Comment by kgeist

It converges to "almost deterministic" on highly predictable outputs (i.e. code) with the right sampling params (say, you only sample the most probable token without randomness/high temperature) and with self-correction loops