DOE aims to move "FASST" on AI with sweeping new initiative

• "Imagine we had a basic science AI foundational model like ChatGPT for English — but it speaks physics and chemistry," Deputy Energy Secretary David Turk said in announcing the initiative.
• Combine that "with the world-class laboratory test facilities we have at [DOE] labs and you will get a sense of the incredible potential here," he said, adding it is already happening with fusion ignition research at Lawrence Livermore National Lab.

Why it matters: The DOE has world-class supercomputing, a powerful scientific infrastructure and experience working with dual-use technologies that position it to power AI advances for science and national security.

• "It is arguably the most important AI initiative yet from the Biden administration" considering the ambition, scale, funding and focus squarely on AI, says Divyansh Kaushik, a VP at Beacon Global Strategies who focuses on critical and emerging tech.
• "The president's budget request for $455 million is a starting point but it remains to be seen what DOE can do with that amount of money and they certainly will need a lot more if you compare to private sector investments," Kaushik says, adding it will arguably require tens of billions of dollars over five years.
• One former Congressional staffer briefed last year by the national labs about their AI initiatives concurred, telling Axios that level of support was discussed.

Zoom in: FASST focuses on how the DOE can leverage its supercomputing resources, data it collects from scientific work at its facilities and the know-how of scientists at the 17 national labs to "supercharge a new AI initiative that will build on this infrastructure," Helena Fu, DOE's director of the Office of Critical and Emerging Technologies, tells Axios.

• The initiative is "going to be focused on solving mission-critical challenges that the private sector won't be investing in or can't invest in based on the specific interests that the U.S. government has," she said. Those include DOE's national security missions and the use of AI for scientific discovery.

It also emphasizes producing scientific data for the development of next-generation AI models, or to solve specific problems and scientific challenges, including automated labs.

• "It's not going to be necessarily the large language models today that are going to give you travel recommendations and help you write essays or summarize points," Fu said.
• It's less about the general models that power chatbots and more about "how you use models that have the language of biology or chemistry or physics to solve those kinds of scientific challenges."
• In a partnership with Microsoft, the DOE designed and developed a new battery material that uses 70% less lithium. It was a proof-of-concept of how AI could help to speed the discovery process.

Supercomputing is also a central part of the initiative. It aims to develop the next generation of computing architectures to support AI and to tackle the technology's growing energy demand, which Energy Secretary Jennifer Granholm recently told Axios' Hans Nichols is a "problem."

• The DOE is already home to four of the world's fastest supercomputers — including the fastest, Frontier.

Another focus will be the development of "trustworthy and secure AI."

• The department, through the National Nuclear Security Administration, has been evaluating and red-teaming models as part of the White House AI executive order.
• More broadly, the initiative aims to develop foundation models to understand both how models work to determine whether (and what) those markers of trustworthiness are; and develop models for DOE's science, energy and security missions.

Zoom out: The DOE initiative aims to create models tailored to specific applications around science, energy and national security.

• One area where AI could first be applied is in nuclear stockpile modernization. The DOE is also uniquely positioned to train models on classified data, for example, analyzing large amounts of classified data.
• Another is in efforts to protect the electric grid, what a recent DOE report calls one of the "most complex machines" on Earth, and make it more efficient.

Yes, but: Exclusively housing classified models with security applications under the DOE could risk giving U.S. adversaries "a single point of attack to extract classified data if they were to gain access to the frontier model trained on classified data, so advanced cybersecurity and security by design methods will be critical," Kaushik says.

Go deeper:

Department of Energy dives into promises and perils of AI (Axios)