Open Weight Definition Provides New Perspective on OSAID Debate

The debate over what constitutes open-source AI continues to evolve, with significant input from both developers and industry leaders. The Open Source Initiative (OSI) has been striving to establish a clear definition with its Open Source AI Definition (OSAID), but disagreements persist about what should be included. In response, the newly formed Open Source Alliance (OSA) has introduced its own framework, the Open Weight Definition (OWD), which seeks to address the unique challenges of AI development.

The OWD aims to strike a balance between closed and open-source AI by setting standards for what counts as "open source" in AI, especially for large language models (LLMs). It focuses on three key areas:

Key Components of the Open Weight Definition (OWD)

1. Model Weights Accessibility

The OWD emphasizes the importance of making model weights—the numerical values that connect nodes across AI layers—accessible to developers and researchers. These weights, determined during training, are crucial for understanding and improving AI models.

2. Dataset Information

While the OWD does not mandate full access to training data, it highlights the necessity of providing detailed information about the dataset's contents and collection methods. This transparency helps in assessing the model's performance and potential biases.

3. Architecture Transparency

The framework encourages the disclosure of the model's architecture, which can aid in further enhancements and modifications by the community.

Amanda Brock, CEO of OpenUK, supports the OWD, viewing it as a step toward better global collaboration. She critiques the OSI's OSAID, suggesting that the OWD's focus on disaggregating AI components is more practical. Brock's perspective is that defining the openness of individual elements like data, weights, and models is more effective than a broad, potentially unfit definition.

The OSA, led by founder Sam Johnston, aims to expand the traditional Open Source Definition (OSD) to include AI-specific elements like weights, proposing an "Open Source 2.0." This move comes amidst ongoing debates about the applicability of the OSD to AI, particularly as it relates to data and model weights.

In response to the OWD, Stefano Maffulli of the OSI pointed out that communities, such as the Linux Foundation, have already developed their own definitions of open weights. Additionally, Heather Meeker, a prominent open-source lawyer, highlighted the fundamental differences between software source code and Neural Net Weights (NNWs), arguing that the principles of open-source software licensing do not easily apply to NNWs due to their nature as learned knowledge stored in large matrices.

Meeker proposed the Open Weights Permissive License as a way to share NNWs under an open-source-like framework, emphasizing the original goals of open-source freedom. However, she acknowledged the challenges in applying these freedoms to NNWs, given their non-human-readable and non-debuggable nature.

Maffulli emphasized that the OSI's definitions, like those of the Linux Foundation, are community-driven, reflecting the collaborative nature of open-source development. He noted that the OSI is monitoring how AI practitioners apply these definitions in practice.

Despite these efforts, Meeker expressed doubt about the likelihood of any single definition becoming a de facto standard, citing the influence of regulatory frameworks, privacy regulations, and market dynamics.

The ongoing debate underscores a broader challenge: defining what truly constitutes open-source AI. While there is consensus that simply labeling an AI model or dataset as "open-source" does not suffice—as seen in the case of Meta's Llama—the community remains far from a unified definition.