16 March 2026

78 new episodes (53 hours listening) published by the 38 podcasts on the list. 1 highlight clip below (2 mins total).

AI 16 March 2026 · 1 clip

Dwarkesh Podcast

Terence Tao

Episode

Terence Tao – How the world’s top mathematician uses AI

Terence Tao, billed by the episode as the world's top mathematician. He predicted in 2023 that by 2026 AI would be a trustworthy co-author if used correctly, and he has been tracking AI-assisted solutions on the Erdos problem list, roughly fifty of about eleven hundred at the time of recording.

Hosted by

Dwarkesh Patel

20 March 2026 · 1h 23m · 1 clip below

contrarian

Terence Tao puts an AI tool's success rate on any given unsolved maths problem at 1 to 2%, and says the fifty Erdős problems that fell had almost no prior literature.

Systematic sweeps produce that rate; the wins that circulate are picked after the fact out of thousands of parallel attempts, so the visible record is a filtered one. He wants standardised challenge sets built outside the labs, because a vendor reporting only its successes leaves nobody able to price the capability.

2m 23s · from 43:46
My observations
- Tao reports the rate per attempt, not per dollar; ten thousand parallel attempts at a 1% hit rate still clear the problem, which is the half of the bull case his number leaves untouched.
- Erdős left these problems stated once and never worked up, so what the models were rewarded for was finding an obscure result and bolting it to a standard technique.
- Any capability figure sourced from a model provider is a maximum rather than an expectation until the sweep behind it is published, and the denominator is what a board would want established before the number appeared in a paper.
- model-capability

16 March 2026

AI 16 March 2026 · 1 clip

Terence Tao puts an AI tool's success rate on any given unsolved maths problem at 1 to 2%, and says the fifty Erdős problems that fell had almost no prior literature.