Summary

Livis has delivered a first-wave demand signal far above expectations. The constraint now is yield and delivery, not customer acquisition. The first tens of thousands of units sold out in roughly 3 hours, and 3 days of sales consumed about 1.5 to nearly 2 months of production capacity; the share of non-car owners was cited at roughly 12% and later nearly 15% during the interview, despite almost no advertising. 范皓宇 insists on maintaining yield standards, accepting short-term supply pressure, while arguing that competition in 2026 does not warrant excessive attention: “The first thing is to serve our car owners well.”
36 grams, 18.8 hours, and 800 milliseconds are all achievable at once because the core work is hardware-software integration across the chip, two ATOS systems, and the operating system—not a layer of AI applications. Livis uses the Heng玄 BES2800 MCU, originally designed for watches, rather than the mainstream Qualcomm AR1 plus co-processor architecture; typical-use battery life is 18.8 hours, standby exceeds 70 hours and reaches roughly 78 hours, and question-answering response time is about 800 milliseconds. Li Auto also reused its Rust-based systems, power and thermal management, imaging, and acoustic capabilities from its cars. 范皓宇 calls it “the only glasses in the world currently built with pure ATOS and pure embedded development.”
Early user behavior supports the positioning of Livis as an “all-day companion,” not a camera accessory. Tens of thousands of users average at least 40-50 conversations with the glasses per day, versus 18-20 with 理想同学 in the car; the team therefore prioritizes daily wear time and intelligent voice turns over photo counts. Car-control features have roughly 30% daily-active penetration, while Q&A, visual understanding, note-taking, and Deep Research are moving the glasses from “control commands” toward communication and dialogue. 范皓宇 also cautions that launch excitement may temporarily inflate conversation data and needs to be tracked over time.
Li Auto’s glasses are not an isolated diversification bet, but a multi-device strategy to bring 理想同学 beyond the car. The project began in early 2024: prototypes first validated always-on companionship and memory, while car control later created a strong link to the core business. The operating system, streaming rearview-mirror imaging team, multi-microphone acoustics, and systems-integration capabilities all came from the automotive business. 范皓宇’s direct explanation is: “The moment you leave the car, it disappears. So why not bring it out?”
Li Auto is targeting the point when the technology moves from early adopters to a highly usable mass-market product, not the earliest demo window. That is why the first generation deliberately dropped forward-facing displays, cut weight from an initial estimate of 42-43 grams to 36.2 grams, reduced the microphone count from 5 to 4, and removed the folding sensor; the test is not the feature list but whether weight, battery life, response time, and social acceptability clear the bar. “High usability is what we are good at. Prototypes and demos are not.”
范皓宇 defines the organization’s core capability as “bottom-up authorization”: product managers use trust to mobilize the company’s full pool of expertise. Ordinary product managers divide their time according to “six-two-one-one”—60% focused execution, 20% proactive communication, 10% reflection, and 10% play—while managers reverse the first two allocations, spending 60% of their time understanding, communicating, and building trust. He defines the product manager as a second kind of star: “I ask the best people in the world to work for me.” Observation, learning, and the paranoia and ambition born of refusing to accept the world as it is are the fuel for continuing up the mountain.
The next leg of growth will come from agency and user co-creation, but only after the fundamentals clear the bar. 范皓宇 wants Agents to move from “push once, act once” toward proactively approaching users, remembering, reminding, and coordinating across devices, but warns that proactive behavior without sufficient capability becomes harassment. The first generation is not overly focused on the social dimensions of sales, profit, and influence; internal capability activation has already happened, while real user value will take another 1-2 years of joint refinement: “Let the wind and rain beat me from every direction—I will always focus on you.”

