What we need is steering so that we can have models with different personalities, not just different prompts (because context is subject to forgetting), but this will never happen with closed-weight models, I'm not sure if it's even feasible at scale.

Yet another reason why the future is open weight.