Lin Haibin, the former head of large model training at ByteDance, has officially returned to the company's Seed team, where he'll once again oversee the development of distributed training infrastructure. He left in December to join Safe Superintelligence (SSI), the startup founded by former OpenAI chief scientist Ilya Sutskever — making his round trip less than six months.
Lin, a CMU master's graduate who studied under renowned database expert Andy Pavlov, previously led the open-sourcing of the reinforcement learning training framework veRL and the 10,000-GPU distributed training system MegaScale. His swift return signals a shift in the flow of top-tier AI talent between China and the US.
Before joining ByteDance's Applied Machine Learning (AML) team in 2020, Lin worked on the MXNet framework at Amazon AWS. At ByteDance, he spearheaded the development of GPU recommendation training systems and the ByteCCL collective communication library, and tackled the challenge of coordinating tens of thousands of GPUs in parallel to build the MegaScale training system. With Doubao, ByteDance's large language model, handling a high daily volume of API calls, and multimodal and agent research demanding extreme training cluster stability, Lin's familiarity with ByteDance's tech stack directly supports iteration on the underlying distributed training system. Additionally, the open-source reinforcement learning framework he developed, veRL, has been adopted by Alibaba's Tongyi Qianwen and the Shanghai AI Lab.
ByteDance's rehire policy allows former employees to reapply after three months, provided they have no compliance violations; if they've been gone less than a year, they retain their previous rank and salary. Lin's return comes amid recent incentive measures for the core AI team. ByteDance has offered Seed employees the chance to purchase special options dubbed "Doubao stock" at $13 per share. The battle for AI talent has escalated into a contest of options, autonomy, and computing power. Around 70 Seed lab members had previously left for external AI model makers, and ByteDance is trying to reverse the outflow through clever equity design.