Temperature Parameter
The temperature parameter sets the degree of randomness the model allows in token selection. Low temperature: the model takes the safest, most predictable path, selecting the statistically most probable next token. High temperature: the model reaches for less probable words, less obvious constructions, less expected turns. Engineers call this "creativity," with the same skeptical quotes, though what it actually means is that the model is less anchored to what it already knows. The Peirce volume reads the parameter through his cosmological doctrine of tychism: the stochastic variation the temperature introduces is a structural analogue of the chance-element Peirce identified as essential to creative thought. Variation, however, is not sufficient for creativity — selection is also required, and selection remains a human contribution.
In the AI Story
At low temperature, the model's output is maximally determined by its training data. The output is the safest extrapolation of observed patterns. At high temperature, the model introduces randomness that produces outputs not determined by any prior state — unpredictable, in the strict mathematical sense, from the inputs.
Edo Segal describes the high-temperature regime in The Orange Pill as "the machine getting stoned" — a vivid but philosophically incomplete characterization. The stochastic variation is real, not pseudo-random. It is genuine indeterminacy introduced at the point of token selection, and it produces outputs that are, in Peirce's cosmological vocabulary, genuinely tychistic.
But tychism without selection produces noise. The creativity Peirce analyzed requires both variation (the unpredictable deviation) and selection (the evaluative judgment that retains illuminating variations and discards random ones). The machine has the variation. It lacks the selection — at least, the kind grounded in confrontation with reality.
The human-AI collaboration therefore exhibits a division of labor mapping onto Peirce's evolutionary cosmology. The machine supplies tychism; the human supplies selective pressure. Neither alone performs the complete creative operation. The machine generates stochastic variation; the human evaluates against the standards of experience, domain knowledge, and the specific demands of inquiry. The productivity of the collaboration depends on the quality of the human's selection — which depends on capacities the machine cannot replicate.
Origin
The temperature parameter derives from Boltzmann distributions in statistical mechanics, adapted for language models to control the concentration of the probability distribution over next tokens.
The Peirce volume's reading of the parameter through tychism is its original contribution — connecting an engineering control with Peirce's cosmological doctrine of genuine indeterminacy.
Key Ideas
Genuine stochastic variation. Not pseudo-randomness masking determinism — real indeterminacy at token selection.
Structural analogue of tychism. The variation Peirce identified as essential to creative thought, now instantiated in silicon.
Variation without selection. The parameter provides material for novelty but no mechanism for distinguishing illuminating variation from noise.
Human selection required. Creativity in Peirce's full sense requires evaluative judgment the machine cannot supply.
Appears in the Orange Pill Cycle
Further reading
- David Ackley, Geoffrey Hinton, Terrence Sejnowski, "A Learning Algorithm for Boltzmann Machines" (1985)
- Ilya Sutskever et al., papers on sampling temperature in neural language models
- Andrew Reynolds, Peirce's Scientific Metaphysics (Vanderbilt, 2002)
- Edo Segal, The Orange Pill (2026), Chapter 4