Aaronson's blog post on an apparent AI-assisted Riemann Hypothesis proof mocks 'stochastic parrot' critics
Summarized from the Hacker News discussion because the article couldn't be read
- Commenters say the claimed proof was built with Lean plus other proof techniques and a human mathematician, which one commenter argues undercuts the 'stochastic parrot' label.
- A commenter reports that the author of the claimed proof posted a top-level sketch of it in a paper, and that the author seems to understand it.
- A commenter says the formalization is about 17M lines of code, and that checking the hypothesis statement and the kernel's acceptance covers the trust needed.
- Commenters point to a separate Hacker News thread with 1000+ comments on Alpöge's result, which suggests the claim has already drawn wide attention.
Hacker News opinions
Aaronson's jab is aimed at people who said LLMs just repeat locally coherent phrases and that scaling can't produce real intelligence. If a GRH proof doesn't count as intelligence, the goalposts have moved a long way.
Lean doesn't make you read all the code. You only need to check that the hypothesis is formalized correctly and that the kernel accepts the proof. That's the whole value of Lean.
Nobody is verifying 17M lines of code. Calling this a breakthrough is premature until someone outside the claim checks it.
I'm not sure the headline is mocking anyone in particular. It reads like a sneer at the 'stochastic parrot' crowd, and that crowd earned it.
Chain-of-thought probably matters more than RLVR for reasoning. The model can work its way to an answer instead of guessing it in one shot from training data.
They're still next-token predictors at the core, but the whole system does more than that. It's like slime mold solving a maze: no single piece is smart, but the system is.
Scott is a theoretical computer scientist, so he's probably watching his own career prospects shrink. The people at the top will do fine for a while, and then the floodwaters come for everybody.