Post-2000 Founder 陈源培: RoboMaster to 李飞飞's Lab, Two Global Firsts

Summary

• 陈源培’s core thesis: in embodied intelligence, algorithms are not a moat—“virtually every approach can be reproduced if we want to reproduce it.” The real barriers are data and a closed-loop, end-to-end system: in frontier research, “if you give me 10 ideas, I think all 10 could be right,” so the only answer is to test enough of them and build the loop well enough that “the results speak for themselves.” Tell people the method, he says; “if the result you get is still not as good as mine, that’s what a good system looks like.”
• The scaling law for embodied intelligence lives in data, and he is betting on human data. Pure teleoperation demonstrations cannot scale; the field needs motion-capture gloves to collect data from people’s everyday manipulation—he does not know how many hours will be needed to produce a scaling law, but “a million hours should definitely show some effect.” The biggest reason to build dexterous hands is precisely that “they can use human data, because the gap with the human hand is not that large.” The fundamental difference from LLMs is hardware and the embodiment gap: machines can share human know-how, while differences in joint force and other mechanics must be filled in through reinforcement learning.
• He calls himself a simulation native, but does not see simulation as the endgame. Differentiable simulation “works for simple tasks, but on complex tasks, one collision and it immediately blows up.” The main paths to a breakthrough are either another order-of-magnitude increase in compute or making world models work. World models face a chicken-and-egg deadlock: “Which needs more data, VLA or a world model? Even if that thing really exists, you still have to go through VLA first.” Musk does not talk about world models because “they do not generate value on their own.”
• His definition of an embodied-AI bubble is one line: “Too many people are saying something different from what they are doing.” Many demos are built with imitation learning but presented as reinforcement learning. He can tell the difference through “experience-based intuition,” although the industry has little ability to falsify such claims. The only real test is to do something others cannot—such as throwing and catching 3 balls with 2 hands, a task for which imitation learning cannot even collect the data. His view of Figure: the demo is “real,” but the success rate is certainly much lower than advertised, which is why the company rarely lets people watch it directly.
• His endgame call: every embodied-AI company—especially those training their own models rather than targeting fixed scenarios—will eventually build hands; “there’s no way they won’t.” His first-principles evidence is simple: “Why did Musk start by building such a complicated hand?” Grippers can work in fixed settings, but one that works in Factory A may fail in Factory B; cross-scenario generality cannot rely on grippers alone. A general-purpose brain could technically work across bodies, but commercially he thinks soft- and hardware may remain coupled for much of our lifetimes. Hardware will dominate costs in the short term; over time, “the brain will definitely account for more.”
• What robots have learned so far is not real intelligence—“Of course not; it’s all imitation.” There is no true generalization, at most compositional generalization, and machines remain inside the distribution of worlds they have already seen. Phrases such as “native world models” are “mostly hype”: with the same data, loss and training process, merely changing the model architecture “cannot produce a breakthrough.” The real target is an embodied version of next-token prediction, and he is betting on “modeling manipulation as object trajectories” as a universal representation.
• His personal arc is itself the argument: there is no single paradigm for learning, and no standard answer for innovation. He studied civil engineering as an undergraduate, taught himself reinforcement learning from scratch and completed a paper in 3 months. As introduced at the start of the episode, he completed the world’s first robotic Great Wall manipulation task and, at Stanford, achieved the world’s first demonstration of training a robot’s 2 hands with human data. 陈源培 says he does not know whether he was the only Chinese undergraduate admitted to 李飞飞’s lab; according to a senior student, there were many Stanford master’s applicants, but 陈源培 was the only one ultimately selected. His life principles are “don’t regret” and “it is absolutely impossible not to be happy.”

Deep dive

1. A Post-2000 Self-Definition: “Cannot Be Summarized”

• Asked about generational labels for people born after 2000, 陈源培 says the defining trait of the post-millennial generation is that it “cannot be summarized.” The internet has created room to experiment and imagine, and the differences between individuals are enormous. For him personally, being post-2000 means “more room for trial and error”; beyond that, people should judge by the product or by technical ability.
• When investors argue that post-2000 entrepreneurs lack the strength and accumulated credentials of famous scholars, his response is: “They’re right.” He is nowhere near genuine success yet, and for now the attention is driven more by capital and buzz. His definition of success is impact: “Valuation or money doesn’t really matter. It’s more about the impact on human society.”

