From AI PCs to super-intelligent bodies, how Lenovo delivers personal AI experiences

One year after launching the AI ​​PC , Lenovo yesterday upgraded its personal AI agent " Xiaotian " to a personal super agent " Tianxi " . Also launched on the same day were the enterprise super agent "LeXiang", a new generation of inference acceleration engine, and four AI terminals equipped with Tianxi , including a mobile phone, a tablet, and two notebooks.

In last year's "AI PC Report " , Weijin Research proposed that the commercial essence of AI PC is to fully deliver the user experience of personal AI to users . It relies on two killer applications, AIOS based on large models and intelligent agent technology , and an inference engine based on hybrid computing power.

Upgrading " Xiaotian " to " Tianxi " gives it the three major features of perception and interaction, cognition and decision-making, and autonomy and evolution. The core change is that it gradually evolves from binding the intelligent body to the device to binding it to the individual, truly improving the personal AI experience. In Yang Yuanqing's words, it is now a "cognitive operating system", " the general entrance to the needs and problems that individuals and enterprises need to solve urgently " , " the more you use it, the more it understands you " .

This "cognitive operating system" is getting smarter, upgrading from passively responding to conversations to autonomously using tools, and can also interact with other intelligent agents; it is also gradually moving from vertical fields to modular combinations, and even general intelligent agents. OpenAI 's o-3 already has a certain ability to use tools autonomously, and the large model context protocol MCP and the intelligent agent interconnection protocol A2A are widely accepted.

Relying solely on large models, "illusions" continue to accumulate in multi-step reasoning, making its experience still not good enough. If the knowledge and tools it uses are more personalized and specific, and even rooted in the user's personal device data and work and living environment, then the AI ​​experience it delivers will be closer to the user's reality.

Lenovo invited guests to demonstrate how their personal agent plans their trips at the press conference. It retrieved their scheduled trips and made preliminary schedules. The agent also broke down the trips, planned routes, and even knew the geographic location of the guest's departure date. It called applications on the guest's mobile phone. Another guest demonstrated how to call documents and data on a personal computer and generate a page of PPT with charts based on her style preferences .

The user's personalized data is stored in different applications and even different devices. Last year, they could only be activated on the device where "Xiao Tian" was located; this year, it can be mobilized across devices, and in the future, it can even include non-Lenovo devices, as long as the device has "Tian Xi". They are all "long-term memories" of user behavior by intelligent entities.

Lenovo also demonstrated that the intelligent agent can understand the user's past consumption habits and recommend travel plans that meet the user's expectations based on hotel dynamics. In the future, the user's personal intelligent agent can even communicate with the personal intelligent agents of relatives and friends to discuss how to travel together.

Personalized AI experience requires personalized data, which means data security and privacy risks. Microsoft launched the Recall feature last year, which can remember and understand everything the user does on the computer by taking screenshots. After calming down the privacy controversy, it was postponed to launch recently. Ultimately, the intelligent agent on the computer will be able to remember and complete everything the user assigns.

One way to protect users' personal data is to use completely local reasoning. Last year, a number of AI PCs upgraded their computing hardware. Since last year, DeepSeek has tried to squeeze the computing power of existing AI chips to the limit and has also open-sourced the solution. China's domestic AI ecosystem is on the wave of software and hardware collaborative optimization.

This time, Lenovo cooperated with Wuwen Xinqiong and others to develop an " inference acceleration engine " that can compress the performance equivalent to the cloud-based lightweight inference model o1-mini into an AI PC . At the venue, Lenovo asked three Lenovo AI PCs with the same configuration to solve the 2024 college entrance examination math problems at the same time. The first-generation model failed to solve the problem, but the second-generation model, equipped with the " inference acceleration engine " , not only doubled the problem-solving speed, but also reduced memory usage and energy consumption by 40% .

Lenovo plans to fully integrate the " inference acceleration engine " into the next generation of Lenovo AI PCs , allowing them to smoothly run large models with tens of billions of parameters locally, far exceeding the minimum threshold of 7 billion parameters proposed last year . Yang Yuanqing predicts that in the next 12 months, the comprehensive capabilities of edge AI will at least triple compared to now.

Placing it on the client side or running it in a hybrid AI mode can also alleviate the cost problem of the intelligent agent being too expensive. Even if it develops in the direction of high autonomy and high-frequency interaction, swallowing up more tokens and computing power, users don’t have to feel bad.

Another way to address privacy risks is trusted hybrid AI reasoning. Lenovo Group Chief Technology Officer Tolga Kurtoglu revealed that the MCP protocol has been embedded in different agents ; when the task is particularly complex, the system will also start an agent-level architecture, first by a " consultant agent " to decompose and schedule the task. Under the hybrid computing power architecture, the task scheduling system will keep the end-side computing power accessible or sensitive tasks locally, and schedule the desensitized data to other devices. In the end-to-end scheduling architecture, the system-level middleware is responsible for ensuring data security and privacy. When the agent uses user data, there is also AI to identify and prevent possible risky behaviors.

During the demonstration, a guest asked Tianxi to help her post a note on Xiaohongshu. The agent retrieved its mobile phone album, recommended a group of photos with relevant content, and prompted that one of the photos had a privacy risk and suggested not to check it.

In the enterprise intelligent agent "LeXiang", there is also a similar security protection. Whether in reasoning, calling external models through APIs , or calling tools through the MCP protocol, a security protocol must be passed. The same is true for mutual calls between intelligent agents. The intelligent agents within the local private environment and the mutual collaboration with third-party intelligent agents must also cross a security protocol.

In fact, interconnecting agents or using advisory agents to orchestrate tasks and deliver results is more secure than directly using tools and resources on the device being called. This is what Google's inter-agent interconnection protocol A2A does. It divides the content sent by agents into tasks , messages ( including user context, instructions, errors, status or metadata, etc.) and artifacts ( the results of tasks, a single task can be split into multiple artifacts to achieve). The local agent initiates a call request to the remote agent, delivers tasks and messages, and the remote agent feeds back messages and artifacts, all of which are protected by a layer of security with the resources on the device where the agent is located.

Under this hybrid reasoning engine and security settings, personal agents connect users' operations, applications and data on mobile phones, tablets and computers, forming " long-term memory " and becoming more and more familiar with users; enterprise agents also have almost unlimited memory based on the company's private data and knowledge base, and can continue to evolve. When facing this silicon-based companion, you can't just hire it, you must continue to train it.