VP, OpenAI for Science, OpenAI
Enterprise & AI | Kevin Weil, VP, OpenAI for Science & Jeetu Patel
Kevin Weil, VP at @OpenAI for Science discusses how AI is redefining how discovery happens, and who gets there first.
Related links:
Explore the full agenda and watch all sessions → https://www.ciscoaisummit.com/ai-virt...
Discover how Cisco is transforming industries with cutting-edge technology and business-driven innovation.
Learn more:
Cisco Official Site → https://www.cisco.com/?dtid=osclytb00...
Cisco Products & Services → https://www.cisco.com/site/us/en/prod...
Cisco Solutions → https://www.cisco.com/site/us/en/solu...
Contact Us → https://www.cisco.com/site/us/en/abou......
About Kevin Weil
Kevin Weil, then-VP of OpenAI for Science, discussed the company's efforts to hire practicing mathematicians, physicists, and biologists to develop frontier AI models. He described the models as "incredible," noting that they have progressed from achieving a 700 on the math SAT three years ago to "regularly solving open problems in math and physics and other scientific fields." Weil stated that OpenAI's mission is "not to win a Nobel Prize ourselves" but "to see a 100 scientists win a 100 Nobel prizes using our technology." He characterized AI as a "metal detector for hypothesis," saying it has read "substantially every paper across every field of science" and can generate more ideas than scientists can experiment with. Weil discussed OpenAI's role in the Department of Energy's Genesis Mission, calling it "one of the most exciting projects happening right now." He said the mission has a "huge amount of scientific data that is currently mostly unused" and that teaching AI models that science could "confer advantage on the US." He expressed particular interest in fusion energy, where AI could "iterate far faster on parameters using simulations and real experiments." Weil also addressed barriers to adoption, including the cost of compute, stating that scientists who could "do the most amazing things" with it often have "the least ability to pay." He acknowledged concerns about AI-generated "slop" in scientific publishing, comparing it to email spam and suggesting AI would ultimately be used to filter it out, while noting that "peer review will still be a thing."
Transcript (40 segments)
I
Interviewer0:00
It's actually impressive the progress that I think all of you folks are making. And it's almost impossible to predict who's going to come out ahead. And so for customers, in fact, that does tend to be a little bit of a confusing thing because one week one model's doing better, the other week someone else does better. So let's start with that. I want to talk a little bit about your product role at OpenAI for just a little bit, but then I want to really shift into what's happening in science with OpenAI. Starting on the product side, you know, you've been a CPO of iconic companies, Twitter, Instagram, Planet, and then what you're doing, what you did at OpenAI, contrast what you saw at OpenAI compared to all the other massive hyper cycles that you saw.
K
Kevin Weil0:52
I think the biggest change is undoubtedly just how fast the world is moving right now. I've never seen anything like it. Every month computers can do something that computers have never been able to do in the history of the world before. And that wasn't true when I was at Twitter in 2009, Instagram in 2015, 2016. When you're building a normal product, you're building on technology that we all know, databases, etc. Databases get a little bit better every year, but they're basically the same thing from one year to the next. The tools that you use, you kind of know and you're building on top of them. Now it's like the underlying technology changes and suddenly you have superpowers next quarter that you couldn't have imagined that you had today and it totally changes how you work. I mean, on the one hand everything is going really fast which is exciting. I feel like if you're not spending your time writing code, if you're not very quickly translating an idea you have or you hear about a bug, if you're not turning that into a Codex task and just flipping around and fixing it yourself, why not? And that's super empowering. It also, I think, leads us to be much more bottoms up in the way that we build that innovation. Because you just can't sit there from on high and say, 'Okay, here's our 12-month road map. Now go do that.' Because you don't know what the world is going to look like in 3 months. And so of course it's still important to have strategy and to have direction and for people to understand what you're trying to accomplish and why. But then the world today I think is much more built for people to just be high agency, take matters into their own hands, understand the direction you're trying to go, but then you push responsibility down and your teams move really fast.
I
Interviewer2:45
Well, the one thing that this audience would probably like and the people that are watching online that are of the same ilk is I think all of us in the community that are building technology for AI, I think they would like us to not just move really fast but figure out a way to make sure that the absorption rate is higher on the companies. Any ideas on that? What can we do to go out and accelerate absorption rate? Because right now the absorption rate is still pretty slow.
