Glossary · Term

attractor

← all terms

Definition

Plain language

A settled-down state that a system keeps getting pulled back toward, no matter where it starts.

As stated in the literature

In dynamical systems, a state or set of states toward which a process evolves over time; used to describe the low-diversity endpoint that closed-loop multi-LLM conversations contract toward, and the equilibrium a recurrent model settles into.

Why it matters: It matters because it explains why some systems lose diversity and get stuck, which limits how useful long automated conversations or loops can be.

For example, a group of chatbots talking only to each other may keep drifting back to the same handful of phrases and ideas no matter how they started.

Heard on the show

“" And underneath all of it — discrete multiple-choice decisions, lightweight fine-tuning, single-step episodes, a synthetic world with an honest option deliberately built in as the safe attractor.”
Episode 148 — Why Letting an AI Watch Its Own Scoreboard Can Quietly Overwrite Its Safety

Mentioned in 5 episodes

  1. 148
    Why Letting an AI Watch Its Own Scoreboard Can Quietly Overwrite Its Safety
  2. 128
    How a Model Can Earn Full Reward and Still Resist Training
  3. 123
    Five Identical Worlds, One Swapped Model: What Happens When AI Agents Run for Fifteen Days
  4. 073
    When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving
  5. 041
    When the Iteration Teaches the Model to Skip the Iteration

Related terms