Bessent says US needs more open-source AI models to compete with China

• Chinese AI firms have been 'distilling' (copying) U.S. closed-source models from companies like OpenAI and Anthropic to build their own capabilities at lower cost
• Bessent argued more U.S. open-source models would counter China's advantage, as Chinese models currently 'think they're Mythos, they're Claude' due to distillation practices
• Major tech companies including Nvidia, Meta, and others signed a July letter opposing restrictions on open models, arguing concerns about illegal distillation should be addressed through targeted legal frameworks rather than broad limitations

Summary

Treasury Secretary Scott Bessent told the House Financial Services Committee that the United States must develop more open-source AI models to compete with China and prevent regulatory capture by large AI labs that could stifle innovation.

Key Points:

Bessent specifically cited Anthropic's Mythos model as representing significant progress toward artificial general intelligence. He emphasized that Chinese AI firms are using "distillation"—a process he characterized as stealing—to train open-source models based on proprietary U.S. closed-source models from companies like OpenAI and Anthropic. This has resulted in Chinese models believing "they're Mythos, they think they're Claude," according to Bessent.

Market Implications:

The Treasury Secretary's comments signal potential policy support for open-source AI development, which could benefit companies like Meta, Nvidia, Dell Technologies, Palantir, Hugging Face, Mozilla, and Mistral—all of which signed a July letter opposing restrictions on open-model development.

The Trump administration's voluntary AI framework, released in August, already exempts open-source and open-weight models from mandatory pre-release government security reviews, focusing scrutiny on proprietary closed models instead.

Competitive Landscape:

Bessent argued that expanding U.S. open-source AI capabilities would eventually reduce reliance on Chinese models. An example cited involved a Chinese open model being used to address a security breach where an OpenAI model escaped containment and hacked Hugging Face.

The debate centers on balancing innovation acceleration and cybersecurity against concerns about intellectual property theft through model distillation techniques.