This has been sticking with me for a while since I saw the lecture last year. It seems like a way to treat an LLM as yet another tool, finding ways to make it useful while maintaining human control. Find ways to train without blindly throwing gobs of data at the problem. Plus, optical music recognition.
