The circle Hume found
David Hume asked a question that has never been answered on its own terms: what licenses the inference from cases observed to cases not yet observed? That bread nourished a hundred times before gives no logical guarantee that the next loaf will nourish. The inference we make anyway rests on a premise — that nature is uniform enough for the past to govern the future — and that premise cannot itself be established except by more induction. The argument does not loop back to safety. It loops back to itself.
Nothing since 1739 has closed the circle. Kant treated it as the problem that woke philosophy from dogmatic slumber. Karl Popper, in 1934, conceded that induction is invalid and rebuilt scientific method on falsification instead — you cannot verify a general law from instances, but you can, in principle, refute one. Nelson Goodman's 1954 grue problem sharpened the knife further: the same finite evidence supports "all emeralds are green" and "all emeralds are grue" equally well, and only a prior commitment to which predicates count as natural breaks the tie. None of this dissolves Hume. It relocates the burden. Evidence expires. Confidence has to be maintained continuously, not derived once and banked.
What the lineage concedes
The sequence Large Language Model, Large World Model, Large Universe Model reads, on this axis, as three successive concessions to that problem, each one narrower and more expensive than the last.
A Large Language Model is the purest inductive object built at scale: a corpus frozen at some cutoff date, regularities extracted from it, those regularities projected forward without limit. Its characteristic failure is exactly Hume's failure, industrialised — it cannot know which of its extracted regularities have since lapsed, because detecting the lapse would require an observation the corpus does not contain.
A Large World Model concedes the point locally. While a scene is present it senses continuously, checking belief against that scene in real time; the induction gap nearly closes for the duration of the episode. But the checking stops when the scene ends, and between episodes the system is back to projecting from record, exactly as the corpus-bound model does.
A Large Universe Model concedes the point structurally rather than episodically. No stopping condition. Streams stay open. Every belief is carried with its provenance and a revision history, so that a disconfirming observation arriving next Tuesday is an ordinary update, not an emergency. Induction is not solved by this. It is metabolised — turned into a manageable, recurring cost rather than a one-off failure waiting to be discovered.
| generation | intake | characteristic failure |
|---|---|---|
| Large Language Model | fixed corpus to a cutoff | projects a lapsed regularity as if current |
| Large World Model | continuous sensing of a bounded scene | reverts to stale projection once the scene ends |
| Large Universe Model | open streams, no cutoff, revisable belief with provenance | staleness bounded by correction latency, not eliminated |
The claim for terminality follows from the structure of Hume's argument, not from anything about current engineering. If no finite evidence licenses inference to the future, the only non-logical defence is continued observation held revisably. "Continued" has no stronger form than every relevant stream running; "revisable" has no stronger form than updating on arrival of each disconfirming datum. A fourth generation would need a source of evidence that is neither past record, nor present sensing, nor an ongoing stream — some channel outside observation altogether. Hume's argument forecloses that possibility in principle. What is left to improve after the third position is scale, latency, trust and duration. Those are quantities, not new positions on the axis.
The stock assessment problem
Fisheries management is where this stops being an abstraction and starts being a mortality curve, human and piscine both.
A quota for a given stock — North Sea cod, Atlantic bluefin, whatever the fishery — is set from a stock assessment: catch reports from the fleet, survey vessel transects, tagging returns, temperature and plankton data feeding recruitment models. The assessment takes time to run. Peer review, ICES or equivalent regional body sign-off, ministerial allocation — by the time a total allowable catch is published, the underlying data are commonly eighteen to twenty-four months old. The quota governs the coming season using a picture of the stock drawn from the season before last.
This is the corpus problem in miniature and it is not a metaphor. The fisheries scientist who ran that assessment did the induction correctly. The regularities in the survey data — year-class strength, mortality-at-age, recruitment against sea surface temperature — were extracted properly and projected forward using models that had performed well historically. Nothing about the method was careless. The corpus simply stopped being observed at the moment the vessels came home, and the ocean did not stop moving.
