Amazon

Contribution to AI

Recommendation at the scale of a catalogue with millions of items was solved in public. The 2003 paper on item-to-item collaborative filtering described moving the expensive work offline — precomputing similarities between products rather than between customers — so suggestions could be served instantly to tens of millions of shoppers, and it became the standard industrial reference for recommenders long before feeds were a cultural complaint. Mechanical Turk, opened in 2005, mattered more than intended: when ImageNet's compilers needed millions of images labelled in 2008, the piecework market that made that feasible already existed, and supervised learning's data-hungry decade rested on it. The deeper contribution was renting computation. EC2 made machines something billed by the hour, and GPU instances from November 2010 meant a group with no hardware budget could train on the same class of silicon as a funded laboratory; a great deal of later research exists because that was true, and the managed training services that followed removed the cluster administration sitting between a model and its use. Voice hardware took speech recognition out of the browser and into rooms, with a microphone array listening at a distance and the recognition happening in a data centre. The one-year moratorium on police use of its face-recognition service in 2020 was also an early instance of a large vendor withdrawing a working product on stated ethical grounds.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).