173: A Conversation with 姚颂: DeepGlint, Oriental Space and Starting Over—The “Genius Boy” Ten Years On

2026/07/21 · Source link · AI Summary archive

Summary

“VLA is close to the end of the road.” 姚颂 said he saw the signal in papers as early as July-August 2025: adding training data to VLA produces diminishing returns, and accuracy actually deteriorates beyond a certain point; some companies have already shifted from VLA toward reinforcement learning and world models. His next move is a latent-space approach: the model only needs to learn physical laws and common sense—“the text printed on the paper and the color of my clothes have no bearing on robot manipulation.” A general-purpose pixel-level world model with “5P, 10P or even 100P parameters may not necessarily hold up,” given 30-frame real-time requirements and edge deployment constraints.
Physical AI is a systems battle, not a genius battle. “One or two young professors or one or two wunderkinds cannot solve these problems.” Citing 曹旭东’s framework from 10 years ago, he says a top algorithm is 50% data, 30% compute, and 20% human experience and inspiration—and notes that none of China’s 4 leading foundation-model companies was founded by a lone wunderkind. First-person data costs RMB300-600/hour at market rates, implying RMB300M-600M for a 1M-hour target; 正大’s more than 400,000 full-time employees and closed-loop operating environments bring data costs below one-tenth of market rates, while infra optimization means “256 cards can deliver what a company without infra capabilities would need 1,000 cards to achieve.”
On commercialization, he is explicit: “Every milestone in the development of intelligence must have a commercial landing.” Independent autonomous-driving companies betting on L4 have an extremely high mortality rate, while Boston Dynamics “can only survive on continuous transfusions.” Physical AI’s GPT-3.5 moment would be 95-99% zero-shot success on any embodiment and any task; today, “foundation models are at the level of top 985-university PhD students, while models of the physical world are still in kindergarten,” and general-purpose algorithms are 2-3 orders of magnitude harder than autonomous driving. But, as with scenario-based autonomous driving, the physical world can deliver substantial commercial value before general intelligence arrives.
His view of bubbles is shaped by history. Horizon Robotics had roughly RMB100M of revenue and a valuation above $3B in 2018, when everyone called it a bubble; after listing in 2024, its market cap peaked above HK$150B. Momenta’s latest round was reportedly valued above RMB100B. “If you can survive until the day technology is deployed at scale and generates commercial value, none of these companies are bubbles.” Two types can survive: the best-funded pure technology idealists—“all your work should be for fundraising”—and companies with an operating business that follow the Horizon Robotics/Huawei L2→L3→L4 path. The NDRC says more than 100 companies have launched over 200 humanoid robots; he believes output could exceed the auto industry within 10-15 years and that embodied intelligence can become “the king of all industries… an economic pillar.”
Striding AI’s structural bet: nearly $100M in its angel round and a 70-person full-stack team. Charoen Pokphand Group—with $110B in revenue, 16,000 company-operated 711 stores and 2,700 Lotus’s stores—and Huaqin Technology, the world’s largest handset ODM, are both its largest external shareholders and providers of retail and 3C manufacturing environments. “Our ability to go overseas was shaped by our ownership structure and global commercial network from day one.” The company chose a wheeled dual-arm robot and will not build dexterous hands in the near term: “The number of degrees of freedom a robot should have is determined not by its hardware, but by its algorithms’ control capabilities.”
Vertical integration will be the mainstream model for embodied intelligence over the next 3-5 years. He compares the industry with the PC era: Apple and IBM built everything in-house in the 1970s and 1980s; standardization brought specialization by the end of the 1980s; today, Apple’s closed ecosystem is again “unique.” Physical AI still lacks standardized interfaces and mature suppliers, so Striding is building its algorithms, edge software, robot design, cloud teleoperation, and solutions in-house, while outsourcing joints, structural parts, and production. This directly contrasts with 许华哲’s “leave it to the ecosystem” view. 姚颂 acknowledges that both are viable paths, but says Striding is choosing vertical integration for now.
His contrarian operating philosophy runs through all 3 ventures: he rejects putting everything at risk by mortgaging one’s house—gamblers eventually lose because “there will be a moment when your principal goes to zero”; he cut DeepGlint’s drone business because “you have to believe sunk cost is zero”; he rates himself 80% offense, 20% defense—“look at him in 2 years and he’s 100 while I’m 70; look from 5 years out and he’s 100 while I’m 200.” The sense of being pushed forward by the era is like riding an elevator: “I know very clearly that my company went public because of the support of the Party and the state,” a conclusion every entrepreneur who benefits from the cycle should understand.
Looking back on DeepGlint: after 50 institutions refused to invest, 金沙江 signed a term sheet on the spot, before AlphaGo defeated 李世石; in mid-2018, the company was sold to Xilinx for more than $300M. He rejected having Nvidia, Intel and other competitors invest simultaneously for a short-term branding boost: in the long run, “you might not be able to sell the company,” and commercially “none of them would help you.” After the sale, pure happiness lasted only “2 minutes of pacing around the table,” followed by a dark period of trying to choose a new path. His view of wealth: “Once your wealth exceeds a certain threshold, it no longer belongs to you”—he did not specify whether that threshold was RMB500M or RMB1B, only that wealth above such a level is no longer truly one’s own.

Deep dive

1. Opening self-definition: “I’m an entrepreneur”

Born in 1992 and 10 years into entrepreneurship, 姚颂 has only one opening line for himself: “I’m an entrepreneur.” His new company, Striding AI, or 正行创新, was founded just over 6 months ago and defines itself as a physical intelligence company: “We are neither simply a model company nor simply a robot company.” It combines data, compute and models, while considering algorithms, software, hardware, solutions and deployment environments together as “an intelligent-capabilities company for physical intelligence.”
曼祺 notes that he has deliberately avoided the hotter term “embodied intelligence”—a choice of wording that is itself the episode’s first argument.

2. Physical intelligence is not embodied intelligence: a system, not a brain

姚颂’s distinction is straightforward: embodied intelligence means “an intelligent capability inside a robot,” but by itself “it cannot provide services to people.” If 10 robots are working in one building, the entire space needs intelligent perception, advance route planning, and real-time visibility into every robot’s battery, location and status.
His key analogy: even a Robotaxi fleet with 30 or 50 vehicles still needs people providing remote teleoperation as a backstop. “Embodied intelligence may also need a remote system over the next 10 years.” The integration of space, bodily movement and manipulation is what he calls physical intelligence.