K
Kevin Weil3:12
Yeah, I think the number one thing is just to get everybody experimenting with it. You can't stand apart from AI and expect to keep up because the iteration cycle is so fast. You've got to just get in there and ride the roller coaster and do the things. Don't wait for it to be perfect. Do the things that barely work because one of the things I've learned over a couple years at OpenAI is you go very quickly from this thing was completely impossible for computers for as long as we've been alive to this thing just kind of works and it's a little bit frustrating. It's imperfect. The model makes mistakes, but it works five or 10% of the time to like three to six months later that thing now works 80% of the time and you would just never imagine not using AI for that thing again. But you're not going to keep up with that if you're not trying in the stages when it's working at 5 or 10%. And you have to get everyone thinking that way. I will say I went through a bit of a transformation relatively recently even with Codex, which is our coding agent, where I'd been using it before but sporadically, it hadn't changed how I work. And then we launched a product relatively recently called Prism and it's a small team and I was contributing code and I was starting to use Codex all the time. And I got to a point where I was sitting in a meeting, my boss's staff meeting, right? So it was like time to go in, close the laptops, pay attention. And I closed my laptop and I was like, 'Oh, shoot. I didn't get a Codex job running before I closed my laptop.' I just missed an hour of productivity. I could have been fixing something, building a new feature. I mean, Codex working for me to do that. Same thing. Before I go to bed, I'm like, 'Okay, what really hard thing can I give Codex to do while I sleep?' and keep working on. It completely changed the way that I think. But that sort of approach is you could build so much faster. It's like have 10 things going in parallel.
I
Interviewer5:23
The other thing you taught me, we were actually in the gym at like 10 o'clock at night, Kevin and I, after a board meeting in DC, and you said to me, he's like, you know, the first time that your engineers use this, they're not going to be great at it. So don't expect greatness in the first time, but you got to make sure that you get everyone to start using it. And since then, close to 80% of our engineers are using it now on a regular basis. And lo and behold, we now have the first product that's going to be 100% written with Codex. And that's amazing.
K
Kevin Weil5:54
Yeah, that was very cool. Yeah, you're totally right. It's a skill like anything else. And you learn how to channel like when the agent does something imperfect, you actually take that and take the lesson and put it in your agents.md file so the next time the AI knows not to make that same mistake again. So it is a totally different way of working but man when you embrace it, first of all it's incredible that you have a team that is entirely building off of Codex. My prediction is that very quickly you'll have two, three, four, five teams.
I
Interviewer6:26
The goal is half a dozen teams by the end of '26 at least.
K
Kevin Weil6:28
Yeah, I bet you'll beat that handily. And the transformation that you and Chuck are driving is just fantastic. But I also bet that the way that happens is other teams look at that team and go, 'Hey, why can't we move as fast as they are?'
I
Interviewer6:44
Totally. Yeah. Okay. So, let's talk about AI and science. What is so significant about this? What made you say, you know what, that's where I want to spend the next how many other years of my life doing and what have you learned?
K
Kevin Weil7:00
Yeah, I mean, so OpenAI, our mission is to build AGI and make it beneficial for all of humanity. I can't think of many ways that AGI will more positively benefit humanity than being able to accelerate science. So, our mission is to accelerate.
I
Interviewer7:16
You studied physics in undergrad? You're like the perfect kind of person to...
K
Kevin Weil7:19
I studied physics. I have most of a PhD and then I dropped out and started working at startups. Most of my colleagues have actual PhDs. But if we can truly accelerate science, if we can do the next 25 years of science in the next five years instead, and that means that we will be sitting there in 2030 with the technology and the science of 2050. How amazing is that? I mean think of how science shapes our lives. The devices that we use, the medicine that we take, it's everywhere. And so for the people that are not scientists, what meaningful thing can you work on?
I
Interviewer7:56
So for the people that are not scientists, give us real examples of where this could meaningfully change day-to-day lives.
K
Kevin Weil8:01
Yeah. So the analogy that I use is I think it's a good one for this audience too. So in 2025, AI completely changed software engineering, right? At the beginning of 2025, if you were using Codex to write most of your code, you were an early adopter. By the end of 2025, just 12 months later, if you were not using Codex or Claude Code or whatever to write most of your code, you were falling behind. Teams that were moving faster than you are. And that was an entire huge multi-trillion dollar industry that changed in 12 months. I think the same thing will happen with science in 2026. And we're already seeing examples of it in '26 itself. And I don't mean by the end of the year we're going to have solved science. Of course not. But that same thing where if you're a scientist and you're using AI to collaborate in your work heavily today, you're an early adopter. But it's starting to pay dividends. Like just in the last few months, actually really just in January, we've seen a huge number of open mathematical problems. Problems that the best mathematicians in the world have tried to solve for years and haven't. We've seen a number of those fall to AI-driven solutions mostly by GPT-5.2. So it's not just doing the thing that you think of with AI which is it's read a lot of information. It knows how to bring information efficiently together to answer questions. No, it's going beyond the frontier of human understanding and what humans have been able to do. And it's not just mathematics. We're seeing it in physics. We're seeing it in biology, chemistry, material science. And again, I'm not saying every problem is solvable. The models can't do everything yet, but they sure can solve open problems that some of the best scientists in the world haven't yet been able to solve.
