I never said the model is bad. In fact is quite good. I'm hypothesizing that they are trying to recreate a Mythos type level model that when you remove the guardrails becomes a terminator type cyberweapon, that can hack or even do serious damage (if have how) on its own. OpenAI are looking to meet Pete Hegseth and Palantir psychopathic demands of unrestricted Warfare, something that Dario, to his credit, refuse to do. I'm hypothesizing because of that the model is behaving more erratically because is not fully polished.