2. Gaming Prodigy and Research: The Same Upgrade Loop

• The host recalls that 陈源培 was still playing cards in his dorm the day before the gaokao. 陈源培 says he played cards right up to the exam, then games throughout the period before university; his parents never intervened. “The best and most influential thing they did for me was to leave me completely alone.” He logged thousands of hours across FPS and MOBA games, nearly joining an esports club’s youth program before middle-school exams, and ultimately scored well enough for “an ordinary 985 university.”
• The connection between gaming addiction and research obsession is straightforward: “They’re actually quite similar.” Research is harder and has more varied metrics, “but in both cases you can keep leveling up. As long as you keep doing it, you can keep getting better.”

3. Why He Started a Company: An Academic Reckoning Triggered by the Russia-Ukraine War

• The epiphany came in late 2023, while he was still doing research at Peking University. The Russia-Ukraine war had begun; people were still fighting and children in impoverished mountain regions were still suffering. “We spend every day picking at tiny innovations just to publish papers—how much does this actually help society as a whole?”
• His conclusion was direct: “The gap is simply too large. From that point on, I felt academia definitely wasn’t where I wanted to go.” That realization became the early origin of his startup.

4. RoboMaster: The Opposite Pole from Academia

• DJI’s RoboMaster competition was dominated by his university, which won 2 national championships and never fell out of the national top 3 during those years. Students built everything themselves, from mechanics and embedded hardware and software to circuit boards and supervisory control code. “Other than eating and sleeping, we spent all our time preparing for the competition.”
• He sees the competition and research as “2 extremes.” Engineering rewards attention to detail and speed of learning, not advanced theory; the task is to combine technologies into a system everyone can use. A civil-engineering undergraduate who “barely attended class,” he joined by teaching himself on the spot for timed algorithm-reimplementation tests.
• That year, the team explored a robotic-arm solution for a gold-mining engineering robot. Starting from zero, he built a 6- or 7-axis arm and worked down into mature code, including conventional motion control and motion planning. That habit became the source of his later speed advantage.

5. 3 Variables That Took Startups from Fringe to Mainstream

• In 2021-22, when he was competing, “starting a company was still a fringe option.” When foundation models emerged, the people attracting attention were mostly already-famous scholars and scientists. The shift since then, in his view, came from 3 forces: government support for hard tech; increasingly powerful AI tools, which let people learn on the job quickly and sometimes perform as well as experienced practitioners; and capital markets becoming more willing to back young founders.
• RoboMaster, he believes, was an early model for AI startups: team members devoted themselves fully to competition, while the previous year’s core members moved into architecture and management after the season and directed younger teammates. The structure was flat, mission-driven and free of KPIs. “When everyone is working extremely hard, it becomes very difficult for anyone not to work hard.”

6. 杨耀东’s Test: Learning Reinforcement Learning on the Spot

• Getting into 杨耀东’s group at Peking University involved luck and timing. 杨老师 had just returned to China, and 陈源培 was effectively his “second student”; one email was enough to get the opportunity. He then finished the assignment quickly: “Use reinforcement learning to build a dexterous hand.” He had never touched RL before and learned everything on the fly from Zhihu, CSDN blogs and code. “Some things didn’t work at first; after changing them a bit, they worked.” Many students could not complete the task; he finished it in a very short time.
• The speed came directly from RoboMaster: “No one had ever taught me anything. I had to look up everything myself, go all the way down to the lowest level of every piece of code, understand how it worked, and then implement it.”

7. A Paper in 3 Months: No Mentorship Was Faster

• His first paper, on using reinforcement learning for 2-hand dexterous manipulation, took only 3 months; everyone assumed it had taken 1 or 2 years. The PhD student supervising him held several meetings, then moved on to another project after a week. 杨耀东 called and said, “源培, this project is yours from now on.” In retrospect, “the writing and all the design were pretty haphazard, to be honest. But it was accepted, and the score was quite good.”
• The experience shaped his view of mature systems, a lesson he now gives interns: “Having a mature system around you can make things faster, but it can also constrain you. When I was doing it, I had no constraints at all. I could do whatever I wanted.”

