
When we train AI there are two broad directions. One is pre-training and similar ways of piling more and more of the world’s knowledge into one place — but the downside is the original face of things gets blurred. You don’t know where a hallucination came from; provenance disappears. That’s a big problem.