Summary

张翼’s core timing call: Robots are more worth getting an early start on than pure AI because software-hardware integration iterates slowly and first-mover advantages compound. “Can you spend RMB1B today and build a robot—possibly like 宇树’s—with locomotion this good? Impossible… It’s not a money problem; it’s a time problem.” He therefore used his and his team’s own money to “hold our breath for 3 years” and get the product to consumer grade. He only raised external capital at the end of last year, closed 2 rounds within months, and should already have put robots in 300 Shanghai homes.
The commercial anchor is childcare, not housework. In Shanghai, a domestic helper costs RMB7K-8K/month and handles both light and heavy chores, while a robot costing RMB20K-30K cannot compete. But a childcare nanny who speaks English and plays chess can cost RMB30K-40K/month—and may be impossible to find. F2 enters through childcare plus light housework and is currently offered as a RMB3K-4K monthly trial rental; early units needed servicing after 1-2 days, while now they need service only after 1-2 months and users still use them for several hours a day.
Data is the ultimate moat; the brain is not. Competitors can reproduce a good model in 1-2 months, while data is slow and scarce (Nvidia’s Dream Zero is only 12B—or “some number in the teens of billions”). Of the 3 data types, the decisive ones are corner cases and “interaction with living things” that data factories cannot capture; 80% of home scenarios involve interaction with people. The strategy is to place robots in homes at cost early and get the flywheel turning.
Every technical trade-off points to consumer-grade cost. He chose wheels over bipedal legs for indoor safety and stability, with a larger battery and longer runtime; a 2-prong gripper over a dexterous hand (“you can’t find a dexterous hand anywhere in the world that can last more than 2 months,” with an interface reserved for a future swap); and full-stack in-house development that brings joint cost to about RMB1K versus RMB8K from third parties. Industrial-grade specs are cut away: “good enough for the home—OK, ready.”
World models gave him a genuine surprise. Zero-shot handling of things it has never seen shows “tens of percent” robustness—something VLA previously found extremely hard—and is “more advanced in approach.” Robustness is nowhere near 100%, and “much of it is still at the paper stage,” so both paths run in parallel; only in the past 2 months has the company brought in many world-model specialists.
His bubble view: there is a bubble, but the mania has not reached full intensity. Online education companies once raised $1B-$2B per round; today, embodied-AI rounds with the same headline figures are denominated in RMB, a 10x gap. The bubble will burst, but the industry will move “upward through twists and turns,” ending with “every company in the world becoming a robotics company.” Crossing the cycle requires a sustainable business model; “losing several billion yuan a year is easy.”
Double Reduction’s management lesson: anxiety makes strategy short-term. “I may win this month, but lose the next 2 years.” Zhangmen’s 1-on-1 business held 70% of market share, yet it fought on 2 fronts after group classes appeared and lost the larger market; failure “is actually an external force helping you make a choice.”

Deep dive

1. Double Reduction Wiped Out the Business Overnight: From More Than 90,000 to Under 1,000

The news arrived while he was in the car. An investor forwarded a document and asked, “Is it real?”—“At that moment, everything went blank.” It took half a month to confirm it was real: he first assumed it was temporary and would require only a small reduction, until a key investor made it explicit: “This is real.”
The numbers were brutal: the company had grown to RMB10B in scale that year, with enormous monthly outlays; “once there was no revenue, the whole company would blow up.” An investor warned, “You can’t drag this out any longer. If you do, I won’t be sitting here today.” In 2.5 months, headcount fell from more than 90,000 to under 1,000, and a tuft of white hair appeared at the back of his head.
The hardest image was this: across branches in 50-plus cities nationwide, HR teams said goodbye to everyone, then shut off the lights and walked out; meetings of hundreds went on one after another, with executives breaking down crying mid-speech.

2. After 1 Year of Drift: AI, Hardware and the Home Intersect

For a time he was lost: he had no reason to go to the office for 6 months—“the company had nothing to do, and I didn’t dare go either.” His conclusion was that life should still take part in at least a small piece of the world’s transformation; “even a tiny contribution is meaningful.”
He went to Silicon Valley to talk with Jiao Tong classmates, when ChatGPT was at 3.0 but not yet 3.5, and thought, “There was an opportunity.” Add his electronic-information-engineering background, the hardware base from building smart-car electronics designs as a student, and years of education experience understanding children and family users: these 3 strands pointed to family robots, and it “had to be bigger than before, or I’d have no motivation.”
The line that finally convinced him: “20 years from now, every household will buy a robot.” The key was that no one was doing it yet: “Whatever I do, it must not be something many people are already doing; otherwise adding me has no meaning.”

