Emmanuel Candès
How to put valid statistical uncertainty intervals around any model's predictions — including language models — without assuming you know what kind of error distribution to expect.
Candès is one of the founding figures of conformal prediction, a statistical framework that produces calibrated prediction intervals — guaranteed coverage of the true value at a specified confidence level — without making distributional assumptions about the underlying model. The framework predates the LLM era by years and is essentially model-agnostic, which means it transfers to language models without modification: you can wrap any LLM in conformal prediction and obtain rigorous statements about the reliability of its outputs. For ai100, which reports comparative scores across engines, this is the statistical foundation needed to attach actual confidence intervals to those numbers rather than relying on the field's habit of reporting point estimates as if they were facts.