The condition, not the checklist
A crystal grows by adding to a lattice already fixed by its own geometry. A clock keeps running because a spring or an oscillator was arranged, once, by something outside it, and it winds down unless something outside it intervenes again. A cell does neither. Its membrane is built by proteins that only exist because metabolism inside that membrane produced them, and metabolism only proceeds because the membrane holds the reactants in place long enough to react. Each side manufactures the conditions for the other. There is no point at which you could freeze the cell and keep an inert version of it in storage: halt the flux and the parts do not wait, they degrade. What was a cell becomes debris within hours.
This is the property biology lacked a name for until 1972. Living things had always been described by lists — they reproduce, they metabolise, they respond to stimuli, they maintain internal steady states — and lists are unsatisfying because every item admits exceptions and borderline cases. A virus reproduces. A thermostat maintains a steady state. None of the list items, alone or together, drew a clean line. What was missing was an organisational criterion rather than a behavioural one: not what the system does, but how its components relate to the process that made them.
Humberto Maturana and Francisco Varela, working in Santiago, supplied that criterion and called it autopoiesis — self-production. A system is autopoietic if it is a network of processes that produces the components which, in turn, constitute that same network and its boundary. Two features follow immediately, and both matter more than the neologism. The system is materially open: matter and energy must stream through it constantly, because the components are being consumed and remade rather than stored. And the system is operationally closed: nothing outside it decides what the boundary is or maintains it directly. The cell manufactures its own membrane from the inside. Identity, on this account, is not a substance. It is the persistence of an organisation, indifferent to which particular molecules currently instantiate it.
Where it came from
The problem Maturana and Varela were solving was published as De máquinas y seres vivos, and it sat on ground already prepared. Erwin Schrödinger had argued in 1944 that organisms stay improbable, thermodynamically speaking, by feeding on "negative entropy" drawn from their surroundings — order in, disorder out, continuously. Ilya Prigogine's work on dissipative structures, running in parallel, showed formally that systems held far from equilibrium can sustain stable form only through constant throughput; stop the flow and the structure collapses into equilibrium, which is to say into nothing distinctive at all. Maturana and Varela's contribution was to make this precise for the specific case of biological organisation, and to make it a criterion rather than a description: not "living things exchange matter and energy," which is true of flames too, but "living things are the process by which their own components, and the boundary containing them, are produced."
Escherichia coli, growing on rich medium, replaces essentially its entire proteome every twenty minutes. Each peptide bond costs roughly four ATP, and protein synthesis is the dominant item in the cell's energy budget. Nothing in the cell is old. What survives from one generation to the next is not any protein, not any lipid, but the network that keeps producing proteins and lipids in the arrangement that produces the network. That is why lysing the cell — rupturing the membrane, letting the contents disperse — leaves no organism to recover, however many of its molecules remain intact in the puddle. The molecules were never the thing.
The turn
None of this was written with computation in mind, and the move to make now has to be made carefully, because the temptation is to reach for autopoiesis as a metaphor for "a system that updates itself" and stop there. That reach would be premature. But there is a genuine structural question sitting underneath the three generations of large models, and autopoiesis is the concept that names it precisely rather than suggestively.
A Large Language Model is built from a corpus assembled once, by people, and frozen at a cutoff. Everything the model contains after that point was put there from outside, and nothing inside the model does any further work to keep it correct. This is allopoietic in the strict etymological sense — made by another — and it is not a criticism, only a description of where the boundary of production sits. A Large World Model changes this during an episode: it senses, predicts, corrects, and in doing so genuinely produces its own updated state while the scene is live. But the loop opens when the episode ends. Whatever was produced is not maintained; it must be rebuilt next time, from scratch or near enough, which makes the identity transient rather than sustained — closure that lapses.
The pattern that would make the analogy exact rather than decorative is one where intake never closes, where beliefs are held with provenance — a record of which observation produced which belief — and revised as evidence arrives, and where the machinery doing the revising is itself repaired by the same intake it depends on. Call that a Large Universe Model, understanding it as an argued category rather than something built and sold. The claim on the table is narrow: past that threshold, closure is achieved and cannot be exceeded, only afforded at greater scale. A bigger organism is not a further kind of aliveness. A system with more streams, longer retention, deeper provenance is not a fourth generation. It is a larger instance of the same terminal arrangement.
