Pytorch’s Shift from Meta to Linux Foundation: For Good or Bad?

PyTorch's technical governance will lie in the hands of a select few making it an undemocratic proposition

Think about setting AI free only to invite a flood of questions. Surpassing the whys when you finally reach for hows, the only option remains open-sourcing. In the hope of contributing its verve to open-sourcing, Meta has announced passing on its brainchild and labor of love, PyTorch to the Linux community, also known as the Python foundation, that Meta has founded recently. This step will essentially, as per Meta's statement, will "ensure that decisions will be made transparently and openly by a diverse group of board members for many years to come." Developed for deep learning and machine learning implementation, PyTorch has become a well-known framework to power natural language and computer-vision projects, which include the Tesla Autopilot project. Therefore, it comes as a step in the right direction to put PyTorch under community control to avoid potential conflict of interests, that would possibly arise if PyTorch continues to be in Meta's control. Apart, from Meta's basic idea of democratizing PyTorch aims to bring significant improvement to the framework as a whole. Acknowledging the staff shortage, Meta engineer Soumith Chintala, said to Protocol, "hundreds of people" at Meta working on PyTorch, that the team is not big enough to manage the demands of PyTorch's global user community."

What is PyTorch project?

In 2016, to cope with the growing demand for inference models, Facebook required a better-performing AI framework, making it move its AI systems to PyTorch. As an implementation of Torch library, go to framework for Tensor computations, and its tape-based autograd, FB made PyTorch Facebook's default AI framework to optimize user experience, irrespective of the device and operating system they use. From deploying personalized recommendations to improvising computer vision, the PyTorch framework worked like a charm in innovating the fastest clip with greater flexibility, also helping keep the Facebook engineers close to the PyTorch community.  It has become the leading framework with around 1,50,000 projects and 2,400 contributors on GitHub. " The creation of the PyTorch Foundation ensures that decisions will be made transparently and openly by a diverse group of board members for many years to come. The new PyTorch Foundation board will include many of the AI leaders who've helped get the community where it is today, including Meta and our partners at AMD, Amazon, Google, Microsoft, and Nvidia" says Meta in one of its blogs. Built with an open-source, community-first philosophy, it said, it will not shift its policy even after the transition – for researchers and developers to be able to contribute generously so that the advancements in the technology are shared for the larger benefit of the community.

Wide open, yet a closed frame

At the surface level, the move looks directed toward the benefit of the Pytorch community. But, the devil lies in details. As per the statement of Meta's engineer to Protocol, though Meta would be transferring its commercial and marketing efforts onto its Pytorch community, the technical governance, as to how the model should be developed, will lie in the hands of a select few, for eg., Microsoft overseeing PyTorch integration with the neural network ecosystem, ONNX; Nvidia controlling GPU related aspects of PyTorch. Meta's statement that it would continue to invest in the PyTorch project and retain it as its main framework, makes one suspect a hidden agenda.

Being primarily a research platform, it indeed raises questions as to why there is not one from the academic community on the Governance panel. Outcomes in the machine learning domain depend on what technical choices are made while designing the frameworks. Given the commercial interests of big techs, it might be counterproductive to over-optimize for only developing enterprise-specific models, ignoring the use cases like running an algorithm on the desktop or say making a machine learning model for purely non-commercial purposes. Putting the control in fewer hands, experts worry, would corner out small and medium players – an utterly undemocratic trump card.