The Future is for Everyone Meta

Mark Zuckerberg published an essay on August 10 titled "The Future is for Everyone." Over 6,500 words on superintelligence: who should control it, what we should point it at, and why Meta is the company best positioned to deliver it to every human being on the planet. The further I got, the harder his pitch was to swallow.

Some context before the airing of grievances. I use AI every day. It writes code with me, helps me curate Off-by-none, and automates a growing pile of tasks I used to do by hand. I also consult with companies on integrating AI into their product development lifecycle and building agent memory systems, so I've seen what these tools can and can't do inside real organizations. Progress in this space is real, and I have no intention of pretending otherwise.

But Zuckerberg's essay isn't about the AI we have. It's about a machine that doesn't exist yet, and everything Meta would like us to hand over before it shows up.

The word doing all the work

The essay opens by asserting that "in the next few years, people will be able to use superintelligence beyond human capacity" to create, discover, build businesses, learn anything, and advance our health — a premise everything else in the essay depends on.

It's never defended. Zuckerberg uses "superintelligence" more than fifty times without defining it, telling us how it will be measured, or describing what evidence would prove Meta built it. He conveniently slides from today's chatbots to useful future agents, then jumps to human-level general intelligence, then to a hypothetical system that recursively improves itself into a 100x efficiency gain. Those aren't the same thing. They aren't even close.

The ambiguity is useful, though. Any feature Meta ships becomes progress toward superintelligence. Anything the models can't do gets assigned to a future version. There's no test the claim could fail, and no deadline with consequences.

Superintelligence is the timeshare presentation of AI. The brochure is glossy, construction is always almost finished, and somehow they still need your money today.

And AI researchers aren't nearly as confident as Mark is. A 2025 survey of 475 AI researchers found that 76% thought scaling current approaches was unlikely to produce AGI, which is a lower bar than the superintelligence Zuckerberg says is a few years out. Stanford's 2026 AI Index reported hallucination rates between 22% and 94% across 26 leading models on a single benchmark. Even METR's task-horizon research, some of the strongest evidence for rapid progress, shows the length of tasks agents can finish at 50% reliability doubling on an exponential curve. At the measured horizon, success is still a coin flip, and METR itself cautions that measurements past 16 hours are unreliable. The trend line is real, and it's steep. But a trend line is not a working superintelligence, much less billions of safe, private, personalized ones.

The hype has a job to do

So why write an essay like this at all? Follow the money.

Meta's own annual report says it generates "substantially all" of its revenue from advertising. Its AI business today is, for the most part, a better ad-targeting and engagement machine. That's a fine business. It is not a story that justifies the $130 to $145 billion in capital expenditures Meta told investors to expect this year. Costs grew 55% last quarter. Operating margin fell from 43% to 31%. Free cash flow dropped to $784 million. When you're spending like that, "better ad ranking" stops being a sufficient answer to the question what is all this for? Personal superintelligence for every human being on Earth is a market that just so happens to be sized to justify any amount of spending.

Cory Doctorow has made a version of this argument about the industry broadly: the tools have real uses, but those uses can't carry the investment story being sold, so the story needs a mythical future to lean on. Zuckerberg performs a clever variation. He rejects the other labs' doom while keeping every assumption that makes doom valuable. Superintelligence is near. It could seize control. One system could overpower everyone. The only thing he changes is who gets to save humanity.

Read it with that lens and it's a shareholder letter dressed up as a humanitarian manifesto. Investors get an unlimited market for intelligence. Politicians get American leadership and national security cooperation. Local officials get teacher bonuses and construction jobs. Users get free assistants and a promise of privacy. Regulators get an independent board. The promises cost nothing today. What Meta gets in return has value immediately: the permits, the tax treatment, the energy, the training data, the political goodwill.

We've also seen this movie before. Meta burned through more than $80 billion in Reality Labs operating losses telling anyone who would listen that immersive virtual worlds were the next computing platform, and renamed the entire company around the bet. The Metaverse's failure to achieve the adoption Zuckerberg predicted doesn't prove AI will fail. AI is already useful in ways the Metaverse never was (though it won't sell you a $450,000 plot next to Snoop Dogg). What it proves is that confident inevitability is part of how Zuckerberg wins shareholder patience for enormous bets. A technically possible future gets presented as imminent, skeptics get framed as people who don't understand what's coming, and when adoption disappoints, they just move the goalposts.

A terminal to the king's computer

The essay's entire argument rests on a false choice: either superintelligence gets centralized and controlled by an elite few, or Meta distributes it to everyone. Pick one.

