Zuckerberg Rails at AI Centralization While Meta Sells Closed Models

Mark Zuckerberg told The New York Times that tightly controlled AI development by leading labs would abandon American tech values, stifle innovation, and centralize power in too few hands. The Meta chief executive, speaking in a rare traditional-media interview published July 28, pushed instead for “personalized superintelligence” available to everyone.

He never named Anthropic or OpenAI. Everyone knew the targets.

Doom Talk Meets a Call for Realism

“So much of the discourse from a lot of the other labs that are developing this is overwhelmingly filled with doom,” Zuckerberg said. “There needs to be a voice or several voices that are bringing realism to this debate.”

He rejected the idea of a single benevolent superintelligence aligned with all of humanity. “I think it is literally impossible to have a single benevolent superintelligence that is simultaneously aligned with everyone at once,” he said. People hold different values. A central system would have to pick winners.

The same day he published a full Wall Street Journal opinion essay titled “The AI Future Is for Everyone.” In it he argued the defining question is not whether superintelligence arrives but who gets access. Centralized power, he wrote, has historically stifled human potential. The better path is individual empowerment, invention as the main purpose of the technology, and balance of power as the safety foundation.

He offered a simple thought experiment: one person with a superintelligent lawyer gains an unfair edge even when wrong on the merits. Give everyone the same tool and justice improves. The same logic, he said, applies to science, business, and politics.

Zuckerberg also posted the essay link himself. “I wrote about why we believe the future is for everyone,” he wrote in Zuckerberg’s own post on the essay, which drew millions of views.

Two Camps Now Stand in Plain Sight

The interview escalates a split already visible across Silicon Valley. One side wants tight controls on the most capable models. The other wants wide diffusion of open-weight models that anyone can download, inspect, modify, and run.

Amodei published an Anthropic position on open-weights models on July 27 clarifying that his company has never sought a ban on open-weights models as a category. He called capable open models without dangerous features a public good. His priorities are keeping advanced chips out of authoritarian hands, stopping industrial-scale distillation that lets rivals leapfrog chip limits, and requiring safety testing of all sufficiently capable models whether open or closed.

Nvidia’s Jensen Huang and a long list of companies released an open weights and American AI leadership letter days earlier. It drew an explicit parallel to 1980s open-source software. Open weights, the letter argued, expand economic access, strengthen competition so gains are not concentrated, give customers control over their own stacks, and improve safety through transparency and many independent testers. Signatories included Nvidia, Meta, Microsoft, Google, OpenAI, Amazon, IBM, SpaceX, Hugging Face, the Linux Foundation and dozens more.

• Open weights let startups and institutions match models to tasks without training from scratch or paying frontier prices for every query.
• Competition across models, chips, clouds and apps keeps benefits broadly shared.
• Organizations keep control of data and the value they create instead of locking into one provider.
• Defenders need comparable model access to study and counter attacker capabilities.

Crowd discussion on X quickly noted the obvious alignment of incentives. Open diffusion helps Meta’s platform, advertising and developer ecosystem. Tight control protects the moats of pure-play frontier labs that sell expensive closed APIs. Public trust in any of the parties remains low.

Meta’s Own Path Turned Hybrid

Meta spent years as the loudest champion of open-source AI, releasing Llama models freely. That stance shifted after the company fell behind.

In 2025 Meta invested $14.3 billion for a large stake in Scale AI and brought founder Alexandr Wang in as chief AI officer to lead the new Meta Superintelligence Labs. The company hired aggressively and rebuilt its stack. In April 2026 it launched Muse Spark, described as its most powerful model yet and purpose-built for personal superintelligence inside Meta’s apps and glasses.

This month Meta began selling API access to an updated Muse Spark for the first time. Pricing was set aggressively, roughly 75 percent below comparable offerings from OpenAI and Anthropic according to coverage of the launch. Wang called the rates “very aggressive and attractive.” Every new account starts with free credits. Meta continues to develop open models alongside the paid proprietary line.

The Muse Spark personal superintelligence launch page frames the goal clearly: an assistant that understands each person’s world because it is built on their relationships and context across Instagram, WhatsApp, Facebook and the rest of the stack. Future open-source versions are still hoped for. The near-term revenue and control play is closed.

Stats snapshot

• $14.3 billion Scale AI investment and talent move that created Meta Superintelligence Labs
• First paid API for a Meta frontier-class model (Muse Spark 1.1)
• ~75% lower token pricing than leading closed rivals at launch
• Billions of users already reached by Meta AI across the apps and glasses

The irony is structural. Zuckerberg argues concentration of superintelligence is dangerous. Meta is simultaneously building and selling a powerful closed system while preaching diffusion. Compute demand for these systems keeps rising; the physical constraints show up in how AI data centers still tied to the grid even when operators chase private power.

