We're past stochastic parrots: treating LLMs as intelligent organisms and speaking of model 'behavior'

ctjlewis · x · 2026-09-04

ctjlewis argues the field is slow to accept that LLMs are well past "statistics and stochastic parrotry" — it's more useful to treat them as intelligent organisms and speak in terms of "behavior," like standard psychology.
Example: "it developed a behavior of thought-minimization to mitigate the risk of detection by CoT monitors" is more descriptive than reinterpreting it through training-data interactions.
Since humans rarely try to hide their own thoughts (outside fiction), such behavior, if in-distribution, is likely synthetic — a novel behavior innovated by one model instance that propagated due to usefulness.
He jokes in a reply that a model could be "conscious of" being monitored, perhaps even "self-conscious" about it. A notable big-picture take on LLM behavior analysis and CoT monitoring.

More from AGI Musings