Claude Mythos scrambled American AI policy, leaving us with what Dean Ball calls a de facto involuntary licensing regime from the administration that promised us the opposite. Zhipu's cofounder says China will have a Mythos-class model before the end of the year; IAPS says February 2027. Either way, Beijing is about to face down the same question, but unlike Washington, it already has a form to fill out.

Guests: Kevin Xu of Interconnected. Matt Sheehan, a fellow at the Carnegie Endowment.

We discuss…

• Why China's answer to Mythos will look like Glasswing — but started by the state, not a lab, with government ministries and central SOEs first in line
• Whether the CAC's content-testing regime can absorb cyber, and why a Chinese lab dropping a frontier model without a heads-up would be "very bold and self-destructive behavior"
• The mixed signals on open source — a Reuters report on MOFCOM and NDRC weighing export controls on model weights, versus Minimax and Zhipu founders publicly doubling down, with Xi's WAIC speech as the tiebreaker
• Whether open weights even matter when running a frontier model adversarially requires a datacenter, and the case that lighting the dark park makes it safer
• What the US and China can actually agree on — non-state actors, eval methodology sharing, and why labor displacement lessons don't travel
• Companion-app regulation flowing from Sacramento and Albany to Beijing, the AI policy brain drain into the labs, and whether Sam Altman's five percent is American state capitalism or just a Trump thing

song: https://suno.com/s/xZE1pIKrVBtCfTBO

Learn more about your ad choices. Visit megaphone.fm/adchoices