AI models reportedly prove the Generalized Riemann Hypothesis in 17.4 million lines of Lean
The stochastic parrots have apparently proved the Riemann Hypothesis
Levent Alpöge says he got Anthropic models to produce a Lean proof of the Generalized Riemann Hypothesis, running to 17.4 million lines of code. Scott Aaronson bets the result will hold, noting that no AI result with a Lean certificate has yet been retracted, though he warns that Lean's kernel bugs and type-theory soundness gaps urgently need fixing. Mathematicians, he adds, won't be satisfied until they can reconstruct the ideas themselves.
In my view, these issues can and should urgently be fixed! Until they are, a sufficiently motivated AI could almost certainly exploit them to prove a false statement!
- Mikhail_Edoshin
There's a book, "Impro", about a theatric studio that practiced improvised sketches. One thing they struggled with was a normal conversation. People have them all the time yet somehow artists could not find the right approach.
One day the head of the studio (the book author) had an insight. She asked the actors to drive the conversation so that they always bump the status of themselves or the other party in any direction, but only a little.
This did the magic. Not only the actors understood her perfectly, the conversation started to flow and could go endlessly while staying amusing and captivating for spectators. (Look up the book for examples.)
After they discovered that secret they started seeing the pattern (status games) everywhere up to the point they asked themselves whether there is anything else in human communication.
This was a British studio, the sketches were mostly comedy.
- esjeon
Lean + Stochastic predictor = Brute-force theorem solver
P.S: AlphaGo was pretty much the same in the spirit btw.
- sc077y
If this isn't intelligence, then I don't know what is
- hoppp
Thats a lot of lean. Can anyone comprehend how it works in it's entirety?
- alterom
"We should still fix the leak"
I think this approach and conclusion applies to a lot more than the immediate context of that conversation.
- janalsncm
I think when people set out to create a chess playing computer they thought we’d end up with something that we could deeply understand at an intuitive level. Instead, when you look at the Stockfish evaluation of a position, you do get an “explanation” but it’s really just the aggregate of a bunch of calculations, which are themselves stacked on top of the “stochastic parrot” NNUE evaluator.
This leaves us in a very unsatisfying position as creatures who want answers to “why” questions. Intuitive answers are incomplete, and complete answers are either fundamentally impossible or completely incomprehensible or both.
Feynman said something similar with regards to magnetism, but I feel that understanding a chess position is an even better example. Magnetism depends on facts about the universe which at their most basic level “just are”, whereas even a complex chess position is theoretically entirely explainable.
- hansvm
Derailing slightly, a variety of HN posts (the blogs/sites they reference) have started banning me with a 403 page from nginx, including this author. My IP reputation is supposedly fine. Does anyone have any ideas where I should look next?
- elorant
Wonder how many mathematicians could reach similar breakthroughs if they had near infinite compute.
- jstummbillig
I wonder he good we will be at not understanding anything to an ever increasing degree, with absolutely no hope of ever catching up. We will be super bad at absolutely everything within sciences and engineering and completely unable to contribute anything of substance, outside our own moods and quirks.
Can humans be okay with that? Can we be happy cats?
- stymaar
Tangent: LLM have stopped being stochastic parrots since RLVR was introduced in their training pipeline.