I am not a fan of "AI." Ironically, I took a class in "AI" in 1983, and "neural networks" as the earliest versions of LLM's were called, were barely mentioned. We read and wrote programs in Lisp. I declined an offer to work in an AI lab in preference for more solid engineering of CAD/CAM, and later learned more about Neural Networks from an ACM lecture in 1986. I generally viewed them as for perception only. In 1992, I watched a PBS Nova presentation about the Cyc computer project, to compile all knowledge symbolically. Much work was being done in Texas. (Wikipedia doesn't show Cyc as having been updated since 2017.). That sort of thing was what I would have bet on.
I've always believed in symbolic reasoning to be essential, and now there are papers to prove it. LLM's can't solve simple problems because they lack symbolic reasoning and planning skills. LLM's are primarily good at pattern recognition, the original problem, since programs couldn't easily be written to understand things like speech or text. Symbolic reasoning can actually solve these problems, and uses orders of magnitude less energy to do so.
No comments:
Post a Comment