3 Aug
2023
3 Aug
'23
10:44 a.m.
idea: it could be forced to say “maybe” before low-confidence words idea: low-confidence tokens could be masked in further inference, so it doesn’t amplify guesses into errors