My working hypothesis here is that one of the most interesting ways to interact with an LLM is to give it operations on meaning rather than requests for text. Things like intersect two bodies of thought, project the result toward a third, distil to a paragraph. I have been testing this since gpt-3 in a telegram bot called Emerson, and it has gotten quite a lot better since.
It gives me the feeling of manipulating meaning like clay, of getting to new insights over bodies of knowledge I couldn't possibly hold in my head even if I tried.
The link is a writeup by Fable of the latest try to turn this into a methode, and our exploration on how that resonates with work done in different fields.
The part I'd be most curious to see torn apart is the claim that the verbs people use on activation vectors and the verbs cognitive scientists use on concepts are the same seven. Also, I am curious if anyone has had success (or failure) with similar methods.
My working hypothesis here is that one of the most interesting ways to interact with an LLM is to give it operations on meaning rather than requests for text. Things like intersect two bodies of thought, project the result toward a third, distil to a paragraph. I have been testing this since gpt-3 in a telegram bot called Emerson, and it has gotten quite a lot better since. It gives me the feeling of manipulating meaning like clay, of getting to new insights over bodies of knowledge I couldn't possibly hold in my head even if I tried.
The link is a writeup by Fable of the latest try to turn this into a methode, and our exploration on how that resonates with work done in different fields.
The part I'd be most curious to see torn apart is the claim that the verbs people use on activation vectors and the verbs cognitive scientists use on concepts are the same seven. Also, I am curious if anyone has had success (or failure) with similar methods.