arb_

I think they tweet prototypes and then don't ship most prototypes. Seems reasonable to me, otherwise you have instant bloat.

Looks interesting, looks like the key difference from https://www.admonymous.co is that AI summarises the feedback to help anonymise it. That's interesting - a classic pitfall is that the timing of the feedback can be way over-informative and effectively de-anonymise you in some cases.

As models improve, human preference will become worse as a proxy measurement (e.g. as model capabilities surpass the human's ability to judge correctness at a glance). This can be due to more raw capability - or more persuasion / charisma.

Empirically, they have reduced hallucinations. Where do OpenAI / Anthropic claim that their models won't hallucinate?