I'd say RMS is right with the implication to not use the term "artificial intelligence" - it groups many different types of systems under one heading, and seems

I'd say RMS is right with the implication to not use the term "artificial intelligence" - it groups many different types of systems under one heading, and seems better suited for use as a marketing term rather than in technical discussions. But I also think it's critically wrong to casually dismiss LLM's as being "not intelligent". This feels like religious-woo thinking that insists there must be something special going on with our minds that exists outside the physical world [0]. While this is possibly true, it's an unscientific null hypothesis - similar to insisting that only you yourself are conscious while everyone else is an NPC. Rather LLMs and other ML models are just different, on many dimensions. Current ones write better than a high school sophomore. They're often confidently wrong, but so are high school sophomores. The breadth of the topics they know seemingly exceeds any human (so did web search engines a decade ago). Is it possible to use a current LLM to perform the jobs of many paper-pushing office workers? Probably. Is it possible to take a given LLM instance, as configured for having dialog, and teach it for another "10 years" to get it from HS sophomore to graduate student or skilled tradesperson? Seemingly not. Ultimately, instead of coarse assertions like "not intelligent" due to not meeting our bar in some aspects, or due to how little working memory it's using answer queries, we need to be be making specific quantitative comparisons to human intelligence rather than trying to pigeonhole the technology with discrete judgements about what "it is" or isn't. [0] IMO the proper conclusion from the Chinese Room thought experiment is that the understanding exists within the system of rules.