My take since day one:

(1) Model capabilities will plateau as training data is exhausted. Some additional gains will be possible by better training, better architectures, more compute, longer context windows or "infinite" context architectures, etc., but there are limits here.

(2) Training on synthetic data beyond a very limited amount will result in overfitting because there is no new information. To some extent you could train models on each other, but that's just an indirect way to consolidate models. Beyond consolidation you'll plateau.

(3) There will be no "takeoff" scenario -- this is sci-fi (in the pejorative sense) because you can't exceed available information. There is no magic way that a brain in a vat can innovate beyond available training data. This includes for humans -- a brain in a vat would quickly go mad and then spiral into a coma-like state. The idea of AI running away is the information-theoretic equivalent of a perpetual motion machine and is impossible. Yudkowski and the rest of the people afraid of this are crackpots, and so are the hype-mongers betting on it.

So I agree that LLMs are real and useful, but the hype and bubble are starting to plateau. The bubble is predicated on the idea that you can just keep going forever.