Gary Marcus argues for neurosymbolic hybrid AI approaches that combine LLM's for pattern recognition with symbolic reasoning for planning and generalization. I've been on this page (as were most AI researchers then) since the 1980's.
https://garymarcus.substack.com/p/even-more-good-news-for-the-future
But why should I want to aid the supercop predictive thought policing future?
Erase what I just said. Keep on doing it wrong.
Marcus presents papers showing that traditional approaches work better for very targeted problems, including protein folding and databases.
People forget about databases and think LLM must be the best for all sorts of problems because it can "think" across all boundaries.
No comments:
Post a Comment