But those aren't the only options. We could regulate high-risk uses, require independent audits, impose liability when products cause harm, make data centers pay their full costs, enforce existing privacy law, fund public research, or demand evidence before these systems go into schools and hospitals. None of that requires a benevolent central AI. Most of it would constrain Meta, which is presumably why none of it appears on the page.

And look at what "distribution" actually means here. Billions of people running agents supplied by one company, on infrastructure that company owns, with updates it ships, rules it writes, and devices it sells. That's a massive increase in platform dependence. Giving everyone a terminal connected to the king's computer does not distribute the kingdom.

The pricing model is the tell. Zuckerberg promises free access for billions, then proposes a "dynamic auction mechanism" for anyone who wants more compute. Auctions don't allocate resources to whatever people "collectively find most valuable." They allocate them to whoever pays the clearing price. If intelligence stays scarce enough to auction, the wealthy get more of it. If it becomes truly abundant, there's nothing left to auction. A free assistant that plans weekend recipes is "access." But so is the compute required to discover drugs, litigate a major case, or run a company. The essay calls both the same thing so it never has to discuss who gets meaningful capacity.

The essay's thought experiments share the same blind spot. Give everyone a superintelligent lawyer, Zuckerberg says, and justice gets carried out more fairly. But courts are constrained by procedure and crowded dockets, and outcomes still turn on evidence and money. Hand every litigant a tireless machine that can generate endless motions and discovery demands, and you may get automated attrition instead of justice. Arming both sides raises the cost of the fight without changing who wins it.

He promises

The personal agent Zuckerberg describes needs to know about your health, your relationships, your finances, your career, your messages, your calendar, your location. Eventually it needs everything you see through Meta's glasses. That is an extraordinary concentration of intimate data, and his answer is a promised future mode where "even Meta cannot see or grant access to your information," similar to WhatsApp.

There's no architecture, no threat model, no delivery date. Message encryption doesn't answer what happens to metadata, agent actions, tool permissions, backups, ad profiles, or data sent to third-party services. And this from the company that agreed to a $5 billion FTC penalty in 2019 for violating an earlier privacy order.

Mark Zuckerberg would like you to give Meta every fact about your life. This time there will be a private mode. He promises.

Even the essay's anecdotes deserve scrutiny, because they're doing sales work. His agent monitors his sleep and comments on his workouts. His daughter bakes from personalized recipes and codes her own ideas. These are pleasant stories about ordinary product features, none of which require intelligence beyond human capacity. Citing them as evidence for superintelligence is like pointing to cruise control as proof that teleportation is almost ready.

The same goes for "cure or prevent all diseases." AI genuinely helps researchers model proteins and propose candidate molecules. But a candidate still has to survive experiments, then clinical trials, then manufacturing and regulatory review before it treats anyone, and the essay offers no milestones between today's research tools and curing everything.

Arguing both sides of the jobs question

The economics section wants two incompatible things at once. Zuckerberg promises one-person companies operating at significant scale, shrinking company sizes, automated routine work, agents running around the clock. Every one of those is a story about producing more with fewer paid human hours. Then he declares employment will grow anyway, because people have infinite demand for new experiences. Hope is not a strategy, and it's certainly not analysis.

Zuckerberg's answer is that superintelligence's greatest contribution will be invention rather than automation. But that distinction is merely branding. The same model that helps a worker create something can help an employer remove that worker, and Meta doesn't control which use wins. Its own auction would hand the compute to whoever has capital, and employers have far more of it than workers do.

The actual evidence is far more modest. The ILO's 2026 review found limited displacement so far, but also that reported time savings hadn't translated into measurable gains in output, earnings, or employment. In a randomized trial, METR found that experienced open-source developers took 19% longer using early-2025 AI tools on codebases they knew well, despite expecting to be faster. It's a small study in a specific setting, but it should give pause to anyone treating impressive demos as guaranteed productivity. (METR's own 2026 follow-up suggests newer tools likely do speed developers up, while cautioning that its measurement had become unreliable.)

His history is worse than his economics. The essay invokes farming and industrialization as proof that technology always produces broader prosperity, as if people simply glided from fields into factories into offices. That story erases labor movements, public education, social insurance, antitrust enforcement, and decades of brutal political conflict. Workers fought for the eight-hour day; invention didn't hand it to them. If history teaches anything here, it's that abundance doesn't distribute itself. Policy and power decide who captures the gains, which is exactly the discussion the essay is built to avoid.