3. Why Robots, Not AI: The Compounding Time Advantage of Software-Hardware Integration

The core logic: AI is pure algorithmic software and iterates too fast, so any early-entry advantage gets erased; once hardware is involved, the pace slows dramatically—“Can you spend RMB1B today and build a robot—possibly like 宇树’s—with locomotion and stability this good? Impossible… It’s not a money problem; it’s a time problem.”
That led to the plan to “hold our breath for 3 years” (先憋三年): use his and his team’s money to keep the team running, build quietly without telling stories, bring the full machine to consumer grade and productized, and come out only once it had an aha moment and was genuinely surprising. The second reason came from his previous rise and fall: “If I raise again, I need to be very sure this can work; otherwise there’ll be a mess to clean up.”
He did not raise on the market until year-end, when he had confirmed the product could enter homes and was popular in trials; he then closed 2 rounds within months. The team has about 100 people, mostly engineers.

4. To C Only: The Positive Feedback Loop of Brand Premium

He does not get hung up on industrial versus home use: “A lot of To B technology ends up without much of a moat; in the end it becomes a price war and is hard to make money.” In cleaning robots, To C companies earn very high profits, while To B sales in China are “a real struggle.”
The To C flywheel is experience → brand recognition → the brand itself commands a premium → profits reinvested into better products: “You form a positive feedback loop. I’ve always worked To C; this feedback model suits me better.”

5. 300 Households Should Have Participated in Co-Creation

They each sent out 5K-6K questionnaires and conducted phone interviews in China and North America, but did not settle the feature trade-offs until the end of last year (about 3 months ago): “He said folding clothes was worth RMB500 and sweeping the floor RMB2K; once you added them up, they still didn’t cover the machine’s cost.” The product definition had been stuck for a long time.
Home rollout came in 3 stages: first, restrict the robot to a few tasks; then let it play freely; finally, let users drive the feature requests. Chess came from users—“Kids stare at a computer playing chess for 2 hours and their eyes are ruined”—and the clamp-style 2-prong gripper lets it play different board games. Children played hide-and-seek every day, so the team strengthened it into a dedicated feature.
Retention is the hard KPI: when robots first entered homes last year, a service visit was needed after 1-2 days; now maintenance comes only after 1-2 months, while users still use them for several hours every day. He watches renewals and referrals—“a 50% referral rate, plus another 30%”—because that shows whether the product is good, not how many units it sells today, which could just be a one-off wave.

6. F2’s Pricing Anchor: A Childcare Nanny at RMB30K-40K a Month May Be Impossible to Find

F2 has 2 functions: childcare and light housework. It reads picture books, keeps children company during piano and violin practice, identifies melodies by ear, corrects fingering and posture in real time, converses, plays chess, looks for LEGO pieces, tosses a basketball, tells bedtime stories, turns off the light and leaves. Light housework means putting toys away and picking up trash; the heaviest chore is the kitchen, which he thinks is still several years away.
The pricing method is simple: “The price of a product depends on what you benchmark it against.” Benchmarking against a housework helper doesn’t work—RMB7K-8K/month in Shanghai, covering light and heavy chores, versus a robot costing RMB20K-30K. But a childcare nanny who speaks some English and can play chess may cost RMB30K-40K/month and still be impossible to find; childcare is where the value case holds.
It is currently rented at RMB3K-4K/month. After launch, the model may be “cheaper to buy outright, with a subscription to unlock higher-end functions,” since some functions consume compute. The elder-care study produced a surprise: retirees said, “Help me with the kids”—not companionship for the elderly, but help with grandparents’ grandchildren. People over 80 are a separate segment, needing care and kitchen functions.

7. Surprises and Scares: A Robotic Arm for Play, and “Dog Duty”

The biggest surprise was “playing with children,” a use case the research never surfaced. Because of the robotic arm, “none of the toys at home have hands; this is the first time the child has a hand.” A toy car can zip out, be retrieved and zip out again; it can keep a child occupied for 2 hours, with a different way to play possible the next day. The best interaction window is from middle kindergarten through elementary school.
The scare became a feature: one family’s dog “couldn’t sleep all night,” and the owner said, “Forget it.” The team later developed “dog duty” (搞狗)—feeding the dog and teasing the cat like a feather duster, keeping the pet occupied for 1-2 hours. His abstraction: “A family robot looks like a product, but what it really sells is a service.” Services are expensive and hard to standardize; physical AI can standardize them.

8. Wheels and 2-Prong Grippers: Everything Gives Way to Consumer Grade

Wheels versus bipedal legs have different strengths and weaknesses, so he will not generalize, but indoors wheels win on value for money, safety and stability. They brought a biped home to test; when it was lifted, its feet started flailing, making it unsuitable for a household. Most Chinese homes are single-level, so wheels are enough; a weighted base allows a larger battery and longer runtime.
He does not build a 5-finger dexterous hand for empirical reasons: “You can’t find a dexterous hand anywhere in the world that can last more than 2 months.” The joints are too small and break under daily impacts. The solution is a 2-prong gripper with a reserved interface: “Once dexterous hands mature, we can swap one in quickly. Many companies are working on this, so we won’t.”

