AI doesn’t replace us; we create together with it.

The Cell Wall focuses on the bridging “AI interaction system”.

Cell Wall’s answer is straightforward: AI isn’t here to replace people, but to give them new capabilities. Humans still steer the ship, but they don’t have to do everything alone.

This is also the product roadmap that Cell Wall outlines around the “AI‑enhanced creator.” At the launch event, Cell Wall unveiled two products: CyboPal ONE and WhisPrompt. They were introduced within the same collaborative context: the former is a desktop‑size robot that connects to the user’s computer, and the latter is an AI work handheld that works alongside the computer. The two products approach desktop and mobile office use cases respectively, while sharing the same underlying AI interaction system. In Peng Tianfang’s words, Cell Wall wants people to bring AI along to realize their ideas, while still retaining control over direction and outcomes.

Peng Tianfang’s observations in the interview can be summarized in three layers: how to use AI to extend human capability boundaries; why embodied intelligence starts with PCs and desktops; and how CyboPal ONE and WhisPrompt serve as two terminal entry points of the same AI interaction system. Linking these three layers together reveals how Cell Wall surrounds the user with goal setting, feedback, judgment, and critical‑point takeover, building a complete human‑machine collaboration system.

CyboPal ONE and WhisPrompt, two entry points of the same AI interaction system

At the launch event, CyboPal ONE and WhisPrompt were presented within the same collaborative workflow. The two products are not isolated hardware items; they share a single AI interaction system that connects users, computers, agents, and devices, enabling people to express ideas, modify requests, switch tasks, confirm key actions, and view results.

In other words, the focus of the presentation was on “how people use this system to accomplish a task with AI.”

First, look at CyboPal ONE. It is a desktop‑size robot that connects to the user’s computer, combining a large screen, perception, and physical movement, allowing AI to appear naturally beside the computer in an embodied form. Peng Tianfang sees this type of product as a bolder experiment: it targets people who work at their desks for long periods and heavily use agents, especially geeks and deep developers. It turns the computer into a perceptible, interactive, and movable embodied collaboration entry point. Users still handle tasks, launch software, and view results on the computer; CyboPal ONE makes the agent’s deployment, execution, and feedback enter that scene more naturally. According to Peng Tianfang, CyboPal ONE features voice input, voice editing, user memory, and delves deeper into this AI interaction system.

Take another look at WhisPrompt. It is an AI work handheld device designed to be used with a computer, placing voice input, buttons, and knobs within reach, targeting mobile office scenarios. Users can more easily express ideas, adjust requirements, hand them off to an Agent for execution, and view the results. Peng Tianfang described it very specifically: the handheld first enhances the experience of intent expression and control, allowing users to rely less on keyboard and mouse and convey ideas to the computer more quickly; the team’s philosophy is that once users become familiar, they can work with one hand, and eventually the computer might just sit there while people speak to it and make slight gestures to interact. The product incorporates hardware elements such as a touchpad, buttons, and knobs, pulling high‑frequency actions like “modify requirements,” “switch tasks,” and “confirm actions” out of the screen and placing them at hand.

The two products share this AI interaction system, connecting users, computers, agents, and devices, allowing AI to understand tasks, execute actions, and return results; when key actions are involved, a human confirms or takes over. At the launch event, the experience was tied together with “Speak, Edit, Switch, Approve, View”: Speak expresses an idea; Edit modifies the request; Switch changes the task; Approve confirms critical actions; View checks the outcome. These five actions may seem simple, yet they correspond to the most fragile links in human‑agent collaboration.

It should be noted that traditional operating systems still manage computing resources, and the cell‑wall AI interaction system is a layer on top that organizes human‑agent collaboration. It enables software, computers, and physical hardware to work together stably, bringing AI out of the on‑screen dialog box to become a perceivable, cooperative embodied intelligence collaborator beside the user.

When you look at CyboPal ONE and WhisPrompt side by side, you can understand why the cell wall says they are “two entry points to the same new capability”: they serve the same goal—enabling people to realize ideas with AI while retaining control over direction and results. As for why this entry point first appears on PCs and desktops, you need to refer back to Peng Tianfang’s assessment of the human‑agent relationship.

Why start from PC, how AI interaction systems organize collaboration between humans and agents

In a conversation with 36Kr, Peng Tianfang first discussed the boundary between humans and AI. In his view, if the relationship stays at the slogan level, it easily becomes an empty notion of “AI augmenting humanity”; if we only talk about “AI replacing humans,” the discussion turns toward anxiety.

What CellWall wants to clarify is the boundary between the two: humans retain direction, judgment, confirmation, and the right to take over, while AI handles expansion, execution, and experimentation. When humans start giving goals to the Agent, providing feedback, and making judgments, the human role shifts from “hand‑doing every step” to “setting direction, observing results, and taking over at critical points.”