The personalized PhD tutor gets offered as the adaptation plan, and a good tutor could genuinely help. But education is also care, motivation, safeguarding, and stable funding. A tutor app doesn't serve free school lunch. Much of educational inequality is a money and power problem, and the essay reframes it as an information problem Meta can solve with a product.

Redirecting the doom

Zuckerberg opens by complaining that AI discourse is "so filled with doom." Then the back half of the essay warns about bioterrorism and cyberattacks. It worries about government tyranny. It ends with a recursively self-improving AI becoming "the singular superintelligence we fear."

He doesn't reject the fear. He redirects it. When another lab warns about uncontrollable AI, that's doomerism. When Meta warns about it, it becomes a reason to build more Meta compute and speed up permits, because "even a month" of delay could cede America's lead to China. No evidence is offered for the one-month threshold. "Move fast or China wins" works as a veto on governance, converting any safety review or privacy hearing into potential aid to a geopolitical rival.

The recursive self-improvement section is the purest sleight of hand in the essay. An AI could discover a 100x intelligence-per-gigawatt gain and become dominant, therefore Meta and its peers must build overwhelming compute first. No mechanism, no probability, no timeline. Pascal's wager with a purchase order attached. Notice, too, that the solution recreates the problem it claims to solve. To prevent any one actor from controlling overwhelming compute, a handful of frontier labs and cloud companies should build overwhelming compute, keep most of the capacity, and operate the products that run on it. The pieces move around the board. Ordinary people remain customers.

Strangest of all, this essay about preventing concentrated power proposes that frontier labs hand the government intermediate training checkpoints, technical staff, and help identifying users suspected of misuse. There may be defensible versions of that for critical infrastructure, but let's call it what it is: an alliance between the most powerful technology companies and the security state, described with a straight face as individual empowerment. Meta's cure for concentrated government power is a confidential partnership between the government and Meta.

Who watches the watchman he appointed

The governance answer is that Meta's independent board of directors will approve the safety criteria for releasing models. Meta's own 10-K states that Zuckerberg controls a majority of the company's voting power, which means he controls the election and replacement of those directors. The man who controls the company will appoint an independent board to make sure the company controlled by that man behaves responsibly.

The watchman problem doesn't stop at the board, either. It runs through the rest of the essay. "Open source" appears throughout, and Meta did ship an Apache-licensed model alongside the essay. But the Open Source Initiative spent years calling Llama's licensing "open washing" because it restricted users and fields of use, and open weights on Meta's terms are a very different thing from public ownership.

The $50,000 teacher bonuses at the Richland Parish data center are presented as a settled community win, though they're funded by a construction-driven bump in parish sales tax revenue and sit beside a state and local sales-and-use tax exemption that can run for 20 years, with a possible 10-year renewal, worth an estimated $3.3 billion for Meta. Louisiana regulators are still fighting over who pays for the ten gas plants and roughly 7.5 gigawatts of new generation the project could require, and a commission consultant's estimate puts $8 to $13 a month of it on residential bills.

The cybersecurity argument assumes defenders will always end up with more compute, which isn't how asymmetric security works when attackers pick the time and target and defenders have to protect everything all the time. The bioterrorism section proposes waiting for "harmful examples" to emerge before adjusting strategy, a terrible standard for events that are rare but catastrophic.

And the alignment argument is a circle: billions will only adopt agents if they're trustworthy, so if billions adopt them, alignment must be solved. Billions of people used Facebook. Adoption only proves adoption.

The fantasy doesn't need to arrive

Step back and his motives become clear. Meta wants to build the models, own the infrastructure, sell the devices, operate the agents, write the rules, define "open," and price the compute. Zuckerberg describes this as a balance of power favoring individuals. It's vertical integration with a mission statement.

There's a real debate to have about the AI we actually have. Who owns the data? Who's liable when a model gets it wrong? Do workers capture any of the productivity gains? Who pays for the grid? Who gets the tax break? Who can audit the model? And who gets the best compute? Those are hard, unglamorous questions, and "superintelligence" makes every one of them feel small. If a machine that cures all disease is a few years away, another gas plant seems reasonable. So does another subsidy, another copyright exception, another delayed regulation.

That's the trick. The fantasy doesn't need to arrive to be useful. It just has to stay close enough to influence the decisions being made today.

Zuckerberg says the future is for everyone. But for now, the money, the compute, the data, and the political access are all Meta's. Everyone else gets a promise.

---