Beijing Academy of AI annual conference

context: BAAI (Beijing Academy of Artificial Intelligence) is a non-profit research laboratory and a domestic leader in AI research. BAAI’s annual conference has been nicknamed the ‘AI Spring Festival Gala’ due to its impressive line-up of top-level AI talent.

The 2023 BAAI (Beijing Academy of Artificial Intelligence) conference was held in Beijing's Zhongguancu, from 9-11 June. According to BAAI, Wu Zhaohui 吴朝晖 MoST (Ministry of Science and Technology) deputy minister and Yu Yingjie 于英杰 Beijing deputy mayor gave speeches at the opening ceremony. 

More than 200 AI scholars and practitioners participated in the conference (some online). A particular focus was on AI risks and ethics, with speakers including AI safety advocates Stuart Russel, Max Tegmark and Geoffrey Hinton. Other guests were OpenAI CEO Sam Altman and researchers from other top-level global AI companies, including Meta, DeepMind, Anthropic, HuggingFace, Midjourney and StabilityAI.

Domestic participants included 

  • Andrew Yao Chi-Chih 姚期智 Tsinghua University Institute for Interdisciplinary Information Sciences dean, known for the famous ‘Yao class’ at Tsinghua University
  • Zheng Nanning 郑南宁 Xi’an Jiaotong University and Chinese Academy of Engineering academician, who recently won China’s highest AI award
  • Zhang Yaqin 张亚勤 former Baidu president

BAAI released the new WuDao 3.0 series of large models at the conference (the WuDao 1.0, released in 2020, was the world's largest model by the number of parameters at the time). The series includes

  • Aquila models for text generation, including a base model, chat model and code generation model
  • Emu model for image generation
  • FlagEval model evaluation system

BAAI has made these models open source, stressing that openness and transparency are important to its philosophy. The Aquila models are available under an Apache Licence 2.0. According to BAAI, this makes them the first open-source large language model with support for commercial licensing and compliance with local data requirements. (CP note: the pre-training data are not open-source and it is not clear which ‘local data requirements’ BAAI refers to. According to the model’s GitHub page, training data include regular internet sources like Wikipedia and code repositories, making it questionable whether these would be in line with proposed regulations for generative AI. Further details on training data will be announced in late June, BAAI says.)