The real story this week isn't the robotics PR or Snapchat's content moderation—it's that open inference is quietly eating closed-model lunch. llama.cpp's relentless weekly releases (five in one cycle) plus work on long-context optimization and standardized reward evaluation infrastructure signal that the hardware bottleneck for running capable models locally is basically solved; what matters now is the software stack that makes it practical. ReToken and OSReward aren't flashy, but they're the unglamorous infrastructure that lets you actually *deploy* AI at scale without vendor lock-in—the kind of thing that only gets attention after it's already won. Watch the inference optimization papers and open tooling velocity, not the house-cleaning robot announcements; the former determines who owns the $100B compute margin in three years.