Each time a new version of ChatGPT or Claude comes out, the software is tested against easy-to-understand benchmarks like how fast and accurately the artificial intelligence can pass a physics exam or a medical board test.
Now the startup Atomic Canyon is working with the Idaho National Laboratory to create the best version of that testing for nuclear energy, Latitude Media has learned.
The company is set to announce a deal with the federal laboratory tomorrow to determine benchmarks that will calculate how effective different software programs are for designing and operating atomic power plants. The standards will be publicly released.
Founded in 2023, Atomic Canyon got its start cataloging documents at the Diablo Canyon nuclear plant. The startup helped sort through mountains of papers as California’s last atomic generating station prepared for relicensing after the state abandoned its plan to permanently phase out fission energy. The company is now in discussions with utilities representing “40% to 50% of the U.S. nuclear fleet,” CEO Trey Lauderdale told Latitude Media.
“A lot of these utilities, their IT teams came and gave everyone broad access to AI tools. There was a period of excitement where everyone said, ‘This is great; look at this cool toolset,’” Lauderdale said. “But then people started to try to do more advanced functions within their selected technology vendors’ ecosystems and realized the limitations hit way faster than expected.”
The potential for AI to improve on the inefficient, costly building practices for reactors has been hyped for much of the last two years.
The International Atomic Energy Agency cited generative AI as a serious innovation in 2023. In July, The Nuclear Company, a developer looking to improve the economics of building whole fleets of large-scale Westinghouse AP1000 reactors, inked a deal with software giant Palantir to design an AI program to equip construction teams with the equivalent of an “Iron Man” suit of information. That same month, Google announced a deal with Westinghouse to use its AI software to improve the nuclear developer’s own workflows.
But Lauderdale warned that “it takes a lot of effort and research and development to create and build applications that will actually work reliably and at scale.” When Atomic Canyon started, he said, “we didn’t take a broad ‘Our AI solves every problem’ approach. We were very methodical and very focused.” After visiting several nuclear plants and talking to them about their software needs, he added, it became clear that the first area of focus was just “helping them find documents effectively.”
With all the hype in AI and the growing competition to deploy it within the nuclear space, he said, it’s critical to establish some industry-wide standards that can cut through the noise.
“We believe it’s our responsibility as the lead in generative AI and nuclear to create the first benchmark for the industry in how these large language models fare when they’re used,” Lauderdale said. “To do that, we wanted to make sure we had a neutral third party to join us. That’s where Idaho National Labor came in.”


