How to Build Viral Products: Lessons from Kevin Weil, VP, OpenAI for Science

Thank you for having me.

Yeah. Okay, I'll do the fast one. I was doing my PhD in physics at Stanford. Thought I was going to be a physics professor for the rest of my life. Met my wife who was a Mayfield fellow. I think there's a Mayfield program here at Berkeley. Basically, she'd like gone through college working at startups. I had no idea that that even existed, but she kind of opened my eyes to it. And so I was like, "Oh man, that sounds way more interesting." You know, I mean, I can write some code and ship something to a million people tomorrow instead of, you know, maybe making a contribution to physics over 40 years. So, I started working at startups. A couple you've probably never heard of because they failed. And then I went to Twitter when it was about 40 people. Had no idea what I was doing but joined as an engineer and ended up as head of product after like seven years. Was head of product at Instagram, was one of the co-creators of Libra if people remember Libra the crypto project, also didn't work but there are good stories. And then was president at Planet, Planet Labs building satellites and imaging the whole world every day. And then joined OpenAI as chief product officer about 18 months ago and just moved to a new role focusing on accelerating science, like using AI to accelerate science all around the world because I think if we can do that, if we can do the next 25 years of scientific discovery in five instead using AI, maybe we'll have an even bigger impact than you know building ChatGPT. So that's the nutshell. And I guess since most of you are entrepreneurs, my wife runs a venture fund called Scribble Ventures and I'm very lucky to get to kind of work nights and weekends with her. It's a pre-seed fund and I've always enjoyed getting to do a little bit of investing and a lot of operating because I always find I learn so much from the entrepreneurs that I get to work with and you know hopefully can help them not fall in some of the potholes that I've fallen in over the course of my career.

Super cool.

Oh, that's a good question. I think one of the biggest ones, especially in the early days, is trying to build a product for everybody. Paul Graham says this well, but I think it's really true. You'd rather have a product that a thousand people absolutely love and can't live without than a product that 10 million people are kind of like, "Yeah, I guess that's better." Because you have to remember that most people don't live their days thinking about your product and they already have habits and habits are hard to break. And so you need to have that product that they're just like, "Oh my god, if this thing went away, I would be so sad." Even if that's a small number of people, because you'll learn a ton from that small group. And then you can expand from there. Whereas if lots of people like you just a little bit, you get a lot of people trying, none of them really stick around and you don't have a product that works. So like find that thing, like it's better to start small and get your foot in the door with something that people really love. And I think people make the mistake of trying to like, you know, Facebook didn't start out going after the whole world right? It started out going after colleges and I think it wouldn't have succeeded if it tried to be everything to everybody early on.

I think I don't know if there's a, there's obviously not a rule because different products go after different segments. It's really like what problems are you trying to solve? And that's another thing I learned over and over again from Kevin Systrom in particular at Instagram is like if you can really tightly articulate the problem that you're trying to solve then a lot of these other things, you know exactly how the product looks and who you're building for, you know what's your ICP or what customer segment are you looking for, a lot of that actually kind of becomes obvious if you really tightly define the problem you're going to solve. If your problem is really broad, like, oh, we're gonna make it easier for people to share content. Well, that doesn't help you at all because you haven't really, that's, you know, I don't go around saying like, gosh, I wish somebody would help me share content, right? So, like what user problem are you trying to solve? Make it as precise as possible. Make sure you understand why that's a problem, for who that's a problem, and then a lot of the stuff actually kind of falls out the other side. But it takes work to, you know, it takes user research and other things to really understand the problem you're solving.

Well, I think it's really cool, right? When you think back 12 months ago, so go back to November of '24. Most people that were writing code were writing code themselves and just like we were kind of doing things the way that we always have done things. You go 12 months to today and it's completely different. I would imagine that most of you, most of your companies are using Claude, Codex, Cursor, you know, one of those for like 90% of what you do. And if you're not, you probably should be. It's amazing how much changes in 12 months with AI. And the cool thing about that is it's not just your engineers that can build stuff. There's very little reason for anybody to make like a PowerPoint deck anymore, or Google Slides or whatever. Like product docs are very different now. Nobody should be writing a PRD. You should be building stuff. Whether you're a product manager, a designer, I mean, our finance team at OpenAI builds tools. They vibe code them with Codex and like solve their own problems where in a previous world you'd be like, "Oh, can I just get an engineer?" And of course, nobody's prioritizing the work that finance does because that's just how it works. It's not the, you know, it's usually not your sort of do or die thing. So it's hard for them to find engineers and they end up with these super manual workflows that never change because they don't have the resources to build this stuff. But guess what? Now they can. And that's amazing. So everybody at your company can now be building stuff. And maybe they're not all building, you know, production code, but they're building tools and they're solving their own problems. What a cool world that is. And like it was 12 months ago that the world was completely different. And I think that's a fun, like it's just a really cool, we're just in a really cool part of the cycle because I guarantee you in 12 months the world is going to be completely different in some other way. I have some guesses but I could be completely wrong. But that just means if you're an entrepreneur, every single product that we use whether it's hardware, software, like any of the big billion user software products, the phones that we all have, like everything will be reinvented over the next five years. Because everything that has huge scale today was built pre-AI and post-AI everything can be totally different. So you know are some of them going to reinvent themselves, some of the big incumbents maybe they could but history would suggest that most of them won't and that it's actually a time for disruption and that's a massive opportunity for every single one of you in the audience and I think that's really exciting.

