Hu Yuanming: Meshy AI, Taichi, Graphics, and Commercialization

2025/08/08 · Source link · AI Summary archive

Summary

Meshy AI has proved that 3D generation can support a fast-growing vertical AI business, but Hu Yuanming does not see it as the endgame. The company has roughly 4 million users, 2.5 million monthly website visits and about 55% market share in the US; revenue comes mainly from subscriptions and APIs, growing about 20% month over month and more than tenfold over the past 12 months, with Meta as its largest customer. The investment takeaway is that category leadership and a self-funding flywheel are in place, but Studio-grade assets still often require post-processing in tools such as ZBrush, so the technology has not yet fully penetrated the workflow.
Meshy found the sweet spot of a market too small for giants to prioritize but large enough for a startup to survive. The gaming industry is nearing $300B, while facing time diversion to short video, stagnant gameplay and business models, longer AAA production cycles and soaring asset costs; on the consumer side, 3D printer shipments are growing about 20% a year, while users lack models. Hu’s first-principles logic is blunt: “If people are already paying for 3D models, and we can generate them at close to zero cost, then we are rich.”
The real bet is whether this team can keep crossing categories, not simply how large the 3D TAM becomes. Hu believes 3D generation alone could support a public-company-scale business, but ultimately wants to build a technology platform akin to NVIDIA; the longer-term unifying idea is “AI for fun,” expanding from games and printing into generative entertainment that lets ordinary people experience another life. The time constraint is explicit: find the next growth curve before Meshy’s exponential growth turns linear and then slows.
The shift from Taichi to Meshy was a life-or-death pivot that used commercial reality to overturn technical vanity. Taichi spent roughly 18–24 months searching for a business model, with low-margin outsourcing as its main source of revenue; Meshy was the team’s third product, attracting about 1,000 users within 8 hours of launch before the team went all in. Hu’s rule for founders: “Fire yourself every three to six months”; if three months later you do not think your former self was “an idiot,” you have not grown.
His next-generation technology thesis is not to abandon simulation, but to demote it into a 3D inductive bias within learning systems. Isaac, MuJoCo, Drake, PyBullet and Genesis all remain constrained by the sim-to-real gap; pure synthetic data “will definitely not work” for training a foundation model from scratch, though it can augment real data. The more viable path is to retain hard constraints such as object persistence, non-overlap and camera projection, while handing friction, materials, boundary conditions and stress-strain curves to data-driven modules. Taichi 2.0 could become infrastructure for this hybrid simulator.
In Hu’s framework, open source is not a revenue model but a hiring, customer-acquisition and trust channel for technical companies. He moved from “open source is great,” to questioning why a commercial company would give away its best technology, and then back to investing in Taichi; his conclusion is that “open source does not make money,” but good projects attract top talent and customers, effectively using R&D as marketing. The prerequisite is a complementary commercial loop; otherwise the open-source unit quickly becomes an organizational cost.
A CEO’s core job is not to be liked, but to respect facts, choose the right market and lead the company to win. Hu roughly allocates his time 40% to recruiting, 30% to roadmap and strategy, and 30% to hands-on business; his talent criteria are hungry, humble, smart and clarity. The company uses blind tests of old and new models to resist self-congratulation, because “most effort in the world yields no return.” He defines a good business as a large market that remains non-consensus, where the team has an advantage, with interest and timing as additional factors. An IPO is a testable goal for the next three or four years, but a means rather than the purpose of entrepreneurship.

Deep dive

1. Hu Yuanming Reduces His Self-Definition to “I Am Myself”

Faced with labels such as Tsinghua’s Yao Class, MIT, Taichi and Meshy, Hu Yuanming’s shortest answer was: “I am myself.” More specifically, he is driven by interest and a sense of mission, and draws energy from creating things.
His logic of action has a high interest threshold: he can devote “1,000%” effort to things he likes, while for things he does not like, “I would rather not do it.” That explains both his productivity and the pain of a research mismatch during his first year at MIT.
He puts talent, physical condition and even effort itself under the heading of luck: “Maybe 99% of it is luck.” Being lucky enough to like what the era needs is what makes sustained hard work possible.