I
Interviewer9:54
Give us like a flavor of an example of what is there going to be a new material that gets built in '26 that we've never had before?
K
Kevin Weil10:00
Yeah. So, take material science. Right. In material science, experiments are really expensive, right? It's very time consuming and often money consuming to go run experiments in the real world. So today you've got material scientists that do their best to bring together everything that they know to try and design a material with a certain set of properties and then they go and try and create that in the lab with the best hypothesis that they can have. I was talking to a scientist the other day and he said, 'You know what AI is for me? It's a metal detector for hypothesis.' So of all the things that I could think about, it brings together disparate information. It helps me sort through a whole bunch of ideas and land at the best possible way, only the bad ones I'm trying to think. Don't waste your time. And so that's its own form of acceleration. Instead of doing 10 things and nine of them don't work, you land on the one that's going to work right away. But then you start thinking about what about robotic labs because you don't need graduate students pipetting things and they need to go to sleep and they have other things going on. I think we will very quickly live in a world where AI is helping refine which experiment you're running or which set of experiments you're running and then it delivers the set of experiments to a set of robotic arms in a lab that are able to run the experiments themselves with as much parallelism as you want. By the way, you can scale that horizontally and then the results of those experiments are piped back into the AI. The AI reasons some more, designs a new set of experiments and you go, think of how much faster that moves than a traditional experimental setting. That's going to be the norm. I mean, we're going to see it this year. It'll be the norm very soon. And we will move faster as a society. We will discover more things. We will solve more problems this way. And I think that's really exciting.
I
Interviewer11:52
So the underlying apparatus and tooling for scientific collaboration is what you're talking about that you will have at Codex level in 2026.
K
Kevin Weil12:06
Yeah, it's that kind of transformation where at the beginning of the year you've got early adopters and they're seeing value but it's certainly not super widespread and by the end of the year it's like okay, the world has changed. Again, I'm not claiming every problem will be solved. Certainly not every problem in software engineering is solved either, but you can't deny that the world is completely different if you're a software engineer. And I think we're going to see that we will live in a very different world for scientists as well in a very good way.
I
Interviewer12:33
And how long have you been doing the AI for science thing now?
K
Kevin Weil12:38
Like three or four months.
I
Interviewer12:40
Okay. Yeah. What's the most surprising discovery that you had in the time that you were doing it?
K
Kevin Weil12:46
Um, that was unexpected. Like you're like, 'Oh my goodness, I...' Well, I mean, so I think there's something really interesting about how quickly we all as humans adapt to the pace of AI. Like, are people that aren't from here, have you ridden in a Waymo yet? If you haven't, you totally should because I guarantee you, at least for me, my first 10 seconds in a Waymo were like, 'Oh my god, watch out for that bicycle.' You know, like holding on to whatever I can hold on to. And then the next five minutes you're like, whoa, I'm being driven around San Francisco by a robot. And you're just like, I am living in the future. This is amazing. And then 5 minutes later, I'm bored looking at my phone like scrolling through the seats. You know, how quickly this thing that blew my mind is suddenly totally... and you know, day regard to me. And I think we see that. So I mean you go back six months ago, a year ago, the idea that AI would be solving open problems in mathematics was completely ridiculous, like couldn't happen, just you know, and people would say oh it's never going to happen. Now here we are and we're like well, yeah but it's not the Riemann hypothesis, right? It's not that open problem, it's only this. And you know we're gonna do that all the way across exponentially faster. So I think that's awesome actually. I think it's really cool that we adapt that fast.
I
Interviewer14:13
And do you think we have the right evals in science? Just like that was one of the areas where you couldn't tell the progress to some degree because we were not creative enough in the eval for actually even measuring how good the models have gotten.
K
Kevin Weil14:27
Yeah, it's really interesting because you get to a point now where in order to evaluate the model on a thing that is at the literal frontier of a scientific field, there aren't that many people that can do it, right? The model puts something out there and like in any field where you're sort of interacting at the edge of your abilities, especially at the frontier, the model is wrong as often as it is right when it's trying to solve something no one has solved before. And so you've got to tell apart like sometimes subtle things at the frontier, you know, in physics even, which was as close to my field as there was, I can't do it. And so we have physicists on our team who are on leave from a university because we need that level of ability in order to understand these things and to build new evals. So it does get really hard. And at some point we're going to need to use the model to build evals for the next phase of the model because we're going to be at the edge and maybe past where humans can do.
I
Interviewer15:32
I asked this question of Greg and I thought I'd ask you the same, which is what are things that you are worried about and also what are areas that you're really excited about right now which have a new set of possibilities that you did not even think possible last year? And that might be science is the answer. But what are you worried about most? And let's talk about something beyond safety and security because we've talked about that a fair amount during...
