AI News

This arXiv paper challenges the assumption that progress in large language models alone will deliver reliable systems for consequential quantitative decisions, including pricing risk, allocating capital, triaging patients, and containing network intrusions. It argues that language models are trained on human descriptions of the world, which encode quantitative records only imperfectly. Information omitted during that description cannot be recovered by a downstream model, regardless of its scale. The paper further identifies reproducibility, lineage from each output to the source records, and calibrated uncertainty as requirements for consequential settings that a language substrate cannot provide by construction. It proposes Large Quantitative Model (LQM) as a distinct model class designed around these requirements.