Defender @defenderofbasic.bsky.social · Jan 30

if you peek inside the mind of language models, we can see that (1) fire is red (2) grass is green (3) apples are slightly more red than green (4) if grossness had a color, it would be way more green than red

4 likes 2 replies

?

Replies

C

Chris Schuck · Jan 30

I'm tech-illiterate so I guess I'm basically asking what you mean by semantic embeddings, in this context....

C

Chris Schuck · Jan 30

So is the idea that you get a visual of the LLM's aggregated amount of associations of a prompt with various words (based on how often the two were statistically correlated in its training material),such that you can compare the degrees of correlation visually)? Or does it go deeper than that?