Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
OpenAI just SOLVED MATH....
TL;DR
OpenAI appears to have a real math first this time — Wes says an unreleased OpenAI general reasoning model produced a publishable, novel result on Paul Erdős’s 1946 unit-distance problem, and mathematicians including Melanie Matchett Wood and Tim Gowers reportedly verified it as genuine.
The breakthrough wasn’t “alien math,” but a cross-field jump humans missed — the model seems to have connected discrete geometry with algebraic number theory, using a higher-dimensional lattice projected down into 2D “like a shadow sculpture,” rather than inventing fundamentally new mathematics.
The most unsettling quote is about us, not the model — Harvard’s Melanie Matchett Wood said that if the same experts who checked the proof had been assembled earlier and given the same amount of time, they likely could have found the counterexample themselves.
This matters because it was a general LLM, not a bespoke math system — citing OpenAI’s Noam Brown, Wes emphasizes that the model was a general-purpose reasoning model, not specially targeted at math, not wrapped in elaborate scaffolding, and may have essentially one-shotted the key idea.
The story lands harder because OpenAI was burned on math just 7 months ago — after an earlier embarrassing claim turned out to be non-novel, OpenAI reportedly sent this result to nine top mathematicians, including Noga Alon, Thomas Bloom, Daniel Litt, Victor Wang, and Melanie Matchett Wood, before going public.
Wes’s bigger takeaway is that AI’s edge may be stitching disciplines together at scale — he connects this result to Google DeepMind’s AlphaGeometry, AlphaProof, Gemini Olympiad results, Isomorphic Labs, and AlphaEvolve, arguing that cross-domain reasoning may unlock discoveries in biology, materials science, medicine, and even AI research itself.
The Breakdown
From embarrassment to a real Erdős moment
Wes opens with the contrast: seven months ago OpenAI got dragged for claiming a model solved a never-before-solved problem, only for that proof to already exist. This time, he says, the response is the opposite — critics and mathematicians who were skeptical before are now signing off, calling this a legitimate, novel result tied to a central conjecture in discrete geometry.
The deceptively simple dot problem
He walks through Paul Erdős’s 1946 planar unit distance problem in plain English: place dots so as many pairs as possible are exactly one unit apart. A line gives you eight connections with nine dots, a 3x3 grid gives you 12, and the longstanding intuition was that clever grid-like layouts were basically optimal as the number of dots grows.
The “shadow sculpture” analogy for how the model won
Instead of beating the grid with some easy-to-draw 2D trick, Wes says the model found an infinite family of better layouts by building something like a higher-dimensional lattice and projecting it down into 2D. His analogy is those messy sculptures that look meaningless until light hits them at the right angle and the shadow suddenly becomes a clean image — the higher-dimensional structure is the mess, the 2D arrangement is the shadow.
The sponsor detour: Recall as an AI memory layer
Mid-video, Wes pivots into a long ad for Recall 2.0, describing it as an AI encyclopedia that stores YouTube videos, PDFs, X posts, and articles in a knowledge graph. The pitch is that its “agentic chat” can combine your saved material with the web, switch between GPT, Claude, and Gemini, and surface exact timestamps from videos; he offers code WEST25 for 25% off until June 1, 2026.
The line from Harvard that really stuck with him
Back on the main story, Wes highlights Melanie Matchett Wood’s comment: if the experts who later verified the proof had simply been assembled earlier and asked to find a counterexample, they probably could have done it in about the same time. For Wes, that’s the eerie part — not that AI invented incomprehensible mathematics, but that it exposed how fragmented human expertise can be.
Geometry people, number theory tools
His core interpretation is that geometers attacked the problem with geometric intuition, while algebraic number theorists already had tools living in a different part of math. He fumbles through imaginary numbers and complex coordinates with plenty of disclaimers, but the point is clear: the model bridged two human knowledge silos that weren’t naturally talking to each other.
Why this is bigger than one proof
Wes leans on Sebastian Bubeck and Noam Brown here: the model didn’t invent a brand-new branch of math, it performed like an amazing mathematician and did it with a general-purpose LLM. Brown says it wasn’t targeted at this problem, wasn’t a heavy scaffolded system, and the companion paper suggests the core proof emerged in one shot before humans refined the exposition with Codex.
General reasoning models are starting to look like research tools
He closes by widening the lens: Google’s earlier AlphaProof and AlphaGeometry were specialized, but now general models like Gemini and OpenAI’s reasoning systems are getting medals, finding proofs, and doing novel research. That, for Wes, is the headline — not “AI solved math,” but that broadly capable models are becoming engines for cross-disciplinary discovery, with humans still needed to judge what matters and push the best ideas forward.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
