In January, the National Science Foundation announced the creation of an open-source AI program to fulfill a commitment of President Biden’s AI executive order. The National Artificial Intelligence Research Resource (NAIRR) pilot, according to the NSF, “is a vision for a shared national research infrastructure for responsible discovery and innovation in AI.” The pilot would serve, in other words, as a hub for citizens to find the algorithmic, data, and computing resources to support AI education and research and development. While the program is promising, it spreads meager resources too thin across competing priorities. To ensure its long-term viability, the pilot should focus exclusively on education and talent-development initiatives.
NAIRR certainly fills a need. Running and accessing the required computing power (“compute” in AI argot) is expensive, and data are in short supply. The pilot’s open-source vision could be a boon to research by allowing cash-strapped students to run and train otherwise-unaffordable AI models and enabling professional researchers to broaden their discoveries. A public program like NAIRR could thus propel progress by expanding access to the necessary tools.
The program’s dual focus on education and R&D could thwart this progress, however. NAIRR is slated to receive only $430 million, which likely isn’t enough funding to sustain the type of big-ticket, leading-edge research that the program’s congressional sponsors (Senators Martin Heinrich, Todd Young, Cory Booker, and Mike Rounds) are seeking. Jack Clark of Anthropic, a private sector AI heavyweight, recently testified that the costs of developing these leading-edge AI systems run in the hundreds of millions and could climb into the billions by 2025. With its budget, NAIRR simply won’t be able to bankroll more than a few big projects per year. After the program’s resources get spread across administration, education, and small dollar grants, the pilot may fail to deliver even one effort that meets its sponsors’ vision.
It’s questionable whether directing NAIRR resources for R&D is even the best way to encourage groundbreaking AI innovation. According to a 2023 survey that asked 410 AI researchers what resources most hinder research, 90 percent rated “specialized knowledge, talent, or skills” as the most severe limitation and the biggest cause of project abandonment. Fewer researchers felt compute (52 percent) and data (51 percent) limitations were similarly debilitating. While costly compute causes research friction, human-capital deficits are the biggest constraint on AI progress.
Using NAIRR for talent development is the lever most likely to advance the national interest on AI. According to McKinsey, most AI companies struggle to find the needed talent to develop and diffuse AI tools through the economy. Private-sector AI adoption is unexpectedly falling as a result; between 2019 and early 2023, the proportion of companies that reported adopting at least one AI product fell from 58 percent to 50 percent. By August, recent publicity had yielded only a 5 percent recovery. Without a robust talent pool, organizations have dialed down their AI ambitions.
By focusing on education and talent development, NAIRR would meet the demands of the moment. Moving forward, the program’s current education and talent plan offers a sound starting point. Recognizing that resource costs can constrain talent development, NAIRR will provide a select group of students and educators a digital portal where they can search relevant resources, data sets, and models, and submit code for processing.
To enhance its impact, NAIRR should focus on filling post-secondary educational-resource gaps, giving students the tools to lower the prohibitively high costs of data and compute. With robust congressional funding, NAIRR could also build AI cybersecurity educational “sandboxes” (isolated testing environments, currently limited in number) to enable students to test, attack, and probe models in a safe and legal setting. The program could even extend resources to the 50 percent of secondary schools without computer-science funding.
At minimum, Washington should treat NAIRR genuinely as a pilot program, scrutinizing the return on investment of both the R&D and education pillars in their current forms. Based on pilot results, Congress should prepare to alter legislation to ensure future iterations pursue the most valuable programs.
The federal government has a unique chance to help propel AI progress. By focusing NAIRR on talent development, America can boost AI research, diffusion, and, ultimately, abundance.
Photo: Vertigo3d/E+ via Getty Images