The next stop for on-device intelligence: a recap of the Personal AI × Physical AI closed-door exchange
On October 16, the closed-door exchange "The Next Stop for On-Device Intelligence: Personal AI × Physical AI," hosted by OpenBMB and ModelBest, was held at INNO100 in Shenzhen. Speakers from Looki, Guangfan Technology, ComindX, LYNOOK and Mira Light shared their explorations and thinking from product practice. Zeng Guoyang, co-founder and CTO of ModelBest, gave the opening introduction and moderated the roundtable, joining the guests in a deep discussion about the future of on-device intelligence.
The exchange centered on one core question: how do you build a good piece of Personal AI hardware?
Product form | How does it fit into daily life? Earbuds, glasses, a pendant, a portable drive or a desktop terminal? Starting from usage scenarios, the discussion weighed different product forms across perception, interaction, battery life and cost.
Personal system | How does it understand one person better over time? Which information is worth remembering, when should it be recalled, and how should memories be updated and deleted? How should the model layer, memory layer and application layer work together, and how do you achieve efficient device-cloud collaboration?
Scenario value | What are users willing to pay for? How much value can life logging, personal assistance and desktop companionship each create for users? Drawing on product practice and user feedback, the discussion covered willingness to pay and business models in different scenarios.
Five teams shared their product practice, technical paths and current thinking around these questions.
Looki | How Personal AI fits into real life: Looki's exploration of Personal Intelligence
Gao Xinlu, COO of Looki, argued that as model capabilities keep improving, how ordinary users benefit from them in daily life is the key question Personal Intelligence must answer. Beyond model capability, he noted, AI also needs to understand a person's real experiences and circumstances, and much of that information happens offline.
Based on this judgment, Looki set three requirements for its hardware: a near-first-person perspective, multimodal information capture, and the ability to be worn for long periods with little burden. The team studied different forms and ultimately chose a pendant, weighing weight, power consumption and wearing comfort. With the hardware form settled, the team still needed a reason for users to keep wearing it, so it chose recording as the product entry point: people already want to preserve life experiences, and AI can lower the burden of recording and organizing them.
Along this product path, recorded content further becomes searchable memory. Gao Xinlu described one user's experience: during a trip, the user lost a hat, reviewed the Looki recordings and found it had been left on the high-speed train, then contacted the station and recovered it. Building on this, the team wants to connect past memories with the current situation, so AI can proactively surface information that matters to the user and gradually connect to agents that can execute tasks. Recording, memory, insight and action thus form a path that reaches deeper into daily life.
To support that path, the cost of processing continuous data is a key challenge. Looki's approach is to first use low-cost signals to decide when something is worth capturing, then organize the data into layered personal memory, and finally decide which capabilities to invoke based on the current goal and context. User control runs throughout: the more automated the system, the more users need to know what it is doing and to be able to choose, modify or reject results.
Gao Xinlu said the goal of Personal Intelligence is to lower the barrier to using AI, so users can get ongoing help in daily life without repeatedly composing instructions and supplying background.
Guangfan Technology | The last mile of Personal AI, and how an AI OS empowers wearable hardware
Zhu Guangwen of Guangfan Technology argued that turning a model's understanding and reasoning into services users can call on at any time is a key problem for making Personal AI real. As on-device models and cloud capabilities advance, he noted, sensing entry points and computing no longer need to sit on the same device, which gives wearable devices close to a person's senses and always ready to interact a chance to carry more Personal AI functions.
Zhu Guangwen also pointed out that the hardware entry point varies with user habits. Some people prefer glasses, others would rather wear earbuds. Based on this, Guangfan focuses its R&D on software capabilities shared across hardware, hoping to use an AI OS to connect personal agents, application services and compute for different devices.
Earbuds are the starting point on this path. Zhu Guangwen explained that they have mature demand and an established electronics base, making smart upgrades relatively easy; brands and supply chains are also fragmented, leaving room for third-party solutions that serve hardware makers. Choosing earbuds reflects both why users would use them today and how a solution can be adopted by more manufacturers.
But there is still distance between connecting a device to a model and completing real tasks for users. Tasks such as hailing a ride or booking a ticket require connections to external services, and the openness of applications, partnership mechanisms and invocation costs all shape a product's service capability and application boundaries. Adapting downward to different chips and hardware platforms, connecting upward to applications and services, and then organizing these capabilities into an experience users can use is what Guangfan calls the "last mile."