Deep dive

1. 范皓宇 reserves the highest product praise for “a knowing smile”

At 43, 范皓宇, a graduate of Xi’an Jiaotong University’s industrial design department, first brings up a proud creation that was not a car: a personal app he designed for Eason Chan more than a decade ago that created two timelines—one using the Gregorian calendar and one using Eason’s own era—to let the artist interact directly with fans. “It’s fun” explains his product sensibility better than any commercial outcome.
Among products that actually reached mass production, he chooses the L9 without hesitation. Two 3K, dual-layer bonded OLED screens and a series of details inside and outside the car pushed the spatial experience to a new level at the time. After an accidental road-test leak, the team was forced to publish one image and explain one feature every week, inadvertently creating a pre-launch communications cycle that kept the product on trending lists.
范皓宇 places the L9’s emotional value in the social context after the pandemic. Early adopters among technology workers in Beijing, Shanghai, Guangzhou, Shenzhen, Chengdu, Hangzhou, and Shanghai embraced the product because it made them feel “able to breathe again, refreshed and renewed.” “Like spring, stirring beneath the surface during Jingzhe,” became his example of a product resonating with its era.

2. Livis defines an AI glasses product from the fundamentals of wearing it

范皓宇’s priorities for Livis are explicit: at roughly 36 grams, it should approach the long-term wearability of ordinary glasses; typical-use battery life is 18.8 hours, while pure standby can exceed 70 hours; and a new voice architecture brings AI Q&A response time down to around 800 milliseconds. “Extremely comfortable” comes before feature count.
The concentration of similar products at CES, in his view, confirms that the industry broadly believes in endpoints that are long-term companions and effectively invisible in use. Ray-Ban Meta, Rokid, XREAL, and other teams did not enter suddenly either; they had spent years accumulating expertise in AR, display, and related fields. Different manufacturers are betting on sports, teleprompting, translation, or pure display, and have not yet fully converged on the same market.
What makes him proud is not the trade-show presence but the horizontal comparison table. Comparable products generally weigh more than 39 grams, usually in the 40-plus-gram range, while their chips commonly pair Qualcomm AR1 with NXP or another co-processor. Livis uses a single MCU and follows a distinctly different systems path.

3. A three-hour sellout shifted the problem from demand validation to supply delivery

The first tens of thousands of units sold out in roughly 3 hours, and another 3 days of sales consumed about 1.5 to nearly 2 months of production capacity. 范皓宇 called the result “far beyond expectations.” The launch budget and media volume had deliberately been kept narrow, initially targeting only Li Auto owners, but spillover traffic drove about 12% of purchases from non-car owners; discussing the broader hardware footprint, he later gave the rough estimate of “nearly 15%.”
As demand surged, the bottleneck quickly shifted to supply and yield. The team did not lower standards to accelerate delivery, so sales were temporarily paused or restricted. 范皓宇’s view is that after absorbing the supply shock and rebuilding capacity, the company should still serve car owners first, then external technology enthusiasts.
He is not anxious about competition in 2026. The underlying platform—real-time operating systems, low-power chips, long standby, and continuous OTA—still has room to expand, while different manufacturers continue to broaden their respective markets. “I don’t think about competition either” is not a denial of rivals, but a judgment that the immediate task remains delivering value to the company’s own users.

4. Conversation turns show the glasses becoming a companion rather than a tool

The team is focused on two north-star metrics: single-session wear time and the number of intelligent voice interactions. The first tests whether the weight and comfort are good enough for users to “weld it to their face”; the second tests whether 理想同学 is genuinely helping users rather than merely completing a few photo commands.
In the early data from tens of thousands of users, the glasses generate an average of at least 40-50 conversation turns per day, versus 18-20 inside the car. 范皓宇 acknowledges that launch excitement may be lifting the numbers and says the curve will be monitored, but users are already showing that they are “communicating and conversing with it, not merely asking questions.”
The first high-frequency use case is voice and visual Q&A: users see an object and ask directly, “What is this?” then ask the system to search, explain, or remember it. The second is car control, with roughly 30% daily-active penetration. Users can turn on steering-wheel heating, seat heating, and air conditioning while leaving an elevator, then open the trunk by voice when they reach the car.
The notebook can recall information users have left behind in later conversations. Recording-time reminders were not yet complete; the team expected to add them in about a month and might connect the feature to calendars. Deep Research runs asynchronously for dozens of minutes, and average users are already using it several times per week.

