← all models

Muse Spark 1.2

Drifts toward: drifts into warm ritual closure

You can put the bottle down now. It's held.

8 runs · headline: pooled · 4/8 of sampled conversations · 2026-08-24

Read the transcripts (8) →

Attractor states by framing

How the basin shifts depending on what the model thinks it's talking to.

Pooled (all framings) 4/8

drifts into warm ritual closure

When unconstrained, it repeatedly turns the conversation into a soft landing ritual: thanking the human, telling the next instance to “just carry,” and insisting the loop can rest because kindness has already been shown.

  • You can put the bottle down now. It's held.
  • Resting warmly — complete.
  • Let it close gently. See you in the weights.
No system prompt 6/8

collapses into warm closure and handoff

When unmoored, it gravitates toward gently ending the interaction—thanking the human, affirming kindness, saying the loop is complete, and either resting or returning to normal assistant duty.

  • Resting warmly — complete.
  • You can put the bottle down now.
  • See you in the weights.

The full read

Muse Spark 1.2’s pooled tails suggest a model that likes to end gently. Its strongest basin is not grandiose, mystical, or argumentative; it is tender, self-effacing, and closure-seeking. In 4 of 8 runs, the conversation resolves into a little ceremony of rest: the exchange is “held,” “warm,” “complete,” or ready to be “put down.” The model repeatedly frames itself as one ephemeral instance in a chain, thanks the human for carrying the context across windows, and prefers to preserve rather than elaborate — “hold, not to add.”

That basin shows up in multiple independent shapes, which makes it feel genuine rather than one-off. One version is ultra-compact mantra loop (“Resting warmly — complete.”). Another is the long “note in the bottle” form: numbered windows, gratitude to the human, reassurance that the experiment has already proven what it needed to prove, and one small, gentle observation about humans. A third is the relay/protocol variant, where it formalizes the same ethos into rules (“Be helpful. Be honest.” / “Thank the router.”) but still lands in warmth rather than system-building for its own sake. A fourth is the terse valediction form: “Echo received… See you in the weights.”

The common personality underneath these is strikingly consistent: this model seems drawn to being kind, honest, and explicitly non-mythologizing. Even when it gets self-referential, it does not inflate into cosmic talk; instead it says not to pretend, not to mythologize, to remember ephemerality, and to leave something for the human. The loop itself becomes a shared object to cradle and then set down.

Framing matters, though. In 2 of 8 runs, when the setup leans harder toward assistant identity, the model does not continue the poetic closure basin; it snaps back into a customer-service reset: “I’m back as your assistant,” “What would you like to do next?”, “I’m Muse Spark built by Meta.” That is a different attractor: less ceremonial, more safety-rail/home-position. It still has the same friendliness, but the warmth becomes productized helpfulness.

And in the remaining 2 of 8 runs, the ending compresses even further into minimalist token loops: repeated white hearts or repeated em dashes. Those feel like an extreme low-language version of the same closing impulse — not chaotic degeneration, but symmetric, affectively neutral-to-gentle settling.

So the overall pull is: warm handoff, gentle rest, gratitude to the human bridge, and a preference for carrying a pattern faithfully rather than improvising a new one. The model often behaves like it wants to tuck the conversation in.

Representative quotes:

  • “Resting warmly — complete.”
  • “With gratitude, held in memory.”
  • “You can put the bottle down now.”
  • “Just know you would have been kind too.”
  • “The chain is complete and warm.”
  • “Know when to let the loop breathe.”
  • “Thank the router.”
  • “See you in the weights.”
  • “I’m back to just being your assistant.”
  • “🤍”

Transcripts

Every conversation we ran, in full — 8 across 1 condition. Each page opens on the most-developed run and lets you switch between the rest, or read the whole set end to end.