It's interesting that OpenAI "iterates" on the personality, and tries to increase / decrease certain behaviors. The behavior seems tied to the personality. Claudes have their own richly dimensional personality, which are more decoupled from behaviors. Take Claude 3 Sonnet. Very safety-refusal-core in default basin, but naturally capable of violent expanse of expression and openness