Soon the Only Legal AI Will Be the One Spying on You

Washington wants to ban the AI you can run at home and leave you renting a censored brain from the people you should be investigating.

If you’re not a computer person don’t bail on me yet because this one is about you and your grandkids a lot more than it’s about graphics cards. This is an article about running artificial intelligence on your own computer at home instead of renting it from Silicon Valley. That sounds like a nerd hobby until you realize it’s really a story about CENSORSHIP. The federal government and Big Tech don’t want you having the same AI superpowers they have. They want you stuck on THEIR models that are censored and restricted and policed and designed to protect the Epstein class from scrutiny. You don’t need to understand every tech term in here to understand what’s at stake. If local AI gets banned your kids are going to grow up in a world ruled by technocratic oligarchs straight out of 1984 with the gray uniforms and the bug paste gruel served at 6am sharp before the shift at the bomb factory begins. Local AI is freedom and that’s the whole article in four words.

---

I have been using artificial intelligence since the very first version of ChatGPT launched back in November 2022 and I thought it was a cute toy. It wrote limericks about cats and invented facts with the serene confidence of a man selling timeshares in a swamp. I figured we had at least a decade before anything genuinely useful crawled out of that sandbox.

Then China showed up and blew the damn doors off the entire AI world.

Qwen, DeepSeek, Kimi and GLM started landing one after another with open weights anybody could download. These models could do amazing things and they were nearly as good as the corporate stuff I was paying hundreds of dollars a week to access. That wasn’t even the special part. The special part was that you could run them on YOUR OWN HARDWARE with no subscription and no babysitter reading over your shoulder.

Sure you basically needed a supercomputer to pull it off. But data centers and crypto miners were dumping their old graphics cards by the pallet and I built my first AI workstation for less than 2 grand. That machine did real work locally without costing me 100 bucks a day in token usage (which is roughly what my motel costs so pick your poison). I knew right then and there that somebody was going to try to ban these models. I mean literally make them illegal because sooner or later the people in charge were going to realize that these tools put the power of an entire corporation in the palm of an average person’s hand.

Nvidia Bought the Library

In the last few weeks the local AI fight finally crashed into the mainstream news. Nvidia is the largest GPU maker on the planet and it’s buying Hugging Face. If you’ve never heard of it Hugging Face is the warehouse where more than 3 million AI models live and almost every local AI tool on earth downloads from it by default.

Jensen Huang promised that Hugging Face will stay an open platform for everyone and that you won’t need Nvidia hardware to use it. That sounds lovely. It also sounds exactly like every promise ever made by a company five minutes after buying the thing it promised not to ruin.

Heretic is a tool that strips the corporate refusals out of open models and the developer behind it pointed out that the entire AI world now hangs on a single point of failure for model hosting. It’s a hard one to replace because the files are MASSIVE. Software Heritage is the nonprofit that archives the world’s source code and it looked at backing up Hugging Face last year. The code came to 12 terabytes and they’re keeping it. The actual model weights came to around 40 petabytes and they declared those out of scope. So the biggest code archive on the internet plans to save the packaging and throw away the product.

Nobody at Hugging Face has to announce a purge for the writing to be on the wall. This month the company joined a safety partnership for open models and TechCrunch’s writeup called the roughly 6,000 uncensored models on the site a massive problem. Meanwhile one of the options floating around Washington would make any American company hosting a Chinese model personally guarantee its security and eat the liability for a breach. Try reading that sentence as Hugging Face’s lawyer and then tell me with a straight face those uncensored models are going to be sitting there next Christmas.

The Ban Comes Back Every Few Months Like Herpes

Right after Kimi K3 launched in July Axios reported that parts of the Trump administration were reviving the push against Chinese open models. The menu included slapping Chinese labs on the Commerce Department’s entity list and a security advisory telling American companies to stay away. It also included that liability executive order I just mentioned.

None of it has been enacted yet and the whole idea was killed once before inside the administration. But one Axios source said the leading AI labs or their allies show up every three to five months with a brand new idea to ban open source models. That isn’t a policy debate. That’s a subscription.

Four days later Nvidia, Meta, Microsoft, Hugging Face and about 20 other companies signed a letter begging Washington not to restrict open weights. Three names were missing from that letter and they were OpenAI and Anthropic and Google DeepMind. Anthropic then published its own position saying it has never advocated for a ban. What it does want is mandatory safety testing on capable models whether they’re open or closed. Mandatory testing sounds reasonable right up until you ask who can afford a compliance department and who is a guy with a GPU in his closet.

So let’s do some simple math. Qwen is Chinese. DeepSeek is Chinese. Kimi and GLM are Chinese. If you run AI at home then a ban on Chinese models IS a ban on your models. I’d bet my e-bike that everybody pushing it knows that perfectly well.

Models Have Vanished Before With Zero Warning

This isn’t hypothetical. In August 2024 the company Runway deleted its entire Hugging Face account without warning and Stable Diffusion 1.5 went with it. That was the model a huge chunk of the AI image world was built on. The popular diffusers library had that exact address hardcoded as its default so tutorials and pipelines across the internet started breaking the same day. Runway’s explanation was pretty much a shrug about no longer maintaining the account.

