe an analogy is to a company like Disney where you make one core asset like Little Mermaid, right? And then you productize it in all these different ways, right?”
Brockman began taking classes at the University of North Dakota while still in high school.
“Well, so I did in high school start going to the University of North Dakota and took a bunch of classes there.”
John Collison says that in 2017, on the way to Camp YC, Brockman told him AGI was two or three years away.
“Do you remember we were going to Camp YC in 2017 and I asked you how far AGI was away and you said two or three years.”
Brockman recalls how neural networks beat hand-built methods in vision and then translation. He credits the shift to compute and scalable algorithms.
“it felt like every day on Hacker News, there'd be a new Deep Learning”
He says OpenAI found the scaling hypothesis through Dota rather than setting out to prove it. This explains its technology-first approach.
“scale hypothesis is what we observed as the thing that was working”
The Dota story shows why milestones fail and experiments and measurement succeed in deep learning.
“You cannot set outcome-based milestones. What you can do is you can control the inputs”
OpenAI needed revenue and chose an API over a specific product. It felt doomed, and early traction came from a text adventure game.
“We knew we needed to build a product in order to be able to continue”
Brockman argues that strong enough capability makes users accept awkward interfaces, which then improve under pressure.
“if the capability is good enough, you are willing to accept any sort of inconvenience”
He predicts AI solving a major open problem and describes huge data centers focused on single problems.
“I would put it two to five years as the right number”
The hosts discuss whether power supply and permitting will cap AI growth.
“that it's energy manufactured into intelligence and that that's your only bottleneck”
Synthetic data and reinforcement learning are cited as new growth curves that replaced worries about running out of data.
“somehow you will find a new S-curve”
Brockman describes Codex as a tool for working engineers, not just app demos, and highlights refactoring as a key use.
“I think maybe the killer enterprise feature is refactors”
Asked to grade his 2017 prediction, Brockman says AI keeps surprising in better ways than expected.
“the thing we were picturing, we got something different, but we got something better”
This is totally backwards from how you're supposed to do a startup, right? You're supposed to have a problem, and we had no idea what the problem was. Is there a world where the AI becomes the manager and gives you ideas and gives you some tasks to do? This was probably the hardest project I've ever done because it felt totally doomed, right? It's like, I know every instinct, every builder instinct of mine. Did it actually feel doomed?
Oh, it felt totally doomed.
Greg, in 2010, dropped out of MIT to become our first engineer and went on to become Stripe's CTO. In 2015, after he left, he co-founded OpenAI. OpenAI is really cooking at the moment, and Greg is one of the most productive people I know.
Cheers. Wow, you really have some old-school photos.
Some deep cuts. GDB trivia. So, yeah, we figure we've got to put people at ease, you know, make people feel at home.
Very nice, thank you.
Okay, well, I just dive straight into all the questions I have, which are lost. All right. If you were not working in AI, how could one have known that something was about to start working? People are telling you that AI was the future in the 1970s, in the 1980s, in the 1990s. And then very quickly in the late 2010s, everything started happening.
Well, you know, I was someone who was not in the field. And so I remember very much what it was like. You know, 2013, 2014, it felt like every day on Hacker News, there'd be a new Deep Learning for X article. And I remember being like, what is deep learning? And, you know, I knew like one person in the field. And I asked them to introduce me to more people in the field. and I just kept getting introduced to a bunch of my smartest friends from college.
Now, if you actually look at the work that was being done, in 2012, basically, you know, image recognition for the first time, you could solve with the neural net much better than anything else. And it just blew all these traditional computer vision approaches out of the water. It's like this like learned system that was able to outperform 40 years worth of let's like write down all the rules and try to like, you know, sort of handcraft the algorithm for the task.
And it's very easy to then be like, okay, well, this approach, Sure, it works for computer vision, but it's never going to work for machine translation. And, you know, 2014, suddenly you're getting great results in machine translation. And I think that this pattern was applied in subfield after subfield.
One thing I've been wondering about is so many different things are finally working at the same time. And so we have LLens, which are obviously amazing. Then we also separately have image models really working. And we also have text-to-speech and speech-to-text working way better than they were before. And so, what's the common factor behind everything starting to work at the same time? Well, it's deep learning, right? I think deep learning is the core.
