In the interview transcript, he seems aware that the field is doing RL, and he makes a compelling argument that bootstrapping isn’t as scalable as a purely RL trained AI would be.