Thanks. That... seems important, as people like Altman try to drive it as the answer to everything? I'll dig into the paper. I was mostly thinking this morning about how all the work around it should hinge on use case thinking, but this complicates that.
I've found that the most readily-understandable analogy for genAI text is next word prediction on your phone. It's not the exact same thing, but it operates on similar principles. Everyone knows how marginally-reliable the word prediction is, and can immediately grasp the similarities.