CONCEPT
Gibbs Temperature (in AI)
The single most direct inheritance of Gibbs’s statistical mechanics in deployed AI systems: the temperature parameter governing how tightly a language model’s output distribution concentrates on its most probable tokens, doing in software exactly what temperature does in a canonical ensemble—controlling how sharply probability clusters around the lowest-energy states.
When
Gibbs defined the canonical ensemble, the temperature was not a decorative parameter. It did real physical work: at low temperature, the probability distribution over microstates concentrates sharply on the lowest-energy configurations, and the system behaves predictably, reliably, in its ground state. At high temperature, the distribution spreads across many states, the system ranges freely, and behavior becomes less predictable but richer in variety. This is the exact behavior of a language model’s sampling temperature, which users of modern AI tools encounter as a slider between “focused” and “creative.” Lower the temperature, and the
large language model concentrates probability on its few most probable next tokens, producing safe, deterministic, predictable—and often repetitive—text. Raise it, and the distribution flattens across less probable continuations, producing surprise, variety, and sometimes nonsense. The user adjusting this dial is, without knowing it, operating the control that Gibbs placed on the