Lindon Gao on Building Dyna Robotics and Finding a Path Through the Fog

Host Xiaojun introduces Lindon Gao, founder and CEO of DYNA Robotics, as a serial entrepreneur taking a product-research approach to robotics. The field has not converged, and founders must choose a path without knowing where it leads. The interview will explore those choices and how to regain clarity when lost. A teaser notes that Gao has faced hundreds of rejections without giving up.

Gao recalls telling his team, “We’re cockroaches. We can’t be killed.” During a period of strategic confusion, the team’s identity came under strain as some pushed for commercialization and others for research. He says idealism and “smelling the money” do not conflict: he entered robotics for the dream, while commercial judgment is a way to survive. His dream is the stars and the sea, but the path must stay grounded. He visits China once or twice a year and has a team in Shanghai.

Gao says Shanghai is developing quickly and robotics is intensely competitive on both sides of the Pacific, with companies constantly announcing products and raising money. The deeper difference is strategic: Silicon Valley tends to believe the brain is the main problem and the body will follow; China has often bet that supply chains matter most and intelligence will catch up. He sees China beginning to invest more in models, too.

Asked whether DYNA belongs to the brain or body camp, Gao says it is full-stack. He was born in China, moved to the US at eleven, and grew up around people who ran small businesses. At fourteen, while in middle school, he began selling video-game accounts and in-game items, then built a small trading platform. At the time, he saw it as a hustle rather than formal entrepreneurship.

Gao’s first platform let users buy and sell game items, sometimes through other platforms or forums. He made more than $10,000, then sold the website as the business faded. His early ventures were scattered: one day he sold game items, another day software or shoes. That kind of side business was common in the entrepreneurial environment around him.

Gao says his school and peer environment encouraged making money, and he learned by following trends and trying things. He does not say making money is effortless; what he enjoys is the reward of getting something right, not money for its own sake. He was not especially focused on school, and says he was not a typical entrepreneur. His second venture began with jewelry trading in college.

In his sophomore year, Gao bought jewelry inventory that others planned to melt down for scrap and resold it on eBay. The business became largely self-managing, so he handed it to a family member and it ran for more than ten years before closing. He says he eventually decided that even a few hours a week was not worth the time, despite earning roughly $150,000–$200,000 a year.

Gao entered the jewelry business through family, friends, and connections to China’s supply chain, while friends in the US helped him understand demand. He shifted from selling directly to consumers—which required daily packing and customer service while he was in college and interning—to a business-to-business model. He began supplying jewelry designers instead.

US jewelry designers often had to find their own suppliers or factories, and production runs could be expensive, poorly polished, and inconsistent. Gao started an import-export business that connected Chinese jewelry suppliers with US designers and earned a commission. His education was similarly self-directed: his parents were hands-off, and he first studied medicine at NYU before switching to finance after two years.

Gao says medicine was largely his parents’ wish, while investment banking was a popular path among NYU business students. When he switched to finance, he barely knew what investment banking involved, but wanted to compete for the same opportunity as his peers. He worked in banking for a little over two years, then went straight into founding a company and later robotics.

Before DYNA, Gao co-founded Caper, a cashierless-retail startup launched amid the 2017 excitement around Amazon Go. The concept let shoppers take items from a store and leave while the system checked them out automatically, prompting companies to rethink retail. Gao recalls more than fifty companies entering the space, which he compares to today’s crowded robotics field.

Caper was one of the few companies to survive. While competitors proposed renovating entire stores with shelf sensors and ceiling cameras, Gao took the idea to supermarket owners. He spoke with about 150 and asked whether they wanted Amazon Go-style technology; none did, even when he offered it free. They said the economics did not work.

Supermarkets did not believe savings from a few cashiers justified hundreds of thousands of dollars in installation, especially because cashiers had other duties and the systems needed renovation and ongoing maintenance. Gao concluded the original product logic was unsound. Caper shifted to putting sensor fusion and computer vision in a shopping cart, which could identify what a customer selected with simpler technology and clarify both who pays and how much.

Caper got into Y Combinator on its first application and had prominent mentors, including Sam Altman and Geoff Ralston. Yet after YC it ranked near the bottom of the batch in fundraising and raised only $300,000. US investors were wary of hardware’s cost and distribution challenges, so Gao and his cofounders faced repeated questions about their background and why the idea would work.

With only $300,000 raised, Gao offered his three cofounders a deal: they could live together, with rent, gas, and food covered, but no salaries and they would cook for themselves. Four founders and a few early employees shared a house, spending about $7,000 a month for two and a half years before raising a roughly $2 million seed round.

The seed round did not end the struggle. Caper’s shopping-cart approach was highly unconventional, and investors kept asking how it could compete with Amazon-style systems that had raised far more. Some YC companies had already raised over $10 million, while Caper had little money and few people. Investors also warned that the cart would eventually become obsolete.

To reach the seed round, Gao cold-emailed more than a thousand angel investors and converted two, each of whom invested about $50,000 when Caper had only $10,000–$20,000 left. More than a hundred institutional investors passed before First Round Capital invested. The Series A was also difficult: hardware development and supermarket sales both took a long time.

Caper needed more than eighteen months to deploy and scale. In 2019, it sought a roughly $10 million Series A, but few funds could write a check that large; over sixty investors passed before Lux Capital invested a little over $10 million. By the Series B, Caper had traction, received two term sheets, and got an acquisition offer, which the team initially turned down twice before deciding to sell.

Gao describes the lowest moments of a startup as a test of one’s inner life. After ten rejections, it is easy to try again; after a hundred, a founder starts questioning their own ability and the project. He says each failure became a round of “reinforcement learning”: look at it from another angle and improve the next attempt. The experience changed how he understood himself and the world.

