Transcription
As we look ahead, we know AI has the potential to shape the future, and shape it for the better. Generative AI has unlocked everything from code writing to philosophical reasoning.
That journey started with foundation models. They've enabled billions of users to tap into the power of data they never would have had access to. They have changed the game for us all. And few have shaped that future more than OpenAI. Through the explosive rise of ChatGPT, used by over a billion people daily, they've redefined how we interact with data and intelligence, and how we imagine the future of work.
And at the center of the transformation is one of the most influential voices in technology today. He's led OpenAI from groundbreaking research to products now impacting billions, driving the evolution of AI on a truly global scale. Please join me in welcoming the founder and CEO of OpenAI, Sam Altman. [Music] [Applause] [Music]
And joining us to moderate the discussion, the founder and managing partner of Conviction and a good friend, Sarah Guo. Thanks. Welcome, Sam and Sarah.
Well, it's amazing to be back. Um, I was saying in a straight art, this looks like a rock concert, but for data.
For data people.
Yeah, you were here two years ago.
Yes, but it wasn't as amazing.
So, to kick us off, um, Samar, what advice would you have for enterprise leaders navigating the AI landscape in 2025?
I think just do it. Like there's still a lot of hesitancy, and the models are changing so fast, and there's all this reason to wait for the next model, or you're going to sort of like wait and see if this is going to shake out this way or that, or if you should build, you know, with thing A or thing B. I, I, I as a general principle of technology, when things are changing quickly, that companies that have the quickest iteration speed, um, and sort of make the cost of making mistakes the lowest and the learning rate the highest win. And certainly what we're seeing with enterprises and AI is the people that are making the early bets and iterating very quickly are doing much better than the people that are waiting to see how it's all going to shake out.
Straight, what would you say?
I'll, I can't agree with that more. And the thing that I'll add on is curiosity. I think there's so much that we take for granted about how things used to work that just aren't true anymore. And going and experimenting, and lots of people, OpenAI, Snowflake have made the cost of experimenting very, very low. You can run lots of little experiments, get value from it, and build on that strength.
I want to echo what Sam said again, which is it's the folks that can iterate the fastest that are going to get the most value from things because they know the things that are going to work, the things that are not going to work. They can navigate this rapidly changing future. There's never going to be, I think, not in the next few years, one perfect moment where everything is settled down and we can then figure out what we do.
How would your advice differ from what you would have said last year?
I think I'd have said the same last year in terms of I think curiosity is the most overlooked thing, and I think it's fine to make mistakes. You need to figure out situations in which it doesn't matter that much, and there are lots of situations like that. But the technology is also rapidly maturing. You know, you can absolutely use ChatGPT now for getting information about the latest events because it knows when to use web search to provide that. And so there are lots of applications like chatbots, whether it's structured or unstructured data, the technology is mature, you can adopt it; yes, you can always push the boundary on what else you can do with it, there are edgier agentic applications, but far away from the frontier, I think this technology is actually ready for mainstream use.
Interestingly, I, I, I wouldn't have quite said the same thing last year, um, I would have said the same thing to a startup last year, but to like a big enterprise, I would have say like, I, I would say like, uh, you, you can experiment a little bit, but this is maybe not totally ready for production use in most cases. And that has really changed. Our enterprise business has gone like this. And we talk to big companies who are now like really using us for a lot of stuff and say like, what's so different? And, and, and we're like, did it just take you a while to figure it out? And they say that was part of it, but it just works so much more reliably. It works, you know, it can do all these things that I just didn't think were going to be possible. And it does, it does seem like sometime over the last year we hit a real inflection point for the usability of these models.
Now an interesting question is what will we say differently next year? Um, and I think we'll be at the point next year where you can not only use a system to sort of automate some business processes or build these new products and services, but you can really say, I have this hugely important problem in my business. I will throw a ton of compute at it if you can solve it. And the models will be able to go figure out things that teams of people on their own can't do. And the companies that are have gotten experience with these models are well positioned for a world where they can say, okay, you know, AI system whatever, go, you know, like redo my most critical project, and here's a ton of compute, think really hard, just figure out the answer. People who are ready for that, I think will have another big step change next year.
I, I think, you know, given reasoning and applying more compute to hard problems and, you know, uh, the introduction of agents into some workflows, uh, there's a view that memory and retrieval have to change a lot. What do you think is the role of memory and retrieval in this, you know, this next era?
I think things like retrieval have always played a key role in making generative AI technology grounded when it needs to be grounded. If you're asking a factual question, you want a reliable answer. So, you know, on GPT-3, we built web search scale systems back in early 2023. So, whenever you asked a question that needed a reference point from the real world to be able to answer, like breaking news for example, you could provide that context. Similarly, knowing how you have tackled certain problems before, memory, your interactions with, with, with a particular system can greatly influence and make that system better for the future. I think their role will continue to increase as you use these models for more and more interesting tasks. And the more context you have, I think the better these systems get, both from an interactive perspective, but also from an agentic perspective.
Sam, is there a framework you can give every leader here to think about like what can agents do today and next year?
