Lindon Gao: Finding Clarity in the Fog of Robotics, Disagreements & Choices, and Wang Yangming

Summary

• Dyna founder Lindon Gao calls the company a “Product Research” outfit and describes its approach as “all-out.” He says product loops should steer research: deploying a task is not the same as closing a commercial loop, and “robots needing a babysitter is a very big product problem.” Last year, he decided not to scale a folding business with more than 10 customers and instead set a new target: have a robot produce in 24 hours what one person produces in 8. He says Dyna has already surpassed it.
• Dyna is betting on egocentric, first-person data, and says it has found a “very strong scaling law” in a pretraining dataset at the 1-million-hour scale. The company has shifted from prioritizing high-quality data to seeking broader, more diverse data. Lindon admits that he selectively listened to LLM practitioners’ ablation findings to reinforce his “all in on high-quality data” view; he later concluded that the first step should be finding what can scale from broader data. He argues that betting on teleoperation, UMI and similar formats is, to some extent, betting that robot hands will not work—and that robot hands are an engineering problem, not a research problem.
• His contrarian commercial view is that demand for general-purpose B2B robots will exceed demand for B2C home robots; the home is a difficult, distant product direction, not the first market to tackle. His case: a 4.5% US unemployment rate means businesses employ 95.5% of the US workforce; assuming a household of 3, the US may need only around 100 million home robots, while businesses’ demand for capacity can keep growing. Dyna’s first deployment setting is hotel laundry rooms: Lindon says the US has 100,000 hotels, each needing 2–5 laundry workers, for a TAM of several billion dollars. He calls the laundry room the “smallest unit” on the path to the home.
• His view on “brains vs. bodies”: the brain sets the floor for what a robot can do; the body sets the ceiling for how well it can do it. He believes supply chains may matter more over a 10-year horizon. Dyna has a 40–50-person hardware R&D and supply-chain team in China, and its workflow robot has entered initial production at the hundreds-of-units scale. The previous generation was not mass-produced because, in his view, it simply could not work.
• He estimates Dyna is still 1–2 years from product value that can scale, and expects the industry to converge further over the next 1–2 years. He points to GPT-3, released in API form in June 2020, and ChatGPT, launched as a chat product in November 2022, to show why product form matters. He says the company will show its next step next year, and argues that robots do not need to reach physical AGI first: product definition and how people use them matter just as much.
• Late last year, the team split over whether to prioritize deployment or research, while Lindon was weighed down by the need to be a “CEO who doesn’t make mistakes.” He admits that when he did not know the next strategic step, he followed the market and watched what others were doing, leading to technical choices that did not reflect first principles. The team eventually focused on workflows, aligning research and deployment around the same strategic point. His takeaway: “Making decisions to avoid mistakes is, by definition, not a good decision.”
• A serial founder from the start: at 14, he made tens of thousands of dollars selling in-game items; his previous company, Capa, built cashierless retail carts, struggled to raise after joining YC, and was eventually sold, after which Lindon spent 3 years at Instacart. At the hardest point, he cold-emailed more than 1,000 angel investors and converted just 2. The 4 founders lived together, burned $7,000 a month and lasted 2.5 years. He once told his team, “We’re cockroaches. We don’t die.” The core lesson: products must come back to ROI—“If the ROI doesn’t work, it isn’t a product, and there is no product-market fit.”
• Working on himself is a throughline of the interview: he says he was diagnosed with ankylosing spondylitis at 18 or 19, and later treated the illness as an experiment, looking for patterns by changing his sleep and diet. Reading Meditations, Wang Yangming and others also helped him through the hardest parts of building a company. His motivation has shifted from external validation and proving himself to “living more fully as who I really am”; he calls the failures he defines for himself “failures of self-cultivation.”

Deep dive

1. The US-China divide in robotics: brain-first, body-first and “all-out”

• Lindon starts by cooling expectations about both markets: “Both sides are pretty caught up in the hype. There are releases every day, and every company wants to raise more money.” But he sees a difference in strategic thinking: the US is brain-first, on the view that “the brain is the only problem; the body will sort itself out,” while China is body-first, believing “the supply chain is the only problem; the brain will sort itself out over time.” He adds that China has recently started investing in models, too.
• Asked where Dyna fits, he says: “All-out.” He later sums up the company’s approach as product research: research steered by a product loop.