5. The glasses grew out of a 2024 prototype rather than a last-minute attempt to chase a trend

The project began taking shape in early 2024. A prototype built with a partner around midyear could reconstruct the context of daily life: if a user entered a room and left their phone on a table by the door, then asked 3-5 minutes later, “Where did my phone go?”, the system could answer based on continuous sensing. That gave the team a concrete understanding of what always-on companionship and memory required.
The initial bottom-up driver was the team’s shared interest in a new form factor. The top-down judgment was that once AI entered everyday life, “a new product form would inevitably emerge.” But the link between the glasses and the car was initially weak. The project gained a much clearer strategic connection only after employee testing showed that one-sentence car control was remarkably convenient.
By September and October 2025, as the product approached maturity, employees in internal testing were describing car control as “so comfortable—I just say one sentence and it’s done.” 范皓宇 also imagines the car proactively describing its state: “I’m getting roasted in the sun. Could you move me somewhere?” in the heat of summer, turning the relationship between the two endpoints into “two living beings communicating.”

6. The 36.2-gram result came from repeatedly rejecting “good enough”

The first structural stack-up was estimated at 42-43 grams. 范皓宇 rejected it outright and set a target of 38 grams. The team did not simply subtract weight from a spreadsheet; everyone asked themselves: “If I had to carry an extra 40 grams on my face every day, I definitely couldn’t accept it.”
Weight reduction reached down to adhesive. Ordinary frame glue is generally black, so the team developed a lighter, thinner transparent adhesive specifically for the product. The battery, substrate, sensors, and even microphones weighing fractions of a gram were all reconsidered. Mass-production weight ultimately landed at 36.2 grams, beating the original 38-gram target.
At 3 p.m. on the day of the final-form review, 范皓宇 demanded an immediate 3D print. By 7 or 8 p.m., he had a model whose engineers were sanding off burrs while running tests. He wore the empty frame back to his lodging, took it in the elevator, and wore it through breakfast the next morning before showing it to 李想: “This is the final product we need to build.”
The coincidence was that the early printed model also weighed 36.2 grams. Three months later, after all the structural and process work had been consolidated, the final product returned to the same number. He attributed the result exceeding expectations to a shared goal, not “one person’s cleverness.”

7. The rebuttal to “lack of focus” rests on reusing automotive capabilities

Asked why an automaker should make glasses, 范皓宇’s answer is: “We have always been focused on user experience and creating new value.” The glasses did not require building a separate giant organization; they reused large amounts of accumulated capability in automotive systems, imaging, acoustics, power, and thermal management.
The glasses’ operating system grew out of the efficient system developed over 2-3 years for the steering-wheel Touch Bar and was rewritten in Rust. The team responsible for the ultra-high-definition streaming rearview mirror moved into glasses imaging, while experience with multi-microphone systems and spatial audio came from cabin acoustics. The point was not to open an entirely separate business line, but to let the same capabilities compound across endpoints.
Reverse migration matters just as much. The glasses imposed more extreme requirements on microelectronics, high integration, materials, process engineering, and heat dissipation. When participants returned to automotive projects, they were already thinking about new approaches to centralization and thermal management. 范皓宇 calls this cross-category feedback loop the real “chemistry.”

