This is using think time compute and reinforcement learning. I think this is going to plateau even faster than the initial LLM scaling though.