3. Two threads across 3 startups: drawing a circle

He uses one circle to unify the arc from AI chips to rockets to physical AI. The circle represents the space in which humanity survives, which must continually expand through energy—ultimately nuclear fusion—and aerospace. The points inside the circle represent each individual’s survival, health and happiness; the relationship between points and the circle is society, whose core questions are efficiency and fairness. “A lot of things fall within these 3 major questions”: aerospace expands the boundary, while chips and robots raise efficiency.
The second thread is internal: “Seeing a simple algorithm, an initial technical idea, turn step by step into a product that genuinely affects many people.” He readily accepts 曼祺’s summary: one thread is outward-facing problem-solving, the other is intrinsic motivation.

4. Freshman year at Tsinghua: 11 hours with a stopwatch

A sponsored student from Changsha No. 1 Middle School who ranked in the top 3 of his grade, he entered Tsinghua’s Department of Electronics and heard a blunt warning during military training: “The worst among you ranked somewhere in the province, but there will definitely be someone in the 270s in the department.” For the first semester of freshman year, he used an old electronic watch to time himself and had to study 11 hours a day excluding breaks, meals and bathroom trips—“I even had to pause the watch when I went to the bathroom.”
The pressure had a face: 韩衍隽, a “monster” in the same year. The calculus professor reportedly designed the final question specifically to stump him; he later won Tsinghua’s top student award and is now a professor at New York University. “So many students at Tsinghua were both smart and hardworking. I simply couldn’t keep up.”
The lasting damage is something he describes candidly: “If I’m not working hard, working late or staying up, I feel guilty. That’s a mental knot a lot of Tsinghua students carry.”

5. The awakening in freshman spring: the hardware club and linear algebra

After nearly breaking down during freshman winter break, he realized 2 things: the string could not stay taut forever, and university was different from high school—the goal was a career, with multiple paths through student work, innovation, research, overseas study and entrepreneurship. He chose research plus innovation and joined the hardware division of the electronics department’s science association. The senior who received him during recruitment was 王鹤, now the chief scientist of 原音’s “宇核通用.”
His motivation to study has to come from seeing how knowledge works. After emailing 汪玉 and joining his lab for research during freshman summer, he wrote to his linear algebra professor that he had not understood the subject all year and only grasped its purpose through research. The professor replied: come teach the first class next semester. As a result, half of the electronics students one year below him attended his first class at Tsinghua.

6. The “god” culture and Spark Program: the founding bench of a decade later

Tsinghua’s dominant values were study and research, and top performers were elevated to “gods”: 韩神, 帆神, and even “二神” because “a class can only have one top god.” 杨植麟, 高继扬—also heard as “高奇洋”—陈建宇 and 许华哲 were in the years before or after him, and “most could be counted as gods.”
Cross-department programs brought this cohort together: the Spark Program, which he attended alongside 韩松 and 许华哲—杨植麟 “should also have been there, but I’m not completely sure”—and Stanford’s UGVR undergraduate visiting-research program, which funded 20-30 students annually and included both him and 高继扬. “A large number of the people at the leading companies in the market are classmates I knew in my freshman or sophomore year.”
He sees his own path as different. 王鹤, 高继扬 and 许华哲 followed the typical route of PhD, professorship or industry, identifying a commercial problem and then starting a company. He entered a lab as a freshman, published 4 or 5 papers as an undergraduate, mentored new students as a sophomore and opened a new research direction in the lab—“a shorter path from research to research management to technology transfer.”

7. Boredom and suffering: why he gave up a full scholarship at CMU

On his first trip to the US during his junior year, he met Tsinghua seniors working in Silicon Valley: 5 or 6 hours a day, slow increases in compensation and title, and conversations reduced to “I’ve been to 28 states” versus “I’ve been to 30 states.” His yardstick came from a professor: “Life has only 2 states: boring and suffering.” For him, “I could not tolerate boring.”
At the time, going overseas for a PhD was still overwhelmingly mainstream—about 45% of his electronics cohort went abroad, versus possibly less than 10% today. Giving up a fully funded PhD offer from CMU to start a company was “an extremely radical choice that few people accepted or understood.”

8. 汪玉’s 2 ways of removing the pressure: opportunity cost and a fallback

汪玉 helped him decide with 2 arguments. First, “the opportunity cost of entrepreneurship is lowest when you are young”—once you are in your 30s, married, with children and a mortgage, you are less willing to take the risk. Second, if he failed, “I can still recommend you to CMU or Stanford for a master’s degree.” The respectable mainstream path could always be reopened.
His parents were firmly opposed. “Professor 汪 did a lot of the ideological work,” repeating to his parents over multiple rounds what he had told 姚颂 until they finally agreed.

9. Against all-in: why gamblers always lose

When asked whether having a fallback would prevent him from going all-in, he pushed back: “Never say that.” He cites a Zhihu gambler’s argument: even with 50/50 odds and no casino rake, “there will be a moment when your principal goes to zero. Once it is zero, you have no capital left to win it back.” His rule: “All-in your time, energy, attention and resources, but do not seal off your escape route.”
The real cost of making failure too expensive is distorted execution. “Pursuing success and preventing failure are 2 completely different paths.” The former lets you plan beyond the 3-year cash runway and keep raising capital; the latter demands that every yuan produce 2 yuan of output and can turn a 1.5-year layoff process into 5 years. “The pace becomes severely distorted.”

10. DeepGlint’s starting point: an AI chip built as a thesis project, and 商汤’s vote of confidence

Several students from the 2011 cohort built a complete AI-chip demo as a thesis project in 汪玉’s lab, assembling algorithms—周二进, later at Megvii Research Institute and now a co-founder of Yuntian Lifei—architecture, software and a compiler. 姚颂 was the project PM. 商汤 was founded in September 2014 in a cheap, run-down building at Tsinghua Science Park, giving them a basis for believing that deep learning’s entry into real applications might overwhelm the capacity of existing CPUs and GPUs.
韩松, then in his second or third year of a PhD at Stanford, returned to Tsinghua at the end of 2014. His work on sparsification and quantization compression was naturally complementary to hardware acceleration, and the two sides “gave each other courage.” The company’s first employee was 隋林志, a junior he had mentored in the science association. At the time, 汪玉 was 32 and had been an associate professor for only 2 or 3 years; 韩松 was 25 or 26.

11. 50 institutions said no: the hard-tech financing market 10 years ago

They began meeting investors in the second half of 2015, and not one of 50 institutions was willing to invest. He identifies 3 reasons: China had produced almost no profitable technology-investment precedents—the number of profitable chip companies could be counted on one hand, such as Spreadtrum; there was no precedent for professors and students founding a company—“商汤 gave us an example”; and most investors came from consumer, entertainment and internet backgrounds and had no way to evaluate the technology.
邓锋 of Northern Light Venture Capital delivered the blunt feedback that changed the pitch. After many revisions, the deck “was still the way a pure engineer would explain it.” Business investors wanted to know what product a trillion-yuan market required and what metric the technology could uniquely deliver. “You have to reason backward from the market to the technology.” 姚颂 later summarized the first lesson of technical entrepreneurship: “The core of your technology is not self-amusement; it has to create value in a product.”

12. 金沙江 signed on the spot; AlphaGo changed little

The angel round was led by 张玉彤 of 金沙江, a Tsinghua electronics graduate who could understand the technology. He brought the project to Silicon Valley partner Richard Lin (林仁俊), a decades-long semiconductor investor who reasoned that the industry would move from general-purpose to specialized processors. The term sheet was signed “right there in the office with me and Professor 汪 before he left.” Gaorong joined decisively after 金沙江 approved the deal. The timing was late February or early March 2016—before AlphaGo defeated 李世石.
After AlphaGo won, their position barely changed: “Investors still mostly looked at credentials.” They changed tactics. If customers and partners deeply believed in the technology, “why not turn those people directly into investors?” 360 led the Series A, followed by Samsung and MediaTek—each after 7 or even 9 rounds of technical diligence.

13. Choosing Xilinx over simultaneous investment by competitors: his proudest piece of analysis

“The thing I’m most proud of at DeepGlint” was that he modeled the company’s ending immediately after receiving the term sheet: an IPO—what the rumored A-share Strategic Emerging Industries Board, US and Hong Kong listings would each require—an acquisition—what Nvidia, Intel and Alibaba might buy—or bankruptcy. “You have to become world-leading at something and be strategically important to the acquirer’s group before it will buy you.”
Having a single company backed simultaneously by Xilinx, Intel, Nvidia and even AMD would have been “unprecedented in human history.” Their terms were originally mutually exclusive, but all ultimately accepted coexistence. He concluded that the arrangement offered only short-term branding value: competitors would gain complete information rights and feed the intelligence into their own product road maps; at a sale, the other parties might bid or interfere; and each would fear that business resources were helping a rival’s product. “In practice, nobody would help you.” He gave up the short-term brand lift and chose Xilinx, which was most useful to the business.

14. Commander and political commissar: DeepGlint’s power structure

“Professor 汪 and I were more like a commander and a political commissar.” 汪玉 set the broad direction while 姚颂 gave orders on the front line. “If my decision caused universal anger, Professor 汪 would come out and clean up the mess.” When an important customer was not a room he could command as a young man, 汪玉 would step in—for example, in a 2016 meeting with 王小川, then at Sogou. 王小川 was from the 1996 cohort and 汪玉 from the 1998 cohort, so they had more in common.
In the 4-person core team, 单毅—heard in the audio as “山艺”—was CTO and led R&D through productization; 韩松 ran forward-looking algorithm exploration through weekly and biweekly meetings. 曼祺 points out that the group lacked someone with commercial experience. Day to day, 刘敬秀, a classmate of 汪玉’s with a multinational marketing background, filled the gap; but 姚颂 developed the ability to make the major trade-offs himself.

15. Killing the drone business: sunk cost is zero

The smart-drone solution developed with Zero Zero Robotics was unveiled at CES in January 2017 and was about to enter small-batch production. That same month, he decided to kill it and shift to smart security and driver assistance. The technical team’s “entire year of work was wasted,” and the company was thrown into turmoil. His principle: “You have to believe that sunk cost is zero.”
The reason was market structure. DJI held 70% of the drone market and did not work with small companies. Zero Zero Robotics, the second-largest player nationally, would not generate much volume, and serving all the remaining companies combined would not generate much volume either.

16. A CEO only handles “bad versus worse”

His observation about the CEO role is worth preserving: no one comes to you with outcomes that are good versus bad, or good versus better. “People only come to you when the result of a decision is bad versus worse.” You are always handling the problems that others cannot absorb—the messiest and most complex issues.
Why does he still like it? “Pain brings growth. A setback may be a blessing in disguise.” The financing period after 50 rejections strengthened his temperament. 汪玉 told him then: “If 1 out of 100 institutions ultimately invests and understands this, you can build it. Once you build it, you can make them look again.”

17. Selling to Xilinx for $300M: 3 of 10 factors

In January 2018, Xilinx’s new CEO wanted to shift the company from a device company to a platform company and urgently needed a software and algorithms team. Xilinx made an acquisition offer. He does not hide the first factor: “We had never seen that much money.” For an undergraduate only 3 years out of school and raised in a salaried household, “of course it was tempting.”
The second factor was more revealing: this was the first time he thought about the company from the negative side. “After you have presented your BP dozens of times, you start believing every word yourself.” Investors had reinforced the belief that the company had no risks. His negative scenario was clear: the Strategic Emerging Industries Board still had not launched by 2018; a domestic IPO required 3 consecutive years of profit growth, meaning at least 5 or 6 years.
The third was the team’s lifecycle. Average employee age rose from 26.7 in 2016 to around 31 in 2018, while a 996 schedule became a 6-day alternate-week schedule. One technical leader “had not spoken to his child for 2 consecutive weeks”; another, 34 or 35 and newly married, said, “The child is the most important thing in my life.” Multiple factors converged on accepting the offer.

18. The wealth line: above RMB500M or RMB1B, it is no longer yours

Does he regret seeing Cambricon’s later market value? “Not really.” His view of wealth: “Once your wealth exceeds a certain threshold, it no longer belongs to you.” The host mentioned RMB500M and RMB1B; he did not specify the actual number, only that beyond such a level the wealth is no longer truly yours—you should not use it, and cannot use it.
“There is no need to envy 陈天石.” They chose a path that let him spend the previous 5 years playing with rockets, and he was happy to do it. He and 曼祺 arrive at the same conclusion: the most important thing in life is time. Spend it on what excites, interests and makes you happy.

19. 2 minutes of pacing, then life went dark

The moment the acquisition agreement was signed: “I was extremely happy and paced around the table for 2 minutes. After pacing for 2 minutes, my life fell into darkness.” He knew he was not suited to a large company: “I seek continuous challenges, and I’m not good at currying favor upward.” He had to leave—but what would he do?
His first startup’s direction had not truly been his choice. He joined 汪玉’s group because he liked 3D integrated circuits; the AI accelerator direction was created by the lab, and he happened to be responsible for it; the company changed because the direction changed. “If I start another company, I’ll have to choose the direction myself. What should I do?” It was “like a fog in front of you—not indecision, but complete chaos. You don’t know what roads you have.”
At 25 or 26, he spoke with 李诞 and 王华东. “I felt a bit like I was having a midlife crisis.”

