The Recursive Economy: AI Self-Improvement and Scarcity

Skip to essay

RESEARCH ESSAY / VERSION 1.0

THE<br>RECURSIVE<br>ECONOMY

What happens when intelligence becomes an input into the production of intelligence?

AI becomes economically recursive when one model measurably shortens the work of building its successor.

Version 1.0 · Released 15 August 2026<br>recursive-economy.pages.dev

Read the argument ↓<br>Download PDF ↓

THE NUMBER TO WATCH

How many weeks did the current model take off the schedule for the next one?

No public benchmark answers it.

01<br>THE NUMBER LABORATORIES DO NOT PUBLISH

02<br>WHAT CURRENT SYSTEMS CAN ACTUALLY DO

03<br>FROM CHEAP ATTEMPTS TO ACCEPTED PROGRESS

04<br>JOBS AND WHO GETS THE GAINS

05<br>CAPITAL AND THE STATE

06<br>WHAT WOULD CHANGE MY MIND

07<br>CONCLUSION

08<br>REFERENCES<br>Suggested citation↘

Approximately 21 minutes at a continuous reading pace.

ESSAY INDEX

VERSION 1.0

Back to top ↑

Note to the reader#

Evidence and sources are current through 15 August 2026 . The forecasts are conditional. For feedback or comments, contact tre.numbing085@passfwd.com.

THE NUMBER LABORATORIES DO NOT PUBLISH#

There is one number I would pay to see from a frontier AI laboratory: how many weeks did the current model take off the schedule for the next one?

This essay is about that number.

A model can lead every public benchmark and leave its laboratory's release calendar untouched. It might write useful functions or summarize papers while researchers spend just as long choosing experiments, interpreting failures, and deciding what belongs in the training pipeline. Another model can look less impressive in public while quietly preparing runs, tracing bugs, and proposing changes that survive review. If enough of that work reaches the next training stack, the second model is doing something more important than scoring well. It is helping build its successor.

If the answer to my question is zero, AI research tools may still be commercially valuable. They have not changed the pace of frontier development. If the answer is six weeks, and the successor removes still more time from the following cycle, intelligence has become an input into producing intelligence in the economically important sense. The loop does not have to be autonomous before it matters.

I do not think the public evidence shows a closed loop. It shows pieces of one: systems that search for algorithms under hard evaluators, rewrite the software around a fixed model, sustain longer coding tasks, and carry out parts of a research project. The missing result is an audited account of what survives human review and moves the date of a broadly stronger successor.

For the next several years, I expect people to set the research agenda and approve consequential decisions while machines take over more coding, experiment setup, debugging, and local evaluation. Each model generation will improve the tools around the next. Call it managed compounding. If accepted AI-assisted R&D rises but, against a defensible baseline, time to a fixed successor-quality threshold does not fall, the thesis is wrong.

The effects would arrive before an autonomous scientist. Frontier laboratories would spend heavily on inference as well as training. Review and evaluation would become larger constraints. Firms in exposed industries could raise output without matching growth in hiring. Models are cheap to copy; advanced chips, grid connections, proprietary experiments, and trusted institutions are not. Control of those complements would decide where much of the gain lands.

The phrase recursive self-improvement often evokes a solitary program inspecting its own weights. Frontier models are built by organizations, not solitary programs. A laboratory joins models to data, code, evaluators, compute, security procedures, and the accumulated judgment of its staff. Decades of research on information technology make the same broad point: the useful unit is often the technology together with the organization that learns how to use it.1

I therefore use a deliberately organizational definition:

Recursive self-improvement begins when work done by one AI system survives review, enters the process that builds a later system, and shortens the critical path to a broadly better successor.

The model need never inspect its own weights. A faster training kernel can qualify. So can a repaired data pipeline, a better evaluator, or an experimental result that changes the design of the next model. A benchmark trick that vanishes during integration does not qualify. Nor does a large pile of generated code that leaves the critical path unchanged.

FIGURE 1Recursive self-improvement as an industrial loop.<br>The definition suggests a ledger. Which AI-assisted changes survived review? How much reviewer and inference time did they consume? Which stages got shorter? Release intervals alone can mislead: a lab may wait for a market window, spend longer on safety, or train a larger system. A useful...