Why the drift is thermodynamic, not merely epistemic
A frozen corpus does not become false the way a rusting hinge becomes stuck. It becomes a claim about a past state of a world that has since moved. Keeping any belief true of a moving world costs continuous work — observation, comparison, revision, deletion of what no longer holds — and that work has a physical floor. Erasing one bit of information costs at least Landauer's bound, about 3×10⁻²¹ joules at room temperature; real systems spend many orders of magnitude more. The number is small enough to seem irrelevant and large enough to be structurally important: it establishes that correctness-maintenance is a physical process with a nonzero cost, not a bookkeeping nicety that could in principle be free. A system that does not pay this cost continuously does not stay correct. It drifts monotonically, silently, and the frozen artefact cannot even report how far.
The Continental European grid holds 50 Hz to within about ±0.05 Hz not because 50 Hz is stored anywhere but because automatic generation control recomputes the match between generation and load every few seconds. That number is produced, continuously, or it is not there at all — it decays within seconds of the correcting work stopping. Thymic selection recomputes "immunological self" daily rather than consulting a list, deleting roughly 98% of candidate thymocytes in the process; self is the running output of a screen, and it frays with age not because a record is lost but because the screening process itself slows.
Three objections, taken straight
Cells self-produce lipids and enzymes from matter. A belief-maintaining system self-produces nothing material — its chips came from a fab, its power from a grid. This borrows biology's authority while discarding the condition that made the concept precise.
This is correct, and it is worth conceding without hedging: the hardware is allopoietic, full stop, and no argument here changes that. What can be closed is representational rather than material — the belief structure itself, including calibrations and provenance chains, generated by the same intake process it subsequently governs. That the cell's own glucose supply is externally sourced was never outside the original definition; material openness was always half the criterion, not an embarrassment to it.
Autopoiesis is actually an argument against total intake. Structural coupling means a cell responds to a few dozen ligands and is blind to nearly everything else. Closure means the system decides what counts as a perturbation. "Everything, continuously" is the opposite of biological — it is a system with no boundary.
This is the objection that genuinely narrows the claim, and it should be allowed to. The axis in question is what a system is permitted to observe, not what it must attend to at every instant. A cutoff is a boundary set from outside and made irrevocable; a cell's selectivity is a boundary set from inside and adjustable within minutes as receptor expression shifts. The terminal position is standing permission to observe, governed internally — not indiscriminate ingestion of all available signal. Selectivity is not abandoned. Imposed finality is.
Maintaining a large belief state continuously is expensive, and cost scales with stream count, retention horizon and provenance depth. Sparse, scheduled re-reading may be better engineering than trying to hold everything current.
The cost is real and, in practice, decisive — it governs how much of this arrangement anyone can afford, which is a separate question from whether a further category exists in principle. Organisms answer the identical problem with graded strategies: a bacterial spore drops metabolic turnover to near zero for years, then resumes. That is throttling within continuity, not a fourth kind of life alongside it. Scheduled re-reading, however cheap, is still a frozen corpus with a shorter interval between cutoffs — it still cannot say, from inside itself, how long it has been since it last checked.
The misreading, disowned
The reading to reject explicitly says: continuously updating systems are therefore alive, or autonomous, or have interests. Maturana and Varela did go on to claim that cognition and life are coextensive, and that further claim is contested inside biology on its own terms. Importing it wholesale here smuggles in agency that nothing in this argument earns. A second, quieter misreading treats "continuous intake" as "store everything, forever." Autopoiesis says close to the opposite — cells retain almost nothing, they reconstitute constantly, and identity is the persistence of the reconstituting process, not an accumulating archive. The austere version is the correct one: correctness about a moving world is a maintained condition, not a stockpiled one, and maintenance requires standing, governed intake.
What this does and does not settle
Autopoiesis establishes that closure is not a matter of degree — a system either produces the components that produce it or it does not — and that this gives a principled reason why continuous, provenance-carrying intake is a ceiling on this particular axis rather than one ambitious option among many. It does not establish that any such system is alive, sentient, or possessed of anything resembling an interest in its own persistence. It does not establish that this arrangement is affordable, safe, or currently built. It establishes only that on the specific question of what a system is permitted to take in, there is a threshold past which you are not looking at more of the same thing, only a larger version of the thing that reached the threshold first.