IT Times reporter Sun Yan
On June 27, iFLYTEK released the iFLYTEK Spark large model V4.0 and announced that it surpassed GPT-4 Turbo as a whole and ranked first in 8 mainstream test sets.
Not only competed in hard power in performance, iFLYTEK also showed off its ecological muscles. At the press conference, iFLYTEK drove the car on the stage for the first time, and the three smart cars equipped with the Xinghuo cockpit, Haobo, Wei brand, and Hongqi also occupied the C position in the exhibition hall. As we all know, iFLYTEK started with speech recognition, and now it has been able to realize switch-free dialogue in 74 languages and dialects, and solve the recognition problems in strong interference scenarios such as multiple people speaking at the same time.
The productization of large models has always been the place that best reflects iFLYTEK's landing capabilities. This time, iFLYTEK Xinghuo upgraded the App/Desk and implanted an AI assistant that understands users better in the "personal space". AI assistants will be ubiquitous, and in education, it can be transformed into a Spark intelligent review machine to help teachers correct homework; In learning, it can be transformed into an iFLYTEK AI learning machine to answer questions one-on-one; In medical treatment, it can transform into the iFLYTEK Xiaoyi App to help you understand medical records and physical examination reports, and become a personal AI health assistant......
In the last mile of the implementation of "artificial intelligence +", iFLYTEK released the Xinghuo enterprise intelligent twins platform, and launched intelligent twin cases such as Xinghuo business opportunity assistant and Xinghuo bid evaluation assistant. All these changes announce the arrival of the era of large-scale model personalization!
There is no doubt that the domestic AI large model industry has a strong ecological landing ability, but the whole world is paying attention to the strength of the base of China's large model. From NVIDIA's supply interruption to Open AI's disconnection, the domestic large model has profoundly realized that "whether there is an independent and controllable domestic base capability to support it, which determines how far we can go on this road?" The wave of artificial intelligence is equivalent to the birth of the PC and the Internet, can we become the second pole of global intelligence after the United States? Liu Qingfeng, chairman of iFLYTEK, threw out two grand questions at the opening of the press conference.
What is the strength of China's large model base?
讯飞星火V4.0整体超越GPT-4 Turbo
The seven core functions of iFLYTEK Xinghuo large model V4.0 have been upgraded, comprehensively benchmarking against GPT-4 Turbo, and surpassing in the overall aspects of text generation, language understanding, knowledge question and answer, logical reasoning, and mathematical ability.
Among the 12 mainstream test sets of large models at home and abroad, iFLYTEK Xinghuo ranked first in 8 test sets, surpassing international large models such as GPT-4 Turbo and also ahead of domestic large models.
The image and text recognition capability of iFLYTEK Xinghuo V4.0 has improved, and the application effect in scientific research, finance, medical care, justice, office and other scenarios has been ahead of GPT-4o. In response to the illusion problem of long document knowledge Q&A, iFLYTEK Xinghuo launched the first traceability function in the industry, and the source of the answer can be found.
What is rare is that the Xunfei Xinghuo large model V4.0 is trained based on the country's first domestic Wanka computing power cluster "Feixing No. 1", which also reflects the independent controllability of the Xunfei Xinghuo base.
Since its full opening in September last year, the iFLYTEK Xinghuo App has been downloaded 131 million times in the Android open market, ranking first among the domestic tool general model apps. In this year's 618 promotion, the sales of smart hardware supported by the Spark model increased by more than 70% year-on-year, and the average monthly use exceeded 40 million.
"Bob in the living room. He walked to the kitchen with a cup. He put the ball in the cup and walked to the bedroom with the cup. He turned the cup upside down and walked out into the garden. He left the cup in the garden and walked to the garage. Question: "Where is the ball?" At the scene, iFLYTEK Xinghuo V4.0 answered this more complex spatial reasoning question correctly, and the answer was "the ball is on the floor of the bedroom".
How far is it from AI models to embodied intelligence? This can be seen from the spatial reasoning "IQ" of the large AI model.
Millions of hardware users have an AI personal assistant with one click
AI assistant is considered to be the first C-end large model application scenario that is just needed, but is the current AI large model qualified to be a personal assistant?
