Google Brain

Contribution to AI

The 2012 result that made the argument was almost comically literal: nine layers, a billion connections, ten million frames sampled from YouTube, sixteen thousand processor cores, and a unit inside the network that had learned on its own to fire at cats. Nothing had been labelled. What that established was not a vision technique but a claim about method — that pouring computation and unlabelled data into an old architecture bought capabilities that cleverer, smaller models did not, and that an organisation with datacentres could therefore attempt experiments universities could not afford. The rest followed from taking the claim seriously. DistBelief made distributed training ordinary inside Google; the framework that succeeded it was released under an open licence in November 2015 and became, for several years, the default answer to how anyone trained and deployed a neural network. Sequence-to-sequence learning from 2014 matured into the translation system that displaced Google's statistical machine translation in 2016, the moment a flagship consumer product was handed over wholesale to a neural network. The 2017 attention paper then removed recurrence from sequence modelling entirely, and nearly every large language model since has been a variation on it. Custom inference silicon running in datacentres from 2015 started the industry's move to purpose-built accelerators. The 2016 residency programme, meanwhile, showed that researchers could be trained in a year without a doctorate, and rival laboratories copied the format within two.

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