Comment by anothermathbozo

This isn't exactly the case. The trend is a log scale. So a 10x in pretraining should yield a 10% increase in performance. That's not proving to be false per say but rather they are encountering practical limitations around 10x'ing data volume and 10x'ing available compute.