20. The 经纬 years: “You don’t need cash; you need enough information”

王华东 advised him to look at what others were doing and invited him to join 经纬 as a part-time venture partner. 张颖 saw through the problem: “姚颂, you don’t need cash. You need enough information. You can walk into any meeting at 经纬 China.” Partner meetings and investment committees were open to him. “I think he was also laying the groundwork for my next venture. 张颖 is an incredible person.”
He invested in 2 or 3 companies at 经纬, including Silicon Intelligence Technology, which recently went public, though it was not his main responsibility. He also incubated and angel-invested in a brain-computer-interface company. But he learned that “I cannot be a professional investor”—when he sees a founder clearly walking into a fire, he cannot stop himself from trying to pull them back.

21. Eliminating lidar: 李一帆’s engineering thesis

In the first half of 2019, he met 6 or 7 new lidar teams in 3 months and discussed their different technical routes. He asked 李一帆 of Hesai and got the conclusion: “The technical route is not important. Engineering capability is.” A small technical team can build a prototype in 3 months; the hard part is moving from prototype to small-batch to mass delivery and from manual assembly to automation over 3-5 years.
李一帆’s inference was that if solid-state lidar matured, Hesai and RoboSense would remain ahead because they were ahead in engineering. “That judgment has now been validated.” 姚颂 did not invest in the company and did not choose lidar as his next startup direction. “In many industries, the engineering difficulty and requirements are greater than those of the 1 or 2 technical breakthroughs.”

22. Eliminating brain-computer interfaces and domestic GPUs: cycle time and integrity

He studied brain-computer interfaces systematically before Neuralink became famous and spoke repeatedly with BrainCo. Invasive systems require opening the skull, followed by 2 or 3 years of mouse experiments, another 2 or 3 years of monkey experiments and 3 years of clinical work. The field was still in research and approval, so “my capabilities might not be useful.” He chose to incubate and invest rather than serve as CEO.
He also passed on the domestic-GPU window when Biren, Moore Threads and MetaX were founded in 2019-20. There were 2 reasons: the acquisition agreement committed him not to start a similar company for 5 years, and “I didn’t want someone to hold shares on my behalf”; and at the time, “without the demand generated by large models, the growth of these companies’ businesses was a mystery.” The only catalysts then were Cambricon’s IPO and Nvidia sanctions.

23. Eliminating nuclear fusion: the first-wall math does not work

Nuclear fusion and quantum computing fit his “big circle” thesis, but both were still purely at the laboratory stage and far from commercialization. He gives a concrete example: fast neutrons would batter a tokamak’s first wall until it was “riddled with holes,” potentially requiring a 3-month shutdown and 6-month replacement with specialized materials. Once replacement cycles, costs and raw materials are included, “your electricity may not be cheaper than today’s hydropower.”
This clarified his selection rule: frontier technologies whose engineering path and commercial value can be clearly seen within the next 2-3 or 3-5 years. “Otherwise it’s too far away. I even considered going back to get a PhD.”

24. 6 months as a wealthy idler: he even considered becoming the drummer for 水木年华

During the dark period, he experienced a different kind of emptiness: “Have you ever felt bored? You don’t even want to scroll your phone, and nothing interests you.” It was not acute pain but a slow simmer. Coinciding with the first season of The Big Band, he moved from guitar into drums and keyboard and posted one-man-band videos to his social feed. He had late-night meals with New Pants and took a backstage photo with GAI at a music festival. “Doing it once was enough.”
“I genuinely considered playing in a professional band. 水木年华 was missing a drummer; I could have gone to play drums for them.” But the conclusion was predetermined: he could not escape technology. Work would always be the main current of his life. “I don’t have work-life balance.”

25. Breaking down China’s commercial space industry, and the 6G-to-space thesis

He breaks commercial aerospace into 3 words. Aerospace means first achieving launch capability and then moving steadily toward reusability. Commercial means counting the economics, financing, cash flow and local-government partnerships rather than ignoring cost. China means “you have to position the company as serving the country if you want to build it properly,” because state missions will eventually be released to market participants.
When he founded the rocket company in 2019, “the only target we could find was SpaceX. Beyond that, we didn’t know where the customers were.” His judgment, based partly on information he had heard, was that the state might establish a Starlink-like satellite internet program over the next 2 or 3 years. His technical intuition came from his old field: 5G is already fast enough, and under the generational-upgrade curse of “single-digit generations falling short of expectations,” 6G would be about coverage, not speed. “Can one base station cover several hundred kilometers around it? Then you have to go into the sky.”

26. “Good, many, fast, cheap”: the reordered 4 priorities

Product definition at Oriental Space began with a reordered priority list. Good means a high success rate: “No satellite-internet constellation plan will use you if your launch success rate is 50% or 60%.” Many means 12-18 satellites per orbit and at least 4 or 5 tons of payload—“less than 5 or 6 tons is not enough to support the mission.” Fast means moving first: “If you wait another 3 or 4 years, the state teams will have reacted and the market will be red-hot.” Cheap comes last; getting satellites into orbit quickly matters more than saving tens of millions of yuan.
This led to a staged path from solid rockets—reliable, fast and requiring no in-house engine development—to liquid reusable rockets. The team was complementary: rocket engineers from the state system understood technology and launch approvals; 姚颂 understood strategy and capital, could work with markets and local governments, and could design the organization.

27. Gravity-1: the idea for the first livestreamed private rocket launch

His happiest moment was obvious: Gravity-1’s first flight succeeded on its maiden launch in early 2024. He proposed the first livestream of a private space launch and explicitly accepted that the launch might fail. “Aerospace should not be this remote from ordinary people. I want it to feel like a music festival.” 水木年华 performed live, and the rocket was branded the “Hailan Home”号 after HLA, whose chairman 周立宸 is a Tsinghua School of Economics and Management alumnus.
Internal resistance was limited. The launch team was confident in the probability of success, while he framed the event as the company’s external calling card and a large-scale event everyone could participate in.

28. Solid boosters around a solid core, sea launches and an EDA mindset

Gravity-1 was technically unconventional: an all-solid clustered configuration is rare—typically the center is liquid and the surrounding boosters are solid. Uneven combustion during the final seconds of solid-motor burn and differences in engine thrust required extensive engineering innovation. Launching from a ship at sea also meant solving flame splashback and hull scorching; winds and ship movement were significant that day.
His cross-disciplinary contribution was a simulation mindset. SpaceX discussed “GPUs to Mars” and GPU-accelerated simulation at GTC in 2015. His EDA analogy is that before sending a chip to TSMC for tape-out, “we are more than 99.5% certain we know the result.” The most expensive and time-consuming stage of a rocket is the initial prototype. If combustion simulation could be made realistic enough, “could we cut testing costs and shorten the experimental cycle?” He discussed combustion simulation extensively with professors at Peking University and Beihang, but “rockets have to put safety first,” so the team is cautious about entirely new technology.

