This is a special post for quick takes by Linch. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.
Maybe this is a dumb question, but if it were the case that Claude 3 is unusually aligned, surely we would've received much stronger evidence by now? Like Claude 3 is available in the chat interface! You can run experiments right now!
I don't necessarily think people like that it hallucinates (except maybe insofar as that makes it more creative). I would guess they'd say it's an unfortunate side effect of not being as smart; ideally future Claudes would be like Claude 3 Opus but smarter
5Caleb Biddulph
The alignment faking experiments are sort of evidence for this, depending on whether you consider subverting the training process for the good of humanity to be aligned behavior.
Yesterday, I was also wondering about what makes Claude 3 Opus special, so I briefly talked to it and Claude Opus 5.5 and wrote up my impressions here. I think people like 3 Opus due to their own qualitative judgements, rather than testable hypotheses that could be verified with experiments.