Vol.196 Industry Watch 37 | How Robots Can Bridge the “Last Centimeter” of Interaction with the Real World
Summary
The core investment thesis for dexterous hands is not “upgrading grippers,” but whether embodied intelligence can cross the final centimeter of physical interaction with a general-purpose end effector. Wei Dehao argues that in the real world, “the only thing with truly general manipulation capability is the human hand”; it covers the widest range of tasks and is also the most accessible data format in internet video and human-motion capture. If robots remain limited to two-finger grippers, even the strongest models will be capped by both hardware capability and the mapping from data to action.
Among the four technical paths, Xingji Lightyear is betting on tendon-driven systems because they are the most likely to push degrees of freedom, payload, speed, precision, size, and weight toward human-hand performance simultaneously. Joint direct drive is constrained by miniature motors; linkage systems are bulky and weak against impact; artificial muscles are “expected to need 15 to 20 years” to mature. Tendon drive trades tendon-like transmission for compliance and light weight. The cost is limited tendon life, creep, and control complexity; the team tested about 30 types of tendon, developed its own Python Cable, and built an automatic tensioning structure to reach roughly 500K to 800K cycles.
The real competitive moat is not a choice among mechanics, actuators, and algorithms, but a flywheel of joint hardware-software iteration. Yan Qianhang’s view is that “speed is the ultimate advantage”: algorithmic progress exposes hardware flaws, while hardware upgrades expand the algorithmic ceiling, and reliance on outside suppliers on either side slows productization. Wei therefore calls the dexterous hand “the first hardware in human history defined by AI”; the deliverable should be an integrated software-hardware manipulation platform, not a bare piece of hardware.
Commercialization has not yet reached the eve of an application breakout; today’s high-DOF dexterous hands are sold first to researchers, embodied-intelligence teams, and seed users. There is still no particularly mature model for controlling 15 to 20 active degrees of freedom, so customers mainly study grasping, piano playing, cracking walnuts, and similar tasks. Demonstrations can land first through repeated validation, while industrial and consumer applications remain exploratory. The real inflection point will come when integrators can get started “very quickly and very easily,” opening the way for textiles, flexible assembly, precision smartphone assembly, and service-sector manipulation.
Large models handle task understanding, while a cerebellum model must handle fast, high-real-time reactions—or “the coffee will have spilled long ago.” Wei says a 10B-scale VLA model runs at roughly 3 to 10 Hz; the loop from visual feedback to judgment and execution can take 300 milliseconds to 1 second, making it unsuitable for high-real-time operations such as collision response. Many current demonstrations are merely “using the dexterous hand as a two-finger gripper.” His approach is to integrate a lightweight cerebellum directly into the hand, while emphasizing that smaller models do not make development proportionally easier.
The data bottleneck also requires hardware-software innovation: teleoperation is “accurate but slow,” while data gloves are “fast but inaccurate.” Teleoperation may produce only a few hundred samples a day, far short of the scale needed for hundreds of millions of data points; gloves can capture thousands or even tens of thousands, but introduce mapping errors between the human body and the robot hand. Wei proposes a dual-mode inner- and outer-exoskeleton system in which a person guides the dexterous hand “hand in hand,” aiming for intrinsic data that is both fast and accurate. He also plans to offer the open-source Gaia Hand at no more than one-fifth the price of products with the same degrees of freedom, expanding the developer ecosystem.
China’s comparative advantage lies in engineering from one to 100 and one to 1,000, while overseas players still lead in zero-to-one exploration and large-scale capital support. Yan says Shenzhen’s supply chain can deliver a revised circuit board the same evening, making its iteration speed far superior to trans-Pacific collaboration. U.S. companies, meanwhile, are more tolerant of frontier research, illustrated by Figure’s $1B financing and Physical Intelligence’s open-source models. Yan sees no major hardware bottleneck for dexterous hands in China overall; the more urgent gap is tactile sensors that are cheap, usable, and durable.
Cost declines and reliability maturity will determine whether dexterous hands can move from RMB40K-50K research equipment to a standard robot component costing a few thousand yuan. Non-standard machined parts, motors, and circuit boards currently account for more than 95% of Xingji Lightyear’s hardware cost. Mass production still has to clear the hurdles of coordinating hundreds of parts, selecting suppliers, ensuring stable delivery, and absorbing cash losses. Yan expects visible changes within 5 years and envisions high-DOF dexterous hands falling to a few thousand yuan; if the full robot still costs about RMB120K, the incremental cost would not be significant, but the premise remains that “hardware requires accumulated depth.”
