Image source: Generated by Unbounded AI

If at CES in 2024, AI will be displayed more as an independent highlight by a few companies; at CES 2025 this year, the integration of AI and the consumer electronics industry will be more extensive and in-depth, just like this year’s CES. The theme of CES is "Dive In".

Take smart glasses products as an example. At this CES, from AR manufacturers such as Rokid, Thunderbird, Xreal, and INMO, to cross-border players such as Xingji Meizu, Raytheon Technology, and Dapeng, to Haliday , Vuzix and other new companies, Chinese manufacturers have staged an "AI Hundred Mirrors War" thousands of miles away in the United States.

At the CES exhibition, "Smart Emergence" found that from stringless guitars, AI facial masks, to rings, crutches, AI bicycles... all kinds of hardware products in life have become large-scale AI models. The new "face" of the terminal. Not to mention all types of consumer electronics that have embraced large models early - from AI glasses, AI headsets, to AI companion robots, to AI PCs, mobile phones, learning machines, etc.

△The theme of this year's CES is "DIVE IN"; Photographed by Su Jianxun

AI is everywhere in this "Spring Festival Gala" of the consumer electronics industry; but behind the enthusiastic appearance , the AI ​​hardware industry needs in-depth "cold thinking":

How many mountains do hardware manufacturers have to cross from "using AI" to "making good use of AI"? When it is no longer uncommon for products to be equipped with large AI models, can AI still bring selling points and premiums to products? Large AI models continue to iterate, can the empowered intelligent hardware keep up?

The founder of an intelligent hardware manufacturer also questioned the soul of "Smart Emergence": When most players on a track use AI, does it mean that the threshold for this matter is too low?

At CES 2025, "Smart Emergence" interviewed four smart hardware manufacturers that actively embrace AI large models, namely Future Intelligence (AI headphones), Xueersi (AI learning machine), Li Wei Ke (AI glasses) and INAIR (AR glasses), they shared their current practices, explorations and challenges in applying AI large models in their respective segments.

Xueersi CTO Tian Mi: It is difficult for Chinese users to pay for AI software, and a better way is to combine software and hardware

1. In China's smart hardware industry, no manufacturer has yet been able to truly put large-scale end-side models into product lines. They all run on the cloud. Because China's current terminal-side chips are immature, they cannot yet run large models.

2. But in the next 2-3 years, I predict that there will be some simple large models that can be run on the device side, and the remaining complex operations will rely on the cloud.

3. It has been less than a year since we launched a large model on Xueersi's hardware products, and we spent the previous two years exploring. We found that AI software is difficult in ChinaChinese users will not pay for an APP if it is implemented alone. They feel that AI technology is not valuable.

The combination of software and hardware is a good way for consumers to feel it. We have put various AI functions into Xueersi learning machines. Actual user data proves that the most frequently used are various AI applications, such as AI correction, AI lectures, and interaction with the intelligent assistant "Xiaosi".

△Xueersi’s AI learning machine source: Enterprise authorization

4. At the beginning, we hoped to train a large model of our own from scratch, but it took a while I found that more and more better open source base models are coming out, and it is actually very uneconomical to do pre-training by yourself. Later, we added a lot of special knowledge in the education field to the world's best open source large model bases for retraining.

Our approach is to cut off the pre-training of general knowledge, but not save any other steps, including pre-training of professional knowledge, fine-tuning and reinforcement learning, which are all ongoing.

5. Compared with previous AI models, large models have greatly improved Xueersi’s product capabilities, which are mainly reflected in two points. First, the accuracy of work (such as AI correction) has been greatly improved. Improvement, the performance is better and stronger; the other is that tasks that could not be done before can now be done.

6. The continuous training and reinforcement learning technology of large models is very difficult and requires very smart talents to explore and try in constant experiments. This field requires talents who both understand algorithms and can do engineering, and their research and development skills must be strong.

