The Real Story This Week: Open-source inference just became the actual battleground. While everyone fixates on Claude export bans and policy theater, llama.cpp shipped five consecutive updates—that's execution velocity that matters. The Ford story accidentally reveals the actual hardware problem: engineers aren't being replaced by AI, they're being replaced by *bad implementations of AI*, which then get rehired to fix it. That's not a future scenario—it's happening now. And those papers on policy composition and optimization? They're the unsexy foundation that makes this repetitive llama.cpp iteration cycle productive instead of pointless—watch how dexterous manipulation policies get reused, because that pattern is coming to every AI stack, and whoever nails composability wins the infrastructure layer.