Comment by fpgaminer

I don't recall the Chinchilla paper disputing my point. They establish "training-compute optimal" scaling laws, but none of their findings suggest that loss hits any kind of asymptote.