Guoyang Zeng
CTO of Model Best
Zeng Guoyang, started to learn programming at the age of 8 (Sam Altman, the CEO of OpenAI, also learnt programming at the age of 8), won the gold medal in the National Youth Informatics Competition in his second year of high school (50 people in the country), and was guaranteed to be admitted to Tsinghua University by the Asia-Pacific Informatics Competition Gold Medal, and then he joined Kuangyi as an internship in his third year of high school. In his freshman year, he won the first prize of Tsinghua University Challenge Cup and the first prize of Capital University Students Challenge Cup. In his sophomore year, he joined the NLP Lab of Tsinghua University and has been engaged in research and development related to big models, and he is a key member of the team of Wudao-Wenyuan Chinese pre-training model. After graduating from the Department of Computer Science, he served as the deputy director of the Language Big Model Acceleration Technology Innovation Centre of Wisdom Source Research Institute, and has rich experience in AI project development and management. In 2021, as a co-initiator, he created the OpenBMB open source community, and is one of the main authors of the model training acceleration and inference acceleration BMTrain and BMInf, and one of the main completers of the two phases of CPM-Ant and CPM-Bee. one of the main completers of big models.
Topic
Faceted Agent Exploration, Embracing Large Models Group Intelligence Emergence
Based on continuous exploration of technological frontiers and years of industry accumulation, we have embraced a dual-engine approach driven by "large models + Agents." Centered around the CPM series basic large models, we have launched innovative achievements driven by large models, including the AI Agent "three-carriage" system, which comprises: the AgentVerse platform driven by large models, the powerful AI Agent application framework XAgent, and the multi-agent collaboration development framework ChatDev. AgentVerse enables collaboration among multiple models and dynamically adjusts the composition of groups, achieving a synergy effect where 1+1>2. Experimental results demonstrate that this framework can effectively deploy multi-agent groups, outperforming single agents, and manifesting cooperative social behaviors.
