After leaving for half a year, distributed systems expert Lin Haibin returns to ByteDance, rejoining the ByteSeed team.

ME AI News: According to Dongcha Beating monitoring, Lin Haibin, the former head of training for ByteDance’s large-model team, has officially returned to the Seed team and will continue coordinating the development of the distributed training foundation.

Lin Haibin left in December last year before joining OpenAI, the superintelligence company SSI (Safe Superintelligence) founded by former Chief Scientist Ilya Sutskever. From his departure to his return has taken less than half a year. As an AI expert with a master’s degree from Carnegie Mellon University (CMU) and a background of having studied under the well-known database expert Professor Andy Pavlo, Lin Haibin previously led the open-source reinforcement learning training framework veRL and the Vanka-scale distributed training system MegaScale. Lin Haibin’s rapid return reflects a change in the flow direction of top-tier large-model talent between China and the US.

Before joining ByteDance’s AML (Applied Machine Learning) team in 2020, Lin Haibin had worked on MXNet framework R&D at Amazon AWS. During his time at ByteDance, Lin Haibin led the development of a GPU recommendation training system and the collective communication library ByteCCL, and he also led a team to tackle the challenges of coordinating parallel collaboration across tens of thousands of GPUs to refine the MegaScale training system. At present, the Doubao large model’s average daily call volume remains at a high level, and multimodal and agent research places extremely high requirements on the stability of training clusters. Lin Haibin’s prior familiarity with ByteDance’s entire technical ecosystem enables him to directly support iteration on the underlying distributed training system. In addition, the open-source reinforcement learning training framework veRL, which Lin Haibin led the development of, has been adopted by Alibaba’s Tongyi Qianwen and the Shanghai AI Laboratory.

ByteDance’s policy on returning personnel states that employees who have been away for at least 3 months and have no records of violations may apply to rejoin the company; if they have been away for less than 1 year, their job level and salary remain unchanged. Lin Haibin’s return to ByteDance echoes the recent incentive measures targeting core large-model teams. ByteDance has opened a special option subscription right to Seed department employees with the “Doubao Shares” priced at $13 per share. At present, the large-model talent war has been upgraded into an all-around contest for “options + autonomy + computing power.” Previously, about 70 members from the Seed Laboratory moved to external large-model companies, and ByteDance is trying to reverse talent outflows through specially designed equity arrangements.
(Source: BlockBeats)