Unsurprisingly then, a coterie of open-source-is-what-the-OSI-says-it-is advocates (note: I’ve historically been in this camp) are castigating Meta for calling its Llama large language model (LLM) open source, despite restrictions that fall short of the OSI’s Open Source Definition. The industry’s response has been a collective shrug. See, for example, “Why Meta LLaMA Models Are Open Source” — a title that must drive OSI folks crazy. Part of this stems from, as one HackerNews commentator says, the idea that “Meta, through the Llama models, has done more for open source LLMs than just about anyone else.” Rewinding even further, open sourcerors can look to Apache Cassandra, React, GraphQL, PyTorch, and other Meta projects that met the OSI’s bar for open source.
It’s hard to get too grumpy with a company that has created some of the industry’s most important open source projects.
And yet some people are very grumpy, despite the fact that there was (and is) no settled definition for open source in AI. Yes, the OSI finally released a definition of open source for AI, the Open Source AI Definition 1.0, but, as with cloud, the OSI is playing catch-up, and its definition has disappointed some of its most ardent supporters by not dictating that training data also be open.