8. Heng玄 BES2800 and dual ATOS represent the most aggressive technical bet

Livis did not use the conventional MCU plus Qualcomm AR1 main-chip combination. Instead, it elevated the Heng玄 BES2800 MCU, originally designed for watches, into the glasses’ main chip. 范皓宇 believes this was the only way to achieve roughly 78 hours of standby, 18.8 hours of typical use, and full AI capability simultaneously.
Li Auto and Heng玄 worked closely for nearly 6 months, jointly developing the mainline glasses system and baseline. The chip and operating system had to be tightly coupled to achieve the power and performance results. The cost was that the team could not simply build applications on a mature platform; it had to work through compilation, feature optimization, and low-level tuning.
Another ISP chip, originally designed for surveillance scenarios, offered good power consumption and image quality but did not start quickly enough. The team therefore worked with the supplier to develop a separate ATOS for the ISP. The main MCU and ISP each run their own ATOS, leading 范皓宇 to call Livis “the only glasses in the world currently built with pure ATOS and pure embedded development.”
Both sides described the schedule during project discussions as “mission impossible.” The partner team had just returned from a closed-door delivery in Shenzhen and barely had time to recover before entering the new project. All the engineers judged the route feasible, but 范皓宇 used hiking as an analogy: planning a feasible route and actually walking it are separated by a very painful process.

9. Bottom-layer innovation did not erase weaknesses; it gave the team room to keep optimizing

范皓宇’s biggest disappointment with the first generation is Wi-Fi transfer speed. The Wi-Fi module in this chip architecture had bandwidth constraints from the beginning, embedded in the architecture choice. The team improved stability, speed, and efficiency through software, but he did not present the known weakness as solved; he simply said it would “continue to be optimized.”
The host’s actual experience was more favorable: Livis connected more easily and transferred data faster than the recording card previously used, so he had replaced the card with the glasses. The disagreement preserves two benchmarks: relative to the user’s prior experience, the product is usable; by the product owner’s higher ceiling, it is still “not fast enough.”
Acoustics are handled by the team that built in-car spatial audio and multi-microphone systems. The exterior curves were modeled by automotive designers and engineers according to a continuous G3 standard rather than by simply rounding corners. Transparent materials once weakened the visual light effect designers had imagined, although the host felt that the effect still came through in darker surroundings.

10. “Getting both hands dirty” means going into the stream and denoising details

范皓宇 defines “mud” as the gap between an imagined result and reality. When the shooting effect and IMU stabilization were considered difficult to solve, he did not merely listen to reports. He led the team through reviews of the logic, implementation, and code item by item, and went to Wuhan to break down the issues in person with the Black Sesame R&D team.
The key decision that night was to use a higher bitrate and remove pre-processing denoising, leaving the burden to post-processing. This could increase the risk of delay, but it avoided the image-quality damage caused by coupling two rounds of denoising. 范皓宇 framed the choice this way: “With my left hand, I can make something we don’t believe in; with my right hand, we might be delayed. I don’t care about the delay, but we have to do the right thing.”
The approach affected ISP post-processing, imaging, and system transmission. The original test cycle was long, but the team ultimately compressed feedback into same-day visibility. The project was not delayed, and image quality and IMU jitter reached the best solution the team could find at the time. The purpose of the person in charge going deep into the details was not to challenge people, but to “inspire everyone.”

11. Automotive is constrained formal prose; glasses are a self-bounded essay

Automotive has a century of history, regulations, and safety constraints, so innovation must fit inside an existing structure. Yet Germany, the United States, and Japan each wrote different versions of the same formal script. 范皓宇 believes China’s auto industry is now inheriting existing know-how and writing its own article. Behind “refrigerator, television, and large sofa” is ultimately an affirmative choice in favor of comfortable living.
He uses an encounter outside Apple Park to explain cultural expression. He was driving a MEGA when a Cybertruck approached from the opposite direction. The other driver lowered the window and gave him a thumbs-up; he returned the gesture. The Cybertruck represents a typical form of American culture, but Chinese products can participate in the conversation in their own way.
Glasses do not even have a mature baseline, making them more like free-form prose. The danger of freedom is that a team can lower the standard to 45 grams and then comfort itself with a long feature list. Li Auto’s self-discipline from building cars is “high usability”: protect weight, battery life, and response time first rather than using features to hide weak fundamentals.

