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.
— Samuel Butler, Erewhon (1872)

A system that consumes megawatts to produce useless output has participated in Factor Prime's physics without participating in its economics. The recursive property of computation—that it can improve itself—does not, on its own, create economic value. The mechanism that connects thermodynamic depth to value is selection.

I use "selection" as a shorthand for a family of filters—training, deployment, capital—that determine which structures survive. "Gradient" here is colloquial, not a literal derivative. The point is that there is a direction: structures that pass the filter persist, structures that fail it do not.

Mud Pies and Why Cost Is Not Value

The problem is the mud pie. The labor theory of value, in its naïve form, holds that the value of a good is determined by the labor required to produce it. The standard refutation is simple: if I spend eight hours making mud pies, my labor does not make the mud pies valuable. Effort alone does not create value. Value requires that the effort produce something people want.

The same objection applies to thermodynamic depth. If I run a neural network training job for six months, consuming megawatts of power and generating vast quantities of entropy, I have certainly created a structure with high thermodynamic depth. But if the resulting model fails to reduce uncertainty about anything anyone cares about, the depth is economically worthless. I have made a computational mud pie. The energy was dissipated, the entropy was produced, and the money was spent, but no value was created.

This is not a hypothetical concern. The history of artificial intelligence contains repeated episodes of expensive failure: models that overfit their training data, systems that performed well on benchmarks but failed in deployment, research programs that consumed years of effort and produced nothing usable. The thermodynamic cost was real in every case. The value was not.

The resolution lies in a concept that economists have understood for decades but rarely connect to thermodynamics.

Armen Alchian, writing in 1950, proposed a reframing of economic behavior that dispensed with the assumption of perfect rationality.(Alchian 1950) Firms, he argued, do not maximize profits because their managers have solved the optimization problem correctly. Firms that happen to behave as if they were maximizing profits survive and grow; firms that do not, shrink and fail. The market selects for behavior that approximates profit maximization, regardless of whether any individual decision-maker intended it. The selection mechanism does the work that optimization theory attributes to calculation.

Richard Nelson and Sidney Winter extended this insight into a comprehensive theory of economic change.(Winter 1982) Their 1982 book argued that firms are best understood as carriers of routines, patterns of behavior that persist over time and are modified through search and selection. The routines that produce good outcomes spread. The routines that produce bad outcomes are abandoned or displaced. The economy evolves through a process analogous to natural selection, with differential survival and reproduction doing the work of improvement.

In markets, selection operates through profit and survival; in training, through loss and retention; in deployment, through adoption and replacement. The criteria differ, but the logical structure is the same: costly variation filtered by a test.

Thermodynamic depth is necessary for value but not sufficient. What matters is whether the structure survives the test of deployment.

Depth measures the computational work required to produce a structure. Selection is the mechanism that determines whether that work was worthwhile. A structure with high depth and no selection pressure is a mud pie. A structure with high depth and strong selection pressure is a trained model that reduces uncertainty at scale. The depth is necessary because someone must pay the cost at least once per frontier increment. Shortcuts generally mean borrowing previously-paid depth through copying, distillation, or transfer, not avoiding the cost at the frontier. The selection is necessary because there is no way to know in advance which structures will prove useful.

Selection as a Multi-Stage Filter

Selection is an observable process with measurable outcomes.

When a model is trained, updates that reduce loss are retained; updates that increase loss are erased. This is selection over parameter updates given an objective, not selection over organisms. But it is a filter: an irreversible rule that preserves some configurations and destroys others. When a model is deployed, user adoption rates separate functional products from abandoned experiments. A fraud detection system processing millions of transactions daily survives because it reduces chargebacks by measurable amounts. A system that increased chargebacks would be disabled within hours. When capital flows, fund returns separate strategies that work from strategies that fail. The venture capital firm that backed the winning approach attracts subsequent capital; the firm that backed the losing approach does not.

