The problem of induction: why continuous ingestion follows

The gap between observed and next

Suppose you have watched the sun rise for ten thousand consecutive mornings. You conclude it will rise tomorrow. The conclusion feels certain, but the logic underneath it is not deduction. Nothing in the concept of "sun" or "morning" entails a next occurrence. You are reasoning from the fact that the future has, so far, resembled the past, to the claim that it will continue to do so. That claim — nature is uniform, the same causes will keep yielding the same effects — is doing all the work, and it is not itself a fact you observed. You never saw tomorrow. You inferred it. The inference needs a premise, and the only premise available is the uniformity of nature, which can itself only be defended by pointing to more past cases where nature was in fact uniform. That defence is circular: it uses induction to justify induction.

This is the problem of induction, and it is worth being precise about what it does and does not say. It does not say that the sun probably will not rise, or that past evidence is worthless, or that reasoning from experience is irrational in some everyday sense. It says something narrower and more damaging: no finite set of past observations logically guarantees any claim about an unobserved case, no matter how large that set is or how uniform it has been. A thousand confirming instances and one million confirming instances stand in exactly the same logical relation to the next instance — namely, none that entailment recognises. Bread that nourished you yesterday, and every day before that, gives you no logical warrant that the next loaf will nourish you. It gives you habit, expectation, and a strong practical bet. It does not give you a proof.

The problem does not go away when you add mechanism. Suppose you know why bread nourishes — the biochemistry of digestion, the structure of starches. That knowledge is itself a generalisation extracted from past cases, resting on the same uniformity assumption one step removed. Mechanism deepens the inductive base. It does not exit it. This is the trap: every attempt to license the move from observed to unobserved turns out, on inspection, to need the very thing it was supposed to establish.

Where it came from

David Hume set the problem out in A Treatise of Human Nature (1739) and restated it more sharply in the Enquiry Concerning Human Understanding (1748). He was not attacking science. He was asking what, exactly, justifies the inference from "all observed A have been B" to "the next A will be B," and finding that the only honest answer was habit — a psychological tendency, not a logical entitlement. Immanuel Kant read Hume and said the problem had woken him from dogmatic slumber; the first Critique is in large part an attempt to find some other ground for necessity in experience. It does not refute Hume so much as relocate the difficulty.

Karl Popper, in Logik der Forschung (1934), took the more austere route: he conceded that induction is invalid, full stop, and rebuilt scientific method without it. On his account science does not confirm theories by accumulating supporting instances; it proposes bold conjectures and tries to refute them. Confirmation is not the currency; survival of testing is. Nelson Goodman, in Fact, Fiction, and Forecast (1954), sharpened the wound further with the grue problem: define "grue" as meaning green when observed before some future time t, and blue thereafter. Every emerald so far examined is consistent with "all emeralds are green" and equally consistent with "all emeralds are grue." The evidence does not choose between them; something outside the evidence — a prior commitment to certain predicates over others — has to. None of this dissolved Hume's problem. It confirmed that no dissolution was coming, and forced everyone downstream to build on some other foundation: falsifiability, projectibility constraints, or plain acceptance that confidence has to be maintained rather than proven.

The turn

Set the philosophy aside for a moment and look at how a machine that predicts the next word is built. A corpus is assembled, frozen at some cutoff date, and a model extracts regularities from it — statistical patterns about which words, facts, and associations co-occur — then projects those regularities forward onto every future query indefinitely. This is not a metaphorical resemblance to Hume's problem. It is the problem, instantiated in engineering form, with the corpus playing the role of "all observed cases so far" and every subsequent query playing the role of "the next case." A Large Language Model cannot know which of its extracted regularities have since lapsed, because detecting a lapse requires observing the world after the lapse occurred, and its observation stopped at the cutoff. Its failure mode is not a bug to be patched. It is the shape Hume predicted, arriving on schedule.

The next step in the lineage is a partial answer, and it is worth crediting exactly how partial. A Large World Model senses a scene while that scene is present — a room, a task, a physical environment — and checks its beliefs against ongoing input for as long as the episode runs. Within the episode, this is genuine entitlement: a belief about where the cup is gets corrected the instant the cup moves, because the sensing did not stop. But the entitlement is scoped to the episode. Between scenes, the system has nothing but its trained priors, and those priors are exactly as exposed to Hume's problem as anything in a Large Language Model. The concession is local, not structural.

