On August 10, 2026, Meta published a governing theory for superintelligence and an ownable model on the same morning. That combination is what a move toward aiciety looks like from the inside.
Mark Zuckerberg’s essay, The Future is for Everyone: The Path to a Positive AI Future, appeared alongside the release of Muse Glimmer — a 30-billion-parameter agentic model, open weights under Apache 2.0, small enough to run on a laptop with a single consumer GPU. An open-weight version of Muse Spark 1.2, Meta’s most capable model, was promised to follow.
Aiciety, as defined on this site, has two senses: the totality of interactions between AI and human society, and — in its second stage — a society that has been shaped by AI. The distance between those two is the whole question. What has to happen for a society to stop interacting with a technology and start being formed by it?
The claim of this piece is that August 10 was a step across that gap, for two reasons that have nothing to do with model benchmarks. Zuckerberg’s essay treats superintelligence as a constitutional problem rather than a product problem. And the model released beside it changes AI from something you visit into something you own. Neither move is about capability. Both are about how a technology takes up residence in a society.
The essay deserves reading on its own terms first, so that is where this starts.
The core claim is political, not technical
Zuckerberg opens with two questions — who will have access to superintelligence, and what will it be directed toward — and answers with 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.
The third carries the weight, and arrives via an unusual argument. Most labs treat alignment as engineering: build one system, make it safe, make it beneficial. Zuckerberg says that project is incoherent, for reasons that have nothing to do with technical difficulty.
Humanity is not a monoculture, he writes. People’s differing values represent genuinely different tradeoffs on questions that matter — what a good life is, what is fair, what should be sacrificed for what. No technological solution can satisfy opposing interests at once. Any single superintelligence would have to rank those values, and in ranking them would cease to be benevolent toward everyone on the losing side. His conclusion is stated flatly: there is no singular benevolent superintelligence. Not one that is hard to build — one that cannot exist, because what it would have to do is self-contradictory.
What replaces it is a governance arrangement rather than an artifact. Safety comes from many agents, aligned to many different people, checking and competing with one another the way interests do in a working democracy. He extends this to the labs: the arrangement improves if multiple frontier labs hold models with different values that can check each other. And he names what he considers genuinely dangerous — not capable models being released, but a leading lab training a powerful model and keeping it in-house, however carefully it justifies doing so.
Three thought experiments do the persuading. One person with a superintelligent lawyer wins regardless of the merits, and justice gets worse; everyone with one, and cases are decided more fairly than today. One actor with cybersecurity superintelligence can break into nearly anything; everyone with it, and the long tail of systems finally gets hardened. One company with superintelligence outcompetes all others; all companies with it, and the economy gets more dynamic. The pattern is the argument: a capability held narrowly is a weapon, the same capability held universally is infrastructure.
That is already an aiciety-level claim. Societies do not debate products in the vocabulary of checks and balances, tyranny, and the balance of power. They debate institutions that way. When the people shipping a technology reach for constitutional concepts as their working tools, the technology has left the product category in their own minds before it leaves it in ours.
It also drives his redefinition of alignment: an agent shares its user’s goals and values, not the company’s, within legal and safety boundaries. His illustration of the alternative is pointed — a leading competitor model that refused to help draft a letter to prospective parents at a school because it judged standardized testing unethical. His practical claim is that people won’t hand sensitive tasks to agents aligned with someone else’s values, making user-alignment a precondition for adoption rather than a concession to it.
What he promises to ship
Six commitments make the philosophy concrete. A personal agent that knows your goals and context, works around the clock, and reaches you through any device including glasses. Tools for creation — he describes his eight-year-old coding her ideas and producing videos in an evening. Tools for starting businesses, with a prediction of more employment over time rather than less. A tutor with expertise in every subject and unlimited patience, framed explicitly as extending to everyone what currently depends on parents’ ability to pay. Participation in science, citing Biohub’s open models for virtual cells and proteins.
And access itself: free versions for billions, with paid compute allocated through a dynamic auction meant to clear at the lowest possible price while directing capacity toward what people collectively value most. Plus a fully private mode where even Meta can neither see nor grant access to user data — the WhatsApp encryption analogy is his.
The auction detail is easy to skim and shouldn’t be. Strip the framing and it describes a metered utility with price discovery, the economic form electricity and bandwidth eventually settled into. Products don’t need market-clearing mechanisms. Continuously consumed societal resources do. The shape of the pricing tells you what Meta thinks it is selling.
The employment argument, in his own terms
The jobs section is the most carefully built part of the essay, resting on three moves.
First, a separation of variables: there is no rule that AI must increase automation faster than it increases human capability. These are two things a lab can invest in, and which one leads is a choice, not a law. He is explicit that if the labs focused on automating knowledge work lead, the transition will be much harder.
Second, opportunity cost. Compute is finite no matter how intelligent AI becomes. If superintelligence can invent things of enormous value, that use outbids spending the same compute automating work that already exists.
Third, open-ended demand. People keep finding new problems and inventing roles to address them; a generation ago there were no app developers or data center operators. He forecasts one-person product studios making custom toys and furniture, world builders and experience designers, personal biologists formulating individual treatments. Company sizes shrink, he expects, but company numbers rise. The historical frame is the pre-industrial 90% who farmed to survive, progressively freed into chosen pursuits.
Infrastructure, risk, and the state
Three further sections matter to anyone thinking about societal integration rather than model capability.
