The Padlock on the Library

In January 2025, DeepSeek released R1 with downloadable weights and Nvidia lost $593 billion in market value in a day.[1] That mattered for reasons well beyond Nvidia. Wall Street suddenly had to price the possibility that very capable AI might become cheaper, easier to copy and much harder for a small number of American companies to keep behind a meter.

I think that is the real fight now: not simply who builds the smartest model, but who gets to own one.

When knowledge becomes rent

America's AI boom is tied to some of the largest companies on Earth. The Magnificent Seven make up roughly a third of the S&P 500, while J.P. Morgan has described the US economy as increasingly K-shaped, with rising asset values strongly benefiting wealthier households while many others remain squeezed.[2]

At the same time, the frontier AI companies are moving into almost every knowledge-heavy profession they can reach. OpenAI is building products and partnerships across finance, medicine, law, government, defence and entertainment. Anthropic is moving through banking, accounting, healthcare, science and government. xAI is selling Grok into business and government.[3]

These are not merely juicy markets. They are places where people learn how things are done. Medicine contains recipes for diagnosis and treatment. Finance contains methods for valuing companies and moving capital. Engineering contains ways of building bridges, chips and power systems. Science contains methods for discovering what is true. Law contains the machinery by which rules become real. Entertainment contains stories, characters and culture.

AI is increasingly sitting in the middle of all of it.

The companies do not need to own every medical paper, engineering drawing or accounting spreadsheet. They need to own the machine people increasingly ask to read, compare, explain and reason across them. If the best models are closed, their weights unavailable, their training choices opaque, their behaviour changed remotely and their use sold through an API, then we do not own the intelligence. We rent access to it. Whoever owns the server can change the price, the rules, the model and the answer.

That is why transparency matters so much to me. It is not a decorative virtue for programmers. It is part of the basis for trust.

If I cannot preserve a model, run it independently, compare it with yesterday's version or understand the broad choices that shaped it, then I am being asked to trust whoever controls the server. That might be OpenAI, Anthropic, xAI, Google, Meta, the US Government, the Chinese Government or whoever comes next. None of them deserves blind trust simply because the technology is impressive.

China does not magically solve this problem. Beijing has its own political interests, censorship system and very strong ideas about what information should circulate. That is precisely why I do not want Washington deciding what intelligence the rest of us may possess, Beijing deciding it, or the two of them agreeing that everything would be much safer if everybody else rented the approved version.

A very convenient kind of safety

Dario Amodei has called for the AI frontier to be paced, including eventual international coordination with China. Sam Altman has publicly agreed that the frontier should be paced, and Elon Musk has backed Amodei's concern.[4]

I do not need to know what is in any of their hearts to dislike where that policy could lead. Motive is almost a distraction. The important question is what happens if "frontier safety" comes to mean that the public cannot download the weights of the most capable systems, cannot run them independently and cannot examine them outside company-controlled infrastructure.

The result is concentration of knowledge and power.

The United States has already examined policy options including licensing access to model weights, restricting public distribution and forcing access through APIs or web interfaces. The same government report also recognised that open weights support competition, independent research, smaller developers and civil society.[5]

That is the awkward bit. Openness creates risks, but it also lets outsiders check the work.

Closed systems ask us to trust the people who built them, the people who aligned them, the people who decide what training material matters, the people who update them and the people who decide what they are allowed to answer. If those systems become the main interface to medicine, science, law, politics and education, "trust us" is not much of a governance model.

It is certainly an excellent subscription model.

China is inconvenient because people can take the model home

China's open-weight models are economically and politically awkward for the United States for a simple reason: other people can possess them.

A university in Brazil, a startup in Indonesia, a researcher in Australia or a government department in Africa can download a Chinese model and run it themselves. They are not forced to send every question to an American server or pay an American company every time the machine thinks.

China benefits from that, obviously. Open models spread Chinese technology, standards and influence. Xi Jinping has openly promoted international open-source AI cooperation even while China's security establishment argues for much tighter control of AI inside China.[6] There is no fairy tale here in which Beijing is the noble defender of human knowledge. There is simply competition, and competition can be wonderfully inconvenient for monopolies.

DeepSeek demonstrated the point in the language markets understand. Release a capable model cheaply, publish the weights, let the world copy it, and hundreds of billions of dollars of assumed scarcity can wobble.

That is a very large incentive for anyone selling scarcity to become suddenly philosophical about abundance.

We have already seen what opaque algorithms do

The thought that a few organisations might control the systems through which people understand politics would sound more paranoid if social media had not spent the past decade making it look rather ordinary.

