Anthropic’s Claude Mythos Reduces Export Controls To Blind Projections

Anthropic has developed Claude Mythos, an AI model trained on AWS Trainium 2 chips, which are not classified as frontier hardware. The article reports that the success of Trainium 2 in producing a capable model suggests that frontier-level AI may not require the most advanced chips.

According to the article, the training of Claude Mythos on Trainium 2 demonstrates that current export controls may be based on flawed assumptions. The US government has primarily focused on restricting Nvidia's high-end chips to prevent rivals from achieving AI breakthroughs. However, Mythos — trained on a readily available chip — undermines the premise that Nvidia chips are the sole gateway to frontier AI capabilities.

The outcome has implications for global AI chip policy and may lead to a reassessment of export restrictions. Anthropic has not commented on the broader policy implications.