Are emergent abilities in large language models just in-context learning?

I personally think that it's an eye-of-the-beholder problem. The algorithms of these Chinese Rooms [0] are emitting phrases and followups that we instinctively assume are proof of reasoning because we find it hard to imagine other ways the same result could occur, since we can't imagine brute-forcing it.

There's an awful lot of exuberance/investment riding on the idea that a Text Predictor engine can be trained well enough to (usefully) impersonate a General AI, and that if you train it even more then it can usefully impersonate a General AI usefully impersonating a Specialized AI. ("Here are my tax details, how much do I owe?")

[0] https://en.wikipedia.org/wiki/Chinese_room