On the evening of April 21, the 4th ATEC Technology Classic (ATEC2023) offline competition came to an end. This year's competition takes the large model as the technical base and focuses on the proposition of "science and technology to help the elderly", which is the first large-scale model full-link application competition based on real scenes in China. ATEC2023 offline competition was in the form of a 48-hour extreme challenge, and four players from Southeast University, Harbin Institute of Technology, and Huazhong University of Science and Technology won the championship and took home a prize of 1 million yuan. The full interpretation of the competition questions and the code display will be broadcast in the first programmer technology program in China, "Burn! Genius Programmer 4".
This year's ATEC Science and Technology Elite Competition (ATEC2023) was hosted by the Chinese Institute of Electronics, organized by the ATEC Frontier Technology Exploration Community, with the participation of 17 universities including Tsinghua University, and the joint proposition of Tsinghua University and Ant Group.
After 3 months of auditions, 16 contestants who stood out from more than 3,000 contestants in 1,901 registered teams within 48 hours, relying on large model technology, in real business scenarios such as Alipay intelligent assistants, to solve a series of practical problems encountered by the elderly in remote medical treatment.
"Large model + close to industrial reality", these two highlights have made the number of applicants for this competition and the proportion of 985 and 211 colleges and universities hit a record high. The core challenge of this competition is to solve the three core problems that need to be solved in the training process and application of large models in industrial real scenarios: usability, security, and high performance.
Zhang Zhiqiang, one of the questioners and technical director of the basic intelligence department of Ant Group, told reporters that the questions were based on these three core questions, testing the players' cracking ideas and abilities, such as letting the model learn to retrieve documents and summarize the content of documents to make accurate answers. Let the model learn to collect user requirements and call tools to solve practical problems, in the model inference stage, master the balance between model effect and inference speed, so that the model can be used for production, and pursue the model effect and iteration efficiency at the same time, but also pursue the online inference speed of the model, so as to make the computing cost controllable.
It is worth mentioning that this competition takes the large model as the technical base and takes the real business scenarios close to the industry as the proposition, reflecting the trend of domestic Internet companies represented by Ant to accelerate the development of AI and AI applications. For example, the medical scenario of the ant large model involved in the proposition of this year's competition ranked fifth in accuracy in PubMedQA, the authoritative global medical evaluation list, with the smallest model parameters, ranking ahead of GPT-4.
"The contestants are facing challenges from real applications, and ATEC is a good training ground for technical youth at a time when large-scale models have shifted to large-scale applications. Professor Shen Chao, chairman of the ATEC2023 Judging Committee and professor of Xi'an Jiaotong University, emphasized, "It is the original intention of the organizing committee of the competition to give meaning to code, let technology benefit society, and pay attention to value propositions such as science and technology helping the elderly." ”
Financial Media Reporter: Luo Hui
Correspondent: Wang Yu
Editor: He Qiao Lu Yanli