Everyone swears to share the power

On 10 August 2026 Mark Zuckerberg published “The Future Is for Everyone”, six and a half thousand words on Meta's vision for what it calls personal superintelligence ^[1]. The essay holds that the greatest risk of advanced AI is that its power concentrates in a single entity ^[1]. It orders the alternative into three principles: individual empowerment as the source of prosperity, invention as the primary purpose of superintelligence, and balance of power as the foundation of safety ^[1]. The thesis did not stay on paper. That same day Meta released Muse Glimmer, an agentic model of thirty billion parameters under an Apache 2.0 license, which runs whole on a 24 GB consumer graphics card and works offline once downloaded ^[2].

Five days later Dario Amodei, Anthropic's chief executive, published on X his longest defense of the opposite position, inside an exchange with the investor Gavin Baker ^[3][4]. The larger phenomenon of which these two texts are the best available example is not a rivalry between laboratories. The constitution of superintelligence is being drafted in public, by its interested parties, and both camps swear to share out the power while each centralizes a different layer. Latin America has no seat in that drafting. It is, at the same time, the world's largest third party among those affected by what gets decided there.

Two camps that swear the same thing

The distribution camp arrives with artifacts as well as arguments. Zuckerberg promises free versions of his tools for billions of people and describes superintelligence as a “tool of invention” rather than of automation ^[1]. He announces besides that Meta will “resume releasing some open source models soon”, the line the company itself had closed at the end of 2025 ^[1][16]. The announcement came fulfilled from the factory: Muse Glimmer understands more than a hundred languages, plans multi-step tasks, calls external tools and retries on its own when one fails, all on a gaming laptop, with no cloud and no subscription ^[2].

The letter that preceded the essay had drawn that camp's size. Jensen Huang opened his X account on 24 July with “Open Weights and American AI Leadership”, an open letter asking Washington not to restrict downloadable weights ^[8]. It went out with 25 signatures and doubled in a day, with OpenAI, Google, Microsoft and Meta inside ^[8]. Anthropic and Amazon never appeared ^[8]. Amodei clarified three days later that his company has “never advocated for a ban on open-weight models”, a precision that arrived once the absence was already being read as protectionism ^[10].

The institutional camp had been writing since earlier. Amodei had published in June “Policy on the AI Exponential”, which proposes subjecting frontier models to mandatory technical testing, “like airplanes”, with state authority to block an unsafe deployment ^[5]. Demis Hassabis proposed on 14 July a FINRA-style standards body, funded by the industry, to which laboratories would hand their models up to thirty days before release ^[6]. Sam Altman and Elon Musk each described it, separately, as a serious framework ^[6]. On 28 July, 1,178 employees of the frontier laboratories asked the US government to support “the technical and governance tools needed to deliberately pace” automated AI development ^[11]. The signatures include Amodei's, Ilya Sutskever's and Shane Legg's, and OpenAI and Anthropic endorsed the letter as companies within hours ^[11]. None of these pieces presents itself as control for control's sake. All of them are sold as counterweights.

What each mechanism concentrates

The underlying disagreement between the two camps is not whether power should be shared out but which layer of the system shares it. Amodei's reply calls false the choice between concentrating through regulation or distributing without it ^[3]. Institutions, he wrote, at their best “vest power in ideas rather than people”, and thereby decentralize it ^[3]. His countercharge against open weights is structural: they help, he concedes, but they “shift the concentration somewhat to those with the most compute and chips” ^[3]. The uncomfortable datum in that sentence lives in the open-weights letter itself, which was organized by the maker of the chips ^[8].

The symmetric argument cuts just as clean. Publishing weights disperses the artifact irrevocably, because a downloaded copy cannot be recalled, but it leaves the production of the frontier exactly where it was: with whoever can pay for the compute. Meta's distribution is, besides, a company's promise and not a property of the system ^[1]. The Muse line was closed at the end of 2025 and reopened nine months later, both times by the decision of a single person ^[1][16]. Institutions carry the inverse defect. The authority to test, license and block gets deposited somewhere, and that deposit is a concentration which, once constituted, becomes an object of politics and of the bureaucracy that administers it. Nothing in the proposals on the table is institutionally new: the mold is an aviation agency founded in 1958 and a securities self-regulator from 2007, aimed at new subject matter ^[5][6].

