The most advanced large-scale models are beginning to be subject to export controls like enriched uranium

Last Friday, a letter sent two of the world’s most powerful AI systems offline simultaneously. A US Department of Commerce export control order prohibits any foreign citizen from accessing Anthropic’s Fable 5 and Mythos 5 models. Due to the inability to identify user nationality in real-time, Anthropic did the only thing it could do: shut down these two models, which were just released three days ago, for everyone worldwide.

For the first time, humans have incorporated an intelligent agent in the form of bits into the same export control framework as enriched uranium.

The script of enriched uranium is being replayed
Export control has only been applied to two types of things in history: hardware and formula. As we know, there are uranium enrichment centrifuges, high-end lithography machines, military grade encryption algorithms, and so on.

Their commonality is physical scarcity. Without providing you with the equipment and drawings, you won’t be able to create it. Effective regulation is due to the physical boundaries of the controlled items, which can be intercepted at customs and traced back in the supply chain.

But Fable 5 is a set of weight parameters. It can be infinitely replicated without going through customs, without the need for containers, and there is no physical action of “smuggling”. In theory, it can wake up on any server worldwide. The US Department of Commerce has found that all traditional tools are ineffective, and it cannot intercept Fable 5 at the border. It can only cut it off at the source.

The truly regulated objects are the “ability density” condensed by this set of weight parameters: code generation ability, reasoning and planning ability, and cross domain knowledge calling ability. When these abilities are dispersed in the brains of countless engineers, they have never triggered regulation. But when they are compressed into one model, when a person can invoke all of these abilities with just one prompt word, compression itself poses a threat.

This is the precise mapping of enriched uranium logic in the digital world.

Uranium ore is distributed throughout the Earth’s crust and has never been regulated; But when it is concentrated to a certain abundance, it becomes the most closely monitored substance on Earth. The same applies to the model: when a single capability is dispersed across open source code repositories, technical Q&A, and academic papers, it is free. When all these capabilities are condensed into a single point callable interface, it crosses that threshold, and the cost of crossing that threshold is being shut down.

The history of enriched uranium provides a mirror for understanding this matter.

In 1938, Hahn and Strassman discovered nuclear fission in Berlin. In the following decade, uranium transformed from an obscure laboratory element into the most sensitive strategic material on Earth. In 1946, the United States passed the Atomic Energy Act, which placed all nuclear technology, materials, and knowledge under government control. Private capital has been expelled from the nuclear energy field, international exchanges among scientists have been cut off, and even basic physics data has been classified. A pure natural element that has been shackled by politics ever since.

Eighty years ago, the regulation was based on a cold logic: some forces were too powerful to be controlled by any entity that did not consider national interests as the ultimate consideration. Eighty years later, the same logic may be activated again, with the target shifting from nuclear fission to forward propagation of neural networks.

In the next decade, three things will happen
The regulation of enriched uranium gave rise to a new international governance structure in the 1950s: the International Atomic Energy Agency, the Nuclear Non Proliferation Treaty, and the Supplier Group. Once technology regulation is established, it will irreversibly move towards institutionalization, multilateralism, and permanence.

AI is no exception. In the next decade, there is a high probability that three things will happen.

1. Ability review will move towards institutionalization.

Every new cutting-edge model will not only undergo security red team testing before release, but also undergo third-party compliance review authorized by the government. The review criteria will not come from within the company.

The evaluation of model capability will shift from “scoring” to “checklist”. Each ability on the list that may be abused will trigger additional regulatory requirements. The “abundance” of the model – parameter size, inference depth, cross domain generalization ability – will be accurately measured and threshold set, just like the concentration of uranium-235. Models exceeding a certain threshold automatically trigger export control clauses.

2. The jurisdictional boundaries will become blurred.

Fable 5’s commands are targeted at “foreign citizens” regardless of their geographical location. The regulatory tentacles of the US government have extended to every user worldwide for the first time.

A developer in Singapore, using an API from a US company, is subject to US export control laws. The compliance obligations of an AI supplier for a company based in Berlin do not depend on German law, but on a single order from the US Department of Commerce.

This unilateral expansion of jurisdiction will force non US companies to rethink their AI supply chains: the US suppliers you rely on may one day be required by the US government to cut off your access. The answer given by the Fable 5 incident is very clear: it can.

3. The technological path will move towards fragmentation.

When the closed source frontier model faces the risk of repeated power outages, the global AI industry will be forced to move towards a dual track system.

One track is the closed source frontier model of the United States, which is subject to export control constraints and carries the risk of being taken down with every release. Another track is open source models, locally deployed models, and models from non US jurisdictions, which may not be as advanced as the former, but they are not threatened by US government power outages.

The market share of open source models will not only be driven by performance, but also by the security attribute of ‘not being unplugged’. In the past three years, open source models have been catching up with closed source models in terms of capabilities; In the next decade, open source models may have a structural advantage over closed source models in terms of reliability.

The deepest crack lies in the property rights system
All the above deductions are based on an unanswered fundamental question. The deepest crisis exposed by the Fable 5 incident is that digital civilization has yet to establish a property rights system for “intelligence”.

Legally, models are sold as a service. You pay, I use my assets to do things for you. My assets are always mine, what you buy is only its output. This logic has been operating in the traditional service industry for thousands of years and has never encountered any problems.

But AI is different. After spending three months tuning all internal tools, training employees, and writing hundreds of automation scripts that rely on the specific output format of Fable 5 based on its specific behavior patterns, your enterprise has effectively turned Fable 5 into your production material.

But legally, it is still Anthropic’s service. It can be reclaimed on any day, and the compensation you can receive does not exceed the subscription fee you paid in the past month.

You have invested real production materials and received legal protection at the service level. The difference between the two is the loss suffered by all global corporate customers when Fable 5 was taken down. This loss did not appear on any balance sheet, did not trigger any insurance claims, and there are no legal provisions that can cover it.

It took three hundred years for humans to establish a legal system regarding ‘property’. A piece of land, a factory, and a patent all have clear ownership, transaction rules, and dispute arbitration mechanisms. But this system has a default premise: property is tangible, or at least has a traceable carrier.

A model has been unplugged, you can’t do anything. It was neither stolen nor destroyed. It still exists, just not for you to use. This is a completely new form of deprivation: what is deprived is not the object, but the right to use. And legally, the right to use has never belonged to you.

Enriched uranium has been regulated for 80 years and remains the most sensitive technological asset of humanity to this day. The regulation of AI has just begun, and its endpoint may be a permanently divided digital world. In this world, the smartest model may not necessarily be the one that can be used. The model that can be used must be the one with the clearest property rights. Not being taken away, at a certain historical juncture, is much more important than being ahead for a while.