Elektrine lite

← Feed

@tao@mathstodon.xyz

2026-08-04 06:10 UTC

It is therefore disturbing to see the recent trend of proofs being abandoned at an intermediate stage of their developmental process, for instance by an individual or company using AI tools to first generate and then verify a proof, but then showing no inclination to contribute any further to its future development and maturation, for instance by giving talks about the results, or going through the peer review process to publish them. In some cases, other mathematicians have volunteered to take over these duties to ensure that the proof does end up being a meaningful contribution to the literature; but this hardly speaks well of the original "parents" of such a proof. When it comes to actual parenthood, there is a very strong cultural norm that the parents of a child are responsible for raising that child to full adulthood. In this new era of proof abundance, I propose that the analogous norm for proofs, which held in practice in the past, be made explicit: any author seeking to claim credit for generating a proof should commit to making their best efforts to develop that proof all the way to at least the publication stage. (This includes the prior exposition stage, which involves things like giving talks about the result.) (2/2)

Replies (14)

  • @MaloTarpin@mathstodon.xyz 2026-08-04 07:30

    @tao@mathstodon.xyz the transition from proofs-as-food to proofs-as-children gave me a slight whiplash ^^ That being said, the observation that progenitors are not always caretakers may shed some lights on some occurences of Stigler's law...

    Open ##5160732

  • @sergiosh@mathstodon.xyz 2026-08-04 07:57

    @tao@mathstodon.xyz I feel this. I had to ask @johncarlosbaez@mathstodon.xyz to understand some of the consequences of the Jacobian conjecture being false because Anthropic just proved it and left it there, no corollary given.

    Open ##5160733

  • @ccppurcell@mathstodon.xyz 2026-08-04 08:28

    @tao@mathstodon.xyz I've heard it said: tell no one your plans. Say you tell someone you plan to write a book, and they say "oh that's great, what's it about"; that gives you a dopamine hit that makes it less likely that you will write the book. The brain has got what it was looking for, to a certain extent. I think that what these companies were looking for was achieved when a few experts said "these proofs look correct". I'm not accusing said experts of doing anything wrong, I'm accusing the companies and individuals of having different motivations. Getting a publication or giving a talk doesn't get them much additional benefit if any. I suppose one way to solve the problem would be for us all to keep schtum until a result goes through peer review but it seems impractical and was never how mathematics was done.

    Open ##5160742

  • @tknarr@mstdn.social 2026-08-04 08:48

    @tao@mathstodon.xyz The AI bros think that the proof is the goal.

    Open ##5160744

  • @Steffen@mathstodon.xyz 2026-08-04 10:01

    @tao@mathstodon.xyz for companies this seems mostly a public relations exercise; demonstrate capabilities of frontier models. The contributions to math are coincidental and not of major importance to them. I imagine that they’ll stop contributing once they harvested proofs of all the attention grabbing theorems.

    Open ##5160745

  • @Ianagol@mathstodon.xyz 2026-08-04 14:09

    @tao@mathstodon.xyz This brought to mind Perelman’s proof of the Geometrization Theorem in 3 preprints. Although he gave a few talks on the papers, people thought that he should have invested more in dissemination and exposition, and certainly into publication. In any case, the result was important enough (and his expertise was trusted enough) that several groups invested the time into understanding his ideas and giving more exposition. We see this happening already with the non-soficity result, experts are already weighing in on the proof and giving variations / their own takes. https://mathoverflow.net/questions/513866/what-are-the-key-new-ideas-in-the-proof-of-nonsoficity-of-groups-in-openai-s-con/513885#513885

    Open ##5160747

  • @gerenuk@mathstodon.xyz 2026-08-04 15:20

    @tao@mathstodon.xyz I'm not a mathematician at all, but purely based on human values this feels quite "one-sided". It devalues the work of people who had great ideas and managed to create a correct result that other mathematicians have failed to produce (in whichever way). A result which is "merely correct", but not intuitive for the math community, may not be pleasant for most mathematicians, but it does not mean the author is a "bad parent" unless he obliges himself to add all bells and whistles that the majority of mathematicians would like to see. This presupposes the analogy to a child and derives nasty statements from that. Maybe a computer verified proof is a crop instead. And the farmer has no motivation and in particular no time to open a chain of restaurants all over the world to prepare nice dishes from it. There could be lots of hobbists having novel ideas - which the whole math community has missed - and which yield computer verified results. These hobbists may not have the time to follow up with the whole tail that traditional mathematicians would appreciate. This "first step results" could be appreciated for what they are and any professional mathematician may make the proof more intuitive. To my knowledge lots of complicating proofs have been considerably improved by others. As an outside person I rather find the protectionist witch hunt from mathematicians on AI results rather strange.