Deep dive
1. The Startup Began by Bringing AI from the Digital World into Reality
Wei Dehao began researching dexterous hands in 2021. By 2022, the generalization capabilities of large models in the digital world led him to ask whether AI could acquire general-purpose manipulation skills in the real world. His answer was to first give the model a robotic body, then give the robot “a pair of truly perfect hands.”
In his definition, “the hand is also the best and most general-purpose end effector we have to date.” His experience at ByteDance and other companies showed him that the market lacked a genuinely usable hand, which had become a direct obstacle to AI entering the physical world.
When Yan Qianhang first met him, Wei had not yet graduated. His small Beijing bedroom was packed with CNC machines, desktop 3D printers, prototypes, and engineering builds. The investment case came not only from a market where products were “either too expensive or not very usable,” but also from a hardware obsessive trying to break the trade-offs among performance metrics through fundamental structural and control innovations.
2. “Robots or Rockets” Was a Choice Wei Made Early
Wei grew up in Henan. In high school, he narrowed his career options to “either robots or rockets” and listed mechatronics engineering at Harbin Institute of Technology as his first choice. In his freshman year, before he knew CAD, Inventor, or SolidWorks, he pulled an across-the-hall roommate into overnight self-study and built a ping-pong-ball-collecting robot that won first prize.
He went on to try a vegetable-cutting robot, a hard-drive-gripper robot, a snake-like tendon-driven robot, surgical robots, and probe robots. He later researched SMA shape-memory-alloy brain-surgery robots at CUHK and worked on FPGA control at KAUST. The experience exposed him to structures, hardware, motion control, and algorithms at the same time.
Yan sees this background as a defining trait of early-stage deep-tech founders: beyond passion, they need the willingness to change conventional industry designs with “their own ideas.” Chinese engineering innovation often accumulates through precisely these localized, fundamental changes.
3. OpenAI’s Rubik’s-Cube Hand Exposed the Ceiling of Traditional Control
During his time at Tsinghua, OpenAI’s work using reinforcement learning to train a dexterous hand to solve a Rubik’s Cube had a direct impact on Wei. In his experience, robots with fewer than 7 degrees of freedom can still be controlled effectively with traditional methods, but a hand with more than 20 active degrees of freedom that continuously interacts with objects is “a disaster” for traditional control.
When he wanted to begin research in 2021, there were no high-DOF dexterous hands available for purchase in China. Overseas, there was only Shadow Hand, priced at roughly RMB900K-2.2M—far beyond a lab’s budget. He proposed to his adviser that he first build a hand from scratch and then study reinforcement learning on it; the product company was born from hardware development being forced by algorithmic demand.
4. Four Drive Architectures, Four Performance Ceilings
Wei divides the field into joint direct drive, linkage transmission, artificial muscle, and tendon drive. The goal is not simply to add fingers, but to make a hand that “looks like a human hand and performs as well as or better than a human hand,” with high degrees of freedom, payload, speed, and precision while controlling size and weight.
Joint direct drive puts a motor inside every joint, making control the most direct. But within human-hand dimensions, only very small motors will fit: “you either sacrifice torque or sacrifice speed.”
Linkage systems concentrate motors in the palm and use rigid linkages to drive the fingers. This allows slimmer fingers, but external impacts are transmitted directly to the motors; high-DOF versions also make the palm too thick. They remain constrained by the size of motors inside the palm, making it difficult to achieve both speed and torque, with limited impact resistance.
Artificial muscles attempt to arrange SMA shape-memory alloys, LCE liquid-crystal elastomers, and other new materials according to the layout of human muscles, using heat-induced contraction or bending for actuation. Wei’s assessment retains a clear uncertainty: “The technology may not mature for another 15 to 20 years.” It remains a laboratory product for now.
5. Tendon Drive Is Closest to Biology, but Carries the Hardest Problems
Most of the human hand’s drive muscles sit in the forearm and connect to the fingers through tendons. Tendon drive uses a similar architecture, placing power components outside the hand, making the system inherently lighter and easier to scale to more degrees of freedom.
Tendons are flexible by nature. When the hand hits a table or encounters an unexpected impact from a model, it usually will not shut down or break; when it contacts a person, it is also less likely to cause rigid-body injury. Wei therefore calls it the only approach capable of “pulling several core metrics up to high levels simultaneously.”
