Post #654502
2026-03-03 15:13 UTC
Can thinking really be defined as a flat sequence of conversational tokens?
Feels like LLMs are missing something more chaotic where different principles could be pulling the chain-of-thought in different directions which could even be contradictory.
Not sure if this can recreated with several LLMs playing different roles for that internal "monologue". How does that conversation get resolved? What hierarchies can emerge and be dissolved?
Replies (1)
-
@bobmagicii@phpc.social 2026-03-03 16:25
@t_var_s@phpc.social its not a flat list though. there's levels and stuff, half of the tuning of an llm is just deciding how deep to let it consider the crap you said. thats where its fake intuition tricks people. f22 is nested in military planes nested in planes nested in vehicles. 1 setting and rng can be the difference between it being willing to talk about a fighter jet or being ultra confused about machines that can fly. like if it rng's this request "im not going deep on that path im quirky"'