The regulator already in office signed no essay

While the two camps publish essays, power over the models is exercised through a channel that appears in neither. A June executive order created a voluntary thirty-day testing regime for frontier models at CAISI, the US government's new standards center ^[7]. The framework for developers of frontier models was finished in early August and was not made public ^[13]. The Center for American Progress describes it as a de facto licensing system, with a warning that sums up the record: a program is voluntary in name only when the administration punishes whoever does not bend to its requests ^[13].

The summer's dates sustain that reading. Export controls reached Anthropic's models on 12 June, and coverage of the case records Fable 5 back online on 1 July ^[13][9]. OpenAI launched a feature on 26 June clarifying that it was starting with “a limited preview” at the government's request ^[13]. July's coverage described, besides, the White House deciding discretionally who accesses which frontier model ^[14]. On 22 July the White House science office accused Moonshot AI of having distilled Fable 5 to build Kimi K3, and the Treasury threatened sanctions ^[9]. Two days later the open-weights letter came out, whose ninth paragraph asks that distillation —the mechanism edition 004 dealt with— not be confused with misappropriation ^[8][17].

Amodei himself declared he was “very supportive” of the approach the press attributes to the administration: pre-deployment testing for frontier models and testing also for open weights “when they get closer to the frontier” ^[3]. The institutional camp's thresholds exist, and they exempt. California's SB 53, which Anthropic supported, leaves every company with annual revenue under five hundred million dollars outside its hardest obligations ^[12]. Europe, meanwhile, stopped debating. Since 2 August every AI that converses with people must identify itself as such, and synthetic content must carry a machine-readable mark ^[15]. Legal analyses underline that the obligation reaches whoever serves European users from anywhere ^[15].

The public argument is over how to share out a power that is, meanwhile, already being exercised unshared.

The concentration of compute is real and had better not be relativized, because open weights are born in the same factories as closed ones. Its downstream effect is, however, the opposite of every concentration known: the product leaves free of charge and every downloaded copy passes beyond its maker's reach. A funnel whose flow is given away does not work as a bottleneck for the one who drinks. Muse Glimmer running offline on a laptop in Barranquilla asks nobody for monthly permission ^[2].

Authority behaves the other way around, because it produces nothing and interposes itself in everything. A pre-release testing regime reaches open weights exactly when they begin to matter, which is when they get close to the frontier ^[3]. An export decision already switched off for weeks the model half the region depends on for work ^[13][9]. The chip coalition proposed in Amodei's essay divides the world between allies who share and adversaries who are denied ^[5]. Latin America appears on neither list, and not appearing has a precise meaning: its access is a residue decided somewhere else.

The question the region should be asking itself is whether its ceiling of available intelligence will be set by regulations of countries where it does not vote. The honest answer is that it has already begun. Europe's Article 50 has bound, since August, any regional builder serving users in Europe ^[15]. The interfaces the region rents operate under a framework its own regulator chose not to publish ^[13]. The missing link is the most valuable one: if the testing of open weights hardens, the free copy that today crosses borders without asking permission will need a license to exist, and that license will be processed in Washington. The only thing no future rule can close retroactively is what is already downloaded.

The institutional counterpoint has to be conceded whole before it is objected to. The thresholds are designed to exempt the one climbing: everything the region builds today sits under SB 53's five hundred million, and the testing regime Amodei defends brakes the one on top more than the one catching up ^[3][12]. Against a technology that concentrates by structure, a counterweight on the frontier actor also protects the small. The objection is not against the counterweight but against the seat: an exemption without a chair is a favor and not a right, and regulatory favors have the half-life of the politics that grants them. Brussels took three weeks to reach the region's builders.

THE PLAY

1. Treat the open frontier as strategic inventory, not as a service. What is published cannot be unpublished, so downloading, mirroring and keeping local custody of the weights the region uses turns that property into a dated asset: every weight archived today is ceiling guaranteed tomorrow, whatever the coming license says.
2. Build your own evaluation before your own model. Knowing what these models know in the region's terms —its languages, its laws, its markets— is the one seat nobody can deny you at a table where you do not vote, and edition 003 showed that layer is already the real regional advantage ^[16].
3. Write distribution into your own law while others write restriction into theirs. The Brazilian framework remains unvoted in the Chamber; a jurisdiction that guarantees by law the right to run open models becomes the natural refuge of all the world's exempted, and that refuge is an industrial policy that costs not a cent of compute.