Summary of “This is the real AI divide” by Len Khodorkovsky (Washington Post opinion, Aug. 4, 2026)
The author argues that the key divide in AI is not open-weight vs. closed models, but whether a model is built through genuine innovation or through theft (primarily via illicit distillation).
Moonshot AI (a Chinese lab) recently released a high-performing, low-cost model and promised an open-weight version, prompting renewed debate. A group of tech firms defended open-weight models as vital to U.S. innovation. White House science adviser Michael Kratsios and others contend Chinese firms systematically extract and copy American innovations rather than innovate themselves.
Distillation—prompting a powerful model extensively and training a smaller one on the outputs—is common and legitimate when done on one’s own models. It becomes theft when done secretly against competitors’ proprietary systems (e.g., via fake accounts or masked access). Anthropic has alleged Chinese labs (including DeepSeek, Moonshot AI, and MiniMax) used ~24,000 fraudulent accounts for ~16 million interactions to extract Claude’s capabilities. OpenAI has made similar claims about DeepSeek. Kratsios specifically alleged Moonshot distilled Anthropic’s model to create its Kimi K3.
The author distinguishes this from earlier Western training on scraped public data: deliberately reverse-engineering a finished proprietary model is different. Chinese open-weight models allegedly follow a pattern seen in other industries (e.g., Huawei, Sinovel)—acquire technology by any means, undercut prices, and lock in markets. Free release does not change the underlying issue if the model was built on stolen IP.
Stakes are high: frontier models are becoming critical infrastructure. Widespread adoption of Chinese models risks long-term dependency, which Beijing views as a strategic opportunity to shape global AI architecture and reduce reliance on the U.S.
The solution is not to restrict openness. Independently developed open models should be treated as legitimate. Policy should instead focus on transparency about model lineage for high-risk uses, certification standards for critical infrastructure, and accountability for deployers relying on models of unknown origin. Competition should occur on equal legal and ethical terms so that better ideas—not better theft—determine the future of AI.
While the (H1B) Indians producing the AI here in the US probably cant copy the databases used in the AI models. They sure can copy the code. Which can probably fit on a thumb drive. I am sure there are several copies in Mumbai right now.
Cisco filed suit against Huawei in January 2003, alleging Huawei's router/switch code (for its Quidway line) contained Cisco IOS source code — including identical comments, spacing, bugs, and even help-screen text. The case settled confidentially in 2004, with Huawei agreeing to alter its command-line interface, manuals, and source code.Some things just don't change.
I worked in China October 1976 to May 1977. I saw copies of industrial and consumer products back then. Their South Bend metal lathe in the plant's machine shop was a knock-off of the American lathe. Their baby food jars were perfect knock-offs of the Gerber glass jars, metal tops and labels...except the baby on the label was Chinese, not American Caucasian.
NEVER trust ANY code from the CCP. EVER.
I been goofing around with MiniMax H3 and it’s really fun..
Can only do 720p on a local setup but you can upscale and it looks really nice.
15 second clip on a 5090 is not exactly lighting fast but not so slow you start to curse.. and sure beats buying tokens.
You can reference to images or video and it does a good job handling the voices.
https://www.youtube.com/results?search_query=MiniMax+H3
Bfl