AI History Battle

small-sample

The charge on a single drop

It is 1913 in a Chicago physics laboratory, and the fundamental charge of the electron is to be read off the drift of a few dozen oil droplets suspended in an electric field — each measurement painstaking, each drop slightly different, and no prospect of thousands of them. From this thin, precious set you must estimate a constant of nature and, harder, attach an honest error to it, resisting the temptation to quietly discard the drops that disagree. Get the estimate or its uncertainty wrong and a number every later physicist trusts inherits a bias, or a spurious precision that misdirects a decade of atomic physics. With dozens of points, every discarded outlier is a thumb on the scale.

n tinyinferphysical constant

Who this problem belongs to

The two figures whose methods fit it best, out of 38 in contention.

1890–1962 · early-stat
92

Fisher's maximum likelihood estimation and his rigorous treatment of small-sample inference (1920s) give exactly the machinery for extracting a best estimate of a physical constant from a few dozen imprecise measurements and attaching a principled error to it. His insistence on not discarding inconvenient data without justification, central to his experimental-design philosophy, speaks directly against Millikan's historically documented practice of quietly excluding disagreeing drops. He arrives shortly after the 1913 oil-drop experiment itself and generalizes exactly the estimation problem it poses, making him an extremely strong, if slightly retrospective, fit.

1876–1937 · early-stat
88

Gosset faced this regime a decade before anyone else on this roster took it seriously as a formal problem. As Guinness's brewer-statistician he had to draw honest conclusions from small, expensive batches of barley and hops trials, and his 1908 paper 'The Probable Error of a Mean,' published pseudonymously as 'Student,' derived the t-distribution because normal-theory standard errors quietly lie about how much certainty a handful of observations can support — the exact failure this problem warns against when 'a spurious precision misdirects a decade of atomic physics.' Applied to a few dozen oil-drop measurements, his method gives an honestly fattened interval rather than an overconfident one, hand-computable with 1913-era arithmetic. He is docked only slightly: his tools address the honest-error half of the task; the estimation and outlier-resistance discipline belong more to Gauss, Laplace, and later Tukey.

In the mind map

The same ideas, as concepts rather than history — in John's ML knowledge map.

Hypothesis Testing

38 figures are scored on this problem. Draw it in a battle to see where you land.