I think the LLM problem mostly stems from an insufficiency in our own language when we describe "intelligence" or "knowledge".

[vv@solarpunk.moe]

I think the LLM problem mostly stems from an insufficiency in our own language when we describe "intelligence" or "knowledge". We generally think of intelligence as a singular concept that exists on a scale. Of course intelligence is more nuanced than that, and many folks have talked about how people have "different types of intelligence". But I think that there's even more to it than that.

To attempt to be more precise, I believe the information storage and retrieval process that LLMs undertake is accomplishing some sort of facet of cognition, maybe you could call it "linguistic sequence processing". But they do not have any other sort of intelligence.

I think the problem we are having with our societal reaction to LLMs is that we haven't encountered something that exists with only this capability, so we don't have a model for it and its limitations. We just expect it to act more and more like a human because that's the only thing that it is similar to.

Probably the best comparison I can think of is early chemistry and how it relates to alchemy. Certain things could be done that looked incredible and you could extrapolate given those results that you could do anything with this technology, and it would even make the concept of value obsolete because you could turn lead into gold.

Or photography. You can capture a perfectly realistic depiction of reality onto a piece of paper. That sounds like it would make art obsolete, it sounds like it might be a threat to reality if it can exist on paper as well as in real life. People might start only looking at photos instead of going out and experiencing real life.

But technology generally isn't that simple, and hopefully we will eventually have a more nuanced understanding of its limitations and stop trying to use it everywhere as if it was, or ever will be, an "AGI", and scale it down so it isn't so resource intensive. I'm convinced that there's no need for it to have as much training data as it does, if it's used more effectively as a small component of a system. I also really don't think it will significantly replace people's jobs, once people realize how limited it is (and that's already starting to happen as employers realize that firing people for AI doesn't work). Hopefully, it can instead be a tool that helps us interact with computers more effectively. But that will take a pretty big change in how we perceive this technology as a society.