4 ms·
"Interestingly, the base pre-trained [GPT-4] model is highly calibrated (its predicted confidence in an answer generally matches the probability of being correc
by letmevoteplease 4y ago
"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 https://openai.com/research/gpt-4
[2] https://i.imgur.com/cxPgkhD.jpg https://i.imgur.com/cxPgkhD.jpg
- furyofantares 4y agoThey 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.