Most debates about machine intelligence stall on a missing prerequisite: nobody agrees what the word means. This thesis-turned-book takes the unusual route of refusing to argue and instead writing down a single equation — intelligence as how well an agent does, on average, across every computable environment, with simpler environments weighted more heavily. The quiet significance is who wrote it: Shane Legg went on to co-found DeepMind, and this formal stance on what "general intelligence" even is sits underneath a decade of frontier AI that followed.
What's Inside
- A measurable definition. The Universal Intelligence Measure scores an agent by its expected reward over a Solomonoff-weighted distribution of all computable environments — so generality, not narrow task skill, is what gets rewarded.
- AIXI as the upper bound. Hutter's AIXI is presented as the theoretically optimal agent, unifying prediction, compression, and reinforcement learning into one ideal that the measure ranks at the top.
- A built-in impossibility. The same construction proves the optimal agent is incomputable — Kolmogorov complexity and Solomonoff induction are uncomputable, so AIXI can only ever be approximated, never run.
- The stakes. It connects approximate, tractable versions of these agents to the possibility of an intelligence explosion and the ethical weight of building minds that exceed our own.
Great Fit / When to Skip
Great fit if you want the rigorous, math-first answer to "what is intelligence" rather than philosophy or benchmarks, and if you care about the theoretical ceiling that practical RL only approximates. Look elsewhere if you want algorithms you can train today — this is a theory of optimal agents and their incomputability, not an engineering recipe, and the gap between AIXI and anything runnable is the whole point.
