From Silicon Valley to DC, the tech world is suddenly obsessed with one concept in AI: Distillation

Artificial intelligence distillation has moved from a technical niche into a major national security debate. The process involves training smaller, more efficient AI models using the output from massive, advanced models. While developers value it for its cost efficiency and performance, critics now argue it serves as a vehicle for intellectual property theft.

The tension spiked recently after the Chinese lab Moonshot AI released Kimi K3, a model that competes with top-tier U.S. offerings from Anthropic and OpenAI. White House advisor Michael Kratsios publicly accused Moonshot of using industrial-scale distillation to copy Anthropic's Fable model. This accusation highlights a growing concern in Washington that foreign competitors are bypassing the heavy investment required for model development by simply harvesting the work of American companies.

However, a coalition of major tech firms including Nvidia, Microsoft, and Meta issued a joint letter on Friday pushing back against potential government restrictions. These companies argue that distillation is a standard industry practice used for validation and model improvement. They warn that blocking these methods could hurt American competitiveness and drive innovation to other regions.

The debate is further complicated by the fact that many U.S. companies also use distillation to make their own systems more affordable. As the cost of running large models rises, the industry is split. Companies like Anthropic are actively fighting the practice to protect their business and security interests, while others argue that the free flow of technology is necessary for progress. Meanwhile, legal experts note a contradiction in the current stance, pointing out that U.S. firms themselves face lawsuits for training their own models on protected creative works.