How AI is accelerating scientific discovery today and what's ahead — the OpenAI Podcast Ep. 10

Maybe the most profound way that people are going to feel AGI in their lives is through science.

The acceleration that is going to come from these tools is going to change science.

Yeah, the mission of OpenAI for Science is to accelerate science. So the question is, can we help scientists do the next, say, 25 years of scientific research and scientific discovery in five years instead. Science underpins so much of what we do and how we live. And if we can make progress go faster by putting our most advanced models into the hands of the best scientists in the world, we should do that. And that's what we're trying to do. You could ask like why now? Why didn't we do this a year ago? Why aren't we doing this a year from now? One of the big reasons is we're just starting to see our frontier AI models being able to do novel science. So we're starting to see examples where GPT-5 can actually prove new things. Maybe not yet things that humans could not do, but things that humans have not done. So these are like these little existence proofs of GPT-5 being able to break out past the frontier of human knowledge and into the unknown. And if there's one thing that I've learned from now, you know, a year and a half or so at OpenAI, it's that you go very quickly from the model can't do something, to the model can just barely do something, and it's not great at it yet, but you see these early examples. And then, you know, six months later, twelve months later, all of a sudden you couldn't imagine doing this thing without AI. And I think science is in that initial phase where we're seeing real acceleration for scientists that are using AI. Sometimes novel, you know, not yet maybe large breakthroughs, call them small breakthroughs. And that just says that there's so much potential in this space.

Yeah, I mean we're seeing examples every day and they're across the range of sort of the scientific frontier. You see examples in mathematics, in physics, astronomy, life sciences, like biology. Alex, I mean, you've worked on some of these. Maybe it's a good time to talk about some of the physics stuff that you've seen.

Yeah. I think one of the cool things, so you've got examples like Alex's where it was probably not something that he like he could have done it himself over you know eventually, but GPT was able to do it faster. That's acceleration on its own. And there's something qualitative about that even as well, because if you can explore, instead of exploring two paths over the course of a week, if you can explore 10 paths in parallel in, you know, an hour, all of a sudden there's a lot more ideas that you can try. And that's also acceleration. We also see examples in like literature search, which you don't think of as maybe like deep scientific innovation, but it's really important to be able to understand, you know, has somebody worked on this problem before? And if so, is there something I can learn to speed up my own work? So, and we've seen interesting examples where there was one, I might get the details of this wrong, but we were talking to this researcher, and he was saying he was exploring this particular idea in like high-dimensional optimization. And he was like, man, you know, this thing I'm working on, it's interesting, but somebody must have worked on this before. I can't be the first person to have had this idea. I just can't. But I can't find any examples. And then he had given it, he'd sort of given a description of what he was working on to GPT-5. And GPT-5 found an example from, I think it was like economics or something, a completely different field that used completely different terminology, so no keyword lookup would have ever worked. GPT-5 did sort of a conceptual level literature search. Found somebody's PhD thesis in German. So also a completely different language, you know, it was like basically lost to time, but this person had done really interesting, sort of related work that helped him in his research. And so, you know, that's another area. So you can talk about the acceleration that comes from just like novel proofs and GPT-5 being able to do something on its own or guided by an expert. But there's also these examples of acceleration in calculations and literature search and all of them contribute to accelerating science.

Yeah, that's another thing that we've heard from professors and researchers that we've talked to is there's so much you have to be so specialized today. And so sometimes it gets hard to explore an area outside of your main area. There's one particular mathematician we were talking to said, you know, one of my last papers, I knew there was an area that I wanted to go follow it off in this direction. But it wasn't my specialty and it would have taken me a long time and I just kind of ended up feeling like, you know, maybe that's not the most efficient place for me to spend my time. Now with GPT-5, I'm gonna go back and explore that because I've got a coworker effectively, a collaborator who has read just about every scientific paper that's out there and is a pretty meaningful expert on just about any topic you want. And I think I'm going to be able to go explore these adjacencies in a far better way with ChatGPT than I could have on my own. And so that's also a fascinating new take, right? It can help you go deeper, like you were saying, and it can also help you go more broad.

