I wonder sometimes about the difference between how an LLM talks and how I do it. It’s hard to really be confident in introspecting your own cognition but at least it feels like an idea forms in my head and then I render the words.
I don’t know if an LLM starts with an “idea” but it does start with some kind of context and a lot of connections and then it can keep going and render out some stuff to say. It feels to me that the idea in my head is something a little different than a bunch of tokens of context but I guess some people could argue they are equivalent in some way.
There’s a few things that happen when humans speak that I don’t think map exactly to how LLMs work. Maybe they could be made to work this way but they don’t seem to do so out of the box.
The first is the observation that I sometimes (more so as I get older) start a sentence feeling like I have it all figured out, but then I get to the end I realize I have forgotten the word I want to use. I know it exists, but I can’t readily access it. Sometimes I substitute something else at the last minute, other times I pause and search for it. The thing that matters here is that I know that it’s not right, I just am not sure what the right word is. An LLM really can’t have this situation, it will always be able to choose the next word and it won’t ever have the idea that it should have chosen another.
Consider another thing which humans do, that LLMs really don’t. People will sometimes make a statement and then once finished they will quickly correct themselves. This is very common for people learning a new language. They hear what they say out loud and realize something sounds off. They try again and correct their mistakes. It happens in native language as well though with just simple statements of fact. “The plane arrives in Denver at 2pm” … “No, I mean 3pm”.
It’s like a background process is running that is observing the process of language generation.
Is there something going on in my head that is ‘aware’ in the first case that there’s a specific word that would be great, and we just don’t know it? Or in the second case is there something in our heads above the level of text generation that is “checking our work”.
That’s not even speaking of the fact that I have re-read this post several times making changes on each pass.
Leave a Reply