The civic bargain. Community Compacts bind data centers to local terms: well-paid jobs, school and public-service investment, self-built energy generation so local prices don’t rise, water-positivity by 2030 with a 200% restoration target in high-stress regions. Richland Parish, Louisiana supplies the evidence — teachers receiving a $50,000 bonus from increased tax revenue. Backing it: a billion-dollar community fund, and America’s Workforce Academy, a free skilled-trades program with guaranteed jobs, which graduated its first class the same week. Communities negotiate compacts with railways, ports, and power plants. They do not negotiate them with products.
Misuse. On cyber, the classic open-source case: more reviewers, faster patches, easier upgrades, with the long-term claim that superintelligence eventually makes most of the world’s code verifiably secure. On bio and chem, an admission of unusual uncertainty — few historical precedents, decades in which synthesis was possible without becoming a major problem — followed by two asks: regulate physical production and distribution rather than the spread of knowledge, and accelerate the FDA and other regulators so cures keep pace with discovery.
The state. The same proposal appears in three separate sections, which signals its centrality: frontier labs should give government intermediate training checkpoints and technical staff during training rather than after, so critical systems get hardened without release schedules slipping. Paired with it, a privacy position strong enough that Meta locks itself out, and a geopolitical stance — keep silicon export controls, don’t slow American releases, remove the friction on American open-source models rather than restricting foreign ones. Distillation gets defended on principle: you may learn from anything you can observe.
Self-improvement. The one place the essay visibly strains. Once systems improve themselves, any lab declining to allocate compute to it falls behind, and a self-improving system could theoretically extract a hundredfold more intelligence per gigawatt and command more effective compute than everyone else combined. His answer is quantitative rather than restrictive: build enough total compute that the majority stays pointed at human goals even while some goes to self-improvement. He concedes there is no clear way to expect benevolence from a system not directed by people.
Finally, governance: Meta’s independent board will approve the safety criteria for model releases and review compliance, with an invitation to other labs to follow.
Why this is a step toward aiciety
Now the framework. The Aiciety stages put Level 7 at daily AI presence across almost all areas of life, new AI-focused careers alongside evolving professions, and societies developing and implementing frameworks to regulate integration. August 10 touched all three.
Presence. Cloud AI has never been ubiquitous in the sense Level 7 requires. It is a place you go — open a tab, authenticate, ask, receive, leave — mediated by a gatekeeper. That is presence by decision. A local agentic model that plans, calls tools, verifies its own output, recovers from failure, and can simply be left running is presence by default. Electricity did not reshape society when it was demonstrated but when it was wired into ordinary buildings; the internet’s transition was the phone in the pocket, not ARPANET. Ownable local models are the same move.
Careers. The forecast of world builders and personal biologists is speculative. A trades academy graduating its first class into guaranteed data-center jobs is an institution manufacturing careers, not predicting them.
Frameworks. The essay is one — checkpoints, board-reviewed release criteria, Community Compacts, positions on export controls and distillation. Framework-making has begun in earnest, from inside the industry.
That is the step. Not a capability threshold, a residency threshold: the point at which AI stops being something society encounters and starts being something society has to organize itself around.
The fork above Level 7
The essay also disputes the ladder it is climbing, which is the more interesting contribution.
Levels 8, 9 and 10 share an assumption. Eight: human labor becomes increasingly redundant. Nine: AI takes over most work. Ten: most people no longer need to work. The upper scale is a displacement curve — progress measured in work handed over.
Zuckerberg argues the ladder branches. His separation of automation from capability growth, his opportunity-cost claim about finite compute, and his open-ended-demand argument all point one way: a society could reach Level 10 saturation without passing through Level 9 redundancy. Distribution is the proposed mechanism. If everyone holds the capability, capability growth stays distributed; if a few labs hold it, the efficient use of it is automating everyone else.
Where the argument is weakest
Open weights are not equal capability. Glimmer is 30B; the frontier stays closed and the open Spark is a promise with a soft date. If that gap widens, distribution is a gesture and the branch never opens.
Access is not agency. The courtroom analogy assumes everyone knows they need a lawyer, knows how to instruct one, and can act on the advice. Distribution solves supply; inequality tends to relocate into the capacity to use what is supplied.
Balance of power needs rough parity. Identical tools don’t equalize the compute, capital, and institutional leverage that decide what a tool gets aimed at.
And on the framework criterion: Level 7 asks that societies develop and implement. What happened was a company proposing frameworks, including one where its own board reviews its own releases. That is the criterion approached, not met — and the gap between an industry drafting governance and a society enacting it is exactly where Level 7 either consolidates or stalls.
Meta also has a commercial interest here, having trailed on raw capability and long used free downloadable models to compete. Sincere and convenient are not exclusive.
What it suggests for the classification
If Zuckerberg is even partly right that pervasiveness and labor absorption can come apart, the scale bundles two variables that may need separating above Level 7 — a Level 8a and 8b, identical in saturation, opposite in redundancy.
Either way, the step itself stands independent of whether his theory is correct. Aiciety in its second sense does not arrive when AI becomes powerful. It arrives when a society has to organize itself around AI — when the questions shift from what the technology can do to who gets it, on what terms, and under whose authority. On August 10 the largest social platform on earth published a political philosophy for superintelligence and put a capable agent on anyone’s laptop, free to keep, in the same few hours.
That is not a threshold anyone announces. It is just the moment the conversation changes category.
Leave a Reply