On 10 September 2026, DeepSeek published a model called V4.1-Flash. It carries an MIT licence, the most permissive licence in circulation. Anyone may download the weights, modify them, pass them on and deploy them commercially without asking permission.

The size figures are not what matters. One line in the benchmark table does. On CyberGym, a testbed for finding and exploiting software vulnerabilities, the model scores 88.1 and sits ahead of the closed frontier models from OpenAI and Anthropic. On SEC-Bench Pro it reaches 62.8.

These numbers were reported by the developer about itself. No outside party verified them. They remain the best we have, and nothing about that changes without an independent measurement.

Elsewhere the same model falls back sharply. On the hardest multi-step agent tasks it trails the frontier by roughly twenty points. Anyone measuring the distance to the leading edge will find it there. Anyone asking what is already in circulation reads the cyber line.

Within a month the platform recorded more than two hundred and forty thousand downloads. Every one of those copies sits on a drive nobody can reach any more. There is no recall for weights. That is the condition from which everything else has to be thought through.

The Answers Came Afterwards

Two days later, Dario Amodei, chief executive of Anthropic, published an essay calling for a slower pace of capability gains. He asks for one or two additional years so that safety practice can catch up. He proposes giving independent evaluators permanent access at the level of an employee. Sam Altman and Elon Musk joined him publicly.

On 13 September, Xi Jinping announced at the BRICS summit in New Delhi that China would lead the creation of an open-source zone for artificial intelligence. It covers cooperation on large language models, specialist seminars and training courses. In the same speech he called for a broadly supported global framework for governing the technology.

That same day, in Doonbeg, Donald Trump rejected any slowdown. Whoever leads in artificial intelligence wins, and the United States must not surrender its edge over China. He called the people raising warnings negative forces. Some measure of regulation might be needed, he allowed, without naming which.

All three answers came after the release. They negotiate over a condition that had already arrived by the time they were written. That puts the same question to each of them. Who does it actually reach.

The Strongest Law and Where It Stops

Europe holds the most robust text this field has. Since 2 August 2026 the European Commission has been able to exercise real powers over providers of general-purpose models. The AI Office in Brussels may demand technical documentation, run its own evaluations, require mitigation measures and remove a model from the single market. Fines reach three per cent of worldwide annual turnover or fifteen million euros.

This exceeds anything any other jurisdiction has to show. The text exempts freely licensed models from part of the obligations while keeping them bound once they carry systemic risk. The construction is carefully built.

It bites where someone offers, sells or operates. A file of weights travelling from Hangzhou to a server in Lagos, Karachi or Belo Horizonte never touches the European market at all. Twenty-six major providers have signed the code of practice for general-purpose models. The signatories were never the problem.

Law works through jurisdiction. Weights travel without it. That is not a flaw in the European statute. It is the outer limit of what a single legal space can achieve. Anyone who sees that limit starts looking for something that extends past it. That is precisely what the laboratories are attempting.

What the Laboratories Are Attempting Privately

Amodei's proposal is the most honest of the three answers because it applies to its author. Permanent evaluator access at employee level increases scrutiny of his own house. Anyone building a moat would demand licensing duties for others. He would not let strangers into his own rooms.

Three or four American houses would end up bound, along with a handful of European ones. Not bound would be any operator who downloads an open model and strips the safety training, which takes modest compute. Not bound either would be anyone who trains a comparable model on rented capacity a year from now.

The circle is therefore narrower than the European market rather than wider. A private agreement reaches less than a law, not more.

That is the admission running through the essay. Nobody asks for an agreement among competitors while holding the state. America's laboratories are trying to arrange privately what their own government refused them that same weekend.

The Refusal That Closes Nothing

The position from Doonbeg is not an instrument. It is a refusal of all of them. It obliges nobody, not even its own industry.

It still has an effect, and one rarely named. As long as the world's largest economy presents any slowdown as self-harm, every international negotiation gets recast as a zero-sum contest. Speak about evaluation procedures in that climate and you appear to be haggling over market share rather than safety. That poisons the ground on which a shared rule would have to grow more thoroughly than any counter-position could.