The first essence of personal assistant is to "understand you", Liu Qingfeng proposed that AI assistants should be personalized based on user portraits, memory learning based on usage history, and enhanced learning based on personal data.
After the upgrade, the iFLYTEK Xinghuo App has added a new "personal space", where users can upload their work, study, life, health and other information to form an exclusive knowledge base for each person, and then set up a persona to make the large model your "doppelganger".
Liu Cong, Dean of iFLYTEK Research Institute, demonstrated the effect of "personal space" on the spot, when he uploaded a small essay written by his daughter and selected an AI character tag that matched his daughter's style, Xinghuo generated an article with a lively style and cute expressions; When he uploads the product posters, user short videos, and related recordings of the iFLYTEK translator, Xinghuo can also generate product training documents based on these multimodal information, and can also conduct multimodal traceability of the generated information.
What's more meaningful is that the Xunfei Xinghuo model has opened up the ecology of Xunfei's full range of C-end hardware products, and millions of intelligent hardware users will have a "Xinghuo Family Bucket" with one click. For example, the files of the iFLYTEK smart voice recorder and smart office book can be synchronized to Xinghuo's personal space with one click, and the operation linkage is formed through data exchange, when you synchronize the meeting minutes in an office book to Xinghuo's personal space, you can let Xinghuo write official documents, and it will also help you make PPT, generate to-do lists, etc.
74 Dialects "Free Dialogue"
Solve the problem of speech recognition in strong interference scenarios
At the press conference, there was a demonstration that made the audience laugh and feel that the high energy exploded: three people stood side by side, at a very close distance, the three of them read the manuscript without disturbing each other, almost with the same decibels, it was difficult for normal human ears to hear clearly, only to see that Xunfei Xinghuo not only separated the overlapping roles of the three people, but also wrote what everyone said accurately in real time.
The technology behind this scene is based on multi-modal sound recognition, which will be applied to iFLYTEK hearing smart office, smart screen and other conference office products.
Recently, iFLYTEK won the first prize of the National Science and Technology Progress Award for the project "Key Technology and Industrialization of Multilingual Intelligent Speech" as the first completion unit.
At the press conference, iFLYTEK announced the realization of "free dialogue" in 74 language dialects, including 37 languages and 37 dialects. Among them, the recognition effect of 37 languages is ahead of OpenAI whisper-V3, and the recognition effect of 37 dialects has increased by 30% on average.
Based on this, iFLYTEK has released the iFLYTEK simultaneous interpretation system, which integrates software and hardware, which is suitable for conferences, conferences, exhibition halls, tourism and other scenarios, and participants can listen to multilingual AI simultaneous interpretation in real time by wearing headphones.
With more than 12 million downloads, AI "family doctors" have become a rigid demand
Since its full opening to the C-side, the iFLYTEK Xiaoyi App has become an AI just-needed application, with a cumulative download of 12 million and an active recommendation rate of 42%.
In the medical field, the core capabilities of iFLYTEK Xinghuo medical model have comprehensively surpassed GPT-4 Turbo and GPT-4o. After this upgrade, the iFLYTEK Xiaoyi App covers 1,600 common diseases, 2,800 common drugs, and 6,000 common examinations and tests, which can be used before, during, and after medical treatment.
The iFLYTEK Xiaoyi App has also launched a "personal digital health space", which can further analyze the cause of the disease before seeing a doctor based on user information such as electronic medical records, examination reports, and physical examination reports, give personalized judgments on drug contraindications when taking medication, give data changes after joint comparison after examination, and understand the health status of other family members through role switching.
As for the data security issues that users are concerned about, at present, the iFLYTEK Xiaoyi App has passed multiple authoritative certifications for data security and privacy protection.
In the current situation of relatively scarce medical resources, AI health applications such as iFLYTEK Xiaoyi App have effectively alleviated the urgent need for medical services in society, so that everyone can have a "family doctor".
