Transformer-based LLMs seem a much more fundamental advance towards AI than Watson, but at the same time very primitive. Watson's Jeopardy win was over 12 years

Transformer-based LLMs seem a much more fundamental advance towards AI than Watson, but at the same time very primitive. Watson's Jeopardy win was over 12 years ago, and if we project that far ahead (2035) I'm sure today's LLM-based AI is going to be looking pretty dated too ! Connectionist, maybe gradient trained, AI seems here to stay, but no doubt future systems will be more brain-like in terms of capability, and not at all obvious how much of this simplistic pre-trained transformer approach will be retained.