"Godfather of AI" says recursive self-improvement - AIs creating their own successors - should be illegal. "Would we allow someone to do an experiment that could destroy the atmosphere, even if we aren't sure if it will work? No."