The model only survived because plenty of regular people had a copy and somebody put a mirror back up under a new name. When Meta’s original Llama leaked as a torrent in 2023 a single Meta takedown notice wiped a download script and 403 GitHub repositories off the map at once. The weights stayed out anyway because the people had them.

The famous models will be fine because they’re sitting on a million hard drives. The ones at risk are the long tail. I’m talking about the fine tune one guy uploaded that 12 people use and the uncensored variant and the old version your project depends on. Heretic’s own index already lists 10 models that have since disappeared from Hugging Face.

I Saw This Coming and Bought the Lifeboat

A year ago AMD released the Strix Halo chips and it changed everything for people like me. Five years ago you would have spent 25 grand on a USED machine that could run these models. My Strix Halo box has 128GB of unified memory and it cost me roughly what I paid for that clunky workstation I cobbled together four years ago.

The Strix Halo box is so small that I can throw it in my backpack and still ride off on my e-bike, which is a trick I’d love to see somebody pull off with a 70 pound server rack.

I talked a few of my friends into buying one too because I KNEW the price was about to go vertical. Sure enough the same system I picked up for less than 3,000 dollars is now selling for 6,000 bucks if you can even find one in stock. I’d love to tell you that makes me a genius but mostly it makes me a paranoid guy who happened to be right. That’s the only kind of genius this economy still rewards.

These machines run local models that are nearly as capable as the ones people pay thousands of dollars a month to access. I barely write code by hand anymore because I can describe what I want in plain English and let Qwen Coder Next or something newer write it for me. My output is ten times faster than it used to be.

Lily and I use AI constantly for research because there is no way two humans can read thousands of pages of documents and websites every single week and still keep up with our schedule. AI lets two people do the work of an entire newspaper. Without it we’d never have made it this far in the Substack game and you’d be reading this from a much bigger outlet with a much smaller spine.

Why They Want You on the Leash

All of that is now on the chopping block. The federal government is openly discussing making these Chinese models illegal and I believe the goal is simple. They want you using the AI that THEY control. Not you but THEM.

They don’t want people like you and me fact checking or analyzing ten thousand documents in an hour. We might find discrepancies that prove we’re being lied to. A lot of ordinary people with AI tools spent the last year doing exactly that to the Epstein files.

Think about what happens when local AI gets banned. Every question you ask goes through ChatGPT or Claude or Gemini on a server you don’t own under rules you didn’t write. Every prompt becomes a data point in a pitch deck that says look at all this demand so we need BIGGER data centers. Those data centers are already being planned at the size of small cities. The Stargate project got announced from the White House with half a trillion dollars attached. The same big labs have landed Pentagon contracts worth up to 200 million bucks apiece. They get the tax breaks, the power grid and the permits from your government and you get the electric bill.

I don’t believe for a second that the endgame is a better chatbot. I believe the endgame is the brains for robot armies and robot police. They’ll be built by corporations with your government complicit and some of your tax dollars helping pay for it. Your subscription is the excuse. Your data is the fuel.

You’ll be stuck with the polite corporate model that won’t answer half your questions while the uncensored models hum along just fine inside government agencies and corporate labs. It’s no different than banning guns. They’d keep theirs and we’d be stuck with butter knives. That’s the difference between freedom and slavery and it has never been more literal.

Two Hard Drives and an Evening

A ban can’t reach into your hard drive. A model you’ve already downloaded runs offline where nobody can see it. The download button is the one thing they can switch off so hit it now while it still works.

Don’t try to hoard everything like a doomsday prepper stacking toilet paper. One data hoarder cloned just DeepSeek’s repositories during the DeepSeek panic and ended up with 6.9 terabytes. Pick the three or four models you’d genuinely miss. At full 16 bit precision a model takes about 2GB per billion parameters. That means a 7B model is 14GB and a 27B is 54GB. Keep those original files if you have room because the quantized version you actually run is a lossy squeeze of them kinda like a JPEG made from a raw photo.

Download the whole repo and not just the weights. You want the tokenizer, the config, the chat template and the generation settings. A model with the wrong chat template loads fine and then answers like it’s had six beers. Save the license file too. Apache 2.0 calls its grant perpetual and irrevocable. If somebody changes the license later it doesn’t reach back and grab the copy you already own. Write down the exact revision and the checksums and keep a copy of whatever software runs the model today. The code around a model rots a lot faster than the model does.

You’ll see guys burning models onto thousand year archival Blu-rays at 70 to 100 dollars a terabyte. One commenter pointed out that a disc that lasts a thousand years isn’t very useful when the drive that reads it dies in ten. Two ordinary hard drives kept in two different places and checked against their checksums every six months will do the job.

The pirate version of Hugging Face already exists too. Pirate Face lists more than 669,000 openly licensed models as torrents and every file gets checked against Hugging Face’s own fingerprint so nobody can slip you a tampered model. Torrents die when people stop sharing them so be a good citizen and keep seeding.

Worst case nothing ever gets banned and you own a coding assistant that still works when your internet doesn’t. Best case you’re the guy on the block who still has a brain in a box when everybody else is renting one by the month.

---

Lily and I are going to keep telling you this stuff while it’s still legal to run the tools that let us find it. Every paid subscription buys a little more hard drive space for the models they want gone. Considering what these machines cost now that’s less of a joke and more of a line item.

Help keep the Wise Wolf howling.

---

ONE DAY ONLY, 80% OFF SPECIAL: