As an enterprise that has been established for 24 years, iFLYTEK has experienced difficulties in the landing of artificial intelligence and gradually walked out of the road of positive cycle. Now it tries to replicate its successful experience in the field of large models, combining large models with existing mature businesses.
Text | Xu Xin
Ed. Zhao Yanqiu
The large model market has entered the stage of exploration and application.
Liu Qingfeng, chairman of iFLYTEK, clearly asserted that in the field of large models, "whoever can land in the application scenarios that have the rigid needs of society can be the first to form a virtuous circle of self-hematopoiesis." Based on this judgment, iFLYTEK has continued to move at the application level.
On June 9th, at the Spark Cognitive Big Model V1.5 Upgrade Conference, in addition to the capability upgrade of the large model, the display of application scenarios was the focus of the conference, and iFLYTEK spent a large amount of space to introduce the application of combining with the large model in the fields of education, medical care, industry, and office. A month ago, as soon as the Spark model was launched, it simultaneously demonstrated potential applications in education, office, automotive, digital employees and other directions.
Judging from the existing trends, the rigid market that iFLYTEK is looking for for large models is closely related to the two paths it has previously explored in the landing of artificial intelligence. The first is to carry artificial intelligence technology into intelligent hardware products that have been verified by the market such as learning and office. The other way is to use artificial intelligence technology to explore and solve the rigid demand problems of medical, industrial, education, finance, automotive and other segmented scenarios, which is the key industry market of iFLYTEK and the place where large factories and various artificial intelligence companies hope to show their strength.
As an enterprise that has been established for 24 years, iFLYTEK has experienced difficulties in the landing of artificial intelligence and gradually walked out of the road of positive cycle. Now it tries to replicate its successful experience in the field of large models, combining large models with existing mature businesses to achieve healthy hematopoiesis.
01
Large models with hardware
iFLYTEK's mature C-end hardware ecosystem is accelerating the combination with the Spark model, and has become a pioneer product for the exploration and landing of the large model.
At the Xinghuo Cognitive Big Model V1.5 Upgrade Conference on June 9, Wu Xiaoru, President of iFLYTEK, introduced that iFLYTEK's learning and office products such as AI learning machine and smart office book have been blessed by Xinghuo Cognitive Big Model and have been favored by users.
At present, all walks of life in the industry have paid attention to the landing of large models to the extent of a bit of "involvement", from investment circles to large model providers are eagerly exploring the application scenarios of large models.
To some extent, this phenomenon of large model "rolling landing" can be considered in the previous difficulties encountered by domestic artificial intelligence products at the landing level. In the past ten years, artificial intelligence companies have always faced application difficulties, on the one hand, they have invested heavily in algorithms and technology research and development, but it is difficult to earn money in artificial intelligence technology and algorithm and software products, and it is difficult to balance between huge investment and revenue. Liu Cong, president of iFLYTEK Research Institute, bluntly said that many artificial intelligence companies earn 1 billion to lose 2 billion.
Domestic software products are not easy to sell, and the paid ecology still needs to be built and cultivated. At the same time, if only technology is provided, it can only play the role of the end in the AI ecosystem, and cannot be the control node that affects the entire ecosystem. An investor who pays attention to the SaaS market has even made a judgment on the domestic SaaS market - in China, sometimes selling software alone cannot solve the actual problems of customers, enterprises cannot position themselves as software companies, and sometimes need to provide software and hardware integrated solutions.
In order to better achieve healthy hematopoiesis, almost all artificial intelligence companies in the past are combining software and hardware to build a more benign enterprise development route. As a veteran artificial intelligence company, iFLYTEK has also experienced a difficult process of exporting artificial intelligence capabilities to customers and users.
In 2016, iFLYTEK began to form the Consumer Business Group, which set a goal of accounting for 50% of the total revenue within three years. With the successive release of intelligent hardware products, the outside world also has questions about whether iFLYTEK can cultivate ToC soil. In 2019, company executives admitted in an interview that there are still many skills to be completed in the process of shifting technology from ToB to ToC.
