Who is John Galt when the prime movers train their machines on everyone else’s mind?

If you spend enough time building software — untangling a legacy estate, or pushing patches into a library you do not get paid to maintain — you eventually meet the industry’s private religion. For decades Silicon Valley borrowed Ayn Rand the way it borrows a whiteboard marker. The engineers and founders were the prime movers: Rearden, Dagny, Galt. The bureaucrats, regulators, and rent-seekers were the looters.

Large language models have put that fable under load. We are living through a stress test of Atlas Shrugged that Objectivism did not write. The people building the machines look like heroes in the lab and something closer to Orren Boyle in the hearing room. The people whose work those machines digested are being asked to treat a taking as progress.

The Reardens of the GPU

At first glance the fit is almost too clean. The researchers and engineers pushing neural nets are doing what Rand said producers do: they apply reason, mathematics, and brutal quantities of compute, and the result is a jump in what a single mind can finish in a day.

The political reaction fits the novel too. Licensing schemes, pre-deployment permits, and “safety” agencies that want a veto over who may ship a model look, from a distance, like the State Science Institute circling Rearden Metal — a revolutionary product treated as a public hazard because it would rearrange the existing order.

That is only the first half of the story.

The data heist

Rand’s defense of industrialists rested on a concrete claim, not a mood. You own the product of your mind when you have turned it into a work: a book, a patentable process, a piece of code released under terms you chose. She was not arguing that every idea on earth is a privately fenced pasture. She was arguing that you do not get to seize the mill and call the seizure civilization.

That is where the LLM era splits the Randian cast list.

Frontier models were not trained on a neatly licensed warehouse of volunteers. They were trained on industrial crawls of text, code, and images — much of it public-domain or permissively licensed, much of it not, and some of it taken from libraries that everyone in the building knew were pirate archives. The legal fight has started to sort those piles. Training on a work can be called transformative in one courtroom and still sit next to a finding that storing millions of pirated books was not a research inconvenience but a theft. Anthropic’s settlement with authors, on the order of a billion and a half dollars, is what the second pile looks like when it finally meets a docket. The New York Times and the Authors Guild versus OpenAI and Microsoft is what the first pile looks like when nobody will admit which theory of fair use survives contact with a product that competes with the original market.

From the open-source side the insult is specific. MIT and Apache were offers to use and modify code. They were not a silent grant to mint a closed appliance that answers questions with a statistically melted version of that code and then bills by the token. Stack Overflow and Reddit did not lock their APIs because they suddenly hated curiosity. They locked them because the bargain had changed: the commons was being reclassified as ore.

To a strict reading of Rand, this is the part of the script nobody cast. Are the labs heroic because they built the furnace, or are they looters because the charge they poured into it was other people’s unlicensed work? The honest answer is that they can be both in the same fiscal quarter.

The modern strike is not a valley

In the novel the producers walk out. They take their minds to Galt’s Gulch and let the looters discover that slogans do not pour steel.

Something like that is happening. It is just not a gulch.

Authors sue. Artists reach for Glaze and Nightshade — tools that cloak a style or poison a scrape — and watch the other side answer with strippers like LightShed. Platforms stop handing out firehoses. Developers change licenses after the fact, which is legally messy and morally readable: this was not what I offered you. That is a strike of the mind in the only form a networked civilization still permits. It is refusal, not disappearance.

Take the analogy as far as it will go and then stop. Galt could drop off the map. We cannot, not if we still want banks, packet networks, foundries, and a cloud account. AI did not invent the surveillance state. It compresses what governments and platforms already collect — identity, payments, location, metadata, filings — into something that can be queried like a model of the population. A hidden high-tech enclave that stays economically alive and legally invisible is a period piece. Romantic exit fails.

Withholding does not. You can be fully visible and still refuse a training license. You can be fully visible and still stop publishing the raw corpus you used to give away. You can be fully visible and still make the next decade of work expensive to steal. Observation raises the price of hiding. It does not write the next library for you.

There is a colder limit. The models already ingested a large part of the public stack. A strike in 2026 is mostly a fight over the margin: new books, new code, new measurements, the next paper. It will not uninvent the weights that already exist. Anyone selling the fantasy of a total walkout is selling the novel, not the constraint.

The clerical convenience and the Kafka factory

A fair objection lands here. LLMs make red tape feel soluble. A two-person shop can turn a rulebook into a checklist, draft a model card, fill the vendor questionnaire, and sound like it has a compliance department. The old cartoon of the producer drowning in triplicate while the looter speaks fluent regulation is weaker if everyone rents the same tireless clerk.

Believe that, and stop there, and you will miss the other direction of the same tool.

The scarce resource in a ministry used to be the human who had to write the instrument. That scarcity was a brake, however ugly. The brake is gone. A small staff can now emit interlocking definitions, annexes, evaluation protocols, and guidance that only resolves if you were in the room when the adjectives were chosen. Complexity that once looked like a rushed memo now arrives polished enough to pass as law. Generating the forty-seventh subcategory of “AI system” is cheap. Deleting it is still politically expensive. The corpus of rules grows like a model trained with no stop condition.

So the technology does two jobs at once. It cheapens the form for the firm that already knows how the gate works. It cheapens the gate-making for everyone who lives downstream of the form. Cheap compliance for the incumbent and cheap sludge for the independent developer are the same machine pointed at opposite desks.

The moat was never the PDF. The moat is permission, liability, capital, certified evaluators, compute thresholds, and the quiet fact that interpretation is a scarce resource. An LLM can write the application. It cannot grant the license. It can explain the safety case. It cannot make “state of the art” mean the same thing for a lab with a policy shop and a maintainer with a day job.

The crony moat

Rand’s special contempt was reserved for the businessman who invites the state in so the state will freeze his rivals. That type did not go extinct. It learned to speak safety.

Some frontier labs talk like they want an aviation authority or an IAEA for weights. Some of the same voices that warn about open models also happen to be the voices that would be most inconvenienced if open weights stayed cheap. Inside the industry the split is no longer theoretical: closed labs argue that capability must be gated; the open-weight camp argues that the gate is the point. You do not need a conspiracy board. You need the ordinary observation that rules requiring eval infrastructure, reporting staff, and a Washington relationship will not fall evenly on Meta’s Llama line and on a student fine-tuning in a rented GPU cluster.

From the trenches this is familiar. First you train on a commons you did not pay to assemble. Then you discover that further competition should require a permit. Rearden built a metal. Boyle built a committee.

Who owns the increment

The AI era does not ask you to choose between “the engineers are geniuses” and “the scrape was a taking.” Both can be true. The uncomfortable sentence is the one Silicon Valley’s Rand fans were not trained to say: a prime mover who needs a living commons of code and writing cannot treat that commons as unowned raw material and then ask the state to keep the resulting engine behind a fence.

Galt’s old question still works as a taunt. It is the wrong question now.

The right ones are narrower, and they are the ones a working engineer actually meets. Who owns the data the model was trained on when the source was a pirate shelf? Who owns the next work, after the last decade has already been melted into weights? Who is John Galt when the directive was drafted by a model, interpreted by a model, and enforced against the people who cannot afford the model that large — and when disappearing into a valley is no longer a plan, only a chapter heading?

The strike that can still succeed is not invisibility. It is the refusal to donate the next increment of mind to a system that has not yet decided whether you were a partner or a deposit.