AlexNet beat the last hand-built vision system by ten points
The paper measured every part it leaned on, which is why fourteen years on its survivors and its discards are still easy to name.
A machine translated Russian in 1954. By 1966 the arithmetic had turned against it.
Warren Weaver's cryptographic analogy made fully automatic translation look close, and the two arguments that finally stopped it, one about meaning and one about cost, were each reading real evidence.
Rebuild ImageNet's test set and the top model loses eleven points
A fresh test set built by the original recipe, no adversary needed, exposes how much of any accuracy number belongs to the data it was measured on rather than to the model.
The EU sorts frontier AI models by one compute number it calls an imperfect proxy
The threshold began binding on 2 August 2025, yet the AI Office that enforces it cannot levy its first fine before August 2026.
Every trained model is rolling downhill on a computed slope
Minimizing x² by hand turns “training” concrete: the derivative points downhill, and a single step-size number decides whether the run settles at the bottom or flies apart.
A transformer never reads the letters in strawberry
The model reads integer tokens from a table built by counting letter pairs across a corpus, and that table decides how it handles spelling, digits, and rare scripts.