2. A Peach Tree Made Him Suspect from Childhood That the World Was a Simulation

As a child living in a school dormitory, Hu watched the peach tree below and kept asking: when he was asleep and no longer observing it, was the tree quietly growing, or did it only fast-forward into its expected state at the moment he looked again?
He soon encountered a recursive problem: if rapidly observing many objects increased the world’s computational load, why did the observer not feel the world slowing down? His answer was, “I myself am also being simulated by this world”; an object inside the system cannot distinguish how the parent system is implemented.
He imagined the Planck constant as the simulator’s floating-point precision, the speed of light and light cones as mechanisms limiting the cost of particle interactions, and the Big Bang as the initial condition. He explicitly placed all of these in the category of “very possible, but impossible to prove or disprove” conjectures.

3. He Discovered Mechanics in Code Before Learning It Was Newton’s Second Law

His parents taught computer science, so Hu encountered computers as early as kindergarten and primary school. He downloaded roughly 6MB of Visual Basic 6.0 from a software site and began writing basic physics simulations.
When a spaceship moved directly in response to arrow keys, he felt that it was “unrealistic,” so he changed the keys to control incremental movement. Only later did he learn that the concept he had rediscovered was Newton’s second law.
In Crazy Machines, a tennis ball passes through levers, pulleys, fans and a toaster before finally hitting a cat. That showed him how simple components could combine into complex machines. His own spaceship game added thrusters, weapons and upgradeable components, but was later lost when his hard drive failed.
His interests then expanded to RPG Maker XP and Ruby. What ran through all these projects was not “learning to program” itself, but the urge to “create a world using addition, subtraction, multiplication and division.”

4. Competition Trained Execution; Physics Simulation Trained Him to Question Models

Computer competitions such as NOI in middle and high school taught him to “write the code and get it right” under tight time limits and intense pressure. That competitive-programming ability remains the foundation for handling urgent problems.
In middle school, he recreated a mass-spring system inspired by World of Goo, using balls and springs to build soft bodies. In high school, wanting to avoid real dominoes collapsing when touched, he turned to studying rigid-body simulation on a computer.
It was through collisions and friction that he discovered many models were only approximations. His research then extended from fluids as an undergraduate to deformable objects, elasto-plastic objects and gas simulation during his PhD.

5. The Most Important Output of Yao Class Was Seeing, Again and Again, That “There Are Always People Better Than You”

Hu graduated from Tsinghua in 2017 and still regularly meets with his Yao Class classmates. His biggest gain from the four years was not the coursework but a group of peers who could casually win ACM World Finals gold medals and consistently score full marks.
The most striking example came in his first computer-application mathematics exam. The professor directly tested material he had not had time to teach; one classmate repeatedly said before the exam that he was going to fail, then scored 98.5. Hu calls that kind of setback “eye-opening”: the earlier you learn that you are not as capable as you thought, the earlier you can grow.

6. His Actual Undergraduate Publications Were All AI-Related, but Early Models Were Far Less Stable Than the Myth Suggests

When domestic opportunities in graphics research were limited, he reproduced large numbers of SIGGRAPH papers in his dorm room, breaking them down from simulation to rendering. He also learned that papers often omit key implementation details, so “you cannot be too superstitious about what papers say.”
After AlexNet, ResNet and AlphaGo arrived in succession, many of the best students moved into AI. At Microsoft Research Asia, Hu worked on white-balance prediction, with the result published at CVPR 2017. In another project combining RL and GANs, he trained more than 1,000 models to obtain one result that was merely “barely usable.”
The experience gave him an early view of the stability limits of RL and GANs. In his view, sparse reward had still not truly been solved by the time of the interview; the common approach remained to pretrain and then fine-tune with RL, rather than have RL “theoretically learn everything well” from scratch.

7. He Nearly Dropped Out of MIT in His First Year, Mainly Because His Research Interests and Constraints Were Misaligned

When he entered MIT in 2017, he remained firmly committed to graphics. In his view, Boston’s industry leaned more toward biotech, while MIT was slower to embrace AI than Stanford, which was closer to major industry centers; more people were still working on traditional foundational problems such as compilers, simulation and 3D printing.
His first adviser’s research interests did not match his own, while a funding project imposed additional constraints. He worked on topics he cared about while completing tasks he did not enjoy. Although he remained highly productive, he became miserable enough to consider dropping out.
The crisis sharpened his self-understanding: when given work he genuinely likes, he can devote himself to it completely; if he is required to work on something he does not believe in for a long period, the environment itself needs to change.

