Meta Faces Criticism for Redefining OpenAI, Calls for Transparency

The Open Source Initiative criticizes Meta for redefining “OpenAI”, urging transparency in training algorithms and model weights.

The Open Source Initiative (OSI) group told the FT that Meta Platforms is changing the meaning of “open” and making it difficult to tell which models are open.

Companies must “disclose training algorithms and other software used to develop them,” as well as what are known as “models’ weights,” which are numbers that describe how strong the signal is between neural networks, to make AI models open source.

The OSI, which is putting out guidelines next week, said that companies must also be open about these things.

Stefano Maffulli, the executive director of OSI, told the FT Meta’s decision to call its AI “open source” is “extremely damaging” at a time when regulators like the European Commission are supporting the growth of “true open source technologies.”

Maffulli said that Google and Microsoft no longer promote their models as open source, but talks with Meta Platforms “failed to produce a similar result.”

Dario Gil, the head of research at IBM, also stated that Meta Platform’s publications, which explain the technical details of its model developments, lack clarity.

He did say, though, that the company’s models are better than “black box models,” which don’t let the public see how they work on the inside.

FT reported that Meta Platforms’ license also stops its competitors from using its family models, which goes against what “open source” means, even though the systems are free for anyone to download.

Meta Defends AI Model Approach Amid Open Source Debate

Meta Platforms responded to the claims by telling FT that current definitions of open-source software do not cover the complexity of today’s rapidly improving AI models.

We pledge to continue collaborating with the industry to develop new definitions, ensuring the safe and responsible service of everyone in the AI community.

In 2023, the owner of Facebook made its model Llama 2 public. At the time, it said that this would let “a generation of developers and researchers” test its models.