A 3% gain that might have been fixed by deleting one bad data point

yacineMTB · x · 2026-07-20

A training-data joke with a very real sting.

The author mocks the kind of result where someone designs a complicated new neural network, measures success in trainer steps instead of wall-clock time, and gets a 3% gain—when the same improvement might have been achieved by removing one bad sample from the training corpus.

Original post →

More from Fun

Fun channel →