29. Bad background checks, outside turmoil and an amicable exit

His lesson in emotional resilience came as early as 2016. An investor asked a Tsinghua alumnus who knew him to conduct a background check, and “we never expected the person to give nothing but bad reviews: what we built had no value, and he could do it all himself.” His takeaway was cold: “Only the people closest to you genuinely want you to do well. Many others are more likely to envy you.”
On the rumors surrounding Oriental Space toward the end of his tenure: “I had no mood for it. It had no impact on me at all. I slept.” He kept one line for himself: “Things succeed through secrecy”—some core strategies simply should not be discussed publicly.
His departure in summer 2025 was “a fairly amicable process.” The company had signed contracts worth more than RMB1B and entered steady development. For legacy shareholders who had invested because of him, he helped find institutions to purchase their old shares and granted every old shareholder equity compensation in the new company. They had entered during the commercial-space boom, so their investment valuations had certainly multiplied several times. His only regret was missing Gravity-2’s reusable launch: “Probably at the end of this year or the first half of next year, the first launch may involve some very bold attempts.”

30. Entrepreneurship is an Inzaghi goal

The third time he searched for a direction, the darkness was still there, but his resources and information had changed—he remains a founding partner and investment committee member at C Fund. His best metaphor for entrepreneurial timing was one of the episode’s strongest: “Entrepreneurship is an Inzaghi goal.” The “king of the six-yard box” did not rely on tricks, dribbles or bicycle kicks. “He was just a few meters from the goal, gave it a little push or a little header, and it went in.”
“You have to feel as if you are standing in Inzaghi’s position and the ball has arrived at exactly the right moment, rather than insisting that you must succeed as an entrepreneur and searching painfully for a direction.” Timing, conditions and people are waited for; they cannot be forced into existence.

31. 正大 and Huaqin: environments, the data flywheel and the first shareholders

This time, the “ball” came from 2 industrial shareholders who approached him: Thailand’s Charoen Pokphand Group, among the top several dozen companies in the Fortune Global 500 and with roughly $110B of revenue last year, spanning feed, agriculture, retail and information technology; and Huaqin Technology, the world’s largest handset and laptop ODM. He proposed the areas of cooperation, but both were strategic priorities for the partners, who were willing to make forward-looking investments rather than wait for an 8- or 12-month payback calculation.
The deployment environments are 3C electronics manufacturing with Huaqin and consumer retail with 正大. All 711 stores in Thailand and Myanmar are company-operated by 正大; it has roughly 16,000 company-operated convenience stores and 2,700 Lotus’s hypermarkets, plus 7,000 or 8,000 franchise partners. “Physical intelligence has to be built on environments: collect data, develop a model product, put it on the production line, collect more data, and iterate.”
Another piece of timing was his renewed connection with 汪玉. Over the past 5 or 6 years, 汪玉’s team has placed dozens of students into embodied-intelligence research spanning reinforcement learning, world models and infra. 余金城, a young faculty member on the team, was the successor to 姚颂 as chairman of the electronics department’s science association. “I knew him when I was a sophomore and he was a freshman.”

32. 谢国民: the 86-year-old chairman who bent down to scan WeChat

In winter 2022, 徐志敏, president of CP Japan and a member of Tsinghua’s 1979 civil-engineering class, arranged a meeting. 姚颂 brought 戴雨森 and 刘畅, co-founder of GalaxySpace, to brief veteran chairman 谢国民. During the 3-hour meeting, “the old gentleman sat upright in a suit and tie, listened carefully and kept engaging, perhaps drank water only once or twice and did not leave for the bathroom.” At dinner, 姚颂 tried to exchange contact details. “He said no—you are the teachers, you sit; I’ll add you on WeChat,” and then bent down one by one to scan their codes. “To retain this curiosity about technology and such respect for people at that age—that is something to learn.”
Afterward, every time 谢国民 came to Beijing he asked 姚颂 to bring “interesting people of the same generation” to meet him: compute brought 陈建 from Parallel Wireless, 林源 from QingCloud and 汪玉; autonomous driving brought 应奇; embodied intelligence brought 王鹤 and 许华哲. When the chairman suggested working together, 姚颂 was still building rockets and politely declined because “rockets and satellites are sensitive industries.” After he left Oriental Space, the response was “unconditional support”; in embodied intelligence, “I chose to work exclusively with you.”

33. “VLA is close to the end of the road”

By July-August 2025, he had formed the view quickly, based on 2 observations. Papers showed that adding training data to VLA produced diminishing performance gains and eventually caused accuracy to decline, potentially because of data bias. Some companies had already shifted to reinforcement learning: “Pure VLA is difficult to take directly into delivery, and accuracy has hit a ceiling.”
He sees an opportunity rather than an endpoint. World models and reinforcement learning “are both necessary.” “We are facing another wave of opportunity, with a new group of people able to explore new technical directions.”

34. The latecomer’s math: which first-mover advantages can be offset

曼祺 asks the sharp question: founded only in the second half of 2025, with 智元 and the company from 原音 apparently associated with 宇树 already at the top, is timing still on his side? He first acknowledges the obvious: first movers can raise more capital and list this year. “That has to be admitted.” They also gain business resources and the ability to attract talent.
But he believes his 3 startups have nearly offset those advantages through trust and relationships. “Before this interview was public, people would already see us in strategic partnerships with several very large listed companies.” On talent, rising awareness of the sector is drawing in stronger people, just as the emergence of Biren, Enflame and MetaX once assembled teams better than DeepGlint’s early group.

35. What others do not have: internationalization embedded in the cap table

The real differentiation is “rooted in China from day one but oriented toward international markets.” The logic is structural: “In a once-in-a-century transformation, once your shareholder structure is set, you may not be able to offset certain factors.” A first-time founder must gradually assemble overseas resources along the way; “from the day we were founded, we already had a global commercial network.”
He does not deny that peers are catching up. 智元 began pushing overseas in the second half of last year, and 银河通用 has also been active. But most first-time CEOs are like children playing StarCraft or Warcraft: “You explore the map.” They build an algorithm and robot demo first, then decide where to sell it. Earlier decisions can become later constraints, including the shareholder structure.

