“Richard Sutton views pretraining, even when used as an initialization for reinforcement learning, as being 'tainted with human bias' and 'fundamentally off course'.”