Y
Hacker News
new
|
ask
|
show
|
jobs
by
idiotsecant
16 days ago
The problem is that all the frontier models tend to be
more
sycophantic when confronted with emotional support issues.
1 comments
agnosticmantis
16 days ago
I believe sycophancy is a side effect of RLHF and whatever reward function it explicitly and implicitly optimizes.
link