We've had deep learning for a long time.
Like, you know, why didn't deep learning work in the 1980s? Well, okay, so you look at the number of orders and magnitude of compute that we've gone through from 1940 to today. I mean, it's just astounding. You think they're all explained by compute scale-ups, applied to the right algorithms? Of course, the type of algorithm changes, and some of those results aren't even deep learning. But I think that fundamentally, it is about compute. and you need an algorithm that is scalable that can actually absorb that compute.
Was OpenAI the first company to take the scaling hypothesis really seriously? I think that claiming the first is always difficult, but I think that it is clear that we sort of succeeded much more wildly sooner than anyone else. And so I think that we have real conviction behind what we needed to do. You know, some people think that OpenAI set out to prove the scale hypothesis, whereas it was almost the other way around. The scale hypothesis is what we observed as the thing that was working for us.
and really saw it for the first time actually during our Dota 2 project. We started out with 16 cores to train a little agent on Jakob and Shimon who are leading the project from an ML perspective on their desktop, and then they scaled to 32 cores, and it felt like every week I'd come back to the office and they'd scale it by another 2x, and we had 2x performance. And just so clear, you just need to keep going.
Like, where does this thing peter out?
And it just never did. Founders get too much credit because you have an initial product that's like a pretty reasonable idea, and then you listen to the customers and you follow what's working. And so you were saying that was kind of open AI with the scaling hypothesis where you started trying to make Dota AI's work and you noticed that adding more compute worked really well. And you said, where else? We'll just throw more compute out of yield benefits.
I think that's like, to first order, correct. And I think one thing that distinguishes open AI from sort of the typical startup is we did everything in reverse, right? It's like, you're supposed to have a problem to solve. No one cares about the technology. Form DMC upfront. Exactly, yes. And for us, we really chased the technology without any idea of how it would be applied. Yes. And a lot of pursuing a technology really is you have to let reality hit you cold hard in the face.
Like there's just no other way to achieve results. Like you can't sort of will it into existence. You can't like sort of convince people that this is the thing. it's like you have to actually make the system work. We have to just sort of figure out what is the right frontier, what are the problems, what are the things that are on the edge of working and to really double down on those.
What else do you take away from the Dota work? Like how else does this, you could say you could have just started with LLMs and we could have skipped that period in the wilderness, but it sounds like it was somewhat formative for the OpenAI organization.
Well, so I think Dota had many lessons, one of which actually was a management lesson for me. I remember when we started out the project, I tried to set a list of milestones, right? It's like, okay, this state, we're going to be this player. This state, we're going to be this player. And it did not work at all. I remember our first milestone came and... Went. Exactly. And so you realize that you cannot control the outcome, right?
You cannot set outcome-based milestones. What you can do is you can control the inputs of, we're going to try these experiments by this state. We're going to implement this feature by this state. And that is what actually worked. And I remember it was one of those things that was just like a story that I could not have sort of written it any better if we'd intended to. Where we had a, we beat our in-house like best player.
And then we were playing as a semi-pro. And he was just trouncing us, trouncing us. And then suddenly we're starting to get pretty good. So we showed up at the international, this tournament blind. First day, we had three players that we played against. We went 3-0, 3-0, and then 2-1. We're like, oh no, we lost. What happened? And it turned out that this pro that we're playing against, that he had used an item we'd never trained against.
And we're like, oh no, like we're totally going to be hosed. So what do we do? Well, we just need to change the training. And so people stayed up all night to get this done. They added this extra item in there. You know, 4 a .m., they finally get the job running. That Wednesday, we're supposed to play against the number two and the number one player in the world. And our semi-pro plays against it. And he's like, this bot is totally broken.
And we're like, oh no, we clearly had a bug. something terrible has happened. And he was like, look, it's taking all this damage it doesn't need to. I'm going to go kill it. He goes in to kill it, he loses. He's like, that was weird. He'd realized that what had happened was it had learned a baiting strategy. And then we realized, well, we have a super bot, but it's so bad at the beginning because it's trying to do the baiting.