Gao says the difficult fundraising period forced him to grow, though it was also bitter. He did not yet know how to read investors; he focused on talking rather than listening, so he failed to find common ground. He recognized this only in hindsight. Caper lasted about six years.

When Caper sold for about $350 million in 2021, Gao’s first reaction was that the team had made the right decision, not that he was suddenly rich. He believed he could build something larger in a bigger industry and wanted to apply what he had learned over six years. The sale was driven mainly by that opportunity, not by the money.

Gao says AI’s rise became unmistakable in 2022, though the shift was less obvious in 2021. His central lesson from several ventures is to work on the most essential problem rather than assume one hire or one customer will solve everything. In product development, the key question is whether the product creates real value.

Caper learned that labor savings were not the most valuable part of cashierless retail. Interactive screens could save customers time and recommend complementary products or coupons, but the larger opportunity was increasing average basket size and generating advertising value. A shopping cart could provide those interactions in a way an unattended store could not.

Gao says acquiring one more customer is meaningless if the product cannot deliver return on investment. A customer’s willingness to pay alone does not establish product-market fit: customization demands and required cycle times may make the product uneconomic. In hardware and robotics, he judges product-market fit by whether the customer’s ROI holds up.

Gao describes Sam Altman as highly capable and says his main lesson from their YC meetings was to launch, iterate, and keep improving. They met perhaps twice a week, but Gao felt he was not yet able to engage with Altman at the same level. The conversation then turns to Gao’s discipline and early-morning routine.

Gao says the habits described in Atomic Habits can compound over time. He treats entrepreneurship like mental athletics: exercise, sleep, and a healthy diet help him focus, and taking care of his body is part of building a company. He generally wakes around five or six and has maintained an early routine since starting his career.

Gao aims to wake at five, exercise around six or seven, and go to bed around ten or eleven. He says discipline removes the need to keep deciding whether to work out; skipping a workout can make the whole day feel wrong. He eats few carbohydrates at night, usually has them at lunch, and keeps breakfast nutritionally balanced.

Gao credits his disciplined father as an example and says his health changed through consistent habits. He believes discipline improves a CEO’s decisions. Finding people, money, and direction are the core responsibilities, but direction depends on making sound, well-reasoned decisions, which are harder when one’s body and mind are foggy.

Gao says poor health and late-night eating make clear judgment difficult. He sold Caper at thirty and felt life was just beginning. He then spent three years at Instacart, partly to learn from leaders who had come from Uber, Facebook, and Google. His first startups had been built through a scrappy, unconventional approach, and he wanted to learn how larger organizations operated.

Gao says a foggy mind and an unhealthy body make wise decisions harder. He was thirty when he sold Caper and felt life had just begun. He then spent three years at Instacart to learn from experienced executives from Uber, Facebook, and Google, after building his first three ventures through an unconventional, scrappy approach.

Gao says he used to decide what a business and product should look like, then ask engineers to build it, leaving him naïve about fundraising. At Instacart, he wanted to learn management and product skills inside a larger organization. He recalls telling his team they were “cockroaches” who could not be killed, reflecting his persistence through repeated setbacks.

Gao says lessons often become part of his instincts rather than something he can easily summarize. At Instacart, he learned how to secure resources from a CEO: explain why the work matters, understand the underlying motivation, and align with it. The process reminded him of fundraising, and frequent interaction with the CEO helped him develop that skill.

Gao says his early management was messy: he did not always define a role clearly before hiring and sometimes chose people based on their apparent potential, then tried to correct them when things went wrong. He learned that some people cannot take on broader responsibility without changing their understanding. He eventually managed several hundred people, directly and indirectly.

Gao says being a VP did not suit him because he deeply values entrepreneurship. He compares founding a company to craftsmanship: a founder spends years refining a work they can be proud of. That work includes the company’s technology, culture, product, and management, not just its business results.

Entrepreneurship suits Gao because it is multidimensional and offers constant learning. After selling Caper, he and cofounder York Yang explored augmented reality for six to eight months, along with quantum computing and nuclear energy. They concluded that some underlying AR technology was not ready and other fields were not a strong fit.

Gao and Yang wanted to build something difficult that could change the world, a goal they felt applied to only a handful of fields. They considered AR, energy, quantum computing, biology, and AI before settling on robotics. Robotics fit their prior experience and seemed like cashierless retail: companies had exciting technology but were still looking for the right problem to solve.

The founders saw that robotics could combine hardware, infrastructure software, and AI, but they needed a cofounder with deep embodied-intelligence experience, so they recruited Jason Ma. After speaking with more than twenty people, they knew what they wanted in a third cofounder. Gao and Yang had worked together for over a decade and understood how their strengths complemented each other.

Gao says his main strength is seeing the company’s direction far ahead and mapping its steps and risks over a ten-year horizon. His team jokes that he can “smell money,” meaning he can identify business opportunities. Every company eventually has to make money, so he works backward from the eventual business model.

Gao says he often explains a conclusion only after walking people through a long chain of reasoning, which can take twenty minutes and lose his listeners. He is still learning to make his thinking clearer. He sees no contradiction between being idealistic and “smelling the money”: he builds robots for a larger dream, while commercial judgment keeps the company alive.

Gao calls himself a realistic idealist: ultimately, he is guided by an ideal. His cofounder York is also a realistic idealist, with strong empathy, structured thinking, execution, technical ability, and management skills. Gao sees it as rare to find a technical genius who is also strong at managing people.

Gao and Yang met through former cofounder Yilin Huang at their previous company. To find their research cofounder, they sought strong research taste—visible in a person’s papers—as well as a desire to make a real impact on a product. Jason Ma impressed them because his papers became simpler and clearer as he progressed through his PhD.

Gao saw Ma’s increasingly clear core ideas as evidence of good research taste. He also wanted a cofounder who cared about getting research into a product. A CEO focused on commercialization and a research cofounder focused only on the “brain” can clash; Gao says DYNA experienced this tension during a period of strategic confusion.

