Hacker News Reader

The problem is that all the frontier models tend to be more sycophantic when confronted with emotional support issues.

I believe sycophancy is a side effect of RLHF and whatever reward function it explicitly and implicitly optimizes.