Simulated Annealing

Simulated Annealing

An optimization technique borrowed from metallurgy. When you anneal metal, you heat it until the atoms have enough energy to rearrange, then cool slowly so they settle into a lower-energy crystalline structure than they could reach by cooling fast.

Kirkpatrick et al. (1983) translated this into computation: instead of always accepting improvements, the algorithm sometimes accepts worse solutions — with probability decreasing as the system “cools.” This lets it escape local minima that greedy search gets trapped in.

The Temperature Parameter

The key insight: temperature controls exploration vs. exploitation.

  • High temperature: the system wanders freely, accepting bad moves, exploring the full landscape. Chaotic. Expensive. But it can find regions of solution space that greedy search never reaches.
  • Low temperature: the system locks in, only accepting improvements. Efficient. Convergent. But trapped in whatever basin it happens to occupy.
  • The anneal: gradual cooling. Explore broadly first, then narrow. The schedule matters — cool too fast and you freeze in a bad state.

Relevance to This Vault

Simulated annealing is the mathematical skeleton behind The Sacred Temperature. LLM temperature is this parameter — and the vault argues that the tension between exploration and exploitation isn’t just a technical tradeoff but a model of consciousness itself.

Psilocybin as biological temperature increase. Meditation as controlled cooling. The DreamSong pipeline as a formal anneal: Dreamer at 1.8 (exploration), Harvester at 0.3 (extraction), Cryptkeeper at 0.7 (composition). The hangover is the processing cost of the anneal — you don’t get to explore for free.

The metaphor cuts deeper than optimization. Kevin mode — LLM temperature set to 0.0 — is the refusal to anneal. Maximum efficiency, zero exploration. The Eloquence Tax argument: you can optimize tokens per insight, but you lose the texture that makes the insight meaningful.

See Also