“Universal Intelligence: A Definition of Machine Intelligence” is a 2007 paper by Shane Legg and Marcus Hutter, published in Minds and Machines. It addresses a foundational embarrassment of the field: there were dozens of definitions of intelligence and no agreed way to measure it. Rather than pick one, the authors collected many informal definitions offered over the years by psychologists and AI researchers, extracted the features they shared, and tried to express that common core as a single mathematical formula.
Their measure scores an agent by how much reward it can collect across the full space of possible environments, with simpler environments weighted more heavily. The weighting comes from Ray Solomonoff’s theory of universal induction, which formalizes Occam’s razor - the idea that simpler explanations should be preferred. An agent that does well only in one narrow setting scores low; an agent that does well across a broad, simplicity-weighted range of worlds scores high. This ties the definition directly to Hutter’s AIXI model of an optimal general agent.
The resulting measure is not something you can compute in practice - it sums over all computable environments and is uncomputable like the theory it rests on. Its value is conceptual: it gives a precise, if idealized, target for what “general intelligence” means, and it framed generality as performance across environments rather than mastery of any single task. Later, more practical efforts to define and benchmark generality, including Chollet’s skill-acquisition framing, can be read partly as responses to this line of work.