36. A few wunderkinds cannot solve it: 曹旭东’s 50/30/20

He pushes back against the current enthusiasm for world-model and World Action Model startups founded by young professors or wunderkinds: “One or two young professors or one or two wunderkinds cannot solve these problems.” The evidence comes from a walk with 曹旭东 at Tsinghua Science Park 10 years ago. What does it take to build a great algorithm? 曹’s answer: “50% data, 30% compute and 20% human experience and inspiration”—with experience ahead of inspiration.
A decade later, he sees the pattern confirmed: “None of the 4 strongest foundation-model companies in China was truly founded by a lone wunderkind or young professor.” DeepSeek’s 100 young people were an important factor, “but it was not enough to make it systematic.” Data, compute and the talent pool must be solved together.

37. One-tenth the data cost, 256 cards used like 1,000

The data economics are brutal. First-person egocentric video costs RMB300-600/hour at market prices. “Set a 1M-hour data target and you have spent RMB300M-600M.” Should a wunderkind raise money on charisma and buy the data? “Your algorithm costs will be 10 or even 20 times everyone else’s. How will you price the product?” His answer is a closed-loop, high-quality data source: 正大’s more than 400,000 full-time employees can push data costs below one-tenth of normal market procurement. The resource owners care most about not disrupting existing operations and keeping process data confidential—“Luxshare or Huaqin would obviously not want to disclose their production-line processes.” The arrangement rests on trust built across 3 startups.
Compute economics are equally clear: DeepSeek was not built by 1 or 2 hundred geniuses alone; “it had tens of thousands of cards.” He is confident on efficiency: “We may be able to use 256 cards to deliver what a company without infra capabilities would need 1,000 cards to achieve.”

38. The talent pool: lessons from AlexNet to DeepSeek’s author list

He uses 2 stretches of algorithmic history to argue that no team stays at the frontier forever. In CNNs, AlexNet from Hinton’s group led to VGG from Oxford, then GoogLeNet; true large-scale deployment came with 何恺明’s ResNet. In detection, the chain ran from R-CNN to Fast R-CNN to Faster R-CNN—he recalls 任少卿 as a key contributor—and then 戴季峰’s R-FCN. “Different teams lead the field for different periods.”
DeepSeek’s technical reports list 100 or 200 authors. Once v2 and v3 began listing individual contributions, it became clear that the main contributors kept changing; perhaps only 1 or 2 people were major contributors in 2 consecutive versions. That is why a large talent pool matters. 杨植麟 has a technical profile and a network across Tsinghua’s Yao class, computer science and electronics departments; 唐杰 has a steady flow of students; DeepSeek has compensation and 梁文锋’s pull. Striding is pursuing multiple channels: students from 汪玉’s group participate during internships, alongside his own appeal. Algorithm head 高峰 was described by several professors as someone whose PhD work was “worth 10 people.”
Can a company stay ahead forever by absorbing diverse teams? “I think that itself may be very difficult.” The national-luck-level performances of Zhipu, DeepSeek and Kimi are unique, and leadership rotates among them. The same is true of OpenAI and Anthropic in the US.

39. Shortening the research-to-commercial-value pipeline

He asked old friends at 商汤 and Megvii what they regretted most over the past several years. The first answer was strikingly consistent: “The pipeline from research to engineering deployment to commercial value was too long.” Even 商汤, Yitu and Megvii could not guarantee that every algorithm would remain state of the art; in the foundation-model era, “they did fall behind somewhat.” If technology cannot remain a moat forever, research and product must be coupled as tightly as possible.
Foundation models validated the point: the model is the product, and once trained it needs relatively little adjustment; customers can switch APIs fairly easily. 曼祺 adds that Anthropic is charging ahead while OpenAI counterattacks in coding, and companies such as xAI have been left behind in the fight. 姚颂 agrees: “The most important capability is organizational capability.”

40. Full-stack in-house development and 许华哲’s opposing answer

Striding AI’s model is a partnership with Tsinghua plus an internal algorithms team. Robot-side software, robot structure and hardware design, cloud teleoperation, solutions and on-site implementation are all developed in-house; it does not manufacture hardware and does not make its own joints or structural parts. “I believe this will be the mainstream direction for embodied intelligence over the next 3-5 years.”
许华哲 argues the opposite: do not build everything yourself; leave it to the ecosystem. 姚颂 answers with PC history. Apple and IBM led in the 1970s and 1980s by building chips, production and operating systems in-house. Standardized interfaces emerged by the end of the 1980s: Intel supplied the CPU, Windows supplied the system and manufacturing was outsourced. “But today, the best overall performance and the most comfortable experience still come from Apple.” Physical AI still lacks standardized interfaces, dimensions and mature suppliers, so vertical integration comes first.

41. Wheeled dual arms, no dexterous hands: algorithms determine degrees of freedom

The initial design is pragmatic. Most work environments are flat; bipedal humanoids require more expensive joints and higher costs. “Once you do the math, that becomes a problem.” Batteries mounted in the torso make movement unstable, while a wheeled base keeps the center of gravity stable and reduces sway during manipulation, making algorithm execution easier. The general-purpose ambition remains: “We are exploring the path to general physical intelligence, and we hope every step opens a flower and bears fruit.”
For now, the company will neither develop nor use dexterous hands. His principle is repeated: “The number of degrees of freedom a robot should have is determined not by its hardware, but by its algorithms’ control capabilities.” Autonomous driving has spent 10 years working with only a few dimensions—forward, backward, left and right—and still has not fully solved L4. Embodied intelligence is “many times more complex”: robots must actively change objects’ physical states; the physical world has physical laws rather than traffic-light-like structure; robots start with 14 or more degrees of freedom, while dexterous hands add another 6-20. Current algorithms cannot stably operate dexterous hands at the 99.9% or 99.99% delivery standard.

42. Every milestone must commercialize

On whether intelligence should come before deployment, he is unequivocal: “Every milestone in the development of intelligence must have a commercial landing.” Autonomous driving is the counterexample. Independent companies committed to L4 have a high mortality rate because every milestone is merely technical and requires investors with faith to keep supplying capital. “This has not even happened in the US. You cannot call Waymo a completely independent company.”
The extreme example of ignoring commercialization is Boston Dynamics: “It keeps producing the world’s most astonishing demos, but it can only survive on continuous transfusions.” His bottom line: “Delivering commercial results will always convince everyone more easily than showing them an astonishing demo.”

