How open source is shaping AI developments

AI safety

Regarding AI safety, an area open source is suited to address because of the transparent nature of open-source software development, Zemlin pointed to the Coalition for Content Provenance and Authenticity (C2PA), a project focused on ensuring the authenticity of digital content.

“In the world of generative AI, ensuring content authenticity is going to be deeply important,” he said. “C2PA provides a digital immutable watermarking technology that can track content from the creator all the way to the publisher, helping us to understand what’s real and what’s AI-generated.”

Zemlin also highlighted the Model Openness Framework, designed to help organisations evaluate and categorise the level of openness associated with different AI models – even those deemed as open source.

“Because there are so many moving parts in the production and deployment of large language models, we created a grading system to help organisations understand what components are open and included in a model versus what’s not,” he said.

The framework defines three levels of openness for an AI model, starting with level one, the highest level of openness, where all the data, instructions and components used to create the model are openly available.

The second level is “open tooling”, where the majority of the tooling and infrastructure are available as open source, followed by level three, where the data is not open but data cards that describe the data sets are available.

Zemlin emphasised the importance of this risk-based nuanced approach, as it addresses the complexities of the AI ecosystem, where openness may not be binary, but rather, a spectrum across various components.

In support of model production, he pointed to the LF AI and Data Foundation that’s driving a collection of projects to provide open source implementations for every aspect of the machine learning and large language model production process. These projects include Delta Lake, an open-source storage framework for data lakehouses, and Monocle, an AI observability service.

Zemlin also touched on the Linux Foundation’s efforts in open data, such as the Overture Maps project, a collaboration between Amazon, Microsoft, TomTom and Meta to create the world’s largest shared geospatial dataset.

“You see a whole new world where data is now being sold in order to train large language models, and it’s only available to the people who have the most money to buy that data,” he said.

“There’s an opportunity for us to also have an open alternative to closed data, but it’s something that’s going to require resources and a very focused approach. Our Overture Maps effort is an early example of how that can be done.”