As artificial intelligence becomes increasingly central to innovation, the promise of "open-source AI" remains both compelling and contested. This talk delves into the definitional and practical challenges surrounding open-source foundation models, drawing from recent stakeholder interviews, regulatory mapping, and primary research conducted for a forthcoming White Paper by SFLC.in.
We begin by questioning what it truly means for AI to be "open." While software licensing frameworks like GPL, MIT, and Apache are widely understood, the notion of openness in AI models particularly foundation models, remains fluid and vulnerable to "open-washing." The talk highlights how stakeholders from startups, academia, civil society, and industry offer conflicting views on openness, ranging from publishing model weights to full transparency of training datasets.
The second half of the talk addresses the benefits, and risks associated with open-source AI. We explore how open models enhance localization, autonomy, and innovation, while also creating vectors for misuse, data manipulation, and environmental strain. Finally, we touch upon the regulatory gap in India, outlining diverse views on what regulation should look like from light-touch policy nudges to robust governance frameworks.
As artificial intelligence becomes increasingly central to innovation, the promise of "open-source AI" remains both compelling and contested. This talk delves into the definitional and practical challenges surrounding open-source foundation models, drawing from recent stakeholder interviews, regulatory mapping, and primary research conducted for a forthcoming White Paper by SFLC.in.
This talk begins by tracing the history of open-source software licensing—from the Free Software movement and the GNU Public License (GPL) to permissive frameworks like MIT and Apache. These licenses codified a shared understanding of openness: transparent, modifiable, and redistributable code. However, as we enter the era of large-scale AI systems and foundation models, these definitions begin to seem inefficient.
We question what it truly means for AI to be "open." While software licensing frameworks like GPL, MIT, and Apache are widely understood, the notion of openness in AI models—particularly foundation models—remains fluid and vulnerable to "open-washing." The talk highlights how stakeholders from startups, academia, civil society, and industry offer conflicting views on openness, ranging from publishing model weights to full transparency of training datasets.
The second half of the talk addresses the drivers, benefits, and risks associated with open-source AI. We explore how open models enhance localization, autonomy, and innovation—while also creating vectors for misuse, data manipulation, and environmental strain. Finally, we touch upon the regulatory gap in India, outlining diverse views on what regulation should look like—from light-touch policy nudges to robust governance frameworks.