When Meta introduced its Llama 2 AI language model, it stirred up quite the conversations by claiming to be an open-source and free AI model. However, despite Meta’s claims of being “open-source,” the AI model couldn’t meet the criteria outlined by the Open Source Initiative (OSI). Now, Joelle Pineau, Meta’s vice president for AI research, suggests that open-source licensing requirements need to evolve.
The crux of the issue
At the heart of this problem is Llama 2, which the company launched in July as the first “open-source” AI model. However, according to OSI’s requirements, for a service to be considered open-source, it should offer free distribution, full access to the source code, and allow for modifications. While Meta did fulfill some of these conditions, the fact that the company charged a licensing fee for developers with more than 700 million daily users and didn’t permit other companies to train their AI models using Llama made it ineligible for the “open-source” label.
Meta’s response to open-source requirements
When talking about the situation, Pineau acknowledged the company’s limitations but pointed out that their solution strikes a balance between sharing information and protecting Meta’s business interests. “Being open has fundamentally changed how we approach research, pushing us to release only what is safe and responsible from the start,” said Pineau.
She also argued that this isn’t the first open-source project the team has worked on. This is because, in 2016, the company introduced PyTorch, an open-source machine learning library based on the Torch library, which many use in applications like computer vision and natural language processing.
“How we choose to release our research or the code depends on the maturity of the work. When we don’t know what the harm could be or what the safety of it is, we’re careful about releasing the research to a smaller group,” Pineau added.
Furthermore, Meta also highlighted that it prioritizes feedback from a diverse group of researchers and collaborates with various industry organizations, such as the Partnership on AI and MLCommons, to develop benchmarks and guidelines for these AI models.
2023-10-31 15:05:13