Entrepreneurial Thought Leaders (ETL)
00:38:56 · S4, Carina Hong, S3, Will
[00:38:56] S4: But so now with, like, AI advancing so quickly, do you think there's still time to join, a Frontier Lab or, like, learn as you essentially get your training in before you start or do you capitalize now and just start? [00:39:09] S2: One thing is I think you can learn a lot in a very short amount of time. That's I think one thing it's unexpected. Like learning is also... Everything's being time compressed. Right? So you're learning, make that also like more time compressed at the general index of time being time compressed, I think it will be good. The other thing is when I say the first question is to choose a hard enough problem. When it's hard enough, then you give yourself a little bit more buffer to really like go through, I call the learning process of from pre rigorous to rigorous to post rigorous. [00:39:39] S4: Yeah. Thank you. [00:39:40] S3: Thank you. [00:39:42] S5: Thank you so much for being here. I'm Will. I'm a math major. I'm just wondering what you think the mathematician of the future, what their job will look like and whether, you know, I'll be able to apply the skills I'm learning right now in the future.
00:39:56 · Carina Hong
[00:39:56] S2: Yeah. So the skills that you're learning definitely I think will be applicable for the future. The general question, I have a lot of thoughts about math the human human math. So I kind of talked about it. It was a lot of mathematicians too, who see different parts of the spectrum. One, this is all decision tree. One, is math closer to cancer biology or is math closer to chess? I think one needs to kind of have their work view and make a decision on that. If it's like cancer biology, then it has intrinsic value, a problem being solved, a discovery being made, something being formulated as a conjecture. It has intrinsic value because after AI compress a timeline of landing to practical word, then you will basically have math be super useful to scientific applications, various applied scientific and technological engineering applications. So that's what I believe in. I believe it's closer to cancer biology than to chess, but some other people also do math for the fun of it, then to those people it's chess. And then my arguments subsequently will not apply. So if you're with me, that is closer to cancer biology, and AI will compress the timeline to make it seem more useful despite it already is. Cite Hardy's mathematician's apology. Then there's a question of, why do you need human understanding if you can get the drop down? And the most convincing argument I have heard right now is that if you can, for example, have human understanding that you can have paradigm shifting ideas that AI is not capable to come up with. And those human geniuses are going to, you know, discover methods like maintenance sieve for bond gap, for example, to push the entire field forward. And now the best counterargument of that is if you assume AI progress continues, which I think is a capital markets problem Okay. Not a technological problem
00:41:39 · Ravi Belani, Carina Hong
[00:41:39] S1: Yeah. [00:41:40] S2: Okay. If you assume AI progress continue, then it will be able to conjecture. It will be able to build mathematical theories. It will be able to even come up with those out of distribution paradigm shifting ideas. Then there is a sort of... Then the counter to that, which is that, okay, well, where are human mathematicians, as you say? And if you believe that it has intrinsic value, then mathematician will have more of like a dev workflow because coding is an example of something that has intrinsic market value. If you do not believe it, then then it's the chess route. Then I don't think it's going to apply. [00:42:14] S1: But if the capital markets fulfill the demand and AI continues... [00:42:18] S2: The worst case actually is... [00:42:20] S1: Our our our math majors will still be relevant? [00:42:23] S2: I think the worst case is capital markets crash, AI progress hinders. Not a lot of people are sort of, you know, invested in understanding math. And so there's less paradigm shifting ideas, and then we'll be in a sort of plateau situation. That will be the worst case. [00:42:40] S1: And in the best case, if if the markets continue. [00:42:43] S2: The best case continues and a lot more people are going to be mathematicians, developed... This defined as people involving mathematics as a significant or majority part of their workflow. [00:42:54] S1: And and and their need isn't gonna be isn't gonna be supplanted by AI machines in the future?
00:42:54 · Ravi Belani, Carina Hong, S3, S6, S4
[00:42:54] S1: And and and their need isn't gonna be isn't gonna be supplanted by AI machines in the future? [00:42:59] S2: I think if you assume really AI progress continue, then you would actually need more mathematical theories to derive specification, what to prove out of the data alongside with the code base will also develop. So now there's a question of, of course, like, time bridge. So it might be great, and then it might crash. And so that might be a problem. [00:43:27] S3: Thank you. Okay. Thank you. [00:43:31] S6: Hi. Can [00:43:33] S1: you go us... Go go closer? [00:43:34] S4: Oh, yeah. Yeah. [00:43:34] S6: Yeah. To give you some context behind my question, I work with a group named... Called the assumption of Physics, and they bring together researchers from all around the world to try to figure out what is the minimum set of physical assumptions that you can find that re... That gives you physics. And this also implies that we making physics mathematically consistent. And since Axiom is, you know, is interested in math, is Axiom also looking towards physics? Because physics is a field where the interpretation of the physics also really matters, and you also need to back up those interpretations with predictions. So is this something that Axiom is looking towards?
00:43:34 · S6, Carina Hong, Ravi Belani
[00:43:34] S6: Yeah. To give you some context behind my question, I work with a group named... Called the assumption of Physics, and they bring together researchers from all around the world to try to figure out what is the minimum set of physical assumptions that you can find that re... That gives you physics. And this also implies that we making physics mathematically consistent. And since Axiom is, you know, is interested in math, is Axiom also looking towards physics? Because physics is a field where the interpretation of the physics also really matters, and you also need to back up those interpretations with predictions. So is this something that Axiom is looking towards? [00:44:13] S2: Yeah. This is very interesting. So we have a few collaboration projects with applied, like, groups. We haven't... We actually maintain... We started to maintain mass econlib. It's a econometrics library. I believe we have a quantum computing partnership with like another like quantum computing academia group. I believe we have theory of deep learning. Just physics, I think there is a Physics Lib, which is the open source communities library for physics. So we haven't been sort of like navigating that space specifically. I am intrigued by the sort of you described the first order logic way of looking at the minimal set and like what can be built, right, like on top of it in the sort of lean dependency tree fashion, but we should definitely talk afterwards. Yeah. [00:44:57] S1: That's great. Thank you. Last question.
00:44:57 · Ravi Belani, Carina Hong
[00:44:57] S1: That's great. Thank you. Last question. [00:45:00] S2: Hi. I'm Sanvi. Great to meet you. This is just a more minute question, but you mentioned that the future of agents is different than sort of the hard code harnesses we've been seeing, but I was wondering if you could double tap on that and more describe what you're talking about. Yeah, we need trainable harness, right? So now I think there are a few labs that have the ability to do agentic RL for the model. And so the model will be able to have access to all the tools, all the skills, even call itself for smaller checkpoints of itself, right? And I think this is going to be the 2026 new award, or we might have a 2027 new award, and I don't know what that's gonna look like. I think swarm is pretty much up there. Without swarm, we will not be able to get to what we have done. And there's a question of optimizing the cost of swarm, which is going to be the next frontier. And part of optimizing the cost of swarm is you need like very good small models. And because of that, I think the sort of small labs training, genetic RL model story will not end because you probably need a larger one than a smaller one. Easier. It's hopefully easier to train. Thanks. [00:46:07] S1: Well, everybody join me in thanking Carina for sharing so profusely even though... Sorry about that. Especially grateful given this is a day of resilience of of of serendipitous trials. So we appreciate it, Carina. Thank you, everybody. We'll see you next week with Celine from Loyal.