Who Closes the Loop

Loss generally occurs when a player overrates his advantage.

Computation has entered its own workshop. Models write code, propose tests, search designs, synthesize data, inspect failures, and help decide which experiment should run next. The loop is incomplete, but it is no longer wholly conducted by human hands.

That is the consequential claim. It is often inflated into a stronger one: because computation contributes to its successor, improvement must accelerate without limit. Nothing in the present evidence establishes that conclusion. A loop can tighten and still meet diminishing returns. It can move one bottleneck only to expose another.

Steam power also participated in its reproduction by pumping mines and moving coal. Electrical control improved the operation of grids. Machine tools built better machine tools. In each case the productive system acted upon parts of its own means of production, while observation, invention, finance, construction, and authorization remained distributed among people and institutions. What changes now is the range of cognitive work that can pass through the machinery itself.

A Share That Has to Be Measured

Call the recursive share, ρ, the portion of a declared capability-development pipeline performed by computational systems. The denominator has to be named before the ratio has meaning. Code generation, data construction, evaluation, experiment execution, diagnosis, hardware design, fabrication, capital allocation, and release authority are different kinds of work. A measure covering only accepted code cannot be silently promoted into a measure of the whole loop.

Nor is tool use enough. A useful estimate would weight contributions by verified completion and ask whether rising participation predicts faster improvement after controlling for staff, compute, and spending. An accepted patch may save time without changing the rate of discovery. A model may propose a thousand experiments while humans still define the objective, adjudicate ambiguous results, and decide what may ship.

Current telemetry therefore answers narrower questions. Code-acceptance data, developer surveys, and task-horizon benchmarks observe different surfaces.¹ Benchmarks can bound tasks systems might complete, but they do not reveal how a laboratory divides causal responsibility across models, researchers, infrastructure, and management.² No reliable public series yet measures ρ for the complete frontier-research pipeline.

The Growth Question

Romer's growth theory made knowledge unusual because it can be used without being exhausted.³ An idea can enter many productive processes at once. Yet the production of ideas may still face increasing difficulty, scarce talent, complementary capital, and institutions that slow or redirect their use.

Jones distinguishes growth regimes in which additional research effort encounters diminishing returns from those capable of sustaining stronger feedback over a relevant range.⁴ Aghion and Howitt likewise model innovation as an endogenous activity shaped by incentives and creative destruction.⁵ These frameworks clarify the question. They do not make ρ a measured parameter, and they do not transform a demonstration of coding assistance into a theorem about aggregate growth.

Nordhaus asks what evidence an economic singularity would leave.⁶ The historical record does not show a growth rate rising without bound. Automation of invention could change that record, but it could also raise research output while growth remains damped by harder discoveries, physical construction, energy, capital, regulation, or demand. Scaling laws relate model performance to specified inputs under observed regimes.⁷ They are not laws of economic acceleration.

For the acceleration claim, the empirical issue is one of pacing. Does computational participation reduce the calendar time required for a verified improvement? Does it increase validated discoveries per unit of staff and spending? Do bottlenecks migrate from researcher attention toward compute, experiments, fabrication, power, or institutional approval? A rising ρ alone answers none of those questions.

The Loop in Practice

Imagine a model proposing an architecture change, an automated system training and evaluating it, and another process returning the result to the next round of search. Such a loop can run without a person choosing each experiment. It may explore thousands of variants in parallel. But someone still chose the evaluation basket, supplied the budget, accepted the risks of optimization against the measure, and decided whether an apparent gain deserved deployment.

The distinction between runtime and principal does not weaken the economic claim. It locates it. Capital can own a capability that performs part of the work of research and management. The capability can help produce an improved successor. The owner can then allocate more resources to the process on the basis of results the process helped generate. A portion of capital allocation has entered the loop even though the loop neither owns itself nor closes every dependency.

Failure is legible. Greater model use without faster validated discovery would count against the acceleration claim, even if it changed the division of work. Review and integration may absorb the time saved. Faster iteration with no durable improvement would show that search, not consequence, had accelerated. Physical capacity expanding while objectives and approval remain binding would show that the loop had tightened in one place and remained open in another.

Partial recursion is already enough to alter industrial organization. It changes the relative demand for researchers, compute, experiments, and infrastructure, and it allows existing capability to participate in the production of what may displace it. Whether that produces acceleration is unresolved. The unresolved question is more disturbing than a slogan about self-improvement because it can be measured, contested, and capitalized before anyone knows the answer.

Capitalizing electricity therefore reaches beyond storing trained weights. It can also fund a search that changes what the next expenditure will build. Chapter 15 turns to a different conversion: computation whose accepted result helps create a protocol-recognized claim and an ordered history.