The idea of artificial neural networks goes back to the 1950s but was abandoned twice: there was not enough computing power, and not enough data to train on. Through the 2000s the standard approach in image recognition was to design features by hand — an expert wrote filters that caught edges or corners, and a classifier used them.

In 2012 Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton entered the ImageNet competition with an eight-layer convolutional neural network. Three things coming together proved decisive: the 1.2-million-image ImageNet dataset that Fei-Fei Li's group had built over years of labelling; GPU programming, which made it possible to train the network on two gaming graphics cards; and training techniques such as dropout and ReLU that held overfitting in check. The result beat the runner-up's error rate by more than ten points. The network had learned for itself which features to look at.

The effect was fast and wide. Within three years hand-designed features were abandoned in image recognition; the same approach moved to speech recognition, to translation, and — with the transformer architecture in 2017 — to language, of which the generative AI wave of 2022 is the continuation. The observation that scaling data, computation and model size together keeps working ('scaling laws') came to shape the industry's investment logic as well. The small company founded by Krizhevsky, Sutskever and Hinton was bought by Google within months; Hinton shared the Nobel Prize in Physics in 2024 and, in the same years, became one of the loudest voices warning about the technology's risks.