字节跳动创始人张一鸣在上月召开的Seed团队全体会议上做出明确表态,即使暂时落后于竞争对手,公司也不会依赖AI蒸馏技术来改进其模型 [1]。AI蒸馏技术是指利用更先进的前沿模型(通常由其他公司开发)的输出来训练AI模型的方法 [1]。
美国AI公司Anthropic曾指控DeepSeek、月之暗面、MiniMax、智谱和阿里巴巴对其Claude模型进行蒸馏 [1]。值得注意的是,字节跳动并未被Anthropic指控 [1]。这一决定主要受到字节跳动与美国政府因TikTok所有权产生的复杂关系影响 [1]。
ByteDance founder Zhang Yiming has issued an internal directive stating that the company will not rely on AI distillation techniques to advance its models, even if this means falling temporarily behind competitors [1]. The statement was made during a Seed team meeting last month, where Zhang emphasized that ByteDance would pursue independent model development rather than leverage outputs from more advanced frontier models developed by other companies [1].
AI distillation refers to the practice of training AI models using the outputs of more advanced frontier models, typically created by other organizations [1]. The decision reflects ByteDance's cautious approach amid complex geopolitical circumstances involving the company and the U.S. government over TikTok's ownership [1]. While several Chinese AI companies, including DeepSeek, Moon, MiniMax, Zhipu, and Alibaba, have faced accusations from American AI firm Anthropic of distilling its Claude model, ByteDance itself has not been named in such allegations [1].