Well, I mean, I switched from focusing on product and thinking about like ChatGPT and our other products all day, which was the coolest product job I've ever had, to working on science. So, you know, that gives you some sense. I think there's a real possibility that 2026 is the year that the way that we do science completely changes and that when we're sitting here in November of '26, we're like, man, do you remember when you had to like do all of this sort of grunt work in science yourself and AI couldn't even, you know, be a brainstorm partner. We're already seeing ChatGPT being used not just to do the kind of things you're used to seeing ChatGPT do, but to actually push the frontier in science. Like you have math professors who are tweeting, and we see this like once a day at this point, who are tweeting, hey I had this idea I was going to give it to my postdoc, I was pretty sure this theorem was true but I gave it to my postdoc, he was busy, it took him a week, he didn't respond so I gave it to ChatGPT and it just solved it in 10 minutes. And it's not, you know, we're not proving the Riemann hypothesis yet, I'm not trying to sort of get out too far ahead, but it's not even necessarily like a full theorem. It's maybe a lemma, but still this is ChatGPT. This is an LLM, an AI that we all built, moving beyond the frontier of what humans know. It's not yet better than what humans could do. In most of these cases, the math professors are like, "Yeah, if I put like another, you know, few hours into it, I probably could have proved that." But ChatGPT did it in 10 minutes. And that's with today's models. And you have to remember, if there's one thing you remember walking out of here at the end of today, the model that you're using today is the worst model that you will ever use for the rest of your life. They're only getting smarter and they're getting smarter very quickly. And so if a model can prove small novel things today, can help with biological research, can help with material science, just imagine where we're going to be in 3 months, in 6 months, in 12 months, let alone three or four years.

I think you have to be using AI in everything that you do. And if you're not, like you're just going to get beaten by somebody who is. And I think that goes for like the product you're building too. The sort of counterpoint to it's, you know, everything is going to be reinvented, everything that we use is going to be done differently, like docs should not be the same docs in a few years, we should be thinking completely differently and the product probably looks and feels completely different. The contrapositive or whatever to that says if you're building something that kind of looks like the way we do things today, then you're probably not building it right. There's probably a different way. And be comfortable building in a place where AI is like only barely working. Sometimes it's a little nerve-wracking when you're building something and the AI is just like only kind of good enough, maybe not really good enough. But if there's one thing I've learned over the last like year and a half at OpenAI, you go through this phase where AI just can't do something right. Name your thing, there was some point where AI couldn't do it and then you get these glimmers where it's just starting to be able to do it. And like 10, 20, 30% of the time it gets it right and it's kind of annoying, but when you hit 10 or 20 or 30%, very quickly you go to like 90%. And then very quickly you go to like, oh, of course I use AI for this and I will never not use AI for this for the rest of my life. So that sort of existence proof of like it kind of works, you go very quickly to it really works. And so I think that area where it kind of works is a great place for entrepreneurs to be building because if it kind of works now, it's really going to work in 6 or 12 months. And if you're the first one to realize that, then you're kind of riding the frontier, which is where you want to be as an entrepreneur.

Risks with the development of AI? I mean, yeah, sure. There's risks with any new technology. There's probably even more risks with AI because it's so powerful. The thing that I think is different with where we are today, having seen the inside of Twitter from the very earliest days and if you go back to 2009 for folks who remember, Twitter was being used like during the Arab Spring and people were using it to communicate when the government had shut the internet down and all these other things and we talked about Twitter as the free speech wing of the free speech party and speaking truth to power and we celebrated that in a major way. We didn't realize at the time that there was another side to that coin and all the things that gave users that power also could give, you know, other people, governments, etc. the power to do bad things with it. We just didn't realize, you know, it was like sort of the first time that we had mass digital media like that. You could, you know, look at any of the various missteps that Facebook has had over the years. I think we're much more, what's the right word? Wary is maybe not the right word, but at least like we sort of know to expect it in a way that I think back in, you know, 2009, 2010, we were all just collectively well-intentioned but naive. And so when we launch a new model, I mean, we have an entire safety team. It's like a large number of people that spend all their time doing AI safety research and then working on models before we launch them to make sure that we round all the edges, that you can't jailbreak the models, that like they don't answer questions they shouldn't answer, etc. We work with third parties both to red team our models upfront. We work with government groups who have responsibility who take the models and try and like break them before we launch them to the public. So, it's just like a much more robust program than anything that I've ever seen before. Which isn't to say that we won't make mistakes. I'm sure we will make mistakes, but it's a system designed to make sure that we make like more small mistakes rather than a big mistake.