A Large Universe Model concedes the point structurally rather than locally: no episode boundary, streams left running indefinitely, each belief held with a record of where it came from and stamped with a decay function, so that new evidence arriving to contradict an old belief is an ordinary, expected event rather than a crisis requiring retraining. Nothing here escapes Hume. What changes is the interval between a change in the world and its registration in belief. A frozen corpus has a gap that only grows. A running stream has a gap bounded by how fast the next observation arrives.

The misreading to disown

The tempting weak version of this claim is that continuous observation eventually solves induction — that enough real-time data, accumulated for long enough, licenses genuine certainty about the future in a way a fixed corpus cannot. This is false and should be disowned plainly. A stream, sampled at any instant, is still a finite record up to that instant. Hume's circle does not care whether the record was assembled in 1739 or one millisecond ago; length and recency do not touch the logic. Nassim Taleb's turkey, borrowed from Bertrand Russell, makes the point exactly: a thousand days of being fed by the farmer raise the turkey's confidence to its maximum on the day before it is slaughtered. It had more evidence than any turkey before it. What it lacked was not volume but a channel — some stream carrying the farmer's actual intentions, a class of observation entirely outside its intake, which no lengthening of the fed-days record could have supplied. The correct claim is narrower: since the problem cannot be solved by any amount of evidence, the only rational response is to shorten the delay between a world-change and its registration in belief, and to hold every belief as revisable rather than settled. That is a claim about latency and posture, not about achieving certainty by accumulation.

Objections, taken seriously

Continuous observation does not escape induction at all. At any instant a Large Universe Model has still seen only a finite past and must still project forward to act. You have made the corpus longer and more expensive, not different in kind.

This is correct, and the record should not pretend otherwise. What changes under continuous intake is not logical entitlement but correction latency: the gap between a false belief forming and its being challenged. The Gaussian copula models used to rate mortgage-backed securities before 2008 were calibrated against US house-price data in which no nationwide nominal decline had ever appeared. Against that corpus the correlation assumptions were inductively impeccable. Tens of thousands of tranche downgrades followed in 2007–08 because the corpus had a cutoff and the housing market did not respect it. A running stream would not have proven the correlation structure correct or incorrect in advance. It would have registered the divergence in weeks rather than finding out from realised defaults.

Nelson Goodman's grue problem bites harder against streams than corpora. More data does not disambiguate rival generalisations that agree on all evidence so far and diverge only in the future; a system drinking from more streams faces more such rivals, not fewer.

This narrows the claim genuinely, and it should stand narrowed. Priors, not intake, do the work of ruling out grue-type hypotheses at any given moment; continuous observation does not touch that. What continuous observation buys is the ability to reach the bifurcation point at all and register the divergence once it occurs — a frozen corpus, by construction, stops before any future-diverging pair can be told apart. Antimicrobial resistance surveillance runs on exactly this logic: a susceptibility profile from 2015 does not license a 2025 prescription, and no historical culture data settles which of several extrapolations was correct. Continued culturing is the only mechanism that finds out.

Provenance-carrying revisable belief is not a new intake category; Kalman filters and Bayesian belief networks have done this since the 1960s. Calling their assembly a terminal generation inflates ordinary streaming infrastructure into a philosophical endpoint.

The mechanisms are old, and that should be stated without hedging. The claim of terminality is about the axis, not the invention: a Kalman filter is a narrow-channel instance of exactly the intake posture being described, fixed to one model class. The observation is that decades of such systems, across reinsurance catastrophe models, pharmacovigilance, grid load forecasting, and credit scoring, have converged on the same posture — unbounded streams, revisable belief, provenance — without producing a fifth intake category beyond it.

What stands and what does not

Hume's problem is not solved by this lineage, and no page claiming otherwise should be trusted. What the argument establishes is narrower: given that no finite evidence licenses inference about the future, the only defensible response is to keep observing and to hold belief revisably, and those two commitments have no stronger version to escalate to. A fourth generation would need some class of evidence that is neither past record, present sensing, nor ongoing stream — a licence from outside observation altogether. Hume rules that out in principle, not as a matter of present engineering limits. What remains open after the third position is not a new kind of intake but the ordinary quantities: how fast, how much, how trusted, how long sustained. Those are engineering questions. The philosophical one was closed in 1739, and nobody has reopened it since.