On July 12th, DFM-2 model and innovative technology application results conference was successfully held in Suzhou.
(Press conference site)
At the meeting, Gao Shixing, chairman and CEO of, and Yu Kai, co-founder and chief scientist, respectively shared the future strategy of and the dialogue AI innovation technology represented by the DFM-2 model. Partners such as Mercedes-Benz, SAIC-GM-Wuling, Great Wall Motor, United New Energy, PATEO Internet of Vehicles, Unicom Intelligent Network, PCITECH, Hisense, Boss Electric, Haier, Changhong Meiling and other partners were invited to attend to explore the in-depth integration and landing of large model technology and industry scenarios.
(Gao Shixing, Chairman and CEO of)
Gao Shixing said that dares to become the "source" of original new technologies, with the two-wheel drive capability of original innovation and application innovation, focusing on the "cloud + core" strategy, taking conversational AI as the core, combining DFM-2 large model technology with comprehensive full-link technology, and constantly improving the standardization ability of AI software and hardware products and the large-scale customization capability of the DUI platform, quickly meeting the needs of smart cars, smart homes, consumer electronics, as well as finance, rail transit, The complex and personalized needs of customers in digital government and enterprise industry scenarios such as government affairs create a large model of industry language and empower industrial upgrading.
has been deeply engaged in the artificial intelligence technology industry, and its market share in key application fields such as smart cars, smart homes and consumer electronics has been increasing. Since officially entering the field of automotive front-loading in 2019, has accumulated more than 6 million vehicles and empowered 80 mass-produced models; In 2022 alone, nearly 30 million smart home and consumer electronics terminals will be shipped with voice products; In 2022, was approved to build a "National New Generation Artificial Intelligence Open Innovation Platform for Language Computing", covering "speech + language" full-scenario dialogue technology, and is the first approved enterprise in Jiangsu Province. In the future, will deeply cooperate with the upstream and downstream industries of artificial intelligence, work with industry-level partners to build an ecosystem, help build a community with a shared future for AI in China, and promote high-quality economic development.
【Walking with the times: language model + comprehensive full-link technology】
As the core of this conference, Yu Kai introduced the latest application results of language model and conversational AI technology in simple terms.
(Yu Kai, co-founder and chief scientist of)
"Big model + language computing" is the key to the generation of artificial general intelligence (AGI). Language computing in the era of large models has emerged three capabilities, namely scenario understanding, instruction learning, and thought chain reasoning, which has attracted more attention to general artificial intelligence.
Yu Kai introduced that conversational language computing has experienced the era of statistical dialogue, the era of deep learning, and the era of large models. As a technology-based enterprise with source innovation and continuous innovation capabilities, has won a number of international championships in related directions since 2010, is one of the earliest companies in China to start the research and development of a series of conversational artificial intelligence technologies, and carries out collaborative innovation of basic technologies with research institutions such as Shanghai Jiao Tong University Cross-media Language Intelligence Laboratory and Gusu Laboratory. In the process of continuously strengthening the application of the industry, has continuously upgraded the human-machine dialogue system around the dialogue artificial intelligence technology: in 2013, the dialogue workshop was developed to achieve a full-link closed loop, in 2015, the AIOS system realized the cloud integration of software and hardware collaboration, in 2017, it realized large-scale customizable flexible artificial intelligence through the full-link intelligent dialogue customization platform (ie DUI platform), and so far through the DFM-2 language large model to achieve flexible customization of general artificial intelligence, which can carry out large-scale, High-quality, personalized artificial intelligence system customization not only meets the personalized needs of customers, but also can greatly improve the "non-standard delivery" efficiency of standardized software and hardware products.
【DFM-2 Language Large Model】
At the scene, Yu Kai officially released the self-developed dialogue language large model DFM-2 (Dialogue Foundation Model), which Chinese literally translated as "General Dialogue Foundation Model", Chinese taken the initial homonym "Dongfeng", which not only uses China's strategic military force code name to pay tribute to the mainland's core strategic science and technology self-reliant and independent exploration spirit, but also implies the great prospects of the mainland's AI industry, and also shows that the east wind of the large model will be used to help thousands of industries achieve "communication of all things, Take care of everything".
