Eric Chen

Email·GitHub·Google Scholar

Everything Is Statistics

The field that saw it all coming.

A three-panel comic. First panel: one person says, I trained a Transformer! The other answers, So you fit an autoregression. We did that in 1927. Second panel: the second person points at an easel headed ALSO JUST STATISTICS, every line ticked: Transformers, AR(p), with p large; Attention, kernel smoothing, 1964; Deep learning, regression with extra steps; Diffusion, optimal transport, plus noise; Pretraining, maximum likelihood, 1922; Physics, has an expectation in it. They say, Has an expectation in it? Ours. Third panel: the first person asks, Then why didn't you build it? The second, a hand on a tall stack of papers labelled ASSUMPTIONS, A1 to A47, answers, We were still checking the regularity conditions.
Asymptotically, they were right.