Zuck's AI manifesto — with Chinese characteristics

Among the practices Mark Zuckerberg, Meta's chief executive, defends in his new AI manifesto is distillation — the art of training a cheaper model on the outputs of a smarter one, a technique the Trump administration has tried to brand as theft and which he insists is simply how learning works. "You can learn from anything you can observe," he writes, and he is right, more completely than the letter admits. The document itself — 6,500 words under the title "The Future is for Everyone" — is a distillation: it takes China's playbook for artificial intelligence, files off the serial numbers, and wraps what is left in the language of American individualism.

That playbook has a name in the think-tank literature: diffusion. Rather than concentrate capital in a single frontier system and guard it, China's labs — DeepSeek, Alibaba's Qwen, and their kin — release their model weights, let capability spread through the ecosystem, and count on adoption to compound into advantage. The U.S. Congress's own China commission, in a report this spring, drew the contrast plainly: open development and rapid deployment on one side of the Pacific, a handful of tightly held labs hoarding compute and control on the other. Zuckerberg's three principles — individual empowerment as the source of prosperity, invention rather than automation as AI's purpose, and safety secured by spreading the technology as widely as possible — are that same case for diffusion over concentration, retold in the first person. He promises free access for billions, a resumption of Meta's open-weight releases, paused since last year, and, of course, a defense of distillation, while waving away the doom that animates his rivals down the road in San Francisco. Strip the flag and what remains is the argument for open models that DeepSeek's founders could have written themselves — a lab that trained a frontier-class model for a reported fraction of what its American rivals spend.

The distillation is shrewd, not merely cheeky, because the source strategy is winning. In February, by one closely watched measure of what developers actually run, the open models coming out of Chinese labs passed America's: weekly token consumption on OpenRouter, a service that routes queries across competing models, tipped toward the Chinese systems, and the gap has widened since. A partner at Andreessen Horowitz, a venture firm, reckons that four in five American startups now build on Chinese base models. China's industrial plan, branded "AI Plus," aims to push the technology into nearly every corner of its economy by the end of the decade. None of this is because those models top the benchmarks. They are cheap, open, and everywhere, and in software, cheap and everywhere has a long history of beating best.

The Beijing consensus

The tell is the policy. Zuckerberg wants the method without the affiliation, and the seam shows wherever the letter turns to Washington. He backs keeping export controls on advanced chips, calling them the right strategic move to continue; he warns that any rule slowing American model releases by even a month could let foreign models race ahead; he frets that China is bringing a gigawatt of nuclear capacity online every other week while America argues over permits. The manifesto thus asks the government to hobble China's labs at the silicon border while Meta imports their commercial playbook wholesale. It is protectionism outside the building and Shenzhen within it.

The obvious rejoinder is that Zuckerberg is describing his own book and calling it a philosophy. Meta's whole competitive logic runs through distribution — it has always been a company that wins by reaching everyone and monetizing the attention afterward — so a worldview in which distribution beats hoarding is also, conveniently, a worldview in which Meta wins. His "open," moreover, is open weights resting on closed data, a closed platform, and an advertising business, which is a narrower kind of open than the letter's language implies. The second rejoinder is that the bet is not yet settled: Epoch AI, a research outfit, reckons Chinese models still trail America's best systems by roughly seven months at the true frontier. Distribution is winning the ground war; the frontier remains American soil.

Both points are correct, and neither sinks the argument, because the argument is mostly right. The uncomfortable part, for Washington, is that America used to be the diffusion power. The personal computer, the internet, the open web, and Android each won not by being the most polished thing in its category but by being the substrate everyone else could build on; the closed, controlled hand was the one history kept losing with. In this cycle the roles inverted. America's flagship labs went proprietary and vertical, wagering that sheer capability would outrun everything else, and China picked up the open, permissionless playbook that America itself wrote. Beneath its individualist gloss, Zuckerberg's letter is a warning that his own country has forgotten how it won the last three platform shifts.

That is why the letter deserves a closer read than the hypocrisy it invites. The billionaire-hypocrite response writes itself — the ad magnate vowing privacy, the walled company preaching openness — and it is not wrong, only small. It settles a question about the messenger and leaves the one that matters alone: whether the distribution-first strategy beats the frontier-first one. On the evidence the market is already handing over, it does. Zuckerberg defends distillation on the principle that you can learn from anything you can observe, and he has spent the past year observing Beijing. This time, he is right to.