Facebook open sources deep learning tools; head of AI research pledges to “start building things in the open”
Context & Ripple Effects
In mid-January 2015, Facebook's AI group moved from closed internal work to releasing its deep learning tools publicly, with the head of AI research pledging that the lab would 'start building things in the open.' The pledge proved durable rather than rhetorical: by year's end the company had also opened up its AI hardware designs as it positioned itself against Google.
The pattern compounded over the following years — Torch modules and Caffe2 from the lab's first five years, then production systems like Horizon, PyRobot, and the DLRM recommendation model all followed the same open-release playbook established with this announcement.
First-order effects
- External researchers and engineers get direct access to the deep learning tooling Facebook uses internally, lowering the cost of building on Facebook's stack instead of assembling alternatives.
Second-order effects
- Google and rival AI labs are pressured to match the openness or cede mindshare among researchers, since top talent gravitates toward labs whose tools they can use and extend publicly.
Third-order effects
- Open-sourcing core AI infrastructure becomes a standing competitive instrument — commoditizing tools while concentrating advantage in data, compute, and talent — the structure later visible when Facebook released everything from hardware designs to benchmarking models like DLRM.
The trend: Major AI labs converted open-source releases into a recruiting and standards-setting strategy, giving away tools to win the researchers who build on them.