The industry has observed that after years of exploration, iFLYTEK has basically followed the path of carrying artificial intelligence technology into hardware and building an intelligent hardware ecosystem. According to iFLYTEK's 2021 annual report, smart hardware revenue increased by 42% year-on-year. According to the 2022 annual report, a number of integrated software and hardware solutions, including learning machines, smart office books, smart voice recorders, translators, and iFLYTEK hearing smart office service platforms, equipped with artificial intelligence algorithms, have brought sustainable turnover revenue to iFLYTEK.
"It is actually quite difficult for us to explore a way to make products based on artificial intelligence technology and achieve a positive cycle in operation in industries such as education, healthcare, including consumer-related automobiles." Liu Cong said in a previous interview.
After the arrival of the big model, iFLYTEK's existing intelligent hardware ecology has become an exploration and landing scene for the big model. Through the empowerment of large models, the performance of these hardware products has been greatly improved, and the activation rate has increased greatly.
Taking the AI learning machine as an example, in early May, the T20 series of AI learning machines with the blessing of the large model can realize the correction of Chinese and English compositions, and the iFLYTEK AI learning machine can talk to users in real Chinese and English like a speaking teacher. Wu Xiaoru introduced that the activation of the AI learning machine series empowered by the large model increased by 214% year-on-year. Another iFLYTEK intelligent office series has been blessed by large models, and its activation increased by 176% year-on-year. "In this year's 618, the learning machine products reached the sales volume of last year's whole day in 3 minutes, and the personal products of all categories only took 7 minutes to exceed the complete sales of last year," Wu Xiaoru said.
It is worth mentioning that iFLYTEK hears smart screen products. Digital Intelligence Frontline learned that the smart screen with a market size of 100 billion yuan has become a must for mainstream office IT manufacturers because it can bring large-scale revenue to manufacturers, and different types of players are entering.
IDC's previous report said that display hardware devices are increasingly homogeneous, and with the blessing of 5G, AI, cloud and touch technologies, the value of future intelligence is more than the hardware itself. Therefore, manufacturers need to pay more attention to the interactive experience attributes of products, combine intelligent systems with hardware, and create perfect solutions to stimulate market segment demand. Equipped with a large model, iFLYTEK hears that the smart screen has intelligent capabilities such as automatically generating meeting minutes and work plans, and may form a differentiated ability in the market in the future.
Combining large models with mature hardware products, iFLYTEK is accelerating the exploration of the landing of various consumption scenarios of large models to achieve the self-hematopoietic and positive cycle mentioned by Liu Qingfeng.
02
Industry demand is exploding
Exploring the combination of large models and industry scenarios is the focus of domestic large model providers that they attach importance to and have laid out. Alibaba, Baidu, SenseTime, Yunzhisheng and other companies are exploring corresponding products or solutions based on industry scenarios.
At the beginning of the launch of the Xinghuo model, iFLYTEK proposed the "1+N" architecture to lay out industry fields in addition to general capabilities. Specifically, 1 refers to a general cognitive intelligence large model algorithm development and efficient training base platform, and N refers to a dedicated large model version used in education, medical, human-computer interaction, office and other industries.
"The selected industry is the business that iFLYTEK has taken root," Liu Cong said in an interview, adding that the combination of large models and industries is not simply to transfer the ability API of large models to use, but to specifically design product functions and combine the needs of users. Therefore, iFLYTEK took the lead in trying the direction of clearer value and deeper accumulation in its own industry. Each direction of N is an area that can be monetized before, and from these directions, it can be promoted to other areas.
Taking industrial scenarios as an example, artificial intelligence applications have previously brought efficiency improvement and cost reduction in all aspects of "research, production, supply and marketing services". For example, in the sales process, market analysis based on artificial intelligence technology can provide better and more systematic marketing strategies. In the service link, artificial intelligence technology can efficiently answer user questions, automatically analyze and improve customer satisfaction.
The empowerment of the Spark model in industrial scenarios is mainly in the two links of supply and demand docking and enterprise knowledge management. iFLYTEK's previously launched Antelope Industrial Internet Platform can connect the upstream and downstream demand and supply sides of the supply chain in industrial scenarios. With the blessing of the large model, it can understand the needs of the enterprise expressed by users in natural language and provide solutions.
Liu Cong dictated the demand on the spot, "A home appliance manufacturer has more than 20 large and small equipment in the workshop, and is seeking a set of solutions that can further save energy and reduce consumption." The Spark model identifies this demand and accurately gives the scheme and related resources. For example, the installation of equipment with high energy efficiency, optimization of production processes, intelligent control systems, and the name of a company in Anhui that provides energy management system solutions for industrial enterprises are given, as well as the names of the corresponding experts in the field. In order to answer the needs of users, Xinghuo model can scan online or enterprise-owned product and service documents to solve the problems and requirements provided by users and better complete the docking.
Another application scenario of the Spark model in the industry is that the combination of the large model and the knowledge base of the specific enterprise can form the knowledge brain of the enterprise and help the new employees of the enterprise solve various faults and problems in the production process. A new employee of the asset management department of a leading optical sorter industry encountered a problem that the touch screen of the equipment did not work in the final assembly workshop. He directly asked the company's knowledge brain, "In the process of debugging the whole machine in the final assembly workshop, the problem of the touch screen is not working. What are the causes of touch screen failure? "The knowledge brain alerted him to possible temperature causes and cooling problems. The newbie started from this level and quickly solved the problem.
iFLYTEK has accumulated a lot of application experience in this field, from pre-diagnosis medical guidance, prevention and misdiagnosis in the process of diagnosis and treatment, to post-diagnosis patient rehabilitation management.
At the press conference, the scene of the combination of the Spark model and medical treatment was mainly introduced, and the post-medical diagnosis was introduced. Digital Intelligence Frontline learned that there is still a gap between the mainland's medical resources and developed countries, with 250 million patients per year, but only 1.88 million doctors above the intermediate level. As a result, most patients are out of hospital management services after discharge, and cannot obtain scientific medication and dietary safety guidance.
The post-diagnosis medical management platform based on the large model can formulate a very personalized rehabilitation plan for patients, including focus on focus, medication guidance, rehabilitation exercise, discharge follow-up, health knowledge, patient consultation, etc., and can also help doctors strictly track and implement the closed loop and urge patients to implement the plan.
For example, a stroke patient taking anticoagulant drugs requires an initial focus on coagulation indicators. Based on his case, the large model automatically gave a plan to review relevant indicators three days and seven days after discharge. At the same time, because the patient has high blood pressure, the system also requires the patient to pay close attention to heart rate and blood pressure indicators. Once a review schedule is in place, the system also reminds the patient the day before the scheduled one. Based on the various examinations made by the patient and the uploaded report, the system will provide medical guidance, rehabilitation advice and other information at multiple levels in a timely manner.
At present, iFLYTEK's post-diagnosis management platform has tried to cooperate with some leading hospitals such as Union Medical College, Anhui Provincial Hospital, West China Hospital, Wuhan Tongji Hospital, First Affiliated Hospital of Anhui Medical University, and Shandong Zaozhuang Municipal Hospital. It has improved the management efficiency of doctors by more than 10 times, and the follow-up rate and consultation response rate of patients in the process of recovery have reached 100%, which will greatly alleviate the problem of insufficient medical resources in the future.
03
Build a "spark" ecology together
On the occasion of the announcement of the Spark cognitive large model, Liu Qingfeng made a schedule, and the upgrade conference on June 9 was the first date to be fulfilled.
The latest achievements of the Spark model were launched as scheduled, and they are inseparable from the support of the localization chain. "We have a good technical accumulation, a well-established team, each key module is completely independent research and development, and the software and hardware platforms are run on domestic reliable platforms." Liu Qingfeng said.
In terms of computing power support, Digital Intelligence Frontline learned that iFLYTEK has built a special data center at its headquarters, laying the hardware cornerstone for the construction of a large model training platform.
