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

i mean they’re spectra, i think, because unflinching false confidence is also bad, particularly from a machine which is often used for hunting for mistakes

i am not convinced we have good measures for what we want, which is actually "not full of shit"

"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

current generation of models are neurotic, and it annoys me to death

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.

I think it depends on what you’re doing. I do a lot of dev with low risk tolerance and I wish Claude was even more cautious and neurotic. I just switch to Gemini if I don’t need it to be cautious

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.

I can only use Claude or Gemini at work, so I can’t compare with ChatGPT, but Claude is very good about asking before making any changes to any plan

I agree but I’d rather have Claude push back for dumb reasons many times then have it not push back one time when I’m wrong

That's what I use adversarial review with another model for.