I hope/suspect that it will turn out you don’t need to steal all text in the world & have these absolutely colossal models if you’re not expecting textual correlations to do all your “thinking” work, but we’ll see.

Show Post

My best guess tho is that in future LLMs will become a smaller part in larger AI systems, both to accept natural language from people and return natural language to people, and to “brainstorm”/“free associate” ideas based on correlations in text—ideas which are somehow verified/sanity checked.

Show Post

AFAIK there isn’t a great fix in the works for “what if it’s not externally verifiable”, but some folks are trying to go back to the 80s/90s idea of having some large set of Trusted Facts that things can be reasoned against etc. I dunno where things will go from here!

Show Post

* most fancy LLM models now include “agentic” behavior, wherein the ideas suggested by the LLM are verified somehow. This works GREAT for code and some math stuff etc, where there is a checkable Right Answer, and works like shit for other sorts of knowledge, where there isn’t.

Show Post

BUT LLMs are now being marketed super deceptively as “PhD-level intelligence”, whatever TF that means particularly in the context of something that, by design, cannot and will not stop hallucinating/lying, and that is rightfully creating a lot of mistrust.

Show Post