OpenAIHow a reasoning model cracked an 80-year-old math problem — the OpenAI Podcast Ep. 20
At a glance
WHAT IT’S REALLY ABOUT
OpenAI reasoning model disproves Erdős conjecture, showing discovery potential
- The team reports that an OpenAI reasoning model generated a credible disproof of Erdős’s unit distance conjecture, a long-standing central problem in combinatorial/discrete geometry.
- They emphasize that the breakthrough came from a general-purpose model using more test-time compute (letting it “think longer”) rather than a narrow math-only system trained to a benchmark.
- Verification relied on multiple layers of checking—model self-checking followed by several days of review by mathematically trained colleagues—reflecting how “too good to be true” results are stress-tested.
- The proof’s key novelty, as described, was bridging distant fields (notably class field theory/number theory with combinatorial geometry), which they interpret as evidence of genuine creative synthesis.
- The speakers argue this should feel empowering to mathematicians and researchers, shifting humans toward digesting, generalizing, and building new theory while models handle more of the exploratory and computational workload.
IDEAS WORTH REMEMBERING
5 ideasTest-time compute is a major driver of reasoning gains.
They describe a clear pattern: giving the model more time/compute increases correctness, with reported performance approaching ~50% success on the hard unit-distance task under large budgets.
General-purpose capability can yield frontier math results without benchmark-specific training.
The team frames the result as emerging from a broadly capable model (Codex-like tool use, web lookup, Python execution) rather than a system narrowly tuned for one competition or theorem-proving setup.
The breakthrough relied on cross-domain synthesis, not just grinding search.
At a high level, they attribute novelty to connecting class field theory/number theory techniques to combinatorial geometry—an unusual bridge that still required delicate execution.
Validation requires human-style skepticism and multi-stage review.
Their process was: ask the model to check itself, then enlist internal mathematicians to attempt to find errors over several days until confidence rose from “no way” to “likely correct.”
AI outputs can seed rapid follow-on progress by humans.
They claim mathematicians used the construction’s intuition to improve bounds and to tackle other problems (they mention a follow-on disproof of a sum-product conjecture variant over the reals).
WORDS WORTH SAVING
5 quotesMaybe this is the one in 100 times where it's too good to be true, but it's, it's actually true.
— Alexander Wei
And at that moment you feel like, okay, this model is something that's really amazing.
— Hongxun Wu
Everyone had a hard time sleeping because it's so, so exciting, yeah.
— Lijie Chen
I think it should not be intimidating. I just think it should be empowering instead.
— Hongxun Wu
Get GPT Pro subscription.
— Hongxun Wu
High quality AI-generated summary created from speaker-labeled transcript.
