What emergent novelty claims
Some things in the world are not on any list drawn up beforehand. That is the whole of the thesis, and it is worth stating plainly before dressing it in examples. Chemistry was not implicit in physics in the sense of being readable off a table of forces; bonding, valency, the specific stubbornness of water, arrived as a new register of description once atoms arranged themselves a certain way. Life was not implicit in chemistry. Mind was not implicit in life. Each transition produced properties, entities, sometimes laws, that no inventory made at the earlier stage contained or implied. The strong form of the claim says these novelties are ontologically real additions to the world, not just new words for old furniture. The weak form asks for less: only that no enumeration made in advance, however careful, could have listed them. Most of what follows needs only the weak form.
It matters to distinguish novelty from randomness, because the two are often confused and the confusion does real damage. A die throw is random: the outcomes were enumerated before the throw, and one of them turned up. Novelty is different in kind. It is the arrival of an outcome for which no slot existed. The die-throw model of uncertainty, however sophisticated its distributions, is a poor guide to a world that periodically hands you a category the model's alphabet cannot spell.
This is not a claim that everything is unpredictable or that pattern-finding is futile. Most of the world behaves regularly enough that yesterday's categories cover tomorrow's cases. The claim is narrower and more stubborn: somewhere in the tail of experience, the categories run out before the events do, and no amount of foresight fixes that in advance, because foresight is exactly what an enumeration is.
Where the idea comes from
The lineage is nineteenth-century British philosophy of science, wrestling with a problem that had grown too large for either of its available answers. Vitalism, the idea that living matter carries some additional non-physical ingredient, had become untenable by the 1840s; nothing under a microscope needed it. But strict mechanism, the idea that everything reduces without remainder to the motions of particles already catalogued, could not account for why hydrogen and oxygen produce something with the specific, surprising behaviour of water, let alone why arrangements of cells produce a nervous system that notices things. John Stuart Mill, in his System of Logic (1843), named the gap with the term "heteropathic" laws — laws of a compound that are not the sum of the laws governing its parts. George Henry Lewes coined the word "emergent" itself in 1875, distinguishing emergent effects from merely resultant ones.
The idea matured into a school in the 1920s. Samuel Alexander's Space, Time and Deity (1920), C. Lloyd Morgan's Emergent Evolution (1923), and C.D. Broad's The Mind and Its Place in Nature (1925) worked out, with considerable care, a middle path between spooky vitalism and flat reduction: new properties arise lawfully from arrangements of old ones, and once arisen, they are not deducible from the arrangement alone, even in principle, even by a calculator with unlimited time. Broad in particular gave the clearest formulation: a novel property is one that could not have been predicted from complete knowledge of the parts and their laws, prior to observing the whole. That is the definition still in service.
The turn
Every position on the intake axis rests on an enumeration, made once, of what counts as data. That single observation is what connects a century-old debate about chemical bonding to a taxonomy of machine intelligence, and the connection is not decorative. It is the same mechanism at two different scales.
A Large Language Model enumerates twice over. It enumerates the documents gathered before some cutoff date, and, more restrictively, it enumerates the vocabulary of concepts those documents happen to contain. A kind of thing that arose afterwards is not merely absent from its training; there is no slot in its representation space shaped for it. Ask such a system about something that postdates its corpus and it will not report ignorance cleanly. It will interpolate, producing a plausible answer built from neighbouring, older concepts, because interpolation between known points is the one operation it is built to do superbly. The confidence is real; the referent is not.
A Large World Model does better, in one specific way. It is not confined to a fixed corpus; it senses a scene as it happens, so it can meet a genuinely novel kind face to face rather than at second hand through text. But it meets that novelty only through the channels and label sets someone fixed in advance — a camera tuned to visible wavelengths, a segmentation model trained on a fixed set of classes. A novel kind that emits outside those channels does not get misclassified. It is not seen at all. The enumeration has simply moved from the corpus to the sensor suite.
The Large Universe Model is the first position on this axis that gives up enumerating its contents. It does not try to specify in advance what kinds of thing may arrive. It commits instead to a procedural discipline: keep every available stream open, record what arrives even when it fits no existing category, attach provenance to every observation so its source and reliability travel with it, and revise beliefs when a later stream contradicts an earlier one. Novelty, under this discipline, is not a boundary violation that breaks the system. It is an ordinary event — an anomalous reading, retained rather than discarded, waiting for a category that may arrive later, from some other stream, or never.