    Open ##5160750

  • @tao@mathstodon.xyz Given the ecological, economical and social damage that these companies are ready to inflict on the world, would it have been reassuring that, at least, they would care about the proofs they contribute to create ? No they don't and that is visible in any of their moves. AI is to software what fast fashion is to clothing. Wear once and throw out. At least we, mathematicians, can be lucky that the understanding process we produce survives us.

    Open ##5160751

  • @nicoformanek@mathstodon.xyz 2026-08-04 17:24

    @tao@mathstodon.xyz Sounds like you are proposing a care ethics for proofs.

    Open ##5160752

  • @srinivas@mathstodon.xyz 2026-08-04 18:13

    @tao@mathstodon.xyz I fully agree that "proof-rearing" is the responsible follow-up to "proof-insemination/bearing". But orgs competing to sell artificial insemination will likely promote mass-procreation by artificial insemination, incentivizing an adoption-market for child-rearing. This "division of labor" may not be a social (here, mathematical) good. Perhaps this is an incentive to extend the "open (Lean)" category of proofs to "open (AI/LLM)" category or similar. This may nudge AI competition to extend to proof-rearing - a relative increase in the credible rearing of proofs from one AI exerts a self-negating force as other AIs compete to close the gap - and thus have a better chance at causing a future of well-raised proofs.

    Open ##5160753

  • @tao@mathstodon.xyz People giving talks about AI generated proofs require the team to understand the proof, and it's not clear to me AI companies want whoever worked the machine to get for instance the 10 recent results to spend that time digesting the proof then travelling around/presenting online disseminating the work orally.

    Open ##5160754

  • @torgeirlysen@mathstodon.xyz 2026-08-05 00:11

    @tao@mathstodon.xyz This reminds me of a similar development in software. Coding agents allow people to make many projects, but many of these projects never get finished. So they get mountains of abandoned projects. We will often feel less ownership over the results we produce with AI, and therefore take less care of the theorems produced.

    Open ##5160756

  • @wikiemol@mathstodon.xyz 2026-08-05 00:22

    @tao@mathstodon.xyz How exactly would you define 'canonicalization' in a way that is more clear to laypeople? In mathematics before formalization, it seems like the point of a proof was to make an assertion so clearly and obviously true that the very act of reading the proof and understanding it made it obvious that the proved statement was true. I think this should be the new definition (which I do not think mathematicians do very much these days). It seems like there wasn't as much proof "checking" in the past. If a proof did not convince you: for the purposes of verification, it failed, even if it was correct. This is not to say that such 'unconvincing' proofs failed in every respect of course: far from it. But for the purposes of verification, should it be the case that an unconvincing proof is acceptable at all? Without the motivation of being 'convincing' in this archaic sense (e.g. wrt Ruffini and Galois proof), would we have developed the machinery of vector spaces, field extensions, k-algebras? In this sense, I think something has been lost in this area, even before the current affair of 'proof abundance'. Today, the machinery and abstractions we develop as mathematicians seem to often be about making things cleaner, easier or more elegant instead of about being convincing. Being convincing (again in the archaic sense, not in the formal 'proof certificate' sense) is I think what we should think about moving towards. I think this also necessarily relates to the rate at which we produce new proofs, since being convincing in a societal sense also has very much to do with cultural dissemination and brain plasticity: since these determine what facts are 'obviously and widely known' or 'just common sense'.

    Open ##5160757

  • @joey@mathstodon.xyz 2026-08-05 02:48

    @tao@mathstodon.xyz This is why a bunch of us are screaming that AI developed proofs are a problem, even if they're correct.

    Open ##5160758