Data intelligence industry innovation service media
——Focus on digital intelligence and change business
On the evening of April 20, HKUST Xunfei officially announced its 2022 and 2023 first quarter performance reports. According to the financial report data, the company achieved operating income of 18.82 billion yuan in 2022, a year-on-year increase of 2.77%, and a gross profit of 7.684 billion yuan, an increase of 2.0% over the previous year; the net profit attributable to shareholders of listed companies was 561 million yuan. In the first quarter of 2023, HKUST Xunfei achieved an operating income of 2.888 billion yuan. HKUST Xunfei said that the new investment in key research projects in the first quarter affected part of the current profit.
We have noticed that last year, iFLYTEK invested about 800 million yuan in the expansion of cooperation platforms for continuous operation bases such as education and medical care, new product research and development, independent controllability of core technologies, and localization adaptation. HKUST Xunfei said that although the above-mentioned investment has not yet generated income in 2022, which has affected the current operating performance, it has laid a solid foundation for the continued improvement of profitability and the strengthening of the industry leader in the next stage.
Earlier, HKUST Xunfei announced that the "Xunfei Spark" large-scale model achievement conference will be held on May 6. At that time, Liu Qingfeng, Chairman of HKUST Xunfei, and Liu Cong, Dean of HKUST Xunfei Research Institute, will release the cognition of Xunfei Xunhuo. Large-scale models, AI learning machines, Xunfei Hearing, smart office books, smart cockpits, open platforms and other products will also be unveiled at the same venue.
Consolidate the basic business base of the base area, and lead technological innovation with a large-scale cognitive model
At the "2022 and 2023 First Quarter Performance Briefing" held on April 20, Wu Xiaoru, President of HKUST Xunfei, first introduced the performance of HKUST Xunfei last year and looked forward to the company's development this year.
Wu Xiaoru said that although the overall revenue only showed a slight increase, the business in the base area still maintained good development, with a year-on-year increase of 25% in gross profit, and the proportion of gross profit increased to 58% . In 2022, 130 new districts will be added for after-school services in the "double reduction" school County coverage, covering a total of 280 urban and county platforms, 769 new learning machine line-up offline stores, 12 new prefectural high school entrance examinations and 2 provincial and municipal college entrance examinations for educational examination business, and 300 new operating schools for personalized learning manuals , 96 districts and counties have been newly developed for smart medical care. Medical insurance fee control has been verified in many cities. The detection rate of AI is 3-5 times that of the traditional model. 110 new cooperative models of smart cars have been added.
According to the previous explanation of HKUST Xunfei, "base business" refers to the business that can really create long-term value in various industry tracks. Typical base business includes examinations, learning machines, personalized study manuals, smart medical assistants, medical insurance fee control, smart cars etc.
HKUST Xunfei said in its 2022 annual report that through the solid layout in 2022, relying on base business and systematic innovation, in 2023 the company will achieve its goals with higher quality and more calmly on a more solid foundation, and start a high-speed A new stage of quality development.
ChatGPT is called the "iPhone moment" of AI. The generative AI represented by ChatGPT makes it possible for everyone to order computers to solve problems. Naturally, well-known domestic companies with the ability and resources to invest will not miss this development opportunity.
Recently, large-scale models of major manufacturers have sprung up like mushrooms after a spring rain. Since Baidu launched its first shot on March 16, at least Baidu, Alibaba, SenseTime, 360, Momo Zhixing, Zhihu, Kunlun Wanwei and other well-known companies have held press conferences to officially announce the self-developed large language model. , Huawei, ByteDance and other giants also have related models or solutions.
As for innovations in large models, iFLYTEK, which is itself in the first echelon of artificial intelligence, has been drawing wide attention from both inside and outside the industry.
In fact, as early as 2014, HKUST Xunfei launched the "Xunfei Super Brain" project to gradually break through technical difficulties such as machine reading comprehension, logical reasoning, analysis and induction, and knowledge expression, and realize that machines can go from listening and speaking to understanding can think.
In 2022, iFlytek will win 16 international artificial intelligence competitions in the key core technology fields of artificial intelligence, 13 of which involve common sense reading comprehension challenges such as OpenBookQA, QASC, ReClor and other well-known competitions in professional fields. In addition, the Transformer algorithm has been widely used in iFLYTEK's speech recognition, graphic recognition, machine translation and other tasks and has reached the international leading level. It has open sourced Chinese pre-trained language models and related models in 6 categories and more than 40 general-purpose series The average monthly call volume of the library exceeds 10 million.
Source: HKUST Xunfei
The reason why HKUST Xunfei can generate innovative advantages in key technologies of artificial intelligence is not unrelated to its continuous investment in artificial intelligence research and development. We can see from the financial reports for 2021 and 2022 that despite the speed and slowness of revenue growth, its R&D investment has always remained at a certain ratio. The ratio of R&D investment to operating income in 2021 and 2022 is 16.03% and 17.83% respectively.
In addition, it can be seen from the personnel that HKUST Xunfei has always adhered to the true nature of a technology company. In 2022, the number of R&D personnel will reach 9,281, accounting for 61.68% of the total number of employees, and 2.71 times the number of sales. Liu Qingfeng said at the press conference that more than half of the R&D personnel will be related to the innovation brought about by the breakthrough of the cognitive large model.
The competitive advantage and entry point of the large model of HKUST Xunfei lies in the professional field
It is reported that on May 6th, the release of iFLYTEK’s “Spark” cognitive large model not only includes a general large model, but also solutions for education, medical care, office, vehicle and other industries.
