“Has a reasoning model ever come up with a math concept that even seems slightly interesting to a human mathematician?” Full episode w Ege Erdil & Tamay Besiroglu out Thursday.

@EgeErdil2 @tamaybes Look, one can pretend that Move 37 never happened only for so long.

Logic dictates that something—or someone—always had to exist. Assume it was a “someone,” not a “something.” Why would such a being create a world like ours, one filled with pain? The Advent of Time provides a definitive answer.

@EgeErdil2 @tamaybes me on thursday

@EgeErdil2 @tamaybes this is gonna be a fun week for these guys when OpenAI introduces their $20k/mo scientists. but even now, I find it highly doubtful that they've produced nothing worthwhile for mathematicians

@EgeErdil2 @tamaybes Seems like a prompter issue. You need to ask the model to explore ideas across domains and make connections. It can't be that hard to start with a problem, iterate through nearby ideas looking for connections, and refine the most promising ones.

@EgeErdil2 @tamaybes What about FunSearch? Presumably this was interesting to at least some mathematicians? (I recognize this wasn't a pure reasoning model working in isolation, but not sure that's a critical distinction and in any case ... it's from late 2023!)

These kinds of criticisms are in a way unpaid labor for the research labs. It’s difficult to think of a persistent criticism they can’t just decide to work on. I would like to see a nontrivial formal critique because all the behavioral evidence in the world wouldn’t disprove that the next model can’t be an answer to whichever problem.

@EgeErdil2 @tamaybes if a reasoning model does succeed in coming up with a breakthrough insight in mathematics, will it be more optimal for the lab to share the result or keep it hidden ( in order to avoid regulations and such)?

@EgeErdil2 @tamaybes refreshing stuff

@EgeErdil2 @tamaybes interesting? more like math models trying to impress their crush. let’s see if they can charm a human mathematician or just end up in the friend zone.

@EgeErdil2 @tamaybes Wow! Excited for this one

@EgeErdil2 @tamaybes cc: @littmath

@EgeErdil2 @tamaybes bro your content is so good I always clear my schedule after work to watch it in my living room! Like watching on my phone doesn’t do justice to the Dwarkesh podcast experience. I’d pay you monthly subscription to get 2x weekly super long podcasts fr

@EgeErdil2 @tamaybes This seems interesting! While AI is reproducing a lot is a fact now, whether AI models are producing anything new at all is the real question.

@EgeErdil2 @tamaybes The new information report of true might hint more at how valid the recombination are

@EgeErdil2 @tamaybes Can't wait to listen to it!

@EgeErdil2 @tamaybes Love the different viewpoints!

@EgeErdil2 @tamaybes

@ancerj @EgeErdil2 @tamaybes People have extrapolated from it being able to retrieve and package info from its training set to it being able to discover and invent. Is this reasonable?

This is why the persistence of LLM memory is relevant to mathematical reasoning, though. LLMs might have the capacity to produce novel observations or hallucinate ideas, but piecing together insights enough to string together new theories still takes some iterations of prompting and analysis. (ChatGPT has modeled new mathematical shapes that I could not find in existing literature).

@EgeErdil2 @tamaybes if any, I bet human mathematicians won't disclose the discovery was helped by the reasoning model

@EgeErdil2 @tamaybes I’d be interested in seeing a poll from your followers on where they see the timeline.

@EgeErdil2 @tamaybes I wonder how many mathematicians in the world could provide an interesting answer to the prompt "come up with an interesting math concept" - even if given 6 months. I suspect that this is a skill that lies right on the very frontier of math - we're just not there yet.