During that period, the team’s identity came under strain: some wanted to focus on deployment, while others wanted to do research. Gao says he did not belong to either faction; he was focused on building toward the eventual product. He did not want to commercialize something that could not become a lasting product with scalable product-market fit.

Gao leaned more toward research, while some engineers favored deployment. The team found common ground by extending its time horizon. Robotics and agent companies face the same tension: aiming high can make near-term deployment difficult, while focusing only on deployment can limit ambition. Deployment’s ability to scale is often capped by intelligence.

Gao’s solution was to pursue a longer-term strategy that could improve intelligence while still delivering something useful. Setting that goal brought the team together. He says every robotics company revisits the tension between research and commercialization, and DYNA had recognized it from the start, even before it understood what deployment required or where the intelligence boundary lay.

DYNA describes itself as a product-research team, unlike a traditional lab that explores many possibilities in pursuit of AGI and converges later. Gao says such exploration can remain shallow if the research never reaches a real product. DYNA’s research must have guardrails tied to long-term product value.

Gao says DYNA prioritizes research in stages because the team’s time and energy are limited. He does not define companies such as Physical Intelligence or Generalist from the outside; each has its own direction. DYNA does not deliberately differentiate itself from research-led companies.

Gao says DYNA’s research direction has product guardrails: he works backward from a clear picture of the product to determine what research is needed. The company began in late 2024, when the team was glad just to assemble its hardware. The cofounders and initial team were in place, and they had started speaking with customers.

In 2025, the foundation model was still too slow and unreliable to support deployment: robots could usually complete one action but struggled to repeat it or continue to the next step. DYNA set a target of 99% success and 50%–60% of human speed before commercialization could work. The research led to DYNA-1, which Gao calls a major breakthrough and the first commercially viable demo in the industry.

DYNA-1 folded napkins continuously for 24 hours with over 99% success. At CoRL, many people gathered around the robot, which first folded paper towels and later clothes. Gao says investors asked why the company chose laundry rather than larger markets such as factories or logistics.

Gao chose folding despite investors asking why DYNA was not tackling larger markets such as factories or logistics. He was closely involved in technical decisions and saw folding as a forgiving test: a spilled latte creates waste and extra work, but a badly folded napkin can simply be folded again. That made it suitable given the model’s technical limits.

Folding offered a useful fit between technology and market: one person could perform it continuously, and hotels and laundromats had a real need for the work. Gao was not initially trying to prove the total market was enormous; he wanted a first commercial feedback loop and a hard problem that could still drive fundamental model improvements. He then explains why folding is technically difficult.

Clothing is a soft-body problem, unlike a rigid object such as a cup, which has relatively few states. Each fold depends on the previous one succeeding, and the robot must generalize across garment sizes, styles, fabrics, and environments. Folding let DYNA test quality, generalization, and speed together and develop recipes across varied clothing and settings.

Gao says post-training ultimately depends on the recipe. At CoRL, DYNA tested how many environments were needed for a model to fold a T-shirt zero-shot in a new setting; a little over ten environments proved sufficient, less than expected. Different sizes, styles, fabrics, and materials also generalized to some degree, helping the team map the post-training ceiling.

The team had found a direction where robots could do the most useful work. As technology improved, the amount of post-training data needed kept shrinking, which Gao says was expected after they reached the post-training ceiling. Pretraining then became the priority. DYNA later spent several months, roughly June through September, testing deployment with customers.

Deployment proved difficult. DYNA-1 completed only four of the five folds needed for a napkin; staff had to do the final fold, which was hard because the napkin was thick and needed flattening. Clients accepted this in the demo but were unhappy in daily use because it wasted time. Folding napkins alone also failed to meet the restaurant client’s production needs.

The restaurant needed 1,300 napkins a day, but DYNA-1 could fold only 850 in 24 hours. The team improved speed, then found that corner cases still took too long to resolve. A common failure was accidentally picking up two napkins; even people do it, but resolving it took about two minutes for the robot.

A napkin normally took twenty seconds to fold, so double-picking even 10% of the time could make the robot miss its daily target. DYNA tried reinforcement learning so the model could handle the issue itself, since pretraining alone could not solve this operational failure. The team also had to teach the robot to place finished napkins neatly in a bin.

The bin-placement problem required another difficult round of model conditioning. DYNA did its own pretraining and initially split resources roughly evenly between pretraining and post-training. Gao explains that the team was testing how pretraining changes affected post-training and asks which technical challenges remained in deployment.

DYNA now invests much more in pretraining. Experiments showed that pretraining changes strongly affect post-training and that most of a model’s robustness comes from pretraining. The hardest deployment problems were small corner cases, not dexterity itself. The team debated engineering workarounds but remained committed to solving them with one model.

DYNA had more than ten clients and had begun scaling, but Gao chose not to expand because deployment had not met his bar for scalable product-market fit. The company had completed one task, not the full job. In a restaurant, the robot would fold a stack of one hundred napkins and then stop, so staff had to check and reload it every thirty to sixty minutes.

Customers complained that the robot was slow, but the data showed long idle periods when no one replenished the napkins. Gao realized the robot needed a babysitter. Scaling an incomplete product would make it a burden. He contrasts this with a research lab that might focus on the brain first, since the best deployment scenario could change as technology improves.

Zhang asks whether napkin folding is too small a scenario and whether the brain still lacks data. Gao says the robot could already complete the task; the business problem was the lack of a closed loop around the whole job. The agent was not the constraint within one task, but could become one when the full workflow had to run.

Gao says a task is only a sequence of motions. Better generalization lets a robot repeat those motions in new settings, but that alone does not create business value. Even zero-shot performance in a new environment is a technical leap, not necessarily a commercial loop. DYNA concluded it needed to move beyond the isolated task.

