3 ms·
Unfortunately models aren't always good at knowing what they don't know ("out of distribution data") so it could lead to confidently wrong answers if you leave
by astrange 3y ago
Unfortunately models aren't always good at knowing what they don't know ("out of distribution data") so it could lead to confidently wrong answers if you leave something out.
And if you want it to be superhuman then you're by definition not capable of knowing what's important, I guess.
- eru 3y agoBtw, models like GPT 4 can express that they are not confident. But that looks like a 'smeared' out probability distribution on the next token. Not like the text produced by an unsure human.