9. Full-Stack In-House: RMB1K Joints vs RMB8K from Third Parties

Starting with hardware self-development in 2022, the effort gradually became full-stack: motors, torque systems, joints, the full arm, the gripper and the complete robot. The secret to low cost is cutting specs: “This much repeat-positioning accuracy is enough; this much payload is enough.” It does not need to support heavy-load handling in factories; specs are cut to the limit—“enough for the home, OK, ready.” Hollow, large-aperture cable-routing joints are not available off the shelf, so they had to define them in-house.
The bigger benefit is access to every parameter: the current loop and the actual state of every joint. Third parties will not expose that many interfaces; only with all those parameters can the team tune the system to the optimum. The hardest in-house part is the drive: integrating high-density components into a palm-sized package while managing heat. Self-development has limits—gearbox manufacturing had too many process problems and was abandoned, while depth cameras could never be fully self-developed.

10. World Models: More Advanced in Method, but Both Paths Run in Parallel

World models delivered a real surprise: zero-shot handling of things never seen before shows “tens of percent” robustness, something VLA had found extremely hard. Nvidia’s Dream Zero has some reasoning ability and can predict object trajectories, while training is lighter: feed it videos, put a camera on the head and add hand operations.
The caveat is explicit: robustness is “not 100%,” and “much of it is still at the paper stage.” VLA is still iterating and cannot simply be dismissed, so both paths run in parallel for now. Only in the past 2 months has he brought in many world-model people: “There aren’t many people doing this in China; finding this talent matters.”

11. The 3-Way Data Split: Corner Cases and “Living-Thing Interaction” Decide the Winner

The brain-data asymmetry is clear: competition in the brain is intense, but it is “easy to copy”; a good model can be reproduced in 1-2 months, while data takes time and has historically been scarce. Nvidia’s Dream Zero is 12B—or “some number in the teens of billions”—which suggests its data volume is genuinely small. The endgame: “If we fold the same piece of clothing, mine is better because I have more home data than you.”
Data falls into 3 buckets: standard data that data factories can collect; corner cases—“the table is missing a leg while you’re folding clothes,” endlessly variable and learnable only in real homes; and the most overlooked category, data from “interaction with living things” (与活物互动). “80% of what we do at home is interaction with people”; ages, races and genders interact differently, otherwise the data covers only 1% of the world.
On US peer Sunday, the surprise is that a company built in the US is also doing hardware. Camera placement, exterior design, a “Umi” approach to source materials and the ACT solution “do reflect a lot of its own thinking,” but with no product on sale and no open-source release, “it’s hard to evaluate in practice.”

12. At-Cost Scale, Subsidy Wars and the Bubble: Still Short of Online Education’s Mania

He is willing to price aggressively: “Early on, we may sell at cost.” Others have to spend heavily to collect data; at least selling at cost gets the flywheel turning, with data as the moat. A large subsidy war would require someone else to reach full-stack consumer grade first; he thinks that may still take a considerable amount of time to reach their level.
The heat check: online education companies once raised $1B-$2B in a round; he himself raised at most a few hundred million dollars in one round. Embodied-AI financing now uses the same figures in RMB: “It used to be $1B; now it might be RMB1B.” The mania is 10x lower.
His bubble view: “There is definitely some bubble, but without a bubble it is hard for an industry to get off the ground.” The bubble runs ahead of the industry and draws in talent. The end state is foreseeable: “every company in the world will be a robotics company”; a convenience-store owner will be a person, with many robots doing the work below. Getting through the cycle will not be solved by raising RMB2B today; it requires a sustainable business model, and “losing several billion yuan a year is easy.”
The biggest obstacle to reaching thousands of homes is the first flywheel: getting robots into homes, making people like them and keeping them in use. The threshold depends on how much data, and what quality, a model upgrade requires, and whether dirty data can be used. He is not worried about consumer acceptance: when the robot played chess in a park, older people crowded around 3 deep and security guards had to chase it away every time.

13. Anxiety Makes Strategy Short-Term: Win This Month, Lose 2 Years

Cofounder Louis says his entrepreneurial style has “completely changed.” He sees 2 differences: he looks farther ahead and, once he starts, persists longer; and he watches the ups and downs from the perspective of a third party and of the future. “If you look back from the future, this thing is really no big deal”—the boat has cleared ten thousand mountains.
His answer on whether entrepreneurial anxiety is a quality cuts deepest: “When I’m anxious, my strategy becomes short-term. I choose the best answer for now, not the best answer for 2 years from now. So I may win this month, but lose the next 2 years.” Zhangmen’s online 1-on-1 business held 70% market share, but when group classes launched, shareholders split and the company fought on 2 fronts: “You won a battle, but lost the bigger market.”
His message to entrepreneurs who have hit setbacks: “The setbacks you encounter are often just forcing you to do something you truly need to do.” Failure is an external force helping you choose; “we have to make peace with ourselves to live better.” He almost never says, “If only I had known.” His 10-year message is a hope that he will “still be able to believe that a beautiful future is always just ahead.”

Verification Notes

Raw captions render the comparison entity as “语速”; “可能是宇树” remains an unresolved name identification.

广密 on Coding as AGI's Second Act and Models as the Next-Gen OS Elon Musk vs. Sam Altman, AI Job Loss, and OpenAI’s $852B Valuation | EP #247