The Open Weights Party Might Be Winding Down

I’ve spent a chunk of the last couple of years doing something mildly ridiculous for a man my age: downloading enormous language models onto a home server and seeing if my power bill can survive the experiment. It’s a hobby that sits somewhere between model trains and amateur radio, except the parts are free and the forum is r/LocalLLaMA. Yesterday that forum was in a bit of a flap, and for once it wasn’t about VRAM.

China’s Ministry of Commerce put out a draft proposal to add export controls on 23 AI-related technologies: foundation models, agent frameworks, compute clusters, the works. Public comment is open until mid-November. The detail that had everyone talking is that already-open-sourced technology is exempt, but new releases above a certain scale would need to be registered before they go out the door. For a community that has spent two years treating Qwen, DeepSeek, and GLM releases like Christmas morning, this reads like a warning that Christmas might get rationed.

The reaction in the comments ran the full spectrum, from gallows humour (“Qwen 4.0 will be renamed GTA 6.0”) to genuine anxiety about whether the open model pipeline that’s made local AI viable for hobbyists and small businesses is about to dry up. One comment put it plainly: it was nice while it lasted, when Chinese labs were effectively subsidising the rest of the world’s AI research by giving away weights that American labs charge a fortune for API access to. Another, a bit more cynical, suggested the labs had been using the global developer community for bug reports and feedback the whole time, and now the free lunch is ending from both directions.

Here’s the thing that actually interests me, though, more than the doom-posting: nobody in that thread, including people who clearly read Chinese regulatory English for a living, could agree on what the rules actually mean. Someone posted a screenshot of the proposed chip restrictions, and the criteria included things like “possessing floating-point or integer computing capabilities,” which, as several people pointed out, describes basically every processor built since the 1970s. A 486 SX could do integer math. A toaster with a chip in it can probably do integer math. Either the drafting is sloppy, or it’s deliberately broad so bureaucrats can apply it selectively later. I don’t know which is worse.

This is the bit that doesn’t get said enough in these discussions: export control language is almost always vague on purpose, because the goal isn’t to ban a technology, it’s to create a mechanism for approving or blocking things on a case-by-case basis without having to legislate every time. The US does the same thing with chip export rules to China, and nobody over there finds the language particularly precise either. Regulatory ambiguity is a feature when you’re trying to keep leverage, not a bug. It just happens to be maddening if you’re trying to plan a product roadmap, or in my case, figure out whether it’s worth holding off on a GPU upgrade.

There’s a wider irony sitting underneath all this that I think the Reddit thread only half noticed. American AI labs have spent the last year lobbying governments to treat open-weight models as a safety risk requiring oversight, partly because open models from Chinese labs have been embarrassingly competitive on a fraction of the compute budget. Now China is proposing its own oversight regime, officially about national security, and the practical effect might be the same thing US labs wanted: fewer unrestricted releases, more friction, more control sitting with fewer, larger players. Two governments with opposite stated goals converging on a result that mostly benefits incumbents. I’m not saying that’s intentional. I am saying it’s the kind of thing that happens when national security and corporate interest start speaking the same language.

I don’t have a tidy view on whether any of this is good policy. States regulating dual-use technology isn’t unreasonable on its face, frontier models probably do belong in a category above, say, open-source spreadsheet software. But I’ve also watched, close up, what open weights have done for people who’ll never afford a frontier API budget: researchers in under-resourced universities, small Australian businesses bolting a chatbot onto their website without sending every customer query to San Francisco, hobbyists like me running something genuinely capable on a desktop that cost less than a decent latop. If that tap gets turned down, it won’t be the big labs who feel it first.

For now it’s a draft, open for comment until November, and plenty of people in that thread suspect the big labs will rush releases out before any of it takes effect. I’ll believe the sky is falling when I can’t download a GGUF anymore. Until then I’m just going to keep an eye on it, the way I keep an eye on interest rates: not because I can do anything about it, but because it’s going to affect me either way.