@glyph my hypothesis on that is that, by virtue of literally being encodings of lexical fields and semantic proximity, and by virtue of their response being the logical continuation of the user's input, LLMs statistically pick up on and amplify subtle tendencies / biases in the user: if you feed it input that uses vocabulary and idioms semantically linked to low self-esteem, the model will more likely compute a reply with similar undertones, feeding said emotion. they amplify whatever emotion you put in, even accidentally.
(thread here: https://tech.lgbt/@nicuveo/116210599322080105 )
N
nicuveo@tech.lgbt
@nicuveo@tech.lgbt
Posts
-
I wish I could recommend this piece more, because it makes a bunch of great points, but the "normal technology" case feels misleading to me.