Access Is Not Power

Mark Zuckerberg's AI agent monitors his sleep, watches him train, and gives him feedback throughout the day. On weekends it plans recipes for him to bake with his daughter, orders the ingredients, and offers suggestions while they bake.

In "The Future Is for Everyone", Zuckerberg argues that this kind of agent should not remain inside a few institutions. It should go to everyone, with distribution itself acting as a safety mechanism because people with powerful tools can check one another more effectively than a single controlling authority.

Distribution does not decentralize authority. The distributor still decides how much intelligence each person receives, what the agent may do, how its terms change, and whether the relationship continues. Billions of people could use the agent while those decisions remain with one company.

What Meta means by distribution

Meta is promising billions of people a free agent to help with work, health, finances, and home life. More compute would go to those who can pay. For people locked out of frontier systems today, that access could change what they can actually get done.

Meta would control the model, meter the compute, and maintain the system carrying the agent's memory. A capacity shortage could limit the agent when someone needs help with a medical appointment, a bill, or a job search. Meta could also raise the price, restrict what the agent may do, retire the product, or close the account.

The manifesto shows people gaining capability, then treats that gain as proof that power has shifted toward individuals. That is the leap it never defends. Giving everyone an agent could make everyone more capable without giving anyone authority over the company operating those agents. The users may check one another. Meta's answer for who checks Meta is a board empowered to approve release criteria inside a founder-controlled company.

Zuckerberg illustrates the claim with a superintelligent lawyer for each side. The software may be equal while the money and time available to each side are not. The party that can afford more compute can still bury the other one, while Meta controls the system issuing both lawyers.

Who sets the terms

The manifesto is more specific about what happens when there is not enough intelligence to go around:

For those who want to pay to use more compute, there will be a dynamic auction mechanism that will guarantee that everyone gets the lowest price possible for the intelligence and compute they're using while also ensuring the capacity is used for whatever people collectively find most valuable.

Read that promise with the money put back in. An auction gives scarce compute to whoever bids highest, and what someone can bid depends on what they can afford. Zuckerberg takes the winning bids and calls them "whatever people collectively find most valuable." But the auction does not measure what people value. It measures which values arrive with money attached. A person on the free tier has access and no claim on capacity.

A guaranteed free allocation could soften that inequality. It would make the system more generous, but it would not make the allocation belong to the user. Meta would still set the floor, decide when it moves, determine which uses qualify, and change the allocation when capacity gets tight. These are commitments about how Meta will behave, not rights the user holds.

The board can approve the criteria Meta uses when deciding what to release. It cannot guarantee a user's compute, preserve their agent's memory, make that memory portable, reverse an account closure, or stop Meta from changing the terms. In the founder-controlled company Zuckerberg describes, the board leaves every one of those decisions inside Meta. Users get oversight of a release process. Zuckerberg keeps control of the company.

Over several years, the agent would accumulate a working record of how a person lives and what has changed. Meta promises a fully private mode that even the company cannot see into. If the mode works as described, Meta could not read that record. The promise makes no commitment to preserve the memory, keep the product running, or make the record portable.

The strongest case for Meta

Free superintelligence for billions of people would put capabilities in ordinary hands that until recently belonged to well-funded firms and governments. People do not need to own the infrastructure behind email, search, or maps to benefit from those services, and the same could be true here.

That benefit could be enormous. A competent medical, legal, educational, or creative assistant offered free to billions of people would change lives, including for people with no realistic path to frontier expertise now. Dismissing that because the users do not own the infrastructure would confuse structural criticism with indifference to what access can actually do.

Meta has also begun delivering on its open-source promise. On August 10, it released Muse Glimmer, a 30-billion-parameter multimodal agent model whose weights can be downloaded and run locally under an Apache 2.0 license. That matters. A model people can possess and operate without Meta's servers is more than access to a Meta product; it gives users a capability the company cannot later withdraw. Meta also promises a private mode it cannot see into and support for competing agents. Those measures could limit what Meta sees and give users real alternatives.

Muse Glimmer matters because people can download it, run it themselves, and keep using it if Meta changes direction. That is what transferring control looks like. The personal superintelligence in Zuckerberg's manifesto is built on a different arrangement. Meta would still supply its scarce compute, control its terms, and decide whether its accumulated memory can follow the user elsewhere. Glimmer shows what decentralization looks like. The personal-agent proposal stops short of it.

What power would require

Zuckerberg writes that "invention, not automation, will be the greatest contribution of superintelligence." But distribution does not choose what the system is optimized to do. The company selling the compute and writing the product terms does. Automation gives buyers savings they can measure immediately. Invention is slower, less certain, and often creates value for someone other than the buyer of the compute. Buyers will fund the returns they can measure, and providers will design around that demand. An agent built around those incentives could help a person make something new, or it could learn how to hold their attention, trigger their spending, and extract more work.

A user with authority could take the agent's record to another system, inspect how capacity is allocated, and challenge a decision outside Meta. Meta could not change continuity or exit terms on its own. Phone companies once treated a customer's number as part of their service; number portability made it something the customer could carry to a competitor. Personal agents will need an equivalent right for accumulated memory before portability means more than choosing among products Meta permits.

Zuckerberg's vision would put a powerful agent in billions of hands. Meta would still decide whose agent gets more intelligence, what it may do, whether years of memory can leave, and when the relationship ends. The capability would be widely distributed. The authority governing it would remain centralized. A billion users do not become a balance of power merely because each is given an agent.