TensorFlow

In the early 2010s, Google Brain — a research laboratory founded in 2011 by Jeff Dean, Greg Corrado and Andrew Ng — was developing internally a deep-learning framework called DistBelief. It powered critical services such as voice recognition in Android, automatic translation in Google Translate and image classification in Google Photos. DistBelief, however, suffered from architectural limitations: it was too tightly tied to Google's internal infrastructure, not flexible enough for research, and difficult to share with the scientific community. In 2014, a team led by Jeff Dean and Sanjay Ghemawat — two veterans who had also designed MapReduce, Bigtable and Spanner — began work on a generic, open-source successor. TensorFlow was released publicly on 9 November 2015 under the Apache 2.0 licence. Its name comes from "tensors", multidimensional arrays that flow through a computation graph built explicitly by the developer; the framework handles distribution of execution across CPU, GPU or the proprietary TPU accelerators designed by Google. The main API is exposed in Python, with C++ and Java interfaces. Versions evolved quickly: 1.0 in February 2017, with a stabilised API; 2.0 in September 2019, with adoption of eager execution mode and full integration of Keras, a high-level wrapper created by François Chollet at Google in 2015. TensorFlow Lite extended the library to smartphones and embedded devices. TensorFlow quickly became one of the dominant frameworks for deep learning, alongside PyTorch developed by Meta in 2016. It powers academic research, industrial applications — health, finance, automotive — and major Google products, notably the neural translation models and, more recently, the Gemini models. It symbolises the transition from a confidential discipline, embodied in the 2000s by Geoffrey Hinton, Yann LeCun and Yoshua Bengio, to a mainstream software infrastructure. Its 2015 release marks the effective beginning of the generative AI era, which would be extended by the 2017 transformers, ChatGPT in 2022 and GPT-4 in 2023.