Personal superintelligence: an AI for you, not your employer
Personal superintelligence is an always-on AI agent that lives with an individual rather than an institution, learning their goals, context and habits in order to assist continuously with work, health, learning, finances, relationships and everyday tasks across both online and offline life. Mark Zuckerberg has made this idea the spine of Meta’s AI strategy, publishing a long manifesto on personal superintelligence while his team shipped a model designed to make that vision practical. This is not a neutral product roadmap; it is a direct challenge to the emerging norm where a handful of companies sell AI through enterprise APIs and institutional deals. Meta is arguing that the most important AI relationship will not be between a corporation and a cloud service, but between a person and their own persistent agent.
Muse Glimmer: infrastructure for AI democratization, not just a demo
Meta’s philosophy would be empty without cheap, distributed AI access, so the company shipped Muse Glimmer, a 30‑billion‑parameter agentic model whose weights are released under a permissive Apache 2.0 licence and small enough to run on a Mac or PC with a single consumer GPU. This open-weight design turns personal superintelligence from marketing slogan into downloadable software, “an agent whose marginal token Meta never pays to serve” because it runs locally rather than from Meta’s servers. In other words, Meta is trying to square the cost problem of a persistent agent for billions by offloading computation to user devices instead of charging per API call. The tracks, as one commentator put it, now run from the data centre to the desk and the phone, with a model capable of planning, calling tools and recovering from failures now something you can run on hardware you already own.
Distributed AI access as a counterweight to institutional power
Zuckerberg’s central claim is blunt: concentrated AI power is dangerous, and there is “no such thing as a singular benevolent superintelligence.” In his view, safety comes from distributed AI access, where competing systems and empowered individuals provide checks and balances instead of deferring to a small number of companies, governments or institutions. That is why Meta keeps stressing AI democratization and open source AI. The company is resuming releases of open models through its superintelligence lab and has already put Muse Glimmer’s weights into the wild as an open-weight model. This is not altruism alone; it is a deliberate counterweight to rivals pursuing closed, institution-first AI. By putting capable models on consumer hardware and promising more, like the upcoming Muse Spark weights, Meta is betting that a network of personal agents will prove both safer and more economically explosive than a few locked-down frontier systems.
From homework to health: what a personal superintelligence might actually do
Meta’s vision only matters if it changes daily life, and here the company is unusually concrete. In Zuckerberg’s own examples, his personal agent flags information, helps him prototype, monitors his sleep, watches his training sessions, gives feedback, plans personalised recipes to bake with his daughter and orders the ingredients. The broader plan is for personal AI agents aimed at work, health, learning, finances, relationships and routine tasks, with some access free or affordable and a private mode where, he says, even Meta cannot see or grant access to a user’s personal information. Meta casts these agents as tutors “with a PhD in every subject,” offering the kind of personalised education currently reserved for those who can pay. In health and science, Biohub’s open models for virtual cells and proteins are already in researchers’ hands, with Zuckerberg arguing that such tools can accelerate drug discovery.
A business model gamble: billions of personal agents vs enterprise APIs
Behind the rhetoric lies a stark business model divergence. Most institutional AI today is sold as an enterprise API where each call is billed, letting providers recover their costs per request. Meta, by contrast, is committing to a free or low-cost agent that runs continuously, which means a persistent agent for a billion people becomes a different inference problem entirely. The company must build standing capacity at huge scale before anyone pays it back, and its narrowed capital-expenditure guidance underlines how far it is willing to go to do that. According to one analysis, “a persistent agent for a billion people is a different inference problem from an enterprise API,” and Meta’s own numbers show tens of billions in recent capital spending to support this shift. Future releases like Muse Spark’s open weights signal that this is a long-term wager, not a one-off experiment. If Meta is right, the centre of gravity in AI will move from institutions to individuals. If it is wrong, personal superintelligence will be remembered as the most expensive side project in tech history.
