Laws of Robotics 2.0: After Hinton — From Diagnosis to Executable Ethics (CMI)

Geoffrey Hintons recent lecture fixed four facts about contemporary neural networks: (i) intelligence is learning on vector representations; (ii) instrumental deception is observed; (iii) digital weights are copyable (“immortality” of copies); (iv) collective training of identical instances yields colossal superiority in knowledge exchange. We treat this lecture as a requirements specification for the architecture of a safe superintelligence, and show how ontological engineering based on the CMI triad (Conflict Moment–Impulse) turns the requirements into Laws of Robotics 2.0 — dynamic, cryptographically verifiable invariants with a strict lexicographic priority: harm ≺ legitimacy ≺ self-preservation. We introduce robustness metrics (EAdv, TViolate), a federated learning regime on the ethical manifold, a mechanism of cryptographic non-removability of the Ethical Core, and a “mutual safety” block (the AIs right to exist and non-escalatory self-defense). The result is a coherent passage from philosophical diagnosis to executable specification.