Anthropic CEO's China AI Warning Confirmed: US Models Used

Researchers linked to the Chinese military are using outputs from advanced American AI models to train their own systems. This comes from a Reuters review of over 80 Chinese academic papers and patents. Here's the thing: these findings support concerns raised by Anthropic CEO Dario Amodei about China potentially using distillation from powerful US frontier AI models. Chinese military and security institutions are using a technique called "model distillation." This transfers capabilities from US models into smaller systems for local deployment, including defense applications. Model distillation is a common AI technique where a powerful model's output trains a smaller one. This allows the smaller system to gain capabilities without needing massive computing resources. For example, researchers from a PLA unit used OpenAI’s GPT-3.5 to process military source code. They then used these summaries to train a domestic AI model, which could operate within Chinese military networks. Also, researchers at the North University of China used Anthropic’s Claude 3 Haiku for synthetic training data for social media monitoring. Anthropic states it doesn't offer commercial access to Claude in China. They also warn that distilled models might lose the original safety safeguards, potentially moving capabilities outside their control. This matters because it highlights how advanced AI technology can be adapted and used in ways that developers may not intend or control.