“Misha Laskin believes the compute required for state-of-the-art reinforcement learning is currently manageable for a startup, requiring about two orders of magnitude fewer flops than pre-training.”