The open-source exemption As outlined in the executive order issued by the White House in June, federal agencies were charged with developing a classified benchmarking system to determine what qualifies as a frontier model covered by the framework. Federal agencies were also tapped to develop a framework by Aug. 1 that would allow the government early access to frontier models to review them for dangerous cyber capabilities. The framework also specified that a list of “trusted partners” would be permitted access to the models following a period of up to 30 days of government review. Other than the big names that participated in Tuesday’s meeting, such as OpenAI, Anthropic and Google, it looks like most companies will be kept in the dark about what that process looks like. “Appian was not included in the White House briefing. I prefer transparency regarding regulatory methods and objectives,” said Matt Calkins, the CEO of cloud company Appian, in a written statement. “Perhaps the secrecy this time is intended to avoid a public conflict and allow standards to evolve. However, this approach creates a group of insiders while excluding others, which impedes innovation.” Another potential explanation is that the Trump administration simply doesn’t have the technical expertise to understand what it’s supposed to be regulating. SpaceX CEO Elon Musk suggested in an interview with The Economist in July that this was in fact the case — and that AI company CEOs should instead have biweekly calls to review each other’s models and potential associated risks. “The bigger problem I think this [shows] is that they don’t actually know what they’re trying to regulate,” speculated Kristian Stout, Director of Innovation Policy at the International Center for Law & Economics. The White House is looking at the different frontier AI models “and they’re like, how do we even know what it is we’re trying to regulate in advance? What does it mean that it has advanced cyber capabilities until we know it hacked into something?” Stout said. The open-source exemption: The Trump administration will win plaudits for its decision to exempt open-source models from review. The open-source community had been leery that the administration would move to ban or sanction Chinese open-source models because of concerns about security and intellectual property theft. But it’s also difficult to see how the administration could subject open models to a review without permanently setting American companies behind the leading Chinese AI companies. Unlike closed, licensed models, it’s also impossible for the government to force a retraction of an open model once published online. “If they stymie American open-source development at this point, you are ceding a lot of ground to Chinese developers,” said Stout. “American companies are finding that the cost efficiency curves for using open-weight models in their organizations are becoming pretty compelling relative to the frontier stuff from the proprietary labs,” he added. Programming note: Join myself, Stout, Council on Foreign Relations senior fellow Chris McGuire and WP Intelligence Editorial Director Luiza Savage for a live briefing this Friday on the future of open-source AI. We’ll discuss the competitive landscape and the economics of closed vs. open-source models, as well as the national security angle on Chinese models The China angle The flip side to the administration’s exemption of open-source is that national security hawks in Washington are pointing out very real national security threats posed by Chinese open-source models, such as code that carries propaganda or malware that can be triggered deep inside American infrastructure, as WP Intelligence’s Lead Global Security Analyst Josh Rogin and researcher Kendrick Frenckel write in a new report. The concerns are prompting China hawks such as Rep. Ro Khanna (D-California), the top Democrat on the House Select Committee on the CCP, to propose implementing a framework for testing open-source models. Rep. John Moolenaar (R-Michigan), chairman of the House Select Committee on the CCP, has likewise introduced a bill that would establish a list of AI systems created in foreign adversary countries and require federal agencies to strip the listed AI off their own systems. The legislative developments come as the capability gap between Chinese open-source AI models and the licensed U.S. frontier models is rapidly compressing. “The gap is definitely compressing, and in certain ways, it is already closed,” said Angelopoulos. “[Beijing-based Moonshot AI’s] Kimi K3 surpasses Fable-level performance on our web development benchmark and, of course, web developers are more than half of all coders.” “The areas where the Chinese models are beginning to exceed American models are really important,” he said. According to several current and former officials in the Trump administration, Treasury Secretary Scott Bessent and White House Office of Science and Technology Policy Director Michael Kratsios favor a harsher approach to cracking down on Chinese AI models. But other factions within the administration see sanctions against those models as counterproductive to innovation. There’s another key piece to the puzzle here that has been hotly contested in Washington: export controls and compute. Samm Sacks, a senior fellow at New America who recently returned to the U.S. from visits to Chinese labs, told Josh that compute remains “a massive constraint” to Chinese AI innovation. “If you talk to any of the Chinese frontier labs, you’ll hear that this is probably the number one barrier that export controls have really taken a bite,” Sacks said. This newsletter is published by WP Intelligence, The Washington Post’s subscription service for professionals that provides business, policy and thought leaders with actionable insights. WP Intelligence operates independently from The Washington Post newsroom. Learn more about WP Intelligence. | 
