In July, Meta unveiled Llama 2, its large language model, marking a move towards greater openness. However, the degree of this openness has raised questions, especially in the context of open-source software principles.
Meta's Llama 2, while released for free, comes with a license that does not fully align with the Open Source Initiative's criteria. This has prompted discussions about the authenticity of Meta's open-source commitment.
The Open Source Initiative's Open Source Definition emphasizes not only code sharing but also free redistribution, source code access, modifiability, and a lack of ties to specific products. Meta's license does not meet all these criteria.
Joelle Pineau, FAIR lead at Meta, acknowledges the limitations of the company's approach to open source. She believes it's crucial to balance the benefits of information-sharing with the potential risks to Meta's business.
Pineau explains that the extent of code release depends on the maturity and safety of the research. When potential harm and safety concerns are involved, Meta is cautious about sharing research with a broader audience.
Meta aims to provide a diverse set of researchers with access to their work to receive valuable feedback. This ethos guided the release of Llama 2, emphasizing the importance of collaborative innovation.
The company actively participates in industry groups like the Partnership on AI and MLCommons to help establish foundation model benchmarks and guidelines for responsible AI deployment. It acknowledges the need for a collective approach to open source in the AI community.
Meta's approach to open source stands in contrast to industry trends, where AI companies have shifted from open research to a more guarded stance due to competitive and safety concerns.
Existing licensing schemes were not designed to accommodate AI models, which involve vast external data sources. AI models carry different risks, and Pineau suggests that licensing terms should evolve to address these unique challenges.
Industry stakeholders and organizations like the Open Source Initiative (OSI) are exploring the limitations of current open-source licenses for large language models in the commercial domain. OSI is working on redefining open source in the context of AI.
Meta's approach to openness with Llama 2 highlights the complexity of reconciling AI innovation, business interests, and open source principles. It brings to the fore the need to redefine open source licenses for AI models and maintain a balance between safety, collaboration, and competition in the ever-evolving AI landscape.
The discussion around open source in the AI industry is likely to intensify as stakeholders grapple with the unique challenges posed by AI models. Striking the right balance between open innovation and business protection will remain a key focus in the years to come.
Read more about Meta: