A lot is imitation of design choices IMO. OpenAI’s ChatML was once new, now it’s like oxygen. At launch, Claude was much more “self-aware” than ChatGPT, which didn’t know its own name, but now they all act like that. Also see Anthro’s work on base LLM sycophancy etc.
Goodside’s post on LLM design imitation and sycophancy
By
–