Why Meta Said Open AI Infrastructure Serves Its Business

I have praised Meta more recently for sharing AI work, but a viewer prompted me to look further back. I had not given the company enough credit for its earlier open software contributions. In this February 2024 video, I review Mark Zuckerberg’s explanation of the strategy and look at some of those projects.

In a clip, Zuckerberg describes Meta’s long-term goals as building general intelligence, open sourcing it responsibly and making it useful to people. That is an aspiration, not a demonstration that Meta has achieved AGI or an unconditional guarantee about any future release. In earlier public remarks, Zuckerberg said Meta could reconsider its approach if systems approached superintelligence.

I turn to Zuckerberg’s remarks on Meta’s fourth-quarter 2023 earnings call, highlighted in a post by PyTorch co-creator Soumith Chintala. Zuckerberg says that sharing the results of costly research can improve models through outside input, while the work needed to turn a model into a product leaves room for companies to differentiate.

His first argument is that community scrutiny, feedback and development can make software safer, more secure and more efficient. I agree that a large community can find problems and contribute fixes that a smaller internal team might miss. Openness can enable scrutiny, although its actual safety and security depend on review, resources and deployment rather than following automatically.

Zuckerberg’s second argument is that shared software can become an industry standard. If others build on Meta’s stack, the company can more easily integrate innovations from the wider ecosystem. I think of established examples such as Linux, Python and Firefox to illustrate how widely adopted open projects can support a broad community; these are examples, not Meta products.

His third reason is recruiting: developers and researchers may want their work to be used widely and improved collaboratively. I find that appealing because it aligns a company’s business interests with engineers’ desire to share and refine their work. Zuckerberg argues that attracting strong people matters when competing in a new technology area.

Meta’s case also depends on keeping advantages elsewhere. Zuckerberg says unique data and product integrations can distinguish Meta’s services even if general infrastructure such as Llama is shared. He describes a strategy of sharing infrastructure while retaining specific product implementations; that does not imply all user data is made public or that every future model will be released.

A response highlighted projects associated with Meta, including React, PyTorch, React Native, GraphQL, Jest, Flow, Yarn, Hermes, FBT, Prophet and Cassandra. I had focused too narrowly on Llama and underestimated this broader history. The list also mentions Mercurial, a version-control system Meta used and contributed to, not a project Meta originally created.

Meta’s open-source catalog contains many more projects and research assets. I scroll past work such as Segment Anything, Detectron and AudioCraft, which I have discussed as a text-to-audio tool. Individual projects differ in their licenses and what they release, so the catalog should not be read as one uniform set of permissions.

I also notice Apple releasing AI papers and projects, although the examples do not establish that Meta caused that change. I look forward to Llama 3, which Zuckerberg said was training at the time. Meta described Llama 2 as open source, but its community license included restrictions; the Open Source Initiative did not consider that license open source. Whether any future AGI is achieved or released remains an open question.