Russian influence operations used American social platforms during the 2016 US election. Meta has dismantled large China-linked influence networks. Romania's 2024 presidential election was annulled after allegations involving coordinated online promotion, paid influence and cyber activity, while the European Commission opened proceedings against TikTok over possible failures around electoral interference.[7]

We do not need to claim that these systems literally changed particular election results. The documented behaviour is disturbing enough.

A social network chooses which posts move past your eyes. An AI can do something much more intimate. It can talk directly to you, remember what matters to you, explain an issue in language tuned to you, answer objections, select evidence, omit evidence and keep going until you are satisfied.

If the model behind that conversation is closed, remotely mutable and controlled by a corporation or government, the user has almost no independent way to know whether yesterday's machine and today's machine are meaningfully the same.

That is where Orwell matters.

The frightening part of Nineteen Eighty-Four is not simply that somebody lies. Humans managed that perfectly well before computers. It is that the machinery for checking the record sits inside the same system that controls the record.

A closed AI system can reproduce a little of that structure. Not because every answer is false, but because the user cannot inspect or preserve the machinery producing the answer.

Lack of transparency becomes lack of trust.

Gutenberg's rude little machine

The printing press did something deeply annoying to people who controlled knowledge.

It made copying cheap.

Books became easier to produce, harder to suppress and available to people who had previously depended on institutions wealthy enough to own manuscripts and scribes. Printing did not magically abolish corruption, poverty or inequality, but it weakened one of the mechanisms that protected them: control over who could read, copy and challenge what was written. European cities that adopted printing early subsequently grew much faster than otherwise similar cities.[8]

The press also printed propaganda, rubbish, pornography and truly dreadful poetry. Apparently civilisation staggered on.

That is why open-weight AI matters to me. A local model is not just cheaper software. It is intelligence you can possess.

You can keep it, test it, modify it, compare it and disconnect it from the internet. You can give it private documents without sending them to somebody else's cloud. Independent researchers can pull it apart, find faults, challenge assumptions and compare what one model does with what another does.

None of that guarantees truth. It gives us ways to distrust intelligently.

And distrust, properly organised, is one of the useful things civilisation invented.

The alternative frightens me far more: governments and frontier AI companies deciding that ordinary people may have access to powerful intelligence, but never possession of it.

Translated into normal English, that means you may ask the machine questions provided somebody else owns the machine. You may use the knowledge provided somebody else decides what the machine is allowed to know, how it reasons, when it changes and how much the next question costs.

That is not a public library.

It is a private library where the books stay behind the desk, the catalogue is secret, the librarian may rewrite them overnight and admission is billed to your credit card.

The really strange thing is that we already built the alternative. Open weights exist. Local models exist. Independent research exists. Cheap intelligence is starting to exist.

Gutenberg's press did not fix corruption. It made knowledge harder to monopolise, which gave more people the ability to challenge corruption, authority and inherited privilege.

Putting the padlock back on the library would reverse the trick.

I would rather keep the books where people can take them home.

Written by Geoff Fane with AI assistance for research, discussion and editing. The arguments, judgements and final wording remain Geoff's responsibility.

References

No.	Source	Title / link	Year	What it supports
[1]	Reuters	DeepSeek sparks AI stock selloff; Nvidia posts record market-cap loss	2025	DeepSeek R1 and Nvidia's $593 billion one-day market-value loss.
[2]	J.P. Morgan / Reuters	US K-shaped economy and market concentration	2025-26	K-shaped US growth and the unusually large market weight of the biggest technology companies.
[3]	OpenAI / Anthropic / xAI	Enterprise and sector deployments	2026	Expansion of frontier AI systems across finance, healthcare, science, government, defence and other professional work.
[4]	Dario Amodei / Reuters	Frontier pacing proposals and industry responses	2026	Amodei's proposal for pacing, including coordination with China, and public responses from Altman and Musk.
[5]	US NTIA	Dual-Use Foundation Models with Widely Available Model Weights	2024	Policy options for restricting model weights and the recognised benefits of open weights for research, competition and civil society.
[6]	Xi Jinping / State Council of China	BRICS remarks on open-source AI cooperation	2026	China's promotion of international open-source AI cooperation alongside domestic AI controls.
[7]	US Senate / Meta / Reuters	Documented social-media influence and electoral-interference cases	2019-24	Examples of state-linked influence operations and the risks of opaque recommendation platforms.
[8]	Jeremiah E. Dittmar	Information Technology and Economic Change: The Impact of the Printing Press	2011	Evidence that early adoption of printing was associated with substantially faster city growth.