Post by @tao@mathstodon.xyz

2026-08-06 14:40 UTC

I can also envisage a more two-step approach to releasing results, in which one begins with an announcement of results in progress without full release of details, followed after some time by preprint-level artefacts in all the dimensions mentioned above. The announcement would not be considered the primary date of completion, but would help coordinate competing efforts towards the same problem. I am not sure whether this would be a desirable development, but one could also revive the ancient practice of releasing what we would now call encrypted hashes of one's in-progress work as part of the announcement.

Replies (2)

@bwag@mathstodon.xyz 2026-08-06 15:46

@tao@mathstodon.xyz If the results are independently discovered by others within a few days, then why does priority even matter? I don’t think we should be so concerned with how to assign credit to the “first” person to prompt an LLM into producing a particular proof. As you’ve been saying, let’s entirely separate the credit for production and for digestion/dissemination. The first person to write a nice proof outline or give a good talk on a result doesn’t have to be one of the people who first claimed it. This will also lessen the incentives to rush to post half-baked or undeveloped results.

Open ##5165732
@srinivas@mathstodon.xyz 2026-08-06 16:57

@tao@mathstodon.xyz If everyone plays nice, a protocol is enforced by conscience and social norms. When they don't, the best solution via social norms is litigation, which I suspect a research community would dislike more than the general public. Given that there is a canonical representation of a proof, however, it may lead to well-defined 'attribution problem' where there is some notion of "value/criticality" for the various components of a proof in its canonical representation. Attribution is then settled using this "proof-data-structure" and the git-repo submitted to the "attribution engine" for review. AI agents can do proof-work according to the protocol you outline as well, and use the "attribution-engine". The little bandwidth a research community has for litigation can then be used just to guard against the xkcd'esque future where counterproductive proof-work results from gaming the attribution system. This is all assuming that the "proof-content" matters more than whether/not it was AI generated. I'm not entirely convinced that this assumption is valid.

Open ##5165733