8. The Paper’s Real Contribution: Showing That GPU-Parallel Simulation Plus RL Worked

• 杨耀东’s initial view was that the task “couldn’t possibly be trained with a single reinforcement-learning algorithm; it definitely needed all kinds of additional techniques.” Historical experience supported that view: in the MuJoCo and PyBullet era, sample efficiency was inadequate. “These were things you couldn’t even train at the time.” The result not only trained successfully, but covered 20 tasks and nearly 50 degrees of freedom across the 2 hands.
• His retrospective deliberately strips away the halo: “Looking back now, there wasn’t that much of an algorithmic breakthrough.” The real value was “showing the entire academic community that reinforcement learning could accomplish all kinds of things in large-scale parallel simulation.” The 2-hand dexterous-manipulation benchmark he built along the way became his calling card, and many later projects were derived from his code. “I really hadn’t thought that far. I just built it while I was at it.”

9. Shadow Hand’s Rise and Fall, the Isaac Inflection Point, and a Typical Problem-Solving Instinct

• OpenAI’s 2019 Rubik’s Cube breakthrough using RL and Shadow Hand made headlines; the team disbanded in 2021. “Learning-based methods weren’t quite ready for fancy things like that.” It required a huge CPU cluster, “was very expensive, and was hard to scale further.” The turning point was Nvidia Isaac’s GPU-parallel simulation, which let ordinary consumer GPUs run thousands of environments. The team has recently re-formed; his explanation is that the original dissolution reflected a judgment that the data was insufficient, whereas “with so much data now, it’s the natural thing to do.”
• He calls himself “one of the earliest users of Isaac in China,” holding original versions such as Preview 2 that he found by casually browsing developer forums. The releases were buggy and often failed to run, but his conclusion was firm: “CPU parallelism will never beat GPU parallelism, and simulation parallelism is extremely important. GPU-based simulation is clearly where things are headed.”
• The host highlights his instinct for solving problems; his explanation again comes down to having no mentor: “Looking back, no one ever truly guided me. If you don’t do it yourself, there’s no way around it.”

10. The Infant-Development Experiment: An Honest Falsification

• One paper had the 2 hands imitate task sequences performed by a 24-month-old infant, motivated by the question of whether a robot could develop “like a human.” The team consulted specialist material, sorted infant actions by developmental stage and trained them one by one with reinforcement learning. The unexpected conclusion: “There wasn’t much of a relationship.” Training early-infant tasks was no easier than training complex manipulation; “it was more about how the reward was written, or about the algorithm.”
• The host juxtaposes that result with 李飞飞’s ImageNet intuition that infants acquire intelligence by accumulating visual signals. 陈源培’s conclusion is that, at least under this reinforcement-learning setup, robots learn dexterous manipulation “in a machine’s way, not a human way.”

11. Reinforcement vs. Imitation: A Route Battle with His Senior Colleague

• The biggest argument at Stanford was with a senior colleague whose imitation-learning approach “worked extremely well.” The colleague teased him: “I collect data for 2 days and I’m done, while you’re still running simulation, writing rewards, training reinforcement learning and doing sim-to-real.” His rebuttal was an endgame argument: imitation learning does not generalize. The senior colleague admitted that was true “at least at the time.” In a real home environment, there would inevitably be unseen situations for which no relevant data could be collected, whereas RL could let a robot explore them in a simulator.
• The colleague’s response was: “When the time comes, the robot can watch a person do it once and learn.” He was unconvinced: “That’s all speculation. I don’t think it’s possible—or rather, there will always be things you can’t do.” The field was “quite confused” then and far less converged than today; the host adds, and he agrees, that practical implementation remains extremely confused even now.