Oh gosh, I don't know. How about right in the middle in the back?

I think I mean nobody goes to school to be a product manager, right? In the same way that nobody goes to school to be an entrepreneur. So I think there's a mix of it helps to be technical, it helps to have good EQ to be able to put yourself in the shoes of users or your team, you know, empathy and so on. You're also in a product role most of the time you're working with a team and you have some leadership role to play. I won't say you lead the team because I think PMs contribute one way and designers contribute one way and engineers another but you have some leadership role to play and nobody reports to you right so good PMs can get people to follow them. And I don't even like, it shouldn't be about the PM having the ideas it should be about good ideas coming from everywhere but the PM has a strong responsibility in making sure that the team ultimately goes in a single direction so you're not just like saying yes to everything you're sourcing ideas from the team and then sort of collectively saying, "All right, this is the bet we're going to make." And then making sure everybody knows it and is rallied around it and understands why you're doing it and what direction you're going in. So like that's a, I don't know how to describe that skill, but it's a very important one. And then at OpenAI in particular, one of the things that we really look for is just agency. Like no one is gonna, if you wait around and ask someone like, "Hey, am I allowed to do this?" No one's gonna, like if you're waiting around for someone, everyone else is building and you're not. So, at OpenAI, the trick is just like it's the whole you can just do things ethos. You have an idea, prove it out, use Codex, build it yourself tonight, and then, you know, prototype it and show it to the team tomorrow. And I think that's one of the coolest things about being a PM in this age.

Sure. We can try one or two and see how fast I can go.

I mean a lot of it comes down to the labs but I, everybody in any product you build, I don't know if you're a B2B or consumer or whatever but you still have a sort of trust and safety component. People are using your product in some way. Can they use it to do things that you don't want them to do, that they shouldn't do, that weren't what you intended them to do? Depending on your product, that's either a big deal or not a big deal. But that's a big component of what we do, too. There's, you know, super high, you know, like fancy AI research, and then there's also very like down and in dealing with fraud and people trying to get the model to do, you know, things that they shouldn't do. And your product will have those challenges too. I think some of the really big safety research stuff is going to be with the frontier labs, but then every product has its own kind of integrity, trust and safety surface, and you got to get that right, especially as you scale.

I think a lot of people are still using GPT-4 and using some of the non-thinking models. I actually basically every single thing I ask GPT I turn it into thinking mode. Because I find I get better answers even if it means I wait five or 10 seconds for it to do a little bit of thinking. With GPT-5 it does a much better job of thinking as much as it needs to. So if you ask it an easy question it'll give you a pretty quick answer and if you ask it a really hard question it'll think for a long time to give you a better answer. We worked very hard to make sure that it doesn't just sort of try and memorize things because you do get far better answers when it thinks. So I don't think that's as much of a problem anymore. One of the interesting things though is when you're working on the hardest problems, when you're, you know, you're a mathematician trying to do a proof, you're in biology and working at the frontier. The model like a human, if you're really at the frontier, if you're asking me the hardest math problems that I can do, sort of by definition, I'm not getting them 100% right. Right? Maybe I'm getting it 10% of the time. And one of the interesting challenges is if a model's only right 10% of the time and you know you're dedicated, you're trying to get it to solve this problem and you try four times and it doesn't work. That actually doesn't mean that it can't solve the problem. It can potentially. You just need to try more. You need to give it more compute, more thinking time. But that's not easy to see from the outside. Like it's very hard to tell the difference as a ChatGPT user between a problem that the model just can't solve yet and a problem that it can solve 5% of the time. And that's a product problem that we're looking to solve. So that you can enter in harder and harder problems. And if the model can do them, it will just like realize that it needs to think more and more and more. Maybe try multiple parallel agents working at once and ultimately get you an answer. Like basically the conclusion is the model is much better than people even realize. But on low pass rate problems it's hard to tell the difference. And so we're working on exposing that which will help for some of these really hard problems.

Yeah, thank you so much for having me.