Washington went after whoever had an address

I want to tell you about a letter, a kitchen, and a rule that landed in a funny place. It took me a month to work out what I thought about it, so I’ll go slowly.

The kitchen bit first

An AI model is a recipe. A very long one, written in numbers.

Those numbers are called the weights. When a company publishes them for free, we say the model has “open weights”. Anyone can download the recipe and cook at home.

A closed model works the other way. You never see the recipe. You buy the finished dish from the restaurant window, one plate at a time. ChatGPT is a restaurant. So are the big ones from Google and Anthropic.

Now the important part. Cooking at home needs an oven. Ovens, here, means computer chips. Nvidia makes the best ones, and almost everybody’s.

The restaurants got so big that they started building their own ovens in the back. So they buy fewer from Nvidia every year. But a school, a hospital, a bank, a normal company? They download a free recipe, and then they have to buy an oven to cook it in.

So here is the whole thing in one line. The more free recipes there are in the world, the more ovens Nvidia sells.

Hold onto that. It explains everything else.

What happened

On 24 July, Nvidia published a three-page letter. It asked the US government not to put rules on models you can download.

Twenty-five companies signed it. Microsoft, Meta, IBM, Dell, Palantir, Hugging Face, Mistral, and others. Jensen Huang, who runs Nvidia, posted about it. It was the first thing he had ever posted on X.

Three big names were missing: OpenAI, Anthropic and Google. The three restaurants.

The timing was not an accident. A few days earlier, reports said the government was thinking about banning Chinese AI models. The letter never mentions China once. Which is a bit like writing about flood insurance and never mentioning rain.

Buried in the letter was one odd paragraph. It asked the government to go easy on something called distillation.

Distillation is when you taste someone else’s dish very carefully, then work backwards to their recipe. Some people call that learning. Some people call it copying. Remember this word. It comes back.

Then two surprises

Surprise one. Within about a day, the list doubled to roughly fifty companies. OpenAI signed. So did Google, AMD, Cisco, Cloudflare, GitHub and others. Anthropic, as far as I can tell, still hasn’t.

Nobody changes their mind overnight. They sign when not signing starts to look worse than signing.

This also ruined my favourite observation. I had noticed that none of the first twenty-five companies sold a closed model. That was true for about thirty-six hours, which is roughly how long I stay right about anything.

Surprise two, and this is the big one. On 5 August, in a private meeting, the government told the AI companies its new safety testing would not apply to open models. Not even the Chinese ones.

The testing applies to the closed models instead.

So there was no ban. Something stranger happened. The free recipes were left alone, and the restaurants got the inspectors.

Some of the American companies who had been asking for rules against their Chinese rivals were surprised. I was too.

My guess at why

Here is the part I keep chewing on.

A rule needs someone to hand it to.

A closed model has an owner. It has an office, an address, lawyers, a front door. An inspector can knock on it.

Open weights are just files. They are already downloaded. Already copied. Already sitting on computers in countries nobody asked. Banning them would be like banning a recipe that has already been photocopied ten thousand times and stuck on ten thousand fridges. You can write that rule. You just can’t enforce it.

So the rule went where it could actually reach.

It isn’t really about which models are safer. It’s about which ones have an address.

And that’s why the distillation paragraph suddenly matters. If you can’t control the recipe, you control the tasting instead. There is already talk of punishing companies caught copying models that way, using trade rules and blacklists. Those land on named companies, and named companies have addresses too.

Nvidia is fine either way. Huang said in January that one in four things AI writes today comes from a free model. Nothing in August slows that down.

What might happen next

Four guesses. I write them down with rough odds so future me can be embarrassed properly. My usual habit with AI news is to make coffee and wait, and waiting saved me here, because the version I nearly published in July was wrong within two weeks.

1. This stays the rule. Free recipes stay free, restaurants keep getting inspected, and everyone gets used to it. A bit better than a coin flip. It is what the letter was built to get.

2. The fight moves to copying. Everyone stops arguing about downloading and starts arguing about tasting. Someone gets caught, and that becomes the real story. Maybe one in three.

3. The restaurants push back. They argue they are only being inspected because they are easy to find, which is fair. The rule gets rewritten quietly. About one in four.

4. A limit shows up anyway. No ban. Just a line: free up to here, paperwork above it. About one in five, and the one I’d watch hardest, because it would arrive as a boring form. Nobody writes headlines about forms.

I get policy timing wrong nearly every time I try. Treat those numbers as what I thought in August, not a promise.

Why I care, from the cheap seats

I do normal work. Keywords, content, rankings. None of my clients will ever download a model onto a big computer somewhere in Ohio.

But every AI tool I pay for runs on someone’s computers. Those tools are cheap partly because free models exist and keep the restaurants honest about price. The same thing happened when I wrote about an AI running a whole ransomware attack by itself. The headline was never the interesting bit. What quietly got cheaper or riskier underneath was.

So I watch small things. The “powered by” line in my tools, and whether it changes without anyone saying so. The price of anything that felt suspiciously cheap this year. And any terms-of-service update that suddenly cares where a model learned what it knows.

That last one will be the copying fight arriving on my street, months before anyone calls it that.