Entrepreneurial Thought Leaders (ETL)
00:15:44 · Carina Hong
[00:15:44] S2: Yes, I think this is a great question. AI, I think is just changing so rapidly, and then when you set goals, you need to make sure you're not setting an ambitious goal for the next quarter, you're setting an ambitious goal for the year. If you don't plan it in that way, you will be late, you will be like behind the curve, because the frontier is updating very fast, right? So I remember back in March, we were saying, is it possible that we do this formalization of bounded gaps in primes? It's a very cute result, which is there are infinite many pairs of primes that are bounded by a finite number. And what that finite number is keeps being pushed to lower. When you get to two, you get the two prime conjecture, congratulations. But in 2013, when it was 70,000,000 by Itau and Zhang, it's the first time we had to find that number. The world record has kind of stayed in like 2014 for 246. And there was a human mathematician, Julia Stalman, who pushed it to 240. That was very recent. And we're actually in a pretty sort of short, like time crunch race with her in the end. But when we wanted to formalize the result of 246, the result from 2014, most people in the company believed that was not possible in the sense that it will just take way too much work. And they didn't want to start the school. We have to kind of like be like, well, look, at the end of the day, we have to do something and we have to make an end out there. So I think that's one really big takeaway I have is we need to set ambitious goal and end up like at the time of boundary gaps in plans, people consider that natural that we are able to do that. The second thing I think is to be it's a dichotomy. You need to be both want to do everything and okay to let it go and use other stuff. I So think that's kind of a bit, it sounds paradoxical. When I say that you need to want to be doing anything is we take a full stack approach. Like we have our really big provider data set, we don't have data vendors. We figure out data strategy ourselves. We have a model training team, they do intense mid training and post training. We build our harness, we have our harness infra, like we have our versioning control. Almost nothing on the stack are like outsourced to others, except like we don't do pre training. That's sort of the scale of the ambition being full stack.
00:18:04 · Carina Hong, Ravi Belani
And the benefit for that is we should be prepared that for some parts of the stack, we will be beaten by others. And so it's okay to use other stuff when needed, but you kind of need to like cast out your net widely. And I think it's also like today, you might not be at the frontier. In three quarters, you might be. There are a lot of other companies like Harvey saying how they build a frontier model, a frontier model training lab after actually the company has reached significant revenue at scale milestones. So I think that's such an example. It's like, it's not a sprint, it's a marathon. And the reason I say I'm specifically thinking about companies that are like not adaptive enough to advances. So for example, when Jeff came out, we tried that classification model day one, like and then like immediately, I think by the evening, data warehouse, a lot of classification tasks are using Jeff. We're just kind of very interested in trying everyone's stuff without a sense of defensiveness in the sense that we could be building the same stuff, not in that specific example, but in other cases we have seen. [00:19:10] S1: So you're you're you're open... You're you're putting a lot of bets, sounds like. You're open to being wrong on certain ones, but you have enough bets where you have confidence that you'll still be relevant and and need
00:19:10 · Ravi Belani, Carina Hong
[00:19:10] S1: So you're you're you're open... You're you're putting a lot of bets, sounds like. You're open to being wrong on certain ones, but you have enough bets where you have confidence that you'll still be relevant and and need [00:19:19] S2: Yes, and you also have focus. So there are things that you're like, wow, like data, for example, that's my big focus. We think that the self improvement aspect comes from the ability to generate data synthetically. And so these are things that we're very sort of... We hold it tight to our chest. [00:19:35] S1: And do you think you need proprietary data to sustain as an AI company? [00:19:40] S2: I think so. I think it's kind of one of data and benchmarks are the definition of model, I think. [00:19:45] S1: Okay. Data and benchmarks. And do you think most people right now are building for the wrong AI stack?
00:19:45 · Ravi Belani, Carina Hong
[00:19:45] S1: Okay. Data and benchmarks. And do you think most people right now are building for the wrong AI stack? [00:19:51] S2: That's a question that we touched on. There's this old word where it's a hard coded harness calling a bunch of specialized models. Yeah. We think that the word is going to be... At least right now, we see it as a de minimis harness. It was an agenda or orchestrator model. Mhmm. It can call itself recursively. It can call, like, other hard coded tools that look more like even the hard coded harness is actor. So that's the word we see. And a lot of people are building infrastructure for the old stack and we're like, we probably should always be two steps ahead because look, even when you lose on the capability layer, you earn the right to play as infrastructure layer. But if you're not like two or three steps ahead, then you're kind of in a zombie position. [00:20:36] S1: Okay. So you're in the stratosphere. I just want people to really feel this of rarified group of companies. You raised a lot of capital. You're growing now [00:20:43] S2: New lab. [00:20:44] S1: With high ambitions. [00:20:44] S2: New lab. [00:20:45] S1: New lab. [00:20:46] S2: Yeah. This is what they call it. I think there is a... DD DAS has this like 50 companies, new lab list, and we're definitely one of them.
00:20:46 · Carina Hong, Ravi Belani
[00:20:46] S2: Yeah. This is what they call it. I think there is a... DD DAS has this like 50 companies, new lab list, and we're definitely one of them. [00:20:51] S1: You're one of the new lab companies. So for that species of companies, can you give color on the nature of the problems that you have to deal with which other startups might not? For example, when you're growing so rapidly, even when you're expanding the team, does that look different than a typical startup in terms of how you're expanding? [00:21:08] S2: Yeah. I think as CEO, a lot of times like are inevitably in people management. And I think for a new lab, this is especially true because like talent density, right? Like that that defines you as business. And if you don't have the talent mode, then I think it's getting into dangerous now where you need to not be in your lab to start like do revenue crunching. So a lot of, I think my day to day are all sorting out the technical management. And this is a shift since the beginning of time. [00:21:42] S1: And this is something that I think is... I wanna make sure that people really grok this because as a as a 24 year old at a Stanford, you may manage to retain and attract talent [00:21:52] S2: Yeah. [00:21:52] S1: That is world class that others have spent, you know, generations trying to cultivate. Can you talk about who... The CTO that you first brought on board, their background, and how you've managed or any hacks that you've managed to get exceptional talent to join you.
00:22:06 · Carina Hong
[00:22:06] S2: Yeah. So Shubo is a he's a former formerly at Facebook AI Research. He's an IC director there. He's this visionary thinker that I always enjoy just hearing him talk about anything. And I think at the beginning it was a time where also Facebook was kind of crumbling. So it was sort of, I think, opportunistic timing. But to align each person to fit into this team that we have cultivated here at Axiom, sometimes I had to be like understanding of what motivates them, especially it's probably the ideal case if what motivates them is the technical problem that you're working on. And in that case, he has worked on testing AI generated unit tests. He believes in scaling, he believes in like just infinite data and big scale. He was one of the authors, Deep Speech and Deep Voice in 2014. So it's very aligned. And that made, I think, a really good foundation. Of course, we have our like ups and downs during like, the normal course of startup when stress is high, but very proactive, very sort of real communication too. I think later, like when you talk about retention, that is I think what I learned. I think the most difficult thing honestly, as a founder is that there are a lot of days where you just kind of are like, everything is going down. You're like, you wanna take care of yourself, but you also have the team out there and you really want to make sure that people are doing well. And I think to the ability to feel empathy and balance that is something that not necessarily in my academia training was like the emphasis, right? It's always about higher, faster, stronger. Like how can you get that higher score in this like Olympia? Necessarily about like just connecting with people, heart to heart. And I think that's very important in the entrepreneurship journey. So that's one of the unique challenges of I think a new lab. The other aspect of a new lab is you're basically dealing with, I mean technical management, not just people management. So there are a lot of priorities that might be conflicting. A constant dynamic I'm dealing with is you can ambitiously set goals, but you cannot afford to also overlook the risk. So it turns out that the people who set up like very ambitious goals and the people who are best at like catching the risk in execution, very, very few times there's like overlap.