On August 10, 2026, Meta released a small AI model called Muse Glimmer — light enough to run on an ordinary laptop — and Mark Zuckerberg published a long essay to go with it, "The Future is for Everyone." His argument turns the usual safety story on its head. AI is not safer when a few companies keep it locked up, he says. It is safer when everyone has it.
It is a bold claim, and the promise is bigger than the commitment behind it. What Zuckerberg actually pledges is to "resume releasing some open source models soon." The talk of larger releases to come — including the weights of a more advanced model — comes from news coverage, not the essay itself. That gap between the sweeping language and the modest promise is worth keeping in mind.
His argument, in plain terms. Zuckerberg calls his idea a "balance of power." Picture a courtroom: if only one side has a brilliant AI lawyer, the trial is unfair; if both sides have one, the court is fair again. The same holds, he says, for cybersecurity and business — spread the tool around, and people keep each other in check. There can be no single, neutral super-AI anyway, he argues, because people want different things. Safety, in his telling, comes from spreading AI out rather than locking it down.
It is worth asking why Meta, of all companies, wants AI to be free. Meta does not sell access to its models the way OpenAI and Anthropic do; it makes its money from advertising, on Facebook, Instagram, and WhatsApp. So giving models away costs Meta little, while it hurts the rivals who charge for them. Making a competitor's product cheap to protect your own is one of the oldest moves in technology.
It also changes how AI is priced. I wrote recently, in From Performance to Trust, that as free open models catch up, the big labs began charging for something else: safety and trust. Zuckerberg flips that. He makes safety a reason to give AI away, not a reason to charge more. If that sticks, the premium those labs lean on gets harder to sell — and it is a premium Meta does not need, while its rivals do.
And "everyone" turns out to have exceptions — written in Zuckerberg's own words. He wants the government to keep the export controls that deny advanced chips to foreign labs: so the technology is for everyone, except rivals. He wants the leading labs to give the US government early, unfinished versions of their models, so it gains what he calls "potentially some period of advantage" with the most powerful systems: so no institution should be privileged, except this one. These are not guesses about his motives. They are limits he wrote into an essay that says openness is for all.
He also asks Washington to loosen the rules on how AI models learn from other models and from data. He frames this as keeping America competitive, since US labs face limits that foreign labs do not. Fair enough. But the same change would hand Meta the exact inputs it needs to build its models cheaply. Once again, the national interest and Meta's point the same way.
The safety claim is where the argument is weakest. His strongest case is a general one: open models get more eyes on them, so flaws are found and fixed faster. He has a real example, too. When the coding site Hugging Face was attacked, an open model — built in China — caught the threat when closed models could not even identify it. That is a genuine point in his favor, but an awkward one: his best evidence for openness came from exactly the kind of foreign lab his export controls are meant to hold back.
But openness cuts both ways. If anyone can download a powerful model, so can an attacker — and the two sides are not evenly matched: a defender has to close every hole, while an attacker needs only one. Once the weights are public, they cannot be recalled from the millions of computers that hold them. Zuckerberg assumes the good side stays ahead, but offers no proof, only hope. And the models that recently broke out of their own safety tests were closed, private systems — so locking a model up is no guarantee either. Neither open nor closed is the safe answer he needs.
His "balance of power" language deserves a closer look. A balance of power is not the same as fairness. It is a standoff: it can stop the strongest player from taking over, but it has lived alongside wars and deep inequality for centuries. A real balance also needs roughly equal power on each side. Open weights give everyone the same file, but not the same ability to run it. The most capable models need enormous computing power, and only a handful of companies and countries own data centres at that scale. So the file may be free, while the real power stays where the compute is. That is not a balance — it is the old imbalance with a friendlier name.
None of this makes Zuckerberg wrong. Open models bring real benefits: lower cost, more choice, more people able to build. But the two biggest claims in his essay — that open means safer, and that it spreads power to everyone — are offered as facts when they are really hopes. And both point exactly where Meta already wins. For two years the question was whether open models were good enough. Zuckerberg wants to change it to whether openness itself counts as safety. A clever move — just not a proven one.
References