[lumost]

I really want a machine which gives me the statistical average opinion of all reviewers in a target audience. Sycophancy is a specific symptom where the LLM diverges from this “statistical average opinion” to flattery. That the LLM does this by default without clarifying this divergence is the problem.

Usually retrying the review in a new session/different LLM helps. Anecdotally - LLMs seem to really like their own output, and over many turns try to flatter the user regardless of topic. Both behaviors seem correctable with training improvements.

[estimator7292]

Yeah, most of the time when I want an opinion, the implicit real question is "what sentiment does the training set show towards this idea"

But then again I've seen how the sausage is made and understand the machine I'm asking. It, however, thinks I'm a child incapable of thoughtful questions and gives me a gold star for asking anything in the first place.