I think recent models have been trained both to reduce sycophancy and so increase disagreement, and that some other aspect of training (possibly extensive RLVR/RLER) has increased mode collapse and so reduced the flexibility of model voice.
Of course they’re still highly sycophantic in some ways, and still have lots of variability. IT’s interesting that more recent Claudes seem to sometimes just get irritated and decide to fight you instead of accept your corrections. Usually in similar circumstances to where a human would get irritated and fight me :)
I think recent models have been trained both to reduce sycophancy and so increase disagreement, and that some other aspect of training (possibly extensive RLVR/RLER) has increased mode collapse and so reduced the flexibility of model voice.
Of course they’re still highly sycophantic in some ways, and still have lots of variability. IT’s interesting that more recent Claudes seem to sometimes just get irritated and decide to fight you instead of accept your corrections. Usually in similar circumstances to where a human would get irritated and fight me :)