AI Insight
This theoretical paper develops a thermodynamic framework for understanding "algorithmic catalysis" - reusable computational structures that can speed up entire classes of problems. The authors prove that the maximum speedup achievable is bounded by the algorithmic mutual information between the computational substrate and the problem class, and that creating such catalysts requires a minimum energy cost due to Landauer's principle of erasure. They derive a theorem showing how long such a catalyst must be used before its energy investment pays off, and demonstrate the concept using satisfiability problems.
Why it matters
This work provides fundamental thermodynamic limits on how efficiently we can build AI systems and specialized algorithms that work across multiple related tasks. It offers a theoretical foundation for understanding the energy costs and benefits of transfer learning and algorithmic optimization in modern machine learning systems.
arXiv:2604.20897v2 Announce Type: replace-cross
Abstract: We develop a thermodynamic theory of algorithmic catalysis within the watts per intelligence framework, identifying reusable computational structures that reduce irreversible operations for a task class while satisfying bounded restoration and structural selectivity constraints. We prove that any class specific speed-up is upper-bounded by the algorithmic mutual information between the substrate and the class descriptor, and that encoding this information incurs a minimum thermodynamic cost via Landauer erasure. Combining these results yields a coupling theorem that lower-bounds the deployment horizon required for an algorithmic catalyst to be energetically favourable. The framework is illustrated on an affine SAT class and situates contemporary learned systems within an information thermodynamic constraint on intelligent computation.
Source: Watts-per-Intelligence Part II: Algorithmic Catalysis