Author: Lai Wenxin
Editor: Chen Caixian
Few people pay attention to such a phenomenon:
In the early days of the development of large models, "R&D" and "products" often came from the same group of programmers.
Due to the complexity of the system of large model technology, the number of new generation of large model product managers is scarce, and programmers are often both technology and product. Therefore, programmers' thinking about "what problems can be solved" and "how to solve a certain problem" of AI large models also depict the appearance of the first batch of large models to a certain extent.
In other words, programmers are not only difficult to be eliminated, but play an important role in the progress of large models.
Based on this background, the 4th ATEC Science and Technology Championship (ATEC2023), hosted by the Chinese Institute of Electronics and organized by the ATEC Frontier Technology Exploration Community, also focused on programmers and focused on the "dialogue" between programmers and large-scale model technology, exploring how programmers can use large-scale model technology to solve practical problems in real life, such as "technology for the elderly".
Last week, the content of the ATEC 2023 competition was broadcast in the form of a live-action variety show "Burn! Genius Programmer", which allowed viewers to watch the real work of the young generation of programmers using large models to think about application solutions in a 48-hour live broadcast, which attracted widespread attention on major online platforms.
From the perspective of the content setting of the competition questions, ATEC 2023 is unconventional, and when evaluating programmers' large-scale model solutions, it is not based on the existing performance evaluation lists formulated by the academic community, such as C-Eval, etc., but from the real-world user experience, focusing on the problems of the elderly using Alipay in scenarios such as life payment, medical services, and red envelope socialization, directly challenging programmers.
This is also the first programmer competition in China that focuses on how large models can solve real social problems.
48 hours of extreme large-scale model challenge
On the evening of April 21, the final round of ATEC 2023, the first large-scale model full-link application competition based on real scenarios in China, came to an end.
Through multiple rounds of online and offline competitions and layers of screening by judges, a champion team was finally competed. The team members are Zhou Qingsong, who graduated from Harbin Institute of Technology (Shenzhen) majoring in electronics and communication engineering, Wu Dongdong, who is studying for a master's degree in software engineering at Southeast University, Qiu Chenhao, who is studying for a master's degree in software engineering at Huazhong University of Science and Technology, and Wang Haoyu, who is studying for a master's degree in cyberspace security at Huazhong University of Science and Technology.
The ATEC 2023 offline competition is a "48-hour large-scale model extreme challenge", and the 16 players who entered the offline competition will face the whole record of the live broadcast camera, using nearly 50 A100 card resources provided by the event, and the final winner will receive a prize of 1 million yuan.
As the first choice of domestic programmers and college students, this is the fourth consecutive year that the ATEC Technology Classic has been held.
Different from traditional technical competitions, ATEC aims to test the comprehensive problem-solving ability of contestants and their team members by designing propositions that closely follow social values and simulating a real working environment. This competition system design not only tests the professional skills of the participants, but also exercises their teamwork and on-site adaptability, and provides a practical platform for the cultivation of application-oriented technical talents.
ATEC has always advocated keeping up with the current technology development trends and the actual needs of the industry to reflect the challenges in real industrial scenarios. The technical problems and solutions faced by the contestants in the process of participating in the competition are the technical or product pain points that need to be solved urgently in the industry.
By designing test sites around real-world scenarios and data, ATEC also provides an opportunity for industry to observe and select talent. In the first three editions, the propositions of the offline competition were "Wildlife Protection", "Technology Anti-fraud" and "Technology Assistance".
Through the operation of "Burn! Genius Programmer", the industry's first code competition reality show, the ATEC technology community shows the competition and cooperation, challenges and counterattacks between young technology players during the competition, and truly presents the face of China's young generation of technology practitioners.
48 hours of offline competition site
The just-concluded ATEC 2023 for the first time focused on large-scale model technology, with the theme of "Technology for the Elderly", based on real scenarios and data, using the 100-billion-parameter multi-modal large-scale model GLM jointly developed by the Department of Computer Science and Technology of Tsinghua University and Zhipu Huazhang Company, and adopted the form of full-link application investigation, which put forward higher requirements for the algorithm and engineering implementation capabilities of the contestants.