Yeah. And I think especially some of the most interesting research now happens at the intersections of two fields. And again, it's hard for one person to be an expert in two fields, let alone three or four or five. And sometimes it's tough for humans to collaborate. You don't necessarily find the right person, the person doesn't have infinite patience. And here with GPT, you have now the option to have a collaborator that will work 24-7, has infinite patience, you know, has read substantially every scientific paper written in the last however many years. And so it's just a new kind of collaboration that is its own form of acceleration.

Yeah. So actually I don't even know if I've told you this, my own sort of origin story with appreciating what GPT-5 could do, or in this case it was, I think, oh this was almost a year ago, so it was o1 preview maybe. But I was meeting with this guy named Brian Spears, who's a physicist at Lawrence Livermore. It was in DC and we'd never met before. So I didn't know sort of what to expect. I thought maybe I was gonna go in and be talking to him about what was new and what he could do with o1 Preview and why he should give it a try. Little did I know. I sat down and he immediately took control of the conversation and said, let me tell you what I can do with your models and like these are the most amazing things for science and this is gonna change the world and he was like, okay, let me take you through this. And he opened up his laptop. And you know, he works on fusion, right? Lawrence Livermore was the first to do large-scale fusion with positive energy, like super exciting. So he's like, all right, we're gonna take a fusion example. And first I'm gonna start with the undergrad version of this problem. And so he shows me this conversation and he's like, all right, so you've got you know a copper rod and we're gonna bombard it with super high pressure waves. What happens? And you know he's like it. So he answered and o1 preview gives a good answer. It's like okay, cool. So it got the undergrad problem right. And then now let's ask the graduate version of this. Now what happens inside the rod itself as you're doing this? And you know what needs to be true in order for it to generate these certain kinds of shock waves? And he goes through and he's like, okay, so got that right. All right, now let's ask the postdoc level question. All right. Now let's ask the and at this point I'm like, you know, despite having a physics background, I'm just following along for the ride because he's beyond anything I can do. Like, all right. Now let's ask the you just joined Lawrence Livermore and you, you know, kind of question. You've gone through your postdoc, you're a nuclear physicist, and he keeps going. And o1 preview keeps getting the answer right. And then he's like, all right, now let me ask you the you've worked at Lawrence Livermore for twenty years question. And it goes and it gets it right. And then not only that, but it suggests that the only way to go forward is to use these set of simulation tools that are like partially classified or that only Lawrence Livermore has. It's like, you know, I don't have access to these, but if you did, you would want to use these tools. And he's like, look, nothing in here that nothing that I just showed you was something that I couldn't do. But it would have taken me days. And certainly not everybody at the lab can do this. Like the acceleration that comes, that is going to come from these tools, is going to change science. And so I went from like sitting down with this guy who I thought maybe I was going to be sort of talking to him about the value of AI, to him just completely blowing my mind about the potential of AI. And this is a year ago. This is o1 preview, you know. We've come leaps and bounds since then. And the thing that I always try and remind everybody, the AI models that we're using today, as good as GPT-5.1 Pro is. These are the worst AI models that we will ever use for the rest of our lives. And when you think about that, the fact that we're here just implies that the future is very bright.

