Meta Platforms is under fire for using its brand in open source AI. What impact will this controversy have on the company’s future?
In a recent statement, the Open Source Initiative (OSI), a software standards organization, sharply criticized Meta Platforms for confusing the public by presenting their Llama models as open source. According to OSI, this designation is misleading and creates confusion about the true open source nature of the models.
OSI, which claims to have coined the term “open source” some 25 years ago and is its current custodian, declared to Financial Times that Meta Platforms corrupts the definition and creates confusion about which models are truly open.
Open Source requirements
For an AI model to be considered open source, OSI maintains that companies must:
- Disclose the training algorithms and other software used to develop them,
- Be transparent about what we call “model weights”, a numerical parameter that defines the strength of the signal between neural networks.
OSI plans to issue guidance on this next week.
Stefano Maffulli, CEO of OSI, told FT that Meta’s open source AI labeling is “extremely harmful” at a time when regulatory bodies, including the European Commission, support the development of “genuine open source technologies”.
Google, Microsoft and IBM face the issue of Open Source
Maffulli pointed out that Google and Microsoft have stopped promoting their models as open source, but that discussions with Meta Platforms “did not achieve a similar result”.
Dario Gil, head of research at IBM, also lamented a lack of transparency around Meta Platforms’ publications explaining the technical details of their model development. However, he added that the company’s models offer a better alternative to “black box models” whose inner workings are not made public.
Furthermore, the license Meta Platforms uses also prevents competitors from using their family models, contradicting the meaning of open source, despite the fact that anyone can download the systems for free, according to FT.
In response to the allegations, Meta Platforms said FT : “Existing open source definitions for software do not cover the complexity of today’s rapidly evolving AI models. We are committed to continuing to work with industry on new definitions to safely and responsibly serve everyone in the AI ​​community.”
Meta Platforms opened access to its Llama 2 model in 2023, noting at the time that this would allow “for a generation of developers and researchers” to test his models. Today, his Llama designs have been downloaded more than 400 million times.