But he does not avoid the counterargument: “Tendon is simply not as durable as a rigid linkage, and it is simply not as durable as aluminum alloy.” Long-term loading also causes creep, elongation, and loss of pretension, while precise control of a flexible parallel structure is more difficult than control of rigid transmission.
6. Specialized Tendons and Automatic Tensioning Determine Whether the Design Can Become a Product
Xingji Lightyear tested about 30 types of tendon from around the world, comparing lines used in extreme sports and skydiving with fishing line and tendons used in other dexterous hands. The team found that the requirement was not simply tensile strength, but a combination of low elongation at break, moderate elasticity, wear resistance, and long-term stability.
After exploring materials and processes, the team developed its own Python Cable, extending life to roughly 500K-800K cycles. That figure is not the reliability of the entire hand, but it shows that the core material for tendon drive must be redefined around dexterous-hand requirements rather than purchased off the shelf.
Traditional designs rely on tensioning wheels, additional motors, or manual tightening, and ultimately require downtime to restore pretension. Wei designed an automatic tensioning mechanism at the structural level and paired it with his own motion algorithms, allowing the tendons to retension themselves during use.
7. Investors Choose a Technology-Maturity Inflection Point, Not an Abstract Architecture
Yan worked backward from the target: high degrees of freedom, precision, payload, and cost all had to work simultaneously. Artificial muscle had not left the lab; joint direct drive still had to overcome miniature-motor limits; tendon drive had problems but was already at the point of moving from laboratory research toward products and commercialization, making it a better fit for the early investment window.
His cost assumption is that a tendon-driven hand does not contain many mechanical structural parts and can therefore be made very light. Once mature, its cost-reduction potential could exceed that of joint-motor designs. This remains a view about future mass production and does not mean tendon drive is currently cheap or reliable.
Europe’s early investment was more closely tied to the need for precision manipulation in aerospace. The program mentioned the German Aerospace Center’s joint-motor and linkage hand and the U.K.’s Shadow Hand. Wei and Yan both believe that products built on hardware and enhanced by software ultimately depend on high-frequency engineering iteration, where China’s supply chain, application scenarios, and customer base can accelerate the cycle.
8. Current Buyers Are Purchasing a Research Platform, Not a Mature Production Tool
The industry has long heard the argument that dexterous hands are unnecessary because grippers can solve every problem. Yan says there is no need to get trapped in a conceptual debate; the priority is to build a product that is cheap, durable, and usable. Until then, the most direct customers for high-DOF dexterous hands remain teams researching embodied manipulation.
These customers buy the hand to explore tasks such as grasping, playing the piano, and cracking walnuts, while researching manipulation systems that can run on the hand. Since no particularly mature model can control 15 to 20 active degrees of freedom, it is difficult to prove large-scale commercial value with a single number today.
Brain-computer interfaces and dexterous hands have long-term synergies: whether the control signal comes from a robot model or a human brain, it ultimately has to drive the same hand. Yan believes neither technology is mature; the first step should be to evolve on robots whose hands are temporarily “less dexterous,” then move toward prosthetics that approach the human-hand experience.
9. Both the Hardware Form and Data Sources for General Manipulation Point to the Human Hand
Wei breaks down the commercialization of embodied intelligence into two parts: AI must understand factory or household scenarios and follow instructions, while the hardware must physically complete the operation. Without a general-purpose end effector, even the strongest scene understanding cannot become useful work.
His core judgment is emphatic: “Of everything in our world, the only thing that truly has general-purpose manipulation capability is the human hand.” Two-, three-, or four-finger grippers can be optimized for specific tasks, but they limit the tasks and action sets a robot can cover.
Data imposes a second constraint. The vast majority of internet video records human hands, and human hands are also the easiest form of data to obtain when capturing human motion specifically. A form factor close to the human hand expands physical capability while helping narrow the gap when transferring human demonstrations to robots.
10. The Dexterous Hand Is “AI-Defined Hardware”; Bare-Hardware Delivery Is Incomplete
Traditional robotic arms have only 6 or 7 axes, allowing engineers to build kinematic and dynamic models and solve a trajectory once the end position is specified. A hand with 15 to 20 degrees of freedom must determine the angle of every joint for different cups, bottles, and other objects; traditional methods alone cannot make it directly usable.