2. His first big payday at 14: selling in-game items for tens of thousands

• He moved from China to the US at 11 and grew up in what he calls a “very entrepreneurial” environment. In middle school, at 14, he realized “there was money to be made in games” and built a marketplace for game accounts and equipment, making tens of thousands of dollars before selling the site. He did not think of it as entrepreneurship then: “It was just a hustle, a way to make some quick money”—selling games one day, software the next, shoes after that.
• His own explanation of the appeal of making money is worth preserving: “It’s not that I like money itself; I like the reward that comes from getting something right.” Schoolwork? “I wasn’t especially focused on it.”

3. The jewelry business: from hard work selling to consumers to taking a cut in B2B

• In his sophomore year of college, he bought jewelry inventory that someone else was about to melt down and sold it on eBay. Later, tired of packing orders and handling customer service, he switched to B2B: through family and friends, he connected Chinese jewelry suppliers with US designers and took a commission. The business solved a problem for designers: US factories were expensive, and their finishing and mass-production quality were poor.
• It became a self-managing business, run by his family for nearly 10 years. He says he spent 3–4 hours a week on it and made $100,000–$200,000 a year; the transcript does not specify the currency. Eventually he shut it down himself: “I don’t think it’s worth the time anymore.”

4. Two years in medicine, a switch to finance, two years in banking—then straight into a startup

• His educational path was “very free-form.” Medicine was his parents’ idea—the host mentions NYU School of Medicine, and Lindon confirms it. After 2 years, he did not want to continue and switched to finance. He went into investment banking partly by following the crowd: “When I first switched to finance, I didn’t actually know what investment banking was. But I knew everyone wanted to do it, so I wanted to see if I could compete with everyone.” After more than 2 years in banking, he went straight into entrepreneurship, with no transition in between.

5. Capa’s contrarian starting point: 150 grocery-store owners turned down even a free Amazon Go

• Amazon Go set off a wave of cashierless retail in 2017, spawning dozens of companies—more than 50, by Lindon’s count. He says Capa was “one of the few survivors.” The team pitched an Amazon Go-style retrofit to 150 grocery-store owners; not one wanted it, even for free. There was no ROI: spending hundreds of thousands of dollars to save a few cashiers did not pencil out. Cashiers did more than ring up purchases, and renovations and maintenance added costs.
• So they changed the product, putting sensor fusion and computer vision into shopping carts. Cashierless retail has to answer two questions—who pays, and how much? A smart cart could answer both, with much simpler technology: “Who pays? Whoever’s pushing the cart. How much? Whatever you put in it.”

6. Fundraising hell and the cockroach pledge: 1,000 cold emails, 2 angel checks

• Capa got into YC on its first application. Its mentors were Sam Altman and Geoff Ralston, both of whom later became YC presidents and CEOs. Coming out of YC, Capa ranked near the bottom of its batch for fundraising, bringing in only $300,000. The 4 founders and 1–2 early employees moved in together: “I couldn’t pay you a salary, but we could cover rent, gas and food—though you’d have to cook for yourself.” They burned $7,000 a month and lasted 2.5 years.
• With just $10,000–$20,000 left in the bank, he cold-emailed more than 1,000 angel investors. 2 converted, each putting in about $50,000 to keep the company alive. For the seed round, Capa met more than 100 firms; all but First Round Capital passed, and it raised about $2 million. In its 2019 Series A, more than 60 investors rejected it before Lux Capital invested more than $10 million. “Not many founders can get rejected hundreds of times and keep going. I told the team, ‘We’re cockroaches. We don’t die.’”
• Looking back, he says the problem behind more than 100 rejections was not that he had the wrong investor profile in mind, but that he had no profile at all: “At the time, I was focused on talking, not listening.” He was not on the same wavelength as investors and struggled to build rapport.

7. Selling Capa: not for the money, but to do something bigger

• By the Series B, Capa had traction, 2 term sheets and an acquisition offer. After turning the buyer down twice, Lindon decided to sell. The company had lasted about 6 years and was sold in 2021 for 350 million; the transcript does not specify the currency. He then spent 3 years at Instacart.
• His first thought at the time of the sale was “not that I suddenly had money, but that I’d realized we had done something right.” The main reason he sold was: “I thought I could build something bigger, in a bigger industry.” The host notes that AI began to take off in 2022, and Lindon confirms that was when the curve became clear.

