On June 21-22, the WOT Global Technology Innovation Conference was successfully concluded in Beijing. The conference attracted 60+ lecturers and experts from the fields of large model, multimodality, AI agent, cloud computing, big data, architecture, etc., and a total of 800+ participants participated in the meeting. As a comprehensive technology conference brand with a history of more than ten years, WOT Global Technology Innovation Conference has always taken the delivery of high-quality content as a criterion to adhere to, and at a time when large models and AIGC have received widespread attention, the conference has always tracked the latest, hottest and highest-quality content and technology trends, and continued to run wild on the large model and AIGC track.
1. Technology makes dreams come true
In his speech at the conference, Yang Wenfei, Vice President of 51CTO and President of the Institute of Digital Transformation and Technological Strength, said: In the digital era, digital technology represented by AI has become a revolutionary force that changes the world, innovation based on digital technology has become the core driving force for the survival of enterprises, and talent is the key factor supporting innovation. The data shows that the shortage of digital professionals in mainland China continues to widen, while the half-life of digital technology is declining rapidly, and the number of skills required for jobs continues to increase. It can be said that the cultivation of digital technology talents has become an important factor in determining the development of the digital industry.
Over the past 19 years, 51CTO has always adhered to the concept of "technology makes dreams come true", and is committed to spreading the value of technological progress to every enterprise and individual, and helping 100 million digital talents grow for life. So far, 51CTO has cultivated more than 8 million digital talents, providing digital talent consulting, training and certification services for thousands of enterprises every year, becoming the main force in the cultivation of digital professionals in China.
2. Enter a new era of intelligent information acquisition
Dou Zhicheng, Deputy Dean of Hillhouse School of Artificial Intelligence, Renmin University of Chinese, shared "Generative Information Retrieval: Entering a New Era of Intelligent Information Acquisition".
Since the 90s of the last century, the information retrieval technology represented by the "search engine" has greatly improved the efficiency of human beings to obtain information from massive Internet data, and has become an indispensable tool in daily life. The recent emergence of generative large language models represented by ChatGPT is expected to profoundly change the way people obtain information, promote the updating and upgrading of information acquisition technology, and may change or replace traditional search engines. In this context, Generative Information Retrieval (GenIR) has received extensive attention from academia and industry.
In the sharing, Dou Zhicheng introduced in detail some of the work carried out by the Hillhouse School of Artificial Intelligence of Chinese University in the field of generative information retrieval, including large models for information retrieval, generative document retrieval, and retrieval enhancement generation, and discussed the future development direction.
3. Use AI to drive growth
Wang Ke, head of the algorithm of Volcano Engine Large Model Service, shared "Volcano Engine Large Model Service, Driving Growth with AI".
Wang Ke pointed out that only a large amount of use can polish a good model. As of April this year, the Doubao model has processed an average of 120 billion tokens of text and generated 30 million images per day. Doubao App, Cat Box, Button, Dreamina, etc. are all incubated products incubated by ByteDance based on the large model of Doubao. These abundant usage scenarios bring a positive cycle effect to the base model, the bean bag model.
Through the practice verification of ByteDance's internal 50+ business scenarios, the daily large usage of hundreds of billions of tokens continues to be polished, and the Doubao large model creates a rich business experience for enterprises with high-quality model effects. The bean bag model family includes: bean bag general model pro, bean bag general model lite, bean bag role play model, bean bag speech synthesis model, bean bag voice replica model, bean bag speech recognition model, bean bag wensheng diagram model, bean bag · Function call model, bean bag vectorization model.
In addition, the one-stop large-scale model service platform of Volcano Ark launched by Volcano Engine not only has a strong system carrying capacity, but also can ensure business stability and cost controllability. At the same time, it is equipped with a full-cycle security and trusted solution to escort enterprise AI applications. At the same time, Volcano Ark provides the same networking plug-in as Douyin Toutiao, the content plug-in of Toutiao Douyin and the RAG knowledge base plug-in, and professional services are provided by the ByteDance algorithm team to help enterprise AI applications easily land.
Fourth, stimulate the new quality productivity of enterprises
Liu Di, Product Director of Tencent Cloud Database, shared "Cloud Native Database Fully Embraces AI and Stimulates New Quality Productivity for Enterprises".
With the deepening of digital transformation, cloud databases are facing new challenges in terms of availability, cost, and operational efficiency under large-scale operation, and databases have evolved into a "software + hardware + AI" trinity construction method characterized by storage and computing separation, resource prediction, abnormal self-healing, and fault prediction.
In terms of software, Tencent Cloud continues to explore and iterate on the self-developed kernel TXSQL, which is 100% compatible with native MySQL, and has developed many matching new features for general scenarios in games, education, finance, industrial manufacturing and other industries.
