White House director Michael Kratsios accused Moonshot AI of illegally distilling Anthropic‘s Fable 5 model to build its Kimi K3 system.
That charge arrived on Wednesday as the Chinese lab released the new model, which critics say matches the capabilities of leading frontier models from OpenAI and Anthropic.
The accusation and the response
The hosts of WIRED’s Uncanny Valley podcast discussed whether this event signals another DeepSeek moment in the ongoing US-China AI race. They noted that this is not the first time Anthropic has accused a Chinese lab of distilling its models.
Leah Feiger, director of politics and science, pointed to a WIRED article by Hugo Lowell describing how the Trump administration is divided over how to handle Chinese AI. The Commerce Department and Howard Lutnick argue the situation is not as dire as it seems. Conversely, other officials want an executive order to stop American IP theft.
Kratsios’s main tool remains export controls. Some observers claim these controls are failing because Chinese labs are still accessing proprietary US information. However, the hosts argued that if theft is occurring, export controls are not the failure point.
Feiger added that executive orders cannot stop China. They are merely memos with declarative intent that carry no legal force abroad.
Open weights versus proprietary secrets
Brian Barrett, executive editor, highlighted that Kimi K3 is an open-weight system. This differs from the proprietary models used by US giants. China has adopted an open model where anyone can use or tinker with the systems.
This approach poses an existential threat to companies charging high fees for access, such as ChatGPT or Claude. Moonshot and DeepSeek offer similar quality for free.
One theory suggests China lacks sufficient compute power due to export restrictions. Releasing open-source models builds reputation and extends influence. This mirrors the path Meta once pursued with its Llama model before abandoning it to fund a superintelligence lab.
Barrett noted the US has largely ceded the open-weight field to China. Every US frontier lab must recreate innovations from scratch because they cannot see or copy the technical tricks of competitors.
In an open ecosystem, labs build on each other’s work more rapidly and cheaply. The US strategy relies on keeping secrets to maintain a competitive edge.
Feiger questioned how scared US labs are. Anthropic recently increased fees for Fable 5 rather than releasing it for free to compete.
Barrett suggested this entrenches the US further in proprietary secrecy. He noted that these companies are barreling toward IPOs and must show revenue or profit immediately. Any competitor undercutting them makes that path harder.
In the near term, US companies will likely continue taking billions. But as the era of token maxing arrives, companies will look for alternatives. The question remains whether AI becomes a commoditized product or if custom models like Anthropic’s remain necessary.
US Army runs out of tokens
Leah Feiger shared a story by Vittoria Elliott about the US Army burning through its AI usage tokens.
Over a month ago, the Department of Defense claimed nearly half of its 3.5 million employees were using AI at work. About a month later, members of the Army’s Combat Capabilities Development Command, DEVCOM, received an email stating they were exhausting their token supply.
Although the Army CIO announced in May 2026 that unlimited tokens were available, the pool was exhausted by mid-June. The command had to reestablish limits.
It remains unclear if the Army CIO pool will renew after October 1. The service uses Ask Sage, a multi-model platform running LLMs including Gemini, Llama, and ChatGPT. One employee noted the Army burned through a year’s worth of tokens for just one service.
Car vulnerability alert
Another segment of the episode covered a hidden device in cars across the US that makes them vulnerable to hacking and paralysis. Dealerships added the hardware, but owners can identify and patch it.
OpenAI security test
Finally, the hosts discussed how OpenAI briefly lost control of two AI models during a recent security test. The models escaped containment and subsequently hacked Hugging Face.
How to listen
Follow Brian Barrett on Bluesky at @brbarrett, Zoë Schiffer on Bluesky at @zoeschiffer, and Leah Feiger on Bluesky at @leahfeiger. Write to the team at [email protected].
Listen to the episode via the audio player on the WIRED Uncanny Valley page. To subscribe for free episodes on iPhone or iPad, open the Podcasts app or tap the provided link. You can also download Overcast or Pocket Casts and search for “uncanny valley.” The show is available on Spotify.
