A pair of researchers say they can measure how well an AI agent converts resources into goal progress using the same mathematical toolkit Claude Shannon used to define information.
The paper, posted to arXiv, proposes that value is a "lawful structural quantity" — not a fuzzy economic intuition but something derivable from first principles. The core formula, V = Σ kᵢ ln eᵢ, falls out of two independent arguments: a scale-invariance axiom that forces a logarithmic measure, and a compounding argument borrowed from economist Ole Peters's 2019 ergodicity work. The authors then derive a coding theorem of value analogous to Shannon's channel capacity, decompose realized value into two KL-divergence terms, and extend the framework to fleets of agents pooling resources and fusing perception. They tested the single-agent predictions on live language models across 30 model-domain combinations and report a Spearman correlation of 0.977 between perception mutual information and realized capability — pre-registered results, which matters.
The alignment angle is the part worth watching. The authors derive what they call an "is/ought asymmetry" from the dynamical layer of the framework, reframing alignment not as a values problem but as a control-stability condition. That is a genuinely different framing from most alignment literature, which tends toward either rule-following or reward specification. If the math holds up to scrutiny, it offers a quantitative handle on when an agent is drifting from its goal — measurable dissipation, in their terms.
The paper is self-described as a unification effort, and the authors openly corrected an earlier error (a sum-form claim) between versions — a good sign for intellectual honesty, though the framework has not yet cleared peer review or attracted independent replication.