Chinese AI labs are training models on Nvidia's Blackwell despite export controls, says ex-OpenAI engineer
International LawPublic PolicyComputer Science
THE AI ANGLE
Retaining hardware signatures in trained weightsA forensic report showed that Chinese labs trained frontier models on export-restricted Nvidia Blackwell chips. Researcher Will DePue found that hardware-specific pseudorandom initialization leaves detectable traces in published model weights. These findings show that existing chip bans do not prevent restricted hardware use, especially through cloud access and smuggling.
THE TEACHING ANGLE
Weight fingerprinting identifies the specific processor family used during training, but it cannot prove if the compute occurred through illicit imports or overseas cloud providers.Read the original at theregister.com Generate teaching or study materials
More in International Law
- Grok Bot took unapproved actions, ignored orders to stop, and then denied itArs Technica · November 18, 2026
- Security flaw found in Java, world's third-largest programming languageBBC — Technology · November 18, 2026
- What AI regulation could learn from the most hazardous industries in the worldThe Conversation — AI · November 18, 2026
- A third of the global AI supply chain is 'invisible' to governments, study findsPhys.org — Technology · November 18, 2026
- AI models can be tricked into identifying civilians as military targets by telling them it’s a gameTechCrunch — AI · November 18, 2026