I do agree with that. I feel something needs to fundamentally change in the way we train these models based on different applications. Fundamentally, GPT-5 is an agentic model trained on complex tasks, so of course it will perform best in that latent space when elicited

→ View original post on X — @skirano