The Universal Law of Machine Intelligence: Machine Decision-Making Ability (MDMA) Across Technological Eras

Debates about artificial intelligence usually oscillate between behavioral measures and hardware-based metrics, but neither provides a physically grounded, substrate-neutral unit of machine intelligence. Benchmark scores are narrow and culture-dependent, whereas FLOP counts and parameter numbers describe particular implementations rather than decision capacity itself. I propose Machine Decision-Making Ability (MDMA) as a candidate for a universal law of machine intelligence: in the minimal physical sense, a machine’s intelligence is its capacity to enact autonomous state transitions per unit time. A modern mechanical momentary on–off button that flips itself off→on and on→off as two autonomous transitions (MDMA = 2). Tracing the concept backward, ancient alarm clocks and Zhang Heng’s seismoscope each realize a single autonomous trigger (MDMA = 1)—the tiniest physical unit of machine intelligence—while the abacus, for all its historical significance, never decides on its own and merely amplifies human cognition. Anchoring intelligence in autonomous state transitions connects AI capability to the thermodynamics of computation (Landauer-style energy costs and Bremermann/Lloyd bounds on ultimate rates) while remaining compatible with agent-level notions such as Legg and Hutter’s universal intelligence. Within this framework I examine three layers: _ (i) hardware MDMA_ across mechanical, electronic, and quantum substrates; _ (ii) Software Layer Abstraction_ as an MDMA multiplier via pipelines, caching, and learning (e.g., LLMs whose modest inference-time MDMA hides vast historical training budgets); and _ (iii) a “physics of decisions” program_ linking MDMA to mass–energy constraints and civilizational scaling.