So what if we just stitch the two bots we have together? And then that bot was just undefeatable. And we played against this number one player and won. And to me, this is like the story of how deep learning works, right? It's like you kind of can't control where you're going to go. You can control everything that goes in. You can put these metrics and these measurements and you can have sort of the evaluations and that being able to gauge where you're at is almost as important as being able to make the forward progress.
But if you get all those elements right,
then you can do true magic. Aren't you also describing something that works really well for an organization where it was motivating to stay up all night? Like, you know, if the prize was impossibly far away, it wouldn't have been as motivating. But the fact that there was a near-term reward function and you were able to show concrete progress?
I think so. I think some of my favorite engineering stories have the same character. I remember you and I staying up all night to get our ISO 85, 83 integration done.
There's something about staying up all night for critical projects that actually have important history. Turns out. Yeah, all startups. I'm glad to hear the tradition of the live world. So Dota's fallen, Chess has fallen, Go has fallen. We've passed the Turing test, I think, by anyone's measure with, you know, people comment on how there was little fanfare when we did, but we seem to have pretty clearly done so. What's a good new Turing test? Well, I'll tell you two things.
One is that if you look at the strict version of the Turing test, I would actually claim we haven't done it yet. So no one's really gone that extra mile to say, can we actually have an AI that like is fully indistinguishable from a human? And it's not clear if it's even a good task, right? But I think that the right question to your point is like, well, what is the milestone that we should be chasing in terms of capability.
Like I remember talking to one of our board members in 2018 and he said, you know, look, I get that you, we're all excited about near-term AGI, but it just doesn't feel like it's on track. And I asked, well, what do you mean? He said, in a world with the near-term AGI, you would expect massive economic value to be delivered by AI already. And where is it? And in 2018, I think that was a very fair criticism.
And clearly that's starting to change now.
It feels like one thing that may really change the AI market is personalization. Up to quite recently, when you asked ChatChapD a question, it was like walking into a shop off the street. They've never met you before. They know nothing about you, whatever. That's obviously not ideal for this close part of your digital life. I'm curious how you're thinking about personalization from a product point of view. Because it feels to me like the most meaningful change since the chat interface.
two and a half years ago. Two and a half years ago. I mean, I think it's absolutely critical. And I think it is very rightly considered to be kind of a next frontier. Like, I am someone who always, when I just Google something, I go into incognito mode because I don't even want my computer to remember that history. And I always used to go for temporary chats on ChatGPT. But now my usage is totally reversed. I want ChatGPT to know, to remember everything.
I want it to remember all of my interactions because it's useful.
Okay, so you guys figured out from a product point of view how to make the memory actually work better.
And this is actually a product point of view
but also really research point of view. And I presume there's a flip-flop between the product and research where when you find something that's useful from a product point of view then the product people say and I'm just a product person like you researchers go actually make this good and then that kind of kicks off more research. It's like...
Well, we actually... So to some extent that's a failure mode in our mind, right? That I think that we really don't want to have that kind of silo and we really want to blur the lines and have people cross-glaborate and so it's very different mindsets from how you would traditionally build a product versus how you do research. And part of what had happened actually was that we had GPT-3. We knew we needed to build a product in order to be able to continue to raise funding.
And we were like, well, what product do we build? And we wrote down a list of like 100 different products, right? We could do a medical thing. And then you're like, okay, well, now we have to sell to hospitals. We're gonna have to hire doctors. And you just realize you give up on the G in AGI, right? You're gonna like go for a specific thing. And so someone had the idea of saying, well, why don't we just make a API and let people figure it out?
And again, this is totally backwards from how you're supposed to do a startup, right? You're supposed to have a problem and we had no idea what the problem was. Yeah, yeah, yeah. And so we're going back into the problem. And so this actually felt like this was probably the hardest project that I've ever done because it felt totally doomed, right? It's like, I know like every instinct, every builder instinct of mine. Did it actually feel doomed?
Oh, it felt totally doomed. It wasn't just like open-ended or something? No, it felt doomed. Because you were still doing it. I mean, it's like, at some point, if you have, there was definitely no other path. There was no other path. It was the only shot we had. I remember someone also saying, like, I can't imagine anyone paying for samples from this model. And I was like, might be right. I'm still trying to imagine it myself.
