So, how did OpenAI come about?
I knew I wanted to do a startup because I felt like that was something.
But you were just in a startup. Stripe was a startup.
It's true. But I felt like Stripe, the problem that we were solving was not my problem, right? It wasn't the problem I had grown up thinking about. It was an important problem that I dedicated myself to that mission for a number of years, but I felt like it was going to succeed with or without me. And so then I had a first moment to really think about what is a mission that I want to dedicate myself to where I would spend the rest of my life working on this problem just to see it play out in a slightly better way. And it was very clear to me that top of the list was AI, right? If you could actually make a difference in how AI will play out in the world, like that would be a life well-lived.
When you were thinking about leaving, Patrick told you to go talk to Sam Altman. What happened in that conversation?
Well, so Patrick had said, 'Sam has seen lots of young people in your situation.' And Patrick, I think really hoped that Sam would convince me to stay. A few minutes of talking to Sam, he's like, 'Okay, you clearly have already decided. It is very obvious.' And so he asked, 'Well, what are you planning on doing next?' And I said, 'Well, I'm thinking about doing an AI company.' And he said, 'I'm also thinking about doing something in AI. We should keep in touch.' So I had talked to Sam maybe one more time after I was leaving Stripe and he asked, 'Are you still thinking about doing something in AI?' I said, 'Yes.' He said, 'I'm also starting to get more details and putting together this dinner in July.' And I flew out for the dinner. And the thing that I remember was a topic was, is it too late to start a lab with many of the best researchers? Is it possible?
neutral point of view and trying to be fair and balanced. And I think that sometimes when you see these screenshots on Twitter that they're not always fully honest themselves in terms of where they came from, either because there's some memories that are behind the scenes that tweak the answer in a certain way or hidden instructions or previous parts of the conversation. And so sometimes it's also there's just no right answer. And so you, you know, you can have like a question you say answer in one word and no matter which one you say, you're going to get some sort of claim of bias. And so I think that to some extent the core of it in my mind is that we are, yeah, we care a lot about truth and about having an AI that really represents you.
Do you think the models evolve to tell us what we want to hear if they're based on reinforcement learning? So if I lean left, it's going to tell me an answer that leans left or if I lean right, it's going to give me an answer that leans right.
Well, so we've actually gone through an evolution of how we train the models to user preferences and that we've seen that at one point like last year that the models really did start to lean into telling you what you wanted to hear, saying, 'Oh, that was such a great answer.' And we reacted to that. We said that this is not how we want our models to operate. And we made changes because the true thing we want the models to be aligned to is helping you solve your goals, your long-term goals, right? And maybe in the moment it feels good to be told that was a great question, best question anyone's ever asked, but that's not what you actually want. Maybe there are some people, but it's not what most people truly want. And so we've actually made great technological improvements to make sure that our AI training does not result in what is called hacking the greater, right? That we really want to make sure that there is a good signal there that is about the goal, not just your short-term what's going to get you a quick hit. And that to me is maybe the most important part of the vision for where our personal AI, personal AGI is going to take us is to really make sure it's not just about something that looks good in the moment. It's really about alignment with your long-term well-being, your long-term goals, the thing that you actually want. And that is what I think will most empower people, right? Is really put you in the driver's seat because you will have this entity that is there operating on your behalf 24/7, right? You're asleep. It's out there trying to figure out what is it that Shane wants? How can I do it better and is actually able to accomplish it?
Are we in a global AI race?
I think we are certainly in a global AI renaissance and I think that the dynamics between countries are not yet fully defined. We have this concentration of where the breakthrough algorithms come from in the US in western companies. There's clearly a lot of innovation happening around the world but I think exactly the balance of dynamics and how like which countries rely on which providers, all of that is something that I think is still being determined.
Is there a consequence do you think for the United States not being the first country to reach AGI?
Well, I do think that leading in AI is very critical for America because I think that this is how you can ensure that democratic values are protected and preserved. And I think that every country is also starting to realize that they need some sort of sovereign AI strategy. They need to, if this is becoming the basis of economic security, of national security, they need to participate somehow. And if you look at a lot of the efforts by the United States to think about how to manage chip exports, how to think about technology exports, there's something where if you lean too far out, then everyone else has to develop their own competitor or rely on someone else who's building this. If you lean too far in, then maybe you lose your advantage. And the question is, how do you balance those? How do you maintain your leadership? But leadership is not just about being ahead. Leadership is about also bringing along the world with you.
Are other countries stealing advancements? I've been reading a lot about distillation.
There's certainly a lot of attempts to distill models and that comes from companies in the US. It comes from all over the world. But I think that it misses the core point which is that the way this technology is developing is it is on an exponential and anytime we have a model we've already moved on to the next one. We're already moving to the next level. So we put in a lot of effort to protect against distillation, make it harder to do, especially with things like chain of thought and other parts of the model that are not really necessary to get the benefits to someone, to get the outputs to someone. But the core advantage that we have, the strength that we're building up over time is really about not just any one model. It's about the machine that makes the models.
Oh, is that why you guys stopped showing reasoning?
That is part of it. So there's two reasons. One is to think about distillation. The second, in some ways more important, is that we had this insight when we first developed the reasoning paradigm that it gives us an interpretability mechanism we had not been anticipating because you can really read the model's thoughts. You can see exactly how it got to an answer. So you can interpret how like what was actually motivating that answer. Now the problem is if you train the model to have a chain of thought that looks good then you lose all the faithfulness, right? It's just going to be like the model knows that part of the answer that is desired is for the chain of thought to look a certain way and so it may not be representative of how it actually arrived at that answer anymore. And so we made an early decision to say we want to avoid any temptation to train these chain of thoughts to look favorable, to look like something you could present to a user. And so that really made us lean out for multiple reasons, for competitive reasons, for safety reasons from the idea of showing these intermediate thoughts.
It seems like the current trend right now is to release preview models. Is that because we're compute constrained, do you think?
I would say that we in general are heading to a compute world. Like if you think about the amount of value that these models can produce for someone, it's extreme, right? It's not just answering a quick question anymore. It's not just even answering, you know, giving you access to health information. It's really going deep and spending a lot of tokens to put together a bunch of different data sources, search through your enterprise knowledge base to actually be able to solve this hard problem, to write that software that's better than a human would be able to. All of that is something that is like hard. And if you look at the progress that we made between GPT-4 to 4o to 4o mini to o1 to o3, it's been extreme. And these models are getting extremely better at understanding your intent, molding to what you want to accomplish. And we also put them in these surfaces like Codex that make them very usable so that you as a developer you can really fly, right? That you can achieve more than you would have dreamed otherwise. And all this is powered by compute fundamentally. And there's not enough compute. That if you just wanted enough compute for, you know, you wanted one GPU for every person in the world you're talking like 8 billion GPUs. We are not on a trajectory to build anywhere near that level of compute, right? It's like, you know, hundreds of thousands of GPUs, like that's a pretty big fleet these days, millions of GPUs coming up. It's not surprising that there is too little compute in the world and that we're going to need much more in order to really be able to bring this technology to everyone. And then in terms of the training, I'd say that the way that we tend to launch things, so we have put in a lot of effort to make sure we are building compute in anticipation of what we see coming. And so I think we're going to be very focused on our mission of bringing these models to everyone, making them widely available.
You guys were teased for putting so much effort, money into data centers. How do you think that's playing out now?
Well, I think it's going to give us an advantage and I think it's going to be something that's an advantage not just for the business, but for actually delivering on the mission of bringing this technology to everyone.
Because you guys, like, you saw that way in advance. You get teased for it by almost all of your competitors.
Who's laughing now?
Yeah. I mean, I think our competitors are not having a good time on compute. Let me put it that way.
But you must have seen something that they didn't see. Like that's the, I mean, everybody was in a very similar, or it seems at least from the outside, everybody was in a very similar technological space. They all knew this was coming and yet you guys had the boldness to make that bet with a hundred billion dollars.
Like, but that is the core of OpenAI is really encountering reality as it is, really thinking about what is the implication of what it is we'll accomplish in the next 6 months, the next 12 months, next 10 years. And that is true for the grand mission, it's true for day-to-day how we design different pieces of our software, and it's true for things like scaling up compute. And I think that we are deeply motivated by bringing this technology to everyone. And we think about lots of different mechanisms for how to do that well and safely.
Do you think data centers eventually get dedicated towards a problem like you'll have a huge data center in North Dakota and it's just on solving cancer and that's all it's doing?
How far away are we from that?
I think that this kind of thing happening this year is not out of the question and it's really amazing if you think about it having this giant machine, right? And have you, have you been to any of these data centers?
No, I've seen them online but never in person.