Zhu Guangwen therefore proposed a gradual product path: first improve existing functions to establish a basis for continuous wear; then take over some needs that would otherwise require reaching for a phone, reducing the burden of operation; and as services and model capabilities mature, continue exploring native AI experiences. The hardware entry point can vary, but the system and services behind it need continuous refinement.
Three early-stage projects exploring personal memory and everyday companionship: ComindX | Building a self-evolving hippocampus for human-machine interaction
Liu Senhao, founder of Wuwei Erzhi Technology, noted that while general models keep getting better at reasoning, their understanding of a specific person does not necessarily grow in step. In his view, the key to connecting general model capability with individual user needs is memory: a person's experiences, relationships, preferences and goals need to be organized so AI can make more personal judgments in concrete tasks.
Based on this, ComindX is exploring a cross-device personal memory foundation. User information is scattered across computers, phones and different apps, and file synchronization does not automatically carry over the experience and preferences expressed in conversations. If every service remembers only a small piece, users still have to introduce themselves again and again across tools. A cross-device personal memory foundation is therefore needed to bring together all of a user's experience and preferences.
Once information is gathered, the next problem is recalling memories effectively and keeping them updated. Liu Senhao further noted that retrieval of existing records alone may miss implicit relationships, while continuously training constantly changing personal information into a model brings update and cost pressures. The team therefore wants to combine memory organization with model adaptation, so the system can both find relevant information and keep adjusting its understanding as the user has new experiences.
Toward this goal, ComindX is exploring graph memory to organize relationships among users, tasks and knowledge, supporting recall through on-device storage and compute optimization, and then using cloud capabilities to continuously update the personal system. On the device-cloud split, the team proposes that the device handle more of the services related to personal identity and memory, while complex general reasoning goes to the cloud, which supplies relevant context on demand to reduce the overhead of repeatedly passing in large amounts of history.
The design aims to build personal memory that accumulates with the user. In the team's design, even if hardware and models are replaced, a new intelligent service can continue the existing understanding with authorization. By reducing the need for users to re-adapt across services, a personal system can accumulate understanding of the user through continued use.
LYNOOK | From Work Agent to Life Agent
Luo Jingmin, founder of LYNOOK, shared his views on how agents enter personal life. Agents at work, he argued, are usually evaluated around goals and results; a Life Agent that enters the home must also make users want to own it, keep engaging with it, and get entertainment, companionship and self-expression from it.
This choice brings three interrelated questions: who bears the cost of continuous operation, how to build user trust, and how to make complex capabilities understood and used naturally. LYNOOK hopes to answer these with a desktop "Stanford town," letting agents live alongside users as characters in a space.
The town-like presentation is first meant to make autonomous action more understandable. Which character can operate which device, what it is doing, and which rooms its permissions cover can all be shown through spatial relationships users already understand. By making the running process visible, users can judge the scope of authorization more clearly and intervene when behavior does not match expectations.
The impact of that intervention then becomes the basis for the next interaction. Luo Jingmin offered a hypothetical: when a friend visits, the user tells the agent that the friend is a maker who would appreciate the equipment in the garage. The system should retain this correction so the next welcome is different. Shared experiences, user corrections and subsequent behavior are thus connected, giving continued engagement a chance to accumulate value.
To keep this experience running long term, the team proposes a split in which the device handles daily operation and the cloud handles new demands, while borrowing the commercial logic of games and entertainment to explore core users who will pay for deep experiences and self-expression. Whether the product looks good and can be customized and displayed therefore also enters the design considerations. For LYNOOK, these choices serve one direction still to be validated: making the shared experience between user and agent a reason to keep using and paying.
Mira Light | Seen, heard, lit up
Wang Jianle, founder of Mira Light, used the way people relate to friends to explain the team's understanding of AI companionship: first perceive the other person's emotions and state, then understand what is needed at this moment in light of shared experience, and finally offer an appropriate response and action. To form such a relationship, a device needs to appear continuously in daily life, gain context, and have sufficiently rich means of expression.