In 2021, the R&D team released UniDU (DFM-0), a unified generative dialogue understanding framework; In 2022, the team unified all the tasks of understanding, generation, and characterization, and developed DFM-1, a unified generative universal dialogue basic model, as a 1 billion-level large model, and carried out small-scale product applications; Today, the R&D team has formed DFM-2, a large-scale industrial language computing model of tens of billions of dollars, through large-scale upgrading and industrialized customization. In some international tests of public datasets, DFM-2 performance is clearly ahead in a ten-billion-scale model of the same size. DFM-2 is positioned as a "big model of industry language computing", combined with comprehensive full-link conversational AI technology, and has more technical landing capabilities for industrial applications in the vertical vertical vertical domain.
【Rising Wind: Industry Language Big Model with General Intelligence】
Practice has proved that the main contradiction in the current application of artificial intelligence industry is still the contradiction between the general technology of technology manufacturers and the highly flexible personalized needs of B-end enterprises. At present, general-purpose language large models in industry applications often face problems such as difficult inclusion and inaccuracy of private domain knowledge, insufficient accuracy and timeliness of discriminative tasks, data security and computational credibility, and high cost and slow speed. It is difficult to have a universal large language model that can solve all these problems.
Industry language computing big models have more advantages in solving these problems, and generally have three basic capabilities, namely general intelligence and knowledge, the ability to solve industry challenges, and the linkage between large models and full-link comprehensive dialogue technology capabilities to ensure a better user experience.
【General Intelligence and Knowledge】
General intelligence and knowledge capabilities mainly include situation understanding, instruction learning, thinking chain reasoning, common sense question and answer, content generation, mathematical calculation, code generation, etc. DFM-2 ranks first in the general ability evaluation such as C-EVAL, CMMLU, MMLU, AGIEval, Gaokao, etc., as well as in the DialogZoo dialogue comprehension and generation task standard evaluation.
At the scene, demonstrated DFM-2's general intelligence capabilities such as general knowledge quiz, content generation, mathematical calculation and code generation.
(Screenshot of the page of the generic ability)
【Industry Challenge Capability】
focuses on scenario-based landing and has accumulated rich market experience. Taking the Internet of Things field as an example, smart terminals have the characteristics of different computing power, scattered requirements, complex scenarios, many customization requirements, and high requirements for complete interactive experience, which determines that it is difficult for general artificial intelligence technology to be implemented in complex and changeable intelligent terminals and application scenarios. continues to develop and improve its large-scale customization capabilities and product standardization capabilities based on general artificial intelligence technology, supports customers to independently build personalized voice interaction solutions with "thousands of people and thousands of faces", and realizes the "flexible batch manufacturing" of human-machine oral dialogue systems.
The ability to actively respond to industry landing challenges is the key to promoting industrial applications, which mainly include specialized field migration, personalized knowledge and skill customization, miniaturized low-cost deployment, full-link low-latency interaction, continuous update and evolution of private models, and multi-modal perception technology linkage.
(Screenshot of the page of the relevant ability)
At this stage, in response to the challenges of industry implementation, DFM-2 has improved five core capabilities of human-computer language interaction: accurate reasoning decisions enhanced by external sources can provide users with more complex, more accurate and more real-time information services; General semantic understanding based on deep cognition can effectively improve the oral comprehension ability and interaction accuracy of products. Credible active knowledge Q&A based on document understanding, which can be analyzed based on DocDFM to provide users with after-sales customer service, legal consultation and other services; Personalized multi-person interaction for users, support single-device multi-role, flexible switching, and differentiated services; In the face of automatic planning and execution of complex tasks, the tasks performed by the robot can be disassembled and planned through the large model, and the control code of each step can be generated so that the robot can complete the corresponding task.
【Large model technology linkage】
The linkage between large models and conversational language technology can bring users a better experience. The original technology of full-link dialogue system continues to innovate, laying a solid foundation for technology linkage.
(Yu Kai Digital Man)
At the scene, Yu Kai demonstrated the digital human synthesized based on his personal image, demonstrated the digital human generation technology in natural scenes, and realized high-quality digital human generation through small data model training. At the same time, Yu Kai also demonstrated the Cantonese and English synthesized sounds trained based on his speech data, and the personalized dialogue generation and speech synthesis models are seamlessly combined to build a freely interactive personalized digital image.
In the field of speech synthesis, supports zero-sample high-quality speech editing based on generative diffusion models. At the same time, high-quality speech synthesis transfer technology realizes the effect of highly realistic sound reproduction and a single thousand tones.
