"The large model with 1.3 billion parameters has already run through on mobile phones." Still familiar shirts and jeans, or familiar Hubei accent, in the much-anticipated 2023 annual speech, Xiaomi CEO Lei Jun finally announced Xiaomi's big model strategy.
On the evening of August 14, Lei Jun, founder, chairman and CEO of Xiaomi Group, gave his fourth annual public speech at the China National Convention Center in Beijing. In his speech, Lei Jun shared his key growth experiences and insights in Wuhan University, Jinshan and Xiaomi, and announced Xiaomi's scientific and technological philosophy: choose technical fields with long-term value for human civilization, and adhere to long-term continuous investment.
At the press conference after the speech, Xiaomi released a new generation of folding screen flagship Xiaomi MIX Fold 3, Redmi K60 Extreme Edition two new mobile phones, as well as the tablet Xiaomi Mi Tablet 6 Max 14, Mi Band 8 Pro, and the new bionic quadruped robot CyberDog2.
The press conference did not disclose information about Xiaomi cars, but Lei Jun gave a brief introduction to another product AI large model that attracted countless attention.
Lei Jun emphasized Xiaomi's strategy of fully embracing large models, and announced that Xiaomi's large models have been initially run through on mobile phones, and Xiaoai upgraded AI large models and opened invitation tests.
Xiaomi's large model C-Eval ranks ninth
"The large model with 1.3 billion parameters has already run through on mobile phones." Still familiar shirts and jeans, or familiar Hubei accent, in the much-anticipated 2023 annual speech, Xiaomi CEO Lei Jun finally announced Xiaomi's big model strategy.
Lei Jun sorted out Xiaomi's investment in the field of AI. In 2016, Xiaomi established a visual AI team, which was later upgraded to Xiaomi AI Lab; In 2018, the AI image algorithm team was established, mainly serving mobile phone photography and videography; In 2021, Xiaomi Robotics Lab was established.
After 7 years and 6 expansions, Xiaomi's AI team has a scale of more than 3,000 people. Its AI laboratory research directions include computer vision, acoustics, speech, NLP, knowledge graph, machine learning, large models, multimodal and other fields.
In April this year, Xiaomi formed an AI big model team and began to fully embrace large models. Today's first attempt is to upgrade the intelligent voice assistant "Xiaoai classmate" to a large model version.
In fact, as early as August 11, before Officer Lei's announcement, Xiaomi AI large model MiLM-6B has quietly appeared in the C-Eval, CMMLU and other large model evaluation lists, and has won the ninth place in the C-Eval list and the first place in the CMMLU list.
C-Eval List: C-Eval Global Large Model Comprehensive Exam Test List. A comprehensive examination and evaluation kit for Chinese language models jointly developed by Tsinghua University, Shanghai Jiao Tong University and the University of Edinburgh. The examination covers four major directions: humanities, social sciences, science and engineering, and other majors, and can more comprehensively evaluate the model Chinese processing ability.
In the C-Eval evaluation, the average score of Xiaomi's large model was 60.2, ranking ninth among all large models; the ChatGLM-6B jointly developed by the knowledge engineering and data mining team (KEG) of Tsinghua University and Zhipu AI ranked first with 71.1 points; the Chinese performance of GPT-4 ranked second with 69.7 points; and the large model Qwen-7B produced by Ali ranked tenth with 59.6 points. It can be seen that although the Xiaomi large model ranks ninth, it is not much different from the more top-ranked model.
From the perspective of different subject scores, the MiLM-6B model currently shows high accuracy in most liberal arts subjects, but there is still some room for improvement in subjects involving "abstract thinking" such as "law, mathematics, programming, probability theory, and discrete mathematics".
In the "Social Sciences" test subjects, except for education and geography, the MiLM-6B achieved satisfactory accuracy in all eight other subjects. Among the more than ten subjects in the humanities, MiLM-6B performs well on the basis of "history and law".
However, according to the information given on the GitHub project page, the size of MiLM-6B parameters is only 6.4 billion, which is the lowest training parameter among the top ten models. And this also means that among large models of the same magnitude parameters, the score of the MiLM-6B ranks first.
Adapting to mobile phones, large models are "light" is the key
In the first half of this year, generative AI models are undoubtedly the most concerned track in the global technology industry. As early as March and April, many domestic technology giants have launched large-model products that benchmark ChatGPT, such as Baidu's Wenxin Yiyan, Ali's Tongyi Qianwen, etc. In contrast, Xiaomi's pace is much slower.
