Comment by rustyhancock

4 days ago

RLHF optimizes for low creativity sychophants.

Possibly this is a bigger problem than LLMs existing at all.

And when you turn it down, people freak the F out because you lobotomized their "friend".

The market, for consumers, will flock to whatever products surge their neurotransmitters the most. This is the bull case for Anthropic (who are serving business with alternative success metrics) and bear for OpenAI who seem (with all their hiring of Meta execs) to want to go for the masses, who have no objective function for product selection beyond "I like it".