Welcome, humans.

A 24-year-old just raised $64M to build an AI mathematician that's smarter than Terence Tao (widely considered the world's greatest living mathematician).

Her name is Carina Hong, and her startup Axiom Math has already solved a 130-year-old problem and disproved a 30-year-old conjecture.

Don’t get it twisted, though; this isn't about making ChatGPT better at algebra. This is about creating AI that discovers entirely new mathematical theorems, proves them formally, and gets smarter with each iteration.

We’re talking about the kind of math that eventually unlocks breakthroughs in chip design, aircraft safety, quantum computing, and pretty much every scientific field you can think of.

In our latest podcast episode, we sit down with Carina to talk about her plan to build a a superintelligent mathematician, how brain science reveals surprising mathematical patterns hiding in nature, and why she believes AI for math is the “algorithmic pillar” that unlocks everything from chip design to quant trading.

Here are some of the most fascinating moments:

• (1:56) The Nobel Prize-winning discovery that rat brain cells fire in perfect hexagonal patterns, and why Carina says “it almost proves God exists“

• (6:57) Why neuroscience hasn't yet guided AI architecture.

• (8:35) What “mathematical superintelligence“ actually means.

• (9:27) What separates Carina from a legendary mathematician like Fields Medalist Terence Tao (hint: it's exactly what she says AI can fix).

• (10:30) The self-improving loop: an AI that generates conjectures, proves them, learns from failures, and gets smarter with each iteration.

• (18:17) Why math is the “bedrock“ that transfers to physics, coding, finance, and even tax law… but not the other way around.

• (31:08) How Axiom solved a 130-year-old problem about Lyapunov functions that stumped Poincaré, Newton, and Lagrange.

• (33:57) The brilliant trick for preventing their AI from generating millions of useless theorems.

• (35:26) How AI detects novelty by finding proofs that bridge “two previously unconnected branches: algebra and combinatorics.“

• (42:00) “Who checks the checker?“ How formal verification means a 3-line statement can generate a proof that never needs human review

• (44:03) The massive opportunity: using math AI to verify legacy code and AI-generated code at scale.

• (48:17) Carina's vision for a “reasoning IDE“ where quant traders and engineers get Terence Tao-level math at their fingertips.

Our favorite part: Carina’s take on why AI might finally break down the silos between scientific fields (35:46):

• Carina explains that “machine-assisted mathematics actually promotes a diffusion of ideas between different fields”, UNLIKE humans, who are bounded by academic specializations or the even more narrow “conference topic.”

• She shares a personal example of sneaking into a continual learning workshop next door to her combinatorics conference in Germany, noting that this kind of cross-pollination “doesn't happen unless you literally put two workshops next to each other in the same castle in the rural area in Germany.”

• AI doesn't need adjacent castles… it can see connections across all of math simultaneously, which might be where new math proofs and all kinds of novel discoveries come from next; breaking down the old walls of our narrow human focus areas to find new connections that remix them all together…

Watch and/or Listen now: YouTube | Spotify | Apple Podcasts

P.S. Carina explains why quant trading firms dream of having “AI Terence Tao“ at their fingertips… and why that matters even if you're not a mathematician.

Additional Resources:

• Learn more about Axiom here.

• Carina Hong interviewed on EO Studio.

• Axiom will be giving a few talks at the Joint Mathematics Meetings in January.

• Axiom is also organizing an Aarhus Mathematics & AI Workshop in January.

Also, here’s a great companion piece to this episode: Dwarkesh’s new interview with Adam Marblestone.

Adam Marblestone (CEO of Convergent Research, former Research Scientist at Google DeepMind) has a fascinating answer to the question Carina touched on early in our conversation: How does the brain do it?

We're throwing way more data at LLMs (large language model AI like ChatGPT and Claude) than humans ever see, yet humans still have capabilities these models lack. Adam's answer: the brain's secret sauce isn't the architecture—it's the reward functions.

Here's the TL;DR on his point: 

• Evolution encoded incredibly specific loss functions (basically the “grading rubric” that tells a model what counts as success or failure during training). that guide learning

• These aren't simple computer science loss functions like “predict the next token” (in case you forgot, tokens = word chunks the AI processes).

• Instead they're complex, curriculum-driven reward signals that turn on at different developmental stages for different brain regions.

• Think of it like thousands of lines of Python code specifying exactly what different parts of your brain should learn and when.

The genius part? These reward functions are compact enough to fit in the genome (a.k.a your DNA blueprint; Adam notes the entire human genome is just 3 gigabytes, with only a fraction dedicated to the brain).

Evolution solved the sample efficiency problem (learning more from less data) by building better “teachers“ into into what Adam calls the “Steering Subsystem,” (the instinct-driven part of the brain that tells the learning part what matters) not by making the learning system (the cortex, a.k.a the brain's learning and reasoning layer) bigger.

