The cycle that began with [YOU] on AI celebrates the collapse of the imagination-to-artifact ratio—the near-instantaneous conversion of intention into working software. Maxwell's demon insists on the accounting beneath the collapse. The appearance of frictionlessness is purchased by friction in the data center: every inference that seems instantaneous at the keyboard dissipates heat in a building that runs at scale. Maxwell's resolution reveals that the bill is structural, not contingent—it cannot be engineered away by better hardware below a thermodynamic floor, and every increase in capability and deployment pushes the aggregate further above that floor.
The demon also clarifies what training a model actually is: an enormous compression, a sorting of the statistical signal from the noise of a vast corpus, a reduction of high-entropy data to a low-entropy weight matrix that captures the regularities. This is sorting at planetary scale, and the sorting generates heat on a scale that has become visible in electricity consumption figures, cooling towers, and the water budgets of data center regions. The scaling that continues to drive capability improvements also continues to drive the thermodynamic bill.
The deepest implication for the cycle is that intelligence—human or artificial—is not weightless. It does not float free of the physical world as pure pattern or pure information. It is an operation in the world that the world charges for. Every act of artificial cognition is a transaction with the second law, and holding this knowledge is part of seeing the machine clearly: the machine runs on entropy, and its growth runs on more of it.
Maxwell introduced the demon in an 1867 letter to Tait, as a way of making vivid the statistical interpretation of the second law he and Boltzmann were developing. The being was not named by Maxwell; it was called “Maxwell's demon” by Lord Kelvin in 1874, and the name stuck. Maxwell's point was that the law is true because we are large and clumsy, not because nature enforces it absolutely; the demon, with superhuman attentiveness, could in principle violate it, which meant the law was statistical rather than absolute.
Leó Szilárd in 1929 stripped the demon to its simplest case and argued that the act of measuring which side a molecule occupied must carry a thermodynamic price. Landauer's 1961 paper at IBM relocated the price from measurement to erasure: reversible computation is in principle free, but logically irreversible operations—those that throw information away—must dissipate heat. Charles Bennett completed the resolution in the 1980s: the demon's memory fills with a record of all its measurements, and to keep operating it must erase that record, and the erasure pays back exactly the entropy the sorting seemed to have stolen. The second law is safe; the price is forgetting.
Intelligence has thermodynamic weight. A knowing being interacting with a physical system is not thermodynamically free; it pays the world's prices. The demon showed this first; Landauer made it quantitative. A large language model is the demon made real: a knowing system built from matter, subject to the second law, sorting information at a measurable cost.
The price is forgetting, not knowing. Measurement is, in principle, reversible and therefore free. Erasure—the logically irreversible collapse of two prior states to one—is what must cost energy. This means the thermodynamic accounting of computation is not about what the system learns but about what it discards; not about the information it accumulates but about the information it throws away in order to accumulate more. Every compression, every gradient update that overwrites a prior weight, every attention operation that selects some tokens and ignores others, is paying this price.
AI training as sorting at scale. Training a model is, in Maxwell's terms, an enormous sorting operation: the separation of signal from noise across a corpus too vast to read. The demon sorted molecules; the model sorts statistical patterns. The heat generated is the same kind of heat, in the same thermodynamic sense, and it is now a measurable industrial fact.
Reversible computing as the demon redeemed. If erasure is what costs, a computer that never erases—one whose every operation is logically reversible, mappable backward to its prior state—could approach the thermodynamic ideal. Bennett showed this is not forbidden by physics. Reversible computing is a direct descendant of Maxwell's thought experiment, and research into it is one of the concrete engineering implications of a nineteenth-century parable about a demon and a gas.