The Selection Gradient

The machines being of themselves unable to struggle, have got man to do their struggling for them: as long as he fulfils this function duly, all goes well with him—at least he thinks so; but the moment he fails to do his best for the advancement of machinery by encouraging the good and destroying the bad, he is left behind in the race of competition.

A failed search can leave something worth knowing. Architectures, datasets, objectives, and evaluation methods are tried because the useful route is not known in advance. An unsuccessful attempt may give the next search a better place to begin; another may leave only an expense. Both consumed resources. Distinguishing them requires an account of what was learned and whether anyone can make use of it.

I use selection for the institutions and procedures through which these decisions are made. Training objectives, deployment tests, customers, regulators, professional offices, and capital markets preserve different objects for different purposes. The singular word is convenient so long as it does not persuade us that they have made a common judgment.

Three Filters, No Final Judge

Alchian described markets as an evolutionary environment in which firms can survive without having solved an optimization problem in advance.¹ Nelson and Winter later treated firms as carriers of routines altered through search and differential survival.² The analogy is useful because it shifts attention from intention to retention. It becomes misleading when survival is mistaken for proof of social value.

Training retains parameter configurations that perform better under an objective and an update rule. Deployment retains a product only if it works tolerably amid latency, distribution shift, integration, user behavior, and failure costs. Capital allocation determines which programs receive another attempt. None of these filters is neutral. An objective can reward the wrong proxy. Customers can prefer a harmful convenience. Capital can sustain a loss-making project in expectation of monopoly or strategic power.

Passing the training test therefore does not entail usefulness. Adoption does not entail legitimacy. Profit does not entail public value. Selection joins expenditure to consequence, but the criterion and the authority behind it decide what sort of consequence survives.

For a stipulated pipeline, the bookkeeping can be written:

C = E · η

where energy flow E and hardware efficiency η yield a rate of computation C. If κ represents task-relevant work per operation and S the conditional share that clears the declared filters, then:

T = C · κ · S

This is a chosen model of verified task throughput, not a universal production function. Its terms require units, thresholds, and a population of attempts. S cannot be read from market survival in the abstract, and multiplying the terms does not show that each is independent. The expression earns its keep by forcing the analyst to say where an apparent gain vanished.

An observed training bill measures what one route cost. A benchmark records performance against a chosen test. Deployment can show whether the output reduced error, time, or loss in a specified setting. Capital markets reveal which expectations attracted funding. These observations belong to different registers. The temptation is to let prestige travel upward from the most measurable one until the entire chain appears validated.

Who Chooses the Test

The question hidden inside every selection gradient is who gets to choose it. A hospital may value sensitivity, safety, explanation, and recourse differently from a consumer application. A platform may optimize engagement while users bear the degradation. A public authority may preserve redundancy that a private operator would call inefficient. There is no view from nowhere that turns all of these into one scalar.

Physics adds discipline without supplying the criterion. Computation remains tied to chips, power, cooling, and time. A search that fails still consumed scarce capacity. Open weights can diffuse a capability without repaying the organization that financed the first search. A cheap copy may nevertheless depend on data, insight, and infrastructure produced by an expensive predecessor. None of this grants the predecessor a natural rent. It explains why frontier search requires financing even when successful artifacts become reproducible.

The same distinction clarifies the capitalizing-electricity thesis. Physical throughput can be made productive as cognition only through an arrangement that gives the output a task, tests it, preserves it where appropriate, and permits it to act. The arrangement may discover usefulness. It may also manufacture demand, entrench a gate, or optimize what should never have been the objective. Selection is necessary to economic consequence and insufficient to justify it.

Perez offers a way to ask whether the conversion is reorganizing more than a few firms. If learned inference is becoming a key input, its effects should appear in design, investment, organization, and the location of complementary scarcity. That proposition begins with cost decline. It does not end there.