Feds kick off National AI Research Resource with pilot program live today
A year to the day after it was proposed, the National AI Research Resource (NAIRR) is coming online — at least in pilot form — as a coalition of U.S. agencies and private partners start to apply billions in federal funding toward public-access tools for aspiring AI scientists and engineers.
NAIRR is the Biden administration's answer to the sudden rise of AI in the global tech scene, and the concentration of its resources and expertise among a relatively small group of tech giants and privately funded startups. In an attempt to democratize the tech a bit and keep the U.S. competitive with its rivals abroad, the feds decided to dedicate some money to making a variety of resources available to any qualified researcher.
The National Science Foundation, Department of Energy, NASA, NOAA, DARPA and others are all partners in the effort, both providing resources (like datasets and consultation) and working with applicants in their areas of expertise. And more than two dozen major tech companies are also contributing in some way. The whole thing has an $800 million per-year budget for the next three years, subject to congressional approval, of course.
In a panoply of statements, executives from OpenAI, Anthropic, Nvidia, Meta, Amazon, and Microsoft commit a variety of resources, expertise, free access, and so on to the NAIRR effort.
The resources that will be made available haven't been listed anywhere. Instead, the overall organization will accept applications and proposals, which will be evaluated and assigned resources. Think of it more like a grant-making process than a free supercomputer.
As the NSF's Katie Antypas put it, NAIRR "will provide the research community access to the computing the data, the models, the software, and the training resources that are necessary to advance the AI ecosystem. The NAIRR pilot is really needed because the resources needed to even begin participating in the ecosystem have become increasingly concentrated and inaccessible to many, many communities that are really essential for developing a healthy and responsible AI ecosystem. And so the pilot is the first step to bridging this gap."
She gave three examples: a researcher looking at large AI models who needs large scale computing resources and has no way of accessing it; a teacher who wants to let kids do AI-related homework (like training custom models) but needs resources like virtual notebooks and compute time; and someone looking at predicting climate and weather events, who can access NASA and NOAA datasets and combine them with hosted models.
For the two-year pilot period, there will be four focus areas:
NAIRR Open, the most general category, involving "access to diverse AI resources," presumably for research and projects that don't fit the narrow categories following.
NAIRR Secure, focused on AI applications that need privacy and security, such as medical and critical infrastructure. This part is led by the NIH and Energy, unsurprisingly.
NAIRR Software is more focused on tools, platforms, services, and interoperability.
NAIRR Classroom is about outreach, education, and training.
It may surprise some that there is no outwardly military research category (considering the presence of DARPA and the DOD in the partner agencies), but remember this is a civilian research effort led by the Executive agencies. Presumably any military research will be siloed, and the military agencies are there to coordinate and delegate, or offer such resources as are appropriate in this case.
The idea is that if someone has a worthwhile idea about how to apply or advance AI in any sector, there should be a domain expert and a check waiting. But it won't be like a public library, where you walk into your local AI center and someone sets you up with an H100 for a half-hour slot. (That said, I wouldn't be surprised if there were some kind of library-centric outreach program.)
You'll be able to peruse the list of resources at the NAIRR Pilot page starting today, and although there are no hard numbers yet, the leaders of the project said that only 25 to 50 proposals will likely be accepted in this initial pilot period, with hundreds more spots opening up in the spring when more systems come online.