Vibeleaderboard
← All Intel
Intel / video

Inside OpenAI’s Breakthroughs in Mathematical Reasoning

Source
youtube.com
Author
a16z
Date

Context

Two mathematicians now at OpenAI say AI models help most when a promising approach exists but is too fiddly for a person to risk pursuing. In an a16z conversation, they describe an experience mathematicians know well: you try an idea for hours or weeks, give up, and sometimes learn later that someone else made it work.

They report two strengths. One is literature search: one of them says GPT-5 found a reference within minutes for a problem on the Erdős problems site that friends had spent hours on, and they found about ten more such cases. The other is executing detailed arguments, where the handles small conditions humans lose track of. On the unit distance problem, they say the general approach was familiar to mathematicians, but the model made the right bet on a finicky route, and its released summarised reasoning looks like a mathematician pruning a search.

These are accounts from people who work at the company, in an informal discussion, so they describe how these two view the model's contribution, not a verified assessment of the proofs.

Terms in this piece · Glossary
  • Model — A computational system that uses learned patterns to produce an output from information supplied to it.
Read the source www.youtube.com
More from a16z
Recommended reads
Comments

Checking sign-in…

Loading comments…