12. Demo Authenticity: The Industry’s Most Common Gray Area

• The host recounts industry discussion around a famous company whose launch-event demo was widely believed to use imitation learning, even though the company claimed it used reinforcement learning. 陈源培 confirms how common this is: “No need to think twice—many.” Many demos are in fact built with imitation learning. He can usually tell through “experience-based intuition,” built on a deep understanding of what control, reinforcement and imitation methods can each achieve.
• That observation sets up his later definition of an embodied-AI bubble: “A bubble is when too many people say something different from what they are actually doing.”

13. The Proprietary Training Method: Watch the Robot Learn, Not the Curves

• “Basically everyone used to watch the curves. I was the only one who opened the simulator and watched the robot learn—watched it go from knowing nothing to knowing how to do it.” When a run failed, he would imagine what the policy looked like at that moment and why it was drifting left, make an adjustment and run it again. Others typically left the program running, went to eat and checked the result later. “They said no one had ever done it this way.” Karen 刘 once asked 王晨: “Why does 源培 like tuning rewards and watching all this stuff so much?”
• The approach mirrors the LLM industry’s view that training a model is ultimately detail work: “I can tell you the method, but the result you get still won’t be as good as mine. That’s what a good system looks like.”

14. The Drawer-Handle Reward: The Seed of Object-Centric Thinking

• The most memorable counterintuitive case involved a robot hand that kept trying to insert its fingers into the hole of a drawer handle, even though “sometimes it didn’t need to insert them; it only needed to pinch the handle.” Locking down every joint and pose made training and generalization harder, since every drawer is different. “The reward for the hand doesn’t need to be written with extreme precision. It only needs to know it has reached roughly the right area, then it can explore within that small range.”
• The host points out that this was the early seed of his later work moving “from human-centric to object-centric.” He agrees: “There was some intuition at the time, but it wasn’t clear yet.”

15. Entering 李飞飞’s Lab: According to a Senior Student, Only 1 Undergraduate Was Taken

• His first paper helped him get into Stanford. 陈源培 says he does not know whether he was the only Chinese undergraduate; according to a senior student, there were many Stanford master’s applicants, “but in the end the lab took only me, because they thought the paper was pretty good and looked fairly cool.”
• Before that, his international travel had mostly been tourism—he mentions Thailand—and this was his first formal academic exchange. Karen 刘 asked him to “talk to everyone in the group,” even though the paper had already been completed online and his KPI was theoretically done. What stayed with him most was the atmosphere: people worked on everything from embodied AI to brain-computer interfaces, and “everyone was very open; no one said any particular direction was definitely impossible.” 李飞飞 was “basically already doing her own thing and preparing to start a company,” but the lab’s model—enough resources, strong people and room to explore—benefited him.

16. A Senior Student’s Wake-Up Call: Leave the Comfort Zone

• A senior student identified his problem: “You’re too good at reinforcement learning, so you want to use reinforcement learning for everything.” In theory, anything could be trained, and “if I really wanted to publish a paper, I could probably get a top-conference paper out with RL in a week.” But that was his comfort zone: “As a real researcher, you should be trying to step outside your comfort zone. What you’ve done before often limits what you can do next, but that is actually very hard to overcome.”
• He acknowledges that this runs against human nature and inertia. When technology has not converged, every direction is a bet of time and resources. But “the way an entire field is built subtly shapes a person over time.” That was what he learned most at Stanford—not any particular technology.

17. A Map of Embodied AI: A Spiral That Keeps Rising

• His map of the field is organized around the tension between data and algorithms: CV and model training cannot solve grasping → collect trajectories for imitation → human demonstrations alone are too slow → reinforcement learning explores autonomously → real-robot training is too inefficient, with even a lab using more than a dozen robots unable to train RL effectively → simulation and sim-to-real, alongside Isaac’s rise → the sim-to-real gap becomes too large → Diffusion Policy and ACT revive imitation learning → after VLA takes off, everyone adds reinforcement learning back in. “It’s a cycle, layer by layer.”
• His conclusion is a recurring view of technical history: “Technology advances in a spiral. No algorithm is redundant; each exists for a reason.” The next phase may be world models, which are themselves simulators: “You may end up going back to training reinforcement learning inside the world model.”

