U.S. agencies say Chinese AI firms ran millions of routed queries against leading U.S. models (ChatGPT, Claude, Gemini, Grok) to extract behaviors and generate synthetic training data, then used that material to train their own frontier models. The firms allegedly used account‑spreading, proxies, third‑party aggregators and cloud hops to evade detection and terms‑of‑service limits.
— If true, industrial‑scale distillation creates a practical leak vector for frontier AI capabilities that undermines export controls, reshapes AI industrial policy, and raises new surveillance and legal questions about model access and misuse.
Simon Curtis
2026.10.06
57% relevant
While the essay is broader than AI, it identifies frontier technology (AI, clean energy) as arenas where the long contest plays out; that connects to existing threads about how China actively shapes technological ecosystems (including model development and norms) to gain strategic leverage — i.e., technological practice as part of long‑term statecraft.
Farrell Gregory
2026.10.06
78% relevant
The article describes a Chinese ecosystem that uses distillation from American frontier models and a substrate of specialized firms, reinforcing the existing claim that Chinese capability growth relies in part on distilling/deriving from U.S. frontier work.
BeauHD
2026.09.08
100% relevant
NSA/CISA/FBI advisory (PDF) naming tactics and companies DeepSeek and Moonshot AI that distilled multiple U.S. models to train R1/R3 and Kimi models.
← Back to all ideas