8. After Changing Advisers, He Shifted from “Publish More Papers” to “What Problems Are Worth Solving”

His Tsinghua senior, Jiajun, suggested making the simulator differentiable, which led to ChainQueen and DiffTaichi. Junyan repeatedly emphasized that research must create impact, rather than merely complete a technical exercise.
Hu subsequently joined the groups of Frédo Durand and Bill Freeman. Both advisers were relatively hands-off, giving him substantial freedom. MIT’s concentration of talent, platform and the certification value of a PhD ultimately made staying the better choice.
If he could advise his first-year PhD self, he would say: “Spend more time first thinking about what kind of work is important.” He agrees with the idea of spending more time defining the problem and less time solving it.
His practical qualification is that before publishing enough papers and “padding” enough papers, people may not know what an important problem looks like. Only after seeing papers generate no excitement and few citations does one develop judgment about direction.

9. His Four Counterfactuals About the PhD All Point to Expanding the Radius of Exploration

In the summer of 2020, he went to NVIDIA to work on Taichi-related projects. Looking back, he believes that accepting Kaiming He’s offer to join FAIR and entering the AI frontier earlier might have helped him more in the long run, though he does not view NVIDIA negatively.
He would still have worked on Taichi, but would not have restricted it to physical simulation. If he had served Transformer optimization earlier, Taichi might have occupied a position similar to Triton. Taichi 2.0 is a renewed exploration of that road not taken.
He would still have set “graduate in one year” as a knowingly unrealistic target, because “aim for the moon and you may land among the stars.” By his third year, nearing graduation, he no longer knew what else a PhD could offer him. His papers had enough impact, so he was allowed to graduate early.
His final counterfactual is that he would have spent less time coding behind closed doors and more time speaking with people from different fields. He repeatedly cites Richard Hamming’s warning: closing the door produces more output in the short term, but “five years later you definitely won’t know what to do.” After becoming a founder, the corresponding action is to keep listening to customers.

10. The Appeal of Computer Graphics Is Creating a Complete World from Basic Operations

Hu defines computer graphics as “using addition, subtraction, multiplication and division to create a virtual world.” CPUs and GPUs only perform basic operations, yet can generate images like those in Cyberpunk 2077 and The Witcher 3.
He describes the field’s evolution in stages: early work focused on visualization and photorealistic rendering; around 1995, work on smoke and similar problems pushed the field toward physical simulation; the technology later spilled over into robotics, computational photography, super-resolution and computer vision.
Graphics also laid the hardware foundation for AI, but its own commercial momentum is weaker. Hu’s view is that when a problem becomes extremely clean, its benchmarks are fully systematized and the solution is reduced to arranging increasingly complex modules, it often means “there are relatively few frontier-expansion opportunities left.”

11. Taichi Fit One Billion Particles into a 3090 Through a Compiler, Not a Single Algorithm

Taichi initially aimed to solve simulation infrastructure: making simulations easier to develop, faster to run, more memory-efficient, compatible with existing ecosystems and naturally supportive of differentiable programming, so fields such as robotics could compute gradients directly.
The signature PhD result was running one billion particles on an RTX 3090 GPU with 24GB of VRAM. The memory budget meant that each particle could not exceed 24 bytes, making ordinary data structures impossible to fit.
His extreme compression scheme packed x, y and z into 11, 11 and 10 bits respectively, forming one 32-bit integer. Through quantized fields, the compiler automatically quantized Taichi data structures, avoiding the need for users to write low-level bit operations that would be extremely difficult to maintain.
CUDA, Python, LLVM, MPM and GPU programming had all matured at the same time, creating the moment for Taichi to “make a splash.” It also required a young person with the fearlessness of someone who did not yet know what could not be done, willing to build a compiler alone.

