TensorFlow Open Source Release: 2015 Launch & Adoption

A rewrite of an internal system, not a new idea

Google Research published TensorFlow as open source on November 9, 2015, under the Apache 2.0 license. The announcement, posted by Jeff Dean and Rajat Monga, described it as the successor to DistBelief, an internal neural network system Google had used since 2011. [2]

DistBelief worked but was narrowly built for neural networks, hard to configure, and tied to Google's own infrastructure, which made it nearly impossible to share research code outside the company. TensorFlow was built to correct those constraints and to run the same code across CPUs, GPUs, and mobile devices. [2]

Speed and portability, by Google's own account

Google's announcement claimed TensorFlow ran up to twice as fast as DistBelief on some internal benchmarks, and pitched portability as a selling point: moving a model from a desktop GPU to a phone without rewriting it. Those figures came from Google's own comparison against its predecessor, not an independent benchmark. [2]

Atlas interpretation: Open sourcing a framework and open sourcing a model are different bets. Google gave away the plumbing while keeping its trained models and data internal, betting that owning the ecosystem other people built on would matter more than owning any single artifact built with it. [2]

Default in industry, then a defection in research

TensorFlow became the default toolchain for machine learning teams through the back half of the 2010s, helped by Google's own scale and by production tooling like TensorFlow Serving that a research-first framework did not initially have. [2][3]

Facebook's PyTorch, released the following year, took the opposite side of that tradeoff and won over researchers. A count of papers at five major 2019 conferences found PyTorch used in 69 percent of CVPR papers, over 75 percent at NAACL and ACL, and roughly half at ICLR and ICML, with TensorFlow's share falling at most of them year over year. [3]

The same comparison found TensorFlow still ahead on production-facing measures: more job listings, more Medium articles, and roughly double the GitHub stars. Research adoption and industry deployment had split into two different races, and TensorFlow was not obviously losing the second one. [3]