Previously, Xiaomi Group President Lu Weibing said that Xiaomi will actively embrace the big model, but "(the big model) will be deeply integrated with the business, but will not develop independently like OpenAI"; In June this year, Wang Bin, director of Xiaomi's AI Lab, revealed that Xiaomi will independently develop a common language large model, but will not launch a product similar to ChatGPT separately. He said that Xiaomi's primary goal is to build a basic model with tens of billions of parameters and apply the large model to existing scenarios such as Xiaoai, the Internet of Things, autonomous driving and robots.
This press conference confirms that these statements are true. Lei Jun said at the meeting that Xiaomi is currently developing some new technologies and products, and will show them to everyone after polishing.
It is enough to see that Xiaomi's AI strategy is very clear. The research and development of large models is not to provide a chatbot that helps people write articles, write code, and find information, but to use the capabilities of large models to empower Xiaomi's original smart products.
In his speech, Lei Jun announced the first product to be added with Xiaomi's large model: Xiaoai. Lei Jun said that 110 million active users around the world use Xiaoai every day, and Xiaoai will have 6 billion parameter computing power and become smarter after accessing the big model.
Lei Jun proposed the concept of "lightweight, local deployment" when talking about the differences between Xiaomi's large model and other large models. As a mobile phone manufacturer, Xiaomi hopes that large models can be deployed locally and "can be used on the mobile phone."
According to Lei Jun, as early as the beginning of the year, Xiaomi was already thinking of a lightweight large model. At present, Xiaomi's self-developed 1.3 billion parameter large model has run through the Snapdragon platform, which can be comparable to the industry's 6 billion parameter cloud large model in some scenarios.
Xiaoyi vs Xiaobu vs Xiaoai
The big model has become a new battlefield for mobile phones
In China, people's daily life has become inseparable from smart phones. But this year's outbreak of large models such as ChatGPT has made major mobile phone manufacturers realize that smartphones can be smarter and help people do more.
Previously, each intelligent voice assistant could only complete simple tasks, and with the "intelligence" of a large model, the voice assistant will become a real assistant, which can communicate naturally with users as they wish, and may even help order takeaway and air tickets.
In this area, Huawei has taken the lead in taking a step. Based on Huawei's Pangu L0 base model, Huawei has built a large amount of scene data and a finely tuned L1 layer dialogue model for end consumer scenarios, and used this model in Xiaoyi, a smart assistant.
With the blessing of the big model, Xiaoyi has become "smarter, more capable, and more intimate". In the demonstration of the Huawei Developer Conference, users asked Xiaoyi to "find a seafood restaurant with a high rating nearby, preferably with a preferential package suitable for four people", Xiaoyi can accurately understand the user's intention, and directly call the Meituan service to find the relevant restaurant; When a user gives Xiaoyi a picture and says to Xiaoyi, "I want to participate in the technical forum and navigate to the address on the map", Xiaoyi can accurately identify the address information in the picture and open the navigation App for navigation.
Also in August, another mobile phone giant OPPO announced a new Xiaobu assistant based on AndesGPT, which will have AI large model capabilities, with stronger dialogue capabilities and content creation capabilities, and can write the content that users need according to the copywriting needs, and AI capabilities such as induction and summarization will also be greatly enhanced.
AndesGPT is a generative large-language model based on hybrid cloud architecture built by OPPO's Andean intelligent cloud team. It once ranked fifth in the overall list of CLUE1.1 in the Chinese language comprehension assessment benchmark, and first in the large-scale knowledge graph question and answer KgCLUE1.0 list. In the C-Eval review list in June this year, AndesGPT ranked fifth.
It can be said that domestic mobile phone manufacturers have "rolled up" in the field of large models. Recently, the technology blogger "Bad Commentary" posted a poll on Weibo: Who has the smartest smart assistant? Among them, Xiaomi got the most votes, and Huawei ranked second. Many people in the comment area said, "The voice assistant has to be Xiao Ai", "Xiao Ai is really reliable", "The one who chooses Xiaobu is outrageous, Xiaobu is a nerd"...
However, people's general impression is likely to change after various voice assistants have blessed large models. According to users who have tried the new version of Xiaoyi, Xiaoyi, who uses the Pangu large model, has an IQ comparable to that of a 7- or 8-year-old child. And how smart the upgraded Xiaoai and Xiaobu are, it will take more user experience to know.
Author|Zhang Gravel
Editor|Gong Zhouzhou
Produced by unnamed science and technology