The quantification matters. Training selection can be measured against a holdout set on a defined metric—accuracy, loss, calibrated error, or task success rate—and compared across alternatives. Deployment selection can be measured in retention and revenue: a customer service agent that reduces ticket resolution time from 15 minutes to 3 minutes while maintaining satisfaction scores creates measurable value. One that increases resolution time or decreases satisfaction is removed. Capital selection can be measured in funding velocity: the research program that attracts follow-on investment is selected over the program that does not. Each stage produces numerical outputs that can be compared across alternatives. Selection is a filter that determines what survives.

The test has multiple stages, and the stages operate on different timescales.

Training selection (Sᵗ) operates at the level of parameters. Gradient descent adjusts the model's weights to minimize a loss function. At each step, the algorithm compares the model's predictions to the training data and nudges the weights in a direction that reduces the error. Configurations that reduce loss are retained. Configurations that increase loss are corrected. The feedback loop operates on the timescale of milliseconds to seconds. Measurable proxies: out-of-distribution performance decay, evaluation reliability, loss curvature.

Deployment selection (Sᵈ) operates at the level of products. A model that performs well on training loss may still fail when exposed to real users with real problems. The distribution of inputs shifts; edge cases appear; latency and reliability become binding constraints. Users adopt the model or abandon it; revenue flows or does not; the model is retained, updated, or replaced. The feedback loop operates on the timescale of days to months. Measurable proxies: retention uplift versus baseline, error-cost reduction, latency-adjusted task success rate.

Capital selection (Sᶜ) operates at the level of organizations. Investors observe which companies and research programs produce models that survive deployment selection. They direct funding toward the approaches that appear to be working and away from the approaches that appear to be failing. The feedback loop operates on the timescale of months to years, and it shapes which directions of research receive the resources to continue. Measurable proxies: follow-on funding rates, revenue multiples, talent concentration.

Each stage filters the space of possible structures. The structures that pass through all three filters are the ones that accumulate economic value. The structures that fail at any stage are discarded, regardless of their thermodynamic depth. As a conceptual factorization, selection is multiplicative because it operates as a sequence of gates:

S = P(passᵗ) · P(passᵈ | passᵗ) · P(passᶜ | passᵗ, passᵈ)

In practice, each term is estimated by proxies: the value of the factorization is that it forces you to ask where the pipeline is failing. A model with perfect training metrics but zero deployment traction has S approaching zero. A model with moderate training metrics that clears deployment and capital filters accumulates value despite its modest depth.

This framing allows us to name the failure mode precisely: costly computation that fails to clear the market test. The category includes vanity training runs that produce models no one uses, research programs that pursue technically impressive but commercially irrelevant objectives, and infrastructure projects that consume capital without generating returns. The thermodynamic work was done; the selection gradient was misaligned; the value did not materialize.

A bonfire produces entropy. So does a power plant. The difference is that the power plant's entropy production is coupled to a load that does useful work, while the bonfire's entropy production dissipates into the night. Selection is the coupling mechanism that distinguishes productive computation from mere dissipation.

What Does the Physics Add?

A predictable objection: this is simply R&D and product-market fit, dressed up in thermodynamic language. What does the physics add?

The physics adds three things. At the most immediate level, it explains why the standard marginal-cost story fails for frontier models. Open weights compress the scarcity premium on a given generation, but they do not erase deployment costs or the cost of advancing the frontier. The rent is in timing, control, integration, and deployment economics, not in the mere existence of a file. The frontier keeps moving, and reaching it requires paying the thermodynamic price again. The physics grounds the economics.

The framing also links capability progress to energy and infrastructure constraints. If computation were purely abstract, progress would be limited only by cleverness. But computation is physical, and the physical constraints—power, cooling, interconnection, chip supply—tell you which bottlenecks are candidates to bind first. The physics constrains where to look.

And it supports the regime thesis. If Factor Prime is genuinely a new factor of production, then the economy should reorganize around it in predictable ways: capital concentration at the energy-compute interface, geographic clustering near cheap power, institutional adaptation to the new cost structure. The physics makes the prediction testable.

