If AI Can Do the Math, What Do the Mathematicians Do? (ft. a16z Researchers)
As AI solves hard math problems, discovery may outrun human understanding, but the panel argues that understanding, taste, community and joy of discovery remain human work, and that there will be more of it, not less.
Watch on YouTube ↗Charlie's summary
- ▶ 5:38
Two frontiers are separating: the discovery frontier and the human understanding frontier
Ittai describes one frontier where new results are produced and another where humans actually understand them. Thurston's 1994 essay made understanding the goal of math. Ittai says that over the last half year the two have split: AI produces proofs, including Lean proofs of major problems, and mathematicians are “struggling to finally get to a human understanding” of them. Whether they converge again is open.
- ▶ 10:13
AI solved in an hour what took Ittai two years, and Justin could suddenly produce conference papers on demand
Ittai spent about two years on metric space embedding problems. This summer a model made a non-trivial improvement on one in about an hour. Justin says GPT-5.6 was a phase transition: it could solve open questions from serious conference papers, work he compares to what a very good grad student does in a PhD. Justin's takeaway is that you must now choose what is worth publishing, since you can no longer publish every result you get.
- ▶ 16:18
Lean makes AI-assisted proofs checkable, turning verification from a bottleneck into a compile step
Lean is a programming language for logic: if the proof compiles, the kernel has checked that the chain of statements is complete. Formalizing used to take top mathematicians months. Justin says that in the last five months you can ask an AI for a Lean proof of a typical theory paper's main theorems, then check that the stated theorem is the one you meant. That gives more confidence than pre-AI review ever did. Scott adds that formalizing was prohibitive because every foundation must be checked “turtles all the way down”.
- ▶ 21:57
Models prove correctly but write badly: they labor over easy steps and skip the hard ones
Dan says AI papers spend pages justifying obvious steps, then dismiss a hard one in three lines. Humans therefore only bother to decode a paper once a Lean proof shows the result is correct. Ittai explains why: checking a proof is an easy reinforcement signal, but a good explanation needs a theory of mind of the reader, which is hard to score. He expects the gap to last a while, with humans doing much of the work of understanding. Dan expects models to write well within a year.
Give your email to unlock the rest of the key points — for every episode.
Do more with Charlie
Free — sign in with your email, no password.