No Priors: Artificial Intelligence | Technology | Startups
00:11:00 · Vlad Tenev, Sarah Guo
So there was all this pressure to succeed academically, and for me, my math PhD program, at some point it became kind of this challenge, like what's the hardest thing, how can I do better at the hardest thing, and it was this almost like competition with myself, with my parents who were also very academically strong, like, oh, what was the hardest subject for you at school, dad? Oh, chemistry. Let me just do really well at chemistry. Yes. Know, what was the most advanced math that you did? And I think math has this, like, ladder of achievement. It's almost like playing a really good video game, like, you always know what the next hardest thing is, and it's almost like how quickly can you go to algebraic geometry or algebraic topology and can can you can you learn that earlier and kind of understand it. And I think what appealed to me when I got to grad school was this idea that if you are... If you create a new mathematical theory, you know, maybe you'll be remembered for a really long time and you know, just like kids in school everyone learns you know, Gauss and Riemann and all these greats. If you create a theory, maybe kids will be learning and sort of like thinking about your work hundreds of years from now. So that was kind of what appealed to me. [00:12:25] S1: Just given that background and, you know, you should also explain what Harmonic is as an effort, what is your reaction to the revolution in math that's happening with AI today?
00:12:38 · Vlad Tenev
[00:12:38] S0: I think on the one hand it's very exciting and sort of predictable because the rise of AI models and the ability to verify the outputs put together has led to kind of automation of the math pipeline and I think it's really really exciting, right? I think that the power and the tools, I mean, the math is changing so quickly, and, you know, math as a field hasn't changed very rapidly at all in thousands of years. I mean, when I was in grad school, most of the work was still being done in person with like chalk and a chalkboard. It's like there was no collaboration. People weren't even using computers. So now math has kind of converged with coding. Mathematicians are using the same tools, they're using version control, they're using Lean for verification, and the field is accelerating very rapidly. About a year ago, so Harmonic, which is a startup that I co founded a couple years ago with the goal of building mathematical super intelligence. So last year we created a model called Aristotle that was one of three that got a gold medal on the International Math Olympiad. And then subsequently to that, we rolled it out to the public and we had people using Aristotle to tackle unsolved math problems. And so it was used to solve these Erdosche problems. There were these other mathematicians that were using it for new results and so forth. And so then I started thinking a lot about, okay, what's the future of math? How should these tools be used? And I think for me, it was very important for the tools that I build, not just Harmonic, but also Robinhood to be empowering to individuals, to people. And so the one thing I don't love about this latest wave of AI models solving math problems is that it's too centralized, right? It's like AI companies with their big data centers and their employees using secret models to sort of do all this work. And I think what's more attractive and what would ultimately win out in the end as a model is rolling these things out and actually putting them in the hands of people and having the users make interesting discoveries and inventions?
00:15:29 · Sarah Guo, Vlad Tenev
[00:15:29] S1: So, I am probably like a pretty fundamental believer in decentralization as as a philosophical premise, and I would love to believe that's true. I think that there are a lot of people who work at large centralized AI companies who would say, like, if I'm if I may push back a little bit that this is Cope and Centaur play will just be outclassed by, you know, more compute in an agent loop. Like, how do you, like, do you believe that's true, or what will be the human input here that is productive in math? [00:16:04] S0: Well, I think I think it's sort of like two separate things, right? One is what's the role of the individual mathematician? And the other is, is this gonna be like a product that people use, or are the AI labs gonna be producing the products? I mean, have thoughts on both of them, but on the second one, it's much more interesting. I think that a world where you have the AI labs on a proprietary basis making all the discoveries, and you have like the OpenAI or Anthropic data centers that are just running code, you know, managed by the employees of those companies, and you know, they're vertically integrating and producing all the drugs and selling the drugs to people, making all the mathematical and physical discoveries and publishing all the papers. If all of that is happening and it's owned by the AI labs themselves, I think that's a bit of like a dystopian outcome and unlikely to happen for a couple of reasons. Not just like technological and market, but I think also if that happens to a significant degree, people will, they'll
00:16:04 · Vlad Tenev, Sarah Guo
[00:16:04] S0: Well, I think I think it's sort of like two separate things, right? One is what's the role of the individual mathematician? And the other is, is this gonna be like a product that people use, or are the AI labs gonna be producing the products? I mean, have thoughts on both of them, but on the second one, it's much more interesting. I think that a world where you have the AI labs on a proprietary basis making all the discoveries, and you have like the OpenAI or Anthropic data centers that are just running code, you know, managed by the employees of those companies, and you know, they're vertically integrating and producing all the drugs and selling the drugs to people, making all the mathematical and physical discoveries and publishing all the papers. If all of that is happening and it's owned by the AI labs themselves, I think that's a bit of like a dystopian outcome and unlikely to happen for a couple of reasons. Not just like technological and market, but I think also if that happens to a significant degree, people will, they'll [00:17:28] S1: lose their It is license not acceptable. To Yes. I [00:17:31] S0: think it's much more, I think you could have the same outcome of Navier Stokes being proven or the partial Riemann hypothesis or any one of these things, and it'll be much more accepted if it was like a random lay person using [00:17:52] S1: It might be celebrated. Yeah.
00:17:52 · Sarah Guo, Vlad Tenev
[00:17:52] S1: It might be celebrated. Yeah. [00:17:54] S0: Think for sure. And I think that's how it's gonna happen. That's the likely scenario one year from now. I think you also have confounding factor of a lot of the people that are going into these AI labs are former researchers, and so they're like excited by doing research, and I think they're confused as to why this is being poorly received. From their perspective, it's like, I'm a researcher, I was brought into a research organization to do research, and now all these academics are mad at me. But yeah, I think that, yeah, and I've said this to folks that work at the companies, I think it would have been much better if they found some mathematicians and kind of collaborated with them to not make it seem like there's this dark data center dystopian future where one company is gonna make all of the scientific discoveries. [00:18:50] S1: Can we talk a little bit about both the, you know, new ways you wanna extend ownership as part of Robinhood and then also how Robinhood is intersecting with what's happening in AI? Maybe we can just start with what you guys are doing with tokenization or your vision for it.
00:18:50 · Sarah Guo, Vlad Tenev
[00:18:50] S1: Can we talk a little bit about both the, you know, new ways you wanna extend ownership as part of Robinhood and then also how Robinhood is intersecting with what's happening in AI? Maybe we can just start with what you guys are doing with tokenization or your vision for it. [00:19:06] S0: Yeah, so tokenization is really, it's a new innovation and the idea is almost similar to stablecoin. So the way a stablecoin works is you take some dollars or treasuries, put them in a wrapper, put them in a box, and then you mint and burn tokens against the box, and the traditional assets stay custodied at a traditional institution like a like a bank, and then the tokens can trade freely on DeFi. So they can move anywhere around the world, they can trade in fractional amounts, they can be swapped in different protocols or lending pools, and that opens up just [00:19:48] S1: They have ton software characteristics.