> we don't know where the ceiling is.

The ceiling for current AI, while not provably known, can reasonably be upper bounded to human aggregate ability since these methods are limited to patterns in the training data. The big surprise was how many and sophisticated patterns were hiding in the training data (human written text). This current wave of AI progress is fueled by training data and compute in ”equal parts”. Since compute is cheaper, they’ve invested in more compute but failed scaling expectations since training data remained similarly sized.

Reaching super-intelligence through training data is paradoxical, because if it were known it wouldn’t be super-human. The other option is breaking out of the training data enclosure by relying on other methods. That may sound exciting but there’s no major progress I’m aware of that points that direction. It’s a little like being back to square one, before this hype cycle started. The smartest people seem to be focused on transformers, due to getting boatloads of money from companies or academia pushing them because of fomo.