The 1985 Antarctic ozone finding illustrates the shape of the problem exactly. Ground-based Dobson spectrophotometer readings at Halley Bay recorded ozone concentrations low enough that, on one widely told account, an earlier automated process had treated comparable satellite readings as implausible and filtered them out. Whatever the precise instrument history, the general shape recurs everywhere novelty shows up: the category did not exist to receive the observation, so the observation, if retained at all, sat as an anomaly until a human noticed the anomaly itself was the finding.
What the terminus does and does not mean
Novelty of kind is unbounded. There is no reason to think the world has stopped producing new sorts of thing, and every generation that follows will break some catalogue nobody has written yet. But there are only three ways, temporally speaking, for an observer to be exposed to evidence of anything: through a record made earlier, through sensing something present, or through a stream that stays open indefinitely. Past, present, ongoing — the three temporal quantifiers are exhausted. That is the sense in which the third position on the axis is terminal. It is not that intelligence has a ceiling. It is that a specific ladder, built around how far back and how continuously a system can look, runs out of rungs once it reaches "everything, continuously." Beyond that point, improvement takes the form of more instruments, better provenance, longer retention — not a fourth temporal category of intake, because there is not one.
Three objections deserve to be taken on their own terms, because each catches something real.
The first says emergence may be an illusion at the level of metaphysics. Jaegwon Kim's causal exclusion argument suggests that if a higher-level property does genuine causal work, it competes with its physical base for credit; if it does no independent work, it is merely a redescription. On this view, nothing new appears, only new vocabulary for rearranged microphysics. The reply is to concede the metaphysics entirely and keep the argument. Whether a new coronavirus variant is ontologically fresh or simply a rearrangement of nucleotides that a sufficiently powerful microscope could in principle have anticipated, no assay designed in 2019 had a channel built for it in 2021. In-principle derivability from physics is not the same thing as prior enumeration in a corpus, a schema, or a label set. Deflate emergence as far as the metaphysics allows, and the practical fact of unlistability survives untouched.
The second objection is sharper and genuinely narrows the claim. Any actual Large Universe Model's streams are themselves an enumeration — of instruments, of channels, of provenance fields someone had to design. Neutrinos existed for billions of years before any detector registered one. A new kind may radiate into no existing sensor at all. "Everything, continuously" is therefore an aspiration, not a coverage guarantee, and it should not be stated as one. The honest version of the claim is that closure operates over classes of intake, not over the coverage achieved by any given configuration of streams. That configuration is always incomplete and stays incomplete. What the terminal position offers is that closing a gap never requires leaving the class — a new detector is simply a new stream, absorbed by the same discipline of provenance and revision — whereas closing the equivalent gap for the earlier generations requires switching intake regimes entirely, from recollection to presence to persistence.
The third objection turns the concept back on the argument itself. A three-generation ladder is a taxonomy, produced at a particular moment by particular people, and by the concept's own logic taxonomies fall to novelty. Claiming that no fourth intake class can appear looks like exactly the confident enumeration emergentism exists to embarrass. The reply is a restriction, not a rebuttal: the ladder is not an inventory of observation methods, which could indeed be added to; it is the exhaustion of a quantifier over time. Exhausted quantifiers do not fall to novelty the way species lists do, because there is no fourth tense. What remains genuinely open is the choice of axis. Intake is one axis. Others — intervention, embodiment, the cost of being wrong — may generate their own ladders, with their own termini, or none at all. Terminality is claimed here for intake, and nowhere else.
The misreading to disown
The common misreading runs: if new kinds keep appearing, forecasting is hopeless, so do not bother building anything. That confuses two different capacities. Nobody can predict the unpredictable-in-kind; that is close to a tautology. What a disciplined intake system can do is shorten the gap between a novelty's first appearance in some stream and its recognition as a novelty, and keep the raw observation, with its provenance, so that recognition remains possible even years later when a category finally arrives to receive it. Omicron was not predicted. It was noticed, within days, as an anomaly in a diagnostic assay never designed to detect variants — a dropped target in a three-gene PCR panel — and the anomaly was retained and escalated rather than discarded as noise. That is what the discipline buys: latency and retrievability, not clairvoyance. Emergentism is an argument against closed catalogues. It is not an argument against inference, and it was never meant to license giving up on either.
What is left standing
The concept establishes that enumeration fails against novelty of kind, always, given enough time. It establishes that three temporal quantifiers exhaust the ways evidence can reach an observer, and that a system built to keep all three open has nowhere further to go on that specific axis. It does not establish that such a system sees everything; coverage remains partial and provisional, bounded by whatever streams happen to exist. It does not establish that the three-generation ladder is the only ladder worth building. It says, modestly, that on the axis of intake, this is the top rung, and the next improvement is not a new kind of rung but a better grip on the one already reached.