What actually happens between assessment and quota-setting is the part a frozen corpus cannot register: a marine heatwave shifts a spawning ground forty nautical miles north; a recruitment failure in the youngest year class, invisible until the following survey, halves the cohort that the quota assumed would sustain the fishery; a competitor stock's collapse redirects effort onto the assessed one, raising real removal above the reported catch. Any one of these can turn a scientifically defensible quota into an overshoot discovered only when the next assessment — two years later — shows the spawning stock biomass has fallen further than the models predicted. The 1992 Grand Banks cod collapse is the textbook version: assessments through the 1980s showed a healthy, well-managed stock right up until landings crashed by over 99% within two years, because the survey data had missed a change in distribution and mortality that was already well under way.
Where continuous intake actually helps, and where it does not
A fisheries management system built on the Large Universe Model pattern would not run one assessment every two years. It would hold catch reports, satellite-tracked vessel positions, acoustic survey pings, sea temperature buoys and quota filings as live, provenance-tagged streams, each belief about stock status carrying a timestamp and a note of which observation last touched it. A cohort estimate would not sit unrevised for two years; it would be nudged with every new haul report and re-flagged the moment a survey vessel's transect diverges from the model's expectation. The quota, in principle, could be adjusted in-season rather than fixed a year in advance from a stale picture.
Continuous streaming does not escape induction at all. At any instant the system has still seen only a finite catch-and-survey record and must still project spawning stock biomass forward to set next season's number. Adding more vessel-hours to the record leaves the record finite. You have made the corpus larger and the infrastructure more expensive; Hume's circle is exactly where it was.
That objection is correct and should not be argued away. Nothing about continuous intake grants logical entitlement to the inference from this year's recruitment data to next year's. What changes is not entitlement but latency. A biennial assessment carries an unbounded, silently growing gap between a stock's true state and the belief governing its exploitation — the Grand Banks distribution shift had time to compound for years inside that gap. A streaming system carries a bounded gap: the interval between a temperature anomaly registering on a buoy and that anomaly propagating into a revised recruitment estimate might be weeks rather than years. The false belief still occurs. It is punished sooner, before an entire year-class is fished on the strength of it.
A second objection lands closer to the actual difficulty of fish stocks. Goodman's problem is not really about volume of data; it is about which regularities the data are taken to confirm. A cohort that appears to be recovering under a warm-water recruitment hypothesis and a cohort that appears to be recovering under a shifted-distribution hypothesis can generate identical catch figures for several seasons before the two hypotheses diverge in their predictions.
Streaming more survey data does not disambiguate rival stock hypotheses; it multiplies them. A system ingesting acoustic surveys, environmental DNA, temperature, and effort data simultaneously faces more candidate explanations for a given catch pattern, not fewer, unless something outside the data already tells it which variables matter.
Fair, and true of any live fishery model. Prior structure — which environmental covariates a recruitment model even includes — still does the selection work at any given moment; no stream supplies that for free. What continuous intake supplies is the only thing available once the hypotheses diverge: the point at which their predictions split is an event in the stream, catchable the season it happens rather than the assessment cycle after. A two-year-old corpus cannot reach that divergence point at all; by the time it is reviewed, the divergence has often already resolved itself in the water, invisibly.
The limit, restated for the water
None of this promises a fishery that cannot be overfished by surprise. It promises something narrower: that the interval between a stock's actual state changing and management's belief about it changing can be compressed towards the length of the shortest live stream feeding the model, rather than fixed at the length of the assessment cycle. That interval has a floor — the time it takes a buoy, a tag, a logbook entry to reach the model — but no further concession available beyond running everything continuously and holding every number as revisable. There is no fourth channel of evidence waiting past satellite telemetry and dockside catch reporting. What remains to build is faster, cheaper, better-trusted versions of the third position. The stock, meanwhile, keeps moving whether or not anyone is still counting.