Another objection: open source collapses pricing, so thermodynamic depth cannot sustain rents. This is partially true. Diffusion compresses rents within a generation. But open source does not eliminate the cost of advancing the frontier. It socializes the benefits of whoever paid that cost. The distilled model is still downstream of the original thermodynamic expenditure. And the frontier keeps moving. The company that open-sources last year's model while training next year's model at scale retains its advantage. The search cost must be paid by someone, and whoever pays it first captures the option value of the next generation.

The Production Function

We can now state the production function for Factor Prime with more precision. Define:

E = Energy flow (joules per second)
η = Hardware efficiency (useful operations per joule)
κ = Algorithmic efficiency (task-relevant work per operation)
S = Selection filter (probability of clearing training, deployment, and capital gates)

A candidate functional form for verified task-throughput T at a defined quality threshold:

C = E · η (compute throughput: operations per second)

Call the cumulative search work that produced a structure its depth: depth accrues when energy dissipation is organized into a search process that reliably yields useful structure.

T = C · κ · S (verified task-throughput is compute, scaled by algorithmic efficiency and filtered by selection)

The function is not additive. You cannot substitute more energy for less selection, or more throughput for less algorithmic efficiency, beyond certain limits. The variables interact: high energy with low efficiency produces heat, not computation; high compute with low selection alignment produces waste. The production function has a multiplicative structure, where each variable must exceed some minimum for the output to be positive.

The selection term S acts as a filter that converts physical expenditure into economic output. A model with high depth and poor deployment performance has low effective task-throughput. A model with moderate depth and strong deployment performance has higher effective task-throughput.

Markets, in this framework, are selection filters with accounting attached. The price system aggregates information about which outputs people value and transmits that information back to producers. Firms that produce outputs aligned with the selection gradient earn revenues that exceed their costs. Firms that produce outputs misaligned with the gradient incur losses and eventually exit. The market does not care about thermodynamic depth per se. It cares about whether the structure produced by that depth solves problems people will pay to have solved.

Once proxies become targets, selection drifts. Benchmarks become gameable; leaderboard rankings diverge from deployment value; the gradient must be renewed by contact with reality. This is Goodhart's law applied to capability evaluation. A model optimized for MMLU or HumanEval may or may not solve problems users care about. The real selection gradient is deployment: does the model reduce uncertainty that someone will pay to have reduced, at a cost that permits sustainable operation?

This is why the mud pie objection does not defeat the thermodynamic theory of value. The theory claims that thermodynamic cost, filtered through selection, creates value. The cost is necessary because there is no way to produce a complex structure without incurring irreversibility somewhere. The selection is necessary because there is no reliable way to know in advance, at frontier uncertainty, whether a structure will prove valuable without testing it against the demands of users, markets, and deployment environments. The two are complementary, and both are required.

The implications for capital allocation follow directly. If Factor Prime is the product of energy, efficiency, algorithmic leverage, and selection, then investment should flow toward:

Cheaper and more reliable energy supply for computation
More efficient hardware that extracts more operations per joule
Better algorithms that extract more task-relevant work per operation
Tighter coupling between model outputs and real-world selection pressures

The fourth item is often neglected. Much of the discourse about AI focuses on capability benchmarks. These benchmarks capture some aspects of selection, but they are proxies. The real selection gradient is deployment: does the model solve problems that users will pay for, at a cost that permits profitable operation, with reliability and latency that meet real-world constraints? The selection term is what distinguishes models that accumulate economic value from models that accumulate only citations.

Factor Prime is now defined: energy, structured through computation and disciplined by selection, that produces economically useful uncertainty reduction. The work is irreversible, the selection is unforgiving, and what emerges from both is the only form of value that cannot be counterfeited. The question is whether this factor behaves as the theory predicts, whether it meets the criteria Carlota Perez established for a new techno-economic paradigm, and whether the economy is reorganizing around it in the ways the regime thesis implies.