The competition is hosted by the Chinese Institute of Electronics, organized by the ATEC Frontier Technology Community, co-organized by Tsinghua University, Zhejiang University, Xi'an Jiaotong University, Shanghai Jiao Tong University, and Ant Group, with the participation of 12 universities, including Peking University, Nanjing University, and Nanyang Technological University in Singapore.
As one of the proposition parties of this competition, Tsinghua University participated in the proposition and organization of the online competition, oral defense and offline competition.
Ren Ju, the head of the evaluation team and an associate professor at Tsinghua University, pointed out after the competition: "We hope to encourage technical practitioners and learners to be down-to-earth, pay attention to practical applications, and reject lofty buildings with a way of investigation close to the real industrial scene; every year, we will anchor a proposition with social value to remind all industry peers that technology should benefit the society; and at the same time, with the help of scientific and technological competitions, and even extreme challenges, cultivate the attitude and appearance of young scientific and technological personnel to persevere and forge ahead." ”
The competition attracted many young and promising masters in the industry, and the number of registrations for the event hit a record high, with a total of 1,901 teams and more than 3,000 players registering, and more than 51% of the applicants came from 211 colleges and universities. The contestants come from Tsinghua University, Peking University, University of Science and Technology of China, Huazhong University of Science and Technology, Sun Yat-sen University, Harbin Institute of Technology and other universities, with an average age of only 26 years old.
After a three-month online knockout competition in the four tracks of "knowledge introduction of large models", "tool learning of large models", "AI news detection" and "network security large models", 16 teams finally stood out from the competition of 1,000 people and successfully entered the offline competition.
Tsinghua University and Ant Group jointly launched the proposition of "Technology for the Elderly", starting from three major sections: "Life for the Elderly", "Smart Healthcare" and "Safety Protection".
Specifically, these 16 teams need to explore how to use the natural language interaction provided by the large model based on the real scenario of Alipay's intelligent assistant, taking the life scenarios commonly used by the elderly on Alipay as an example (life payment, medical services, red envelope socialization, etc.), so that the elderly can easily complete the desired operation without learning the cumbersome App operation method.
Under the comprehensive investigation, the contestants need to solve a series of practical problems encountered by the elderly in remote medical treatment through large-scale model technology in the 48-hour extreme live broadcast.
Zhou Qingsong, one of the champion teams, believes that the theme of this year's competition "Technology for the Elderly" is of great practical value, which can lower the learning threshold of smart tools and help the elderly and other people with learning disabilities to handle business with simple conversations with the help of AI large models.
When the elderly take the "Agent" express
Why is this year's ATEC focusing on "Technology for the Elderly"?
Zhang Zhiqiang, the technical director of the basic intelligence department of Ant Group, the author of ATEC, told Leifeng.com that population aging has become an important theme of the current national demographic change and development trend. On the one hand, the rapid development of technology has brought a scientific and technological gap to the silver-haired people, on the other hand, the imbalance in the distribution of population in urban and rural areas has brought about the problem of under-nutrition, and the needs of life, medical care and security have become the three important propositions to meet the needs of the silver-haired people and improve the pension environment.
How to rely on advanced large-scale model technology to realize smart pension for the elderly has naturally become an important challenge for scientific and technological workers.
Based on the theme of "technology for the elderly", the contestants need to solve two major problems in the competition: technology and application.
Player code page at the time of the match
From a technical point of view, the first key point of assessment faced by the contestants is retrieval. The model needs to obtain documents through network search engines or internal retrieval, and then answer medical-related questions based on the documents.
The second key point of assessment is the invocation of tools. The language model can only reply to text, but if the contestants' solutions can call on the tool to complete the actual task, the score will be higher. For example, after entering the natural language requirement "Help me book a second-class ticket from Hangzhou to Shanghai at 10 o'clock tomorrow", the model can directly book the train ticket.
In addition, the detection of rumors and abnormal questions at the technical level are all security issues in the use of large language models, which are also important assessment points of this year's ATEC Technology Elite.
