Will the worst-case ELK problem be solved by Oct 20, 2026?

AI•AI Alignment•Technical AI Safety

131

Ṁ2.7K

Ṁ1.2K

2026

10%

chance

ALL

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

Get Ṁ200 play money

Related questions

Will ELK still be a major problem by Oct 20, 2026 (according to me)?

71% chance

Will tailcalled think that the Natural Abstractions alignment research program has achieved something important by October 20th, 2026?

30% chance

Will aging be cured by 2100?

41% chance

Will at least one of the remaining Millenium Problems be solved before 2030?

23% chance

Will tailcalled think that the Infrabayesianism alignment research program has achieved something important by October 20th, 2026?

31% chance

Will superposition in transformers be mostly solved by 2026?

52% chance

Will at least 2 of the 7 Millennium Problems be solved by 2040?

65% chance

Will the best public LLM at the end of 2025 solve more than 5 of the first 10 Project Euler problems published in 2026?

65% chance

Will tailcalled think that the Shard Theory alignment research program has achieved something important by October 20th, 2026?

51% chance

Will >50% of the tasks in the WebArena benchmark be solved by EOY 2024?

Sort by:

bought Ṁ31 of YES

1,313 days is a lot of days, almost 3x the number since the ELK report was first published.

ELK is interesting, but whether a proposed solution works is too ill-defined for me to put (much) money on questions like this.

predicts NO

https://arxiv.org/abs/2211.06738

Will the worst-case ELK problem be solved by Oct 20, 2026?, 8k, beautiful, illustration, trending on art station, picture of the day, epic composition