With companies such as Apple and Xiaomi developing and upgrading their own AI big model technologies, AI big model technology has become a key trend in the field of intelligent hardware.
Recently, Li Ying, Vice President of Baidu Group and CEO of Xiaodu Technology, stated in an interview with Southern Metropolis Daily that Xiaodu Technology, which strategically positions itself as "AI+hardware", has released the AI native operating system DuerOS X based on the Wenxin big model and has implemented applications in multiple products on screen devices.
AI big models rush to land on intelligent hardware. At the end of July, Apple released the first iPhone AI version of Apple Intelligence. The new software is currently only released in the developer beta version of iOS 18.1, with features mainly focused on writing tools, Siri, photo albums, and other aspects. This update has not yet integrated ChatGPT functionality, and Apple stated that this feature and more updates will be officially launched next year. On the same day, Xiaomi officially announced that the "Big Model Xiao Ai" application for Xiao Ai students is undergoing a full upgrade, covering core categories such as mobile phones, tablets, TVs, speakers, and cars. It supports functions such as natural Q&A, image editing, and external wake-up defense.
As early as April 16th this year, Xiaodu released its first AI native operating system DuerOS X based on the Wenxin big model. Currently, it has achieved big model "brain swapping" in products such as the Tiantian AI tablet robot, Xiaodu learning machine Z30, Tiantian friend machine, and Xiaodu intelligent screen X9 Pro.
The AI native operating system DuerOS X is equivalent to giving Xiaodu a new 'brain', which not only understands users' voice more accurately, but also understands their gestures and facial expressions input. "Xiaodu Technology CEO Li Ying recently said in an interview with Nandu reporters that as early as the beginning of this year, all of Xiaodu's screen devices had upgraded their voice chat function with big model technology. After the upgrade, Xiaodu's multi round conversation ability was greatly enhanced. According to the latest data in July, the number of chat interactions increased by 7 times. Li Ying stated that Xiaodu's strategic positioning is "AI+hardware", and it will use AI to redefine smart speakers/screens, smart education products, and deeply cultivate applications in industries such as smart elderly care and smart hotels.
Children are the group that chats with Xiaodu the most
Southern Metropolis Daily (hereinafter referred to as "Southern Metropolis Daily"): What are the significant improvements in user experience for Xiaodu after switching to the new "brain", the DuerOS X operating system?
Li Ying, Vice President of Baidu Group and CEO of Xiaodu Technology (hereinafter referred to as "Li Ying"): Taking the Tiantian AI tablet robot as an example, it is Xiaodu's first intelligent hardware product equipped with DuerOS X based on the Wenxin big model. The Tiantian AI tablet robot has shown significant intelligence improvement in multi round and complex intelligent voice interaction. Different from traditional smart speakers, it can only receive a single voice command to play, pause, adjust volume, and so on. Now it is possible to understand users' colloquial expressions, and at the same time, its ability to understand context is better. Xiaodu will also actively initiate questioning, achieving multiple rounds of intelligent interaction.
From the backend perspective, one of the groups that chat with Xiaodu the most every day is children. For example, they may ask: What if my classmates in kindergarten don't play with me? Xiaodu would suggest that he could share snacks and toys with his classmates more, giving everyone some time to get to know each other. If children continue to ask about this situation, should they tell their parents? Xiaodu will also calm his emotions and suggest that if he can feel more comfortable, he can communicate with his parents. After applying the large model, the number of chat conversations in Xiaodu significantly increased.
In addition, Xiaodu can flexibly call multiple agents, also known as intelligent agents, to complete complex tasks. There is a vivid metaphor that Xiaodu is not serving you alone, but a group of "people" standing behind you. Users can invite any expert they need, such as legal experts, travel experts, health consultants, astrologers, and so on.
In addition, we have also planned the capability of "holographic field of view". It can accompany you to play with rock paper scissors, give you dressing advice based on the weather, conduct food health analysis, and so on. For example, by taking a photo of a medicine bottle, it can tell you the effects and instructions of the medicine. Large models enable our 'brain' to simulate human like emotions, empathize with users, develop long-term shared memories, and truly provide a sense of companionship. These will be gradually applied to Xiaodu's entire range of products as technology iterates.
Southern Metropolis: How to solve the problem of cost increase caused by increased interaction volume?
Li Ying: Xiaodu uses the MOE architecture, which is a model routing architecture that fully considers performance, cost, and speed. We have the smallest model tiny, and even complex problems will ask the strongest model ERNIE4, with characters, speed, and so on in between. The benefits of model routing architecture are that firstly, I can choose a model that is more suitable for my different needs, and secondly, it solves the cost problem.
Of course, there will definitely be a certain increase in cost, and we will take it into account in the hardware. We adopt the AI+hardware+scenario mode, configuring different model capabilities on different hardware, so some new devices have the highest overall configuration and richer model capabilities. The upgraded part of the original equipment is directly provided to users, which is equivalent to releasing the capabilities of the large model more universally to users.
Southern Metropolis: After the DuerOS X launch event, Xiaodu has successively launched the Tiantian AI tablet robot and the new Xiaodu learning machine. Why prioritize the implementation of education?
Li Ying: The application of big models in the education industry has received great attention, which is also one of the most valuable landing scenarios for big models in our opinion.
Traditional learning machine products generally gather various learning content, resources, and tools as much as possible according to the disciplinary dimension. With the addition of AI technology, many high-end learning machine products have almost implemented various AI functions such as precision learning, essay guidance, oral practice, interactive reading, etc., covering everything from problem solving, homework modification to quality education. But what if the child still doesn't want to learn or can't learn?