8. The core lesson from his early startups: everything comes back to ROI

• His guiding principle is to focus on what matters most. Do not assume “if we hire this person, the problem is solved” or “if we land this customer, we can raise.” In robotics, a company may find something it calls PMF, but customer customization and the required cadence can leave the product short of what robotics needs to deliver. His rule is blunt: “If the ROI doesn’t work—especially in hardware and robotics—I don’t think it’s a product, and there is no product-market fit.”
• Cashierless retail taught him that product lesson. Everyone came in with a hammer, looking for nails—in this case, ways to save on cashiers. But the real value was screen-based interaction: recommend one snack when a shopper picks up another, or offer a coupon for buying 2 more. “The product’s core value is increasing average basket size, not saving on employees.” They also found an advertising opportunity, neither of which was possible in a cashierless store without a screen.
• He admits he was not ready to absorb what Sam Altman was teaching him at the time: “I wasn’t capable enough to get on his wavelength.” Sam might meet them twice a week and repeatedly urge them to do one thing: “Launch, launch, launch, launch and iterate.”

9. The CEO as a mental athlete: how discipline improves decision-making

• He usually gets up at 5 or 6 a.m., exercises from 6 to 7, and goes to bed at 10 or 11. He eats on a regular schedule, rarely eats carbs at night, and has kept up the discipline for about 10 years. He traces the idea to Atomic Habits: good habits stop taking up “mind share.” “It turns into an obsession: if I don’t exercise today, I feel uncomfortable.”
• The connection to decision-making is “100%.” Of a CEO’s 3 jobs—hiring, fundraising and setting direction—setting direction takes the most brainpower, because every decision has to get to the essence of the problem. “If your head is foggy, you eat late-night snacks every day or you’re unhealthy, it’s hard to make good choices.”

10. Three years at Instacart: management lessons for a founder who learned by doing

• One major reason he joined Instacart after the sale was to learn: “There were a lot of executives from Uber, Facebook and Google at Instacart. I was someone who had made it by taking an unconventional path.” The most useful lesson was “how to get resources from a CEO”: explain why your work matters, frame the issue, understand the underlying motivations and align with the CEO.
• He managed up to 300 people directly and indirectly. Over time, he realized “some people can’t be changed”; some cannot take on a larger scope without going beyond their current understanding. Does being a VP fit his personality? “Not at all.” He compares entrepreneurship to an artisan making something: the company is the work, built from technology, culture, product and management.

11. Why robotics: narrowing down the things that could change the world

• After selling Capa, he and his longtime collaborator 杨士源 seriously explored AR for about 6–8 months. They concluded that “some of the underlying technology might not be ready yet.” They also looked at quantum and nuclear; biology did not feel like a fit. Their filter was to take on something difficult that could change the world: “There aren’t many things that can truly change the world—maybe you can count them on two hands.”
• They landed on robotics because the industry felt familiar: “Everyone was looking for nails with a hammer.” Robotics also matched their background across hardware, software and AI. The team lacked someone senior in embodied intelligence, so after talking to around 20 people in Silicon Valley, they found Jason.

12. The taste test for a research co-founder: papers that get simpler over time

• His first criterion for a third co-founder was research taste, which he thinks you can read from someone’s publication history. Jason’s papers in his first and second years of a PhD were “relatively fancy”; later, they became “simpler and cleaner, with the core idea much clearer.” Lindon saw someone continuing to grow—and could see that trajectory in the papers.
• The second criterion was pragmatism: a person who cares about academic research but also wants to ship the work and have a real impact on the product.

13. “Smelling where the money is” and a dream of the stars: Lindon’s self-portrait

• He says his strength is thinking ahead: he can picture the company’s vision 10 years out and the strategic steps needed to get there. His team says he can “smell where the money is”; he sees that as another expression of direction-setting. “Every company eventually has to ship and make money. I can see what the money-making version looks like, and then I know what groundwork to lay.” His weakness is speaking too broadly: he can spend 20 minutes walking through a chain of reasoning, while the other person struggles to follow.
• Asked whether “smelling where the money is” conflicts with idealism, he says: “No. I’m a very idealistic person. I didn’t get into robotics to make money. I think ‘smelling where the money is’ is a way to survive. My dream is the stars; my path has to be grounded.” He calls himself a practical idealist, and praises 杨士源 for his empathy, structured thinking, ability to execute, and strength in both technology and management.