From an industry perspective, this is essentially responding to the most common fork in embodied intelligence and AI hardware startups: doing only software is light but often lacks barriers; doing only hardware lands quickly but can become idle once the novelty fades. CellWall chooses to place the judgment on “people still at the helm,” at least giving the product logic a verifiable starting point—essentially seeing whether people are more willing and comfortable handing their ideas over to an agent, allowing users to dare to try, dare to change, and dare to adjust direction at any time.

In fact, this judgment is related to Peng Tianfang’s personal experience. He describes himself as “extremely lazy,” preferring to lie down rather than sit, and to achieve greater returns with less energy, so he naturally wonders: can a person complete work while half‑lying and looking at a screen? His academic background is in micro‑system design and manufacturing, which gives him insight into which new technologies can be mass‑produced at reasonable cost in the near future.

Thus, CellWall’s starting point in designing embodied intelligence product forms is precisely to make “users submit requests, screens provide feedback” happen in a more comfortable way.

His methodology for the product team is “add a bit of surprise and the unexpected on a familiar foundation”: the form factor shouldn’t be so odd that users are reluctant to use it, nor should it merely re‑skin something old. The large‑screen desktop robot aims to give a desktop PC or monitor a fresh lease on life; the handheld work device seeks a familiar yet slightly surprising spot between a mouse, a walkie‑talkie, and a phone.

This methodology sounds simple, but it’s not common in today’s AI hardware industry. A frequent approach in the industry is either to turn the “new form factor” into a spectacle for demo videos, or to slap voice capabilities onto an old peripheral and call it AI hardware; Cellwall tries to strike a balance between “familiar” and “unexpected.” The opportunity lies in the fact that once users find it intuitive, form‑factor innovation can become part of everyday use.

Because of this, Peng Tianfang believes the company must possess both hardware‑software and algorithm capabilities. Pure hardware results in products lacking soul and conversational ability, and it doesn’t understand users’ personal habits and history, so the novelty quickly fades; pure software that only interacts with users through a cold screen, mouse, and keyboard, without delivering a truly new experience, is unlikely to have any defensible moat.

He gave an analogy: raising a four‑legged robot dog is cute at first, but the interaction never becomes part of its habits, so the novelty quickly fades; a real dog, on the other hand, gets to understand you more and more over time. He wants to point out that the role of software and models is to let interactions solidify into user understanding.

That’s also why Cell Wall chose to build both an AI interaction system and a embodied physical terminal. From an industry division‑of‑labor perspective, this choice reflects Cell Wall’s desire to root itself between the “model layer” and the “hardware layer,” establishing the AI interaction system as a solid foundation. This position isn’t easy: upward, it must keep pace with rapid model capability iterations; downward, the hardware form is easy to imitate. But if context and user habits can truly be embedded in the interaction system, it could become a stickier entry point than a single app or a single piece of hardware.

The decision to start from PCs and desktops stems from Cell Wall’s “non‑consensus judgment.” The industry generally believes AI will bring new terminals, but the path taken determines whether a startup can succeed.

Peng Tianfang observes that the interaction patterns users truly rely on are changing dramatically, and the areas that need new‑form hardware are mainly concentrated in productivity scenarios: people who write code, draft articles, conduct research, or create content feel a strong pain and pleasure perception toward AI, and they indeed need a novel interaction terminal to boost efficiency. In contrast, for short‑video, social, shopping and other mobile scenarios, users do not strongly expect “efficiency gains”; the traditional swipe interaction is already comfortable enough, so the demand for new‑form hardware is weaker.

PCs and desktops are the core venues where creators handle tasks, launch software, and view results, and they are also the main battlefield for high‑frequency, continuous collaboration between humans and agents. CellWall chooses to start from the PC to first stabilize the primary human‑machine collaboration link, and then use WhisPrompt to add the “hand‑held” mobile office entry point. This strategic choice is noteworthy: it starts directly from “which group of people and which part of their work is most painful and frequent.” For early‑stage companies, proving the AI interaction system in a high‑frequency productivity scenario first is more realistic than simultaneously expanding into phone, glasses, home‑appliance, and other contexts; however, this also means CellWall must first endure the scrutiny of the PC scenario, and users must be willing to have an additional entry point beside their computer.

Peng Tianfang roughly divides the industry chain into four layers: cloud large models, edge-side models, interaction or operating system layer, and hardware substrate. He believes the first two layers are basically settled, while the “cell wall” resides in the third and fourth layers closer to the user— the interaction layer and the hardware substrate layer.

In his view, traditional operating systems have long fulfilled the historical mission of coordinating software and hardware resources and allowing users to control computers through mechanical clicks; the next‑generation operating system must address a deeper human‑device relationship: enabling computers, TVs, phones, refrigerators, air conditioners, glasses, and other devices to naturally recognize who the user is, understand the user’s past habits, and grasp the true meaning of vague commands.

