Are we really still having the "LLMs can only regurgitate what was in their training data" argument? Where did the counterexample to the Jacobian conjecture come from?