“The Commerce Department on June 12 barred any foreign national from using Anthropic's latest frontier models without a license. Since Anthropic couldn't verify users' nationalities in real time, it turned the models off for everyone. Every buyer everywhere learned that the most capable artificial intelligence on the market has a shutdown switch.
One week earlier, the White House issued a memo demanding that AI companies allow no national adversaries to "prevent use of, disable or degrade, or materially modify" the systems on which the nation's war fighters depend. Fair enough.
But when Anthropic shut down its Fable 5 and Mythos 5 models, American companies realized that they hadn't been buying a product from AI providers. They had been buying permission to use that product, and this permission could be revoked.
Those companies need stable and reliable AI usage that follows rules a permission-style model can't, one of which is that the AI product must survive its supplier.
Buyers aren't waiting, and "open-weight models" have begun to appear. These are programs that users can download and run under their own control. Hugging Face, an open-source platform with downloadable AI models and data sets, hosts more than 113,000 derivatives of Qwen, an AI from the Chinese company Alibaba. China's Moonshot AI, Kimi K3, wasn't fully open-weight until late this month but is already ahead of all but two closed models on the usage leader boards.
What Alibaba and Moonshot sell isn't a better model but a keepable one. Their releases have censorship and distillation problems, and Washington's instinct is to ban them. But no one can ban an idea that has already produced a hundred thousand derivatives. Besides, a ban would also cost America the category it invented. Meta's Llama was the model that kicked off the modern open-weight boom, only for Qwen to take the lead. Beijing may close this window itself, but for now it's open.
The market is splitting in two. The narrow slice is the frontier, and the American models of this kind are the best in the world. What true frontier AI can do -- the deepest reasoning, the longest context and the most complex work -- nothing in open source can match. The frontier needs to thrive.
But it also is a narrow slice of the AI world. The broader slice is the everyday: the coding, the instruction-following and the bulk of real workloads. Mozilla's first "State of Open Source AI" report found open-weight models near parity in cost, with prices massively down over the past three years. That broad everyday slice is the critical infrastructure for AI, and it needs to work by a different set of rules. A permission-based model is a fine way to buy excellence. But it is wildly inappropriate for the ordinary uses of AI.
Most corporate boards can't answer the basic questions of which models their companies run and who has the power to turn them off. Yet it is financially important. Britain has designated Amazon, Google, Microsoft and Oracle critical third parties for its financial system, since concentrated suppliers pose systemic risk. At least partially in response to this risk, the European Union is planning to spend 200 billion euros. Canada wants to create a locally owned AI backbone, as do dozens of other countries. This is the largest infrastructure build-out since the internet.
The trouble is that none of these are trustworthy if the AI products they use come with someone else's off switch. So demand is emerging for consumer-owned AI products: intelligence you can keep. Anthropic's shutdown made it clear that AI users can't take their supplier's word for it. Companies need to ask themselves what happens if their AI supplier vanishes or is ordered to cut off service. If a company leaves a supplier, do its data and workflow go with it?
That's continuity, verifiability and interoperability. Washington has demanded exactly this for American war fighters. Everyone else needs the civilian version.
We've seen this standard before. It built the internet when anyone could run a node without permission, inspect the protocols and route traffic. Public money funded it, private capital scaled it, and trillions in value got built on free protocols.
The defining question for AI in this era won't be who tops the leader board for usage. It will be whether June 12 can happen to someone else. Anthropic's models came back online, but the doubt didn't go away. We need AI the world gets to keep, not because it's forced to trust its providers but because it never has to. That's the oldest American play there is: building what the world is already lining up to buy.
---
Mr. Krikorian is chief technology officer of Mozilla.” [1]
1. AI Shouldn't Require a Permission Slip. Krikorian, Raffi. Wall Street Journal, Eastern edition; New York, N.Y.. 28 July 2026: A15.
Komentarų nėra:
Rašyti komentarą