Thread

1. Animats · 1790745846 · focus · HN ↗

   This paper wants AI companies to pay human mathematicians to understand AI-generated stuff. That's an unusual ask.

   1. unddoch · 1790749416 · focus · HN ↗

      If youre going to spend 10 million dollars on 10000 agents trying to solve some important maths problem I think it's reasonable to ask for some grants to help digest whatever they came up with. Or you could hire mathematicians and do it in-house, but I guarantee you grants to PhD students are cheaper than silicon value salaries.

      1. ChickeNES · 1790808235 · focus · HN ↗

         Why pay for grants when you can just spend the money on compute more efficiently?

         1. eli_gottlieb · 1790809761 · focus · HN ↗

            Because the compute isn't gonna turn AI slop into a clear, readable paper.

            1. gus_massa · 1790817488 · focus · HN ↗

               Not for now, but it's AI is getting better. A big step is "refactoring" a very long proof into a few intermediate lemas and theorems that are more inteligible and useful for other proof. It may take a few years or decades in some cases.

               Anyway, I expect AI to be better at "refactoring", but for now a centaur is better.

               1. hodgehog11 · 1790859237 · focus · HN ↗

                  Actually, for what you are mentioning, it is getting worse. There was a sweet spot somewhere around the release of GPT-o3, and ever since, the LLMs have been getting more accurate at solving problems, but worse at explaining how, and to hone in on what is interesting. This isn't surprising, as RL strategies shifted from RLHF to RLVR, so priorities during learning changed. I don't expect AI labs to reverse course on this. We can expect AI proofs to become increasingly incomprehensible over time.

                  1. gus_massa · 1790871351 · focus · HN ↗

                     I still remember Garry Kasparov vs. Deep Blue. Now, Magnus vs Stockfish is not even funny. I've seen AlphaGo and AlphaStar, and how their communities reacted...

                     I remember when Mathematica only could tell the answers and you had to type the formula correctly. Now there are apps that solve the exercise from a photograph with all the intermediate steps. We reminded the T.A. to be more alert during the midterms becuse we already had problems. I'm very worry about the magical glasses now, but it's important to be not overreact and be polite with the students.

                     Back to refactoring unintelligible long math proof: Let's talk again in 2031.

            2. GPerson · 1790948784 · focus · HN ↗

               The paper is secondary to developing a general understanding of new ideas. The AI is not going to do that. All of the AI bros in this thread are going to be addicted to super AI Netflix and the math community will be dead.

               1. eli_gottlieb · 1790963439 · focus · HN ↗

                  The paper, both writing it and reading it, are a means by which we develop our understanding of new ideas. We can surely come up with a better means if we try, but "stop trying and leave actual understanding to machines" ain't gonna cut it.

                  1. GPerson · 1790971213 · focus · HN ↗

                     Right we agree.
      2. 6510 · 1790810920 · focus · HN ↗

         They aren't buying tokens, more likely it costs them 1m to run the 10000 agents for 88 hours. Still leaves room for the PhD ofc.

         I'm trying to map it on other fields and cant help but think it is absurd. If some company spends 1m developing a new alloy, should metallurgists be upset when they publish the recipe?

         1. GPerson · 1790948829 · focus · HN ↗

            It was a lot more expensive than this… Just look up the estimates before making stuff up.

Unofficial Hacker News client; not affiliated with Y Combinator.