Abstract
While machine learning has advanced through massive parallelization, weidentify a critical blind spot: some problems are fundamentally sequential.These "inherently serial" problems-from mathematical reasoning to physicalsimulations to sequential decision-making-require dependent computational stepsthat cannot be parallelized. Drawing from complexity theory, we formalize thisdistinction and demonstrate that current parallel-centric architectures facefundamental limitations on such tasks. We argue that recognizing the serialnature of computation holds profound implications on machine learning, modeldesign, hardware development. As AI tackles increasingly complex reasoning,deliberately scaling serial computation-not just parallel computation-isessential for continued progress.