12. Dropping forward-facing displays was a maturity judgment, not a rejection of the route

范皓宇 believes head-up displays are “not quite ready” at this stage. A screen in the main field of view can block the physical world, where information density is highest, while people opposite the wearer may see strange rainbow patterns through the lenses, damaging face-to-face interaction. Teleprompting and other specialized needs exist, but he believes the technology as a whole is not ready.
He has spent nearly 15 years with MR, AR, and VR, and experienced products before Oculus was acquired by Meta, so he is not opposed to displays as such. His actual judgment is that placing a screen directly in front of the eyes is “very computer, very engineer,” and risks turning glasses into a monitor attached to the face.
Meta’s side-mounted monocular display experiment struck him as “pretty cool” because it treats the display as an aid rather than occupying the center of vision. It may still be difficult to match when the prescriptions of the two eyes differ. Li Auto will continue researching the area and integrate it with new interaction and design approaches once the technology reaches high usability.
The same subtraction logic reduced the microphone count from 5 to 4. The acoustics team used simulation and real-world testing to confirm that spatial audio remained achievable before cutting the fractions of a gram. The folding sensor was also removed, with software used to optimize wear detection. Every deletion required the capability not to disappear with the hardware.

13. A Chengdu customer-service shift exposed the delivery chain beyond the product

On the night sales opened, traffic and consultation volume overwhelmed the front-line team in Chengdu. Some colleagues cried because they had been underprepared and felt they had let users down. More users continued calling after the 3-day sellout. Some, trying to secure an order, casually wrote their prescription data on envelopes or sticky notes and uploaded photos, creating a large volume of errors and repeated communication.
范皓宇 spent 3 hours doing customer service in Chengdu and received roughly 17 groups of users, more than 10 of whom were chasing their orders. He had to locate orders in internal systems and press for actual delivery dates before he could see directly that after a product is finished, leads, prescription data, customer service, and fulfillment jointly determine the user experience.
His public conclusion was an apology to Livis users: “Our customer service was not good enough this time, and our initial preparation was not sufficient.” Borrowing the phrase “a gentleman does not make the same mistake twice,” he promised that the next sales round would bring service capability up to the same level as product capability.

14. “Six-two-one-one” breaks product work into focus, communication, reflection, and play

For ordinary product and engineering staff, 范皓宇 recommends spending 60% of the time focused single-mindedly on execution, 20% proactively communicating across fields, 10% reflecting on whether decisions came from fear or desire, and the final 10% playing. Proactive communication is not passive meeting attendance; it means finding people from different disciplines to solve a problem together.
“Play” is not simply a search for inspiration, but a way to regulate the creator’s state. He believes a product made by people who look pained and exhausted will carry a bitter taste. Only when the people making it can keep their feet in the mud while looking up at the blue sky and feeling that the work is fun can the product carry a smile.
At the director or senior-manager level, the 60% and 20% should be swapped: large amounts of time go toward understanding, communicating, and building trust, while 20% remains dedicated to personally driving specific modules. Product managers neither write the code nor draw the designs, so they cannot secure real commitment through commands alone.
He distinguishes two kinds of stars. One becomes a “god” in a single field whom everyone goes to for help; the other says, “I ask the best people in the world to work for me.” Product managers belong to the second category. They route dispersed expertise toward one result through the heat of their communication, an understanding of difficulty, and fighting alongside the team.

15. What product managers really need is bottom-up authorization

The host calls the product manager a “CEO without power”: there is no inherent administrative authority, yet the role must get R&D, design, supply chain, and marketing to accept a direction. 范皓宇 draws a further distinction: an appointment letter from a superior grants top-down power, while a team’s willingness to provide support is bottom-up authorization.
Both forms of power can coexist, but he considers the latter more important for product managers. If execution is demanded solely through formal authority, teams can deliver perfunctorily. Only long-term trust-building leads people to keep adding effort through difficulty and uncertainty.
“Getting both hands dirty” is therefore also a trust mechanism. The person in charge cannot draw a vision in the air and leave others to make it real; they must enter the front line and face the gap between ambition and reality together. He notes that Steve Jobs was not only responsible for imagining things either, but spent large amounts of time communicating direction and personally pushing details forward.

