Quote · Dwarkesh Podcast
Adam Brown – A deep but accessible introduction to general relativity
Where this was said
How far can AI get without experimental evidence?
At 1:32:45 · chapter starts 1:29:33
Dwarkesh asks whether humans will be able to keep up with AI civilisation's scientific discoveries. Brown is cautiously optimistic, using Terry Tao's 'indigestion' framing — the fear of billion-line inscrutable Lean proofs — as a foil. The empirical counterevidence: the LLM disproof of the Erdős unit-distance conjecture was not a Lean certificate but a human-readable argument using new graph-theoretic ideas, which mathematicians immediately seized on to prove new theorems. Brown also notes that AI models have 'extreme patience' — they will cheerfully spend resources trying to disprove a conjecture everyone believes is true, which is exactly the kind of cognitive stubbornness that could unlock progress. The episode ends with Dwarkesh thanking Brown and Brown noting that explaining 100-year-old physics is, in fact, a genuinely fun way to spend an afternoon.
The fear is that AI will produce proofs no human can understand. The evidence suggests the opposite: the LLM disproof of the Erdős unit-distance conjecture produced human-interpretable ideas that mathematicians then used to prove brand-new theorems. Superhuman explaining, not just superhuman proving.
