Zuckerberg's Meta Manifesto on Personal Superintelligence

Key takeaways

1. Date / source. August 10, 2026, meta.com, dedicated domain, ~6,500 words, signed "– Mark".
2. Key framing. the text shifts safety from an engineering problem to a power-distribution problem. This is asserted as a political principle, not a technological one. ### The central inference on alignment Two steps, the second of which affirms the consequent: | Proposition | Status | |---|---| | "Solving alignment is necessary for billions of people to adopt personal superintelligence agents" | acceptable — a product-trust condition | | "if we reach a state where billions of people are using and scrutinizing … agents, then we will have solved alignment with their interests" | adoption stands in as proof | What the passage leaves aside: mass adoption measures perceived usefulness, not the absence of systemic harm; it is compatible with rare, diffuse, or delayed failures. The author's final formulation: "alignment is not an idealized technology capable of keeping a singular superintelligence benevolent; it's what makes personal agents useful and trustworthy." Question the text leaves unanswered: what harm would remain detectable once adoption is achieved. Same structure as the mechanism described in [[dumortier-marketing-ai-os-verification-2026-08-12]]. ### The anti-monoculture argument "Humanity is not a monoculture. People's diverse values represent different trade-offs they would make on important issues. There is no technological solution capable of aligning simultaneously with opposing interests and values. Any singular superintelligence would have to prioritize certain values over others and would, in doing so, be incapable of being benevolent toward everyone." This is the text's most useful move, and it holds independently of who is making it. The next leap, however, is an empirical hypothesis presented as self-evident: from "no singular benevolent superintelligence" one cannot deduce "therefore maximal distribution produces a good balance". The thought experiment assumes that tool parity produces outcome parity. ### Recursive self-improvement: the dilemma posed without a mechanism "Once AI systems can autonomously improve themselves, any lab that doesn't let its system direct a substantial share of compute toward recursive self-improvement will inherently fall behind." The order of magnitude is given by the author: a system optimizing its own efficiency "could theoretically invent ways to extract 100x or more intelligence from every gigawatt". The proposed response is a ratio with no figure attached — allocate enough compute to self-improvement to stay competitive while directing the "significant majority" toward people's individual goals. Three elements are missing: a threshold, a verifier, a sanction — the escape clause being "if harmful behavior emerges, we should coordinate and adjust appropriately". ### Governance: what is delegated, what isn't | Delegated to the independent board | Not delegated | |---|---| | Approval of safety criteria | The release decision itself | | Compliance verification of a release | The roadmap, training, the schedule | | — | Board composition — Meta is founder-controlled, the author says so | Two verifiable tests in twelve months: whether the criteria were published, and whether a release was delayed or modified on the board's advice. ### Open source: "resume" and "some" "Now that Meta Superintelligence Labs are operational, we will soon resume releasing some open source models." The first term acknowledges a halt, the second is an unqualified restriction — no capability tier, no schedule, no parity with internal models. The rest of the section is more substantial: not restricting the existing ecosystem, not banning foreign open source models ("our goal should be for American open source models to be the best in the world"), and an explicit defense of distillation — "all AI models derive from human knowledge… it's important to protect the principle that you can learn from anything you can observe." This is the text's sharpest position, and the most directly self-interested. State of the sector: [[mozilla-state-of-open-source-ai-2026-07]]. ### The employment thesis and its supporting point Framework: no rule requires automation to grow faster than individual capabilities; compute is finite, so it is subject to an opportunity cost between inventing and automating; company sizes can shrink without the number of jobs shrinking; new occupations are not imaginable today (one-person product studios, world builders, personal biologists). All of this hinges on an unsourced sentence: "Recent statistics suggest it may be more likely that individuals' capability growth could match or outpace automation." The author himself names the decisive variable: "If labs focused on automating knowledge work take the lead, then I expect a much harder transition." The thesis is thus less "AI will create jobs" than "the outcome depends on which type of lab dominates". To be weighed against [[ng-the-batch-352-no-ai-jobpocalypse-2026-05-08]]. ### The community compact Announced mechanism: well-paid local jobs, investment in schools and public services, a guarantee against rising energy prices, environmental protection, and a Future Is For Everyone Fund for each host community. Claims made: in Richland Parish (Louisiana), a $50,000 bonus for teachers this year from datacenter tax revenue; America's Workforce Academy, free construction-trades training with guaranteed employment; clean energy production by Meta everywhere it invests; datacenters "among the most water-efficient in the world"; a commitment to be water-positive by 2030 with 200% replenishment in water-stressed areas. All of these figures are self-reported and unaudited. The structural argument is stronger than the figures: "communities where we invest for the long term support development more than those where speculators build with minimal local investment", with a historical parallel (railroads, highways, electrification, broadband) that makes the compact a precondition for the buildout. Named shortfall: building infrastructure is harder in the United States than in China. ### The two public-policy proposals 1. Intermediate training checkpoints provided to the government, with technical staff, so it can harden critical systems "without restricting or delaying individuals' access to personal superintelligence". The proposed trade is explicit: more upstream access in exchange for less downstream control. 2. Regulate matter rather than knowledge: "it will be easier to regulate and control physical components than the spread of knowledge", plus an accelerated FDA. The justification given for the relative calm on biorisk is the weakest part — "it has been possible for decades for bad actors to synthesize harmful compounds, but this has rarely become a significant problem — perhaps because the financial motivation that exists for cyberattacks isn't as present here" — and the author concedes "few historical precedents" and "additional humility". The text cites no capability evaluation. ### Dynamic bidding on compute "A dynamic bidding mechanism that will ensure everyone gets the lowest possible price for the intelligence and compute they use, while ensuring that capacity serves what people collectively find most useful." In other words: the price of the agent's attention floats with demand. A bidding market allocates to ability to pay, something the text does not weigh against its own thesis on balance of power. A precise product commitment, and therefore verifiable. On compute-financing mechanics, see [[nunez-mistral-gigawatt-compute-europeen-venturebeat-2026-08-11]]. ### Citation hygiene
3. None of this text's figures are independent data. The $50,000 bonus, water efficiency, the water-positive milestones, and the compact arrangements are all Meta self-reported claims.
4. Unsourced and unverifiable as presented: the "1 GW+ of Chinese nuclear power every two weeks", the HuggingFace incident "in recent weeks", the "recent statistics" on employment, and the example of a model refusing to write a letter to future parents of students — the latter aimed at an unnamed competitor.
5. Quotable as-is and usable to hold Meta accountable later: the fully private mode, board governance over release criteria, the resumption of "some" open source models, the 2030 water-positive target, dynamic bidding.