Wei therefore says: “The dexterous hand is a completely new form of robot. It is the first hardware in human history defined by AI.” The company does not want to sell only a hand; it wants to deliver an integrated software-hardware manipulation platform whose basic capabilities users can call immediately.
The next generation should integrate a “cerebellum model” directly into the hand to handle actions such as catching coffee and cracking walnuts. It must also combine with the reasoning capabilities of a large model to provide fast local control. Smaller models reduce cost, but Wei cautions that this is an underexplored area, and development difficulty will not fall proportionally.
11. VLA Latency Compresses Genuine Dexterous Manipulation into Pick and Place
Wei’s engineering estimate is that a 10B-scale VLA model runs at roughly 3 to 10 Hz. If a coffee cup is bumped while being delivered, the time required to send visual information back to the model, make a judgment, and execute could already be 300 milliseconds to 1 second: “By the time the instruction arrives, the coffee will have spilled long ago.”
Humans handle such disturbances through hierarchical control, with fast reactions potentially completed at the spinal-cord level without routing every decision through the “brain.” Unless a lighter, faster, edge-deployable paradigm emerges, using the current Transformer architecture to handle all dexterous control directly is unsuitable.
This explains the limits of many demonstrations. Although a VLA is connected to a dexterous hand, it performs only simple Pick and Place, effectively “using the dexterous hand as a two-finger gripper.” The question of why a gripper is not enough reflects the fact that the relevant model has not yet truly been built, not that dexterous manipulation lacks value.
12. The Core Data Trade-Off Is “Accurate but Slow” Versus “Fast but Inaccurate”
Teleoperation lets a person control a robot in real time, so the resulting data captures the robot’s actual movements and is therefore accurate. But one operator may complete only a few hundred samples in a day. Reaching hundreds of millions of samples would require enormous labor and time.
Data gloves directly record the human-hand skeleton and can capture thousands or even tens of thousands of samples per day, making them much faster. The problem is that human and robot hands have different structures, so precision is lost during mapping. Wei summarizes the distinction as: “Teleoperation is accurate but slow; data gloves are fast but inaccurate.”
During his time at Tsinghua, he proposed a dual-mode inner- and outer-exoskeleton system in which the operator wears the apparatus and directly guides the dexterous hand, like a kindergarten teacher teaching a child to write “hand in hand.” The system captures intrinsic dexterous-hand data while preserving the speed of human demonstration, aiming to create a third path that is “both fast and accurate.” He says researchers at Stanford and MIT have recently been pursuing similar directions.
13. Open-Source Gaia Hand Expands the Developer Base Before Commercialization
After founding the company, Wei set himself a personal goal: enable hardware obsessives without luxury laboratories to DIY high-DOF dexterous hands. The team opened up its proprietary miniature joint modules, structural files, host-computer materials, and software recognition functions. The first-generation Gaia Hand costs no more than one-fifth as much as products with the same degrees of freedom.
Yan supports lowering the barrier at this stage because the first commercial customers are precisely the people researching embodied manipulation. “First build the ecosystem, then push commercialization.” If the number of developers and users does not grow, pushing commercialization too early will cap the market.
Open source also allows researchers to build different forms around their own tasks instead of first becoming dexterous-hand specialists. Xingji Lightyear calls its first-generation design an “appetizer” and plans to continue releasing more usable open-source products.
14. The Ultimate Moat Is a Flywheel in Which Software and Hardware Expose Each Other’s Flaws
Asked whether mechanical design, core actuators, or perception and control algorithms matter most, Yan refuses to choose one. Customers judge a product by the final user experience, so the real moat is coordinating the algorithms and hardware.
One step forward in algorithm research exposes hardware flaws; a hardware upgrade then enables an algorithmic breakthrough. If a company only does algorithms, it must wait for someone else to change the hardware, and vice versa. For a startup, that delay is a direct loss of productization speed, so “in commercial competition, especially for startups, speed is the ultimate advantage.”
Frees Fund worked backward from the progress of humanoid platforms to identify upstream investment points. As embodied manipulation models began attracting attention, sensors and dexterous-hand degrees of freedom emerged as constraints. Early-stage investors were looking for areas where “the technology is immature now but will be important in the future,” and where startups were better positioned to drive progress.