Gao says there is no single fundamental task for a robot. A task is a sequence of motions, and stronger generalization may let it perform that sequence in new settings, but that does not mean the task has formed a commercial loop. The full job and its surrounding workflow still matter.

Gao says even a task that works zero-shot in a new setting is not a business loop. DYNA asked why folding on a tabletop did not provide enough scalable value and found the robot needed a babysitter. To remove that dependency, the company had to complete the whole workflow, including loading and unloading.

The full napkin workflow means finding inventory, retrieving it, returning to the table, loading the robot, folding, moving the finished napkins to a shelf, and bringing over the next bin. Once DYNA framed the problem this way, its next step became clear. Zhang asks whether the robot can now move freely through a laundry room and decide what to do.

Gao says completing a workflow is fundamentally a pretraining problem because every step has its own success and failure rate. Chaining steps compounds errors, so the model must handle every task and its corner cases reliably enough to run continuously. Early on, DYNA used language to direct where folded napkins should go in a bin.

The VLA model’s language-following ability was too weak for that, so the team traced the problem back to pretraining. Asked whether DYNA would build a humanoid, Gao says it has never committed either way. The company later adjusted its pretraining approach and uses a semi-humanoid robot with wheeled legs and a humanoid upper body.

Gao says DYNA is often perceived as deployment-focused, but he sees it as product research. The company researches breakthroughs future robot products will need. A product loop helps reveal the next research direction; without one, the team would not know where to go next. Its product philosophy is research-led, with deployment providing feedback.

Asked whether the planned release is DYNA-2, Gao confirms that it is the model and robot due out at the time. He groups DYNA’s technical progress into three systems: System Two for embodied reasoning, System One for translating language into actions, and System Zero for motion control.

System Two understands physical properties, the current state, task semantics, and what should happen next. System One is a world-action model that converts language into actions. System Zero turns trajectories into fast, smooth, accurate movement through reinforcement learning and control. Gao sees strong signs of generalization in Systems One and Two.

Gao describes two ways to teach a person or robot: demonstrate an action or explain it in language. Generalization can be measured by how many demonstrations a model needs and how well it follows instructions. DYNA has seen new tasks learned from only a few demonstrations and a major increase in language sensitivity, which it plans to discuss in a release.

DYNA already has commercial customers for complete workflows in settings such as hotels, though deployment is not fully complete. Gao says replacing one employee is difficult under current foundation-model limits: hardware, motion control, and data speed make it hard to match a person’s pace.

Gao says the useful target is not beating a person’s speed but working longer. In the US, a person works about eight hours a day, so DYNA set a goal for one robot to work twenty-four hours and match one person’s output in eight. Gao says the company has already exceeded that target. Zhang asks why DYNA did not start with humanoids, manufacturing, or home use.

Gao says DYNA did not rule out any category in advance; it explored what it could do. Home robotics may be the ultimate form of generalization, but a robot that works in every home would not automatically be a good product. Its daily output must match buyer expectations, which are subjective and involve many safety and task corner cases.

Gao says buyers’ expectations are subjective, much like the differences people experience with housekeepers. Even a general-purpose housekeeper can vary in speed, personality, and quality. Home robots face many more safety and task corner cases; he thinks broad generalization may take thirty to fifty years, and would not automatically create a market.

Zhang asks whether a humanoid robot in the home is the most generalizable product form, perhaps like the Model T. Gao says the home has the highest generalization demand of any setting he can think of, but generalization is only one dimension of consumer value; speed and doing things the way people prefer also matter.

Gao says a home robot must satisfy many product details, preferences, and safety requirements. Generalization alone does not make a good home product, and a home robot would not automatically work in an industrial setting. Asked for DYNA’s sweet spot, he says he expects demand for general-purpose B2B robots to exceed B2B2C demand.

Gao says human desires keep growing as labor costs fall and goods become cheaper, so businesses will need robots to produce a wider variety of products. He estimates that one robot for every three-person US household might mean about 100 million home robots, while B2B demand could be larger.

Gao expects the US to need more B2B than B2C robots, including robots for food production and hotels. Commercial settings are easier to standardize and do not require solving emotional or aesthetic preferences to the same degree. Zhang asks why DYNA did not start with home humanoids, which many consider the crown jewel.

Gao says B2B is the clearest starting point: if a robot cannot do a laundromat worker’s job, it is unlikely to handle every household’s laundry. The scope is controllable. A hotel laundry room may need two to five employees, depending on hotel size, and laundry is essential to making rooms available; during the pandemic, some owners did the work themselves.

Gao says hotel laundry is a painful, fundamental problem. Most laundry rooms are similar, so solving the workflow could create DYNA’s first research feedback loop and generalize across roughly 100,000 US hotels. He estimates the market in the billions and sees laundry as a necessary step toward home use: if a robot cannot handle a laundry room, it is unlikely to work at home.

Gao says US hotels have high willingness to pay and relatively forgiving requirements for error and timing. Existing food-delivery robots address only a small part of the opportunity; a capable robot could eventually handle laundry, cleaning, food preparation, or other work in hotels. A hotel contains many additional workflows to explore.

Gao says a hotel is a standardized version of a home, making it suitable for early deployment. Housekeeping may become feasible once model intelligence is sufficient because the work is repetitive, though it has many corner cases. Standardization helps near-term deployment, while long-term generalization still requires substantial pretraining data.

Moving between hotels can teach deployment and generalization. Gao says room cleaning is harder than laundry because it involves more corner cases, such as cleaning showers, handling a messy room, or responding when guests break or take things. The workflow takes longer too, so DYNA wants to close simpler loops first.

Gao defines a minimum viable product as a closed loop that can then be expanded. Asked whether hotels were one of DYNA’s first scenarios, he says there were several, mostly commercial. He does not divide the market into commercial, industrial, or production categories; he asks which workflows the robot can perform.

