TensorFlow is open-sourced

On 9 November 2015 Google released TensorFlow, the successor to the DistBelief system it had built for the cat experiment, under the Apache licence. Deep learning had until then been done in academic frameworks, Theano from Montréal, Torch from a group at Facebook and NYU, Caffe from Berkeley, each with its own gaps. Google's decision to give away the framework that ran its own products, with documentation, tutorials and an engineering team behind it, made deep learning something a student could do on a laptop in an afternoon.

It also started a competition. Facebook released PyTorch in January 2017, built on Torch but with a Python-first design in which the computation graph was built as the code ran rather than declared in advance, and researchers found it easier to think in. By 2020 most papers at the main conferences were in PyTorch, and Google's later framework, JAX, followed its style.

The frameworks are on this timeline because they compressed the field's iteration time. A new architecture in 2012 meant weeks of CUDA; by 2018 it meant a few lines. The transformer, diffusion models and the language models of the 2020s were prototyped and scaled in tools that anyone could download.