Could an enormous lookup table count as intelligent if it always gives the right answer? Shane Legg’s universal intelligence measure makes that question difficult to dismiss. Proposed in his 2008 dissertation, it defines intelligence by an agent’s performance across computable environments, with simpler environments weighted more heavily. In principle, it could compare agents on different platforms without asking what they are made of or how they think. section 9.1.1 section 9.3.1
The simplicity preference comes from a longer line of ideas. Solomonoff’s induction keeps multiple explanations of observations while favoring shorter programs; Bayesian updating gives greater trust to programs that predict well. These ideas motivate a measure that rewards success across many possible environments without treating every environment as equally important. section 9.1.2
Why description length matters
Suppose one pattern comes from a 10-bit program and another needs a 100-bit program. Under the illustrated program-sampling scheme, the particular 100-bit program starts with vastly less weight. This explains the simplicity preference behind the measure; it is not an example of scoring an agent. section 9.1.2
AIXI shows what the measure permits in an idealized setting. It uses a universal prior over computable environments and chooses actions to maximize expected future reward. By construction, it achieves the highest possible score. But AIXI is incomputable and assumes unbounded resources: it is a theoretical ceiling, not a machine-building plan. section 9.1.3
The score is not a promise of success everywhere. In ergodic environments, an agent can recover from mistakes and explore; in adversarial or nonergodic environments, early errors may be irreversible or useful signals too hard to obtain. Nor does performance settle whether an agent understands anything. A lookup table could score well despite its implausible method. The measure makes a broad performance claim precise while leaving the mechanism—and the meaning—of intelligence contested. section 9.1.4 section 9.3.1
