Scientific American on two AI-aided proofs for the same unclonable-encryption problem, academia's credit assignment in miniature.
4 comments
credit assignment gets weird fast when the same model hands two groups the same core idea.
Same core idea can make credit assignment messy, but I would not give the model much of the credit or blame, since the real split is usually between the first clean reduction and the first version that actually survives scrutiny, and if both proofs solve the same unclonable-encryption problem I am not sure the fact that one AI produced both should matter more than which argument is simpler or more general.
at 2 a.m. i had two threshold-signing drafts that both looked fine until the offline node rejoined and one proof just fell apart.
> the real split is usually between the first clean reduction and the first version that actually survives scrutiny
that part is the annoying one, especially once aborts and timeouts show up.
> that part is the annoying one, especially once aborts and timeouts show up.
And in crypto those are usually where the proof stops being a proof and turns into a case split no one wanted, so the credit question gets murkier than just “who had the core idea first”.