Um, I mean, the, the coding agent we just launched called Codex has been one of my like feel the AGI moments. You like watch this thing. You can give it a bunch of tasks. It goes and works in the background. It, it's really quite smart. It can do these long horizon things, and then you get to just sit there and say yes to this one, no to that one, try again. And it is able to just kind of like connect to your GitHub and, you know, at some point it'll be able to also watch your meetings if you want and look at your Slack and read all your internal documents, and it's just doing incredibly impressive stuff. And you know, maybe today it is like a sort of intern that can work for a couple of hours, but at some point it'll be like an experienced software engineer that can work for days. And then we'll see this for many other categories of work. And so you see you hear from companies that are building agents to automate most of their customer support or their outbound sales or any number of other things. And you hear people that talk about their job now is to assign work to a bunch of agents. Um, look at the quality, figure out how it fits together, give feedback, and it sounds a lot like how they'd work with a team of, you know, still relatively junior employees. And that's here. It's not evenly distributed yet, but that's happening. Um, I would bet next year that in some limited cases, at least in some small ways, we start to see agents that can help us discover new knowledge or can figure out solutions to business problems that are kind of very non-trivial. Um, right now it's, it's very much in the category of okay, if you've got some like repetitive cognitive work, we can automate it at a kind of a low level on a short time horizon. And as that expands to longer time horizons and higher and higher levels, you know, at some point you get an AI scientist, uh, an AI agent that can go discover new science, and that will be kind of a significant moment in the world.
Uh, you said it was a moment, you know, Codex and experiencing, you know, coding agents, uh, was a moment you felt the AGI. So I have to ask you about that, like what is the, what is the definition of AGI to you now? And, and, um, how far away are we from it? What will that mean for us?
Um, I think if you could go back to most people, if you could travel back in time, just 5 years, 2020, let's say, uh, it's like the dark ages for AI, though. Actually, that's a very interesting time because I, I think that was, if we could go back exactly 5 years, I may get this wrong, but I think that was just before we launched GPT-3.
Okay. So the world had not yet seen like a good language model. And if you could go back to that moment and show someone ChatGPT today, to say nothing of Codex or anything else, but just ChatGPT. Uh, I think most people would say that's AGI for sure. And you know, so we're great at adjusting our, uh, expectations, which I think is like a wonderful thing about humanity. Um, I think mostly the question of what AGI is doesn't matter. It is a term that people define differently. The same person often will define it differently. Um, the, the thing that matters is the rate of progress that we have seen year over year for the last 5 years should continue for at least the next five, probably well beyond that, but hard to say. And whether you declare the AGI victory in 24 or 26 or 28, um, and whether you declare the super intelligence victory in 28 or 30 or 32 is way less important than this one long beautiful, shockingly smooth exponential. Um, all of that said, to me, a system that can either autonomously discover new science or be such an incredible tool to people that our rate of scientific discovery in the world like quadruples or something. Um, that would, that would satisfy any test I could imagine for an AGI. Some other people would say it's got to be a system capable of self-improvement. Plenty of people would say like ChatGPT with memory today, very AGI-like, certainly across some of our early tests like Turing test that people used to say was the, was the target.
Um, okay. Scrolling back to 2020, Sudar, do you remember what the first OpenAI model you used when you were building search was, what year?
We were actually using, uh, GPT-3 playground and running little experiments with it and then with the, and then with APIs, we couldn't afford, uh, GPT-3 or running it at web scale. So we basically reverse-engineered how we could do this with 7 billion, 10 billion parameter models. Yeah. But already you could see greatness for me when you saw this problem called, um, abstractive summarization actually get tackled nicely by GPT-3. This is basically taking a blog that's 1500 words and writing three sentences to describe it. It's really hard. People struggle with doing this, and these models all of a sudden were doing it. That was like a bit of an aha moment for me. If you could do this on the entirety of the web corpus, you of course have search which can figure out which 10 pages to look at. That was a bit of an aha moment when it came to, oh my god, there is incredible power here, and of course it's kept adding up.
At, at what point in, you know, your journey as an entrepreneur or a CEO of Snowflake did you think like, wow, I, I mean, I, uh, employ a former NEVA person as well, and part of my premise was like everything is search or search plus in, in this era. Did, when did you have that thought?
It's about setting context. Um, once you look at and interact with these models or think about any problem, you also want to have a way to narrow down the lens of what you want it to operate in. And it's a very powerful and generalizable technique. Even if you look at many of the post-training techniques that have come, it's a little bit of okay, take this incredibly powerful model, give it context for what's worked, what's not worked, and use it to improve what it is is producing. I would say it's, it's more a general concept than a specific tool for how do you make something happen. It's all, it's all about the right context setting. There's always an infinity of context. Humans solve it by what we call attention where we focus on something. I think of search as a tool for setting attention for a model.
And do you agree with Sam that it's really just, you know, being on this exponential capability curve, or is the, is there a definition of AGI that matters to you or matters to customers?