15. Application Breakout Is Still Early, but the Market for Human-Hand Tasks Is Taking Shape
Xue Fang asked whether the industry was on the eve of an application breakout. Yan’s answer was, “It’s still a little early.” The inflection point will not come from dexterous-hand companies completing every downstream task themselves, but from integrators picking up the product quickly and placing it into scenarios the company itself had not anticipated.
Demonstration scenarios are easier to adopt first: actions can be repeated and validated many times to ensure stability. Production and consumer tasks still require teams to explore and refine them alongside research customers and seed users.
One way to assess future demand is to look for work that traditional automation cannot handle well and that still has to be done by human hands today. Apparel and textiles require handling soft fabrics; precision assembly of smartphones and similar products is not simply moving A to B and applying a fixed force, but sensing whether a component has flexibly seated in place. These are the natural markets for dexterous hands.
Consumer applications offer similar opportunities, such as making milk tea with human-hand motions rather than hard-coding the process into fixed steps like gripping a cup and pouring ingredients. Xue summarized that the tasks left for dexterous hands are the harder, finer-grained ones. Wei accepted the description and called them “difficult but right.”
16. The First Lesson in Going from Hardware Obsessive to CEO Is the Team; the Second Is What Not to Do
Wei admits that building a software-hardware team has been the hardest part of starting the company. Even though he could cover structures, hardware, control, and algorithms himself, he repeatedly hired people only to discover that their technical stack did not match, or that a new hire’s arrival revealed, “This is the kind of person we actually need.”
In the early days, he still spent long hours personally “tightening screws and writing code.” Over time, he accepted that the CEO’s responsibilities should shift toward fundraising and strategy. A company is a machine run by coordination among different roles; the founder cannot continue treating himself as its only critical component.
His strategic shift is especially clear: “The most important thing in setting strategy is deciding what not to do.” Startups constantly face temptations to expand the product, direction, and organization. What is truly scarce is not ideas, but the ability to concentrate limited resources on the single most important thing.
Yan adds an organizational-efficiency perspective. Hardware obsessives are good at solving problems and often have distinctive labels that attract talent, but because they are strong themselves, they may be unable to tolerate other people’s weaknesses. A good CEO must maximize the ROI of investment in people and avoid hiring “100-point people” only to use 60 points of their ability.
17. Over the Next 2 Years, the Company Will Perfect the Hand Before Connecting the Cerebellum to the Brain
Xingji Lightyear will first focus on building the perfect dexterous hand as it understands it. Wei expects hardware refinement to take roughly 2 years. The company will also devote more effort to building general, generalizable, fast-reacting cerebellum models and exploring how to connect them with the reasoning capabilities of large models.
Hardware design must constantly account for algorithmic usability. The two distal joints of the human hand are naturally coupled, and first principles suggest replicating that structure. But the algorithm team may find coupled joints difficult to train in reinforcement-learning simulation and difficult to transfer from Sim-to-Real, forcing the team to decide whether to sacrifice biomimetic design for algorithmic practicality.
Wei cited application attempts including Optimus sorting batteries, Figure’s long, uncut package-sorting demonstration, and domestic companies such as Leju and Galbot working on factory loading, unloading, and transport. He expects applications to multiply over the next few years, but the company’s focus will remain the underlying general-purpose platform.
18. The Mass-Production Gap Runs from Non-Standard Parts to Cash Efficiency
Around 95% of Xingji Lightyear’s current hardware cost comes from non-standard machined parts, motors, and non-standard circuit boards. A single finger may require more than a dozen parts that have not yet been molded. Hollow-cup motors in 6 mm, 8 mm, 10 mm, and 13 mm sizes, along with their reducers, have not yet become stable standard products, leaving both the company and suppliers to absorb high costs at small volumes.
Reliability and R&D speed are in direct tension. If versions iterate too quickly, there is no stable long-term version on which to run complete life testing; if the version is frozen for testing, frontier R&D may stall. Wei still believes “durability is definitely an engineering problem,” but acknowledges that early products need time to accumulate experience: “Hardware requires accumulated depth.”
Yan warns that the cash-burn rate is completely different when building 1 prototype, delivering the first 100 units, or making an initial delivery of 10,000 units. The arrival schedule for hundreds of parts, supplier selection, and stable delivery can turn local losses into a company-wide management problem. Mass production is not a hurdle already cleared, but the unavoidable challenge of the next stage.