In addition, in terms of engineering technology, iFLYTEK's computing power platform can also accelerate the inference efficiency of tens of billions of parameter large models by nearly a thousand times, providing the possibility for the economical and large-scale application of larger and more cognitive large model technologies in the future. In order to ensure independent and controllable, iFLYTEK has established in-depth cooperation with representative enterprises with domestic computing power capabilities, such as Huawei, Cambrian, and Dawning.
iFLYTEK introduced that iFLYTEK Xinghuo can quickly tackle key problems in 5 months, and the algorithm support behind it is inseparable from iFLYTEK's continuous investment and accumulation of cognitive intelligence core source technology and systematic innovation for more than ten years, and it has rich experience in Transformer deep neural network algorithms.
The industry is concerned that the evolution of large model capabilities requires abundant and high-quality corpus and data resources. It is understood that after years of accumulation in this field, iFLYTEK has accumulated a wealth of industry corpus under the premise of compliance, providing massive industry text corpus and user feedback data for training industry cognitive large models that reach the level of human experts, and also providing scene guarantee for innovative application research and development and pilot promotion based on large models.
In order to promote the continuous evolution of cognitive intelligence large models and facilitate users in different industries to evaluate the ability of large models when combining cognitive large models with industry "landing applications", iFLYTEK has also cooperated with the industry to build a general cognitive intelligence evaluation system.
This evaluation system was designed by the Key Laboratory of Cognitive Intelligence jointly established by iFLYTEK and USTC, and discussed with the Innovation Alliance of the School of Artificial Intelligence of the Chinese Academy of Sciences and the Yangtze River Delta Artificial Intelligence Industry Chain Alliance. On the one hand, the 48 main tasks given by ChatGPT are benchmarked against the 48 main tasks given by ChatGPT, and combined with the actual needs analysis of 4.08 million developers on the iFLYTEK open platform (as of May 6, 2023), 481 dimensions in 7 major directions, including text generation, language understanding, knowledge question and answer, logical reasoning, mathematical ability, code ability, and multimodal capability, are finally formulated to evaluate the capabilities of general artificial intelligence large models in seven directions.
"On the iFLYTEK artificial intelligence open platform, there are currently more than 4 million developer teams, this time the ecology brought by general artificial intelligence will be more prosperous, and we will work with all ecological partners to promote the arrival of the era of large models." Liu Qingfeng said.
In order to build a more prosperous ecosystem, iFLYTEK also intends to open the capabilities of large models to the developer team. iFLYTEK opens the Xinghuo cognitive big model development interface, fully opens the seven dimensional capabilities and 200+ assistant corresponding capabilities to developers, and supports rapid integration of multi-terminal access, supports private deployment, and builds an ecosystem with the majority of developers.
Liu Qingfeng believes that the prosperity of the general artificial intelligence ecosystem will bring new empowerment from every angle of production and life, and profoundly change today's business model of piling manpower and time. A very important factor affecting the speed of industrial development is talent, and whether it can discover and cultivate leading talents in China's general artificial intelligence era faster and more systematically is a key issue related to the future development of Chinese intelligence.
To this end, iFLYTEK joined hands with the first batch of 22 national key universities such as Peking University, Tsinghua University, and China University of Science and Technology to launch the Spark Model Scenario Innovation Competition and the iFLYTEK University AI Spark Camp to cultivate leading talents in the era of general artificial intelligence for the future of the industry.
On June 9, the day of the press conference was also the 24th anniversary of iFLYTEK, Liu Qingfeng's voice was already a little hoarse at the end of the session, and he emotionally returned to his original intention, "To truly use artificial intelligence to benefit every industry, every family and everyone, this is the reason why we stay up late and work overtime, our voice is hoarse, but our eyes are still shining with passion." ”
To solve the rigid needs of large models, the really useful features are emphasized by him again. "We must have both short-term ambition and long-term perseverance to win in long-distance running." Solving the rigid needs of society and building a better world with artificial intelligence is the basic industrial logic and scientific research laws we should follow. ”