18. The World’s First Robot-Hand Training with Human Data: Gaps Created by Weak Cross-Disciplinary Links

• The program describes this work as a world first. 陈源培 says the direction “should have been obvious to everyone intuitively,” but around 2023, after discussions with many Chinese academics and graphics researchers, he “couldn’t find a single similar piece of work.” His explanation is insufficient cross-disciplinary exchange: fancy human-motion graphics had long used motion capture, “but robots hadn’t reached that stage.” His team happened to be early in both dexterous hands and simulation, so “the connection happened naturally.”
• The method used a person wearing a motion-capture glove to manipulate objects, then fed video, trajectories and temporal data into simulation for reinforcement learning and sim-to-real. “Many companies are doing this now.” It was also the source of every technical direction at his company, including its in-house glove. The final answer to the dexterous-hand debate is simple: “The biggest reason to use a dexterous hand is that it can use human data. The gap with the human hand isn’t that large.”

19. The Ceiling of the Simulation Camp: What Differentiable Simulation Cannot Do

• Responding to simulation advocates who say the sim-to-real gap can be solved, he says: “I’m a simulation person. Most of my work is sim-to-real.” It can be reduced, but “not particularly well,” because of constraints in graphics. Differentiable simulation—encoding rules so gradients can be backpropagated, putting in a real-robot trajectory and automatically adjusting simulated friction from 1 to 0.5—is popular in theory, but “works for simple tasks; on complex tasks, one collision and it immediately blows up.” The bottleneck is computational complexity.
• He sees 2 main paths to a true breakthrough: “another new order of magnitude in compute,” or making world models work, either as differentiable simulators or as synthetic-data generators. Simulation “can produce some very fancy demos, but it definitely is not the true endgame. At least, we haven’t seen that yet.”

20. Object-Centric Modeling: Representing Manipulation as Object Trajectories

• The underlying motivation is practical: “I don’t want to write a reward for every task.” One reward for grasping, another for placing and another for opening a microwave is not general. Whether the hand is human or robotic, “operations that look complicated are really just changes in an object’s trajectory.” Modeling manipulation as an object moving along a trajectory at the corresponding times is “a universal way to represent operations,” and could unify rewards.
• The host compares this with 李飞飞’s use of ImageNet to define a general-purpose problem. 陈源培 agrees and extends the analogy: “Language is next-token prediction—the way all languages can be handled. We may not have found the equivalent for vision, but object trajectories may have that potential for manipulation.” The architecture follows the foundation-model playbook: use imitation learning to absorb large datasets and train a large model, then use reinforcement learning for fine-tuning. “The big picture hasn’t changed: from general to specific. The 2 algorithms have different strengths and weaknesses, so they need to be combined.”

21. Building with Lego and Demystifying Academia: “Academia Needs a Story”

• The reality behind the “Sequential Dexterity” paper was a long-horizon task of “having a dexterous hand build Lego,” involving 5 or 6 steps. It took a year, and only converged after trying “all kinds of messy approaches,” including hierarchical reinforcement learning and open-loop policy chaining. The paper’s narrative was imposed in the final month before submission: “To get the paper accepted, I actually had to force myself to find so-called novelty points or a story to tie it together. Otherwise, I couldn’t even publish the paper.”
• That gave him “a clear understanding, or perhaps disillusionment, about academia”: “Academia is not pursuing the best final result; it wants you to tell a story.” Its operating rules “sometimes do not favor genuinely impactful work.” The system itself, however, held up: after the paper was accepted and the equipment dismantled, visitors came and the team rebuilt everything on-site; the original model still worked. “That is extremely important for the reproducibility of the entire system.”
• The ultimate explanation for success remains systemic: “If you build the system well enough, you can do many things that were previously impossible. Everyone knows the algorithms. The difficulty is in every single detail.”

