In the past year, the popularity of artificial intelligence large models has been in the ascendant. Affected by ChatGPT, major domestic companies, including artificial intelligence, Internet manufacturers and chip companies, have launched their own artificial intelligence models. So, compared with GPT-4, how big is the gap between domestic large models?On the evening of December 12, Liu Qingfeng, chairman of iFLYTEK, said when participating in the dialogue program in the live broadcast room of Yu Minhong, the founder of New Oriental Education Group, that the gap was about half a year.
The gap between our cognitive model and the United States is about half a year, and it will not be completely thrown off by him, but there is a difficult process to catch up. Liu Qingfeng said that these gaps are mainly manifested in the complex reasoning multimodal ability of large models and the Xi of small samples.
The main reason for the above gap is not that there are not enough domestic algorithms, but because it takes time to train a model with 1 trillion floating-point parameters.
Liu Qingfeng said that the iFLYTEK Xinghuo large model is based on the domestic computing power platform with 1 trillion parameters, which was only built in October, and in terms of time, it is half a year behind GPT-4. In April next year, iFLYTEK Xinghuo will fully benchmark GPT-4.
Since 2023, based on the popularity of ChatGPT and the great value of large models in business, iFLYTEK has taken cognitive large models as the company's key research projects, and built a complete, first-class, efficient and collaborative R&D team.
On May 6, the iFLYTEK Xinghuo model was officially released, along with the application results of the large model in the four major industries of education, office, automobile, and digital employees
On June 9, iFLYTEK Xinghuo was upgraded to V1Version 5, in addition to achieving a breakthrough in open-ended Q&A, multiple rounds of dialogue, and a further upgrade of mathematical ability, also shows the characteristics of maturity and the ability to iterate quickly
On August 15, iFLYTEK released the iFLYTEK Xinghuo cognitive model V20,** new breakthroughs in capabilities and multi-modal capabilities, continuous improvement in performance, and the release and upgrade of the iFLYTEK Xinghuo cognitive model V20 capability for multiple applications and products.
On October 24, iFLYTEK officially released the Xinghuo cognitive model V30, the seven major capabilities continue to improve, the active interaction and AI personality capabilities have made new breakthroughs, and the comprehensive capabilities have surpassed ChatGPT 35. At the same time, the medical model was released, and the overall performance surpassed GPT4.
A few days ago, the third annual evaluation report on large models released by the China Enterprise Development Research Center of the Xinhua News Agency Research Institute - Artificial Intelligence Large Model Experience Report 3In 0, the iFLYTEK Xinghuo cognitive model won the championship of domestic large models with the highest score of 1775, and won the first place in the three evaluation indicators of basic ability index, IQ index and tool efficiency improvement index. It is followed by SenseTime's consultation model and Zhipu's AI-chatGLM model.
With the blessing of the large model, the performance of iFLYTEK's AI hardware has been continuously improved.
In view of the office attributes of iFLYTEK office notebook and iFLYTEK voice recorder, iFLYTEK has brought two heavyweight functions: "meeting minutes" and "discourse regularization". Among them, the "meeting minutes" can further process the content, summarize, extract and summarize the text to form a concise meeting minutes, and the "discourse regularization" function can streamline and optimize the text content of the recording and transcription, and further improve office efficiency.
As iFLYTEK's products for students and education, and are first empowered by large models, iFLYTEK AI Xi machines T20 and T20Pro fully combine the rigid needs of students and deeply customize and develop functions and applications that meet actual needs.
It not only provides an in-depth understanding of students' essays, but also provides refined correction feedback and personalized guidanceIt can also provide children with real-life dialogue like a speaking teacher, provide a complete oral Xi environment, and realize interactive supplementary learning like a real teacher.
At the same time, the new enlightening and interactive AI English Q&A supplementary learning is added, which can make children's English change from "passive learning" to "active learning" through personalized solutions, inspired and guided questions, and interactive learning Xi. So far, the iFLYTEK AI learning Xi machine has completely completed the perfect transformation from accurately identifying weak links in knowledge, to recommending the Xi path of the nearest development area based on the knowledge graph, and then to human-like dialogue counseling and explanation.
Affected by this, during this year's Double 11, iFLYTEK Xi a variety of smart hardware, including AI learning machines, smart office laptops, smart voice recorders, translators, smart keyboards and mice, etc., won the sales champion of JD &Tmall dual platforms, and the cumulative sales of the whole cycle of Double 11 increased by 126% year-on-year.