nucleation

THEORY-1240: Trained, instruction-tuned and preference-tuned chat models keep the nature of the idealized fiction machine: what they assert follows the narrative coherence of the frame, role or story the conversation sets up, not a representation of truth that stays fixed across frames

1 min read

Properties4
StatusProposed
Tagscognition, philosophy-of-language
SourceLIT-178 — Borges and AI
ExtendsTHEORY-547 — A language model idealized as continuing any text with a plausible next word from the collection of all plausible texts follows narrative coherence and not truth, so its false statements are confabulations within the story the context sets up; whether trained and aligned chat models keep that nature is an open empirical question the idealization does not settle

Graph View

Source

Bottou & Schölkopf (2023), Borges and AI, LIT-178, read in full in NOTE-102. The essay takes real LLMs to approximate its perfect language model, and says fine-tuning and RLHF prune it “against its nature” (§§1–2).

Where the evidence stands

What follows from the idealization is THEORY-547, Active, which this entry extends. The essay carries the idealization over to real systems by assertion; its only evidence is one Vicuna-13b transcript pair, and §2 closes by leaving open “whether alignment techniques can transmute one into the other”. The promote_when above is THEORY-547’s condition before the twenty-fourth status review promoted it, carried here unchanged, because it asks what would make the thesis true and not what the source says. Its sentence that the idealization needs no test refers to THEORY-547.

Backlinks