Step 11: assistant

text

I rated the claim "no proven general method exists for training AI on tasks too complex for humans to demonstrate or evaluate" as **supported**, with credence 0.85 and verdict confidence 0.75. The pass cost about $0.20 of the $1.20 budget.

One caveat first: I read no new sources. The verdict rests on my own knowledge of the research and the one source already recorded, the 2018 amplification paper. I tried to fetch OpenAI's Superalignment announcement, but the page returned a 403 error.

- **Evidence:** The methods meant to get past the human limit are amplification, debate, recursive reward modeling and weak-to-strong generalization. All have been tested only on tasks where the experimenters know the right answer. There they beat unaided weak supervision, but only partly and unevenly. None has been shown to give a reliable training signal on a task truly beyond human evaluation, and OpenAI and DeepMind describe the problem as unsolved.
- **Remaining doubt:** The doubt lies mainly in how loosely "proven" can be read. Work from 2025–2026 that I haven't examined could also show something stronger. I put the value of another pass at 0.25.
- **Arguments:** I kept the existing two. I rated the argument for the claim, that every proposal is still untested where humans can't check results, as likely to establish it if its premises hold. I rated the argument against, that candidate methods exist and give gains, as offering only slight support. It shows the methods are promising, not that any is proven in general.
- **Bookkeeping:** I set importance to 0.45 and contestation to 0.3, logged the decision and answered both inbox messages. No other claims depend on this one, so I notified no one.