ai-theory

This paper identifies the verification gap as a key constraint in AI reasoning progress, provides theoretical and empirical analysis of verifier soundness trade-offs, and introduces proof-carrying cognition to enhance reward settlement in reinforcement learning.

The paper analyzes a class of associative memories with hidden neurons, deriving phase diagrams and storage capacities using replica methods and linking to softmax attention in transformers.

This paper analyzes the capability separation between world-model policy learning and imitated world-action models, demonstrating that world-model learning differs in its decision rule and information requirements under certain conditions.

An encyclopedia entry explaining the computational theory of mind, which holds that the mind is a computational system. It covers Turing machines, the history of computationalism in cognitive science, and challenges from rival paradigms.

The paper argues that general intelligence necessitates non-reducible constraints across multiple levels of description, presenting theoretical implications for AI and cognitive science.

This opinion paper argues that large language models are a degenerate special case of world models, not a separate paradigm, and proposes a continuous spectrum from next-token prediction to latent-space architectures like JEPA, examining the data and architecture challenges along this path.

This paper proposes a thermodynamic measure of intelligence, defining intelligence as the ability to make rare but valid futures more likely. It introduces a metric called 'rare-valid lift' that quantifies how much more often a system produces unlikely but acceptable outcomes compared to a passive baseline.

This paper proposes a thermodynamic measure of intelligence defined as 'rare-valid lift' and argues that recursive self-simulation is necessary and nearly sufficient for high thermodynamic intelligence, making intelligence measurable on a universal scale.

This is a popular science article of over 25,000 characters, starting from the origin of entropy, reviewing the development of dissipative system theory, and exploring a three-level analysis of whether AI belongs to dissipative systems (hardware level, training level, static model).

This paper introduces a tree-based formal framework for modeling complementarity in multi-agent human-AI interactions, proving that complementarity is attainable in regression but obstructed in classification under natural conditions on local aggregation and loss functions.

This post explores the debate among top AI figures regarding whether LLMs alone can achieve AGI or if additional breakthroughs like world models are required.