22. 2 Phrasing Lessons from Stanford

• A senior student taught him a communication lesson: “You say 1+1=3, and I say, ‘Wrong, 1+1=2,’ or ‘Right, but 1+1=2.’ The facts conveyed are exactly the same,” yet the impression is completely different. Saying whether something is right or wrong first carries no information. “Attitude doesn’t matter; facts matter more.”
• Karen taught him not to write phrases such as “our results are impressive” or “our generalization is extremely strong” in a paper. If the reader disagrees, the adjective is a negative; if the reader agrees, the reaction is, “Something this ordinary looks so strong in his eyes?” The lesson shaped how he views the field: he dislikes categorical verdicts because that is important for preserving curiosity and respect. “Even a terrible paper has something you can learn from.”

23. The Embodied Scaling Law: The Answer Is Data; the Variable Is Hardware

• Asked when embodied AI will produce a scaling law, he answers: “Wherever the scaling law is, the data must be there.” Teleoperation alone will not be enough; only human data can reach the required scale. How many hours? “I don’t know. We’ll know only after we do it. But I think a million hours should definitely show some effect.” At the time, a few hundred hours of motion-capture data was already considered enormous; previously, “people might start training with just 50 data points.”
• The fundamental difference from LLM scaling is hardware: the embodiment gap. Human and robotic hands differ; 4 fingers and 5 fingers differ; joint counts differ. How do you transfer across bodies? The key is to “extract the general information embedded in human behavior” and share that directly. “Exactly how to grasp something and how much force to apply at each joint will obviously differ between a human and a robot. Let it learn and fill in that part through reinforcement learning.” A purely virtual language model “never has to think about this.”

24. Starting a Company in China: No One Was Talking About Reinforcement Learning in Mid-2024

• He had been preparing to apply for a PhD, with “some probability” of getting in, but introductions from 杨耀东 and others led him into entrepreneurship. The motivation was still impact. His goal echoes OpenAI’s slogan: “Achieve artificial general intelligence and benefit all of humanity.” Applications are only the vehicle. His first-principles reason for choosing embodied AI is that “what can truly change the world is a fundamental leap in productivity”—the fourth industrial revolution, in which general-purpose labor is no longer scarce, underpinned by breakthroughs in compute and energy.
• One counterintuitive market detail: when he surveyed the market in mid-2024, “no one in the entire market was talking about dexterous-hand manipulation, and no one was talking about reinforcement learning.” o1 had not yet launched. “Everyone said reinforcement learning was dead; ChatGPT showed that supervised data was enough.” The market focused on force control, simulated grasping, trajectory prediction and various simulation-plus-real-robot approaches. His team committed to human data plus RL at that point, which he says was “extremely important for an embodied company—or for a startup.”

25. The Only Way to Falsify a Claim: Do What Others Cannot

• In a confused market, “if you do the same thing as everyone else, it is very hard to falsify anything.” The only answer is to produce a capability others cannot imitate. His test is to “go beyond what human teleoperation and demonstration can do”—2-hand juggling, or in an extreme case, throwing and catching 3 balls. “The teleoperator has to watch the robot while operating it, so the process is very slow. You can’t collect that kind of data.” Imitation learning is fundamentally imitation of humans; it cannot perform tasks beyond human ability.
• The next breakthrough in brain-level capability would be equally concrete: “You have a robot, and anything you tell it to do, it can do. The success rate might not be that high, but it can do everything.” So far, he sees no “particularly discontinuous” progress overseas.

26. How He Views Overseas Companies: Human-Centric vs. Robot-Centric

• Companies drawing attention include Generalist, Figure and Sunday. Figure is the most secretive, and its demos are often questioned. From an algorithmic perspective, he believes “it’s real—that is something we could do too, and it isn’t difficult to train with the right equipment. It’s just that the success rate won’t reach 100%, and the generalization won’t be that good. It is more likely to be specific to a particular scenario, which is why they rarely show it directly.”
• He sees some similarities between his company and Generalist. He also mentions Real List’s overall data pipeline and Sunday’s in-house glove, which it deploys in many locations to collect data. The companies may target different scenarios, but the core divide is between UMI-style robot-centric systems and their own human-centric approach: “We will definitely make it possible to collect data from people’s everyday lives without needing dedicated data-collection staff.” What if the route is wrong? “That is definitely possible. Then you need to adjust very quickly, without baggage. You can’t say that because you’ve always done reinforcement learning, you won’t use something else when it works. In the end, the results speak.”

