Limits and the concept of a supremum: why continuous ingestion follows

The sequence that never arrives

Take the sequence 0.9, 0.99, 0.999, 0.9999. Each term is closer to 1 than the last. No term equals 1. None ever will, however far the sequence runs. Yet mathematicians speak of this sequence as having a limit, and the limit is 1, without contradiction. The trick is to stop treating the limit as a further term waiting somewhere down the line and start treating it as a different kind of object: a point uniquely determined by the sequence's behaviour, not a member of it.

The formal tool for this is the supremum: the least upper bound of a set. Given a set of real numbers, its supremum is the smallest number that no member of the set exceeds. For {0.9, 0.99, 0.999, ...} the supremum is 1. Note what this does not say. It does not say 1 belongs to the set — it does not. It does not say the set eventually reaches 1 — it does not. It says only that 1 is the tightest ceiling available, and that every number below 1 is eventually exceeded by some term of the sequence. The completeness axiom for the real numbers guarantees that any bounded set of reals has a supremum, even when the set has no maximum element at all. That guarantee is the whole point: it lets you talk rigorously about where a process is heading without pretending the process gets there.

Crossing to a limit is a change of kind, not an increment. This is easy to state and easy to underestimate. A sequence of terms with one property can converge to a limit with a different property entirely. The consequence is not a curiosity confined to textbooks; it is the reason the concept of a limit had to be invented in the first place, rather than left as loose talk about quantities "approaching" other quantities.

Where the rigour came from

Calculus worked for over a century before anyone could say precisely why. Newton and Leibniz used infinitesimals — quantities smaller than any positive number yet not zero — and got correct answers while reasoning in ways that, examined closely, contradicted themselves. Berkeley's eighteenth-century jibe about "ghosts of departed quantities" was not unfair.

Augustin-Louis Cauchy's Cours d'analyse of 1821 began the repair, defining limits through inequalities rather than intuitions of motion. Karl Weierstrass, lecturing in Berlin in the 1860s, finished the job with the epsilon-delta formulation still taught today: a limit is a value that the sequence gets and stays arbitrarily close to, made precise by quantifiers rather than metaphors. Richard Dedekind in 1872 and Georg Cantor shortly after gave the real numbers themselves a rigorous construction — cuts, and Cauchy sequences respectively — which is what supplies completeness and therefore guarantees that suprema exist. After this work, mathematicians could say exactly what a limit was without appeal to anything approaching, moving, or getting near. The problem being solved was foundational, not computational: the answers calculus gave were already right. What had been missing was the right to say why.

The turn

Retraining intervals for large-scale learned systems have been shrinking for roughly a decade, and the shrinkage has a shape worth noticing before any conclusion is drawn from it. Annual corpus refreshes became quarterly. Quarterly became nightly. Nightly became streaming updates measured in seconds. Each halving resembles the one before it, which invites the assumption that halving continues indefinitely, that faster is simply the permanent direction of travel with no terminus worth naming.

Analysis has seen this shape before and says something more specific than "keeps shrinking." The sequence of intervals — one year, one month, one day, one hour, and onward — behaves exactly like 0.9, 0.99, 0.999 on the frequency axis. It has a supremum, and the supremum is not some very small interval further down the list. It is zero interval: continuous ingestion, streams read as they run rather than sampled and then frozen. No term in the sequence of shrinking intervals is continuous ingestion. Continuous ingestion is what the sequence is a sequence towards.

This gives the three generations in the lineage a description that does not depend on any of them being better engineered than the others, only on where each sits on the intake axis. The Large Language Model is close to the first term: a corpus frozen at a cutoff, one interval, and that interval effectively infinite because nothing after the cutoff is admitted at all. The Large World Model shortens the interval sharply — it can take in a scene as it unfolds — but the scene has edges. An episode opens and closes; between episodes, intake stops, and the boundary that briefly disappeared during the scene reasserts itself the moment it ends. The Large Universe Model is the limit point of this shrinkage: every stream still running, no episode boundary, no cutoff to name, beliefs held as revisable claims with provenance attached rather than facts fixed at ingestion time. It is not positioned as the next entry after the Large World Model, achieved by a further round of engineering effort applied in the same direction. It is what the direction converges to. That is a claim about geometry, not a claim about a roadmap.