In terms of hardware, a method of "combining software optimization with new hardware", such as full-link RDMA network, compiler optimization, three-layer decoupling framework, and L2 cache based on persistent memory, is proposed to optimize the critical path performance.
In terms of AI, Tencent Cloud has launched an end-to-end one-stop parameter tuning service based on AI methods such as deep reinforcement learning and expert experience, which can automatically explore and optimize parameters according to scenarios without human intervention, greatly reducing the tuning time. In addition, it has also built real-time multi-scenario and multi-dimensional intelligent anomaly analysis capabilities, which can detect anomalies in seconds, and the exception processing time is greatly reduced by 80% compared with manual work. In terms of serverless architecture, it is iterative to build AI4DB services and form an "AI cloud native" product matrix.
It is worth mentioning that Tencent Cloud Database TDSQL reached a record 814 million tpmC at a unit cost of 1/3 of the same architecture, and topped the TPC-C test list in terms of performance and cost performance.
Through the comprehensive optimization, integration, improvement, and innovation of the trinity of "software + hardware + AI", the TDSQL solution of Tencent Cloud Database also brings a more extreme scenario experience. Taking Luckin Coffee's practice as an example, TDSQL-C has helped Luckin achieve a cloud-native architecture upgrade with zero replacement and transformation costs, which can flexibly cope with 10 times the peak of sales traffic, while saving a lot of resources and operation and maintenance costs.
5. Large models accelerate the integration of artificial intelligence and industrial digitalization
Xin Zhou, general manager of Baidu Intelligent Cloud AI and Large Model Platform, shared "Large Model Accelerates Industrial Intelligent Transformation".
In recent years, artificial intelligence has ushered in a window period for the development of large models, with ultra-large-scale training data, complex deep models, and distributed parallel training creating this rising transformative force. At the same time, the AI native application development model based on large models is also evolving, from prompt word engineering and RAG at the beginning, to Copilot, and then to the explosive Agent.
In this regard, Xin Zhou said that we need to screen the applicability of the model to business scenarios and comprehensively select the appropriate application development model according to the specific situation.
He said in his sharing that Baidu Intelligent Cloud Qianfan large model platform is committed to building a large model service super factory, providing advanced generative AI production and application whole-process development tool chain, the platform has accumulated 120,000+ users, 20,000+ cumulative model fine-tuning, and 420,000+ thousand sail application development numbers.
As an industrial-level AI native application development platform launched by Baidu Intelligent Cloud, Qianfan AppBuilder is committed to connecting the last mile from large models to business. At the same time, based on Baidu's technology accumulation and self-owned business precipitation, Baidu's intelligent cloud Qianfan large model platform has built a wealth of large model capability components, AI capability components, Baidu's own business components, etc.
Sixth, Baichuan Intelligence's AGI road practice
Xie Jian, co-founder of Baichuan Technology, brought a theme sharing on "Baichuan Large Model Technology and Application Practice".
Xie Jian mentioned that between June 2023 and May 2024, Baichuan released 12 large language models. Among them, Huggingface, an open-source model including Baichuan-7B and Baichuan-13B, has been downloaded more than 7 million times. Among the multiple versions, compared with Baichuan 3, Baichuan 4's capabilities have been comprehensively improved, and in the SuperClue benchmark evaluation in April, Baichuan 4 ranked first among domestic large models.
In addition to improving the performance of large models, Baichuan has also made in-depth explorations in the field of agent technology. On the one hand, thanks to professional search skills, it can gradually deepen through directional search, multi-round search, etc., to approach the core answer of the question, and at the same time present the search results in a structured manner to optimize the information layout; On the other hand, it can guide users to clearly express their needs and accurately locate problems.
Looking to the future, Xie Jian believes that several important trends in technology development include the expansion of large parameters/windows, multimodal and real-time natural interaction, inclusiveness, long-range mission planning and completion capabilities, and self-learning and evolution.
7. Concluding remarks
In addition to the sharing at the main venue, the conference also set up 13 parallel sub-venue activities. From Baidu, Alibaba, Tencent, Byte... and other large-scale model technology manufacturers, to Baichuan Intelligence, 4Paradigm, SenseTime... This year's WOT Technology Conference invited many large-scale model companies to display their own practices.
Hot situation of sub-venues and exhibition areas
Although a comprehensive technology conference needs hot technical topics such as large models, multi-modality, and AI agents, it also needs topics that technical people continue to pay attention to, such as R&D efficiency improvement, architecture evolution, big data, technology leadership, and big front-end. At this year's WOT Technology Conference, we invited experts from JD.com, 51Talk, Jobbang, Founder Securities, China Literature Group and other enterprises to continue to share their wonderful experiences.
Tencent Cloud special session
Although the 2-day conference has ended, the pace of continuous tracking of hot technology trends will not stop, so stay tuned for more exciting.