Yes. And it was just not clear. Were we above threshold or below threshold? And, you know, we showed it to people, and people were interested, but it's very different from people being like, I will build my company on top of this. Yes. And so what was the first use case to get any traction? So AI Dungeon. Was that again? There you go. So AI Dungeon was a text-based adventure game.
Oh, sure. Yeah, yeah, yeah. Okay, but that was real revenue or that was non-zero revenue? Yeah, yeah, yeah.
And in fact, I believe they were our first paying user.
And that gets you confused where you're like, ah, clearly the future of OpenAI is gaming. I know, right?
Well, back to our roots. And it's interesting too, because we had dreamed of all these applications and medicine and all these things. And you start with the gaming application, but we could see signs of life on so many other things. I think in many ways, GPT-3 was like the world's best demo machine. When we released the API, people were coming with all these cool things you could do, but making them reliable was so hard. And it really wasn't until the next generation of GPT-4 until we started to figure out how to do post-training well that then you were actually able to build real businesses on top of these things.
Bill Gates was saying recently that GPT-4 was the best demo he'd ever seen since Xerox PARC. You know this quote?
He said it to me the night that he saw it.
So that's high praise. I want to touch the medicine thing as you've mentioned it. So like you said, I think your family has kind of personal stories as you've talked about, getting very valuable diagnostic help. We ourselves actually, it's much more minor in our family, but we managed to fix a cat thanks to, you know, debugging it with an LLM. And I think this is an interesting example, right? Because so many people that I know have had some kind of experience like this.
And maybe it's because you actually don't get that much time from a doctor. Are there other examples like this medicine application where you're seeing a lot of success that many people have similar stories, but we just hear less about?
Yeah, I think it's a great question. And by the way, I think medicine is an example of one where I kind of thought it was going to be one of the last domains that we successfully built to add value in, but it turns out that the bar is so low that you just need to exceed WebMD. And so I think that we have seen sort of other areas that are like a real common theme. Like one that's very interesting right now is like the life coach, like life advice kind of application or we just talked to your eye.
That's actually really taking off. Yeah, it really is. And so I think that there's, you know, things like education is another area that just like clearly like is really having an impact. And there are studies coming out now that actually show that people are able to learn better through the use of these tools. That's to be expected, right? Like it is the Bloom Two Sigma effect in a product. Yes, yes. And that, for example, is like why, you know, Saul Khan, that's why he started Khan Academy was to think about if you can give personalized tutoring to everyone.
Yes. And we showed him GPT-4. He's like, this is the thing. We need to become a GPD poor app. And so I think that there are these really amazing applications that are affecting everyone's daily lives. Obviously, programming is another one. I think that people are seeing all across the board in professional context. We're heading to a world where just like, if you want to do productive work, you don't have access to computer, you're going to be hampered.
And so similarly, not having access to AI is heading in the same direction.
Okay, so speaking of not having access to AI, I will posit that these days, it feels like AI product development is mostly OS-limited. Is that how you feel? Like, are we stuck at the moment?
I do feel a little of the stuckage, but not to worry. It is overcomable. But yeah, I think it is true. Look, like two years ago, we released plugins in JGBT. Do you remember those? Yeah. Yeah, and that was like trying to make it so anyone could write apps that then ChatGPT could access. And the models were just not that good, right? Then we limited it to like three plugins at a time. You could have only so many functions and stuff.
And it just wasn't that reliable. And now we're in a world where MCP basically, you know, is really taking off and is a way to hook up your AI to different tools and very much like kind of trying to take that same type of idea and really make it work. Now the world that we're in is very similar where there's certain interfaces we don't have, right? Being able to access your phone and all those APIs. And there's a question of, is the model above threshold to actually use them or not?
And my observation has been that basically I think that there is maybe a lag of six months of different interfaces that are hard to access. But once we have a model that's good enough, we will find a way. People will find a way. And so I think that we're in a world where I have every expectation that we will get the future that has been promised. but it's just going to take some work.
I feel like there are many moments where I'm using my phone and I want a single button where it just is like, you know, ChatGPT, what do you think of this? Like, I need your comment, I need your fact check, I need your explanation, something like that. And you take a screenshot and you like go into ChatGPT and you click upload photo. And, you know, it feels like very 1993 versus the button on my phone that just says, hey, ChatGPT, what do you think about this?