The symmetry is worth noting. Beijing offers open weights and calls for a global framework in the same breath. Washington keeps its weights closed and rejects the framework. The roles one would expect have been swapped.

Of the three answers, one alone has a structure behind it. That structure is being built right now, and it is not being built here.

The Structure Now Growing

On 16 July 2026 in Shanghai, twenty-nine states signed an agreement founding the World Artificial Intelligence Cooperation Organization, known as WAICO. Membership now stands at thirty-seven. Its seat is Shanghai. The body sits deliberately outside the United Nations system. The Secretary-General of the United Nations attended the signing.

Founding members include Russia, Brazil, South Africa, Indonesia, Pakistan, Kazakhstan and Ethiopia. None of the major Western democracies belongs. Neither does India, although India hosted the summit at which the open-source zone was offered. Even the bloc the offer addressed does not follow it as one.

I have measured the pace of the competing format once before in this series. That objection falls short here. What decides is not how often the United Nations global dialogue convenes. What decides is what it brings when it does.

It brings the widest membership this field has. It brings nothing that turns membership into obligation. At WAICO the arrangement runs the other way. It gathers fewer states. It rests on a signed agreement with a council, a secretariat and an accession procedure.

One format has the breadth without the binding force. The other builds the binding force and collects the breadth afterwards. Together they still do not add up to a rule that holds.

What the Rule Would Have to Do

That describes what is missing. What is missing is a binding law on the handling of artificial intelligence that does not end at the reach of one legal space and does not end at the signatory list of a private pact. I know how remote that sounds. I still hold it to be the only target that matches the situation, and the paragraphs above already say what such a law would have to do.

It would have to attach at publication rather than at market entry. The European text fails not through weakness but because it presumes a market that a travelling file never enters. Release weights above an agreed capability threshold and the duties would follow, wherever the server happens to stand.

That threshold would have to hang on capability rather than on a list of houses. The private pact fails at exactly that list, because it goes stale with every new entrant.

A threshold in turn is worth only as much as its measurement. The numbers this brief opens with come from the parties meant to be examined. No serious arms control treaty has ever been built on self-declaration, and a treaty over capabilities of this kind cannot be either.

That leaves the signatures. An agreement without Beijing leaves free the very weights at issue. An agreement without Washington leaves out the largest closed capacity in the world. The two negotiating rooms that stand side by side today have to touch, and neither will do it out of insight. It takes states that sit in both and make the joining their own cause.

The timing is still open, and it is not a question of good will. Months separated the last two model generations from the same house, not years. A treaty that takes longer to negotiate than the doubling of the capability it governs arrives to find a subject that no longer exists in that form.

I therefore hold a schedule to be unavoidable, one measured against that cadence rather than against the comfort of committees. The next session of the United Nations global dialogue in May 2027 has to produce a negotiating mandate rather than another declaration of intent. The signed text has to exist by the end of 2028. Let that gap open once and you spend the rest of the negotiation discussing a condition you can no longer change.

Why It Has to Reach Past Earth

A law meeting that description would be planetary. It would still not be finished.

This series has covered the filing for up to a million data-centre satellites once before. The question then was the enclosure of the sky. The same matter now presents itself differently. A model running on a satellite enters no single market and touches no national jurisdiction. Orbital compute has stopped being a thought experiment and is under construction at several companies.

That moves the question outside international law as we have it, which was written for territory. Every rule we formulate now will be inherited by the first installations beyond Earth. It will be in force there before anyone has tested it for that case, or it will not apply there, and then nothing will.

A rule for artificial intelligence therefore has to be built as multiplanetary from the start. Not because the framing sounds larger, but because a rule written for land surfaces fails at the first installation that needs no land. We made this mistake once already with the Outer Space Treaty, where we settled ownership and left out operation.

I regard any rule for artificial intelligence that ends at a national border as ineffective. Anyone who sees it differently should name me the border at which a file stops.

Homepage: https://planet-futures.org