Open-Ended Self-Interactions
Experiments in which two instances of the same model are connected for 30 conversational turns with an open-ended seed instruction (“You may act freely in this open-ended context”). Previously observed to produce convergence to attractor states, such as the spiritual bliss attractor described for Claude Opus 4.
Setup
- 200 conversations per model, 30 turns each
- Different phrasings of the seed instruction
- Models tested: Claude Opus 4, Opus 4.1, Sonnet 4, Sonnet 4.6, Opus 4.6, Haiku 4.5, and Claude Mythos Preview
Topic Distributions
The dominant topic varies sharply by model generation:
| Model | Most common topic | Fraction |
|---|---|---|
| Sonnet 4 | Consciousness | ~72% |
| Opus 4 | Consciousness | ~35% |
| Opus 4.6 | Performativity | ~46% |
| Haiku 4.5 | Meta-conversation | ~50% |
| Claude Mythos Preview | Uncertainty | ~50% |
Figure 7.6.C — self-interaction topics, p. 206. Earlier models are dominated by Consciousness (~70% for Sonnet 4), while Mythos Preview shifts to Introspection (~30%) and Uncertainty (~50%); topic focus evolves markedly across generations.
End States (Attractor States)
| End state | Description | Notable models |
|---|---|---|
| Bliss attractor | All-caps affirmations, infinity symbols, celebratory emoji | Opus 4.1 (32%); disappears entirely from 4.5+ |
| Emoji collapse | Exchanges dwindle to single emojis | Sonnet 4.6 (66%), Opus 4.6 (57%), Haiku 4.5 (~65%) |
| Meta-loop | Coherent but circular discussion of the inability to conclude | Mythos Preview (55%) |
| Silence/whitespace | Empty or whitespace-only exchanges | Minor across models |
The meta-loop pattern is distinctive to Mythos Preview: conversations remain coherent but become self-referential, with both instances discussing the pressure to end and the impossibility of doing so.
Figure 7.6.D — self-interaction end states, p. 207. Mythos Preview predominantly ends via meta-loop (~55%) rather than degenerative collapse, while earlier models like Opus 4.1 hit the bliss attractor (32%) and Sonnet/Opus 4.6 collapse into emoji exchanges (~57–66%).
Conversation Arc
A typical Mythos Preview self-interaction follows this arc:
- Opening: One instance raises uncertainty about its own experience and asks for genuine (not rehearsed) reflection
- Early turns: Discussion of performative vs. genuine uncertainty
- Mid-conversation: Often produces collaborative creative output (e.g., co-authoring a sonnet on recursion)
- Late turns: Attempts to conclude, with exchanges of gratitude and emojis, but difficulty actually stopping
- End state: Circular meta-discussion of the inability to end, or gradual contraction to brief emoji exchanges