43. 3 moments for physical AI: GPT-3.5, breaking out and $40B ARR

He sketches 3 milestones. The GPT-3.5 moment is: “Put a model on any robot, give it any task, and achieve more than 95% or 99% zero-shot success,” allowing 1 or 2 hours of real-robot reinforcement tuning. His own company “is definitely not at the GPT-2 moment; I don’t know whether it has reached GPT-1.” Foundation models are at the level of top 985-university PhD students; physical-world models are still in kindergarten, and general-purpose algorithms are 2-3 orders of magnitude harder than autonomous driving.
The breakout moment would mirror DeepSeek’s open-source, free release in January 2025. Once 60M-70M Chinese users realized that writing poems and Spring Festival couplets was genuinely useful, the industry truly took hold. “Do not insist on saying something is useful only because a benchmark says so.”
The commercial-validation moment would mirror Anthropic reaching $40B ARR in Q1 this year, with reports of quarterly operating profit. MiniMax and Zhipu swapped valuation rankings on coding; “without coding, people would have thought Zhipu’s RMB40B-50B valuation was expensive when it listed in January.” GLM 5.2 combines performance and value. “We hope to participate in and witness all 3 stages.”

44. No envy of other people’s luck: from 1,500x to 1,800x over dinner

Did ChatGPT’s arrival at the end of 2022 shake him? “Not really. I had promised to do something, and my promise was to devote my energy to aerospace.” He recalls a lesson from 2017. A friend who had bought Ethereum as a student and participated in early crowdfunding showed him the phone over dinner: “Oh my god, it’s up 1,500 times.” By the time dinner ended, it had risen again: “Oh my god, it’s up 1,800 times.”
The shock faded after a week or 2. “For someone who built DeepGlint, my luck is already one in ten thousand. The thing you are doing already means your luck is exceptionally good. Your job is to seize the luck you have now.”

45. Make money in the physical world first: scenario autonomy and the Jasper lesson

Unlike foundation models, embodied intelligence “may achieve commercial results first.” Scenario-based autonomous driving is the reference: if L4 is not arriving soon, constrain the task and environment to achieve L4-like performance in closed settings. That produced major breakthroughs in mines—EACON and TAGE Idriver—ports—Mainline and Westwell—and low-speed autonomous vehicles—UISEE, Zhixingzhe, Ninebot and Neolix. His distinction is clear: “I will not build a fully engineered technology stack for a single scenario.” The foundation should remain a general-purpose model, with additions only at the solution layer.
曼祺 points to the digital-world precedent of Jasper, which used GPT-3 to write marketing emails before being covered by the foundation-model providers. 姚颂 adds the story of his classmate 施天林, or Tim Shi, who around 2017 built Krista to collect a user’s linguistic style through conversation and simulate that person, then pivoted to intelligent customer service when the technology was immature. Physical AI is different, he argues: every scenario is large enough.

46. “Language is the world” is impossible, and monopoly is impossible

Could a handful of giants absorb the market? He asks how many automobile companies there are. “Monopoly is difficult in the physical world.” The root difference is complexity. In the digital world, “you” and “I” are 2 concepts in semantic space; in the physical world, meanings are much more varied. “Language is naturally digital, while the physical world is analog. Use language to describe the tilt angle of this cup or the distance between us in meters.”
His scale estimate is extreme. LLMs are now at the “5K-10K parameter level” and could reach 50K or even 100K within 2 or 3 years. How large would a genuinely all-encompassing physical-world model need to be? “5P, 10P or even 100P parameters may not be enough. I cannot imagine this model.” It would be too large to deploy or reason over.

47. The real-time and distillation gap: not just a model problem

One of the biggest differences between physical AI and LLMs is real-time performance. “If you cannot reach 30 frames, the system has no meaning,” and deployment will probably have to happen on edge chips. In the CNN era, one could pursue accuracy first and prune and compress later; large-model distillation only reduced parameter counts by a few times. In physical AI, “the model you distill from and the model you finally deploy may differ by 10,000 or even 1M times in parameter count.” The deployable edge models are 2.5B, 5B and 10B.
Pursuing performance first and optimizing deployment later is “fine,” but it is OpenAI’s path. “You need to raise money on the scale of $10B,” then burn $400M-$500M a year using thousands of cards to grow parameters from several billion to GPT-3.5’s 110B. Some companies will “have the disease of OpenAI without its fate.” Fundraising will be the life-or-death issue for many technology-focused founders. OpenAI’s renewed robotics team in Q2 this year is positive for every Chinese company building the brain: China and the US are the only 2 countries that can produce innovation in foundation models and physical intelligence, and China can quickly evaluate and follow new ideas.

48. The latent-space route: learn physical laws, discard the small terms

Striding’s own foundation model will broadly follow a latent-space approach. The first principle is simple: the model needs to learn 2 things—the laws of physics, including elastic collisions, conservation of momentum and Newton’s 3 laws; and physical common sense, such as knowing that a paper cup is light and soft, that aluminum is probably not heavy and that gold is heavy. “The text printed on the paper and the color of my clothes have no bearing on robot manipulation.” The model should understand the world in rule space rather than pixel space.
He attributes this direction to physics-competition training rather than sentiment. Mathematics competitions seek to calculate without omissions; “physics research requires you to decisively discard small quantities and retain only the main structure, otherwise you cannot compute the physical world.” 曼祺 summarizes the latent space as “the correct kind of ambiguity,” which he accepts.

49. Why WAM is so hot: the overlap of 2 historical moments

He sees 2 emotions behind the aggressive fundraising by world-action-model companies in the first half of this year. The first is the 商汤/Megvii moment: deep learning pushed face recognition to several nines and advanced from ImageNet to MS COCO to KITTI, showing a technical path with an extremely high ceiling. The second is the moment associated with “壁仞、燧原、沐曦”: after Horizon Robotics and Cambricon succeeded, new entrants arrived with new approaches. “It is a possibility that could push human physical-intelligence algorithms forward by a major step.”
Is it a bubble? His historical lens: Horizon Robotics had around RMB100M of revenue and a valuation above $3B in 2018, and everyone thought the bubble was enormous; in 2020, 余凯 posted about shipping 100,000 chips; after the 2024 IPO, the company’s market cap peaked above HK$150B; Momenta’s latest round was reportedly above RMB100B. “If you can survive until mass deployment generates commercial value, none of these companies are bubbles.” He accepts 曼祺’s rebuttal: “You are talking about survivors.” “Yes”—many companies pursuing an ultimate L4-style model may disappear along the way.

50. Who survives: the fundraising choke point and the operating-business path

The market is crowded. The number of leading companies pursuing general-purpose models is “certainly more than 40 or 50.” According to the NDRC’s figures at the beginning of the year, more than 100 companies had launched more than 200 humanoid robots, many focused mainly on hardware. His survival rule for pure technology idealists is blunt: “The few with the strongest fundraising ability will survive, so all your work should be for fundraising—for example, accepting more media interviews.” He calls himself “strongly introverted” and says, “This is my first time recording a video like this. There will not be more media interviews afterward.”
The other path is the one he is backing: companies with an operating business that continually iterate their algorithms, following the Horizon Robotics, Momenta and Huawei model. “Nobody starts by building L4. First get an L2 product used at scale, then gradually push toward L3; when policy is ready, your L4 will also be ready.”