From the perspective of application, whether the language model can provide a better intelligent experience for the elderly is the key to judging whether the contestants' solutions can get high scores. Among them, the intelligent experience needs to closely focus on the needs of the elderly, such as travel, policy consultation, medical consultation, safety capabilities, etc.
In addition, exploring how to reduce the cost of model deployment is a practical problem and technical difficulty faced by large models, and it is also one of the test points for the questioners to design the competition questions and test whether the contestants can fully think and exert their creativity.
"In the past, the model was small and could be deployed independently, but now the model is very expensive to deploy. Privacy-preserving technologies that protect how language models are input or interact with them can potentially save significant resources. We wanted the model to be a power station-like tool, so it would be impossible to have a generator at home. Therefore, it is a very promising technical and engineering problem to put the original model on the computing platform and call or use it in a single computational way. Zhang Zhiqiang explained to the audience during the live broadcast of the event.
The live broadcast of the program explained the scene
Of course, it's not realistic to complete a fully functional Agent development task in less than 48 hours. Therefore, the questioner disassembles the complete link of the agent into several links, and each link provides the corresponding data logic, so that the contestants can solve the tasks of different stages one by one in a limited time.
Zhou Qingsong and his team members won the championship of the "Large Model Extreme Challenge", graduated from Harbin Institute of Technology (Shenzhen) in 2022 with a master's degree in electronics and communication engineering, and is currently working as a senior engineer in a major technology company.
Zhou Qingsong, who specializes in natural language processing (NLP) and data mining, has won the title of "kaggle master" and won the championship in the 2021 ATEC Technology Classic. Two years later, Zhou Qingsong won the championship again at ATEC.
When talking about the purpose of the competition, Zhou Qingsong said frankly that the high prize money was the initial motivation to attract him to participate, and at the same time, the proposition from the real industrial scene also made the "technology competition" more valuable.
What impressed Zhou Qingsong the most was the last question of the offline competition. The topic is based on the medical field, and the contestants need to introduce medical vertical knowledge into the large model, improve the understanding ability of the large model to diagnose and treat related problems as much as possible, and increase the depth and breadth of their medical knowledge.
Specifically, the model needs to be able to provide an accurate diagnosis and give detailed treatment opinions when asked about a range of disease symptoms. The more accurate the diagnosis and the more appropriate the treatment, the higher the score of the contestants.
Zhou Qingsong with his teammates
In Zhou's view, the most challenging part of the whole game was improving the speed of the code. Slowing down your code can significantly increase the time and cost of training your model.
"It's a very extreme race, and we have a limited amount of time to complete a series of tasks, and we rely on the speed of the code that we may not normally care about. Zhou Qingsong told Leifeng.com, "In the time-limited scenario, it is difficult and important to improve the speed of code running, and if it is optimized, it is equivalent to the overall iteration speed will be ahead of other teams." ”
AI First
As the first full-link application competition for large models based on real scenarios in China, the ATEC Classic Competition has focused on large models since this year.
"The large model technology became popular in the first half of last year, and it was already very popular in our offline competition last year, so it was proposed that this year's competition questions must be related to large models, which is a hot spot and a challenge. ”
Zhang Zhiqiang told Leifeng.com, "There are competitions for large models in the industry, but there are fewer competitions for combined applications. After judging each dimension, we believe that natural language interaction must be a very important, future-oriented, and AI product interaction mode, which can make up for some shortcomings of traditional interaction models. The theme of this year's session is to help the elderly with technology, because it is not easy for the elderly to understand the operation logic of mobile phones, and natural language interaction is very simple for them, so the AI products developed based on large language model technology are especially suitable for the elderly. ”
Zhang Zhiqiang, nicknamed "Zero", is the technical core leader of Ant Models, and he and his team are responsible for the research and development of Ant's basic technologies, including language models, knowledge graphs and graph neural networks, which are used in Alipay's face recognition and financial, medical and other products.
At present, the corpus optimization of language models is one of the most important tasks of Zhang Zhiqiang's team, which is responsible for the research and development of 100 billion parameter models and corpus optimization, and the development of large models for the medical industry on the basis of the base of Bailing large language model.
