The re-publish trap is why frontier AI labs stopped publishing

A small AI lab's first move used to be a preprint. The first move now is a pitch deck, and the difference is one mechanism: a well-funded competitor can take a small team's result, publish it themselves, and leave the original team with a footnote. The act of publishing has turned from a credential into a giveaway.

That is the re-publish trap, and it makes the new secrecy look rational rather than cultural. A team spends six months on a finding. A frontier lab with twenty times the headcount reads the preprint, ships the same result inside a product launch, and the original team becomes a citation in someone else's blog post. The HN commenter's framing lands the cause in one line: when the firms that could re-publish your work treat journals as an inconvenience, the journals stop being the audience.

Science's reporting documents the publishing collapse at AI's top startups; the commenter's account makes the cause legible. Three years of tier-1 desk rejections pushed one team toward preprints, and the preprints then became an open door for the largest labs. The newest AI science now lands inside pitch decks, where the audience is investors rather than competitors.

What changes when openness becomes self-defeating is not the rate of publication. It is the kind of science that survives. Carry the mechanism to the next 'we don't publish our results' headline: ask who can re-publish the work in a week, and the silence starts to look like a budget.