undefined | Better HN

0 pointsletmevoteplease3y ago0 comments

"Interestingly, the base pre-trained [GPT-4] model is highly calibrated (its predicted confidence in an answer generally matches the probability of being correct). However, through our current post-training process, the calibration is reduced."[1] The graph is striking.[2]

[1] https://openai.com/research/gpt-4

[2] https://i.imgur.com/cxPgkhD.jpg

0 comments

furyofantares3y ago

They should make the aligned one generate the text and the accurate one detect if it's lying, override it, and tell the user that it doesn't know.

j / k navigate · click thread line to collapse

0 comments

furyofantares3y ago

They should make the aligned one generate the text and the accurate one detect if it's lying, override it, and tell the user that it doesn't know.

j / k navigate · click thread line to collapse