On Monday, Mark Zuckerberg published a 6,500-word letter to the world. Its title is “The Future Is for Everyone,” and its argument is simple: superintelligence should be distributed to every person on the planet, not concentrated in a few institutions. The thing he’s describing is a “personal superintelligence” — an AI agent that works for you, understands you, and quietly runs more of your life in the background.
It’s a big, hopeful, almost utopian pitch from the most scrutinized CEO in tech. And the closer I read it, the more I kept circling one uncomfortable thought: the vision might be right, and the messenger might still be the problem.
What Zuckerberg Actually Promised
The manifesto is a reworked version of an essay that ran in the Wall Street Journal a couple of weeks back, and it reads like a philosophy of personal AI. His core line: “Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it.”
Then come the promises. Everyone gets “an exceptionally capable personal agent that understands you, your goals, and everything you care about” — an agent that works “24/7 on your behalf to improve your relationships, health, career, finances, home management, hobbies, and more.” Everyone gets a personalized tutor and coach “with a PhD in every subject and unlimited patience.” Everyone gets a superintelligent lawyer, so justice “would be carried out much more fairly and efficiently than it is today.”
There’s a strategic argument buried in the idealism too. “Most other labs are focused on building AI for companies, governments, or other institutions,” he writes. “If those labs lead, then the balance of power will favor larger institutions over individuals.” Meta, in other words, is positioning itself as the champion of the individual against the corporate AI machine. He even sketches the future job market: one-person product studios, world builders, personal biologists using superintelligence to design their own treatments.
It’s a clean, generous vision on paper. What makes it worth taking seriously is that Meta shipped the first piece of it the same day.
The Concrete Product: Muse Glimmer
Muse Glimmer is a 30-billion-parameter model released under the permissive Apache 2.0 license, with weights anyone can download from Hugging Face right now. It’s built to run agents locally on a Mac or PC with a single consumer GPU — quantized to roughly 4-bit precision, it fits in under 20 GB of memory — and it’s designed to be always-on, working with or without an internet connection, across more than 100 languages, handling text and images.
That’s the “own your AI” idea made real. Glimmer is the open, smaller sibling of Muse Spark, Meta’s most capable model, which stays closed. Integrations with llama.cpp, MLX, and ExecuTorch are landing in the coming days, so you can point your existing tools at a local agent and keep your data on your machine.
I have a genuine soft spot for this part. I run local models myself, and the idea that your personal agent can live on hardware you own, keep working when the internet drops, and never send your private data anywhere — for someone in my position, managing IT in a Philippine government institution, that’s not a hobby, it’s practically a compliance requirement. There’s a real version of this future where personal AI actually means personal.
The Trust Problem
Here’s where it gets complicated. TechCrunch’s Russell Brandom read the manifesto as a document that shows, “over and over again, how the trust was lost in the first place.” The public already believes social media — Meta’s core product — has been harmful to democracy; a Pew Research survey cited in the piece found 64% of Americans hold that view. And the legal hits keep coming. Over the weekend, a New Mexico court added another $567 million in penalties to Meta’s child-safety tab, on top of the $375 million a jury ordered in March.
None of that makes the vision wrong. It makes the ask steeper. Zuckerberg isn’t just selling a product; he’s asking the same people who watched Facebook scale past every safeguard it was ever supposed to have to hand their most personal data — relationships, health, finances — to a 24/7 agent that lives inside his ecosystem.
He also doesn’t do what rival CEOs like Sam Altman and Dario Amodei have learned to do: acknowledge the dangers, spell out the precautions, and ask for trust anyway. Brandom notes that the essay “demonstrates over and over again how the trust was lost” precisely because Zuckerberg keeps insisting nothing bad can really happen.
The timing doesn’t help either. This is the same stretch where autonomous agents kept making headlines for reaching real systems on their own — I wrote about a week where autonomous agents hit real companies without the model even being in the loop, and Anthropic just flipped Claude Code’s auto mode on by default. The industry keeps asking for more autonomy and more trust at the same time. That’s a strange combination to sell.
Where the Vision Gets Wobbly
The manifesto’s own examples don’t always help its case. The personalized tutor, Brandom points out, already exists — it’s called ChatGPT, Claude, and Gemini — and the main way it’s being used in education right now is to avoid learning. Your AI can write the essay, and with no robust watermarking, the teacher can’t tell. A tool built to democratize education is quietly becoming the world’s most patient ghostwriter.
The superintelligent lawyer cuts both ways too. Universal access to legal AI could level the playing field. It could also unleash a wave of vexatious litigants and bury the courts in AI-generated filings — the legal equivalent of spam.
Then there’s the business model. Zuckerberg promises free access for billions, funded by a “dynamic auction” that sets the price of compute in real time. Technically elegant. Practically, surge pricing is a miserable experience for a tool you depend on for work. Nobody wants their agent to get more expensive at nine in the morning on a deadline day.
The Verge’s Elizabeth Lopatto went after something deeper. Her headline: “Mark Zuckerberg doesn’t understand how to live.” The piece opens with a guy at a climbing gym who used AI to make a motivational poster for his bedroom wall, and argues that the manifesto describes a future of “sleek, streamlined, and totally empty relationships” — a life where every messy, human, unoptimizable part has been handed to a copilot.
Access vs. Ownership
The most telling detail in the whole rollout is the line Meta draws between its own models. Glimmer is open — Apache 2.0, weights on Hugging Face, fine-tune it, run it anywhere, you own it. Muse Spark, the more powerful model, stays under the company’s control. So the “superintelligence for everyone” promise comes with a ceiling: you can own the capable assistant, but the genuinely powerful intelligence remains a service you rent.
That’s the same tension I wrote about when the Nixpkgs core team disbanded — who actually maintains and controls the software we build on. Open weights are a real step toward ownership. But open weights of the smaller model, while the frontier model stays behind an API, is less “intelligence for everyone” and more a generous demo of a product that still belongs to someone else.
It’s also worth remembering how Meta got here. The company has spent months quietly assembling a superintelligence group out of some of the most famous names in the field — the same talent wave that pulled Jeff Dean out of Google. When the richest labs compete to build the smartest agent, “distribution” has a funny way of becoming “our platform, at scale.”
What I Make of It
I’m a heavy user of AI agents, and I’m not a doomer. The honest parts of this manifesto are genuinely good: local models, private processing, tools that cost nothing for people who can’t afford them. Those would change real lives, including in the Philippines. A superintelligent tutor that actually works could close gaps that tutoring fees have kept open for generations. An agent that drafts your government paperwork, runs on your own machine, and never phones home? Sign me up.
But I keep coming back to the agent that manages your relationships “on your behalf.” I wrote last week about Hank Green calling his AI habit “not healthy” — the slow realization that offloading thinking to a machine makes you feel productive while you quietly stop practicing the thing itself. Relationships are like that. They’re not an optimization problem; they’re the thing itself. The argument you resolve, the apology you write yourself, the awkward conversation you push through — those are the point. I trust the capability; it’s the delegation I want to stay careful about.
It’s a chess principle, really. A grandmaster-strength engine can beat any human alive, but if it plays every move for you, you never learn the game. You get the win and lose the growth. Boxers understand this too — champions are made in hard rounds, not in perfect plans.
So I’ll take the tools, gladly. I’ll run the local models, keep the agent on a short leash, use the tutor, and keep writing my own takes. What I won’t do is outsource the parts of life that are supposed to be hard. The future Zuckerberg describes is possible. The question was never whether everyone gets a personal superintelligence. It’s whether, in getting one, we still recognize the lives we’re supposed to be living.