As one of the co-proposers, Ant Group, together with Tsinghua University, decided to use the large model as the technical base of this year's ATEC competition, which also reflects the strength of Ant's "All In AI" in recent years.
In 2023, Ant Group proposed the "AI First" strategy, which is known as the group's three major strategies, along with "Alipay Double Flywheel" and "Accelerating Globalization".
In fact, long before the "AI First" strategy was announced, Ant Group had already made in-depth investments and research in the field of AI. Around September 2022, the company has already clarified the direction of AI research with large models as the core. At the end of 2022, Ant Group officially launched the R&D project of the large model. In November 2023, the Ant Bailing model passed the filing.
In addition to the process of self-developed models, Ant's "AI First" strategy is reflected in its efforts to continuously integrate large models into business scenarios such as life, finance, and healthcare:
Alipay's smart assistant, which is currently being tested in gray and will be gradually opened next month, can provide users with digital life services in the fields of travel, health, and government affairs;
"Financial Butler" Zhi Xiaobao can provide users with high-quality market analysis, position diagnosis, asset allocation and investment education and other professional services;
CodeFuse, an intelligent R&D tool, supports the entire software development life cycle and provides enterprises with full-cycle AI R&D management.
The all-in-one security solution "Ant Tianjian" can provide a one-stop security service solution for large models from detection to defense......
By integrating domain-specific expertise into a common model base, the model can more accurately adapt and optimize the application of multiple vertical industry scenarios. At present, Ant Group's large-scale model technology has shown remarkable application results in many industries such as healthcare, remote sensing, government affairs, and finance, promoting intelligent transformation and efficiency improvement in these fields.
Since 2020, Ant has participated in the ATEC Technology Classic, which has unwittingly become one of the hottest AI competitions in China, building a bridge to connect outstanding technology developers and disseminating the latest AI technology to the public in the form of variety shows.
Write at the end
ATEC 2023 is jointly proposed by Tsinghua University and Ant Group, based on the theme of "Technology for the Elderly", which introduces the scenarios and data of Alipay's intelligent assistant, and the competition questions and database will be opened to many universities across the country in the near future. Tsinghua students will be the first batch to directly learn the practical application of AI large models in industrial scenarios in the course.
It goes without saying that in ATEC 2023, the in-depth cooperation between academia and industry complements each other, with the Chinese Institute of Electronics, the ATEC Frontier Technology Exploration Community, and top academic institutions represented by Tsinghua University providing strong theoretical support and talent training platforms for the competition, and Ant Group providing real industrial scenarios and data, jointly promoting the development of the industry-academia combination of AI large models. This cooperation model not only provides students with practical opportunities, but also provides high-quality talents for technology companies.
As one of the challengers, Ant focuses on the "AI First" strategy, and while attracting AI talents through competitions, it also intensively lays out the application of self-developed models and large models, coupled with the technology and accumulated data assets honed over the years, Ant has gradually built a "moat" in the large model competition, and has unique advantages.
In addition to Ant Group, as the focus of the AI industry shifts from the "100 model wars" to the implementation of AI applications, more and more competitions in the fields of STEM, programming, and AI led or deeply participated in by Chinese technology companies, such as the Huawei Software Elite Challenge, the Huawei Geek Algorithm Elite Competition, the Baidu AI Technology Innovation Competition, the Tencent Advertising Algorithm Competition, and the Alibaba Global Mathematics Competition.
In addition to allowing young people to "enter the battlefield" in advance and train in real cards, real computing power and real scenarios, science and technology companies have opened a competition between enterprises to attract cutting-edge scientific and technological talents with high bonuses and computing resources that are difficult to access in the school.
After all, in the large-scale model competition, talent is the key to conquering the city.
The competitions led by the ATEC Technology Elite are not only a demonstration of resources and technical strength of technology companies, but also an in-depth excavation and training of professionals in the field of AI. Through the competition, companies are able to identify and attract talents who have a deep understanding of large-scale model technology, and at the same time stimulate their innovation potential and promote cooperation with enterprises.
The competition continues, the talent continues, and the competition continues.
The author of this article, anna042023, will continue to pay attention to the development trend of personnel, enterprises, business applications and industries in the field of AI large models.