Note: This is a research note supplementing the book Unscarcity, now available for purchase. These notes expand on concepts from the main text. Start here or get the book.

The Tool That Trains Its Replacement

In mid-June 2026, SpaceX bought Cursor, the AI code editor that millions of developers open every morning, for a reported $60 billion. Three weeks later, Musk’s xAI shipped Grok 4.5, a coding model trained on real Cursor developer sessions and priced at $2 per million input tokens, under half the $5 that Claude Opus 4.8 and GPT-5.5 charge. It landed fourth on the Artificial Analysis Intelligence Index, above every open-weight model. The press branded the launch “SpaceXAI.” Developers called it something else once they did the arithmetic.

Here is the arithmetic. Every keystroke, every accepted suggestion, every “no, not like that” correction a developer typed into Cursor was a labeled training example: a demonstration of how a skilled human turns intent into working software. Cursor’s users spent two years teaching the editor what good looks like. Then the editor’s new owner distilled those lessons into a model pitched to do their jobs faster, cheaper, and without the developer. The tool trained its own replacement, and the people doing the training paid a monthly subscription for the privilege.

This is not a Cursor problem. It is becoming the defining economic loop of the AI era, and it deserves a name.

The Data Flywheel, Pointed at You

Silicon Valley has a friendly term for this loop: the “data flywheel.” More users generate more data, which trains a better model, which attracts more users, which generates more data. Spin it fast enough and competitors can never catch up, because they can’t buy the one input that matters most: millions of humans doing real work inside your product.

The flywheel is old. What changed in 2026 is where it points.

For most of the software era, the tools you used to do your job were sold by companies that wanted you to keep doing your job well. Microsoft wanted Excel users to stay employed and keep renewing. Adobe wanted photographers to keep shooting. The vendor’s incentive and the worker’s incentive pointed the same direction, roughly, because the vendor sold productivity to the worker.

Now the vendor increasingly sells the worker’s output capacity to the worker’s employer, and the worker is the training set. When the company that owns your primary instrument of labor is also building the autonomous system meant to replace that labor, every hour you spend in the tool does double duty. It gets today’s task done, and it teaches tomorrow’s model how to do that task without you. You are simultaneously the customer, the quality-assurance department, and the raw material.

This is the loop the AI coding revolution has been quietly running all along. When 51% of the code committed to GitHub is already AI-assisted, that is not just a productivity statistic. It is a measurement of how much human coding judgment has been captured, labeled, and fed back into the models. The junior developer writing boilerplate under an AI’s supervision is not only doing a job. They are narrating a tutorial for their own successor.

We’ve Fed the Machine Before, But Never Like This

We have been unpaid data laborers for years. When you solved a reCAPTCHA to prove you were human, you were labeling images for self-driving cars. When you searched Google, you taught it which results were good. Tesla built its Autopilot data advantage by watching millions of drivers handle real roads: the same trick, wearing a seatbelt.

But those systems captured a narrow signal. reCAPTCHA learned whether a square contained a traffic light. Search learned which blue link you clicked. Autopilot learned how a foot moved between two pedals. The machine got a thin slice of behavior, useful only for one narrow task.

A coding session captures something far richer: the entire arc of cognitive work. Not just the final answer, but the reasoning that produced it. The dead ends you backed out of. The moment you realized the architecture was wrong and refactored. The taste that told you this variable name was clearer than that one. The whole loop of intent, attempt, evaluation, and revision, which is precisely the loop that separates a professional from an amateur, is now the exportable product.

That is the leap. Earlier data capture harvested the exhaust of knowledge work, the clicks and traces we left on the side. The new generation harvests the engine, the judgment itself. And once a workflow has been fully instrumented, the distance between “AI that assists you” and “AI that is you, minus the salary” collapses toward the substitution threshold: the point where the cheapest reliable provider of a task is no longer human.

Why Owning the Whole Stack Changes the Prize

The Cursor acquisition matters because of what it stacked together. Musk’s world now reaches from the chip fab through the compute clusters to the model to the editor where the work happens. When one owner holds every layer, the data flywheel stops being a metaphor and becomes a wholly owned pipeline: developer types in the editor you own, the session trains the model you own, running on the silicon you own.

Vertical integration like this is why the gains of cognitive-labor deflation concentrate rather than spread. The commoditization of intelligence has already made raw model capability cheap and fungible; US firms now route nearly half their AI tokens to Chinese open models at a fraction of the price. When the model itself is a commodity, the durable advantage shifts to whoever controls the scarce, non-fungible inputs: the workflow data, the distribution, the place where humans actually do the work. Owning the editor is owning the well the whole industry drinks from.

