In July, Meta unveiled its Llama 2 large language model, offering it relatively openly and for free, setting it apart from its competitors. However, the open-source community had reservations about the true extent of Meta’s openness.
While Meta’s license makes Llama 2 free for many, it doesn’t align completely with the Open Source Initiative’s (OSI) requirements for true open source. True open source involves free redistribution, access to source code, the allowance for modifications, and no ties to specific products. Meta’s limitations include a license fee for developers with over 700 million daily users and restrictions on other models training on Llama 2. Some argue that calling it open source is misleading.
Joelle Pineau, Meta’s VP for AI research, acknowledges these limitations. She argues that this approach strikes a balance between information sharing and protecting Meta’s business interests. Pineau emphasizes that their approach has driven responsible and safe AI research within the company.
“Being open has internally changed how we approach research, and it drives us not to release anything that isn’t very safe and be responsible at the onset,” Pineau says.
Meta’s openness has been exemplified by projects like PyTorch, which has thrived since its open-sourcing in 2016. The decision on how much to release depends on factors like the safety of the code in the hands of external developers.
Meta is actively engaged with industry groups to set benchmarks and guidelines for responsible AI. Pineau believes collaboration is key to advancing open-source AI safely.
The debate around Meta’s openness is in contrast to the practices of other big AI companies like OpenAI and Google, which have been more cautious about sharing research.
The conversation also delves into the inadequacy of current licensing schemes for AI models, which differ from traditional software due to increased risks and liability.
In response, the Open Source Initiative is exploring how to adapt licenses for AI models, striving to maintain the core principles of open source.
