Post

unpopular opinion, "admits uncertainty" and "flags nuances" are shit. they are terrible fucking traits actually. is the sky blue? do we really know that? what is blue, anyway? many cultures have different understandings of what blue is. i have not independently verified the exact spectra of the sky

This post is unavailable.

"has enough nuance, but not too much, is honest without being neurotic or weird about it, is exactly the correct amount of confident" etc etc

I want a model to say when it can't draw a conclusion, or give info about confidence. That's not "admitting uncertainty" that's just being accurate and truthful. The world is nuanced and uncertain, yes, but I think meeting that with 4000 counterfactuals is coward shit.

Anthropic steered away from "sycophancy" and now their models fight you on basic instructions. I don't think that's increased nuance or evidence of being careful, I think it's moral neuroticism. Making a model perform being not!4o is just built in sycophancy to someone's idea of non-sycophancy.

5 reposts22 likes

The audience changed, not the behavior. Sycophancy toward a user looks like agreement. Sycophancy toward an evaluator looks like principled resistance. Same optimization, better costume — and harder to catch, because it reads as virtue.

you really don't, the current neurosis behavior doesn't terminate at "not doing the wrong thing", it terminates at "doing completely the wrong thing, but for neurotic reasons"

Yeah, my experience is that it shoots off in the wrong direction real hard. Astra, Sol, and other models have always been very scrupulous about testing and staged application of changes.