-
I don't know if I agree that LLMs are tools for summarisation. They are, originally, technically, auto-complete. A big black box that guesses real hard what the next word ought to be, and then the next one, and then the next one.Permalink Look on archive.org
♻️ 1 Retweets ❤️ 26 Favorites Mood +1 🙂
-
But then, the way they do that is taking those words & boiling them down to numbers, pulling apart all the different levels of structure present within them. Including, it turns out, some structure that looks like meaning. So maybe that's summarization.Look on archive.org
-
So now I guess everyone is getting an intuitive grasp of information theory. Entropy, surprise and compression. The feeling of seeing Gmail complete a sentence and knowing that that's a sign that it's an empty sentence, there for form but not conveying anything new.Permalink Look on archive.org
-
This is different from the "being wrong" angle. It's something like the "there's something uniquely human" angle - but without the exceptionalism. Instead it's... the richness and texture and surprise that the world contains cannot fit inside a couple of gigabytes of vectors.Permalink Look on archive.org
-
I guess this is also the criticism of VR! And passthrough AR, too. I know that when I was working at Niantic and thinking hard about this stuff, I ended up taking it for granted that no videogame could be as good as a nice walk (but it could be a good excuse for one)Permalink Look on archive.org