The ByteSeed AI4S team is considering a spin-off, with core members Xiaowen Zhi and Quanquan Gu leaving to start an AI pharmaceutical company

According to Dongcha Beating monitoring, ByteDance's AI research department Seed's AI for Science (AI4S) team is discussing a new round of organizational adjustments, even considering a spin-off from ByteDance. One proposed plan under discussion is for the AI4S team to transition from the Machine Learning System team (AML) led by Xiang Liang to be under the leadership of Yang Zhenyuan, but sources close to ByteDance revealed that Yang Zhenyuan is not enthusiastic about the transfer arrangement.

Since Wu Yonghui took over Seed, the Seed Lab's Embodied Intelligence (Seed Robotics), Artificial Intelligence for Science (AI4S), and Responsible AI teams have successively merged into Seed. The reporting line of the AI4S team has undergone several changes. Xue Wen's team, originally part of AML, was merged into the AI for Science team led by Li Hang last year, shifting reporting from Xiang Liang to Li Hang. After Li Hang's retirement, the team once again returned to Xiang Liang's system.

More concerning than the organizational restructuring is the loss of the core technical backbone. UCLA Associate Professor of Computer Science Gu Quanquan, who co-led Seed's large-scale pre-training and expansion direction, and Xiao Wenzhi, the lead of the Protenix project in computational biology, and other core members have successively left to start their own ventures. The entrepreneurial focus is on AI pharmaceuticals, protein design, and drug discovery platforms, and has received multiple rounds of funding from top-tier USD institutions.

The team's most significant achievements include the open-source replication work of AlphaFold 3, Protenix, and surpassing AlphaProteo in PXDesign for protein binder design. PXDesign achieved binding rates ranging from 20% to 73% at 5 out of 6 different protein targets at the nanomolar level. However, AI pharmaceuticals differ from internet businesses; predictive model outcomes must undergo multiple validations such as wet lab experiments, animal experiments, investigational new drug (IND) applications, and business development (BD), facing an extremely long feedback and monetization chain. This pressure is also the underlying driving force pushing scientific teams towards asset generation and independent entrepreneurship.