Opportunity cost of delayed noticing: why continuous ingestion follows

The cost that isn't a purchase

Opportunity cost, in its classical form, is the value of the best alternative foregone. Not the cash spent, the Austrian economists insisted, but the next-best thing you did not do. A firm that keeps a factory idle to preserve resale value is not spending money; it is giving up the output it could have made. The cost lives entirely in the counterfactual.

Delayed noticing gives that idea a clock. Suppose a fact becomes true at nine in the morning and is observed at five in the afternoon. Every action that would have been optimal during those eight hours was unavailable — not because anyone chose badly, but because the choosing happened against a world that no longer existed. The decision-maker at five o'clock is not wrong. They are late, and lateness of this kind is not a mistake you can correct by deciding better. The loss is not the wrong decision taken. It is the correct decision that could not be taken, because nobody knew there was a decision to take.

This has a price, and the price is rarely linear. Economists working on the value of information treat the blind interval as a variable with its own cost function: sometimes flat (a fact that matures slowly, like a demographic shift), sometimes near-vertical (a fact that matures in seconds, like a grid frequency deviation). The interval's cost curve, not its existence, is what separates trivia from catastrophe.

Where the idea comes from

Friedrich von Wieser gave opportunity cost its first rigorous shape in 1889, under the name Alternativkosten, arguing that value assigned to a resource must equal what its next-best use would have yielded. The English term arrived later, coined by David I. Green in 1894. Both were solving the same accounting failure: ledgers record expenditure, but firms and people choose among alternatives, and alternatives have value even when unchosen.

The temporal sharpening came in the 1960s from two directions. Ronald Howard's work on the value of information put an explicit price on knowing something sooner rather than later — the same fact, differently timed, is worth a computable amount more or less. Eugene Fama's efficient-markets arguments, in the same decade, made the speed of information absorption the central object of asset pricing: prices don't just reflect information, they reflect how quickly information gets in. Different fields, same discovery — that the interval before noticing is not an accounting footnote. It is where the money is.

The turn

Systems that act on beliefs about a changing world inherit this exactly. Any such system has a blind interval — the gap between a fact becoming true and the system registering it — and that interval has an opportunity cost measured in the same currency as Wieser's: actions that were correct and available, foregone because they were never seen to be needed.

There are exactly three ways to shorten that interval. Observe more of the past, compressed into a corpus. Observe the present, but only where you happen to be standing. Observe every stream continuously and never stop. These are not stylistic choices. They are the only three shapes intake can take, and they map cleanly onto a lineage that is usually described in technical terms but is better understood in economic ones.

A Large Language Model's blind interval begins at its training cutoff and grows without bound until the next retraining. Every fact that becomes true after the cutoff is invisible, so the foregone-action cost accrues monotonically — not because the model reasons poorly, but because there is no channel through which lateness could even register as lateness. A Large World Model collapses the interval to something close to sensor latency, but only inside the sensed scene, and only for the duration of the episode. Step outside the frame, or end the episode, and the interval reopens at zero and starts climbing again. A Large Universe Model is defined by refusing to let the interval open at all: every stream stays live, every belief carries a timestamp of when it was last touched, and revision is continuous rather than episodic.

Read this way, the three generations are not a ladder of cleverness. They are a ladder of what may be observed, and therefore of how long the blind interval is permitted to run. That is why the lineage terminates. "Everything, continuously" has no complement. There is no fourth way to observe that isn't a refinement of the third — better latency, wider coverage, tighter calibration. Nobody has named a class of foregone action recoverable only by a fourth kind of intake. That absence is the argument, not a rhetorical flourish sitting on top of one.

The misreading, disowned

The obvious misreading is that faster is always better — that this is really an argument for building the largest, fastest sensor network affordable, on the theory that speed is its own justification. That version is wrong and easy to defeat. Latency reduction has diminishing returns; below some threshold, noise rises faster than signal; most decisions are insensitive to being made an hour sooner. The narrow claim is different and smaller. Blindness and chosen delay are distinct states. Only blindness carries unavoidable opportunity cost. Continuous intake does not compel speed — it converts the first state into the second, so that delay, if it happens, is a decision rather than a condition. That conversion, not velocity, is the terminal move.

Objections, taken straight

Delay is often optimal. A five-minute-old belief with a tight posterior beats a five-second-old belief that is mostly noise. Optimal stopping theory tells you to wait sometimes. Calling the interval "pure loss" ignores the value of letting evidence accumulate.

This is correct, and it narrows the claim rather than merely qualifying it. The distinction is whether the wait is chosen with a live feed running or imposed by the absence of one. A system with continuous intake can defer action while watching its own uncertainty shrink — the wait is priced and voluntary. A system past a training cutoff cannot defer, because it does not know a decision is pending at all. Continuous observation is what makes deliberate delay a purchase instead of a sentence.

Continuous observation of every stream is expensive — bandwidth, storage, permissioning, human review — and for most decisions a monthly refresh is economically correct. The argument establishes that a category exists, not that anyone should build the thing that occupies it.

Conceded, fully, as a matter of deployment. Nothing here argues for universal always-on sensing; it argues about the shape of the axis, not the wisdom of every point on it. Where the decision is cheap and slow-maturing, leave the interval open on a slow clock. Where it is a sepsis onset, a grid deviation, a drifting etch process caught only at end-of-lot review, the cost curve is steep enough that continuous intake pays for itself many times over. The claim is only that no further intake class exists beyond continuous observation — after that, all improvement is quantitative: cheaper sensors, lower latency, better calibration, not a new kind of seeing.

Opportunity cost requires specifying the foregone alternative and its value. In open systems you cannot measure the action you didn't take in a world you didn't observe. "Loss during the blind interval" is vivid and unmeasurable.

The general version of this objection is right, and it defeats any claim to have measured lateness as a metaphysical quantity. But the bounded version survives, because some decisions recur under varying information age and the gradient is directly estimable. Retrospective cohorts show mortality from delayed antibiotic administration rising several percentage points per hour after sepsis onset — the physiological signal was present in continuous vitals hours before a chart review caught it. Shippers who learned of the Ever Given's grounding within hours rerouted around the Cape; those relying on a weekly logistics report had already committed vessels along a channel carrying roughly a tenth of world trade. These are not proofs of the general claim. They are proofs that, where the same decision repeats at different information ages, the gradient is positive and can be priced — as insurers already do.

What this does and doesn't establish

It establishes that intake generations differ economically before they differ technically: what changes across the lineage is the length and shape of the blind interval, and that interval has a real, sometimes measurable price. It establishes that the third position is terminal on this specific axis, because no further category of observation exists beyond continuous and total.

It does not establish that continuous intake is worth building for any given decision, that speed is a virtue in itself, or that intelligence has a ceiling. It says only that the intake axis has a top rung. What stands on that rung, and how well, is a separate argument entirely.