At MLSys I watched the field celebrate efficiency: better kernels, inference ASICs, training ASICs. That work is not neutral. It is hardware hardening around open-loop learning, and every layer of specialization makes closed-loop learning harder to even attempt.