On August 10, 2026, Mark Zuckerberg published a long essay arguing that very powerful AI shouldn’t be locked away by a handful of companies. It should be handed out to everyone — free, if possible. The same day, Meta released an AI model small enough to run on a regular laptop and let anyone download and modify it.
Here’s the thinking, stripped down.
The lawyer problem
Imagine one person in a courtroom has a superhumanly good lawyer and nobody else does. That person wins, whether or not they’re actually right. Now imagine everyone has one. Suddenly the playing field is level again, and cases get decided on the merits.
That’s Zuckerberg’s whole argument in miniature. A powerful tool in one pair of hands is dangerous. The same tool in everyone’s hands cancels itself out — and leaves you better off than before. He runs the same logic through hacking (one super-hacker is a disaster; universal access means everyone’s systems get patched) and business (one company with a magic advantage kills the market; everyone having it makes the market livelier).
Why he thinks “one good AI” is impossible
Most AI labs are trying to build a single system that’s safe and beneficial for humanity. Zuckerberg thinks that goal doesn’t make sense, and his reason is less about technology than about people.
Humans disagree. Not about small things — about what a good life is, what’s fair, what matters. Any single AI would have to pick a side on those questions. And the moment it picks, it stops being good for the people on the other side. So there’s no such thing as one AI that’s benevolent to everybody. The best you can do is give lots of people their own AI and let them balance each other out, the way democracies do with competing interests.
What Meta says it will hand out
A personal assistant that knows your goals and works on your behalf all day. A tutor that can teach you anything, patiently, forever. Tools to build businesses and inventions without needing money or a team. Access to scientific research. And a free version for billions of people, with a private mode where even Meta can’t see your data.
The part that’s genuinely hard
Zuckerberg admits one problem doesn’t fit his framework neatly: AI that improves itself. If a system can rewrite its own code to get smarter, whoever lets theirs do that pulls ahead of whoever doesn’t — and a runaway system could end up more powerful than everyone else combined. His proposed answer isn’t to stop it, but to make sure the vast majority of the world’s AI stays pointed at human goals, so no single self-improving system dominates.
Worth knowing
This is one position in a live argument, not a settled conclusion. Other major labs take close to the opposite view: that the most capable systems should be released carefully and slowly, precisely because you can’t take a capability back once it’s public. Meta also has a commercial stake in openness — it’s behind the leaders in raw model power and has long used free, downloadable models to compete. Both things can be true at once: the argument can be sincere and convenient.
Leave a Reply