7. The same is the model ability of accessing Xueersi. The hardware form is very important for user acceptance. For example, a learning machine is more convenient for users to learn than a mobile phone. We now have both independent APPs and learning machines. At present, there are many domestic manufacturers, including mobile phone manufacturers, Pad manufacturers, PC manufacturers, and eyewear manufacturers, all of which are adopting Xueersi's API.

Future Intelligent CTO Wang Song: Large models are developing in two directions, one is the base and the other is the end

1. In the future, the wearable device will be a so-called AI agent, which can accompany the user at all times, instead of having to be held in the hand like a mobile phone. It has a variety of sensors that can serve as the user's eyes or ears to sense the surrounding environment and give feedback to the user.

2. At present, the focus of future intelligent iterations is to work towards personalization. We extract useful information from the user's meeting content in a structured manner and store it in the form of a database or RAG to form a long-term memory of the large model. This piece of memory will eventually be associated with the user's personal assistant, which will generate some personalized answers that match the user's preferences based on the user's personal preferences.

3.AI glasses can now run some computing power, such as Ray-Ban Meta is equipped with some local models, which can be calculated in real time through the SOC chip. However, because the SOC has insufficient computing power, AI headsets still use cloud computing power. Almost all the so-called smart headphones we see on the market now rely on the computing power of the cloud.

4. If the computing power is deployed on the client side, the response of large models will be faster, more timely, and more secure. Many users are concerned about data privacy. For example, the meetings held by some investors may be very sensitive and do not want the data to be uploaded to the cloud. Future smart AI headsets will provide this function option. User data does not need to be uploaded to the cloud, but will be stored in the headset or mobile phone.

△Future smart AI headset source: Enterprise authorization

5. The large AI model is currently developing in two directions. One is the large base model, its parameters and data volume It is getting bigger and bigger; the other direction is the terminal side, which is becoming more and more efficient, and security, data security, etc. are also guaranteed. These are two directions, which actually do not conflict.

6. The iteration or advancement of AI capabilities will actually have a huge impact on the future of wearable devices. I predict that in five years, some large local AI models will be able to run on headsets. Once that level is reached, the headset can be used as an independent device, and many interactive scenarios do not need to rely on the mobile phone. This will bring qualitative changes to some experiences at the user interaction level.

7. Currently, there are relatively few AI hardware that can achieve high premiums by accessing large models. This involves the development stage of an industry. At this stage, the so-called smart headphones are actually implemented by software on the mobile phone. I think it may have to develop to a certain extent, and the headset can run some end-to-side models locally before it can truly achieve the so-called smart headset.

To realize a true "smart headset", there are currently two main sticking points, both of which are in the hardware. One is the computing power of the SOC chip. The computing power chip of the headset must be small in size and high in computing power. The computing power is strong, which is difficult to achieve; the other is the battery life problem. If the SOC chip is inserted into the earphones, the power consumption will be very high and the battery life will be very short, which is difficult for users to accept.

Ru Yi, founder of Li Weike: The development cost of AI glasses application is much lower than that of the XR ecosystem, and it will not follow the old path of XR

1. I think that the most dense way for humans to obtain information is through the eyes, so my intuition is that AI glasses are one of the consumer carriers closest to the eyes and are the carrier of voice interaction. This is the best carrier for conversational AI.

2. Killer applications installed on AI glasses will definitely appear in the next two years. This is what Li Weike must do. Otherwise, AI glasses will become a simple "shell" with little value.

3. When I founded Li Weike in 2021, I had a judgment: in the next three years, AI will have explosive growth. But I didn’t expect it to be so soon, 2022It started already at the end of the year and exceeded expectations. So in the spring of 2023, we made a choice-All in AI large model.

When designing products, we have been firmly doing two things. One is to do a good job in AI interaction, and the second is to do a good job in personification so that thousands of people can have different faces.

Of course, what we value more is that we built the entire large model system ourselves, it is complete, and we can continue to iterate. Rather than handing things over to a third-party model company, you have absolutely no control over it.

