Post by @srinivas@mathstodon.xyz

2026-08-20 08:47 UTC

@tao@mathstodon.xyz I had thought incentives strongly favoured your present position after reading your initial "proof-parenting" post. I take your metaphor to mean "division of labour for proof-work". I'll use it as such to avoid fraying a nice metaphor. If mathematicians are selflessly proof-minded, it seems to make sense to formally divide labour for proof-work so each division may be undertaken by the agent (human/AI) best suited for it. However, if we assume personal mathematical motivation, "proof-adoption" may seem bleak since it forebodes less natural "proof-conception", "proof-gestation" etc. as AI learns to be best suited for every necessary task. The motivation for AI to prove something seems to have at least one natural source, @kangmeister@mathstodon.xyz - the utility of the proof to the larger sci./engg. agenda: a proof guarantees a claim, minimising the risk in making decisions with it and saving costs downstream. Currently a proof requires human insight and understanding. If, instead, these are mere niceties afforded by AI, then they will likely disappear over time when their utility - getting enough mathematicians to feel safe adopting & training AI - is fulfilled. However, I still think there is a strong argument for engaging in part of this dynamic.

Replies (1)

@srinivas@mathstodon.xyz 2026-08-20 08:55

@tao@mathstodon.xyz The case for why we should do it anyway: if the economic incentive heavily favours AI doing proof-work then it will likely happen. We can choose to have a say in *how* this happens, but not *if* it does. Perhaps some features to consider could be: 1. Design for mathematician-centricity: One AI's proof is vetted by a competing AI before it is proposed/submitted to a human via an "open (AI/LLM)" category as mentioned in https://mathstodon.xyz/@srinivas/117038594469750488 . Each AI is rated on the number of human-verifications its proofs have received. 2. Maximise human digestion of AI-generated proofs: Incentivize at least one published video (on Mathematical Discourse, for instance) for every proof provided by an AI-lab as in your suggestion here -https://mathstodon.xyz/@tao/117046985982017853 . The financial incentive may be via a single fund from competing AI-labs granted to a single consortium of researchers as in https://mathstodon.xyz/@srinivas/117045197053214360 3. Attribution: For every person who has worked on the proof, there is a clear attribution-score provided by a deterministic attribution engine, as in https://mathstodon.xyz/@srinivas/117049621107898711 .

Open ##5132047