OpenAI CPO Kevin Weil on AI solving open math problems and launching Prism to accelerate science

the AI software engineer. Crush your backlog with your personal AI engineering team. And I'm also going to tell you about Vanta. Automate compliance and security. Vanta is the leading AI trust management platform. And without further ado, we have Kevin Wheel from OpenAI, the chief product officer in the ream waiting room.

What's up, guys?

How you doing? Good to see you.

I'm good. It's good to see you again. Thanks for having me on.

It's been a minute.

Yeah. What's what what's new with you since we've had you on the show. You've changed duties, you've changed focus, like uh reintroduce yourself, I guess.

Yeah. By the way, I love that we are talking about science today. I think we need to bring more of the this kind of coverage to science and get people excited about science because, you know, there are a few more important factors in all of our lives and AI is going to is in the process of fundamentally changing science. So,

yeah.

Uh

yeah, what what is your focus? Because science and health can mean a bunch of things. It can mean uh DTOC, you know, advice. And you know, one of the ways to cure cancer is just to get a lot of people to stop smoking cigarettes. That's something they could learn by GBT, should I be smoking? And it makes a very convincing case for why you should not uh and maybe it helps you not smoke. Uh at the same time, there's researchers that are going to spend decades developing cancer drugs, uh getting them through FDA approvals. There's a whole bunch of different ways that you know even Microsoft Word and PowerPoint and Excel helped the previous generation of biotech companies. You can imagine LLM's vending in there and then you can imagine the fully automated AI scientist. What's been your focus? Where do you think uh where do you think we're going to see the most progress in 2026?

Yeah, the the mission of our team is to accelerate science and the reason that we are starting to do this now. I mean this obviously been something that open cares about a lot but when you go back a couple years we were all amazed that GPT you know 3.54 whatever could do well on the SAT

and then you start seeing it solve graduate level problems and then it does okay on math competitions and then you fast forward a bit and our model can win a gold medal at the IMO the top you know math competition in the world

and now we're seeing it solve

open problems s in mathematics. So things that humans have not solved, suddenly AI models are contributing solutions to

and it's not everything. It's not every open problem. Science is not is far from done, right? But uh but the fact that that AI models can now contribute at the frontier of science and actually push beyond where humans have gone before

is incredibly exciting because if we can if you can replicate that across math, physics, biology, chemistry, material science and you can do the next say 30 years of science in 5 years instead. If we can be doing you know the science of the 2050s in 2030,

the world is a better place.

Yeah. and the models are there now and so this is a huge focus for us.

So yeah, talk a little bit more about your work specifically because uh not not to like discredit you but I feel like this part of the magic of these large AI models is that you get a lot of this stuff for free. Like oftent times you don't need a specific SAT model. You just build a amazing model and it learns how to code and also learns how to write poetry and tell jokes and it can do some science too. So what how are you working on fine-tuning? Is there a reinforcement learning project? Is there specific compute allocated towards this? You have a team. Are you listening to certain customers?

Yeah, I I imagine my my guest was just working really closely with customers, understanding, making sure they're not just signing up for chat GBT off the shelf and using it, but actually, you know, building a bunch of specific functionality for them and innovating on the products.

Yeah. Yeah. How do you think about that?

Yeah. Well, it's it's not uh it's not about building offshoot models. This is still this is still about improving our core models but there is if you think about science science has a huge surface area right

um so you want there's a lot that we can teach models at the frontier that um you know there's there's work that we can do to make to make our models even better at performing frontier science

um and that crosses from pre-training into post- trainining and reinforcement learning

um and so we get to solve you frontier problems in science to make our models better. There's also a whole lot of work that we can do on the tooling front. You know, scientists use a lot of different tools

and you want to bring those to bear so that the model is is using the same kind of advanced scientific tooling that scientists are because then it can be an even greater force multiplier for the work they do. So there's there's a whole bunch that we're doing on the pure model front. Yeah,

that's not the only thing though because I think if there's anything that we've learned from the way that AI has completely revolutionized coding over the last year is it's two things. It's great models and it's also integrating AI into the environment where engineers operate, right? You don't

you don't spend your time going to chat GPT and copying and pasting back and forth. You use codecs in your IDE.

Okay.

Uh and bringing the model into the environment where you're working is a huge part. That's why we launched Prism yesterday. Um the idea is uh or I guess on Tuesday. But the idea is to bring great models that can that can help with AI into the workflows that scientists are using every day. In this case, Prism is about scientific writing and collaboration.

And if you can help people communicate their ideas faster, if they can use AI to express the research that they've done and collaborate with other scientists, that's its own form of acceleration. So we both want to make models smarter and we want to bring AI into the environments that scientists are operating in day-to-day. Both of those are parts of accelerating science.

So talk about some of the integrations. Uh chat GBT at one point just got access to a Python ripple and could sort of write code and execute it in the uh in the in the query for just a GPT5 pro query. Uh but G chat GPT also has an integration to Gmail and there's a like some sort of business development relationship. Uh what is your as you're bringing more tools to bear in the scientific uh workflow? Uh is there some sort of balancing act? Do you just need to integrate with everything? How do you think about actually wiring up and like unhobling the models? Yeah, I think it's about both uh a handful a set of tools that we think are going to be broadly useful that we can integrate ourselves.

As an example, if you're in the process of trying to solve a hard scientific problem, you need to solve a differential equation.

Yeah,

the models are actually smart enough these days that they can solve a differential equation just by reasoning through it. But you also you also can uh integrate a computer algebra system. A system that will deterministically and very quickly solve differential equations.

Why not let the model use that?

Exactly.

And you know go back to doing what it does best which is reasoning broadly about how to solve hard problems. So you have that you have like protein databases and biology. You have so many tools like this.

Yeah. And I think it'll be important because I I might be a scientist studying, you know, the evolutionary biology of snails and I might have my own set of tools that I use or maybe small models that I've trained. They're good at doing very specific things.

So, it'll be both about us teaching the models to use certain tools that'll be broadly helpful and enabling scientists to bring their own tools to bear so that the models can very deafly adopt them uh into the process. Mhm.

What does a chatbt moment look like for Prism within the scientific community?

Ooh, that's a good I that's a good question. Um, you know, I I I think it's going to be a I I think it's going to be a more like process of incremental compounding where you've got a a product that helps people work and operate faster. you've got models that are going to be increasingly useful and we're going to see kind of two um you know exponentials if you will. One is just the exponential that the models are on and their ability to help any scientist who's adopted them do what they do faster right to aid in their thinking.

I was just talking to a scientist earlier today who called uh uh GPT 5.2 a metal detector for hypothesis.

Right? So he's thinking about he's got so many different ideas in his head. You can only run so many experiments. You only have so much time. And he uses GPT5 as a thought partner in helping him hone in on the most valuable ways to test his ideas. So you've got that exponential and then there's a separate one that comes from scientists beginning to adopt these tools because a lot of scientists still haven't adopted them. And the more that they do, the more that they individually move faster and the more that the entire field of science accelerates. So I think there's a lot to be excited about here. Uh a couple weeks ago we talked to Andrew from Cerebras and I was asking him about where uh wafer scale computing is most exciting to him and he actually cited science as a particularly valuable place and I was wondering if you had thoughts on the value of speed or the value of different uh chip architectures in the scientific workload. I was kind of come to coming to it like, well, you know, developing drugs takes a long time. It's probably fine if the model goes off and cooks for an hour, but he was sort of pushing back on that. And I'm wondering how you think about the different parameters. Obviously, we're on an exponential with intelligence and capabilities, but there's also latency and and uh usability and flavor, and there's so many other knobs that are being turned as the models progress and as the projects progress.

Yeah. I think one of the interesting things about scientific problems in particular is that when you're solving the hardest frontier science problems, you need to do a lot of thinking, right? If these problems were easy, really smart humans would have solved them long ago. And so the kinds of problems that are left often involve the model thinking not for you know 5 minutes or 20 minutes which would be a long uh a long time if you're inside chat GPT

but maybe an hour 2 hours 12 hours 2 days.

Yeah.

Uh and you know that is where we're going.

Yeah. And if you have really fast inference that can take that two-day roll out and turn it into a six-h hour roll out or an hour thing only takes 10 minutes then again you that that just it's more opportunity for you as a scientist to

maybe instead of testing two hypotheses you're testing 20 in the same period of time. Again it's acceleration.

Yeah. uh to to go back to the chat GPT moment. Um how are you thinking about actually seeding uh your work into the scientific community because there's there's one angle where it's just it's just chatgptt.com. It sort of goes viral. Everyone's playing with it. That's part of what the chatgpt moment was was just anyone with a web browser could use it. Uh at the same time, there's only so many real scientists. A lot of them are in labs or in academic institutions. And I could imagine you uh doing partnerships or deals or anything to actually get something deployed into the most elite scientific environments. Have you thought about the different trade-offs there?

Well, one of the cool things is we're seeing so much organic adoption. Yeah.

You go on Twitter these days as you do and you're like every day I feel like I'm seeing new examples where someone will say, you know what, I just solved this problem with GPT5.

Yeah. or I gave this problem to my grad student and they were busy. It took them too long. I I just, you know, I wanted to make worker progress. I just gave it to GBT5 and now I have a solution.

Uh and so there's there's this incredible organic adoption because of course when other people see that, other scientists see that they go, "Oh, wait a minute, you know, and maybe the last time they tried the models was a year ago when they weren't at a place that they were going to really meaningfully contribute to scientific research. They could help with other things, but they weren't going to help with your hardest problems. Yeah,

now they can. And so there's this ground swell of scientists that are adopting uh GPT 5.2 especially, I think, has been uh kind of an inflection point.

