This 24-Year-Old Founder Raised $64M to Build World’s First AI Mathematician | Axiom, Carina Hong

1. 0:00 – 2:17

   Intro

   1. SPSpeaker0:00

      Math is full of hardship. Math research is a process of almost a monk praying in the temple day after day. Keep pushing harder and harder. The dopamine hit feeling is a little bit like clean scandal. It feels natural, effortless. You can look at how far you have come. In the competition, so usually the clock is ticking. [clock ticking] I think that feeling is just really amazing. It's exhilarating. If you have a hedge fund, being able to afford the quant researchers, typically paid at 14 million each year starting, but that AI mathematician, um, at $5 each hour. And then you are able to suddenly afford to tackle a market of total trading volume of, say, 8 million, because that market likely has not been deeply studied, and previously it was not considered a target. We are entering an era of mass intelligence, and an AI mathematician is a crucial part of this future. [upbeat music] Hi, I'm Carina, founder and CEO of Axiom. I was a mathematician for most of my life. I did math and physics double major at MIT, where I plunged into the ocean of mathematics, worked on very interesting research projects that settled some open conjectures. And then I was a Rhodes Scholar at Oxford, pursuing degree of neuroscience, and that year I also stumbled into AI at the UCL Gatsby Institute, which is basically the home of DeepMind. After that, I was at Stanford as a Knight-Hennessy scholar, pursuing a joint JD and PhD, then dropped out to start Axiom. And we are building an AI mathematician that is the first model that will eventually evolve to be a self-improving super intelligent reasoner. To build an AI mathematician, you need three pillars: AI, programming languages, and mathematics. Because of this vision, we assembled a world-class team of experts from each of these three pillars coming together. So it's a very interdisciplinary approach. Axiom was fortunate to raise seed round of $64M. Investors value us at $300M valuation. I'm 24 years old. [upbeat music] My background is in combinatorics

2. 2:17 – 4:41

   Why Math Will Save the World

   1. SPSpeaker2:17

      and number theory. It's about patterns, patterns everywhere. In sets, in numbers, there are correspondences that are unexpected. I love just everything number theorist. Uh, obviously Gauss. Reading about Gauss's life was also very inspiring. All these aha moments, all the long nights when he was pushing for a eureka moment, um, just fantastic. I think it's, um, the crown jewel of mathematics, um, as Gauss put it. It's quite beautiful and it feels seamless compared to other fields of math. So it's a lot of symbolic expressions. The symbols, they actually look beautiful on the sketch paper. I, I just remember I wanted to be a number theorist. In the history, every mathematical tool after its invention has led to terrific breakthroughs, not just in fundamental science, but also in real-world application. For example, the invention of abacus, that has led to the bloom of trade and commerce. Think about integrals and calculus. That led to mechanics and thermodynamics. The rest is history. It's the Industrial Revolution. If you think about Babbage engine or the difference engine, that is a math tool to calculate log table faster. That is the prototype of computer. So every mathematical tool kind of sparks the flywheel in real-world applications and in turn requires more computational tools. So there's this theory of Jevons paradox, which is when the price of a tool becomes elastic, you will then have unexpected use cases and applications, hence requiring more tools and making these tools in turn more valuable. By building an AI Gauss at your fingertip, we think there will be so many magnitudes of use cases and markets being unlocked. And pragmatically, if you think about the time spent of a math invention to, say, the real-world application, that has in fact spanned centuries. AI compressed this timeline. If you think about AI mathematicians working together with applied scientists, which by the way, human mathematicians seldom work with applied scientists. Now, AI mathematicians can go on to these applied fields and solve the complex system that has never been theoretically understood, and we think that's incredibly powerful and will shorten the century-long time span to, to much shorter. [upbeat music]

3. 4:41 – 7:33

   Problem Solver to Theory Builder

   1. SPSpeaker4:41

      I grew up loving math. Every time you solve a math problem, you will get that instant reinforcement of, um, this is the thing that you love doing. You want to continue doing it. You get this little dopamine hit every time you solve an Olympic Math question. I think it's in elementary school, at fourth grade. It was about 1,000 really bright kids being assembled into 24 classrooms and try to compete, um, after each exam. It was a little bit stressful just because, you know, after each exam you are being sort of ranked, and then it's a whole system and only the ones that perform outstandingly can go to, um, the next level. But also it was just eye-opening. I was able to read the proof of quadratic reciprocity. I was able to look at the names of the mathematicians that I've never known or never heard of, and there are some French and German mathematicians. So you kind of wonder, you know, what it's like to visit the French and German research institute. It feels like intellectually exploring the world, and that felt really motivating to try harder, do more exercises, and to do really well in the exams. I remember the elementary school Math Olympic problems are made to be quite fun and interesting, formulated actually in the real-world scenario, right? Like, two trucks going across each other and, um, how long each truck is and how long does it take for them to pass each other. Convert that to mathematical formulation to equations that feels trackable and you can solve without paying much attention to what the real-world problem is. I think that process was quite magical to me, right? And then you solve it, and you plug that back in. What does that tell you in the real-world scenario? I think that's very beautiful process, and it's mathematical thinking that is very transferable to other fields. At one point, I do think that I got introduced to research math, and that was eye-opening. I think that research math is a lot more delayed gratification. Research problem is really hard, and it takes you a long time to figure it out. You don't have that dopamine hits anymore. I mean, if you're a Math Olympiad student, you can solve a dozen problems each day and feel good about yourself. A dozen months has passed, you have done nothing. That's usually the, uh, status of research math if you encounter a bottleneck in the problem. So I wanted to be a better mathematician, um, kind of changing from a Problem Solver to a Theory Builder. And for that, I owe, um, to a lot to my mentors that teach me how to do research, how to be patient, how to look around the corners and look for unexpected connections, um, from another field to the current problem. That was why I got into research and continued, um, the fruitful collaboration with, um, Professor Ono and other collaborators. The feeling that you can

