A software standards organisation reportedly criticised Meta Platforms for confusing the public by promoting its Llama models as open source, a term the body claims to have coined around 25 years ago and today serves as a steward of the approach.
The group, dubbed Open Source Initiative (OSI), told Financial Times (FT) Meta Platforms is polluting the definition and creating confusion about which models are truly open.
For AI models to be open source, companies need to “disclose training algorithms and other software used to develop them”, in addition to being open about what is known as “models’ weights” – a numerical parameter defining the signal’s strength between neural networks, argued OSI, which is due to publish guidelines next week.
OSI’s executive director Stefano Maffulli told the FT Meta’s open source labelling of its AI is “extremely damaging” at a time when regulatory bodies including the European Commission are backing the development of “true open source technologies”.
Maffulli pointed out Google and Microsoft have stopped promoting its models as open source, but discussions with Meta Platforms “failed to produce a similar result”.
IBM head of research Dario Gil also bemoaned a lack of transparency around Meta Platform’s publications that explain the technicalities of its model developments. However, he added the company’s models provide a better alternative to “black box models” or models whose internal workings are not made available to the public.
Further, the license used by Meta Platforms also prevents its rivals from using its family models, which contradicts the meaning of open source, despite allowing anyone to download the systems free of charge, FT reported.
Responding to the claims, Meta Platforms told FT: “Existing open source definitions for software do not encompass the complexities of today’s rapidly advancing AI models. We are committed to keep working with the industry on new definitions to serve everyone safely and responsibly within the AI community.”
The Facebook owner opened up access to its model Llama 2 in 2023, at the time noting this will allow “a generation of developers and researchers” to test its models.
Today, its Llama models have been downloaded more than 400 million times.
Comments