Unexpectedly interesting HBS paper!
This study looks at the impact of technology control on external contributions in open collaboration contexts by examining the case of PyTorch, a popular machine learning framework, which shifted its governance from a for-profit corporation (Meta) to a non-profit foundation in 2022. The results show that this shift led to a significant decrease in contributions from Meta but a notable increase from external companies.
The PyTorch project was moved to a foundation in 2022, and that has been a pretty big success (from most any angle you care to look). This paper uses PyTorch as a natural experiment where an already-open-source project changed in governance structure, and what the result was.
The net result they find is a similar level of overall contributions, but increased contributions from hardware companies. They conclude that there was previously a concern on project direction that could hold up certain types of contributions:
openness does not magically create incentives for external participation without costs, but rather shifts incentives between focal [meaning the originators of the project] and external firms. In particular, control rights theory emphasizes that consideration of the optimal allocation of control rights (with respect to overall welfare) depends on the marginal returns to the ex-ante effort of each party
There is a lot of adding structure to intuitively reasonable ideas, e.g. that “Users” have a higher incentive to collaborate or contribute as they capture value by by the API, which is unlikely to change regardless of directional shifts by the project owner. “Complementors” on the other hand benefit when their product is used in conjunction with the framework, and therefore they need more ongoing cooperation, so are more sensitive to the control of the project. In PyTorch’s case hardware manufacturers are complementors, and hence their contributions are expected (and did) increase with the changes in governance.
What’s interesting, as they note in the paper, is that on a technical level the governance of PyTorch hasn’t changed that much. It does seem though a fair conclusion that adding the overall project governance group makes it more likely that such changes could be made, if needed:
Nevertheless, by changing the governance to a model run by a voting board of other organizations and bringing in the LF, Meta’s singular control of the technical direction of the project (and potentially its social status as the creator of the tool) was greatly diluted.
My main concern with the conclusions is the confounder of massive extra interest in generative AI post-Chat GPT. They do address that, an attempt to control by looking at TensorFlow:
usage of AI technologies dramatically increased in December 2022 due to public release of OpenAI’s ChatGPT. And while Chip Manufacturer and Application Developer companies are both affected by this demand shock, it is possible that they are differentially affected by this change in a way that confounds our analysis. To rule out this possibility, we augment our sample by further gathering the external company commits data to TensorFlow, Google’s open source machine learning framework.
TensorFlow feels like not a great baseline due to the different adoption by the research community. I would be mildly interested to see if XLA had any changes in contributions: the Google diaspora have certainly spread the technology via Jax!
Regardless, this is an interesting paper and a good contribution in the larger economics of open source. More foundation led projects are better for everyone, so I’m glad to see research in these areas!