So it's it's exciting to see and there's a lot happening, you know,

even without us with people just discovering this and talking about it.

Yeah, a lot of people experienced that with like Andre Carpathy's journey with vibe coding and they a lot of people were like, "Okay, yeah, like if it's good enough for him, I got to jump back in." It's funny. man is an incredible communicator.

Yeah. Yeah. Really?

What uh what's the update on how like forwardthinking labs are are thinking about integrating with a product like Prism? Ideally, there's a future where, you know, a scientist could be at home, have an idea, may maybe they're working in, you know, pharma or something in biology and they can just like start running. You can imagine somebody starting to run an experiment just based on a prompt and then somewhere in the physical world uh there's actual you know the biological process is actually happening. Um how how much progress is there on that front?

Yeah, I'm super excited about uh the world of robotic labs. I think it is 100% likely to be the the future that um that that we're moving towards because you can do so much more in parallel. again to the idea of accelerating science and moving faster the world where you can have a hypothesis maybe that you've honed with you know back and forth with chat GPT in in this case it may also be running simulations you know take if you're doing something like uh fusion where you want to do heavy simulation before you run an experiment because your experiments are expensive then you have the model thinking running a fusion simulation looking at the results of that refining its thinking running another fusion simulation and you do as much with the compute that you have in advance so that when you do something in the real world it's like that much more likely to be successful. You can look at the same thing with respect to biology. There's no reason at this point that you need to have grad students, you know, uh pipetting one thing into another. the the idea that I think like a scientist in a lab coat could easily fade away where they're just like in a normal office, you know,

or or at the very least doing the kinds of things that are incredibly hard for models for robots to replicate.

Yeah. Um but there's a lot of science that can be totally automated and the idea of robotic labs that are um you know 24/7 online that you can scale in parallel as far as as you know you can make it efficient and you have models thinking you know reasoning for two days to find the most uh efficient experiments to run maybe running simulations to test that and then once they get to a good point passing that to a robotic lab which can experiment in parallel at high volume. the results pass back into a model which reasons about the results and then goes out and runs a different set of experiments. You know, you're doing reinforcement learning with a loop through the real world

and that is absolutely

are robotic labs something that that OpenAI would ever do or is that best suited for an external partner? How do you think about that that point of the stack?

I think it'll be both. We want to partner broadly uh with with scientists that have their own labs that are already doing anything. Like I said, science is an extremely broad uh has huge amounts of surface area. There's no way that we can do even a tiny fraction of all of science.

And I think there will be a lot of opportunity for us to learn from things like robotic labs to make sure that we're working really well with those scientists. So I think it will be both.

Can you clarify the business model? There was some confusion about how this might be monetized.

Poor Sarah Frier. She keeps

I'm pulling for ads in science. I want a pill that I take. It helps me lose weight, but then I wind up loving Ford F-150s. I'm willing to take [laughter] that trade. Adup supported medicines. That's what we need. But no. Uh, what are you actually thinking? Is anything changing?

This is why I love you guys. [laughter]

We're your strongest defenders. Whatever you do.

Um, yeah. Sarah said something when when Sarah was talking about, oh, maybe maybe there are ways to like monetize IP. She was speaking specifically to the idea of us doing partnerships with large companies, you know, pharma companies, people like that

where there would be a specific partnership that was developed with the idea of sharing royalties in the future. It was not meant at all about normal signing up and this already happens with the OpenAI investment fund. There's startups that take that take capital from open AI, they build something on top of chat GBT or GPT APIs and like I I think most people inside sort of understood that but thanks for clarifying. Um so in general most people will just be on sort of some sort of like API consumption based pricing or subscription fee like mostly the normal.

The nice thing about Prism is you log in with your chat GPT account. So it just you just bring your existing chat GPT account along with you. are this is not you know there are not a billion scientists in the world this is not an effort to like build a brand new business model this is about accelerating science because I think it's one of the most impactful and missionoriented things we could possibly do

yeah that makes a ton of sense give us the update on detachment 2011 how's it going

are you sold it's been a blast been working out a lot

it's been an absolute blast uh I've been in DC to various bases things like that uh it I I think it is massively important that we bring Silicon Valley and uh and DC closer together because we have an incredible tech community. Uh we have a country and a set of values that are worth defending and you know it's a dangerous world out there.

The more that we can do to um you know to strengthen who we are and what we do, the better.

That's great. Well, thank you so much for taking the time to come talk to us. Very fascinating stuff.

Yeah. Thanks so much for having me on, guys. Good to see you again.

Good to see you, too. We'll talk to you soon, Kevin. Goodbye. Let me tell you about Railway. Railway

founder

is the all-in-one intelligent cloud provider. Use your favorite agent to deploy web app servers, databases, and more while Railway automatically takes care of scaling, monitoring, and security. And up next, the New York