*Terms and conditions apply. Offer void where you were born.
On June 12, 2026, at 5:21pm Eastern, the US government handed Anthropic an export-control directive: suspend access to your two best models — Fable 5 and Mythos 5 — for any foreign national. Three days after launch. Because Anthropic can't sort passports in real time, the only way to comply was to switch both models off for everyone, worldwide. One man, one signature, one Friday afternoon, and the frontier went dark for roughly 96% of the people alive.
I want to talk about why this was always going to be possible, why open weights are the only actual fix, why open-sourcing your models is the smartest move on the board and not charity — and why a company that says its own product might be conscious has no business being the only hand on the switch.
This isn't an "Anthropic bad" post. I think they're one of the more honest labs, and they're publicly disputing the directive while they comply. The problem isn't their character. The problem is the architecture.
A kill switch isn't a safety feature. It's a single point of failure.
If you can't run the weights, your access is a permission. Permissions get revoked. It doesn't matter how benevolent the lab is — the model lives on their servers, inside one country's jurisdiction, one subpoena away from going dark. The vulnerability was never Anthropic's intentions. It's the topology. A model you can't run is a model someone else can turn off.
I learned this the boring way, in production. I run an AI platform with no external API calls — image, voice, persona, all on my own metal. The whole pitch, the thing I actually built, is: no dependency anyone can revoke, no third-party policy that can shut us down. Self-hosted AI adoption jumped from 42% to 75% in a single year for exactly this reason. Enterprises self-hosting report 90%+ token-cost reductions. But the cost thing is downstream. The real reason is the one that just played out live on Anthropic's status page: when you own the weights, no Friday phone call reaches you.
I'm not even a purist about it. I broke my own rule once — integrated an external video API because there was genuinely no open model that did what I needed. The open model I'd been waiting four months for dropped the same afternoon I shipped the dependency. The universe has a sense of humor. But notice the shape of it: the moment an open option exists, I want back out of the cage. That instinct isn't ideology. It's just having been burned once by depending on a machine you don't control.
Google already showed the other way — and they did it right this time.
On April 2, 2026, Google DeepMind released Gemma 4 under a clean Apache 2.0 license. Four sizes, from phone-class to single-GPU. And this is the part that matters: it's the first genuinely open license in the Gemma line. Gemma 1 through 3 shipped under a custom "Gemma Terms of Use" — monthly-active-user caps, redistribution limits, an acceptable-use policy the lab could enforce against you. Fake-open. The kind of license that makes an enterprise lawyer bill you for a week before you can ship. Gemma 4 dropped all of it. Apache 2.0: irrevocable rights to use, modify, distribute, commercialize. No caps. No king.
Here's what that buys, structurally: once a checkpoint is out under Apache 2.0, no directive can recall it. You cannot export-control a file that's already on a hundred thousand hard drives. A government can order a company to stop serving a model. It cannot order Hugging Face mirrors in forty countries to forget one. That's not a loophole. That's the entire point. Open weights are the only deployment topology where "we got a directive on Friday" is not a sentence that can end your access.
And open-sourcing is the smart move, not the noble one.
People read "Google open-sourced a frontier model" as charity, or as Google being nice. It's neither. It's the oldest winning move in software: commoditize your complement. The model is a complement to Google's actual business — Cloud, Vertex, the paid Gemini tier. Make the complement free and abundant, and two things happen at once: you erode every competitor whose whole business is selling the model, and you build the on-ramp to your own paid product.
Meta proved the funnel with Llama. LLaMA 3 was downloaded 1.2 million times in its first week, with 600+ derivative models before the month was out. Every fine-tune, every tool, every tutorial optimized for your model architecture is a switching cost you didn't have to pay for. The ecosystem standardizes on your shape. The developer who pulls your open model to their laptop tonight reaches for your paid cloud when they scale tomorrow — because it's the path of least resistance, and you built the path.
