Value prediction's where the most gains are to be had. Of course, the irony of using an LLM to execute an LLM is delicious.