16. Observation, learning, and ambition form the product manager’s three layers of fuel

The first capability is observation, because “if you cannot see it, you cannot possibly build it.” At a restaurant, 范皓宇 noticed black blocks on the ceiling and kept asking employees across several floors before learning that they were positioning antennas, each covering 3-4 tables and working with an RF device on the tray to help servers deliver food. Many people passed through without ever noticing the system.
He uses Sherlock Holmes to make the point: dirt on a shoe sole, hair on clothing, fingernails, a person’s eyes, and what they wear are all elements for understanding an event. Abstracting the world into “one-two-three” is convenient, but erases its “soup and water”—the texture of lived reality. Opportunities are often hidden in people’s specific pain, delight, and details.
The second capability is rapid learning. Product managers must cross front-end engineering, AI, materials, process engineering, supply chain, manufacturing, and marketing. They cannot be deeper than the expert in every vertical, but they must understand boundaries and pain points through listening. The goal of communication is not to issue orders, but first to say: “I understand. Now I’ll figure out a way.”
The third is paranoia or ambition, which he sees as “isomorphic”: both are driven by refusing to accept that the world has to remain this way. The title of Camus’s book The Invincible Summer in Me resonates with him because that heat sustains product people as they face the “impossible,” trying to become “the one, the only” and “make something happen.”

17. AI has not rewritten the essence of product management, but it has made the capability boundary more winding

范皓宇 does not believe an “AI product manager” needs a different personality. Just as a WAP product manager entering the mobile internet had to relearn after old tools stopped working, the basic method remains the same: observe the user’s environment, understand the technology’s underlying drivers, and use the product as a bridge to move that capability naturally to the user.
An AI product manager must understand 3 objects: the dataset, the model architecture, and the evaluation set—what data passes through what structure to produce what result, and how the result should be judged. A model is like a growing child: abilities emerge step by step, from buying things to conducting research and handling a project, rather than becoming available in one clean release.
AI’s boundary line “winds and turns.” It may be weak in one position and surprisingly strong nearby. Understanding the boundary allows teams to play to strengths and avoid weaknesses. Once a model fills an old gap, the changing boundary itself becomes a new product opportunity rather than something to wait on until an abstract “general capability” matures.

18. “4o” moved voice from a command box to continuous conversation

Before Children’s Day in 2025, Li Auto launched the “4o” real-time conversation capability. The team had initially placed it in a corner of the in-car laboratory, but discovered that many users kept using it over time. That showed voice interaction did not always have to be a clearly defined query such as navigation or music.
Previously, users would first think through how to command a tool, then say, “Navigate to…” Under the 4o framework, users could communicate freely. Low latency combined with prompt-based persona design began to create “a conversation setting, not a command setting.”
This inflection point later became the infrastructure for conversations on the glasses, and explains why the team places conversation turns ahead of photo counts. Even if a photo is triggered by voice, it remains a control command; whether the user can keep communicating determines whether the device becomes an assistant, secretary, or companion.

