On May 6, Xunfei officially released Spark, which means: a single spark can start a prairie fire.
After the press conference, the market also gave a very positive response, and the stock price directly rose by the limit on Monday.
Compared with Baidu, when Wenxin Yiyan was released, the stock price fell by 10%, and the results are still very impressive. Of course, a few days after Wen Xinyiyan was released, Baidu's stock price also rose a lot, but it's a pity that it basically fell back after more than a month. It can only be said that the market is changing rapidly, and Feng Shui takes turns.
But in fact, Xunfei’s first-quarter annual report is not good. Revenue has declined by 17.64%, and it has been losing money since the fourth quarter of 2022.
Let’s briefly introduce and comment on Xunfei’s Spark model. At the end of the article, there is a full-text video of the press conference and a link to the experience application, which needs to be picked up by yourself.
Xunfei named its large model " cognitive large model", which means that it emphasizes cognition. In fact, it is generally called a large language model in the industry. LLM, that is, Large Language Model, still has a distinction. It shows that Xunfei pays more attention to the model's understanding of the world, and hopes to know and know the world.
In addition, I have to mention one point, which I appreciate more. The Xunfei press conference mentioned that they have specially developed a complete set of detailed test methods : an evaluation system with 7 main directions and 481 subdivided task types. Although it sounds absurd, this awareness is actually very important. The model needs a large sample data set for training, and it also needs a large sample test, so as to fully understand the capabilities and limitations of the model. Constructing such a quantitative test method is actually not simpler than model training, because the model generation is different every time, and many problems are mixed with subjective knowledge, which brings great difficulty to the evaluation.
However, only when the evaluation of the model effect is done well can the continuous improvement of the model be guided.
Education is also the focus of iFLYTEK's business. At present, it may still focus on government projects, and the volume of consumer products is not very large. Looking at the official store, the price ranges from 4,000 to 8,000, and the purchase volume is about 20,000. Consumer-grade products are mainly learning machines, and this business may have a lot to do with the double reduction policy. The double reduction regulations limit teachers, training institutions, and control people, so what about machines? Are intelligent learning machines within the scope of supervision? Does too much screen time have some impact on eyes and mental health? These may create some risks in the future.
Another piece, office is also one of the key functions of Xunfei model demonstration. However, the sales volume of office books is less than half of that of learning machines. The real work requirements are so high, and the proportion of various meetings that need to be recorded, recorded, transcribed, and minutes may not be very high. On the contrary, these may become part of the functions of mobile phones in the future . With one device, who wants to carry another one?
Personally, I don't think the value of a virtual person is much. Do we need an avatar standing on the big screen, lip-syncing? Besides the form, what else does it bring? What is really important is the model behind it, the ability to answer questions, and the organization and understanding of business. These actually test people.
Building an ecology is what every big company talks about, but the size of the ecology actually depends on the size of the cake. The ecology of iOS and Android comes from its huge cake. The WeChat ecology stems from its large number of users, so WeChat mini-programs can flourish. Therefore, for large companies, making the cake bigger, increasing the number of users, and reaching users conveniently and reasonably. As Zhang Xiaolong said, only by having their own values can they truly achieve a single spark and start a prairie fire.
appendix:
Watch the full conference address: https://www.56.com/u63/v_MTgzNTk0NTQw.html
Application experience address: xinghuo.xfyun.cn
No public:
AI artificial intelligence and big data
Share artificial intelligence algorithms and big data, and gain insight into the future together.