Mark Zuckerberg has stopped pretending it is a friendly disagreement. In a 6,500-word essay published on Meta’s newsroom on August 10, titled “The Future is for Everyone,” the Meta chief executive argued that building AI the tightly controlled way OpenAI and Anthropic have chosen amounts to “abandoning our values” in American technology. He named no one. As the Times of India noted, he did not have to.
His sharpest blow lands on alignment, the very idea Anthropic was built around. Dario Amodei has spent years warning about the difficulty of making a powerful model reliably safe toward people. Zuckerberg calls the framing fundamentally flawed. Humanity is not a monoculture, he argues, so no single system can be aligned with everyone’s conflicting values at once. “I think it is literally impossible to have a single benevolent superintelligence that is simultaneously aligned with everyone,” he told The New York Times.
His alternative is what Meta calls personal superintelligence: agents aligned to a user’s own goals rather than a lab’s house values. To illustrate the difference he points to a leading model that refused to help draft a letter to prospective school parents because it had decided standardised testing was unethical. That, in his telling, is not safety. It is a company’s opinion shipped as a guardrail.
Underneath the philosophy sits a sharper accusation. The most dangerous scenario, Zuckerberg writes, is a handful of labs training the most powerful models and keeping them in-house, however much responsibility language is wrapped around the decision. If those labs lead, he argues, power tilts toward large institutions and away from individuals. He backed the words with a product: Muse Glimmer, an open-weight model developers can download and modify, released the same day, with a more powerful version promised soon.
The other labs did not concede. Amodei published his own post the same Monday, restating Anthropic’s position: keep advanced chips away from authoritarian governments, stop industrial-scale distillation, and require safety testing for every sufficiently capable model, open or closed. OpenAI said it supports open-source software and is working with the White House. Neither company would comment further.
What makes the essay awkward is the timing, and Zuckerberg’s critics were quick to point it out. Meta held back the weights of its more capable Spark model earlier this year on safety grounds, and started charging for access. The openness crusade, in other words, has arrived precisely as Meta trails the very labs it is criticising. It is easier to champion giving the best model away when you are not the one holding it.
Still, the argument is not merely convenient. The question of whether frontier AI should sit behind a corporate login or be downloadable by anyone is among the most consequential in the field, and it does not resolve neatly along good-and-evil lines. Zuckerberg’s wager is that the arc of the industry bends toward openness. His rivals are betting their safety case, and their valuations, that it does not.