In fact, in December last year before the full-scale outbreak of ChatGPT, HKUST Xunfei launched a special research project of "1+N Cognitive Intelligence Large Model Technology and Application", where "1" refers to the research and development of general cognitive intelligent large model algorithms and efficient training program base platform, "N" refers to the application of cognitive intelligent large model technology in education, medical care, human-computer interaction, office, translation and other industries. At the press conference, Wu Xiaoru said that iFLYTEK is confident that it will be at the forefront of technology in the development of domestic cognitive large-scale models, and will take the lead in in-depth applications in education, medical and other industries.
The iResearch research shows that the content output capability of the general-purpose basic large model in the vertical professional field is weak and the credibility of the output results is uncontrollable, which requires vertical field know-how and data tuning. At the same time, large-scale model products also have the opportunity to be embedded in various enterprise service systems for upgrading and transformation, including enterprise-level applications such as knowledge retrieval, data analysis, auxiliary programming, digital employees, interactive hardware, and digital humans.
HKUST Xunfei has accumulated a wealth of industry data in the research and development and promotion of cognitive intelligence systems for many years, which provides a solid foundation for the research and development of cognitive intelligence large models and optimization of professional fields, and has the credibility of the industry. Moreover, Xunfei has continued to invest in leading national special projects and the construction and optimization of localized training and reasoning platforms, building a solid barrier to competition.
It is understood that iFLYTEK has accumulated more than 50TB of industry corpus and active applications with more than 1 billion user interactions per day in years of research and development and promotion of cognitive intelligence systems , especially in the fields of education, medical care, and translation. Under the premise of laws and regulations, it provides a large amount of industry text corpus and user feedback data for the training of large-scale industry cognition models reaching the level of human experts, and also provides professional scenario guarantees for innovative application development and pilot promotion based on large models.
Taking the medical industry as an example, iFLYTEK's "Smart Medical Assistant" is the only intelligent auxiliary diagnosis system in the industry that has passed the national occupational physician qualification test. More than 90%, providing more than 530 million AI-assisted diagnosis and treatment suggestions for grassroots doctors. In the professional field of medical care, which is related to the safety of people's lives, AI "serious nonsense" cannot be tolerated.
Source: HKUST Xunfei
Liu Qingfeng said that from the perspective of the overall situation in the industry, larger-scale large-scale models are approaching the critical point. In the process of promoting cognitive large-scale models, iFLYTEK’s breakthrough point lies in professional models. Firstly, a powerful professional model is formed, and then professional knowledge and corpus are connected to each other. Under a unified framework, professional sub-models are trained together to obtain higher and more credible results, and finally a general large model is formed.
"I think HKUST Xunfei's big cognitive model today is not following GPT, but has its own logic. On the one hand, we will have a process of catching up with OpenAI. Leading the way, we will turn back from the professional field and form our respective advantages in the entire general field. HKUST Xunfei has the opportunity to make an advantage in the industry and lead the industry, and we are full of confidence in this goal." Liu Qingfeng said.
At the press conference, many brokerage analysts mentioned how to apply the capabilities of large models to existing Xunfei products and how to enhance the commercial value of related products. Wu Xiaoru said that this aspect is another focus of Xunfei's efforts. Xunfei will use the cognitive large model to upgrade the existing core products and comprehensively improve the effectiveness of product applications. He introduced that the Xinghuo Cognitive Model has three main aspects to improve Xunfei’s existing enterprise service products, including:
Significantly improved human-computer interaction experience: whether Xunfei is in vehicle-mounted interaction, customer service systems in many industries, smart home and other products, the current semantic understanding errors will be greatly reduced, and the delivery effect will be significantly improved;
Product functions such as content generation and correction guidance have jumped: products including English listening and speaking teaching, AI learning machines, Xunfei Hearing, and office hardware will all benefit from cognitive large-scale model technology;
Personalized services are more precise: mainly include personalized education solutions, chronic disease management, intelligent medical assistants, personalized learning manuals and other precise services, large models will significantly improve the quality of existing products.
Wu Xiaoru continued to say that the release of the Xinghuo Cognitive Model will not only significantly improve product performance, but also greatly reduce the threshold for using existing Xunfei products. For example, customer service and English listening and speaking are bound to multiple content resources. Currently, a large number of customized resources are used. When the cognitive model is fully released, it will not only drive a large number of customers to use Xunfei’s products, but also It can extensively empower millions of partners of the open platform, which will bring new opportunities for reshaping the Internet industry structure.
At present, HKUST Xunfei already has a solid and good developer team foundation, and continues to build an ecology around the developer group. HKUST Xunfei’s 2022 annual report shows that as of the end of the reporting period, Xunfei’s open platform has opened up 559 AI capabilities and scenario solutions, gathering 3.981 million developer teams, with a total of 1.646 million applications and more than 5 million ecological partners.
Liu Qingfeng said that HKUST Xunfei will strictly abide by the requirements of the Cyberspace Administration of China, gradually open the large model capabilities to developers through invitations, and empower them to develop more applications through APIs. For developers, especially industry developers , they will also get the first batch of tickets to lead cognitive large models to industrial applications.
The technical dividends of Xunfei's cognitive large-scale model "1+N" will be gradually realized through the business logic of "platform + track" in the future. Liu Qingfeng said that Xunfei has been mentioning "platform + track" many years ago. "Platform" means that the development team obtains various capabilities of the Spark large model through API calls and other methods, and pays according to the number of calls between the platform and the platform. Value sharing is carried out in various ways such as application cooperation sharing, traffic sharing and subscription payment. "Track" means that large-scale models should empower education, medical care, office, vehicle and other industries, and form a pattern of mutual promotion of dedicated model technology leadership and large-scale commercial applications in these fields.
Text: Lewis / Data Ape