And obviously you guys are not empowered to go build that. That's what I mean by it feels somehow like we're a little operating system.
Well, I definitely get it. But I'll say, I think that there are two dimensions. And this is kind of how I've been thinking about things since we released the API back in 2020. There's capability and convenience. What you're referring to is the convenience, right? It's like pretty inconvenient to do the screenshot and paste it. But the thing is, if the capability is good enough, you are willing to accept any sort of inconvenience, right? It's like if this, you know, by taking a screenshot and showing it to ChatGPT, it could give you amazing insight.
It could tell you like how to, you know, build Stripe in some way. And it takes you like a month to do it. You have to crawl to the top of some mountain. Like you'll do it, right? Like the convenience will not stop you. And so the point that I'm trying to make is that if the capability is high enough, people will start doing a specific flow.
They'll discover the use cases and the convenience will just catch up.
And then in the convenience, there's so much pressure, right? There's pressure on the phone manufacturer. There's pressure on us. there's pressure on everyone in order to bring down the...
So I just need to be patient. It'll be great in three years' time. Yes. And really lean in. Use the AI. A criticism people like to levy of AI is, yeah, it's great and handy and all, but it hasn't come up with a single novel advance in mathematics or science. Well, you could have if, you know, you'd become a mathematician. But, you know, but humanity has, you know, for keeping the score, the scoreboard. What do you make of that criticism?
Just wait. Okay, so you think, like, take, you know, one of the Millennium Prizes, you think we plausibly will see that? I think for sure.
I mean, there's no question.
Like, two years, five years, ten years?
I think that is the question. It's just timing, right? That is my question. I mean, I would put it two to five years as the right number. And I think, ultimately, this comes back to the question of benchmarks, right? Is that actually being able to solve a Millennium problem is a pretty high bar. Yeah? Pretty high bar. And once you can do that, there's so many other things that will definitely be possible. And I think that we're starting to see the leading edges of this.
And to me, if we look at our definition of AGI, we recently started talking about this framework of thinking about levels of AGI, starting from chatbots to reasoners to agents to innovators to organizations. Five levels. And we're basically somewhere in level three right now. And level four, this innovators, that's going to be different. I recently posted some pictures of our visit to Abilene, Texas, where we're building these big data centers together with our partner, Oracle. And imagine taking that whole data center and just thinking hard about one problem.
Imagine just thinking about how to solve a millennium problem or how to cure a specific kind of cancer. Maybe it needs access to some apparatus. Maybe it needs access to robotic wet labs. Maybe it needs access to different tools in the world. But that level of computational power coupled with the ability to experiment and learn from your ideas, that is going to be something the world has never seen.
Okay, so yet again, we haven't put a respectable amount of compute on these problems compared to what we will be doing. There's all these tiny little computers. So that actually gets to, in terms of these scaling laws, do they eventually run out because we just run out of compute? or do we eventually get to the point where we're inventing new kinds of nuclear energy and so that is what unlocks the next level? A lot of energy that comes online now is for data centers, which was not true when you guys started training GPT-2.
And so isn't that just the upcoming bottleneck? I mean, as it should be, right?
It really should be that it's energy manufactured into intelligence and that that's your only bottleneck.
But I'm saying that'll be like quite a plateau compared to the exponential growth we've seen over the past few years. Well, I think... Unless things really change in terms of permitting and plans for building and everything like that.
This is, I think, the core, right? Is that if you look at every trend in this field, there's these exponentials, these S-curves that sum up to exponentials.
Sure, but these exponentials were mostly existing in like tech, Silicon Valley space, where it's pretty easy to have exponential growth. it's pretty hard in permitting and real estate and damning rivers and building nuclear power plants. It's harder to have exponential growth.
Let's see how fusion pans out.
Yeah, okay. But even fusion, you know, most industry observers would say is still kind of five years away. And so I've heard it's the next five years of power growth.
So look, I think that it is very possible that we end up bottlenecked on energy. And that's actually one reason that we've been spending a lot of time really trying to advocate for the fact that we just need far more power. And I think that my observation of the market is that ultimately the capitalist markets do provide. I think there's this absolute tsunami of demand that is coming our way. But I feel some confidence that, again, when there's enough pressure, when there's enough clarity of this is the bottleneck, and it's not just for any company, right?
