Google

Contribution to AI

Two systems built for reasons that had nothing to do with intelligence ended up setting the field's conditions. The 2003 file system and 2004 MapReduce papers described how to make thousands of unreliable machines behave as one computer; the open reimplementations that followed turned web-scale training corpora from a fantasy into an ordinary engineering task. The search business then supplied the data and the revenue to use it. From 2011 the Brain group spent both: a network trained on unlabelled video frames that learned to detect cats without being told what a cat was, then word2vec in 2013, which made dense vector representations of words the default input to language systems rather than an exotic choice, then sequence-to-sequence learning in 2014. The pay-off arrived in September 2016, when phrase-based statistical translation was pulled out of Google Translate and replaced with a neural system — the first time deep learning carried a flagship product used by hundreds of millions of people. The TPU, running in datacentres from 2015 and disclosed in 2016, established that it was worth designing silicon for one narrow class of computation, and every accelerator company since has argued against that benchmark. Then in 2017 came the Transformer, published openly, discarding recurrence for attention. Every large language model that followed, including those from Google's competitors, is a variation on a paper Google gave away.

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