4. 7:33 – 8:38

   Why Taste is Important in the AI Era

   1. SPSpeaker7:33

      develop new mathematical theories based on how the flow of past theories go is quite fascinating. You will invent definitions in the natural way. You will link these definitions together to formulate interesting conjectures and, you know, what does, um, natural mean? What does interesting mean? And then after you formulate that conjecture, you prove it in an elegant way. What does elegance mean? These are questions about taste, and I felt like by learning so many, um, branches of math, I started to form my taste about math. And I think in an era where AI is prevalent and can do a lot, taste becomes quite important. It distinguishes between a good scientist and a mediocre one. I think we want to try to understand the problem of taste and intuition better, um, using modern-day machine learning techniques, and that, of course, will be a very difficult technical challenge, but it's something that our team is incredibly excited about.

5. 8:38 – 11:16

   The Hardest Problems Are the Strategy

   1. SPSpeaker8:40

      At the Ross, um, Mathematics Program, actually, when I was 15, it was a beautiful summer where the professors will teach us how to think deeply about simple things. I remember first days are like, "Can you prove that zero multiplied by everything is zero?" I'm like, "Well, this is obvious. Like, why would you want me to prove that?" In fact, you are asked to derive that statement strictly from a set of five axioms, such as zero plus everything is that thing itself, one multiplied by everything is that thing itself, et cetera. That way of axiomatic and deductive reasoning really was inspiring to me. Thought that was just, first of all, a bit crazy that a bunch of us stuck for more than 24 hours on, um, that simple problem set, but also it teaches me reason strictly, rigorously, um, to get to the final destination. Axiom, the name of our company, actually comes with this inspiration. We want to build out the knowledge graph and expand the frontier of mathematics through deductive logic, and we are using the programming language of proofs, Lean, for that. I think that there are so many components, just like math, right? One think worry deeply about simple things. A very short, concise mission statement, AI Mathematician, requires various techniques joining together. For example, my colleague Hugh Lather and his team have been working on applying deep learning to code generation since very early, since, like 2017. Francois Chollet have been pioneering the field of AI for mass discovery using Transformer to solve symbolic integration in 2019, and they show that it work better than computer algebra. CTO Subhas Sengupta have worked in FAIR for many years at Facebook AI Research, pioneered work such as large reinforcement learning models like OpenGo. We believe in solving the hardest problem is the best way to win, and I think this mission in itself is incredibly attractive to talents. I love the fast-paced environment of startups. I love executing. I love executing with the team. I love unblocking others. I love asking for help. You have the sort of very instant reward signal. It's similar to the childhood Carina trying to solve the m- math problems. Just being in the moment, being immersed in how research engineering is done on day-to-day basis is just a once-in-a-lifetime experience.

6. 11:16 – 14:36

   Math is the Sandbox of Reality

   1. SPSpeaker11:16

      Maths really is the fundamental of lots of branches of sciences, and it's also the sandbox for reality where you can try to put a lot of the real world objects into mathematical variables and then formulate the problem in a purely theoretical way. So I thought understanding math will allow me to generalize to other domains, and indeed, in machine learning, we found the transferability of math reasoning to be quite striking. So a model that does really well on math can likely do better in coding. We find this sort of surprising phenomena of math concepts being unexpectedly present in a lot of the applied scientific fields. And if you have a real world problem, you can try to understand that theoretically through solving it mathematically. I think by math is the sandbox of reality, there's another meaning, which is if you think about modern machine learning models, they want to gather real world data for them to try to experiment, generate new knowledge from spatial reasoning data. Math is a digital version of that. We can solve reasoning by being in the digital world without relying on the real world data. So yeah, it is the playground for trying out what we understand about, about reality and the universe. Math is full of hardships. Like, I think, like, math research is a process of almost like a monk praying in the temple day after day. You just hope that the stone that you are looking at will have a flower grow out of it. I think the feeling of being stuck can feel quite depressing just because you don't really know where your work and the math ends and where, you know, your life begins. You kind of blur your identity, um, with your work. I think that's a struggle a lot of mathematicians do face. Think that with AI, um, hopefully AI mathematicians will make the process a lot more enjoyable. The ideal state is obviously one has a spec, a lemma that one wants to prove, and instead of the human banging the head on the wall, um, for days and sometimes weeks, months, the AI mathematician can help prove it, and the human mathematician will just guide the collaboration to the next lemma, makes the journey a lot more exhilarating. I do think that's part of the vision. More like a collaborative end game that we are imagining rather than, say, replacing humans. Five or 10 years is hard to imagine, but, uh, we fully believe that this is a fundamental technology that will turn us into a generational company. [instrumental music]

Episode duration: 14:36