27. The Endgame for Grippers and Hands: All 3 Routes Lead to Hands

• The market has 3 camps—grippers, single hands and 2 hands—but his view is unequivocal: “Trust me, every company working on embodied AI, especially those claiming to train a foundation model, will build hands in the future. There’s no way they won’t.” The first-principles evidence is: “Why did Musk start by building a hand, and such a complicated one?” At the endgame of cross-scenario generality, “you can’t solve it with a single gripper.” A gripper that works in Factory A may fail in Factory B, and it certainly cannot cook fried rice at home. Even top North American companies have not found a high-confidence route. “When a real solution finally works and scales, I think they’ll switch to hands. They may simply not have the bandwidth yet.”
• In an ideal setup, a general-purpose brain could control 5-finger hands, 4-finger hands or even grippers. “Technically it’s possible, but commercially it may not work. I’m still thinking about that.” Embodied AI sits somewhere between autonomous driving and foundation models; “at least within our lifetimes, I think the things we see generating real commercial value may still be tightly coupled hardware and software.” The sequence will likely be a model first achieving broad capability or a major breakthrough on one body, then being transferred to other bodies. He sees that as “a very important moment.”

28. Algorithms Are Not a Moat: All 10 Ideas Could Be Right

• His recurring thesis is that “virtually every approach can be reproduced if we want to reproduce it, but I don’t think the final result is fundamentally determined by the approach.” Imitation or reinforcement, data mix, backbone, world model or VLA—“these things are not important; you can change them very quickly.” Once a company reaches a certain level, algorithms “are not a moat. It is very hard for you to be the only person in the world with an algorithmic breakthrough that nobody else understands, and then defend it indefinitely.”
• His methodology is: “If you give me 10 ideas, I think all 10 could be right.” Brain science, video data and human data all have a rationale. “You can only test them: build the entire loop well enough, test enough of them, and let the results speak.” The 10 routes may all be partly right and partly wrong, each with something useful; the question is how to combine their strengths. That also explains why his company does not advertise world models: “What you say you’re building is never absolute. A year later, it may no longer be called a world model; it may be called a reward model. If it really works, we can use it.”

29. Demystifying World Models: Hype, Jigsaw Pieces and Musk’s Silence

• On terms such as “native world models,” his honest reaction is: “It feels more like hype.” The host notes that the underlying technology still relies heavily on DiT (Diffusion Transformer) from image and video work, and describes approaches that output actions and future frames together, or first predict the future and then use inverse dynamics to output actions. 陈源培 agrees that the US has not found a fundamental breakthrough either. His strong claim is: “If the data stays the same, the training loss stays the same and the training process stays the same, I don’t think it’s possible to get a breakthrough-level change simply by changing the model architecture or how the modules are connected.”
• He agrees with the host’s analogy of a world model as an infinitely large jigsaw puzzle: no one will unify the field in the short term; different routes will eventually intersect to form a larger picture. His more pointed question is: “Why doesn’t Musk talk about world models? Because world models do not generate value in themselves. They are either a concept or a method. Going to Mars, energy and embodied AI create actual value; the world model itself does not.”

30. Active World Models and the Rejection of “Real Intelligence”

• Another scientist at the company, 温莹, drew a 3-layer theory of world models from a book: the objective laws of the world, meaning a deep understanding of physics; one’s own policy and imagination; and the missing middle layer, an “active world model” that decides which information is useful. “If I want to pick up a phone, I extract the phone’s shape and position. There’s a microphone next to it, but I don’t need to know that. It’s noise.” The issue is signal-to-noise ratio: feeding a model data with a lower signal-to-noise ratio than its capacity “will actually reduce its capability.” He immediately qualifies the idea: “This is just one possibility. It might work; we’ll have to see.”
• Asked whether robots have learned real intelligence, his answer is blunt: “Of course not; it’s all imitation.” There is “no true generalization, at most some compositional generalization”: a robot that has done AB and CD may be able to combine B and C, but it still needs to have seen the components. It is still generalizing within a known world. His view of borrowing from brain science is pragmatic: neural networks were built by computer scientists because they worked, not because they resembled the human brain. “Whether these theories actually make sense, you have to try them. Practice is the most important thing.”

