Transcriptly

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

EO14:36Added Aug 31, 2026

If an AI Mathematician can reason and prove on its own, how will the future change? Carina is a mathematician and the Founder and CEO of Axiom. She started the company at 24, and Axiom is building an AI Mathematician with a $64

Watch on YouTube →
Contributed by 刘嘉琪

Transcript

Transcript format

  1. Intro

  2. 00:00Math is full of hardship. Math research
  3. 00:02is a process of almost a monk praying in
  4. 00:06the temple [music] day after day. Keep
  5. 00:08pushing harder and harder. The dopamine
  6. 00:10hits feeling is a little bit like Quinc
  7. 00:12Gambit feels natural, effortless. You
  8. 00:15can look at how far [music] you have
  9. 00:17come in the competition. So usually the
  10. 00:20clock is ticking.
  11. 00:22I think that feeling is just really
  12. 00:24amazing. It's exhilarating. If you have
  13. 00:27a hedge fund being able to afford the
  14. 00:29quant researchers typically paid at 14
  15. 00:32million each year starting but that AI
  16. 00:35mathematician um at $5 each hour and
  17. 00:38then you are able to suddenly afford to
  18. 00:40tackle a market of total trading volume
  19. 00:42of say 8 million because that market
  20. 00:45likely has not been deeply studied and
  21. 00:48previously it was not considered a
  22. 00:50target. We are entering an era of mass
  23. 00:53intelligence and an AI mathematician is
  24. 00:56a crucial part of this future.
  25. 01:07Hi, I'm Karina, founder and CEO of
  26. 01:09Axiom. I was a mathematician for most of
  27. 01:11my life. I did math and physics double
  28. 01:14major at MIT where I plunged into the
  29. 01:17ocean of mathematics. worked on very
  30. 01:19interesting research projects that
  31. 01:21settled some open conjectures and then I
  32. 01:23was a rose scholar at Oxford pursuing
  33. 01:25degree of neuroscience and that year I
  34. 01:27also stumbled into AI at the UCL Gatsby
  35. 01:30[music] institute which is basically the
  36. 01:32home of deep mind after that I was at
  37. 01:34Stanford as a Hennessy scholar pursuing
  38. 01:36a joint JD PhD then dropped out to Dar
  39. 01:40axium and we are building an AI
  40. 01:41mathematician that is the first model
  41. 01:44that will eventually evolve to be a
  42. 01:46self-improving super intelligent
  43. 01:48reasoner To build an AI mathematician,
  44. 01:50you need three pillars. AI, programming
  45. 01:52languages, and mathematics. Because of
  46. 01:54this vision, we assemble a world-class
  47. 01:57team of experts from each of these three
  48. 01:59pillars coming together. So, it's a very
  49. 02:01interdisciplinary approach. Axiom
  50. 02:03[music]
  51. 02:03was fortunate to raise seed round of $64
  52. 02:06million. Investors value us at 300
  53. 02:09million valuation. I'm 24 years old.
  54. 02:15My background is in combinatorics and

  55. Why Math Will Save the World

  56. 02:17number theory. It's about patterns.
  57. 02:19Patterns [music] everywhere in sets and
  58. 02:21numbers. There are correspondences that
  59. 02:24are unexpected. I love just everything
  60. 02:26number theorist. Uh obviously golf.
  61. 02:29[music] Reading about golf's life was
  62. 02:30also very inspiring. All these aha
  63. 02:33moments, all the long nights when he was
  64. 02:36pushing for an Eureka moment. Um just
  65. 02:38fantastic. [music] I think it's the
  66. 02:40crown draw of mathematics. Um, as GS put
  67. 02:43it, it's [music] quite beautiful and it
  68. 02:45feels simless compared to other fields
  69. 02:48of math. [music] So, it's a lot of
  70. 02:50symbolic expressions. The symbols, they
  71. 02:53actually look beautiful on the sketch
  72. 02:55paper. I just remember I want to be a
  73. 02:57number theorist. In the history, every
  74. 03:00mathematical tool after its invention
  75. 03:02has [music] led to terrific
  76. 03:03breakthroughs not just in fundamental
  77. 03:06science but also in real world
  78. 03:07applications. [music]
  79. 03:08For example, the invention of abacus
  80. 03:10that has led to the bloom of trade and
  81. 03:12commerce. Think about integrals and
  82. 03:14calculus that led to mechanics and
  83. 03:16thermodynamics. The rest is history.
  84. 03:18It's the industrial revolution. If you
  85. 03:20think about babbage engine or the
  86. 03:22difference engine that is a math tool to
  87. 03:24calculate log table faster that is the
  88. 03:27prototype of computer. So
  89. 03:29aromathematical tool kind of sparks
  90. 03:31[music] the flywheel in real world
  91. 03:33applications and in turn requires more
  92. 03:36computational tools. So [music] there is
  93. 03:38this theory of Javon's paradox which is
  94. 03:40when the price of a tool becomes elastic
  95. 03:44[music] you will then have unexpected
  96. 03:45use cases and applications hence
  97. 03:47requiring more tools and making these
  98. 03:49tools in turn more valuable by building
  99. 03:52an AI goals at your fingertip. [music]
  100. 03:54We think there will be so many
  101. 03:55magnitudes of use cases and markets
  102. 03:58being [music] unlocked
  103. 04:00and pragmatically if you think about the
  104. 04:02time span of a mass invention to say the
  105. 04:05real world application that has in fact
  106. 04:08spent centuries. AI compress this
  107. 04:10timeline. If you think about AI
  108. 04:12mathematicians working together with
  109. 04:14applied scientists which by the way
  110. 04:17human mathematicians seldom work with
  111. 04:19applied scientists. Now AI
  112. 04:21mathematicians can go on to these
  113. 04:23applied [music] fields and solve the
  114. 04:25complex system that have never been
  115. 04:27theoretically understood. And we think
  116. 04:30that's incredibly powerful and will
  117. 04:31shorten the century long time span to to
  118. 04:36much shorter.

  119. Problem Solver to Theory Builder

  120. 04:41I grew up loving math. Every time you
  121. 04:43solve a math problem, you will get that
  122. 04:45instant reinforcement of um this is the
  123. 04:48thing that you love doing, you want to
  124. 04:49continue doing it, you get this little
  125. 04:51dopamine hit every time you solve an
  126. 04:54Olympic math question. I think it's an
  127. 04:57elementary school at fourth grade. It
  128. 04:59was about 1,000 really bright kids being
  129. 05:02assembled into 24 classrooms and try to
  130. 05:05compete um after each exam. It was a
  131. 05:09little bit stressful just because you
  132. 05:11know after each exam you are being sort
  133. 05:13of ranked and then it's a whole system
  134. 05:15and only the ones that perform
  135. 05:17outstandingly can go to um the next
  136. 05:20level but also it was just eyeopening. I
  137. 05:23was able to read the proof of quadratic
  138. 05:26reciprocity. I was able to look at the
  139. 05:29names of the mathematicians that I've
  140. 05:31never known or never heard of. And there
  141. 05:33are some French and German
  142. 05:35mathematicians. So you kind of wonder,
  143. 05:38you know, what it's like to visit the
  144. 05:40French and German research institute. It
  145. 05:42feels like intellectually exploring the
  146. 05:45world and that felt really motivating to
  147. 05:48try harder, do more exercises and to do
  148. 05:51really well in the exams. I remember the
  149. 05:53elementary school math Olympia problems
  150. 05:55are made to be quite fun and interesting
  151. 05:57formulated actually in a real world
  152. 05:59scenario, right? like two trucks going
  153. 06:02across each other and um how long each
  154. 06:04truck is and how long does it take for
  155. 06:06them to pass each other. Convert that to
  156. 06:08mathematical formulation to equations
  157. 06:11that feels trackable and you can solve
  158. 06:14without paying much attention to what
  159. 06:17the real world problem is. I think that
  160. 06:18process was quite magical to me. Right?
  161. 06:20And then you solve it and you plug that
  162. 06:22back in. What does that tell you in the
  163. 06:25real world scenario? I think that's very
  164. 06:28beautiful process and it's mathematical
  165. 06:30thinking that it's very transferable to
  166. 06:32other fields.
  167. 06:35At one point I do think that I got
  168. 06:37introduced to research math and that was
  169. 06:39eye opening. I think that research math
  170. 06:42is a lot more delay gratification.
  171. 06:44Research problems is really hard and it
  172. 06:46takes you a long time to figure it out.
  173. 06:49You don't have that dopamine hits
  174. 06:50anymore. I mean, if you're a mass
  175. 06:51Olympia student, you can solve a dozen
  176. 06:54problems each day and feel good about
  177. 06:55yourself. A dozen months has passed, you
  178. 06:57have done nothing. That's usually the
  179. 06:59status of research math if you encounter
  180. 07:02a bottleneck in the problem. So, I
  181. 07:04wanted to be a better mathematician. Um,
  182. 07:06kind of changing from a problem solver
  183. 07:08to a theory builder. And for that, I owe
  184. 07:11um to a lot to my mentors that teach me
  185. 07:14how to do research, how to be patient,
  186. 07:16how to look around the corners and look
  187. 07:18for unexpected connections. um from
  188. 07:20another field to the current problem.
  189. 07:22That was why I got into research and
  190. 07:24continued um the fruitful collaboration
  191. 07:27with um professor Ono and other
  192. 07:29collaborators. [music]
  193. 07:30The feeling that you can develop new

  194. Why Taste is Important in the AI Era

  195. 07:33mathematical theories [music]
  196. 07:35based on how the flow of past theories
  197. 07:38go is quite fascinating. You will
  198. 07:41[music] invent definitions in the
  199. 07:43natural way. You will link these
  200. 07:45definitions together to formulate
  201. 07:47[music] interesting conjectures. And you
  202. 07:49know what does um natural mean? What
  203. 07:51does interesting mean? And then after
  204. 07:53you formulate that conjecture, you prove
  205. 07:55it in an elegant way. What does elegance
  206. 07:57mean? These are questions about taste.
  207. 07:59And I felt like by learning so many
  208. 08:02[music]
  209. 08:02um branches of math, I started to form
  210. 08:05my taste about math. And I think in an
  211. 08:09era where AI is prevalent and can do a
  212. 08:12lot, taste becomes quite important. It
  213. 08:15distinguishes between a good scientist
  214. 08:17and a mediocre one. I [music] think we
  215. 08:19want to try to understand the problem of
  216. 08:22taste and intuition [music]
  217. 08:24better um using modern-day machine
  218. 08:26learning techniques and that of course
  219. 08:28will be a very difficult [music]
  220. 08:30technical challenge. But it's something
  221. 08:32that our team is incredibly excited
  222. 08:34about.

  223. The Hardest Problems Are the Strategy

  224. 08:40at the Ross uh mathematics program
  225. 08:42actually when I was 15. It was a
  226. 08:44beautiful summer where the professors
  227. 08:46will teach us how to think deeply about
  228. 08:49simple things. I remember first day
  229. 08:51they're like can you prove that zero
  230. 08:53multiplied by everything is zero. I'm
  231. 08:54like well this is obvious like why would
  232. 08:56you want me to prove that? In fact you
  233. 08:58are asked to derive that statement
  234. 09:01strictly from a set of five axioms such
  235. 09:04as zero plus everything is that thing
  236. 09:06itself. one multiplied by everything is
  237. 09:08that thing itself etc. That way of
  238. 09:11acimatic and deductive reasoning really
  239. 09:14was inspiring to me. Thought that was
  240. 09:17just first of all a bit crazy that a
  241. 09:19bunch of us stuck for more than 24 hours
  242. 09:21on um that simple problem set but also
  243. 09:24it teaches me reason strictly rigorously
  244. 09:28um to get to the final destination.
  245. 09:31Axium the name of our company actually
  246. 09:33comes with this inspiration. We want to
  247. 09:36build out the knowledge graph and expand
  248. 09:38the frontier of mathematics through
  249. 09:40deductive logic and we are using the
  250. 09:43programming language of proofs lean for
  251. 09:46that. I think that there are so many
  252. 09:49components [music] just like math right
  253. 09:51one think very deeply about simple
  254. 09:53things a very short concise mission
  255. 09:57statement AI mathematician requires
  256. 10:00various techniques joining together for
  257. 10:02example my colleague Hugh Leather and
  258. 10:05his team have been working on applying
  259. 10:07deep learning to code generation since
  260. 10:10very early since like 2017
  261. 10:13chart have been pioneering the field of
  262. 10:15AI for mass discovery using transformer
  263. 10:18to solve symbolic integration in 2019
  264. 10:21and they show that it work better than
  265. 10:23computer algebra. CTO Shoen Gupta have
  266. 10:27worked in fair for many years and
  267. 10:29Facebook AI research pioneered work such
  268. 10:32as large reinforcement learning models
  269. 10:34like open go. We believe in solving the
  270. 10:37hardest problem is the best way to win
  271. 10:40and I think this mission [music] in
  272. 10:42itself is incredibly attractive to
  273. 10:45talents. I love the fast-paced
  274. 10:47environment of startups. I love
  275. 10:49executing. I love executing with the
  276. 10:51team. I love unblocking others. I love
  277. 10:54asking for help. You have the sort of
  278. 10:56very instant reward signal. It's similar
  279. 10:59to the childhood Karina trying to solve
  280. 11:02the ma math problems. Just being in the
  281. 11:04moment, being immersed in how research
  282. 11:07engineering is done on day-to-day basis
  283. 11:09is just a once in a-lifetime experience.