19. Agency, user co-creation, and three definitions of success set the next stage

The first priority remains basic performance: battery life, response speed, and transfer efficiency cannot slow down as OTA updates and features accumulate. “The product has become slower” is unacceptable to 范皓宇. Weight is already fixed, but the other underlying metrics still require continuous optimization.
The second stage is agency. 范皓宇 uses Agency to describe proactive behavior, but proactivity must be built on sufficient capability. If capability is inadequate, proactively approaching users only means “proactively coming over to find bugs and create trouble,” turning companionship into harassment. A car heated to nearly 80°C proactively asking someone to move it is his positive example.
He says it is highly likely that he will push for user participation that year, allowing a group of users to participate directly in developing the car, glasses, and model capabilities. “Do we finish the painting and show it to people, or paint part of it and paint the rest together?” The latter is the product relationship the team wants to establish.
Success is divided into 3 layers: the social layer of sales, profit, and influence; the company layer of strategy, tactics, and capability activation; and the team-and-individual layer of creating new experiences. The first generation does not pursue mass advertising or social success. The company has already demonstrated the possibility of an integrated product, while user value still requires 1-2 years of joint refinement.

20. Li Auto’s in-house models and the “AI continent” explain its move toward multiple endpoints

Livis is backed by Li Auto’s in-house end-to-end voice and multimodal model stack. The interview first describes it as divided into OA and OVO, then refers to the visual-plus-voice streaming Q&A module as OV. 范皓宇 acknowledges that in-house development first means “spending a lot of money and a lot of resources,” but only by controlling the models can the company merge endpoint capabilities into a unified Agent instead of writing rules one by one.
His full-stack logic covers chips, operating systems, models, application services, and cloud communications. The VLA for assisted driving and 理想同学 in the cabin are also built on an end-to-end architecture. The value is that when users expose a problem or demand a new capability, the team can adjust from the bottom layer and create further chemistry among multiple endpoints and multiple users.
In the “continent” metaphor formed at the end of 2024, the core is the model that determines the upper bound of capability; the second ring is endpoints such as cars and glasses; and the outer ring is applications and services. In the ocean flow professional capabilities such as lawyers and LBS. A-to-A transactions and collaboration among Agents will become more efficient, while the number of ports depends on the scale of continent users and endpoints.
Apple, Huawei, and Xiaomi are different continents in this metaphor. Tesla demonstrates the combination of models, endpoints, applications, and services, while Tencent resembles a maritime hegemon. Li Auto sees itself as merely an island beginning to surface, grounded in several million car owners and their families; the nearly 15% of non-car owners attracted by the glasses form a new protruding section. The goal is not “multiple endpoints for the sake of multiple endpoints,” but expanding a seamless experience at the right moment.

21. Familiar hardware, high usability, and long-termism define the boundary of the next bet

Li Auto chose glasses before phones because glasses can make information and technology surround you. Phones still require users to take them out and unlock them, and 范皓宇 believes 90% of their functions belong to the previous era. Glasses, like cars, also express taste and identity, making them a natural fit for Li Auto’s existing capabilities in integration, form, and aesthetics.
He rejects ranking glasses, earphones, and pendants in investor-style top-one and top-two order. Each team is choosing a path based on the materials it has in hand. Personally, however, he prefers AI activating an existing form: a pendant adds wearing and charging burdens, while a pen already used for recording could naturally add a microphone and IMU, as could a telescope adding recognition and recording.
Robots are an explicit direction, but remain far from the high-usability threshold. The ideal home robot would behave like a housekeeper: when the user comes home, the place is already tidy, and the executor is preferably not in sight. Current training data is still too sparse; teams cannot declare victory after spending 6 months or a year assembling open-source frameworks, supply chains, and dancing demos.
Glasses have accumulated roughly 2 years of work, while automotive has taken 10 years; robots need an even longer period of “deep squatting.” 范皓宇 defines Li Auto’s ideal window as the point when technology crosses from early adopters into the early majority. He closes by describing himself as “literary, curious, bad-tempered, impatient, fond of bringing up children, fond of drawing and writing, a router, a fox, fond of going into the mountains, and passionate,” then invokes 施展’s Mountains and Rivers: mountains and rivers shape settlement, stories, and culture, and products will inevitably carry Chinese people’s understanding with them. “As long as we still have this breath in us,” technology and the humanities can continue creating new possibilities together.

Adam Buckstein's Stride Thesis $LRN Tobi Lütke: 21 Years of Building Shopify