14. The team’s identity crisis late last year: deployment vs. research

• “For a while, we lost our strategic direction, and the biggest disagreement was between one group saying, ‘We should focus on deployment,’ and another saying, ‘We should focus on research.’” He thinks every robotics company has to confront the same tension. He did not firmly belong to either camp, but at the time leaned more toward research: deployment could make money, but without scalable PMF it would not be a durable product form. The deployment camp included some of the engineers.
• The answer was not to choose one side, but to “extend the time horizon.” Whether deployment can scale is often limited by intelligence, so they set a more distant strategic goal that could both improve intelligence and support deployment. That gave the team a point around which to regroup.

15. What “Product Research” means: research with guardrails

• A traditional research team might explore AGI broadly before narrowing down. But in an open-ended process without deployment, “seeing something is not the same as having understood it deeply enough.” Product research adds guardrails: research must contribute to a long-term product, with priorities set rather than exploring everything at once.
• Asked about companies such as Pi and Generalist, he declines to say what direction they are taking, noting that each company understands its own strategy. Dyna, he stresses, has always been research-led, but with product guardrails: “I have a very clear picture of the product we want in the future. I work backward from the research needed to reach it.”

16. Dyna-1: a commercially viable demo that folded for 24 hours at 99% success

• Dyna’s 2025 product-research goal came from the weaknesses of the foundation models available at the time: robots were slow and erratic, and could usually manage one task, but often failed when asked to repeat it or move to the next step. Lindon set the team a target: reach 99% success and 50–60% of human speed to have a shot at commercialization.
• The team saw Dyna-1 in April or May. Lindon calls it a major breakthrough for the team and the industry—and the first commercially viable demo to be sold by a company for the first time. The robot could fold napkins continuously for 24 hours, with a success rate above 99%.

17. Why choose folding against the grain: tolerance for error and technology-market fit

• Investors asked why the company was tackling laundry rather than factories or logistics, and how big the laundry market could be. Lindon’s logic was that a deployment setting needs room for error. Spill a latte and you create work for someone and waste a cup of coffee. “Folding is different. If you fold it wrong, you can fold it again.” Given what they thought the technology could do at the time, folding was a technology-market fit and an isolated task. Conversations with laundromats and hotels showed that both had dedicated folders and attendants.
• He did not start by asking whether the market was big enough. He asked whether they could close the first commercial loop, and whether the problem was hard enough to improve the model. The technical challenge was deformable objects: cups have a relatively limited set of states, while every folding step depends on the previous one working and tests generalization across clothing. Their tests suggested that more than 10 environments might be enough for zero-shot generalization on T-shirt folding; they also needed coverage across garment sizes, styles and fabrics.

18. The brutal details of deployment: 1,300 napkins, double-grabs and how to stack the bin

• Deployments from June to September exposed the gap between a demo and a product. A customer needed 1,300 napkins a day; Dyna-1 could fold just 850 in 24 hours. Once they improved speed, they found a fatal corner case: sometimes the robot grabbed 2 napkins. Folding 1 normally took 20 seconds; handling 2 could take 2 minutes. “If 10% of the napkins were double folds back then, there’s no way we’d meet the standard today.”
• Folded napkins also landed in the bin randomly and unevenly, making them look unfinished. The team had to condition the model to place them neatly in specific positions. When corner cases came up, they insisted on solving them with the model rather than engineering around them, because they believed scalable PMF required the model to handle them.

19. Robustness comes from pretraining: shifting the resource mix

• The key lesson from deployment, Lindon says, was that “most of a model’s robustness comes from pretraining, not post-training.” That is why Dyna started pretraining its own models. Initially, it split resources evenly between pretraining and post-training; later, it increased pretraining investment substantially. Post-training needed less and less data, which did not surprise him.
• He thinks the problem of grabbing 2 napkins and reducing throughput would be hard to solve through pretraining, so at the time the team considered reinforcement learning and other methods. Getting a napkin into a specified part of the bin and stacking it neatly, by contrast, could be addressed through language and pretraining.

20. The decision not to scale: “Robots need a babysitter”

• The folding business already had more than 10 customers, but Lindon chose not to keep scaling it: “The problem we solved wasn’t fundamental enough. I only completed the task; I didn’t complete the whole workflow.” A typical issue was that after folding a stack of 100 napkins every half hour to hour, the robot would stop, and an employee had to set an alarm to refill it. Customers complained that “the robot was too slow,” but the data showed it had been idle for half an hour. “It wasn’t that the robot was too slow. The robot itself needed a babysitter.”
• The product principle: “If you scale an incomplete product, it will become a burden later.” The product loop in this setting was not complete enough, so he decided not to expand it further.

21. A task is not a commercial loop: responding to “solve the brain first”

• The host poses the research-lab view: shouldn’t the industry solve the brain before defining a use case too early? Lindon brings it back to the core problem. In napkin folding, “the agent itself was no longer the problem”; the workflow still did not form a complete loop. Any task can be broken into a sequence of motions. Generalizing well enough to perform a task in a new setting is real progress in the agent, but it is not the same as closing a commercial loop.
• That led to Dyna’s next step: “move from tasks to workflows.” For napkin folding, the robot has to handle the full process: know where the inventory is, retrieve it, return to the table, fold, place the finished napkins on the shelf, fetch a new bin and repeat. He says the system can now complete the full workflow. It has commercial customers and is starting to work across different settings, but is not yet fully deployed.

22. Dyna-2’s stack: System Two, One and Zero—and signs of generalization

• Dyna-2 differs from Dyna-1 in its model architecture, pretraining dataset, reasoning and even hardware. Lindon describes the system as 3 layers. System Two is an embodied reasoning layer that understands physical properties, current state and semantics, then gives language instructions to System One. System One is a world-action model that turns language into actions. System Zero handles control, using methods such as reinforcement learning to make trajectories faster, smoother and more accurate.
• He measures generalization against 2 ways of teaching a person: demonstration and language. On the demonstration side, some entirely new tasks can now be learned after just a few demonstrations. On the language side, language sensitivity has improved significantly; he says he will share the details at launch. Completing a workflow also requires every step to be reliable: multiply the success rates of multiple tasks and the overall rate falls quickly, so the team needs to understand each task’s corner cases and connect the tasks together.

23. Redefining the commercial metric: not faster than a person, but working longer

• Asked whether robots can undercut human labor, Lindon says that under current foundation-model constraints, it is hard for robots to match one person’s daily output. Hardware, control and data all limit speed. So he changed the metric: “The most important measure of value isn’t doing the work faster than a person, but doing it for longer than a person.”
• His target for the team was for a robot to work 24 hours and produce what one person does in 8 hours. He says it has already exceeded that target.

24. Why not the home: a subjective market and the B2B>B2C population math

• The home may be the end state for robotics, but it comes with time and product challenges. A robot’s ability to generalize across households does not mean it can perform every industrial task; nor does being able to do things at home mean the product works. Expectations in the home are highly subjective. If someone has a bad relationship with a housekeeper, that does not mean the housekeeper is not general-purpose. Speed, temperament, where things are put, safety and corner cases are all product issues. Lindon thinks robots might be able to generalize into every home in 30–50 years, but building a product does not guarantee there is a market.
• He believes demand for general-purpose B2B robots will exceed B2C: “People’s wants are infinite.” If labor costs can be ignored, consumer demand will grow, and businesses will also need more robots to make more products. He cites a 4.5% US unemployment rate and says businesses employ 95.5% of the US workforce. Assuming 1 robot serves a household of 3, the US might need around 100 million home robots, but B2B demand would be larger. B2B is also easier to standardize, and safety can be managed to some extent.
• His view on the path is straightforward: “If you can’t make B2B work, how are you going to make B2C work?” If a robot cannot handle the work of a laundry employee, it will have an even harder time dealing with every household’s different clothes and needs.

25. The hotel-laundry economics: 100,000 hotels, several billion dollars and the “smallest unit” on the way home

• The market looks small until you break it down. Lindon says each hotel laundry room needs 2–5 people, and laundry directly affects whether rooms can be turned over. He has heard of a hotel owner who went to work in the laundry room himself during the pandemic. Most laundry rooms are much the same; solve one workflow and it can generalize across 100,000 US hotels, a TAM of several billion dollars.
• The strategic value is bigger than the revenue: “The laundry room is the smallest unit on the path to the home. If you can’t make the laundry room work, the home definitely won’t work.” He also sees the setting as forgiving, with fewer demands on timing and a strong willingness to pay. Hotels could eventually expand into cleaning and cooking. Room cleaning could work once the model is intelligent enough, but it has more corner cases and is a harder stage. You cannot skip the smallest loop and jump straight to the hardest thing.

