Two Companies, One Essay
Meta has spent the better part of two decades being two different companies. One locked the social graph down tight enough to draw a five billion dollar privacy fine and a decade of antitrust fights. The other open sourced React, PyTorch, and a meaningful chunk of the infrastructure the modern internet runs on, no strings attached, years before “open source AI” was something anyone fought over. Both of those are real. Both of those are Meta.
This week, both versions of the company showed up in the same essay.
On Monday, Zuckerberg published “The Future is for Everyone,” his philosophy for the superintelligence era. Three ideas hold it up: individual empowerment, invention over automation, and balance of power as the actual foundation of safety. His concern is that concentrating advanced AI inside a few labs or companies is the real risk. His fix includes personal AI agents for everyone and a promise to resume open source releases “soon.” That same day, Meta released Muse Glimmer, a 30 billion parameter model, under a genuinely permissive license.
It’s a good essay. It’s also worth reading next to Meta’s actual licensing history, because the two don’t fully agree with each other.
What the License Actually Says
Meta’s flagship Llama models ship under the Llama Community License, and the Open Source Initiative, the nonprofit that maintains the actual definition of the term, has said flatly and repeatedly that it doesn’t qualify. Not a fringe complaint. That’s the standards body saying, on the record, that the language is being stretched further than it should go.
Three details explain why. First, the 700 million monthly active user clause: any company that crosses that threshold has to request a separate license, one Meta can grant or deny at its sole discretion. It’s a short list of companies large enough to compete with Meta, and Meta alone decides who gets in. Second, geographic restrictions that have limited certain uses for entities in the EU, without much public explanation. Third, and most basic: Meta doesn’t release the training data or training code. You get finished weights, not the ability to see what built them. That’s open weight, a real and useful thing. It just isn’t open source, which is probably why Meta’s own Llama 4 announcement quietly started using the more accurate phrase.
Where the Philosophy and the License Disagree
Here’s the part I can’t get past. The essay argues that concentrating superintelligence in a handful of powerful institutions is the central danger of this moment. Meta’s own open source license contains a clause built specifically to control access for a handful of powerful companies, at Meta’s discretion. If the worry is genuinely about balance of power, a 700 million user threshold is an odd thing to write into your flagship license, because it isn’t a technical safeguard. It’s a competitive one, and it protects exactly one company.
To be fair, this week’s Muse Glimmer release lands the same stretch that Chinese labs like DeepSeek and Qwen have been setting the pace in open weight AI, with Llama reportedly falling off the usage rankings. Some of this is probably competitive pressure as much as principle. That’s a real business reason. It doesn’t erase the license terms, but it’s fair to hold both facts at once.
So is Zuckerberg lying? I don’t think that’s the right question. My honest read is he believes most of what he wrote. But belief isn’t the same as practice, and by the essay’s own standard, that 700 million user clause is the exact kind of gatekeeping the philosophy claims to oppose. That’s a real contradiction, not a matter of interpretation, and the OSI has been naming it for two years.
Watch the License, Not the Language
None of this means the underlying instinct is wrong. AI concentrated in too few hands might genuinely be risky. It just means the instinct and the license agreement are answering different questions. One is a philosophy. The other is a contract, and contracts are where companies tell you what they actually mean.
So here’s where I’ve landed: don’t evaluate Meta’s AI strategy by the essay. Evaluate it by the license. The language describes the future they want you to picture. The license describes the one they’re actually building.
Watch the license, not the language.
