Amazon ML Expert: Understanding the True Nature of Open Source Models

Meta and Mistral are positioning their open-source AI models as compelling alternatives to proprietary systems from OpenAI. However, the question arises: what truly defines an open-source machine learning system? Take Meta’s Llama 2, for example. While the company made the model’s weights and evaluation code available, it did not disclose the training data used to create it.

Julia Ferraioli, a machine learning strategist at Amazon, underscored this crucial distinction, emphasizing that the mere availability of a free system does not warrant it as "open." Speaking at the State of Open Con event in London, Ferraioli cautioned that the ability to view model checkpoints or weights does not adequately encapsulate the essence of an open-source machine learning system.

“For a machine learning system to be categorized as open, I need the capability to question it,” Ferraioli asserted. She suggested a litmus test for determining whether a system qualifies as truly open-source: a user should have access to the model, the underlying data, the code, and relevant metadata.

“Models are essentially expansive matrices,” she explained. “Access to all this supporting information is fundamental. If I have this information, I can verify, reproduce, and modify the system. Moreover, I can engage critically with it, which is a vital component of the open-source philosophy.”

The field of open-source AI is rapidly evolving, with new systems frequently emerging amid the generative AI boom. Ferraioli emphasized that machine learning serves as the backbone for generative AI systems, underscoring the need for transparency in how these systems are trained, the data used, and their intended applications for scientists and practitioners alike.

For companies and community organizations endeavoring to open-source their systems, comprehensive disclosure is essential to authentically embrace the open-source ethos. Ferraioli acknowledged that while some may question whether all aspects of a model need to be open-source, providing this access is crucial for a genuine open-source experience.

“Challenging work should not deter us,” she declared. “By dissecting complex systems into their fundamental components, we can establish a meaningful specification for open-source machine learning that emphasizes what truly matters.”

This commitment to transparency and engagement not only fosters trust among users but also propels the advancement of open-source machine learning, enabling innovation and collaboration across the tech landscape.