26. Classify by workflow difficulty, not setting: what “laying eggs along the way” is really for

• He rejects the industry’s usual setting-based categories: “People often ask whether we’re a commercial robotics company. I don’t think that’s the right question. I classify things by workflow complexity.” Loading and unloading a food-ordering kiosk is a simple, short workflow. He says laundry may be level 2 or 3, while hotel cleaning may be level 5, though he has not defined the levels in detail.
• His rule for deployment: “The setting I deploy into has to be within the agent’s intelligence range. The strategic direction I want to pursue next is outside that range.” Laying eggs along the way is not about cash flow; it is a way to test what level the agent has reached, what corner cases it has, and whether those cases can steer pretraining.

27. A pretraining bet: 1 million hours of egocentric data and a scaling law

• Lindon says Dyna may have released the largest dataset in the industry and found a “very strong scaling law” in a 1-million-hour egocentric human pretraining dataset. He calls it an important moment for the industry: over the previous 1–2 years, it had been hard to break through the intelligence ceiling, and everyone said data was the biggest bottleneck. The team’s first-principles question was how to reach a GPT-3- or GPT-4-like moment for a robotics foundation model. His answer: extract what can scale from low-quality data instead of scaling only high-quality data.
• This was a public change of mind. Dyna-1’s best-performing model was trained on high-quality data, and at the time he told the team, “Need to go all in on high-quality data.” He admits he selectively listened to LLM practitioners whose views reinforced his own. After moving to workflows, the team realized what mattered was giving the model the most diverse and broadest possible set of human experiences. Only after reaching internet-scale data does it make sense to run ablations and distill.
• The bet on egocentric data is a 5–10-year call. In his view, betting on teleoperation, UMI or other non-egocentric data is, to some extent, betting that robot hands will not work. He believes robot hands are an engineering problem, not a research problem.

28. Real-world vs. simulation is a false debate: data mixture depends on the training objective

• On the factional debate in China between real-world and simulation-first teams, he says: “I don’t know why this debate exists.” The technical route is a means, not the point. Dyna uses both real-world and simulation data, and “both work well.” The bet should be on the data and model capabilities you ultimately want, not one particular method. The conversation keeps moving: early last year people talked about real robots and teleoperation, by year-end interest had shifted to UMI, and this year it may shift to egocentric data and simulation.
• He compares it to foundation-model companies taking different paths based on their goals: OpenAI leans toward consumer applications, Anthropic bet early on coding agents, and Google emphasizes multimodality. Robotics teams will likewise have different views on pretraining objectives; there is no single right answer. Dyna has 2: reduce the number of demonstrations needed to speed up deployment, and improve language-following as the basic link for connecting workflows.

29. Solving one setting means solving them all? Can larger models swallow setting-specific optimization?

• His leverage argument: “The difference between solving one workflow and solving 100 workflows is really just the data for those 100 workflows.” The product’s core is a system: data goes in, and the system produces a workflow. Once the first workflow is fully understood, many workflows are not that different, though their SOPs and corner cases still vary. The underlying framework for solving them can be shared.
• Asked whether, as with agent applications overtaken by leaps in foundation-model capabilities, setting-specific robotics applications could get swallowed up, he says it could happen—but that is not a contradiction. Anthropic building a coding agent does not mean it can only do coding; the same pretraining and post-training frameworks could apply to Excel and other settings. Dyna keeps pretraining broad and general-purpose. Post-training is where it “lays eggs”: validating whether a workflow works and identifying bottlenecks that should inform pretraining.

30. Robotics’ ChatGPT moment is a product-definition problem; Dyna’s strategy is still 1–2 years away

• Asked whether the industry is nearing convergence, he says he does not want to get bogged down in that debate: “What matters is when robots can truly bring value to people.” For Dyna’s own product and strategic direction, he estimates that is still 1–2 years away. He thinks workflows can develop to the point where a robot handles 10 or 20 at once.
• The big aha of ChatGPT, he says, was that “anyone could see it was useful.” Robots do not necessarily have to reach general intelligence first for everyone to be able to use one. In his view, that is a question of product definition.
• He compares GPT-3 with ChatGPT: GPT-3 arrived in June 2020 as an API, with many developers and companies such as Jasper building products on top of it. ChatGPT arrived in November 2022 as a browser-based chat product. The change in product form cannot be overlooked.

