Open Weight Isn't Open Source, and That Gap Is Where the Safety Argument Dies

By the time NPR checked in this past May, Hugging Face was hosting more than 6,000 "abliterated" AI models, open-weight systems with every safety guardrail surgically removed, up from roughly 600 in 2024. A free tool called Heretic can now do the stripping in a couple of minutes on an ordinary laptop. Meanwhile, the UK's AI Security Institute has been tracking something else: the capability gap between open and closed models, once measured in years, is now down to four to eight months. Put those two facts together and the case that open weights make AI safer or fairer starts to look like a slogan, not a finding.

It's worth being precise about the terms, because the debate depends on blurring them. Open weight, by the Open Source Initiative's own account, means the trained parameters are downloadable and nothing more; the training code and dataset usually stay locked up. Open source AI, under OSI's 2024 definition, requires those too, enough that someone could actually reproduce or fully audit the system. Almost nothing marketed as "open source" today clears that second bar, Llama, Qwen, DeepSeek, and Kimi K3 included; they're open weight. That's not a pedantic distinction: it's the entire reason abliteration works on these models while independent bias audits of them mostly don't.

The evidence:

→ Hugging Face's abliterated-model count has grown roughly tenfold since 2024, per NPR; the tool driving much of that growth needs about two lines of input to strip a model's refusal behavior.

→ The UK AI Security Institute's Frontier AI Trends Report puts the gap at four to eight months, defining "open source" models as those "whose code, parameters and training data are made freely available," a stricter bar than most of the open-weight releases actually narrowing that gap, Qwen, DeepSeek, Kimi K3, clear.

→ Cornell researcher David Widder, via Cornell Chronicle: "maximally open" AI systems don't themselves democratize access or enable outside scrutiny, because running, auditing, or fine-tuning frontier-scale weights still takes resources concentrated in a handful of firms.

→ Mark Zuckerberg's own position has shifted, from "open-source is necessary for a positive AI future" in 2024 to promising to be "careful about what we choose to open-source" as Meta approaches superintelligence-level capability.

My take:

The "essential for safety and fairness" framing conflates two different claims: openness as a property (inspectable, reusable, no vendor lock-in) and openness as a guarantee (safer outcomes, fairer power distribution). The first is often true of actual open source; the second keeps failing to show up in practice, on the releases that dominate the ecosystem. Even the vocabulary does rhetorical work: calling a weights-only release "open source" borrows the four freedoms, use, study, modify, share, without granting the "study" freedom that would let anyone verify the claim made in its name. Compute and expertise, not source availability, are the real gate on fairness; Widder's point about Meta's own PyTorch strategy is the tell, open infrastructure Meta still profits from and steers. Safety fares no better: that capability lag isn't much of a head start when stripping a model's alignment takes minutes and leaves the result sitting, permanently and untraceably, on someone's hard drive. Open used to mean "the toy version." At AISI's pace, it increasingly means "the frontier version, unsupervised."

The tell:

Neither side is really arguing about safety or fairness so much as who gets to compete with whom, dressed in the language of ethics. Zuckerberg wrote "power isn't concentrated in the hands of a small number of companies" in 2024, then started talking about being careful about what to open-source once Meta's own models neared a capability tier that could threaten Meta. OpenAI argues closed is safer while resisting the outside audits that would let anyone check that claim, a pattern Widder flags directly. Call it a coincidence if you like; I'll grant intent is unproven. But watch what each camp does with its own frontier model, not what it says about everyone else's.

What to watch for:

As the open/closed gap keeps narrowing and abliteration keeps getting easier, this fight moves from philosophy to hosting policy: whether platforms like Hugging Face start restricting which frontier-capable weights they'll distribute at all. That decision, not another blog post about democratizing AI, will show what anyone here actually believes.