I was gonna I actually think you made a really important point in the middle of that around the fact that you gave it the hard question. It didn't get it right. You gave it an easier question, it got that right, and then you were able to give it a harder question, it got there is still, you know, as excited as we clearly are about the future here, there's also a very real sense, like when you're giving GPT-5 or any of these AI models a problem that's on the frontier, that's at the limit of their capabilities, they tend to still be wrong a lot. Right. Kind of like any human would be at operating at the level of at the frontier of their capabilities. And it takes you know, it isn't just automatic yet. Hopefully in the future it will be, you know, enter in any hard question and the model answers it. But today there's a lot of back and forth. And the people that are best, the researchers that are best at getting the most out of the models have a sort of patience to go back and forth with them. I think that's natural. It's probably the way that you would work with any two people operating at about the limit of their capabilities. But I think it's important especially for folks listening to this who are doing research with the models to know that it isn't just one shot and it always works. There really is a back and forth and sort of a patience that it takes. And one of the interesting research problems that we're spending a lot of time thinking about is how we help people with, yeah, how we sort of help reduce that cognitive load. Because when you're working on a problem, say the model has a 5% pass rate on some problem. So technically the model can get it right once out of 20 times. But it's really at the frontier, so it's not going to get it right nearly, you know, even close to every time. If you're sitting inside ChatGPT and just entering in this question, you're gonna have to enter it in, you know, what, ten times before you have the odds that it's gonna get the right answer? And that's most people aren't gonna do that and so there's a whole host of problems that the model can solve that people probably try and are like, oh after three tries, it didn't get it right, so let's I'll move on. The model's not good enough yet. And actually it is, but it's just very hard to tell apart low pass rate problems from problems that are too hard. And I think that's actually a really important thing for us to help researchers and mathematicians get past. Because the most interesting problems right now are going to be the ones where the model has a very low but non-zero pass rate. Those are going to be the hardest problems that the model can solve, the best ways that it can help accelerate science. And so that's a really interesting research problem that we're taking on to try and make that a little more automatic, a little less grunt work. But for now, I like putting in the time and really going back and forth with the model does yield results.

I think it's probably just a characteristic of any question that's on the frontier of or sort of at the limit of what the models can do. And back with GPT 3 and early versions of 4, the questions that were at the limit of what the model can do were much more basic. Now they're questions of you know scientific research, but you still, when you're operating at the frontier, the pass rate will be low. And so you gotta kinda like there's value in sticking with it and trying a few different things and taking the parts that it gets right and refining them while telling the model where it got other things wrong.

Yeah, and human and AI together are exactly much more powerful than human alone or AI alone.

Yeah, so we've talked a bunch about these anecdotal examples that Alex has gotten from the time that he spent with his colleagues that we see coming in across Twitter, you know, on a semi-daily basis at this point. And we wanted to sort of bring them together and just write something, publish something about that that lays out the current sort of state of GPT-5 with respect to science. And so what we've got, it's a handful of collaborators from inside OpenAI and I think eight or nine academics from beyond our walls. Across a bunch of different fields, math, physics, astronomy, computer science, biology, materials science. And the paper is something on the order of twelve sections, each one highlighting a different way that GPT-5 is accelerating their work. The goal was not to be, you know, hypey and say everything is solved. It's, you know, it's really to say, look. Like this is what works. This is what doesn't work. Here's what I tried. In many cases, we're sharing the full share links, the conversation, so you can see the back and forth that the scientist has with the model. And it's meant to be kind of a moment in time to say, this is where we are today. And I think we'll look back in six months, twelve months and we'll probably be much further and that'll be exciting. But even where we are today, we've got a section in the paper on a bunch of different examples around literature search. A section of the paper with a bunch of different examples around acceleration, whether it's calculations and other things like that. And then a section where we actually contribute four or five new non-trivial results in mathematics. And a couple of these are small, a couple of them probably could have been papers on their own. And so you go from kind of the mundane but very pragmatic and real bits of acceleration to the more sort of profound GPT-5 actually pushing past the current frontier of human knowledge. And so we're super excited about this paper. It's, you know, I think there'll be a lot more to come. We're not the only lab doing great work, by the way. Google has been doing this for a while, and I have a ton of respect for what Demis and the team have done with AlphaFold and more. I just think we're at a really exciting time. You know, ideas in science often have their moment when you have multiple people coming with the same idea, whether it's quantum mechanics like Alex was talking about or the light bulb. Right now it's very clear that AI is just beginning to change science. And it's going to be an exciting few years.

