On the other hand, we’d been trying to do anything useful with natural language since the 50’s and had thoroughly failed.
That’s really not true. For instance, machine translation and spam detection (document classification) were getting really good by the late 2000s. Image recognition was great beginning the late 2010s.
What we’ve seen in the last few years (besides continual incremental improvements in already-existing solutions) is improvement in the application of generative tools. So far the uses cases of generative models appear to be violating copyright, cheating on homework, and producing even more search engine spam. It can also be somewhat useful as a search engine so long as you want your answer to be authoritatively worded but don’t care if it’s true or not.