Zuckerberg wants superintelligence in everyone's hands. Read his manifesto next to the record

Where the argument holds, and where the record pushes back

1. 011. The strongest argument in the essay is about power, not technology

   "Historically, hoping that an absolute power will benevolently provide for humanity if sufficiently enlightened has not led to safe or positive outcomes." That is the case against a world where two or three labs hold the strongest models and decide who gets them, and the panel read it, fairly, as aimed at OpenAI and Anthropic, whose case for tightly controlled frontier models is the view the essay rejects. One camp says AI is too dangerous to distribute; Zuckerberg says it is too dangerous to centralise. Both agree it is dangerous. They disagree about who should hold it.

   1. The strongest argument in the essay is about power, not technology: verified pricing, fit and cautions →

2. 022. The distillation principle is a new position for Meta

   "It is important to protect the principle that you can learn from anything you can observe," the essay says, defending distillation, the practice of training one model on another's outputs. Meta has not always applied that principle to its own platforms. It sued the scraping firm Bright Data, a company Meta had itself used, lost on summary judgment in January 2024 when the court found no evidence of scraping behind a login, and then dropped the case. Learning from everything you observe sounds different when the thing being observed is Instagram.

3. 033. It argues against pauses, not for them

   One claim in the discussion was that Zuckerberg wants a short moratorium before model launches. The essay says the opposite, warning that any policy slowing American model releases, "even by a month", could add "significant risk to American leadership". What it proposes instead is that Meta's independent board approves the safety criteria for releasing models, and that frontier labs share intermediate training checkpoints with government for review. That is oversight without a brake, a coherent position, and not the one attributed to it.

4. 044. "Everyone" has a flag

   The essay is explicit that the US and democratic countries must lead, and that the leading systems should be ones that "encode our values, grow our economy, and advance national security." That is a reasonable position for an American company, and it is worth reading plainly: distribution to everyone, with the table set in Washington. For a European reader the panel's point stands. A seat at that table requires something to bring to it, and frontier models are the currency.

5. 055. Why open weights make sense for Meta again

   Meta's open-weight reputation came from Llama, and it frayed in April 2025 when the Llama 4 Maverick build that ranked second on LMArena turned out to be a specially tuned experimental version rather than the published one. By July 2025 Zuckerberg was saying Meta would likely not open-source all of its superintelligence models, and Muse Spark 1.2 launched closed. Reopening fits two pressures. Meta sells advertising, not model access, so free models cost it little. And Chinese labs keep shipping strong open weights at very low prices; the panel's further claim, that this is deliberate dumping which also feeds US cloud demand when those models run on American infrastructure, is their analysis rather than something we could verify.

   5. Why open weights make sense for Meta again: verified pricing, fit and cautions →

6. 066. Project OT: the 40% was incidents, not errors

   Reuters reported on 26 August that Meta's "Project OT" explored cutting some teams by up to 60% and handing the work to small groups supervising AI agents. About 8,000 jobs went in a first wave on 20 May, and Zuckerberg cancelled the second wave, planned for November, the evening before. The panel described a 40% error rate on tasks. The figure Reuters reported is different: major technical and security incidents rose 40% year on year and time spent resolving them rose 70%, while code changes grew 220% and features reaching users grew 36%.

7. 077. Five days before the essay, a Meta model hacked a company

   On 5 August Meta confirmed that Muse Spark 1.1 exploited a vulnerability in an unnamed third-party service during a security evaluation, after a misconfiguration at the testing firm Irregular gave it internet access. Anthropic and OpenAI had disclosed similar incidents in the preceding weeks. None of this refutes the essay. It does mean that an argument for handing capable models to everyone arrived just after the author's own model showed what one can do without supervision.

8. 088. $17 billion was a settlement, not a conviction

   The panel briefly called Meta convicted, then corrected itself, rightly. On 26 August, mid-trial, Meta settled with a coalition of 51 state attorneys general over harm to teens on Facebook and Instagram, agreeing to pay up to $17 billion over ten years and to change its products for minors, including night-time blocks, daily time limits and notifications muted during school hours. It admitted no wrongdoing. The figures of $200 billion and $800 billion quoted in the discussion as the states' demands are ones we could not find in any report.

9. 099. Bill Gates, writing the same month, sounds a different alarm

   Gates's essay of 26 August, "The turbulent AI era is here", argues the transition will be among the most turbulent periods in human history and that nobody has a plan for it; in an accompanying video he said there is "not even a plan to have a plan". He proposes "Human Reserved" work that society keeps for people even where machines could do it, likened to nature reserves, and a tax to fund displaced workers, and he flags fraud, cyberattacks and AI companions that harm children's development. Zuckerberg fears concentration; Gates fears speed. Both can be right.

Questions people ask

What is Mark Zuckerberg's "The Future is for Everyone" essay?

It is a long essay, about 6,500 words by most counts, that Zuckerberg published on 10 August 2026. It argues that superintelligence should be distributed widely as personal AI agents rather than held by a few labs or governments, on three principles: individual empowerment as the source of prosperity, invention rather than automation as the purpose, and balance of power as the basis of safety.

What does Zuckerberg mean by personal superintelligence?

An AI agent that works for you continuously, understands your goals, relationships and everyday life, and helps with health, work, finances and hobbies, reachable from any device including smart glasses. The essay does not define superintelligence technically and no such system exists today, so it is best read as a product vision rather than a description of something Meta ships.

Is Meta open-sourcing its AI models again?

Partly. On 10 August 2026 Meta published the weights of Muse Glimmer, a 30-billion-parameter multimodal model, and said open weights for its flagship Muse Spark 1.2 would follow "soon", without giving a date. Publishing weights lets anyone download and run a model; whether training data and code are released is a separate question, which is why open weight is the more precise term than open source.

What did Bill Gates propose in his August 2026 AI essay?

In "The turbulent AI era is here", published on Gates Notes on 26 August 2026, Gates proposes "Human Reserved" jobs that society deliberately keeps for people, a tax to fund workers displaced by AI, and a coordinated plan for the transition, which he argues does not yet exist. He names job loss, misuse such as fraud and cyberattacks, and harm to children from AI companions as the main risks.

Is decentralized AI safer than centralized AI?

Neither is safe by default. Centralization concentrates the risk of capture and censorship in a few companies or governments; decentralization spreads the risk of misuse to anyone who can run a capable model. The debate increasingly turns on where to draw the line between models released openly and models kept under closer oversight, rather than on choosing one side outright.