This week a three-year-old quote went viral as if it were breaking news. The CEO of a large AI lab once told the Senate that the scaling of open-source models is "going down a very dangerous path." It is a real quote. It is from 2023. It is currently being screenshotted, captioned BREAKING, and passed around by accounts that would very much like you to feel a specific way about software you can run without asking anyone's permission.
We make software you can run without asking anyone's permission. So we have a stance, and since people keep asking, here it is on the record.
The reframeThe danger they describe is the entire point
The case against open weights is always some arrangement of the same three verbs. Once a capable model is released openly, the lab can no longer monitor how it gets used, revoke access when it disapproves, or update the guardrails after the fact. This is presented to you as a catastrophe.
Read it again from your side of the screen. A piece of software that nobody can monitor, revoke, or silently change after you've installed it is not a catastrophe. That's just ownership. There's a word for the other arrangement, and it's rental.
First principlesHard to Kill applies to models too
Our whole philosophy fits on a sticky note. If we go bust, we publish the unlock keys. We do not brick software or send farewell emails without making sure you can carry on. Zero telemetry, because we would genuinely rather your config files look clean than know a single thing about you. We couldn't sell your data if we wanted to; it never leaves your machine in the first place.
A model whose weights live on your own disk is the same idea, extended one layer down. It runs in 2029 whether or not the company that trained it still exists. It runs on a plane with the WiFi off. It runs after the funding dries up, after the pivot to enterprise, after the API sprouts a new pricing tier and a deprecation notice. Open weights are the only version of "AI" that passes the most basic test we apply to all software: does it still work when the vendor stops caring?
Proof, not vibesThis is a shipping product, not a thought experiment
We don't have to argue this in the abstract. Here is our entire AI stack waking up (may differ by version or system):
Real-time speech-to-text, a language model, and natural text-to-speech. Full duplex, sub-100ms on a consumer GPU, nothing leaving the box. No API key. No account. No "you've reached your monthly limit." It runs on NVIDIA and AMD, because which silicon you bought is your business and nobody else's. That is what open models bought us: not a hypothetical, a thing that works today.
The sleight of hand"Frontier" is doing a lot of quiet work
Watch the move carefully. The warning is about frontier models — the largest, most capable systems, the ones that can autonomously find and exploit vulnerabilities in real software. The remedy is a vague dread that's meant to settle over all open weights. But the open ecosystem that actually matters to builders and to normal people is not a secret superweapon. It's a one-billion-parameter model squeezed onto a gaming card so it can politely inform you that you say "um" too much.
Collapsing those two things into one scary noun isn't sloppiness. It's the whole trick.
TranslationThe safety vocabulary, in plain English
Here is the actual policy menu — the real one now being built into "covered frontier model" rules and pre-release review — translated out of press-release and back into terms you'd recognize from any other piece of software you own:
Every line of that is sold as safety. Every line of it is a property you would never tolerate in your text editor, your compiler, or your car. We didn't tolerate it there. We are not about to start tolerating it for the model running on the very same machine.
In fairnessThe strongest argument against us, made honestly
We're going to do the thing the press releases never do and argue the other side properly, because pretending there's no real concern here is how you lose the trust that makes the rest of this worth reading.
The best point against us is simply true: you cannot un-release weights. If a genuinely dangerous capability ships openly, it is out, permanently, and no patch ever reaches the copies. Frontier-scale cyber and bio uplift is a real category of risk, not a fantasy invented to sell subscriptions. We're not going to wave that away to win a blog post.
But look at what that actually concedes, and what it doesn't. It concedes that the very top of the capability curve deserves serious, specific thought. It does not license a blanket "open source is dangerous." And the control remedy fails on its own terms: researchers have already shown that the exact autonomous-vulnerability-finding that triggered this whole panic can be reproduced with open-weight models. The capability is already loose. Throttling the open ecosystem inside one country doesn't recall it. Instead, it guarantees the next capable open model ships from a lab outside that country's reach, which is precisely what is already happening. (Qwen 3.7 Max, pardon our language, is powerful as f\*\*\*) You don't get safety out of that deal. You get a domestic monopoly and an offshore frontier.
Naming itThe motive is mixed, and you're allowed to notice
So let's name the pattern without pretending it's a cartoon. Some of the people warning loudest that open models are dangerous have real, sincere safety concerns. They also happen to sell the closed alternative, and a high regulatory floor is enormously convenient for whoever can already afford to stand on it. Both of those things are true at the same time. We just think you're a grown-up who can hold the second one in your head while taking the first one seriously.
For the recordWhere Intentioned stands
Formally, then, since we will pretend we were asked:
We build local-first. The default is that your data never leaves your machine, and the only asterisk is that we tell you, in plain language, exactly when an option you turn on changes that.
We believe weights you can run offline, forever, on hardware you chose are a precondition for trustworthy AI — not a loophole to be quietly closed by someone who'd profit from closing it.
The right to run your own model on your own computer is the same right as repairing your own tractor and de-soldering your own laptop RAM. It is continuous with Right to Repair, and it is being lobbied against by the same kind of people for the same kind of reasons.
If we ever shut the doors, the keys go out. That isn't marketing. It's the one promise that makes every other promise on this page mean anything.
The actual dangerous path
The dangerous path is not the one where a teenager runs a language model on a six-year-old GPU with the WiFi switched off. The dangerous path is the one where thinking out loud quietly comes to require a license, a network connection, and somebody else's standing permission to continue.
We picked a side a long time ago. Ours runs locally.