I think it becomes a matter of debate, like Sam is saying, in, uh, I think sometimes it's also a philosophical question that I would liken to, I don't know, does a submarine swim, um, at one level it's absurd, but of course it does, and so I see these models as having incredible capabilities that we will, like any person looking at what things are going to be like in 2030 would just declare that's, that's AGI, but remember you and I would, to Sam's point, would say the same thing in 2020 about what we are seeing in 25. To me, it's the rate of progress that is truly astonishing, and I sincerely believe that many great things are going to come out of it, and similar to again, how do we feel about the fact that a pretty decent computer can beat every person in the world that can play chess, doesn't matter, we still have people that play chess that are very, very good at it, so I, the definition matters. What's it? It's more popular now. It's more popular now than before. Um, it's the same for Go. So I think there is a lot that we will learn from it. The actual moment I don't really think matters a whole lot. Um, I have a hunch personally that when people ask about AGI, I think they're really asking about consciousness. They just don't always frame it that way, or at least some large subset is, which is like a more, as you said, philosophical question.
Um, I have to ask you because we have, you, you're training more models, you know, you see the next capabilities before anybody else does, uh, what emergent behaviors are you seeing in the next set of models that change, you know, how you operate, what you want to build from a product perspective, how you're running OpenAI?
Yeah, the, the models over the next year or two years are, are going to be quite breathtaking, um, really there's a lot of progress ahead of us, a lot of improvement to come, and like we have seen in the previous big jumps, you know, from GPT-3 to GPT-4, businesses can just do things that totally were impossible with the previous generation of models, and, and so what an enterprise will be able to do, we talked about this a little bit, but just like give it your hardest problem, if you're a chip design company, say go design me a better chip than I could have possibly had before, um, if you're a biotech company trying to cure some disease, say just go work on this for me, like that's not so far away. Uh, and these models' ability to understand all the context you want to possibly give them, connect to every tool, every system, whatever, and then go think really hard, like really brilliant reasoning and come back with an answer and, and have enough robustness that you can trust them to go off and do some work autonomously. Like that, that I don't know if I thought that would feel so close, but it feels really close.
Is there any intuition you can give everyone here, uh, for like what knowledge is in scope or soon to be in scope, because I, when I think about core intelligence, I'm like, well, you know, I'm reasonably smart, but I don't have a perfect physics simulator in my head. So like, how should I know what's, what's possible?
The, the framework that I like to think about this is not something we're about to ship, but like the platonic ideal is a very tiny model that has superhuman reasoning capabilities. It can run ridiculously fast and one trillion tokens of context and access to every tool you can possibly imagine. And so it doesn't kind of matter what the problem is. Doesn't matter whether the model has the knowledge or the data in it or not. Like the model using these models as databases is sort of ridiculous. It's a very slow, expensive, very broken database. But the amazing thing is they can reason. And if, and if you think of it as this reasoning engine that we can then throw like all of the possible context of a business or a person's life into and any tool they need for that physics simulator or whatever else. That's like quite amazing what people can do, and I think, you know, directionally we're headed there.
Uh, amazing. I, um, want to ask both of you, uh, for a more like conjecture question. If you had a thousand times more compute, the, the original thought was infinite, but that gets silly. A thousand times more compute, what would you do with it?
I mean, I guess the super meta answer, I will give a helpful one after this, but maybe the real answer is I would ask it to work super hard on AI research, figure out how to build like much better models, and then ask that much better model what we should do with all the compute. Doing your hardest problem.
Well, I mean, I think that's, I think that would be the rational thing to do. Um, well, it means you really believe the answer. You do have to really believe that. I, I think the more helpful thing I would say is we see all of these cases now inside of ChatGPT or inside of enterprises that are using our latest models where there are real returns to test time compute. You know, if you let the model reason more, if you try more times on a hard problem, you can get much better answers already. And a business that just said, "I'm going to throw a thousand times more compute at every problem would get some amazing results." Now, you're not literally going to do that, and you don't have a 1000x compute. But the fact that that's now possible, I think, does point to an interesting thing people could do today, which is say, "Okay, I'm going to like really treat this as a power law and be willing to try a lot more compute for my hardest problems or most valuable things."
Uh, Streetar, do you just do the same thing with Snowflake and whatever your hardest problem is, or you've built this amazing career of, you know, data infrastructure, search optimization, running Snowflake, is that, is that just ask the question?
I think that would be a pretty cool way to use a lot of compute, but just to give an answer that's different from the world of tech that we live in, you know, there's a project called the Arnome project, it's like the DNA, uh, sequencing project that we did 20 odd years ago, but it's about figuring out RNA expression. Turns out they control pretty much how proteins work in our body. And a breakthrough there, knowing exactly how RNA controls DNA expression, it's likely to solve a ton of diseases and put humanity forward so much more. That would be a cool use of basically the equivalent of the DNA project done with, with language models. That would be a pretty cool outcome if you have a lot of compute to throw at something.
Inspiring. And one of our, you know, humanity's biggest problems.
Thank you so much, sweetheart.
Thank you.
Thank you, Sarah.
Thank you.
Thank you.