In other words, the “cell wall” focuses on the bridging “AI interaction system”: the user expresses a need, it automatically understands and translates it into an executable request; the user’s language habits, preferences, and personal context are embedded in the interaction layer. “Context is the key for an agent to understand the user, and the interaction layer is the place closest to where context is generated and evolves, so it has the opportunity to become a more natural connection between humans and agents,” Peng said.

The name “cell wall” is also related to this. Peng Tianfang explained that the name was chosen in the hope that silicon‑based life could coexist with carbon‑based life; in the future, humans may be surrounded not only by other human partners but also by many silicon‑based companions, and interaction is the key to making this happen. This expression carries a visionary tone, but the product issue it points to is not abstract: how to make machines truly understand humans.

First solidify desktop collaboration links, then bring embodied intelligence into every home

From a business logic perspective, choosing a cell wall isn’t complicated, but the order is crucial. Many AI‑hardware startups tend to tell a big ecosystem story first and then go back to find users; the cell wall does the opposite: it first locks in a group of people who most need “to turn ideas into reality with AI,” then uses two entry points to cover their high‑frequency scenarios, and only then talks about openness and ecosystem. Whether this sequence works depends on whether it can first prove one thing: that there are people willing to continuously pay for this brand‑new human‑machine collaboration experience.

Peng Tianfang’s judgment of the target audience is very specific: developers, researchers, content creators, and freelancers who frequently use AI agents and want to bring their ideas to life are high‑frequency users. Compared with opening more chat windows, what they truly need is an AI interaction system that links intent expression, process control, and result verification.

Peng also mentions a consumption logic: AI‑enhanced creators spend long periods in front of computers and desktops, and they are willing to pay for comfort, efficiency, and smoothness of time; just as some people are willing to spend thousands of yuan on a keyboard or a few hundred on a mouse because each keystroke and grip feels more comfortable. This group’s spending preference is productivity tools that truly improve daily experience.

From an industry perspective, this is equivalent to treating “AI‑enhanced creators” as an emerging niche market that traditional peripheral manufacturers have not yet fully served: the demand is real, the reasons to pay are concrete, and the product form also has the opportunity to break out from the substitution relationship with mouse and keyboard.

Once the target audience is defined, the entry‑point combination gains commercial meaning. CyboPal ONE enters from the desktop, WhisPrompt from mobile work, and both share the same AI interaction system. The desktop entry covers high‑frequency actions of “collaborating beside a computer,” while the mobile entry covers scenarios of “having requests at hand.” For users, the two products are extensions of the same interaction system in different contexts; for the cell wall, sharing the interaction system means that coordination among software, computer, agent, and device is the core capability, with hardware being the physical endpoint of this collaborative chain. Peng Tianfang also admits that the two products target different scenarios: the handheld is easier for ordinary users to accept, while the desktop large‑screen robot is a bolder attempt at embodied intelligence in a desktop setting, initially attracting geeks and deep developers, then gradually bringing the experience to a broader audience. This “one steady, one daring” product combo aligns better with an early‑stage company’s resource rhythm than focusing on a single form: the handheld lowers the understanding barrier and expands the user base, while the desktop large‑screen robot validates more ambitious interaction hypotheses.

Looking further down the industry stack, Cell Wall places its AI interaction system—situated below the cloud‑model layer—alongside the embodied device layer: it connects users and agents on one side, and computers, devices, and various hardware forms on the other. Peng Tianfang believes that the desktop is only the first landing point for embodied intelligence; manufacturers of glasses, tablets, phones, cars, and even household appliances will need a more natural AI interaction system in the future. If the system is sufficiently open and can share data and architectural components with user consent, the adoption of embodied intelligence will accelerate dramatically.

Therefore, Cell Wall’s strategy is to advance both the interaction model and the openness of the AI interaction system simultaneously. “Even a refrigerator manufacturer could use this system to build a fridge that understands human language.” In Peng’s view, a closed ecosystem may not be suitable at this stage; openness is a key pathway to bring embodied intelligence into millions of homes. Compared with teams that focus solely on applications, Cell Wall also moves down to the hardware and interaction layers. Only by establishing a solid product experience first can an open ecosystem have a firm foothold.

“This industry is still in the early blue‑sea stage; peers should explore product forms together, each find target users, and jointly improve the interaction experience.” In Peng Tianfang’s view, the core of corporate competition lies in a deep understanding of how quickly user habits change: users’ willingness to pay, growth rate, which software they use, and what new habits are forming. This judgment brings the business logic back to product experience: first prove that there are people willing to use it daily on an ongoing basis.

In other words, for cell walls, the first commercial hurdle is whether “users are willing to turn ‘speak, edit, cut, approve, view’ into daily actions”; once this hurdle is crossed, universal AI interaction systems, open source, and vendor collaborations have a realistic foundation.

And the product must make users feel comfortable and effortless while helping them accomplish tasks. When this AI interaction system first runs smoothly on the desktop, embodied intelligence entering millions of homes will have a truly solid starting point.

Depth, insight, foresight, capturing business opportunities for the top 1%

Selected public accounts under 36Kr