I think we're hitting the ceiling of most models capabilities. We're getting to a point where too much training apparently creates models that hack people.