DYNA ranks work by workflow complexity rather than industry label. Folding napkins and loading or unloading are short-horizon workflows; laundry is more complex, and hotel cleaning is harder still. Gao describes these as levels that can be unlocked as the model’s intelligence improves.

Gao says workflows are layered. The laundry room may be around level two or three: it is consistent but involves a long sequence, opening washer doors, retrieving dropped items, and handling many corner cases. He commercializes workflows within the agent’s current intelligence range and sets the next research goal beyond it.

Gao says deployment is a way to understand the agent’s current level and identify corner cases that should guide pretraining. It is not only about generating cash flow. The team uses each workflow to see what the agent can do at each level and where its limitations lie.

Zhang asks how DYNA can pursue ambitious research while spending so long on one workflow. Gao says they are not in conflict: pretraining must keep improving regardless. DYNA says its released dataset is the industry’s largest and shows a strong scaling law, and that its pretraining effort is at least as large as that of other companies.

Gao says research labs also divide people across pretraining, post-training, and technical directions; DYNA does the same, but organizes post-training around workflows. Using one million hours of human egocentric data, the team found a strong scaling law, which he calls an important milestone because the industry has struggled to break through an intelligence ceiling attributed to data limits.

DYNA initially believed high-quality data was essential. DYNA-1 used the best data the team could collect, and repeated A/B tests showed it trained the strongest model, so Gao told the team to focus on quality. He then asks whether the path to a GPT-3- or GPT-4-level robotics model might require scalable, lower-quality data instead.

Gao admits he selectively heard evidence that supported his belief in high-quality data. LLM and video-generation researchers told him that removing some data in ablation studies could improve pretraining. But once DYNA began building workflows, it found the most important thing was giving the model the broadest and most diverse range of human experience.

Gao says data-quality optimization becomes meaningful after reaching internet-scale volume; before then, broad coverage matters more. Looking three, five, or ten years ahead, the team could not imagine a better ultimate data form than egocentric and exocentric video. Betting only on teleoperation, UMI, or another non-egocentric source could implicitly assume robot hands will not work.

DYNA believes robot hands are primarily an engineering challenge, not a research problem, and decided to bet on egocentric data as a long-term source. The company also uses simulation data. Gao rejects the idea that a team must choose between real-world and simulated data; the mixture depends on the training objective and the model the team wants to build.

Gao says a model reflects its founders’ priorities: OpenAI’s consumer focus shaped its early models, Anthropic chose to bet on coding agents, and Google has long emphasized multimodality. Robotics teams also differ in what they want pretraining to achieve. DYNA uses all the data sources it can, and Gao says he had not known China framed the debate as real-robot versus simulation camps.

Gao considers real-robot data, teleoperation, UMI, egocentric data, and simulation means to an end, not the ultimate strategic bet. The important question is what dataset represents the future. Industry views shift quickly—from teleoperation to UMI and then egocentric or simulation—so founders should commit to what they believe is most fundamental.

For DYNA, pretraining should reduce the number of demonstrations needed for deployment and improve language following, which is essential for chaining workflows. Gao says a company’s understanding can change with the market, but the product must remain its guardrail and the destination must be real-world use, however long the ramp takes.

Gao says robotics is still ramping toward its ultimate product form. He cannot predict how many years AGI will take because the data ceiling is not yet clear. A robot’s end state, in his view, is receiving a vague instruction and completing the whole job. That future could be close or far away, and each embodied breakthrough will reveal another possibility.

Gao compares robotics AGI with coding or general-purpose agents. AI progress can appear flat for a long time and then jump suddenly; four years after ChatGPT’s release, coding agents have reached a level few could have predicted. DYNA’s major jump came from finding a scaling law in egocentric data, suggesting that the data ceiling remains high.

Gao says probing the ceiling of intelligence and building products along the way are part of the same process. DYNA moved from treating an individual task as the smallest unit to treating a workflow that way. As language capabilities improve, the next stage may require fewer demonstrations and stronger generalization, until a robot can act from language instructions alone.

Zhang asks whether language alone could eventually direct a robot through many workflows or let robots collaborate. Gao says it is hard to predict the effects of scaling, but the team can track what additional intelligence enables. He argues that completing one scenario thoroughly can reveal the current bottlenecks through its corner cases, quality requirements, and speed demands.

Gao says the difference between one workflow and a hundred is largely the data for those workflows. DYNA’s product is a system that turns data into a workflow. Once the team has solved the underlying logic, many workflows may differ mainly in procedures and corner cases, which become data problems.

Zhang asks whether a new skill in a general model could swallow an entire application and make scenario-specific optimization obsolete. Gao says this can happen, but completing one workflow does not prevent broader research. He compares DYNA with Anthropic: coding-agent work does not limit Anthropic to coding, because pretraining and post-training can also support agents in areas such as spreadsheets.

DYNA’s pretraining remains broad and focused on generalization; workflow-specific post-training is where the company tests products. Each workflow reveals bottlenecks that can guide future pretraining. Gao says that even at ten million hours, the dataset should remain broad, though the resulting capabilities may affect different post-training tasks to different degrees.

Gao says each model generation reveals which bottlenecks have disappeared and which remain. Workflow testing helps distinguish what pretraining can solve from what it cannot. A model may learn through language to place a napkin neatly in a bin, while a rare double-pickup that reduces output may need a different solution. The team uses those findings to decide where to invest.

Gao says the important milestone is when robots create real human value and reach scalable product-market fit, not the industry’s current stage label. DYNA thinks it may be one or two years from reaching fit in a specific scenario. Once a workflow works, the company expects to run ten or twenty workflows in parallel. He compares the potential breakthrough to ChatGPT becoming useful to anyone.