Oh man. The five year question is so hard. Here's a crystal ball. Yeah, you know, I think I mean the exciting thing about this field in general is you look back twelve months and you're completely embarrassed by where you were twelve months ago. You know, the idea if I When GPT-3 launched, it was unbelievable. Right. I mean, it I'll speak for myself. It blew my mind the idea that AI could do any of these things. And then somewhere in around like GPT 3.5 and 4, the Turing test, which we had held up for like what, 75 years as the pinnacle of artificial intelligence research. Like oh man the world will be different when an AI can pass the Turing test. We just went whooshing by and like now we just don't talk about the Turing test anymore. And even you look back to the beginning of this year of 2025. And most people were writing code themselves. Most engineers were writing all of their own code. And the idea that you're writing it yourself. And now fast forward, and you've got like the idea that you would do really much of anything without leveraging Codex, Claude Code, GitHub Copilot, you know, any of these tools. They're all incredible. Is crazy, right? You're so much more productive with it. So just in 12 months, and I don't in 12 months software engineering has fundamentally changed. I think over the next 12 months, we're going to see profound changes in the way that science is done, you know, both in the stuff that we can do in silico, in theoretical physics and mathematics, and computer science. And I think we're going to begin to see it in the life sciences, in the physical sciences. That's over the next twelve months. I mean five years?

Well, I think one of the valuable there's so many areas where models can help with life sciences. If you take, you know, biology drug discovery, for example, you have a huge search space. And the more that the models can learn how to prune that search space, the more, even if you're going to end up with a bunch of physical real-world experiments to run at the end of the day, if you can intelligently prune the search space, then you can more rapidly converge on the drugs that are likely to work in particular scenarios. And then you can think about the impact, you know, for that to have real world impact, you need to make it all the way through the regulatory process. That is its own process that AI can help speed up because you end up needing to write these huge papers that bring together you know tons of different findings and so on. So you can take each step of the process. AI can help upfront as you prune the search space and try and find candidates that are more likely to meet your needs and meet the goals that you have. And then as you go through the process to getting this thing out to consumers and making a real world impact, AI can contribute there. And we have pilots with a number of the companies in the space doing that. So it really is fairly broad-based.

I mean, it is an absolute privilege for me to get to come back and work on science. And, you know, I am nowhere near the scientist that folks like Alex and other people here at OpenAI are, but I don't know of something you know I think we talk a lot about AGI at OpenAI, artificial general intelligence. I think maybe the most profound way that people are going to feel AGI in their lives is through science. Yeah, ChatGPT is an incredible tool. I use it tons of times every single day. But AGI inside ChatGPT will be able to do lots of things. But when I can have, you know, personalized medicine if AI models can contribute to science, you know, finding a way to do scalable fusion more quickly. Those kinds of things will change all of our lives. And I think these are very real possibilities at the pace that we're going. So that's why this is the most exciting thing in the world to me to get to work on.

Yeah. Or they're using the free version because you know, of course, that's how everyone starts. And the free version doesn't think for as long, and so it can't solve problems that are as challenging. Yeah, I think that's really real. It's one of the reasons that I think the best advice is to just like keep trying the problems. Even if you're working on problems and as you try them on GPT-5, it isn't super helpful. I wouldn't give up. I would keep trying it every few months. And I think at some point, you know, it's gonna start being valuable if it's not already there today. We talked about sort of thinking time. Yeah, that's another area that we're really excited to see that with GPT-5 Pro. You can get the model. I've seen it think for what, maybe 40 minutes on some of the hardest problems. But you know, it has a certain amount of sort of compute allowance because we have to serve it to many, many, many people. Forty minutes is certainly not a limit on thinking. Like the models can think for two hours, six hours, twelve hours, twenty-four hours. And one thing we continue to see is that pass rate on hard problems continues to improve as you give the models more time to think. Which is like, you know, it's surprising actually the number of times there's a totally reasonable human intuitive human analogy to these things. There are a lot of problems that I can't solve in 20 minutes, but that I might be able to solve if you gave me two hours.

Yeah, and some that I can't solve in two hours, but if I had a day to really think about it and try different things, I might get there. And the models are the same way. So being able to give a much small, you know, there aren't as many scientists in the world as there are users of ChatGPT, if we could find ways to give scientists that really know how to use the models well, just a huge amount of compute. I think that is yet another way that we can accelerate science.

Yeah, a hundred percent. I think if model progress stopped today, just the process of driving awareness within the scientific community and giving people more of the best that the models can deliver, I think we would see a large amount of scientific acceleration, but of course progress is not going to stop, as Alex was saying. And so when you think about the models being able to think for a longer time, being able to train them to do harder and harder scientific tasks. And actually also just, you know, getting out in the scientific community and helping people see what the frontier really is and how they can use the models better to do the work that they're doing. I just like I'm excited to see where this goes over the course of the next six months, twelve months, twenty-four months.

Yeah. Like with many things, there's sort of an intuitive way to understand this is the models get smarter, benchmarks are just a way of testing the model in some sense. And as the models get smarter, you need to give them harder and harder tests because they learn how to ace the earlier tests. So if you take GPQA, which stands for Google Proof QA, it's a scientific benchmark that asks basically PhD level questions across a range of scientific fields. We thought for a long time that was a very hard benchmark to beat. I think it came out in 2023 and GPT-4 originally was like at 39% on this benchmark. Humans, by the way, are at about 70%. But now you fast forward two years and our latest models are nearly at 90%. So they're surpassing the capability of most humans in their field of scientific study. Across every field at once, which is kind of amazing when you think about it. But that isn't, you know, those aren't the hardest questions in the world. And that's one of the reasons that we're focused on new evaluations that ask frontier science and mathematics questions. It's also, you know, we released something called GDP Val recently, which is an eval that tests the model's ability to do economically valuable tasks. So the smarter the models get, the harder the tests that we want to keep giving them. Because, you know, every gap that we see, every place where the model can't answer a certain question. That's feedback for us and gives us a way to improve the model further.

You want to go first?

Yeah and I'll say fusion. Just because if we can actually we have again small scale or I mean large scale but small existence proofs of it. So clearly it can work and the challenge now is to do it at bigger scale, more reliably. Clearly it's possible. We will figure this out, but if we can accelerate it, then you know the world with fusion is a significantly better place than the world without. We solve a lot of problems if we solve fusion. And you know, I'm excited to see if maybe we can contribute in some way.

Yeah. It's incredible. I mean some groups might need to like might be looking to build lots of infrastructure for lots of GPUs, for example.

Yeah. And I think anytime you change something by an order of magnitude, the world changes. I think what we've seen over the past year with the way that software engineering has changed, you now don't need to be trained as a software engineer to write, you know, meaningful amounts of code. That means you can bring, you know, there are like what 30 million software engineers in the world. I think now 300 million, maybe 3 billion people can write software. And that's gonna fundamentally change things. If we can move, you know, if we can make energy ten times more prevalent, ten times cheaper. It will change the world and I think it's a really high potential place for us to apply the intelligence of our models.

Yeah, it's such an important point. The frontier or the surface area of science is massive. And this is not about what we can do within OpenAI individually to accelerate science or to accelerate specific scientific projects. It's about giving scientists all around the world AI, so that they can accelerate their work. That's how we move science forward faster. So, you know, there are pieces I think that we will try and do because it'll help us learn. But the vast majority, like what we really want is to see a hundred scientists win Nobel Prizes using AI.

Exactly.

Certainly it's a sort of a there's a science 2.0 moment happening, I think.