In user research, the team observed that existing companion hardware often lacks a reason to be used every day: interaction depends mainly on the user initiating it, expression is limited, and charging and operation add burden. Once the novelty fades and the device sits idle, the opportunities to learn more about the user also shrink. Mira Light therefore considered high-frequency use, proactive perception and emotional expression together.
Based on this, the team chose the form of a desk lamp. Wang Jianle recalled where the idea came from: at three in the morning, the cats and dogs at home were asleep, and the desk lamp was still on. What would it be like if it could respond? For people who work, study and create at a desk for long stretches, a lamp is already nearby; lighting provides a reason for daily use, and a fixed position makes power, connectivity and continuous interaction easier to arrange.
On that basis, the team wants the lamp to become a companion that can perceive and express: understanding the user through vision and sound, responding through body posture, light and voice, and letting long-term memory shape changes in personality and interaction style. The same gesture of care should feel different as the relationship develops.
To support this experience, the technical implementation needs to connect perception, understanding, decision-making and action. The team described its work on behavior models, mental models and device-cloud collaboration, aiming to let the device choose suitable interactions for real scenarios. Mira Light first targets independent creators and others who spend long hours at a desk, trying to build a companion relationship on top of lighting as an everyday function: establish the basis for ongoing company through daily use, then gradually deepen its understanding of the user through long-term interaction and offer more personal responses.
Roundtable: the future of Personal AI
The roundtable discussion centered on three questions: what hardware form suits Personal AI, how it can understand one person better over time, and what value users are willing to pay for.
On hardware form, several guests stressed the importance of coordination between different devices.
One guest used his own wearing habits as an example: even if glasses have strong visual and voice capabilities, health-related information may still need to come from a ring, a watch or other devices. Different positions yield different information, and a person's needs span multiple scenarios.
Another guest noted that whether hardware is suitable depends on the specific scenario. Phones still play a role in identity, connectivity and compute; a desk can provide a compute node; glasses and earbuds each have interaction advantages. The value a new device adds must be enough for users to accept its weight, battery life and usage burden.
On understanding the user, the guests dug further into the relationship between data and actual experience.
An translation example made the question concrete: the same word may need different expressions on different occasions. Only by first recognizing whether the user is chatting with a child or in a serious meeting, and then composing the translation, can it fit the actual need. Memory must be accurate, and the live environment must also inform judgment.
The discussion also noted that the current environment, past memories and the intent the user is expressing need to work together to help the system understand the task at hand.
But more information also means higher trust requirements. Whether users are willing to hand over data, and whether they can understand and correct AI's understanding of them, affects how far the system can go. One guest particularly reminded the room that even if AI can do many things for people, experience design must preserve a sense of human participation.
Another point the guests returned to was the measure of proactive service. Timely reminders and advance preparation can delight; irrelevant suggestions and overly frequent interaction can become a nuisance. When to appear and in what form is as much a part of product refinement as whether the answer is correct.
On business models, the guests offered various ideas based on different product positioning and user needs.
Some valued the ownable, perceptible value of the hardware itself; some would pay for features that clearly solve a problem; others cared more about companionship, aesthetics and self-expression. For how services charge on an ongoing basis, subscriptions, hardware-software bundles and advertising all became topics of discussion.
One guest suggested that sustained willingness to pay among mass-market users may be harder to build than practitioners imagine, and that advertising is seen as one option to explore. But advertising also conflicts with AI's objectivity and user-centered service in ways that need to be handled.
Another view started from product experience and described a path from "usable" to "good to use" and then to "want to use" in specific scenarios. Translation, teleprompting, recording, or experiences that help people relax and enjoy themselves can all create reasons to pay; the key is whether the value to the user is clear enough.
Conclusion
Personal AI is an important direction for AI to enter daily life. It needs to gradually learn a person's habits, remember important experiences and offer the right help when needed, through long-term companionship with one individual. The value of intelligence is thus expressed in each person's concrete life.
On-device intelligence is an indispensable part of this. Perception close to the user, timely and natural interaction, and local processing of personal data and memory all require on-device capabilities; only by coordinating with complex cloud reasoning can sustained personal service become possible. From models to hardware, from memory to applications, every concrete question discussed at this exchange bears on whether Personal AI can truly become part of everyday life and keep creating value for users.