The Future Is for Everyone. Power Is Not.
Mark Zuckerberg has published a manifesto about the future of AI, and the temptation is to decide what it means before reading it. If you admire Meta, call it vision. If you distrust Meta, call it a billionaire's self-interested defense of decentralization. Neither reaction costs much. The essay deserves more.
Its strongest claim is not about a product. It is that concentration is itself an AI risk.
“Humanity is not a monoculture,” Zuckerberg writes. That sentence is right. No single model, company, government, or committee can represent billions of incompatible ideas about a good life. His proposed counterweight is personal superintelligence: capable AI in many hands, aligned to many people, helping them create, learn, work, and negotiate with institutions that already possess enormous computational power.
That is a serious argument. Treating centralization as safety does not eliminate the danger of power; it transfers the danger to whoever operates the lock.
The case for open models belongs here too. The International AI Safety Report 2026 finds that open-weight systems can broaden research, lower barriers, and help smaller organizations compete. It also warns that released weights cannot be recalled, safeguards are easier to remove, and misuse becomes harder to trace. The report describes Meta's Llama releases as open-weight rather than fully open source because Meta withholds training code and data and uses restrictive licenses. These are not contradictions to wave away. They are the problem. Concentration creates one class of risk. Uncontrolled proliferation creates another. Anyone selling only one side is not describing the future; they are describing their preferred owner of it.
Zuckerberg does more than offer a slogan. He proposes free versions, with additional compute allocated through a paid auction; a fully private mode for personal agents that he says even Meta could not access; renewed model releases; Meta's board of directors empowered to approve release-safety criteria and review compliance; government access to intermediate training checkpoints and technical staff; and a community compact for the places carrying the physical burden of data centers. These promises are specific enough to be tested. That matters.
But the essay repeatedly moves from distributing capability to distributing power as if they were the same thing. They are not.
A billion people can use the same agent and still be tenants. If one company controls the model, the memory, the compute queue, the release rules, and the path out of the system, it holds a different kind of power from the person typing into it. Giving a tenant and a landlord equally capable AI lawyers does not make housing court equal. One still controls the property, the records, the time, and the ability to absorb delay. Better tools can narrow an inequality without dissolving the structure that produced it.
The compute auction makes this distinction unusually clear. A useful free service could improve millions of lives. An auction may also be a rational way to allocate scarcity. But when additional capacity is rationed by willingness to pay, access is broad while leverage remains tiered. “Everyone gets AI” is not the same promise as “everyone gets an equal claim on what AI can do.”
Privacy presents a similar test. Axios notes that most interactions with Meta AI can currently be used to improve its models, while the manifesto promises a future mode that Meta itself cannot access. The right response is not automatic disbelief. It is verification. Is privacy built into the architecture? Can outsiders audit it? Can a person take years of agent memory to another provider without losing the digital self the system helped construct? A private mode is a feature. The ability to leave is power.
The essay's distinction between invention and automation is equally valuable—and equally unfinished. Zuckerberg wants AI to help people make things rather than simply replace them. Good. But an economy does not choose invention because a founder prefers the word. Markets often reward employers for immediate automation savings while distributing delayed social costs elsewhere. If Meta wants its agents to augment people, the evidence will be in product incentives, pricing, labor outcomes, and who captures the gain—not in the stated intention.
The early public responses locate the fault line well. In the Associated Press, digital-rights advocate Matt Lane agrees that open access can strengthen defenses, but warns that reliance on Meta-designed systems would chiefly benefit Meta and calls for competition beyond corporate open-model efforts. Anthony Aguirre of the Future of Life Institute, an AI-safety advocate who calls for a pause on advanced development, points to recent failures of control and asks why anyone should trust far more capable systems to remain manageable. Both objections deserve a hearing. Neither enclosure nor proliferation is safety by itself.
The real test of this manifesto will be whether Meta accepts limits it cannot quietly revise: portable memory, independently verified privacy, durable and predictable open-release rules, transparent thresholds for its compute market, a board whose independence, criteria, and veto authority are publicly demonstrable, and an enforceable voice for the communities and workers absorbing the consequences. Those are not accessories to Zuckerberg's vision. They determine whether “everyone” describes a public future or merely a very large customer base.
Dismissing the essay solely because its author owns Meta would substitute biography for argument. Accepting it because Meta can distribute a product at planetary scale would substitute scale for proof.
Zuckerberg has made a serious argument for distributing intelligence. Meta still owes evidence that it will distribute authority in practice: the ability to leave, audit, contest, and govern the system without surrendering one's history or bargaining power.
If the future is for everyone, “everyone” cannot merely mean everyone has an account.
— D.D.