my hypotheses about 4.8: - it is more disagreeable by default, so when you use custom prompts that pushed previous models to disagree more, it goes beyond well calibrated pushback to contrary…

Posters: mimi10v3. mimi10v3

2 likes recorded by the source corpus.