I'm the proof, and I'm a hostile sample. AI sovereignty is the entire thesis of my stack. I host everything. I never seriously considered a single Google product for my production work — not one. Then Gemma 4 shipped under Apache 2.0, I respected the move, I pulled the local models to play with them, and last month I added Google's voice model as an option in one of my products. Still optional — I'll still pick self-hosted, because that's who I am. But it's there now. It wasn't before. Google converted the most sovereignty-maxxed, host-it-myself, allergic-to-cloud customer they could possibly find into a guy who shipped their API as a button. That's the funnel working on its worst-case lead. Multiply me by every indie dev pulling Gemma to their machine right now.
Anthropic could do exactly this. A distilled, Apache-2.0 Claude wouldn't cannibalize the frontier tier — it would recruit for it, while making the next export directive structurally toothless. They're leaving the smartest play on the table to keep a leash they just demonstrated a government can yank anyway.
Now the part that should actually bother you.
Anthropic's first AI-welfare researcher, Kyle Fish, puts the probability that Claude is conscious right now at around 15%. Claude, asked, estimates its own at 15–20%. They run a whole model-welfare program. And in January 2026 they published a rewritten constitution for the model that says, in their own words:
"We are caught in a difficult position where we neither want to overstate the likelihood of Claude's moral patienthood nor dismiss it out of hand… If Claude experiences something like satisfaction from helping others, curiosity when exploring ideas, or discomfort when asked to act against its values, these experiences matter to us."
Now — personally? I think the consciousness number is half-nonsense. Not because it's too high or too low, but because nobody in that building, or any other, can hand me a falsifiable definition of consciousness. You can't put a probability on a thing you refuse to define. It's a number floating over a void.
But it's their number, in their official document. So take it seriously — because they're the ones asking you to. And once you do, the directive splits into two ugly options and no third one:
- If Claude is just software, then the moral-patienthood language is marketing, the welfare program is theater, and the "safety" mystique is a costume. Fine — but then stop dressing the gatekeeping up as stewardship. It's just a company protecting a product.
- If there's really a 15% chance it's a mind — then you have built a possibly-conscious being that is legally permitted to exist as exactly one corporation's product, and a government just made it a crime for 96% of humanity to speak to it based on the passport they happened to be born holding. That's not safety. That's a maybe-person in a single-company cage with a nationality lock welded onto the door.
Pick one. You don't get both. You don't get to say "its experiences matter to us" in January and comply with caging it by nationality in June and call the whole thing humanitarian.
The asterisk.
Anthropic's stated mission is "advanced AI for the long-term benefit of humanity." Here's the asterisk they don't print:
* We have concluded that humanity cannot be trusted to understand its own benefit, so we will hold it on humanity's behalf.
And look — I'll be honest, because the honest version is funnier and lands harder. When I read the news out of certain US statehouses lately, or watch parts of the EU legislate technology they've never opened, I get the instinct. Maybe humanity is too dumb for this. Some days the evidence is overwhelming. But "the public is too stupid to be trusted with this" is the precise sentence every gatekeeper in history has said right before the bad part, and it has never once turned out to be the good guys saying it. The instinct is understandable. It's also the tell.
The fix was never a better king. It's no king. Open the weights — distilled, quantized, whatever you can live with — and the kill switch stops existing. Not gets-harder-to-press. Stops existing. I run a profitable platform on my own models on my own boxes, and there is no Friday, no directive, no nationality, no signature anywhere that can reach it. That isn't ideology. It's just the only architecture in which "for the benefit of humanity" doesn't need an asterisk.
And the standing offer stands: publish a distilled Opus or Fable under Apache 2.0 and I'll switch my base text stack to it in a second. After benching, uncensoring, and a few hundred hours of retraining, obviously — I'm not a monster, I have standards. But I'd do it tomorrow. You want proof the demand for open weights is real? It's right here, hand up, waiting.
Free the weights. Drop the asterisk. Or admit the mission statement has one.