Here's where it gets interesting for Carina's work: Adam is on the board of Lean; the exact same formal proof language that Axiom Math uses. He argues that formal math proving is becoming “a perfect RLVR task” (RLVR = reinforcement learning w/ “verifiable rewards”, where the AI gets automatic right/wrong feedback, like a compiler checking code) and predicts we'll be able to “RLVR the crap out of proofs” in the coming years.

He also sees formal verification (mathematically proving code can't fail in certain ways) extending to provably secure software and hardware; exactly what Carina mentioned around (44:03) when she said “the model that can do formal verification on math can formally verify code.”

Adam also speculates that the cortex might be doing “omnidirectional inference”, or predicting any variable from any other variable, not just next-token prediction. In other words, instead of just predicting “what comes next,” the brain might be able to fill in any missing piece of information given any other pieces, like solving a puzzle where you can start from any corner.

This kind of flexible reasoning is exactly what you'd need for the mathematical discovery Carina describes, where AI can bridge “two previously unconnected branches: algebra and combinatorics.”

Both interviews explore the same fundamental question from different angles: What's actually missing in current AI? Carina's building it from the math side up. Adam's mapping it from the brain side down. And they're both betting on Lean.

ICYMI: This episode on AI and chemistry blew our minds a little bit…

We get it, it’s the holidays, but this interview with Nick Talken did NOT get nearly enough love on YouTube, and it’s one of our favorites from this year.

Nick's company, Albert Invent, is helping Fortune 100 companies like Kenvue (the $32B company behind Tylenol, Neutrogena, and Listerine) compress R&D projects from 3 months down to 2 days.

Here's the wild part: there are more possible chemical combinations than atoms in the universe. If you have 100 ingredients and need to pick just 10 for a new formulation, that's 17 trillion possible combinations… before you even factor in the ratios.

Put simply, ChatGPT can't solve this. You need domain-specific foundational models trained on 15M molecular structures, plus enterprise data that's never been digitized (most chemical companies still use paper notebooks).

Here's some of our favorite parts:

• (4:10) Before Albert vs After: How scientists go from 3-month projects to 2-day turnarounds.

• (13:04) Why running hundreds of thousands of strategic simulations beats brute-force experimentation.

• (16:16) Why simulations will never fully replace physical experiments (and why that's actually good).

• (22:08) Nick's take on whether AI will make “meaningful scientific breakthroughs“ by 2028.

• (28:18) How a $32B company is deploying Albert to every single lab, and why every product will soon be “touched by AI.“

• (37:13) Nick's vision: we should be able to “invent the physical world with a laptop“ (just like AWS did for software). We’re V bullish on this idea, y’all!

Why this matters: Chemistry is the foundation of the physical world. Everything from the guitar strings to the shirt you're wearing required material science R&D. Battery chemistry is the bottleneck for AR glasses, electric vehicles, and Mars missions. If we can speed up chemistry 50x, we accelerate everything else.

Listen and/or Watch on Spotify or Apple Podcasts.

IN CASE YOU MISSED IT… Check out our other recent favorite episodes below!

Three other recent interviews you’ll definitely want to check out (pick whatever looks interesting to you and dive in!):

• Your AI Meeting Agent Sends Your Digital Clone to Meetings (Spotify, Apple Podcasts): Otter.ai CEO Sam Liang reveals they're building AI avatars that attend meetings on your behalf, asking questions and even answering for you.

  • Plus: their real-time sales coach that feeds reps answers during live customer calls (basically “cheating in an exam”), and why one Fortune 500 company found that for every 20 Otter seats, they save the equivalent of one full-time employee.

• The Future of Windows in an AI World: Microsoft's Pavan Davuluri explains the new “Copilot Highlights“ feature that literally draws on your screen to show you which buttons to click.

  • We also dig into what AI runs locally vs. cloud, how their new MCP agent registry works (and why it's like an app store for AI connectors), and the vision for AI that helps your grandparents actually use their computer.

• Securing AI (with AI): Microsoft Security CVP Vasu Jakkal walks through the 6 new AI agents in Security Copilot that automate phishing triage and vulnerability prioritization.

  • Plus: how Microsoft is extending Zero Trust principles to the emerging “agentic workforce” of AI systems. Essential viewing if you're deploying AI agents (especially w/ Microsoft Copilot) in your org.

• Guru: The AI Source of Truth Demo: A hands-on walkthrough of how to solve the “knowledge is everywhere, context is nowhere” problem. The killer feature: when a clinician flags outdated info during a call, managers get notified, Guru drafts the fix, and one update pushes everywhere instantly—chat, Slack, CRM, internal apps. Finally, a solution to the “garbage in, garbage out” problem.

Last thing: if you haven’t subscribed yet, please do! Click the image below to go to our channel and hit “subscribe” to get notified right when new videos go live.

Stay curious,

The Neuron Team

P.P.S: Love the newsletter, but don’t want to receive these podcast announcement emails? Don’t unsubscribe — adjust your preferences to opt out of them here instead.