Our approach is to redefine the "AI teacher" function, based on a large model enhanced by Wenxin knowledge, where AI teachers simultaneously master both general and professional knowledge; At the same time, in the preview scene of ancient poetry, we will generate immersive images through large models, combined with the vivid and emotional explanations of ancient people, to create a "immersive" learning scene, which invisibly enhances children's memory and understanding of the text. And during the video class playback, children can interrupt and ask questions at any time, and AI teachers can answer them at any time.
For example, in post class practice scenarios, AI teachers do not directly provide answers to incorrect questions. Instead, they gradually inspire, correct, and encourage children to learn problem-solving methods and techniques on their own.
Let me share some data with you: According to Q2 2024 data statistics, the cumulative number of users of Xiaodu Learning Machine exceeds 2 million, with more than half of them being daily active users. These users spend more than 100 minutes a day using it, and the daily usage rate of Xiaodu AI learning function is as high as 97%.
DuerOS X product architecture.
Next, we will focus on the layout of hotels, elderly care, and whole house intelligence
Nandu: What are the current applications of DeurOS in the industry?
Li Ying: Currently, Xiaodu has 46 million self owned brand devices, while there are 700 million devices covered by third-party ecosystems. Equipped with DuerOS smart devices, the monthly voice interaction frequency exceeds 7.1 billion times. These are all industry applications, including smart elderly care, smart hotels, smart cars, smart home appliances, smart wearables, etc. We empower various industries with large-scale models by outputting Xiaodu AI intelligent assistant capabilities or complete solutions.
In the consumer electronics industry, we have established deep partnerships with top brands such as Huawei, Honor, OPPO, Vivo, and Xiaotiancai, empowering over 130 million devices in total. In the hotel/real estate industry, Xiaodu currently covers 1.3 million hotel rooms, including more than 10 well-known domestic and foreign hotel groups such as Huazhu, Jinjiang, InterContinental, First Travel Home, and Yaduo, as well as high-end hotel brands such as Shimao Intercontinental and The Chedi Anlan Hotel, occupying an absolute leading position in the industry; In the elderly care industry, Xiaodu collaborates with industry partners to build a smart elderly care industry ecosystem, with over 40% of elderly users in home settings, and Xiaodu users using it for an average of more than 3 hours per day. In addition, our IoT whole house intelligence has been deployed in 400 cities across the country, serving over 300 million people in total.
Southern Metropolis: What industries will be the focus of layout next?
Li Ying: The industries we will focus on next include hotels, elderly care, and whole house intelligence. Of course, cars and home equipment are also key. Firstly, these are industries where we have accumulated experience, and the arrival of large-scale models will also bring changes to the industry.
A typical case is the recent popularity of China Travel, where many overseas tourists who come to China for travel post videos of their hotel stays on social media platforms, shocked by Xiaodu's English intelligent conversations. These are the capability improvements brought by the upgrade of big model technology to Xiaodu, allowing us to quickly and efficiently add a new language to the product.
Refactoring the 'usefulness' of a large model from three aspects
Nandu: What specific goals have been set after taking over as CEO of Xiaodu?
Li Ying: Xiaodu's strategic positioning has always been "AI+hardware". Since taking over, I have seen the opportunities brought by the technological transformation of big models and the combination of AI and hardware. My goal is to redefine hardware products through big model technology.
I believe that AI assistants are a strategic opportunity at the entry level of the big model era. We must seize this opportunity, build the core competitiveness of Xiaodu in the big model era, and deeply explore the "usefulness" of big models. It mainly includes three aspects of refactoring:
One is the reconstruction of AI assistant capabilities: DuerOS capabilities and ecosystem construction require continuous accumulation of large model capabilities, strengthening the construction of AI native operating systems, and creating a richer and more powerful intelligent agent ecosystem through multimodal perception and natural language interaction. As Robin said, intelligent agents may be the closest and most mainstream way to use large models for everyone in the future. Based on powerful basic models, intelligent agents can be generated in batches and applied in various scenarios.
Secondly, the C-end will redefine hardware with AI: I will continue to consolidate the advantages of smart devices in home scenarios, redefine smart speakers/screens, and smart education products with AI, bring about a greater leap in product experience, explore the continuous evolution of hardware products' intelligence and emotional intelligence, and break through the boundaries of big model capabilities; At the same time, actively opening up user entry level devices may be a brand new track. Thirdly, B-end empowers more industry sectors, such as smart cars, smart elderly care, smart hotels, smart homes, medical health, etc., and collaborates with industry partners to expand the ecological panorama.
Nandu: How do you view the development trend of AI native operating systems in the future smart device market?
Li Ying: Big models are becoming the new core of AI native operating systems. As the core engine, the big model cannot be absent in the operating system kernel. In addition to various large models, operating systems also need to have the ability to build powerful large model services, providing toolchains for model invocation, evaluation, deployment, and more.
As for the future development trend in the smart device market, the first is the further deep integration of AI native operating systems and AI technology. On the one hand, this integration will enable smart devices to demonstrate a higher level of intelligence in handling complex tasks and understanding user needs. On the other hand, AI native operating systems will continuously introduce new technological innovations, such as multimodal interaction, natural language processing, reinforcement learning, etc., to enhance the intelligence and user experience of the system. These technological innovations will drive the development of smart devices towards a more intelligent and user-friendly direction.
Next is the entire smart hardware market. As consumer demand for smart products such as smart homes and wearable devices continues to grow, AI native operating systems will become a key element in enhancing user experience and competitiveness for these devices. In the future, AI native operating systems will be applied in more types of smart devices, from smartphones and tablets to smart homes, smart cities, and other fields, and market demand will continue to grow.