ByteDance founder Zhang Yiming bans Seed AI team from distilling rival AI models even if it means lagging domestic competitors
ByteDance founder Zhang Yiming told its Seed AI team to stop distilling rival AI models, even if it means lagging domestic rivals short-term, per Semafor. The rule covers external distillation only. It reportedly dates to 2023, predating 2026 US-China AI export tensions.
ByteDance founder Zhang Yiming has told the company's Seed AI team to stop using distillation, the practice of training a model on the outputs of a stronger rival model, as a shortcut to improving capability. The instruction holds even if it means ByteDance's models lag domestic competitors in the short term, according to reporting from Semafor and Technology.org. The rule applies specifically to the Seed team's external work rather than the entire company, and ByteDance can still distill from its own previous Seed models internally. Notably, the no-distillation policy reportedly dates back to 2023, predating the period when avoiding US retaliation became a potential commercial priority for Chinese AI firms. That timeline undercuts the idea that the policy is primarily a defensive response to 2026 US-China AI export tensions, suggesting instead a longer-standing engineering bet that original pre-training at scale will outperform distillation shortcuts. Distillation has become one of the central flashpoints in Sino-American AI competition this year, with US labs accusing Chinese firms of training on their models' outputs to close the capability gap without matching the underlying research investment. ByteDance's stance sets it apart from domestic rivals willing to use the technique, at the cost of near-term competitiveness on leaderboards.