31. The next reveal is not the robot’s shape, but “a product for how to use robots”

• He will only go so far in describing the end-state product: “You might think the final form is a robot. I don’t think it’s a robot form; it’s a product for how to use a robot.” He says it is B2B, but the product experience should feel useful to everyone. It will not launch this year, but the next release is worth watching. The strategic next step discussed in the interview will be visible next year.
• Whether the robot is humanoid comes down to ROI: “Will adding these 2 legs create more value? Will that value be far greater than the cost of building them?” He calls the next stage a larger general-purpose market, not “a bigger story.”

32. The biggest misconception: Dyna is not a deployment company

• “When people in the industry hear Dyna, they instinctively think we’re more about the body and the setting. That’s a big misconception.” Early PR focused on Dyna-1 deployment, and the company did not spend much time explaining its research. In reality, it is heavily research-focused. He sees no fundamental difference in research objectives between Dyna, Physical Intelligence, Generalist and Google DeepMind: all are trying to make models generalize, improve language steerability and strengthen few-shot demonstration.
• He says 90% of the company’s capacity goes to the agent. Deployment engineering infrastructure has been kept simple. A dedicated creative team produces demo and demonstration videos; they are not made by the engineering or research teams.
• One reason for staying quiet is a high research bar: the team had language steering working by August or September last year, but did not think it was at 99% and did not want to release it. Another is that when the strategy was unclear, they did not want to go out and tell a partial, stage-by-stage story for PR.

33. Why pursue a commercial loop so early: the deployment flywheel may take 2–3 years

• The point is not revenue, but feedback: “A loop gives you feedback, and feedback tells you what to do next.” He compares it to self-driving: unless you drive every street in San Francisco, you do not know where the capability limits are.
• Deployment itself is hard. Hardware, software, models and controls can all break; sometimes a robot that folded 1,300 napkins one day folds only 800 the next, and no one knows why. That is why teams need an observability stack and support and maintenance stack. Lindon thinks it could take 2–3 years to get the entire deployment flywheel working. “If you finish building the agent first and only then start closing the deployment loop, you’ll find you’re 2–3 years behind.” That is a lesson he took from deploying Capa’s products.

34. Supply chain, mass production and a 10-year bet: the battle over cost and quality

• Dyna built its China team from day one. Lindon’s view: “Without a supply chain, an agent can’t produce a world-class robot.” The company has 40–50 people in China focused on hardware R&D and supply chain, and says it has worked through the necessary groundwork. Its workflow robot has entered initial production at the hundreds-of-units scale. The previous generation was not mass-produced because, in his view, it simply could not work.
• If he had to pick one factor for the long run, he thinks the supply chain may matter more 10 years from now. Customers ultimately receive hardware, and over time competition may come down to cost and quality. The case for building the full stack early is that hardware and brain need to be developed together: a robot body needs at least 3 iterations before it can enter mass production. The supply chain is not yet settled; starting 3 years late means losing 3 years of iteration. Dexterous hands may change, but he does not think they are ready for deployment today.

35. The brain sets the floor, the body the ceiling—a view opposite to the consensus

• The host notes that this runs against the mainstream view. Lindon explains: “The brain determines whether something can be done at all; the hardware determines how well it can be done.” Folding 1,000 napkins in a day, for example, can depend on data, controls and hardware, all of which affect speed and accuracy. Different control algorithms can also produce very different results from the same teleoperation data: one model may move smoothly, another may perform terribly. For B2B deployment, he says, it is hard to hit the required quality and speed without owning the body.
• He sees the brain as driving the body. The research team sets hardware specifications such as camera FPS, global or rolling shutter, angle and field of view. As the brain improves, the hardware may become the limit. His analogy: “It’s like being able to play against a world-class table tennis player, but your body can’t keep up.”
• His view of peers is specific but measured: Unitree has strengths in hardware R&D and closing the system-identification and sim-to-real gaps; AgiBot has ample funding and more room for error, allowing it to cover a wider product line and more agent directions. The industry is now more about mutual respect and learning. He thinks the next 1–2 years could bring further convergence in model architectures, data and capabilities.