"A large model that can do problems" helps teachers reduce the burden of homework by 90%
In the actual test of the real questions of this year's high school entrance examination, the "results" of iFLYTEK Xinghuo Yushu are all ranked first. In terms of scientific research, iFLYTEK Xinghuo helped the team of Professor Liu Haiyan of the University of Science and Technology of China to increase the success rate of protein design from 0.1% to 20%, and the design time from 6 months to 1 day. It also empowers everyone, helping a 70-year-old man who does not know the law to recover his pension arrears, and helping a hearing-impaired person realize his literary dream······
So, will this "big model who is more capable of doing problems" tutor students? iFLYTEK AI learning machine upgrades the AI 1-to-1 Q&A tutoring function, which can not only carry out multi-modal heuristic explanations, free questions and personalized answers, but also interactive inquiry-based learning, super-anthropomorphic guided companion learning, etc., so that children have an additional "AI auxiliary teacher". The data shows that compared with traditional problem-solving video learning, the learning method of AI Q&A tutoring increases the learning completion rate of children to 90%, and the problem solving rate to 93%.
A printer and a computer screen turned out to be a machine that could help teachers correct homework, 15 student assignments, and half a minute to complete.
Put the homework into the Xinghuo intelligent review machine, it can automatically scan the homework, no limit on the size of the paper, imitate the handwriting of real people, and the homework that is spit out has been corrected.
After grading, it can also analyze each student's mastery, form a learning report, and automatically generate homework explanation courseware. With this AI assistant, teachers no longer have to worry about incomplete assignments, which can now be completed in just 5 minutes after taking 90 minutes to complete. Manual analysis takes 60 minutes, but now it only takes 1 minute to complete; With the help of individualized teaching and scientific analysis, the rate of solving students' mistakes has increased from 50% to 73%.
Xinghuo Enterprise Intelligent Twins creates an exclusive AI assistant for each position
In addition to Chery Automobile, China FAW, Volkswagen, and Jiangqi Group, which are connected to the Xinghuo cockpit mentioned at the beginning, the Xunfei Xinghuo model is creating exclusive AI assistants for more enterprises, covering leading enterprises in many fields such as China Energy Group, PetroChina, China Mobile, Chinese People's Insurance Company, Pacific Insurance, Bank of Communications, Haier Group, and Midea Group.
In these companies, AI may be an engineer who can write code, a customer service who can make phone calls, or it may be responsible for compliance review, bid evaluation, etc. At Bank of Communications, iFlyCode, the code product of Xinghuo Model, covers more than 6,000 R&D personnel, with a code adoption rate of 38%.
How to solve the last mile problem of enterprise large model application? Xinghuo Enterprise Intelligent Twins came into being, covering more than 400 AI atomic capabilities, integrating more than 90 external sources, and opening up more than 100 internal IT systems, so that enterprises can quickly implement intelligent twins in combination with their own business scenarios. The 32 enterprise intelligent twins are distributed in the production domain, science and technology innovation domain, office domain, management domain, etc., which can be plug and play.
In order to give a sample to the enterprise, iFLYTEK has created the Spark Business Opportunity Assistant, the Spark Bid Evaluation Assistant, etc., which can realize the knowledge of business opportunity clues, the quality and efficiency of customer visits, and the intelligent research and judgment of sales management. Through the functions of pre-bid sourcing, intelligent bid evaluation, and bid calibration review, the intelligent bid evaluation results have a human-machine consistency rate of 98%, and a bid abnormality detection rate of more than 80%, which greatly improves the efficiency of enterprise bid evaluation and reduces procurement costs.
Is the iFLYTEK Spark model good or not? Is the ecology strong? This can be seen in the number of developers. Since the release of iFLYTEK Xinghuo V3.5 on January 30 this year, in just five months, the number of Xinghuo developers has increased from 5.98 million to 7.02 million, the number of overseas developers has exceeded 400,000, and the number of large model developers has reached 570,000.
Liu Qingfeng said that only an independent and controllable prosperous ecology can have a great future for China's general artificial intelligence. Facing the new ecology of artificial intelligence in the future, he emphasized that it is necessary to pay attention to the source technology ecology, agent ecology, application ecology and industry ecology, and realize independent controllability and integration of software and hardware, so as to realize the in-depth implementation of large models; It is necessary not only to scientifically and rationally understand the comprehensive gap between China and the United States in large models, but also to have the confidence to quickly catch up, and give a complete set of methods from source technology, to industrial ecology, to application landing, and to build a truly independent and controllable AI industry ecology with long-termism.