This is the same move the book flags in Compute Landlords: the builders of the AI era keep discovering that the real power is not in producing intelligence but in owning the toll roads it travels. Cursor is a toll road made of your own labor. Every trip you take widens it for the owner and narrows the shoulder you might have walked on.

“But Every Tool Learns From Its Users”

The obvious objection: hasn’t every tool always improved by watching its users? Spell-check got better because we typed. Maps got better because we drove. Is this really different, or is it just the usual grumbling that greets every new machine?

It is different, on one specific axis: the alignment of interest.

When a tool improves to make you better at your work, you and the vendor are on the same side. A sharper spell-check makes you a better writer and keeps you subscribed. When a tool improves to make you unnecessary, owned by a party whose business case is your obsolescence, the interests split. The better the tool gets at learning from you, the less it needs you. You are not a customer being served. You are a resource being depleted, and the depletion is dressed up as a feature you pay for.

The tell is what happens at the limit. Spell-check that reaches perfection still needs a writer to have something to say. A coding agent that reaches the skill of the developers who trained it needs those developers for exactly one thing: a final tranche of training data before the model no longer requires them. This is the mechanism behind the death of the junior developer that the labor numbers already show. The apprenticeship that used to turn juniors into seniors has been automated into a data-collection stage.

The Involuntary Apprentice

There is an old, honorable arrangement where a skilled worker trains a replacement. It is called mentorship, and it runs on consent and reciprocity: you teach the apprentice, the apprentice supports the craft, the knowledge outlives you. The medieval guild made this a sacred contract.

The flywheel keeps the training and quietly deletes the reciprocity. You teach the model, and the value of what you taught flows entirely to the owner. No credit accrues to you. No stake in the resulting model is set aside. The terms of service you clicked through years ago called your work “usage data,” and usage data has no author. You are an apprentice whose master is a corporation, whose lessons are extracted rather than given, and whose graduation ceremony is your own redundancy.

This is where the Unscarcity framework gets specific, because it treats contribution as something that should be recognized and rewarded, not silently harvested. In the book’s model, meaningful contribution earns Impact, a currency of acknowledged value that flows back to the contributor. The involuntary-apprentice problem is what you get when a civilization has powerful ways to capture human contribution and no honest mechanism to credit it. The data is worth trillions. The people who generated it were billed for the software.

The Unscarcity Lens: Own the Flywheel or Be Owned by It

None of this is an argument against the tools. Grok 4.5 at $2 a million tokens is a genuine gift to anyone with an idea and no budget, and the democratization of building is real. The argument is about who captures the compounding value once the tools have learned everything we know.

Left alone, the loop drives straight toward the two failure modes the book was written to prevent. It accelerates the Labor Cliff by turning every skilled workflow into training data for its own automation, and it concentrates the winnings by handing the flywheel to whoever already owns the stack. More capability, fewer owners: the exact shape of the AI talent and value paradox, where abundance in one layer manufactures a new aristocracy in another.

The Unscarcity response is not to smash the flywheel. It is structural, and it comes in two parts.

First, decouple survival from labor. If your rent does not depend on out-competing a model trained on your own past keystrokes, the flywheel stops being an existential threat and becomes what it should be: a very good tool. That is the entire point of the Foundation, the guaranteed floor of housing, food, energy, healthcare, and compute that makes the automation of your job a change of activity rather than a catastrophe. It is why the book argues for income floors like Universal High Income as a bridge, not as charity but as the precondition for facing the flywheel without fear.

Second, confront ownership directly. A flywheel powered by the collective labor of millions is a commons that has been quietly privatized. Whether the answer is data dividends, worker-owned model cooperatives, or public stakes in the systems trained on public behavior, the question the book insists we stop dodging is the one Cursor’s developers asked the morning Grok 4.5 shipped: if my work built this, why do I own none of it?

The tool that trains its replacement is not a glitch in the AI economy. It is the AI economy, stated honestly. The only choice left is whether the people feeding the machine get a seat at the table it is setting, or just the bill.

That choice is what Unscarcity is about. Start with the framework, or read the book.

References

• Introducing Grok 4.5 - xAI
• SpaceXAI Releases Grok 4.5, a Cursor-Trained Model - MarkTechPost
• The AI Coding Revolution - How coding became the first cognitive job to hit the cliff
• The Substitution Threshold - The line where the cheapest reliable provider stops being human
• The Commoditization of Intelligence - When models are fungible, the moat moves to workflow and control of supply
• Compute Landlords - Why owning the toll roads beats producing the intelligence
• The AI Talent Paradox - How abundance in one layer breeds a new aristocracy in another
• The 2025-2030 Labor Cliff - The crisis the flywheel accelerates
• The Foundation - Decoupling survival from labor
• Impact - Crediting contribution instead of silently harvesting it