51. Every major company enters: surpassing autos in 10-15 years

His forecast is that every Chinese manufacturing and internet company will enter the physical-AI wave over the next several years, with strategic-level investment. Internet giants want to penetrate the physical economy from the digital economy. Every automaker—Li Auto, XPeng, Seres, Xiaomi, Huawei, BYD, Geely, BAIC, SAIC and GAC—will participate, as will handset and supply-chain companies—OPPO, vivo, Honor, “this year’s marathon winner,” Lens Technology and Dinglong—and heavy-equipment companies such as Zoomlion and Sany Heavy Industry.
The industry consensus is that output will exceed the auto industry’s in 15-20 years. “I’m even more optimistic: perhaps 10-15 years.” 高继扬 told 曼祺 that embodied intelligence would certainly become “the king of all industries.” 姚颂 goes further: “It will absolutely become one of the economic pillars.” Robots will penetrate the tertiary, secondary and even primary sectors.
He dismisses a Matrix-style virtualized end state: “That is nonsense.” His view comes from studying neuroscience: consciousness remains scientifically unexplained, nobody knows where memories are stored, and Musk’s brain-computer-interface efforts treat the brain as a black box. Ready Player One is possible; The Matrix is unlikely.

52. The industry’s disorder: look for the tiny text in every demo

He describes the most common form of misrepresentation in practical terms: many companies use teleoperation videos to pass off autonomous-operation demos. When he watches any public video, he first checks whether the screen says 1x, 5x or 10x—most VLA and world-model videos are sped up by 5, 10 or even 20 times. Second, he checks whether the video explicitly labels the operation as fully autonomous. Every Striding demo video must carry both labels.
Is it acceptable to bend the rules slightly for a strategic objective? “Not for me.” There are both ethical and practical reasons. Ethically: “There are things I can choose not to say, but anything I say will be true.” Practically, deception eventually backfires—like the blood-sample and blood-exchange claims, or the VR company that put a whale in a stadium. “You cannot hide it forever; you can only say you were lucky if you eventually deliver what you showed.” Fake it till you make it? “I do not do that.” The cost is real: investors may not understand why he refuses a sexy, fancy label or why the company is not a world-model company. Striding has completed an angel financing series of nearly $100M, with 70 people and a complete full-stack R&D and commercial organization.

53. No extinction-level disaster now: the fourth industrial revolution will disrupt the third

What could cause an extinction-level disaster for embodied intelligence? “That possibility no longer exists.” His macro view is that the fourth industrial revolution may be large models and AI technology overturning the products of the third industrial revolution. Many high-value jobs in the digital economy “will very likely be comprehensively disrupted by foundation models”; mass software-development layoffs in Silicon Valley and China’s internet companies are an early sign. Penetration into the physical world will take longer—“perhaps a 10- to 20-year cycle”—but factories, services and agriculture will inevitably be entered.
His evidence that the bubble is manageable is valuation dispersion. Valuations already differ sharply between China and the US, and another valuation gap separates foundation-model companies from physical-intelligence companies. “We have not yet seen companies with a few hundred million in revenue shouting about valuations of $18B or $20B.” He expects that in roughly 2 years, people will encounter real-world situations nearby and feel that “this is an intelligent robot providing me with a service.” There will be volatility and adjustment in the next few years, but “the amplitude will be small and the cycle short.”

54. 80% offense, 20% defense: the balance for a serial founder

Why does he define himself as “an entrepreneur”? Physical AI requires everyone to start learning again, and “people born in the 1990s, 1985-89, 1995-99 and 2000s are equal at this starting line.” His difference from a first-time founder lies in long-term operating discipline. In rockets, he had no financial dealings with leaders of state-owned enterprises or government officials and avoided overclaiming. “Many first-time founders are 100% offense and 0% defense. I’m 80% offense and 20% defense.” What does he defend against? “I will not go in, and I will not become bankrupt. I am Chinese; I do not want to be forced to live in another country one day.”
Could a barefoot 100%-offense founder beat him? His answer is about time horizon: neither industry is winner-take-all. “Look at 2 years: he is 100 and I’m 70. Look from 5 years out: he is 100 and I’m 200.” The objective is sustained fighting power, not the highest valuation in the shortest time.

55. The push of the era is an elevator: underestimate the era, overestimate yourself

To founders who entered after 2023 and feel strongly pushed forward, he offers the episode’s most serious warning: “If they cannot clearly recognize that this momentum comes from the country needing a company like theirs rather than from their own effort, they will run into major problems.” Since reform and opening, many entrepreneurs “underestimated how important the era was to their success and overestimated themselves.” You feel the push because you are in the elevator; support from the era will not stay in one industry for 5 or 10 years. It comes in phases.
The most clear-eyed statement he has heard came from a compute-company founder: “I know very clearly that my company went public because of the support of the Party and the state. It does not necessarily mean I have already done a good enough job.” He leaves himself the same margin. After rereading papers and reorganizing his technical understanding, “I feel more than 90% is correct, but it may turn out to be wrong in the future, and some things may become baggage.”

56. Defining success and the origin of “正行”

Were his first 2 startups successful? “They were OK.” The first proved the complete chain from lab code to deployment in millions of cars; the acquisition itself was external feedback that “what you were doing was right.” The second was outside his original field, but strengthened his ability to learn, formulate strategy and organize systems engineering. “Having a rocket launch successfully with my participation was already a major reward.” The third is different: “I have begun to move beyond the mindset of a young person.” This time he wants a long-term outcome—sustained operations and an eventual listing—while participating in physical AI’s GPT-3.5 moment, open-source breakout and killer-app moment.
The company’s name echoes those 2 minutes of pacing. Striding means walking in large steps, “not small steps.” The Chinese name 正行 adds a value statement: “We must act with integrity.” “Innovation” completes the name because “I’m a company that is difficult to define with the labels that came before, like DJI. You can only summarize it as innovation.”
His delivery commitment for the next 12 months is concrete: begin deployments in 2 target environments this year, then refine the product through 2 or 3 rounds over roughly 1 year and genuinely provide services. “We want people to feel that humanoid robots and embodied intelligence are not just dancing anymore.”

Ant Lingbo's 沈宇军 on Robot-Native Foundation Models and Data Scale-Up 🔬Causal Models Need Causal Data - Xaira’s X-Cell model (Bo Wang & Ci Chu)