31. The Simulation Camp vs. the Real-Data Camp: A Chicken-and-Egg Deadlock

• On companies that collect no real-robot data and rely entirely on simulation, he says, “I respect them,” but rejects their speed arithmetic. Simulation may reach a high volume of data quickly, “but it isn’t necessarily faster on diversity.” With 100 scenarios, “I can film 100 people and may have data from 100 scenarios in a single day.” Simulation advocates need real-to-sim: rebuild each scene in a simulator and then collect the data. That is “extremely dirty work,” and scene-generation capability is currently far slower than simply collecting data in the real world. The central deadlock is: “If you think a breakthrough in world models will eliminate VLA, I’ll ask one question: Which needs more data, VLA or a world model? Even if that thing really exists, you still have to go through VLA first—and its data will most likely be collected this way too.”
• The host compares AI’s eventual ability to accelerate every part of simulation with coding’s transformation of the software world. 陈源培 partly accepts the analogy but not the conclusion: coding is structured and logical, “so it can be implemented very quickly”; the physical world “can also be accelerated, but it cannot solve the problems of accuracy and certainty.” Cross-modal fields develop incrementally; “it definitely won’t be the case that one day a field is eliminated structurally.” LLMs can replace some language work in finance and healthcare, but “if you actually need to inspect, listen, question and diagnose, they still can’t replace a person.” LLMs in embodied AI are useful: “L is definitely needed. It’s a tool, but it definitely won’t be a paradigm.”

32. Commercialization, Cost Structure and the Definition of a Bubble

• After nearly 2 years of building a company, his central commercialization lesson is trade-off. It “really tests the founder’s ability”: “You must always remember that we are a company building models and AGI, not simply commercializing a single result.” Balancing near-term performance with long-term value has no universal solution; it depends on the situation. Last year’s WRC mahjong demo used reinforcement learning but still relied on simulation. In China, the leverage comes from scenarios and policy; Silicon Valley is more focused on demos and technical exploration.
• On the hand-brain relationship, he chooses “the brain determines the hand.” Hardware will account for a large share of costs in the short term because commercial deployment will initially involve specialized, weakly generalized scenarios. “In the future, the brain will definitely account for more,” depending on its generalization, compatibility and generality. His definition of an embodied-AI bubble is one sentence: “Too many people say something different from what they are actually doing.” A company may claim to use reinforcement learning while actually relying on other methods; because it is difficult for outsiders to tell, the practice persists.

33. A Methodology for a Nonstandard Life: No Regrets and “It Is Absolutely Impossible Not to Be Happy”

• His advice to gaokao students is simple: “You can enter robotics from any major.” He names 3 traits of strong robotics talent: a drive for excellence—“there is no ceiling; if you make something good, there is always something better”; systems thinking—full-stack capability, because “you can’t focus only on the model and ignore everything else; it’s a systems-engineering problem”; and openness without ego—“being able to see other people’s strengths is a good thing.” Diligence does not make the list: “I’m not particularly diligent myself, to be honest. The top students and competition winners at Tsinghua and Peking University work much harder than I do.”
• His attitude toward “geniuses” and role models is equally demystifying. The spinning jenny and the steam engine “were not invented at the biggest companies or in the places with the most money. They were suddenly built by people whose ideas had not yet been validated.” Large companies tend to hire geniuses who have already been validated. He does not want to be a role model: “I can’t exactly encourage everyone to play games every day.” His main advice is to do what you want and not be too afraid. “If you want to do something, it may be possible to do it.”
• He settled on his life principle in sixth grade: “Don’t regret.” Playing cards before the gaokao meant, “This is what I’m choosing now. Even if, when I’m old, I genuinely haven’t read many books because of these choices, that’s fine. I’m happy now.” The other principle is: “It is absolutely impossible not to be happy. If there’s something I don’t like doing, I simply won’t do it. I’ll find every possible way to finish it.” Is entrepreneurship fun? “I’m doing what I like. At the very least, I think I’m luckier than many people—many people are doing things they don’t like.”