This month, the US Department of Energy’s (DOE) released a Request for Information on their Frontiers in Artificial Intelligence for Science, Security, and Technology (FASST) initiative. Mozilla was eager to provide feedback, particularly given our recent focus on the emerging conversation around Public AI.
The Department of Energy’s (DOE’s) FASST initiative has the potential to create the foundation for Public AI infrastructure, which will not only help to enable increased access to critical technologies within the government that can be leveraged to create more efficient and useful services, but also potentially catalyze non-governmental innovation.
In addressing DOE’s questions outlined in the RFI, Mozilla focused on key themes including the myriad benefits of open source, the need to keep competition related to the whole AI stack top of mind, and the opportunity or FASST to help lead the development of Public AI by creating the program as “public” by default.
Below, we set out ideas in more depth. Mozilla’s response to DOE in full can be found here.
- Benefits of Open Source: Given Mozilla’s long standing support of the open source community, a clear through line in Mozilla’s responses to DOE’s questions is the importance of open source in advancing key government objectives. Below are four key themes related to the benefits of open source:
- Economic Security: Open source by its nature enables the more rapid proliferation of a technology and according to NTIA’s report on Dual-Use Foundation Models with Widely Available Model Weights, “They diversify and expand the array of actors, including less resourced actors, that participate in AI research and development.” For the United States, whose competitive advantage in global competition is its innovative private sector, the rapid proliferation of newly accessible technologies means that new businesses can be created on the back of a new technology, speeding innovation. Existing businesses, whether a hospital or a factory, can more easily adopt new technologies as well, helping to increase efficiency.
- Expanding the Market for AI: While costs are rapidly decreasing, the use of cutting edge AI products purchased from major labs and big tech companies are not cheap. Many small businesses, research institutions, and nonprofits would be unable to benefit from the AI boom if they did not have the option to use freely available open source AI models. This means that more people around the world get access to American built open source technologies, furthering the use of American technology tools and standards, while forging deeper economic and technological ties.
- Security & Safety: Open source has had demonstrable security and safety benefits. Rather than a model of “security through obscurity,” open source AI thrives from having many eyes examining code bases and models for exploits by harnessing the wisdom of the crowd to find issues, whether related to discriminatory outputs from LLMs or security vulnerabilities.
- Resource Optimization: Open source in AI means more than freely downloadable model weights – it means considering how to make the entire AI stack more open and transparent, from the energy cost of training to data on the resources used to develop the chips necessary to train and operate AI models. By making more information on AI’s resource usage open and transparent, we can collectively work to optimize the efficiency of AI, ensuring that the benefits truly outweigh the costs.
- Keep Competition Top of Mind: The U.S. government wields outsized influence in shaping markets as its role not just as a promulgator of standards and regulations but due to its purchasing power. We urge the DOE to consider broader competitive concerns when determining potential vendors and partnerships for products and services, ranging from cloud resources to semiconductors. This would foster a more competitive AI ecosystem, as noted in OMB’s guidance to Advance the Responsible Acquisition of AI in Government which highlights the importance of promoting competition in procurement of AI. The DOE should make an effort to work with a range of partners and civil society organizations rather than defaulting to standard government partners and big tech companies.
- Making FASST “Public” By Default: It is critical that as FASST engages in the development of new models, datasets, and other tools and resources, it defaults to making its work public by default. This may mean directly open sourcing datasets and models, or working with partners, civil society, academia, and beyond to advance access to AI assets which can provide public value.
We applaud DOE’s commitment to advancing open, public-focused AI, and we’re excited about the potential of the FASST program. Mozilla is eager to work alongside DOE and other partners to make sure FASST supports the development of technology that serves the public good. Here’s to a future where AI is open, accessible, and beneficial for everyone.