  284. Math is the Sandbox of Reality

  285. 11:16Mass really is the fundamental of lots
  286. 11:19of branches of sciences [music]
  287. 11:21and it's also the sandbox for reality
  288. 11:23where you can try to put a lot of the
  289. 11:26real world objects into mathematical
  290. 11:28variables and then formulate the problem
  291. 11:31in a purely theoretical way. So I
  292. 11:33thought understanding math will allow me
  293. 11:35to generalize to other domains. And
  294. 11:38indeed in machine learning we found the
  295. 11:40transferability of mass reasoning to be
  296. 11:42quite striking. So a model that does
  297. 11:45really well on math can likely do better
  298. 11:47in coding. We find this sort of
  299. 11:50surprising phenomena of math concepts
  300. 11:53[music]
  301. 11:53being unexpectedly
  302. 11:56present in a lot of the applied
  303. 11:58scientific [music] fields. And if you
  304. 12:01have a real world problem, you can try
  305. 12:03to understand that theoretically through
  306. 12:06solving it mathematically. I think by
  307. 12:08math is the sandbox of reality. There's
  308. 12:11another [music] meaning which is if you
  309. 12:13think about modern machine learning
  310. 12:15models, they want to gather real world
  311. 12:19data for them to try to experiment,
  312. 12:22generate new knowledge from spatial
  313. 12:25reasoning data. Math is a digital
  314. 12:28version of that. We can solve reasoning
  315. 12:32by being in the digital world without
  316. 12:35relying on the real world [music] data.
  317. 12:37So yeah, it is the playground for trying
  318. 12:40out what we understand about about
  319. 12:43reality and the universe. Math is full
  320. 12:46of hardships. Like I think like mass
  321. 12:48research is a process of almost like a
  322. 12:52monk praying in the temple day after
  323. 12:56day. You just hope that the storm that
  324. 12:58you are looking at will have a flower
  325. 13:00grow out of it. I think the feeling of
  326. 13:03being stuck [music] can feel quite
  327. 13:05depressing just because you don't really
  328. 13:07know where [music] your work and the
  329. 13:09math ends and where you know your life
  330. 13:11begins. You kind of blur [snorts] your
  331. 13:13identity [music] um with your work. I
  332. 13:16think that's a struggle a lot of
  333. 13:17mathematicians do face. I think that
  334. 13:20with AI um hopefully AI mathematicians
  335. 13:23will make the process a lot more
  336. 13:25enjoyable. The ideal state is obviously
  337. 13:27one has a spec a lema that one wants to
  338. 13:29prove [music] and instead of the human
  339. 13:32banging the head on the wall um [music]
  340. 13:34for days and sometimes weeks months the
  341. 13:37AI mathematician can help prove it and
  342. 13:41the human mathematician will just guide
  343. 13:43the collaboration to the next lema makes
  344. 13:46the journey a lot more exhilarating. I
  345. 13:48do think that's part of the vision more
  346. 13:50like a collaborative endgame that we are
  347. 13:53imagining rather than say replacing
  348. 13:56humans. Five or 10 years is hard to
  349. 13:58imagine but uh we fully believe that
  350. 14:00this is a fundamental technology that
  351. 14:03will turn us into a generational
  352. 14:05company.
  353. 14:13Thank
  354. 14:14you. Thank you.
This 24-Year-Old Founder Raised $64M to Build World’s First AI Mathematician | Axiom, Carina Hong — Transcriptly