So you're absolutely certain, say after a few hundred rounds of self improvement, that the AI will always obey human instructions? It doesn't even always do what you asked now...