Computable Artificial General Intelligence

AIXI's theoretical foundation has significant limitations: While AIXI is the only AGI formalism with mathematical proof of optimal behavior, it is incomputable (cannot be practically implemented) and relies on subjective assumptions about intelligence, making it problematic as a basis for AGI safety claims.

The "Cartesian dualism" problem in current AGI theory: Existing formalisms like AIXI incorrectly assume cognition occurs in isolation from the environment, rather than recognizing that intelligence emerges through interaction with the world—a distinction the authors argue is crucial for developing more realistic AGI models.

A new measurable alternative to compression-based intelligence: The paper proposes "weakness" as a computable proxy for intelligence that can be mathematically proven to achieve optimal behavior, potentially offering a more practical framework for analyzing AGI behavior than compression-based approaches.

Experimental evidence suggests compression may not be necessary for optimal performance: The authors' experiments show that maximizing "weakness" outperforms description length (compression), suggesting that the long-standing assumption that compression equals intelligence may be fundamentally flawed.

Bridging cognitive science and AI could improve AGI safety research: By integrating enactive cognition models from cognitive science with formal AI theory, the authors propose a framework that addresses both the mathematical rigor and real-world applicability needed for more credible AGI safety analysis.