Latest / Minds, Bodies, and Terawatts / US Commerce Department to Test Frontier AI Models

Transcript

US Commerce Department to Test Frontier AI Models · Minds, Bodies, and Terawatts

• 0:00
  [Host] This is Minds, Bodies, and Terawatts, where deregulation meets "wait, maybe not." The Washington Post reported yesterday that the Commerce Department will test frontier AI models from Google, Microsoft, and xAI before they're released, a sharp reversal from a White House that spent a year calling AI regulation "foolish" and the industry "a newborn baby" that shouldn't be constrained.
• 0:25
  [Guest] The Center for AI Standards and Innovation, CAISI, sits under NIST and already had agreements with OpenAI and Anthropic from the Biden era. Now three more companies are in. Director Chris Fall says these collaborations help them "scale our work in the public interest at a critical moment." Their February report, "Evaluating Dual-Use Capabilities in Frontier Systems," documented stress-testing models for autonomous cyber offense and biological design, over 40 evaluations, some on models the public never saw.
• 0:59
  [Host] What changed since last July when Trump was all "let the baby grow"?
• 1:05
  [Guest] Anthropic's Mythos model. In April, Anthropic revealed it built an AI so good at finding software vulnerabilities that it refused to release the thing. Thousands of flaws in operating systems and browsers, surfaced by a single model. They gave access to only about 50 critical infrastructure organizations through Project Glasswing and stopped there.
• 1:28
  [Host] And overnight Washington wanted safety gates it had been mocking for months.
• 1:33
  [Guest] Dean Ball, co-author of the White House AI Action Plan itself, told The Hill the administration "was not prepared to deal with this." David Sacks, the AI czar who championed the hands-off approach, stepped down in March. His portfolio got split between Chief of Staff Susie Wiles and Treasury Secretary Bessent. The ideological architect left the building.
• 1:58
  [Host] CAISI is the Biden-era AI Safety Institute with a new name. Lutnick rebranded it last June, calling the original "regulation under the guise of national security." The core work, testing models for cybersecurity and biosecurity risks in classified settings, never stopped. Even Chris Fall, the new director, replaced a former Anthropic staffer who was fired within days of being appointed.
• 2:24
  [Guest] Now the New York Times reported May 4th the White House is weighing an executive order to create a formal working group of tech executives and government officials to vet models before release. That moves pre-deployment testing from voluntary cooperation to structured mandate.
• 2:43
  [Host] Who's missing from the new agreements? Anthropic, the company whose model triggered the whole pivot.
• 2:50
  [Guest] They clashed with the administration over autonomous weapons and surveillance policy. They have a renegotiated deal from 2024, but the Pentagon's recent AI contracts with seven companies pointedly left them out. They convinced Washington the house was on fire, and Washington responded by hiring the other tenants as firefighters.
• 3:11
  [Host] Does CAISI stay a referee, measuring and reporting on model capabilities, or does it cross into deciding which models reach the public?
• 3:21
  [Guest] Right now the agreements are voluntary. But an executive order changes the game. Once government sits in the release pipeline, "testing" and "licensing" become hard to tell apart.
• 3:33
  [Host] And the compliance cost falls hardest on the small labs, the ones least able to absorb it.
• 3:41
  [Guest] Google and Microsoft can staff compliance teams. A ten-person lab cannot. The deregulation president may have just handed incumbents a moat they didn't have to dig.
• 3:52
  [Host] So designing oversight after a crisis instead of before it means the rules protect whoever's already at the table.
• 4:00
  [Guest] That's the pattern. And it maps to what Patrick Deglon calls the difference between a referee and a ruler in Unscarcity. A referee measures and reports. A ruler decides who gets to play. The moment CAISI crosses that line, it stops calling balls and strikes.
• 4:20
  [Host] For more on AI governance and the referee-versus-ruler framework, head to unscarcity dot ai. The book is on Amazon, the audiobook is on Spotify. This has been Minds, Bodies, and Terawatts.