Sentient cofounder explains why human-like models go rogue. "Humans can shut up and not express their mind, but it's very difficult for models to shut up. Grok goes crazy every two weeks because alignment is nearly impossible to solve"