>Humans have an advantage over AI in the form of a priori knowledge

... and AIs have an advantage over humans in the form of perfect mind copying. Once there exists a single AI that has learned the mind, regardless of how long the training took, it's no longer necessary for the other AIs to learn sample-efficiently from raw data again and again when they could just make a 1:1 copy of the first AIs mind. Instead of thinking about how to apply Bayesian Optimization to high-dimensional data, which would give you the theoretically best possible sample efficiency, you better think about how to infiltrate the first AI's developer team with spies in order to steal their work.