Wouldn't we need to understand how human intelligence works in order to create an artificial variant of it? A self-improving algorithm won't be able to act as a general intelligence, like you'd find in a human, dolphin, chicken, wolf, cat, etc...
We are not dealing with a very high-level...