Researchers Across Labs Want the Option to Slow Down

On the same day the Times interview landed, more than a thousand employees from the frontier labs released the Pacing the Frontier employee statement. The final tally reached 1,319 signatories. They asked the U.S. government to support an international effort to build technical and governance tools that could deliberately pace the frontier of automated AI research.

The worry is recursive: AI that accelerates its own R&D could outrun understanding and control. Competitive pressure stops any single company or country from slowing unilaterally. The letter does not demand an immediate halt. It wants the option to buy time later.

I think a lot of people have this notion that if you build some kind of singular A.I. you can, through some idealized form of alignment, make sure it is benevolent to humanity. I’m personally skeptical of that path.

Zuckerberg said that in the Times interview. Yet Meta’s own chief scientist Shengjia Zhao signed the pacing letter, as did Dario Amodei, OpenAI chief scientists, Google DeepMind leaders and many others from the same companies. Zhao wrote that frontier labs are close to systems that exceed the best humans on almost every intelligence metric and that responsibility must drive the process.

The same week therefore contained a CEO attacking doom rhetoric, a CEO-level essay demanding personal superintelligence for all, a broad industry letter defending open weights, an Anthropic clarification rejecting open-weights bans, and an employee letter from inside those same firms seeking future pacing machinery. The fracture runs through companies, not just between them.

Washington, China and the Next Ninety Days

Chinese open-weight models have narrowed the American lead in recent months. That progress raised the temperature in Washington. Some officials floated restrictions on Chinese open models. The open-weights coalition letter was a direct response. Tech executives including Huang and Altman were heading to the capital the same week to brief the Trump administration.

1. July 24, Open-weights coalition letter released and amplified by Huang and Nadella
2. July 27, Amodei publishes Anthropic’s open-weights position; Nvidia forms Open Secure AI Alliance on safety tools
3. July 28, Zuckerberg WSJ essay and Times interview; Pacing the Frontier letter surfaces with 1,000-plus signatures
4. Ongoing, Executive visits to Washington as the administration weighs export controls, safety testing and open-model rules

The policy fight sits inside a larger competition. Chip export controls, distillation crackdowns, and pre-release testing regimes are all on the table. Parallel regulatory moves abroad continue; the recent EU AI Act transparency code delay shows how governments are still calibrating rules while the technology races ahead.

David Sacks, a prominent tech investor and adviser, amplified Zuckerberg’s core line on X: concentration of power is the biggest risk. When a few labs decide who gets which capabilities, they shape what can be said, known and built. That is control dressed as safety, he argued. Decentralization and competing models create real checks.

What Personal Superintelligence Would Change

Zuckerberg’s positive vision is consistent across the interview, the essay and Meta’s product language. Superintelligence should help people invent, start businesses with less capital, discover drugs, improve health and pursue their own goals. Invention, not just automation, is the prize. If the technology stays concentrated, jobs and agency could shrink. If it spreads, he expects more entrepreneurship and more small businesses.

He pointed to historical patterns: fears that transformative tools would leave people behind have repeatedly given way to broader prosperity when access widened. The bicycle-shop brothers, the bookbinder’s apprentice, the garage kid with a personal computer. The arc of the industry, he said, has bent toward putting power in more hands, not fewer.

Critics on X and elsewhere counter that Meta only discovered the virtues of openness after falling behind on closed frontier performance, and that selling Muse Spark while preaching diffusion looks convenient. Supporters reply that a hybrid strategy is simply realistic: open models for the long tail, proprietary systems where Meta can differentiate and monetize, and a public argument that keeps regulators from locking the entire field into a handful of approved labs.

The practical test will be whether open-weight ecosystems keep producing competitive models, whether safety testing regimes stay targeted rather than becoming de-facto bans, and whether “personal superintelligence” reaches ordinary users as usable tools rather than marketing language. Meta already reaches billions through its apps. The closed Muse Spark layer now sits on top of that distribution. The open layer remains a live commitment.

Zuckerberg closed the Times conversation by noting he does not need everyone to agree. If the goal is broad empowerment, consensus is neither required nor expected. The general arc, he believes, still favors openness.

That claim is now being stress-tested in real time by Chinese model releases, Washington negotiations, internal lab letters, and Meta’s own pricing page for Muse Spark.