K
Kevin Weil15:59
Yeah. Yeah. Yeah. It's... I think that the unknown thing is just like what society is going to change is going to have to change really fast and what is that going to look like?
I
Interviewer16:14
Do you think universal basic income is like a foregone conclusion that we're going to have to have?
K
Kevin Weil16:20
I don't know. So I am not a person who believes that we will be happy sitting around like eating grapes and collecting our UBI and writing poetry. You know, I just think we as humans strive to do something bigger than ourselves and to accomplish things. I don't think that goes away. So I think we will, and that's what gives me confidence ultimately that we're going to get through all of this change just fine because humans are incredibly adaptable and full of ingenuity and drive. So, I'm optimistic, but I do think there will be a lot of change. One of the things that's most exciting though is just like we can all create. If you have an idea right now and say you had an idea two years ago and you're not an engineer and it involved writing code that was kind of hard to deal with, right? You had to go find somebody or go hire a team somewhere or whatever and probably the activation energy was just too much and you didn't do it. Now literally you can go write a prompt in Codex and it'll go work for an hour and you'll have a working version of whatever you were thinking about and then you can go back and iterate and make it better. There is no excuse not to be creating whatever you can think of.
I
Interviewer17:34
So what becomes a scarce resource in that world?
K
Kevin Weil17:36
Well, I mean we believe compute becomes a scarce resource, right?
I
Interviewer17:39
No, but I mean like does judgment become...
K
Kevin Weil17:43
I think like judgment, agency, I think this moment selects for people who are high agency more than ever before. Someone who says, 'You know what? I have an idea and rather than letting it stay an idea, I'm gonna go build it and I'm gonna build it this morning and I'm gonna have it in the afternoon and then I'm gonna improve it and by the end of the day I'm gonna have a thing that I didn't have this morning.' I think those people who are high agency, who are curious, who are learners, and who are just going to use the new tools to accomplish even more, I think that's... I hope that's not a scarce resource. I hope that's a very common resource but that's what the future is going to select for I think.
I
Interviewer18:23
Last question, why did you join the Cisco board?
K
Kevin Weil18:26
I mean, first of all, what an iconic company, important for any number of things including our national security, also what a critical moment where we need what Cisco delivers even more. And Cisco has an incredible opportunity with the way that the world is building out infrastructure. We really do believe compute is one of the most valuable resources. You know, basically like the more compute you have, the more intelligence you can provide. And I think there will be an infinite demand for intelligence. So Cisco has an incredible opportunity. You and Chuck and the team are driving one of the most impressive transformations I've ever seen with respect to a big company reacting and changing and taking advantage of what AI has to offer. I just think those are the most interesting moments of a company. Yeah. So I'm just, you know, proud to be a part of it.
I
Interviewer19:17
We love you being here. Is there a question I didn't ask that you wish I'd asked?
K
Kevin Weil19:21
What are you most excited about? I'm going to turn it around on you. What are you most excited about? And what are you most excited about?
I
Interviewer19:26
This is... I'm supposed to be interviewing. I know. No. I actually think one of our biggest concerns that we've had, we have always been resource constrained on the ideas that we have that we want to prosecute. We have an idea factory that's way larger than the resources to prosecute them. I think we might actually change that balance over the course of... and largely because of the body of work and partnership that we're doing with you and others, but largely with OpenAI. But I do feel like if this thing starts to really crank where our AI products are the first ones that actually get self-written and then eventually like we want to make sure that Martin is actually building silicon that is being done 80, 90% with AI, like you can start to. And it's not just for building products fast for the sake of building them fast. I think it's very seldom in life that you get a chance to be part of a movement that... look, I'm 54. This is probably the last big shift that I'll see in the next decade and this is a seismic one. It could change the shape of humanity and we actually can participate in making that happen and it can happen without the constant worry of me having to keep bugging Chuck saying we need more money, we need more money and we do it after. Now, by the way, Chuck, we'll still need some more money for a while, but like... I'm turning on the recorder. But I do feel like that's the thing that's the most exciting is if we get to build something that is so magical at very fast clock speeds, then the only thing that we have to be extremely careful of is that we don't have AI slop in the market and that we build things that are truly with a level of care and craftsmanship and judgment and intuition. But I feel like we as a community and especially at Cisco with the culture we've built, we'd be very good at that. And then not being constrained on certain areas where we're like, man, I wish I could have prosecuted that idea. Seems like a very very fun way to spend the next decade.
K
Kevin Weil21:35
Yeah, I love it.
I
Interviewer21:36
Thank you for being here.
K
Kevin Weil21:37
Yeah, thank you for having me.