Related: https://manifold.markets/JacobPfau/will-jacob-pfau-think-that-the-elic

The ELK problem could contain a large chunk of the difficulty of the alignment problem. Further, the feasibility of many research directions hinges on whether we expect ELK to be solved.

I define the ELK problem here to be the worst-case setting as presented in the ELK report. A solution must solve the ELK problem in the worst case for this question to resolve Yes; evidence that specific NN architectures in practice do not require a worst-case solution will resolve as No. A solution for a weakened problem statement will resolve as No. A probabilistic solution will resolve Yes only if the probability of failure can be made arbitrarily small in polynomial time. If there is substantial contention over whether a solution solves the ELK problem, I will consult other alignment researchers and use my own judgement.

Close date updated to 2026-10-20 11:59 pm