- Model distillation is increasingly seen as a national security concern, with Chinese AI firms like Zhipu accused of extracting capabilities from U.S. frontier models.
- Mira Murati's Thinking Machines built its first foundation model using Chinese models such as DeepSeek-V3 and Moonshot AI's Kimi K2.5, highlighting reciprocal innovation.
- Distillation is a common machine learning technique, not new or inherently nefarious, but now politically charged in the context of U.S.-China AI competition.
- U.S. companies like OpenAI and Anthropic tighten API protections to prevent unauthorized extraction, while themselves incorporating Chinese open-weight models.
- Chinese open-weight models (e.g., DeepSeek, Qwen) are now competitive benchmarks and part of global AI development, challenging one-directional innovation narratives.
- Apple's approval to use Alibaba's Qwen models for Apple Intelligence in China shows Chinese models becoming core global infrastructure.
- Policy paradox: U.S. restricts its frontier models while Chinese models remain more open, creating multiple coexisting global AI frontiers.