△ Source of Li Weike’s AI glasses: Enterprise authorization

4. The large AI model provides information integration on the web page very well, but the effect is not good if it is directly connected to the glasses. Well, it requires a process of integration.

For example, if I ask the AI ​​glasses how the weather is today, the AI ​​model will not answer directly, but will ask where you are? Therefore, in order for AI glasses to have a good experience, the large AI model must be optimized and adjusted.

5. For our AI glasses startup company, we do not actually need to hire many people to do things related to large models. Our entire modeling team may only have about ten people, but we can stand on the shoulders of giants to fine-tune and optimize.

6. Not only smart glasses, but also any industry that has reached this stage will have fierce competition. A market without competition will not prosper. Competition is necessary to jointly educate the market and penetrate the consumer side faster. In the past few years, when there was little competition in the smart market, the cost of the education market was too high.

7. In the past, the XR ecology was not mature enough, resulting in poor sales. This was largely due to the incomplete ecology and high application development costs. AI glasses will not follow this old path because its development cost is much lower than that of the XR ecosystem. If a suitable scenario can be found, it is possible that one or two developers can build the agent.

Qi Jingxuan, INAIR product design leader: In the future, the AI ​​Agent itself will become an independent OS

p>

1. Since the advent of operating systems, everyone has thought that there would be something like a "little assistant" in the computer to help you solve many things. However, in the past, including Siri, Xiaoai classmates or Google Assistant, none of them did a very good job, and in most cases they failed. Because users don’t know where the boundary capabilities of AI dialogue lie. The advent of the big model changed that, it put all the questions in perspective and allowed all the conversations to proceed.

2. As soon as ChatGPT broke out at the end of 2022, we were aware of this trend. Adding large AI models to INAIR's products was in our plan from the beginning.

INAIR applies large models and most AI glasses and AI hardware currently on the market.Software is different. Their main function of AI is to help users understand the external world. INAIR mainly helps users solve software and system operation problems more efficiently.

3. For INAIR, the use of large models in products is similar to the use of Copilot in Microsoft Windows PCs, both of which are an important selling point. Large models can better solve user experience problems, giving users more natural interactions and a faster and more convenient experience.

4. INAIR cooperates with many large AI models. We found that different large models are good at different things. For example, Doubao may have a stronger ability to understand pictures, and iFlytek is good at ASR (speech recognition) interaction. Very capable. INAIR will use different large models in different scenes.

△INAIR’s AR glasses Source: Enterprise authorization

5. INAIR’s product advantage lies in the integration of software and hardware. In an environment that integrates software and hardware, AI multi-modality can realize a closed loop from perception prediction to interaction, communication to execution.

This is also the advantage of INAIR products. For example, users can read an English paper and have the Chinese translation displayed on the screen of the glasses in real time, or directly ask the agent what the Chinese summary of the paper is. Another example is when watching a movie, you can directly ask the agent questions about the characters or objects in the picture. These are all functions that can only be achieved by a Siri-like role at the system level, which can flexibly call different applications.

The above operations can also be achieved simply by relying on software, but they require mouse clicks, copying and pasting, and switching between different APP pages. The operations are much more cumbersome. This is the difference between software and hardware integration and pure software.

6. Devices integrating software and hardware can also achieve active perception and prediction. For example, if the device sensor finds that the user stays on a certain interface for a long time, the system agent can provide targeted service suggestions. .

7. We hope that the large model (technology) on the end side can be further improved so that the AI ​​large model can be called when AR glasses are not connected to the Internet. One advantage of this is faster response. Currently, the response speed of large model solutions in the cloud is still relatively slow; in addition, if not connected to the Internet, users can better protect their privacy and be more secure when using it.

8. All the hardware forms of users today, whether it is a computer, a mobile phone, or a computer, are all solving the problem between the user and the final application. AI is an application and function in the OS. But in the future, AI Agent itself will become an independent OS to solve the distribution problem of these list-based applications.