36. Organizing in the AI era: flatter teams, about 200 people in the US and the best people on the most important work

• Lindon thinks organizational structure has changed fundamentally. Before coding agents, big companies had many layers; now teams are much flatter, CEOs and senior leaders can influence execution more directly, and work moves faster. Dyna has more of a lab culture: its US team focuses on research, while its China team is mainly hardware. The US team has about 200 people, the host notes, and says it is a little larger than that.
• His management principle is “hire the best people and have them do the most important work.” Most individual contributors have managed people before and chose to return to IC roles because they prefer the work. They can break down their own scope and execute. Smaller teams reduce management overhead and make it easier to pull together around one goal. He sees the smaller-team model as specific to this stage, but thinks future companies may not need to grow as large as before: work that once took 1,000 people may now take 100–200.
• The rise of coding agents has lowered engineering costs and sped up iteration. “Something that used to take a month now takes a week, so you can explore 4 directions at once.” He says a team’s iteration speed also affects how people think about the timeline to AGI.

37. Where self-cultivation began: ankylosing spondylitis and a change in perspective

• He began working on himself while building Capa, more than 10 years ago. Lindon says he developed ankylosing spondylitis at 18 or 19. His immune system attacked his spine, and even sleeping or sneezing hurt. He resented the world and thought he was unlucky. The biggest breakthrough was not that the illness improved, but that “I changed how I looked at it.” He treated the pain as an experiment, spent more than 10 years adjusting his sleep, diet and lifestyle to look for patterns, and says he eventually got the illness under control.
• He applied the same shift in perspective to investors. He used to think investors he could not connect with were ignorant or stupid. Later, he realized their way of thinking had something to teach him, and views he could not accept before became useful prompts. He reads Meditations and Wang Yangming, and has also looked at Buddhism and Christianity. He thinks they all touch on how to live with the world, how to see it and how to change one’s perspective.

38. Overthinking, people-pleasing and a new source of motivation

• He admits that he cares about how others see him and may have a people-pleasing streak. After a talk, he watches the video back and fixates on how he spoke and where he paused. Reading “The person in the world who cares most about what you look like is you” helped him get past some of that.
• The drive behind his fourth startup has changed. He used to judge himself by fundraising, achievements, valuation, customers and revenue; now he wants to “live more fully as who I really am.” He draws on the idea that “a mountain is a mountain,” and the story of Adam and Eve eating the apple of knowledge: as people grow up, rules and other people’s views constrain them. He wants to stop caring about others’ opinions and be free to choose what he wants to do. “Proving I’m not a flash in the pan” used to weigh on him; now, external judgment does not matter.

39. Last year’s mistake: the burden of being a mistake-proof CEO and getting pulled by the market

• During the period when the strategy was lost, “I just couldn’t figure it out.” One reason was the burden of trying to be “a CEO who doesn’t make mistakes”: as a founder who had already succeeded once, he thought people had come to him for his experience, so he could not get things wrong. “Making decisions to avoid mistakes is, by definition, not a good decision.”
• He admits that, without a clear next strategic step, he watched what others were doing and thought Dyna had to catch up. He lacked his own market judgment and first-principles thinking. The team got pushed around by the market and made technical choices that did not reflect first principles, wasting time. Only after finding a direction that made sense to them did the team become stable and cohesive.
• After getting through the difficult period, his first feeling was gratitude—to his teachers, team, wife and children for their support. Some challenges cannot be solved by changing other people or the world; you have to get past them through self-cultivation. Many problems cannot be solved just by working longer; they require a change in mindset.

40. The World Cup, a marathon with no visible finish line, and rapid-fire questions

• He calls robotics the “World Cup”: the strongest players are all in the arena, and he approaches it with respect. The industry is still shrouded in fog; finding your own path through it is rewarding. What if this attempt does not go as well as the last one? “At my core, I’m here to win.” But if he can play with style and enjoy the game, he is happy; each milestone is a checkpoint, not the goal. Founders keep falling into holes and climbing out again, with no idea where the marathon ends.
• Rapid-fire answers: hometown foods are Fuzhou mixed noodles and bianrou; he has 3 daughters, aged 1, 2 and 3; the book everyone should read is The One Thing in Life; the paper that could shape robotics, he says, is one of Jason’s, which redefined what future robot self-learning could look like and was one reason he got to know Jason. The important bet right now is infrastructure, because model iteration speed and the scope for scaling up matter. His off-the-cuff take on “language is the world”: the world is vision, physical properties and all kinds of sensory inputs; language helps robots understand the world and turn language into action.