Thermodynamic Measure of Intelligence | AI Research

This paper proposes a new way to measure intelligence as a physical quantity, moving beyond traditional benchmarks like conversational skill or task success. The author defines intelligence as the "lawful amplification of rare but valid futures." In this framework, an intelligent system is one that uses an internal model to increase the probability of specific outcomes that would be highly unlikely under natural, passive conditions, while ensuring those outcomes remain valid within the rules of the system's environment.

The Core Concept: Rare-Valid Lift

The paper introduces "rare-valid lift" as a universal metric for intelligence. A system acts intelligently when it shifts the probability of future events away from a passive baseline toward a "rare-valid" set—outcomes that are improbable by chance but physically or logically admissible. By focusing on these rare events, the framework distinguishes true intelligence from random noise or simple, repetitive behavior. This approach allows for a unified comparison of diverse systems, ranging from simple feedback controllers and large language models to biological entities and idealized information engines like Maxwell’s demons.

Recursive Self-Simulation

A central premise of the research is that intelligence requires a system to model the world and its own place within it. This leads to "recursive self-simulation," where a system represents not just the external environment, but also its own future actions and internal states. The author argues that this architecture is not just a feature of intelligence, but a requirement. For a system to successfully target rare-valid futures, it must be able to simulate itself as a causal object within the world, allowing it to evaluate the consequences of its interventions before they occur.

Necessity and Sufficiency

The paper provides formal mathematical results connecting this internal architecture to the measurable "lift" of rare-valid futures. The author demonstrates that high intelligence is impossible without high-fidelity self-simulation; if a system cannot accurately identify rare-valid futures in its internal model, it cannot effectively target them in reality. Conversely, when a system possesses both high-fidelity simulation and an effective policy for acting on that information, its ability to produce "lift" approaches the theoretical optimum. This establishes recursive self-simulation as a necessary and nearly sufficient condition for high thermodynamic intelligence.

A Universal Scale

By framing intelligence as a thermodynamic process of trajectory reweighting, the research offers a way to quantify capabilities across different substrates. Whether the system is a human generating text, an AI model, or a physical controller, the framework evaluates them based on the same underlying operation: how they transform information into action to make specific, rare futures more likely. This provides a path-facing account of intelligence that treats various forms of problem-solving and adaptation as special cases of a single, measurable physical phenomenon.