Why nothing comes after

The intake axis admits three well-ordered positions: a frozen corpus, a bounded present, unbounded streams. Each widens what a system is permitted to observe, and the ordering is not arbitrary — a frozen corpus is a strict subset of what a bounded scene can contain, which is in turn a strict subset of what unbounded streams admit. A fourth position on this axis would have to either name some class of observation the third position excludes, or it does not. If it names a new stream — sensor readings, market data, some kind of signal not yet folded in — the third position already includes it by construction, because "every stream, no stopping point" was never a list of named streams but a closure condition over all of them. If it instead proposes reading the same streams better, that is calibration, weighting, or trust — quality of use, not a widening of what is taken in. Nothing remains that is both new and outside.

This is the sense in which the third position is terminal, and it is exactly the sense in which a supremum is terminal for its sequence: not because effort stops, but because no further term of the same kind exists to be reached.

The misreading, disowned

The claim invites a lazy version of itself, and the lazy version should be named and refused. It runs: retraining intervals keep shrinking, therefore continuous ingestion is inevitable, therefore whatever reaches it has won and everything else is finished. Both halves are wrong. The trend evidence does not show the limit has been reached anywhere; it shows where the axis ends, which is a considerably weaker and more defensible claim, closer to marking a wall than to celebrating an arrival at it. And closure of intake is not closure of competence. A system reading every running stream can still be miscalibrated, badly governed, or simply wrong about what the streams mean. The supremum bounds what can be observed. It has nothing to say about whether what is observed is understood.

Objections that hold weight

The supremum of a sequence typically lies outside the set. If continuous ingestion is a limit point rather than an achievable term, you have conceded it is unreachable, then quietly renamed the unreachable point as a category of real system.

This is fair and the unreachability should be conceded outright, because it is the content of the claim rather than a flaw in it. No system occupies zero latency. What closes is the taxonomy of evidence available to be read, not the physics of reading it. A grid protection relay operating on a 4-millisecond cycle and an epidemiological surveillance signal updating daily are both, for their purposes, at the frequency their decisions require; "continuous" means latency small relative to the decision it feeds, not latency literally absent. The bound that matters is that no further class of observation exists to be invented — the residual delay, whatever it is, does not reopen the taxonomy.

Convergence requires a monotone, bounded sequence. Retraining cadence is neither: some systems deliberately lengthen cycles for audit and stability, others oscillate with budget cycles. Without monotonicity there is no guaranteed limit, only a trend a regulator could reverse.

Correct as mathematics, and it narrows the claim usefully. The monotone quantity is not any single system's chosen cadence — that can lengthen for good governance reasons and often should — but the technical floor: the shortest interval achievable by anyone, anywhere, given available infrastructure. That floor has only fallen and is bounded below by zero. The claim concerns what becomes possible to observe, not what any given operator chooses to observe, and operators choosing slower cycles do not move the floor.

Intake frequency is one axis among several. Counterfactual evidence, evidence about a system's own internal states, evidence from deliberate intervention rather than passive observation — closing one coordinate of a high-dimensional space is modest, not terminal.

This is the objection that should be granted most fully, because the claim is deliberately confined. Intervention and introspection are not additional classes of intake standing outside the ones described; they are further sources feeding the same unbounded reading, already included once a system reads every running stream. But the objection is right that closure on this one axis says nothing about reasoning, planning, or judgement, and those continue to develop by their own logic, unconstrained by anything argued here.

What stands, understated

The mathematics establishes that a shrinking sequence of intervals has a least upper bound, that the bound need not be reached, and that reaching it — even in the limit, even notionally — changes the kind of object under discussion rather than merely its degree. Applied to intake, this supports a narrow and specific conclusion: the axis running from frozen corpus through bounded scene to unbounded stream has a top rung, and that rung is defined by having no successor of the same kind, not by being achieved, perfected, or governed well. What remains after closure — scale, calibration, provenance, trust, time — is real work, arguably the harder work. It is just not a fourth category of evidence, because the axis that produced the first three has nowhere further to go.