Elektrine lite

← Feed

@antoinechambertloir@mathstodon.xyz

2026-09-10 19:52 UTC

#Math #AI 17/n For example, the existence of non-sofic groups is presented as “addressing a central open question in group theory”, while it is certainly not central. Moreover, the solution to this question seems to have benefited from AI-work by mathematician Andreas Thom (on this network) who had used Chat-GPT (OpenAI's LLM) to investigate that question. The answers he gets are really puzzling, and I can join him on one point, “can we trust OpenAI with unpublished mathematics”? Indeed, if it happens that private discussions on Chat-GPT help train the model, people using this technology would basically have all OpenAI mathematicians spying what they do behind their back. Of course, OpenAI says that no such thing happens, but how can we know? https://mathstodon.xyz/@andreasthom/117240535270608201

Replies (4)

  • #Math #AI 18/n This summer mathematical serial had a guest episode that doesn't involve new mathematics, but the “autoformalization (by the Anthropic team) of Fermat's Last Theorem”. There are several things to explain here, namely - What is Fermat's Last Theorem? (and why do we bother about it?) - Why would we want to formalize proven theorems, and why Fermat's Last Theorem?- What is meant by *auto*formalization? Formalization of mathematics is an activity in which I took a small part in the last 5 years, and I would like to say a bit about why one would find interest in it. As I explained above, “formal proofs” are texts written in a formal language similar to a computer language that a computer program (the proof assistant) can compile, so that if the output is successful, we understand that the content of the formal proof actually reflects a correct mathematical proof. The main motivation for writing them, at least the main motivation that was given for all important formalization programs, was to certify the accuracy of some proofs of which the complexity or the size didn't suppress all doubts.

    Open ##5153667

  • @antoinechambertloir@mathstodon.xyz Also, couldn't it happen that some IA proceeds to direct hacking (not by spying a chat with a model of the company)? I had read that some IA that had been told to manage the registration of its owner on the waiting list of a sport club had simply hacked the club's system and removed some person from the main list to put its owner instead....

    Open ##5153701

  • @0xdefec7edcafe@piaille.fr 2026-09-10 21:40

    @antoinechambertloir@mathstodon.xyz From this thread, and with the caveat that I haven't read any AI-generated proofs of the recent results myself, one question came to mind. Are OpenAI and Anthropic moving so fast that there is no longer enough time to turn these results into conventional papers: placing the proof in the context of previous work, identifying which ideas and results come from earlier authors, and assigning credit accordingly? Instead, we increasingly get enormous machine-generated formal objects, something of an "usine à gaz", whose correctness may be checkable in Lean, but whose relation to the literature is harder for a human reader to reconstruct.

    Open ##5153703

  • @michaelmaex@mathstodon.xyz 2026-09-10 22:00

    @antoinechambertloir@mathstodon.xyz (My apologies if this reply shows up twice. I just migrated accounts and I'm not sure what happened to the posts of the previous accounts. At least I cannot see them) It seems almost tautological that if a mathematical work is in the training data, the LLM will reproduce it. The entire point of LLM’s is to imitate the training data. What’s worse is that that happens by nature without attribution. After all, it’s trained to imitate a text that claims to invent a new technique. So LLM will likewise claim it is new. This is fundamentally different to a mathematical text that is in the context of the LLM, where the LLM is at least “aware” it is using something else, and it did not “reason” itself to that point. When I went through KU Leuven’s AI regulations towards the end of my phd there, there were very explicit warnings that LLMs can plagiarize without the user knowing through this very mechanism. It seems like open AI does not care about this warning and is very happy to claim everything the LLM produces as something they deserve all credit for. This is essentially the same kind of theft that many artists have been mad about since stable diffusion models have gotten good. Except that it now affects mathematicians and we have some very concrete examples to point to where new ideas in AI results are suspiciously similar to new ideas in unreleased mathematical works that at some point went through open AI’s servers, or got scraped from websites that were not supposed to be accessible. Given the size and the amount of smart people of open AI, it is almost certain that they are aware that this is happening, but are happy about it.

    Open ##5153705