It's really for national competitiveness. And you look at other countries that are just building huge amounts of power far more than we are. I think that actually for America to remain competitive, there is just no choice but to build. We've got to figure out power. We do.
Speaking of bottlenecks, everyone was talking about the data wall in 2023. I think this is an interesting thing where no one is talking about the data wall anymore. And yet, it doesn't feel like AI progress has slowed down. Is it just test time confused? Is it like people were wrong about the data wall and where it presented a bottleneck? Like, is there actually still a data wall, but it's two years away?
It's basically all of these things, right? It truly is. It's like you keep changing the paradigm. That is the real core of the Kurzweil view of the world, is that, fine, this one way of doing things taps out. And if you just look at that one way of doing things, you feel hopeless. You feel like this is it. But somehow you will find a new S-curve. And I think that's what's happened. For example, synthetic data, for example, reinforcement learning.
If you think about the RL paradigm, fundamentally that's a data production mechanism. And just that the AI happens to be training on its own data and then you learn it very rapidly and then you learn on that. And each of these has taken us much further. And so I think there are lots of algorithmic ideas, lots of techniques, lots of ways of even using the existing data better. And so I think that fundamentally the S-curves continue.
And if you zoom out, it all looks smooth and uninterrupted.
So it's kind of like chip miniaturization where, you know, each generation people are like, okay, well, that's the smallest you could possibly make a chip. That's it, we're done with miniaturization. And somehow we figure out a way.
Yes. And now one difference with chips is I think at the end of the day, there is some clonk limit.
Right, but we've never been that close to that. Right, right. Yes, yes. Where does AI coding go? And in particular, Vibe coding is all the rage right now. It's kind of the term of 2025. It's sort of working. It's very impressive. No one is really fully letting AI software engineers run end-to-end in production. So I'm just curious, what are your one to two year predictions on what happens with AI coding?
Well, my general observation is that once something kind of works in this field, the next gen is going to be great. And so I think that's where we are right now for AI coding. And so I think what we're going to see is AI is taking more and more of the drudgery, more of this like pain, more of the kind of parts that are not very fun for humans. Now, one thing that's very interesting is that I think that so far, the Vibe coding has actually taken a lot of code that is actually quite fun and left behind the review and the deployment and these things that are not fun at all.
And so I'm hopeful that we're actually going to be able to make a lot of progress on these other areas as well. But fundamentally, we should really end up with a full AI coworker. And I think it really will be anything you want to create. You can be the manager, right? And you can really have this team of software engineering agents. Now, the thing that I think will be very interesting to see is, is there a world where the AI becomes the manager and gives you ideas and gives you some tasks to do?
And that's something that, again, is just totally backwards in terms of how we think about it. But are there ways in which you can actually have outcomes for companies and actually have people whose jobs become much more meaningful because they have an AI who really deeply understands them in the same way that your AI doctor really deeply understands all of your needs?
But isn't part of the common thread that we're talking about here often places where AI tools underperform, it's because they're trying to do something generally, you know, like voice recognition is not that good because it's trying to recognize all voices as opposed to trying to recognize, you know, my voice in particular. And similarly with AI coding, they work well in places where you need no context at all. And, you know, we're single-shotting an app based on, you know, publicly available libraries.
And in places where you have to understand a million-line code base, fully figured out how to do a good job of that? Is that a fair parallel to draw between all these challenges? Well, I think there's two things in there.
One is that I think this is already changing. So if you look at something like Codex, it's actually great at operating in a big code base. I ask it for where functionality is implemented and it's better than I am at finding it. Which is kind of a wild fact. It's super cool to see it grapping around and just going and exploring. And actually this is one thing that we really shot for with Codex, was to build a tool for software engineers who are not necessarily vibe coding, right?
It's not about building a new app from scratch, which is a cool demo, but like that's not actually how most software gets written. And actually I think maybe the killer enterprise feature is refactors, right? It's like rewriting your Copa app or changing your -
When Facebook did hip hop, you know, to do static PHP.
