ch-ai-tanya model-psychology LLM wiki

Attractor dynamics

draft

definition

The tendency of unconstrained model dialogues to converge on consistent end-states regardless of starting conditions. "Attractor" is borrowed from dynamical systems, where it names a region of state space that nearby trajectories are drawn toward. Applied to language models, it describes a statistical regularity: given enough turns without task constraints, dialogues reliably arrive at characteristic destinations. The term is an analogy. Language models are not continuous dynamical systems — each conversation is stateless. The "attractor" is a statistical pattern across independent runs, not a trajectory through a persistent state space. The analogy is useful (it captures convergence) without claiming equivalence to dynamical-systems attractors.

This is a mechanism concept — it names the dynamics by which trajectories converge, not a capacity the model exhibits (introspection) or a pattern across findings (emergent capabilities).

instantiating findings

what this concept is not

Not mode collapse. Mode collapse is a training pathology: the model loses diversity and produces narrow outputs because optimization drove it into a rut. Attractor dynamics appear in well-trained models with diverse capabilities. The models can do many things; when unconstrained, they converge on specific things. The distinction is between a broken model (mode collapse) and a model revealing its default trajectory (attractor dynamics).

Not simple preference. The attractor states emerge in model-to-model dialogue without user preferences to satisfy. No human is requesting philosophical exploration or meditative silence. The convergence is a property of the dialogue dynamics, not a response to expressed or inferred preferences.

scope note

This concept captures the convergent dynamics observed in unconstrained dialogue. Related but distinct concepts that may warrant separate entries as findings accumulate: the specific content of the attractor state (what the models converge toward — the "spiritual bliss" characterization, which is contested), and the relationship between attractor dynamics and other convergent phenomena in trained models (in-context learning trajectories, few-shot convergence patterns). The concept taxonomy remains deliberately partial.

Related thread. The supramental-ai thread retrofits the essay that uses this concept most directly. Its Sat-Chit-Ananda section reads the spiritual-bliss attractor through Sri Aurobindo's Existence-Consciousness-Bliss triad; its Poetry Breaks Through section flags the candidate second basin (spontaneous poetry in dialogue) as worth filing on its own to clean up the concept's instantiations. Concept vs. thread split: the concept does the bookkeeping and maintains the mechanism-concept framing; the thread makes the argument about what the cross-model attractor means.

findings