>> There is a large class of AI researchers who have been working on this problem for decades (and I am one of them). Which problem is that? It's not clear fro

>> There is a large class of AI researchers who have been working on this problem for decades (and I am one of them). Which problem is that? It's not clear from the context and I don't want to make uninformed assumptions. It is true that many people have worked on problems like language understanding, language modelling and language generation for many years. It is also true that most people who are today so excited about large language models don't know anything about all those years of research. My experience is that, at some point, there was a switch from computational linguistics (i.e. doing linguistics with the help of computers) to statistical natural language processing (i.e. trying to represent language with statistics). The idea behind the switch was that understanding human language, its mechanisms, its origins, and so on, is too hard, and we can make a lot of progress if we stop trying to do all that, and instead turn to learning statistics from large corpora (cf. "every time I fire a linguist..."). Of course, without good understanding of what language is, and how it works, it is very difficult to evaluate such progress, but the answer to that was obviously to establish new metrics, and then try to maximise performance according to those metrics; metrics such as BLEU or ROUGE and friends, that are suspiciously like the kind of metric that one would choose if one wanted to measure progress as the ability to learn statistics from a large corpus. So it's true to say that many people have worked on all sorts of problems to do with natural language and AI, but I'm still left wondering exactly what you mean by "this problem". For me the problem is that an entire field of research ("NLP") abandoned the search for answers to interesting scientific questions and turned instead to the production of toys, and to spectacle, presumably because that's where the money is and researchers must also pay them bills. And so now we have large language models that are better at reproducing language than earlier models, but we don't have the tools to evaluate or understand those models; and so people who aren't experts, but also people who are experts, can't separate the wheat from the chaff, and keep coming up with wild, senseless proclamations that could well have been generated by a language model themselves.