Agent Model From Shanghai AI Lab Predicts Global Weather in One Minute

Shanghai Artificial Intelligence Laboratory, together with Fudan University, the Institute of Software of the Chinese Academy of Sciences, East China Normal University and other institutions, has unveiled the Shusheng Xinghe agent model. Built on the AI4AI paradigm, the model has more than 400 million parameters and can forecast global weather for the next week in about one minute. The team also used the agent to complete a MiniOS operating system setup in 20 minutes.

Based on 769 task records and agent operation logs from 56 researchers, the team conducted an empirical analysis to quantify human-machine collaboration in the new research paradigm. The results show that AI assistance was used in 96.5% of tasks, with 33.2% of tasks relying entirely on AI capabilities under fixed resources and quality standards. AI generated 64.6% of technical proposals and handled large volumes of basic work including code development, repetitive experiments and data statistics. However, humans made the final decisions on 85.5% of technical route choices, 93.4% of research goal definitions and 81.9% of acceptance standard settings, validating a division of labor in which AI executes and humans decide.

Following the idea that "science is the next programming", the team explored a controllable recursive self-improvement (RSI) paradigm. The approach keeps human high-level judgment as a core constraint, using a full-process verification and tracing system and giving agents the ability to proactively ask for help when needed. According to the team, the Shusheng Xinghe model ranked first in two and second in two of eight authoritative benchmarks.