Recently, the China Multimedia Conference (ChinaMM2023) was held in Kunming, Yunnan. The conference focused on the development of technologies and applications in the multimedia field, and gathered professionals in the multimedia field. Yu Bing, senior vice president of Kuaishou and head of the R&D line, was invited to deliver a keynote speech on "Frontier Progress and Thinking of Smart Media Technology in the AGI Era" at the conference, sharing Kuaishou's technological frontier progress in combination with Kuaishou's innovative practice in the field of multimedia technology and related thinking.
Explore business development strategies in stages and continue to promote technology energy efficiency optimization
Since its establishment in 2011, Kuaishou has developed into a national-level short video live broadcast digital community, maintaining a healthy growth trend in terms of users, content, and business. The vigorous development of Kuaishou's community ecology is inseparable from Kuaishou's long-term huge investment and large-scale self-developed technology to support business development.
Looking back on Kuaishou's past experience, Yu Bing said that business development will go through multiple stages such as entry, growth, and maturity. In different stages of business development, Kuaishou adopts different research and development strategies. The innovation and growth stage pays more attention to continuous investment, exploring new products and new cycles; while the mature stage needs to focus on optimizing efficiency and taking into account experience, which not only ensures the competitiveness of product experience, but also ensures the steady growth of business.
Yu Bing took the Kuaishou audio and video business as an example. At present, this business has entered a mature stage, and the focus of research and development strategies has also shifted to improving efficiency, optimizing costs, and pursuing the improvement of computing power, storage and network usage efficiency under unit cost. Based on a series of core technologies such as self-developed video quality assessment KVQ, video enhancement and repair KEP and KRP, and video compression coding algorithm KVC, Kuaishou forms a data-driven video processing closed loop, which compresses video extremely while taking into account the experience.
In 2022, StreamLake, Kuaishou’s technology-to-B business, launched the first self-developed intelligent video processing chip SL200. The characteristics of density and intelligence are in a leading position in the industry. In the MSU2022 World Encoder Competition held in July this year, SL200 won the first place in 16 of the 24 indicators on the 4K and 1080P track. At this conference, SL200 also won the China Multimedia Enterprise Innovation Technology Award. At present, the SL200 chip has been fully used in Kuaishou's live broadcast and short video services, realizing mature technology to empower the industry through StreamLake.
In the field of digital humans, Kuaishou has also made many leading achievements. Kuaishou has self-developed digital human core technologies such as light field scanning and reconstruction, super-realistic portrait modeling, intelligent binding, motion capture and driving, and physical simulation. "Two major solutions; combined with the unique advantages of the content platform in the field of brand marketing, empowering in art, technology, operation, marketing and other dimensions, created a virtual anchor Guan Xiaofang with millions of fans, Mengniu Group's first virtual employee Naisi, etc. well-known IP. In the future, combined with AI capabilities such as intelligent songwriting, image/video generation, and LLM, Kuaishou digital human technology will continue to be widely used in cultural tourism, education, games, live e-commerce and other fields to help customers create high-quality and low-cost 3D and 2D digital people.
Deeply cultivate the field of AI large models, expand the industry-research ecology to achieve two-way empowerment
With the in-depth integration of multimedia and AI technology, Kuaishou continues to cultivate AI technology, promote the development of product form and user experience, and explore the second curve of short video business. Yu Bing believes that AI large model, as the most important revolutionary technology at present, has entered the explosive period and is expected to open the AGI era, and multi-modal content generation and understanding are its core capabilities.
At present, the multi-modal AI large model is expected to open up new technical perspectives for the entire link of video production, understanding, distribution, and consumption due to its outstanding generation and understanding capabilities in text, code, image, and video. Break through the technical ceiling of traditional audio and video coding, break through the traditional thinking of search, promotion and promotion algorithms based on user behavior, drive video content creation from PGC and UGC to the AIGC era, create AI large-scale model-driven video content creation tools, and stimulate creators. Creative space to produce high-quality video content with high efficiency and low cost.
As short videos and live broadcasts are the most typical multi-modal media, Kuaishou also grasps the platform genes, invests heavily in the field of AI large models, and explores technological breakthroughs in all aspects. According to Yu Bing, at present, Kuaishou's AI large-scale model layout system is divided into three levels: based on the "big infrastructure" with high performance, high concurrency, and high computing power, the Kuaishou multi-modal AI "big model" is constructed, and then Create "big applications" in the fields of search, promotion, content creation, user growth, and R&D efficiency.
For example, in the field of search, promotion and promotion, Kuaishou's search and promotion algorithm has reached the international leading level, and related achievements have won honors such as CIKM Best Paper and SIGIR Best Paper - Honorable Mention in the top international academic conferences in the field of information retrieval and data mining. Currently, Kuaishou breaks through the traditional user-based Behavioral technical ideas, explore deeper model networks, develop recommendation models, and use content generation and understanding to explore new paths for deep user interests.
At the same time, with the support of multi-modal AI large models, AI technology and tools can empower film and television creators in an all-round way, helping them stimulate creativity, improve efficiency and content quality in various stages such as creation, shooting, and post-production. The cycle time can also be greatly improved, and blockbuster films that used to take several years to shoot are expected to be completed in a few months.
The development of technology from budding to mature not only depends on the self-research of enterprises, but also requires the empowerment of talents from scientific research institutions in universities. Previously, Kuaishou has successively established joint scientific research institutions with Tsinghua University, Beijing Research Institute of University of Science and Technology of China, and Renmin University of China, and established scientific research cooperation with top universities and laboratories around the world to jointly explore cutting-edge technologies in the fields of audio, video, multimedia, and AI. Scientific research talents.
"Academia has top-notch technology and excellent scientific research talents, while the industry has real application scenarios, the advantages of massive data and large computing power. The in-depth cooperation and two-way empowerment between the two will multiply their value." Yu Bing said that, on the one hand, technological breakthroughs will be used on a large scale in Internet services, generating huge economic and social benefits; on the other hand, real Internet scenarios, massive data, and powerful computing power can also help scientific research Technology is constantly iterating. In the future, Kuaishou will continue to promote the improvement of the industry-university-research ecology, open scenarios, data and computing power to the academic community, and jointly explore new technologies for smart media in the AGI era, so as to empower industry innovation and development with technology.