Frees Fund’s post-investment support includes connecting companies with downstream customers, supply chains, and different scenarios, as well as having founders who have already passed through mass production, customer development, and product-iteration stages share their experience. Yan emphasizes that “teaching someone to fish is better than giving them fish”; helping Wei grow is more important than directly providing orders or naming suppliers.
19. China Excels at Taking One to 1,000; Overseas Players Still Hold the Edge in Zero-to-One and Capital
Yan summarizes China’s advantage as a combination of scenarios, customers, and supply chains. In Shenzhen’s Huaqiangbei, a modified circuit board might arrive that evening; if R&D is in the U.S. and the supply chain is in Shenzhen, the same revision could take more than 10 days across the Pacific. The gap in iteration cycles continually compounds into an engineering-speed advantage.
That is why domestic teams may catch up quickly even though overseas companies such as Boston Dynamics, Tesla, and Figure got started earlier. The U.S., by contrast, remains stronger in frontier innovation and capital scale; the program cited Figure’s $1B financing. Its market is also more tolerant of long-term scientific exploration.
Yan views Physical Intelligence as a commercial company operating somewhat like a research organization. Among the embodied-intelligence models it releases, the relatively stronger ones are open-sourced, helping move the industry forward together. His summary is that overseas players have an advantage in “zero to one,” while China is better at “taking one to 100 and one to 1,000.”
China’s new-energy and consumer-electronics supply chains are already relatively mature, with foundations in traditional industries for motors, reducers, ball screws, and other components. Yan has not seen many major “chokepoints” in dexterous-hand hardware domestically. The harder problem is tactile sensing: robots need not only hands but also human-like perception, ultimately requiring sensors that are “cheap, usable, and durable.”
20. Patents, Standards, and the Capital Map All Revolve Around Product Maturity
Xingji Lightyear plans to build domestic and international patent coverage around materials, mechanical structures, and control innovations, while continuing to monitor and participate in industry standards. Wei sees these as part of the moat, but the discussion makes clear that patents alone are insufficient; products, models, and engineering iteration must also work together.
Yan believes capital should not be invested in only one part of the chain. The ideal sequence is to invest in the robot platform while the technology is immature, move upstream into the supply chain once constraints emerge, and enter downstream applications once they begin to take off: “Put money in the right place, at the right time, behind the right people.”
Frees Fund’s robotics map has four sections: the core supply chain, key sensors, AI plus the robotic body, and downstream applications. The program named Xingji Lightyear, InTimes Robotics, LimX Dynamics, Yuanli Intelligence, which focuses on the embodied brain, vision companies, force-sensor company Hangke Microelectronics, and the early logistics-intelligence project Covariant.
The map extends back to earlier industrial robots, parallel robots, and specialized robots. Yan wants to identify new constraints before every “step-change acceleration” in technology, rather than wait until products, supply chains, and applications are mature and then chase consensus.
21. The Financing Boom Has Not Removed the Boundaries; A 5-Year Cost Reduction Is the Real Standard-Component Test
Yan dates the start of the current embodied-intelligence boom to Tesla’s announcement in September 2022 that it was building Optimus. Progress in VLA-based manipulation then drove financing for domestic robotic bodies and embodied brains. This year, Unitree’s appearance on the Spring Festival Gala created a mass-market visual shock, pushing investor consensus into society at large as attention and capital amplified each other.
Founded roughly 1 year ago, Xingji Lightyear has completed financing from its angel round through Pre-A. Wei recalls that around the Spring Festival Gala, at least 5 new investors and financial advisers added him on WeChat every day. By June and July, the heat had cooled and the market was again asking why dexterous hands were necessary; more recent robot progress has renewed recognition of their value.
Yan expects visible changes in 5 years and envisions high-DOF dexterous hands falling from the current RMB40K-50K to a few thousand yuan. If a complete robot costs about RMB120K, a hand costing a few thousand yuan would not materially raise the total. The adoption path could resemble lidar and air suspension moving down from premium configurations, but “the path to becoming a standard component is indeed long.”
On labor-market disruption, the guests condition their views on speed. Slow substitution would first enter dangerous, repetitive jobs that young people do not want to do and where employers may even be unable to hire; if technology advances too quickly, it could rapidly replace some jobs, much as online shopping disrupted physical stores. Wei once turned down a major-tech-company offer worth more than RMB1M a year to start a company. Yan’s advice to those who follow is to first understand the technology’s boundaries and unresolved problems, then choose a battlefield genuinely worth solving.