I mixed up physical and extended theses. The physical one talks only about computability ("at all"). The extended one requires at most polynomial slowdown ("efficiently").

We are interested in the latter. Exponentially slow AI is of no use.