Invest Like the Best with Patrick O'Shaughnessy
00:08:43 · Sarah Guo
[00:08:43] S2: To the point of perhaps making a bet that others wouldn't, we thought very carefully about what markets are going to matter and what might be different about the founders that we look at in this era if we are right about capability growth and the breadth of impact. That would just be non obvious to other people. Where's the biggest difference from the status quo? How does the framework change? And I'll give you two examples. One of the things that we were really looking for in the first year was application areas, workflows and professions, tasks that we thought were a good fit for purpose for the models, which is a very technology forward approach. Lots of people be like, this is nonsense. You gotta think about the customer problem and working from the customer back is the only way. We wanna do both. And if you look at Harvey and the function of the law, rationally, if you think that we can do next token prediction with language and you knew that in late twenty twenty two, then law is structured language. I'm not a lawyer. But from the outside, I'm like, you need to, like, read a lot of documents. And we had retrieval and you need to generate text. And there's a lot of precedent text both inside firms and in common law and history. That feels like a really good match. We also took a very specific view of what is now possible and then what is valuable within what is possible and then who's aligned with us. Winston and Gabe, they believed that AI would transform the practice of the law in a very AI pilled way. We will do enormously complex work with lawyers. I don't know when. It could be next year or five years from now. But from the kernel of I can look at a landlord tenant agreement in California and answer a question, somewhat trivial, to I can project to doing a Activision Blizzard M and A and doing 85% of the work. That's a leap. The thing that appealed to me in that moment was the ambition of what was possible then and the technical logic of, like, why it would work. I think that's probably just different decision making framework from how other people are approaching in that moment.
00:11:06 · Patrick O'Shaughnessy, Sarah Guo
[00:11:06] S0: What is that group talking about today, the two fifty researcher frontiers group? What's the most interesting thing and the most notable difference between today and six months ago or twelve months ago or something? [00:11:15] S2: It is a violently competitive landscape. Think I that was true twelve months ago, but even more true than it was '24 and then '36. I think people are very concerned that it's a globally competitive landscape, and there's insecurity in that because I think it's a bit narrative breaking as well. I'm going to describe a belief and then question the belief. The belief is with recursive self improvement of AI research models that can improve the models themselves, we are a year, two years away from some sort of exponential intelligence. I think that belief is new within the last twelve months for a lot of researchers. Some driver of it that we are two years away, Andrej Karpathy will actually in a very self aware way say, I thought it was two years away for about ten years. And he thinks that again, to be fair. Who can say? I do think that there is some sense of the major labs are so compute intensive and so large from a headcount perspective now that the sense of contribution of I can move the needle of OpenAI as 200 people. There's not that many researchers. The question of like, how do we get there is really up to every single person. Now if the question is, well, I need $750,000,000,000 of compute spend and we have many thousands of people working on this problem. I think people feel less ownership of the outcome.
00:11:15 · Sarah Guo, Patrick O'Shaughnessy
[00:11:15] S2: It is a violently competitive landscape. Think I that was true twelve months ago, but even more true than it was '24 and then '36. I think people are very concerned that it's a globally competitive landscape, and there's insecurity in that because I think it's a bit narrative breaking as well. I'm going to describe a belief and then question the belief. The belief is with recursive self improvement of AI research models that can improve the models themselves, we are a year, two years away from some sort of exponential intelligence. I think that belief is new within the last twelve months for a lot of researchers. Some driver of it that we are two years away, Andrej Karpathy will actually in a very self aware way say, I thought it was two years away for about ten years. And he thinks that again, to be fair. Who can say? I do think that there is some sense of the major labs are so compute intensive and so large from a headcount perspective now that the sense of contribution of I can move the needle of OpenAI as 200 people. There's not that many researchers. The question of like, how do we get there is really up to every single person. Now if the question is, well, I need $750,000,000,000 of compute spend and we have many thousands of people working on this problem. I think people feel less ownership of the outcome. [00:12:46] S0: Oh, that's interesting. So it's just like a function of getting compute. That's the thing that's gonna push this thing over the line, not our own human efforts.
00:12:46 · Patrick O'Shaughnessy, Sarah Guo
[00:12:46] S0: Oh, that's interesting. So it's just like a function of getting compute. That's the thing that's gonna push this thing over the line, not our own human efforts. [00:12:53] S2: I definitely think there's a large contingent of researchers who would feel that one of two things is now true. What I do doesn't matter anyway because the model is gonna do it. Or two, the only thing that matters is compute scale. And both of those are somewhat disempowering. [00:13:07] S0: So then what are they doing? If I believe both of those and I'm a top five researcher or something. [00:13:12] S2: That says something about your psychology because you want to do something that matters. I think there's also a set of scientists who want to work on the thing whether or not it matters that they're working on it. An entrepreneur asked me yesterday, I would not work on the companies if I felt like we couldn't change the outcomes a little bit for them. He asked me yesterday, would any of the companies that you backed not been backed without you at that round? Would people just said no? 5%, 10% max. They were resourceful, really talented people. They find other investors. There are lots of smart investors in the world who wanna take that risk. Maybe they wouldn't have gone the next $100,000,000 of compute. I don't think every researcher doing frontier work at these top couple labs right now feels like they're essential to the machine.
00:13:12 · Sarah Guo, Patrick O'Shaughnessy
[00:13:12] S2: That says something about your psychology because you want to do something that matters. I think there's also a set of scientists who want to work on the thing whether or not it matters that they're working on it. An entrepreneur asked me yesterday, I would not work on the companies if I felt like we couldn't change the outcomes a little bit for them. He asked me yesterday, would any of the companies that you backed not been backed without you at that round? Would people just said no? 5%, 10% max. They were resourceful, really talented people. They find other investors. There are lots of smart investors in the world who wanna take that risk. Maybe they wouldn't have gone the next $100,000,000 of compute. I don't think every researcher doing frontier work at these top couple labs right now feels like they're essential to the machine. [00:13:56] S0: What do you think is unhealthiest part of everything going on right now? Like, what worries you about what people are trying to accomplish and what are the things that you're most worried will inhibit the future that you wanna see?
00:13:56 · Patrick O'Shaughnessy, Sarah Guo
[00:13:56] S0: What do you think is unhealthiest part of everything going on right now? Like, what worries you about what people are trying to accomplish and what are the things that you're most worried will inhibit the future that you wanna see? [00:14:09] S2: I think compute is one. People, I think, that very well. The version of us having enough compute over the next five to ten years. And people are like very much thinking about like 2032 at this point at scale. I was talking to the leader for infrastructure at one of the hyperscalers earlier in the week. And he's like, there was nothing that is gonna move the needle for us at sufficient scale before 2030. That's depressing. Where can we get sufficient natural gas? Americans and entrepreneurs have been able to build new technologies and new capabilities very, very quickly in the past. I don't think it is a technology or capability or capitalism problem. I think it is a regulatory problem and an alignment problem. And I don't mean like AI alignment. I mean, if you want to build data centers in New York, you need to convince the people of New York they should want data centers there or America should want data centers there. If you want to make the price of nuclear competitive as baseload power, then you need to convince people it's safe and you need to allow enough construction [00:15:21] S0: of AP1000s. Yeah. Rest of cars.