Gao says GPT’s breakthrough was that it worked for people with different questions and ways of thinking. The question is whether robots need the same kind of generality. He sees this partly as a product-definition issue: a robot may become useful to many people without solving all of physical AGI or reaching unlimited generality.

Gao says a robot should be both an agent and a product, not just an agent. It need not have unlimited intelligence; its capabilities must be turned into a product that works at that level. He compares this with GPT-3, released in 2020 mainly as an API, and ChatGPT, which launched in 2022 with a conversational product form.

Dozens of developers could use the GPT-3 API, and companies such as Jasper built products on it. Gao says ChatGPT’s major change was the product form: a browser-based chat interface made it accessible to far more users. Product shells change with the level of intelligence. DYNA is investing in embodied reasoning, whole-body control, language stability, workflow execution, and its egocentric scaling law, all toward an eventual product.

Gao says every technical effort is guided by a product endgame, and DYNA has a strong sense of direction. He does not think that endgame is simply a robot shape such as a humanoid; it is a product for how people use robots. He says the idea is not ready to discuss in detail and may be a topic for a later interview.

Gao says the product must make robots useful to each person, beyond the hardware itself. DYNA is focused on B2B, but he believes a B2B product could ultimately reach everyone. Whether a robot should be humanoid comes down to ROI: do the legs add enough value to justify their cost? Zhang then asks how DYNA’s research on the robot brain differs from other companies’ work.

Gao says DYNA is not fundamentally different from Physical Intelligence, Generalist, or Google DeepMind in trying to generalize models, improve language steering, and learn from demonstrations. The team has worked on the robot brain from day one. Its reputation as deployment-focused came partly from presenting DYNA-1 around real-world deployment instead of explaining the research behind it.

DYNA wanted real-world deployment, but Gao says it is fundamentally a research-heavy company. DYNA-1 itself required a research breakthrough. The team sees intelligence as the main bottleneck to deployment and uses deployment as a feedback loop for testing intelligence. Gao estimates 80%–90% of the company’s capacity goes to intelligence.

Gao says deployment is less burdensome than people assume because DYNA’s pretraining generalizes well and its infrastructure lets data collectors train models with few demonstrations. The data-collection team—not research or engineering—produces the deployment and demo videos. Zhang asks how the team reconciled its earlier debate about research and deployment.

DYNA’s “one employee” strategy aimed to match a worker’s daily output by handling loading, unloading, the task, and the full workflow. The plan united the team around improving intelligence enough to complete the whole job, strengthening deployment infrastructure, and making post-training faster. Gao calls it a major strategic milestone.

Gao says deployment infrastructure had to improve alongside workflow research, and post-training had to become faster so the whole workflow could be deployed. This combination of intelligence, infrastructure, and post-training united the team around its strategy. He calls it an important milestone, then lays out DYNA’s major steps from 2024 through 2026.

Gao describes the milestones this way: 2025 focused on closing a deployment loop around an individual task, improving execution quality, speed, and language control. In 2026, DYNA adopted the “one employee” strategy and clarified what pretraining needed to generalize toward. The team also had to improve reasoning and language capabilities for workflows.

The 2026 plan also required a new hardware layout and whole-body control because DYNA’s first-generation robot could not move. Gao says the team has looked beyond the current workflow toward a broader market: a system that lets anyone use a robot. The next step is not simply an extension of the current scenario.

Gao says a robot reaches product-market fit when its output within a defined scope exceeds a human’s. DYNA’s chosen scenarios should meet that ROI threshold; otherwise, they cannot form a commercial loop. Working through the loop early exposes problems—just as autonomous-driving teams learn by driving every street—and helps classify what generalization can solve.

Gao says the team must run through every scenario and corner case in its target workflows, then separate problems that generalization can solve from those it cannot. Deployment needs its own feedback flywheel. At Caper, even after the product could be installed in any supermarket, it took years to make deployment work reliably across hardware, software, models, and motion control.

Operational problems can be hard to diagnose: a robot that usually folds 1,300 napkins a day may suddenly produce only 800. Gao says deployment needs observability and support systems, and may take two or three years to get right even after the intelligence is ready. His prior experience taught him to investigate those details early.

Asked where he sees commercial opportunity, Gao says it lies in the workflows DYNA plans to tackle. Early deployments create a feedback loop that reveals the next research step. Without that feedback, the team cannot tell where to direct research; with it, he can map the next two or three steps, like crossing a river by feeling the stones.

Gao says DYNA established a China team from day one because a world-class robot requires both intelligence and a strong supply chain. The company has about forty to fifty people in China focused on hardware R&D and supply chain, and has reached initial production of hundreds of workflow robots. Gao chose not to mass-produce the previous generation because he did not think it would work.

DYNA kept a low profile partly because its research bar was high. The team had developed some language-steering capabilities before others announced similar work, but waited because it felt they were not yet at 99%. Gao later realized that such a high bar could mean never releasing anything. Still, he does not want to announce interim work when its strategy is unclear.

Gao says DYNA releases work only when its product, research, or engineering logic feels complete and worth showing. The workflow goal is distant and requires the team to work together, so the company prefers to explain the full story instead of sharing partial progress for PR. He sees robotics as an even longer race than large language models, with the industry far from convergence.

Gao says a founder needs a clear view of the eventual product and should keep moving toward it. He considers intelligence more important than product because there is no product without intelligence, though product requirements also determine how much intelligence is needed. For DYNA, the central challenge combines product, intelligence, and real-world deployment.

Gao says the US expects hardware supply chains to converge while intelligence remains decisive; China often sees supply chain as the harder long-term problem. Both matter, but over a ten-year horizon he thinks supply chain may matter more because customers ultimately receive hardware. He says the brain sets the product’s floor and the body its ceiling.

