U.S.-China dispute centers on model distillation

3 min read

Model distillation, a technique that uses a large "teacher" AI to train a smaller "student" model, is now at the center of a U.S.-China dispute. The method can produce cheaper, task-focused models that run on less powerful hardware and can be tailored for devices, factories, vehicles and private networks. Washington and leading U.S. AI firms have accused Chinese rivals of using distillation to extract capabilities from proprietary models, while researchers note that transferring reasoning traces (step-by-step solutions) can teach smaller models how to solve difficult problems, intensifying debate over consent and capability transfer.

Model distillation can transfer capabilities without creators' consent.