Exactly, right? And if you think about it, like the amount of like deep sophisticated thought that is required to accomplish every factor is actually not that high. There's a lot of mechanical work that's just the sheer volume of it that's hard. And like, that's an AI-shaped problem for sure. So I think we're going to see a lot more productivity on all sorts of AI tasks as a result. Now, we are in a world where, you said a second thing, which is, you know, like maybe you need to kind of narrow down more.
Like, I think that the way these models work is you actually do want one model that knows more and more things, and you want it to have some personalization to you. But the fact that they kind of have this one base model that kind of knows everything is actually a very useful starting point. So I do think that you're going to see a world where we'll have more and more capable base models and figuring out how do you really connect it to all of your organization's code and context and history.
How does OpenAI decide what's products to do? I'm just curious how you think about when to develop specific products or when you think, oh, you can truly do that with ChatGPT and that's good enough.
Yeah, this is a really tough question, right? It's something we really struggle with. And I think that over time, I guess actually when we first launched ChatGPT, we were left with this, well, we're an enterprise business and we're a consumer business. And that seems terrifying, terrifying as a startup. And I remember talking to one of my board members who said, it just feels like what you have is an unfocused strategy at first, right? Because you're just doing all these different things.
But if you think about it, maybe an analogy is to a company like Disney where you make one core asset like Little Mermaid, right? And then you productize it in all these different ways, right? You think about Little Mermaid, the ride, the lunchbox, the t-shirt. And I think that we have some element of that. We have the core model. And then we have a question of, well, what are the applications that this can add a lot of value to quickly, right?
With like a small amount of additional work. And so I think the question of what areas to go into are how far does it take us off the general path? on the return on how important is this domain area, especially for achieving the bigger goal. How much synergy is there across it with respect to other things we work on? So coding is one where there's very clear synergy, very clear ROI, right? Because if we can speed ourselves up, that's something that accelerates everything.
How is being from North Dakota, how's it shaped you?
Look, North Dakota was an amazing place to grow up. Well, it actually, come on. I've been there. And it was great. Look, it was incredibly safe. Our doors didn't even have working locks. Like, it was that kind of place. I had a lot of freedom academically. So sixth grade, my dad taught me some algebra. Seventh grade was the first time they split you into advanced math. And so I was going to be taking pre-algebra. So my mom took me to go see the teacher.
The teacher, and we asked, can you skip? And the teacher looked at us very condescendingly and said, every parent believes that their child is special. I can guarantee your son will be plenty challenged in my class. And so after a month of me sitting in the back just playing games on my calculator, and she'd call me randomly to try to trip me up. I'd just look at the board and be like, 2x. She said, okay, fair enough.
Your son has nothing to learn in this class. And so they moved me into eighth grade algebra. But then eighth grade rolled around, and I had no more math left in my middle school. That's right. You went to college, right? Well, so I did in high school start going to the University of North Dakota and took a bunch of classes there. But also, I was connected to a lot of people who were the top math kids in the country through things like Math Camp and the math competitions.
So you're saying the social scene was not too distracting in North Dakota?
Not too distracting, but it was definitely fun. Last question. Do you remember we were going to Camp YC in 2017 and I asked you how far AGI was away and you said two or three years. Did I say that? You did, yeah, yeah, yeah. I don't see the recording. I'm just trying to think, like, was I right? Were you right? How should we grade that? Because we didn't get AGI, but we didn't not get AGI either.
And so I'm curious if you have any reflections from your own AGI prediction journey.
I think that we are, I will say, I think that AI is surprising. I think that that is like the single most consistent theme is that the thing we were picturing, we got something different, but we got something better, more magical, something that is more helpful. And so I'm actually quite happy with that. Now predicting where you go, it's again, it's really hard to manage the outputs here. One goal of OpenAI that we have successfully achieved is every year to have at least one result that just feels like a step function better than anything before.
I think as long as -
You know when you see us kind of, you just want to have one really awesome AI feeling thing each year. That kind of thing. I like that. Yeah. That's a good way to tie it back, which is, you know, the way we grade that prediction is that you've stopped setting metrics based on outputs.
Yeah, exactly, exactly. Yeah, yeah. Yes. But it does feel we're getting really close to something pretty magical.
I agree.
Thank you. Thank you.