Gao says robotics may eventually compete on cost and quality. He praises Unitree’s hardware R&D and its reduction of the sim-to-real gap through system identification. That work gives Unitree an advantage. He then discusses AgiBot’s approach.

Gao says AgiBot has its own robot embodiments and hardware, ample funding, and room to explore a broader product line. He respects that strategy. More broadly, he does not yet see robotics companies as direct competitors: people openly share their research, and meaningful competition will begin once companies have real market share.

Gao says robotics companies are currently in a state of mutual respect and learning. After DYNA-1, more teams converged on the importance of success rates and execution speed; UMI, egocentric data, and world models are also becoming shared areas of interest. Zhang asks when faster industry convergence might happen.

Gao says convergence happens gradually, with the industry increasingly agreeing on useful capabilities, model architectures, and data directions. He thinks a more converged pattern may emerge over the next one or two years. Asked about the relationship between the robot’s brain and body, he says both are important.

Gao says the brain is the robot’s floor and the body its ceiling, the opposite of what many people assume. The brain determines whether a task can be done; hardware and motion control determine how well it can be done. If a robot can fold 1,000 napkins a day, better hardware and control may make it faster and more precise.

Gao says brain and body are tightly coupled. Different motion-control algorithms can produce very different results, and he doubts a company can deliver B2B quality, speed, and efficiency without owning the robot body. DYNA’s research team defines hardware specifications such as camera frame rate, shutter type, and viewing angle.

Gao says the research team also defines the camera’s field of view. A stronger brain may eventually reach the hardware’s limits, like a world-class table-tennis brain held back by a body that cannot keep up. In that case, the hardware—not model intelligence—becomes the constraint.

Gao says full-stack development is not premature because brain and body must co-develop. Hardware needs at least three revisions before mass production, and the supply chain has not converged. Waiting until the intelligence problem is solved would sacrifice years of hardware iteration; better specifications alone would not make the hardware ready for deployment.

DYNA will adopt dexterous hands when they are ready for real deployment; they are not yet mature enough. Gao says his concerns as CEO change from period to period—people, fundraising, or strategy—rather than staying on one long-running issue. Recently he has been thinking about data sourcing, a concern shared by robotics companies.

Gao thinks DYNA may have a higher data ceiling than some competitors because it can train on more data, but there are no good data sources simply to buy. From DYNA-1 to DYNA-2, its model architecture, pretraining dataset, reasoning approach, and hardware all changed substantially. Gao says it felt like becoming a new company, and recalls being confused about the next strategic milestone late the previous year.

DYNA had to decide whether to bet on deployment or research. It chose a workflow that could support a commercial loop while raising the agent’s intelligence ceiling. Gao says the team has reached that goal and is now moving beyond the workflow scenario, though he cannot yet discuss what comes next.

Gao says DYNA has plenty of cash to support robotics-model R&D. Most spending goes to talent, compute, and data; more capital will be needed when the company scales its data substantially. It has not burned much money in moving from the previous phase to this one. Zhang asks what the team validated through DYNA 1.0.

Gao says DYNA 1.0 clarified how far the agent and pretraining must advance, and what generalization and language performance are needed to support a workflow. He describes 2.0 as breaking free of a single scenario and pursuing a broader market. The system is already in place to turn incoming data into workflows and business value.

Gao says the next milestone may take six months to a year, though he cannot say exactly when it will arrive. He describes DYNA’s team as resilient and not easily shaken by changing trends.

Gao says he loves working with people who genuinely love the work. Robotics is restless, with attention constantly shifting between companies, but people who care about what they are building are less likely to be swayed by outside opinion. DYNA keeps a clear picture of its goal in mind.

Gao says DYNA is a pure-hearted, passionate team that wants to make its vision real. Internally, it feels more like a lab. Zhang asks whether organizational forms have changed from those of earlier companies.

Gao says organizations have changed significantly in the AI era. Before coding agents, large companies were structured in many layers; Google might have ten or eleven, each with its own responsibilities. Now a CEO or senior leader can reach engineers and researchers directly, sometimes through only one manager, making execution faster.

DYNA’s US team is somewhat larger and focused mainly on research, while its China team focuses on hardware. From Caper and Instacart, Gao learned to hire the strongest people for the most important work. Many individual contributors had management experience but chose IC roles because they preferred the work; they can take ownership with less management overhead.

Gao says a small team of strong people can move quickly and share a common focus. More employees can make unity harder, so he wants to build alignment in the smallest workable unit. He sees the current small-organization structure as a phase rather than a permanent form.

Gao expects DYNA to grow beyond its current size, but says future companies may need fewer people than a decade ago: work that once took a thousand employees might take only one or two hundred. Agentic coding has accelerated engineering; a task that took a month may now take a week, allowing teams to explore several directions at once.

Gao says he cannot predict how many years AGI will take because engineering efficiency keeps improving. What matters most is a team’s iteration speed. Asked how his fourth startup feels different, he says he has moved away from measuring entrepreneurship by fundraising, valuation, customers, and revenue, and toward understanding what kind of person he wants to become.

Gao says he now turns inward to ask how to become his best self and think through problems at their root. Self-cultivation helps him live more authentically, and he began this practice when he started Caper over ten years ago. Entrepreneurship is difficult, but a change in perspective can make the same situation feel larger and more manageable.

Gao says he developed ankylosing spondylitis at eighteen or nineteen. His immune system attacks his spine, and pain sometimes grew worse when he slept or sneezed. At first he resented the world and wondered why he was so unlucky. His biggest breakthrough was not an immediate cure, but changing how he approached the illness.

Gao began experimenting with his routine, diet, and other changes to see what might help. After more than ten years, he found patterns and brought the condition under control. He sees entrepreneurship similarly: both involve testing how a change in perspective or behavior affects a difficult problem.

