Adam's convergence proof was wrong and it still trains deep learning
Kingma and Ba built a nearly hyperparameter-free optimizer from two running averages of the gradient, then proved a convergence bound a simple convex problem would later break.
Move a tumor detector to a new hospital and it loses twenty-three points
Every benchmark accuracy measures one model against one distribution, and the distance from that distribution to where the model runs decides whether the number means anything.
Arthur Samuel's checkers program learned by chasing its own deeper guess
On an IBM 701 with a few thousand words of storage, Samuel's program already separated memorizing positions from generalizing over them, and it never beat the champion the legend later handed it.
A network's gradient is a product of local slopes chained backward
The slope a network rolls down is the product of the local slopes along its links, traced from the loss back to any weight in a single backward pass.
A token id is an address into a table the model learned
The integer a tokenizer hands over indexes one row of a learned matrix, and where that row sits in space is the only place its meaning lives.
Employers decide whether New York City's hiring-bias audit applies to them
A research team found 18 posted audits among 391 employers, because the law lets each employer decide first whether it is covered.