[ByteDance's Seed AI4S Team Reportedly Considering Independent Spin-Off, Core Members Xiao Wenzhi and Gu Quanquan Depart to Found AI Pharma Company] According to monitoring by Dongcha Beating, the AI for Science (AI4S) team under ByteDance's AI research division Seed is discussing a new round of organizational adjustments, potentially even considering a spin-off from ByteDance. One proposal currently under discussion involves transferring the AI4S team from the Machine Learning Systems (AML) team led by Xiang Liang to Yang Zhenyuan's division. However, the plan is still under deliberation, and sources close to ByteDance reveal that Yang Zhenyuan is not particularly enthusiastic about the transfer arrangement. Since Wu Yonghui took over Seed, three major teams under AI Lab—Embodied Intelligence (Seed Robotics), Artificial Intelligence for Scientific Computing (AI4S), and Responsible AI—have successively been integrated into Seed. The reporting structure of the AI4S team has undergone multiple changes. Xiao Wenzhi's team, originally part of AML, was merged into the AI for Science team led by Li Hang last year, with Xiao Wenzhi's reporting line shifting from Xiang Liang to Li Hang. After Li Hang's retirement, the team returned to Xiang Liang's division. What is drawing more attention than the organizational restructuring is the loss of core technical talent. Several key members, including Gu Quanquan, a UCLA computer science associate professor who co-led Seed's large model pretraining and scaling efforts, and Xiao Wenzhi, the computational biology lead for the Protenix project, have left to start their own ventures. Their entrepreneurial focus is on AI-driven drug discovery, protein design, and drug discovery platforms, and they have already secured multiple rounds of funding from leading dollar-based institutions. The team's most notable achievements include the open-source replication of AlphaFold 3 through the Protenix project and surpassing AlphaProteo in protein binder design with PXDesign. PXDesign achieved a nanomolar-level binding hit rate of 20% to 73% across five out of six different protein targets. However, the logic of AI-driven drug discovery differs significantly from that of internet businesses. Model predictions must undergo multiple layers of validation, including wet lab experiments, animal testing, Investigational New Drug (IND) applications, and business development (BD), resulting in a lengthy feedback and monetization cycle. This has been a deep-seated driver pushing scientific teams toward asset generation and independent entrepreneurship. [Original Link]