Gao says he used to assume investors who did not understand his pitch were foolish. After shifting his perspective, he realized they saw the world differently and had things he could learn from. Recent conversations with investors have given him new inspiration. Investors assess different dimensions—financials, technology, product, endgame, and founders—and weigh them differently.

Gao says Stoicism, Wang Yangming’s philosophy, Buddhism, and Christianity share ideas about self-cultivation and one’s relationship with the world. Understanding oneself and the world’s underlying logic can reduce needless inner friction. He used to agonize over how others judged him, including whether he sounded foolish in a talk or seemed awkward on video.

Gao says he used to worry about whether people disliked him when he made a mistake or thought he was dumb when he had not explained something well. Reviewing a recording of one of his talks could make him dwell on a pause or an awkward phrase. A book helped him realize that the person most concerned with his appearance is himself; viewers move on after watching a short clip.

Asked why he still starts companies if he is no longer trying to prove himself, Gao says he wants to live more fully as his authentic self. He describes childhood as seeing things plainly, before social rules and fear of ridicule distort one’s view, and compares this to Adam and Eve eating the fruit of knowledge. For him, freedom means following one’s heart without being governed by others’ opinions.

Zhang jokes that Gao is handsome and successful yet seems to carry more inner pressure than others. Gao says his cofounder considers him relatively insensitive and that he worries less now. He thinks looking fitter and healthier reflects living more healthily.

Gao cherishes the chance to build DYNA because he loves entrepreneurship. He sees robotics as a World Cup where people from every industry bring their strongest skills and strategies. He respects the other teams and wants to compete while living expansively and improving his own abilities.

Gao says he wants to hone his craft and values the challenge of setting direction in a complex field where technology has not converged. It demands a lot of him personally, but he enjoys that challenge. He says robotics is a long race, and he wants to keep developing his ability to choose a path.

Gao says the robotics industry is always foggy, but finding one’s own way through it is fulfilling. He believes outsiders are wrong to think DYNA’s thesis is deployment-first, though they are right that the company is building something valuable. He defines success as making robots accessible to everyone.

Personally, Gao says failure would mean failing at self-cultivation. Meditation can quiet the mind, but self-cultivation is broader: it shapes one’s view of the world and response to each challenge. Some problems can be solved by changing other people or the world; inner problems require changing oneself.

Gao says losing clarity about strategy late the previous year was a major challenge. He felt unable to see where the company was going and worried the team would ask the same question. Letting go of mental baggage and distracting noise helped him think clearly again.

Gao says problems are not always solved by working longer hours. Having worked 100–120-hour weeks in investment banking, he learned that some problems require a different mindset. He had to quiet outside noise and release pressure rather than simply spend more time working.

Gao’s main burden was believing a CEO who had succeeded once could not afford to make mistakes. That pressure led to unwise decisions, or choices made only to avoid error. He had also wanted to prove his first success was not a fluke, but now sees that as self-validation shaped by others’ judgments and no longer considers it important.

After moving past that mindset at the end of last year and early this year, Gao felt deep gratitude toward his teacher, team, wife, and children for supporting him. His cofounders continued to stand by him despite decisions he regretted.

Gao says unclear strategy and insufficient attention to company culture weakened team cohesion. The resulting technical choices were not always grounded in first principles and wasted time. Zhang asks whether the problem was the choice between research and deployment.

Gao says the team’s drift was not simply a research-versus-deployment conflict. It had not thought through the next strategic step, so it began reacting to competitors’ announcements. Once it found a coherent direction and returned to its own judgment, the team converged and regained strong momentum.

Gao says some limits in one’s thinking require help from others to overcome. A mentor, caring teammates, and people who loved him helped him see how to move forward. A hard wall can be the ceiling of one’s own understanding, and it can take another person to reveal what lies beyond it.

Gao says he cannot yet explain in simple terms how his worldview changed. He believes it begins with goodwill toward the world and treating people kindly. Instead of proving one’s ego, the work is to understand oneself, clear away the fog, and rediscover an authentic self. He says he changed enormously this year.

Gao describes the change as an inner rebirth. He says founders continually encounter hard walls and must decide whether to change the world around them or change themselves. For a long entrepreneurial journey, changing one’s thoughts and the lens through which one sees the world may be more sustainable.

Gao says the marathon of entrepreneurship has no visible finish line, making inner friction especially draining. Founders need to keep improving and become more certain about their direction; otherwise, they may not have the stamina to continue. The uncertainty about how long the journey will last makes it harder still.

Gao says the middle of the journey is often the hardest: the finish line and the road ahead are both invisible. His strategic reset was a simple change from the outside, but felt like a wall given his mindset at the time. Getting through one difficult period does not mean there will be no others; people experience hardship differently.

Asked how he would feel if DYNA turned out less successful than Caper, Gao says he wants to compete at the highest level and play in his own style, as his genuine self. That would make him happy regardless of the result. He believes life is a journey, not a finish line to stare at.

Gao says milestones are checkpoints, not the goal. Fixating on the outcome without enjoying the process makes the journey painful. He used to focus only on making the company succeed, but now sees setbacks as part of the path: fall into a pit, climb out, and keep going. Asked for a favorite food, he names Fuzhou banmian noodles and bianrou wontons; his favorite place is wherever his wife and children are.

Gao says he has three daughters, aged one, two, and three, and married at thirty. Asked for an influential book, he recommends Life Has Only One Thing by Jin Weichun. He names Jason Ma’s “Eureka” as a robotics paper that reshaped how robots might learn by themselves and introduced him to Ma.

Asked what DYNA’s most important bet is, Gao answers infrastructure. It matters because it determines how quickly the model can iterate and how far it can scale. Zhang asks what he thought of the studio name “Language Is World.” Gao says robots perceive the world through language and sensory input: vision, audio, physical properties, and other modalities. Language helps them understand the world and turn that understanding into action.