【Growth of all things: aiming at the B-end market and empowering industrial applications】
Yu Kai said: "Basic AI technology innovation must eventually enter the industry, combined with scenario applications to become valuable, and large models will become important enablers of industrial digitalization." Big Model + Language Computing will be used as a new generation of conversational user interfaces (DUI) to empower industrial applications." In order to further meet the personalized needs of the market, combined the DUI platform with the DFM-2 model to launch DUI 2.0, complete the upgrade of conversational AI full-link technology, and promote in-depth industrial applications.
At the scene, Yu Kai demonstrated the application of DUI 2.0 based on DFM-2 upgrade in various fields. In the field of intelligent vehicles,'s automotive voice assistant Tianqin system has been fully upgraded to 6.0, supporting multi-modal, multi-intent, multi-sound, and multi-round continuous dialogue in all scenarios. In the field of smart home, the product solution adds functions such as intelligent butler service with logical reasoning capabilities, complex intention dialogue interaction, reading comprehension, and copy generation. In the field of government services, the product solution adds multi-round Q&A and intelligent assistant capabilities based on government document reading comprehension, supporting functions such as policy consultation, business handling, document generation, and business circulation. In the field of medical and health care, new professional field personalities, intelligent consultation, and high-emotion speech synthesis have been added, which can make online consultation more empathetic; In the field of scientific research, program supports literature knowledge Q&A, abstract generation, field literature review, material property prediction and other functions.
In the industrial application introduction session, Yu Kai grandly introduced's self-developed conference office software and hardware products, and Wang Yanlong, product director, made a detailed introduction on site. Among them, the software product "Mai Er Huiji", with the blessing of large model technology, has been fully upgraded, adding functions such as AI summary, A to-do, text editing, and one-click drafting, bringing users a smarter office experience. In terms of hardware, Wang Yanlong focused on different product matrices for large, medium, small and micro, among which, AI transcription microphone speaker M6, the first application of AI directional sound pickup and AI two-way noise reduction functions in similar products, can turn on AI transcription with one click; The C60 AI tracking binocular voice camera has a variety of AI tracking modes and can be controlled by AI voice to meet the needs of different conference scenarios.
Yu Kai said that in the future, will continue to deepen application scenarios, solve industry problems, promote industrial development, and combine with partners' proprietary scenarios based on DFM-2 model capabilities and related technical achievements to create a dedicated model with more industry characteristics and empower thousands of industries.
【Dream Galaxy: Join hands with partners to create the future】
At this press conference, many heavyweight partners of attended the event, and customer representatives from PCITECH, SAIC-GM-Wuling, United New Energy and BOSS Electric brought wonderful theme sharing to everyone.
(Qin Wei, Vice President of PCIDO Technology Group Co., Ltd. - Academia Sinica)
(He Yibo, Deputy General Manager of Technical Center, SAIC-GM-Wuling Motors Co., Ltd.)
(United New Energy Automobile Co., Ltd. - Chief Software Development Engineer - Cai Yong)
(Hangzhou Boss Electric Co., Ltd. - Senior Vice President - Zhou Haixin)
At the scene, held a signing ceremony with Mercedes-Benz, SAIC-GM-Wuling, Great Wall Motor, United New Energy, PATEO Internet of Vehicles, Unicom Intelligent Network, PCITECH, Hisense, Boss Electric, Haier, Changhong Meiling and other partners. In the future, will give full play to their respective advantages with partners, explore new cooperation directions and contents around the DFM-2 model, and continuously improve the market competitiveness and brand influence of both products.
Recently, General Secretary Xi Jinping visited the Suzhou Industrial Park and put forward ardent expectations for the high-tech park to promote high-quality development. Gao Shixing expressed three expectations for the future, as a returnee entrepreneur and a technology enterprise rooted in Suzhou Industrial Park, first, has always adhered to the original intention of serving the country by the technology industry, and this original intention has never changed. Second, will continue to adhere to the source of technological innovation and promote industrial application. Third, the self-improvement of science and technology enterprises is very important to promote industrial development and enhance the mainland's international influence. Science and technology enterprises should combine their destiny with the national economy and national destiny, keep in mind the general secretary's instructions, empower the real economy, and promote industry development, technological progress and national progress.
In the future, will give full play to the advantages of platform technology and language large model, under the guidance of policies, build a "new generation of artificial intelligence open innovation platform for language computing countries", improve the development of innovative technology and the transformation of scientific research achievements, help improve the overall competitiveness of the industry, and strive to grow into a supporting force for expanding innovation clusters.
——END——
Welcome to pay attention to [Chinese Business Tao Strategy], know the personalities, and read the legend of Tao Strategy.