12. The Romance of Taichi Was One Person Playing Scientist, Engineer, Product Manager and CEO at Once

Hu describes the experience in two words: “It felt great.” He went from PTX and compiler intermediate representations to GitHub CI/CD, documentation and Zhihu distribution, personally running the entire full stack.
Because he was himself a simulation user, he could judge product requirements directly. As a scientist, he tracked new technology; as an engineer, he handled architecture, performance, multithreaded compilation and cache; then he integrated the results into a product other people wanted to use.
His most memorable moment was “using my own programming language to implement my own program, then discovering that it was even easier to use than the library.” At the time, it felt “really damn awesome”; he stresses, however, that it was only more usable in a specific niche.
During an internship at Adobe, he spent three or four days adding autodiff to Taichi, turning an MPM soft-body simulator that only supported forward computation into one that could run backward. The logic of the CUDA-based ChainQueen required only two additional lines to become end-to-end differentiable code.

13. He Changed His Mind on Open Source Three Times, Ultimately Positioning It as Technical Marketing

During his PhD, he made every paper experiment reproducible end to end through a single command line. In a graphics community where many papers shipped without code, this reflected his initial firm belief that “open source is great.”
After founding a company, Taichi had documentation, a community and users, but still could not find a commercial loop. He briefly concluded that commercial companies doing open source faced “nine deaths for one life”: if the best product was already free, where would the revenue come from?
Once Meshy’s cash generation improved, he revised his view again: “It is true that open source does not make money,” but a good project can attract the best talent and customers and build trust through code. As with a game, where the same budget can be spent on development or marketing, open source uses R&D to build distribution.
The conclusion is not that every company should open source, but that it must be “very strategic.” Open-source capability needs to complement paid products and strengthen the competitive position; otherwise the organization will soon ask why it is maintaining a department that creates no revenue.

14. The Key Error in Simulation Is Not Newton’s Second Law, but the Boundary Conditions of Reality

Hu does not believe Newton’s second law has failed. The problem is that friction, materials, air resistance and boundary conditions are often approximated crudely; the mechanism of two smooth surfaces rubbing against each other is entirely different from glass sticking to a wet surface.
He argues that simulators should become more data-driven and more neural: high-performance MPM should continue handling the underlying solve, while stress-strain curves, elasto-plasticity and other difficult-to-model components should be learned from data.
The reverse is also true: AI needs inductive bias from the 3D and physical world. RoPE shows how structural priors can reduce the cost of learning. In theory, a pure MLP might learn everything, but in reality there is not “enough compute and enough data.”
His balancing principle is: “Do not overstep your role, and do not underestimate the model’s ability to learn.” At the same time, not every problem should be handed unconditionally to learning; simulation needs more AI, and AI needs more 3D bias.

15. Camera Projection Is His Best Argument Against Learning Everything End to End

In video generation, traditional graphics can complete camera projection with fewer than 50 FLOPs. If AI had to relearn the same mapping, it might require an additional 1 billion parameters, which is clearly uneconomic.
When video models rotate the camera 360 degrees, they often return a completely different world, exposing the lack of physical persistence. Fixing the scene with particles, voxels or another 3D representation can reduce the model’s freedom to rebuild the world from nothing.
Tairan He listed simulators including Isaac, MuJoCo, Drake, PyBullet and Genesis. Hu believes their common bottleneck is the sim-to-real gap: parameters, boundary conditions and solver states are difficult to calibrate all at once.

16. He Agrees That Real-World Data Is the Foundation and Synthetic Data Can Only Augment It

Tairan He cited Sergey Levine’s article “A Spoonful of AGI,” which argues that simulation might actually constrain robotics foundation models. Hu’s response was: “That makes sense.”
His judgment is categorical: if all or most of the data used to train a foundation model is synthetic data, it “will definitely not work.” If something can be synthesized, it follows rules written by people; reality may contain 100 million cases while the rules cover only 10,000, and the rules themselves may deviate from physics.
That does not invalidate general robotic simulation. If simulation is combined with a VLM or another foundation model, the final system may be “neither a simulator nor a neural network,” but a 3D inductive bias embedded with neural networks.
This could include an end-to-end differentiable 3D sampling operator: the trajectory of a thrown ball can be handled by the structural capabilities of a simulator, while complex contact and boundary conditions are delegated to data-driven adaptivity.

17. The Ceiling of Physics Simulation Is That Local Worlds Are Computable, but the Entire World May Not Be Predictable

Hu offers a carefully qualified conjecture: if reality itself is a simulator, a simulator built inside it is unlikely to be stronger or faster than the parent system. That makes it difficult to predict a stock market containing every participant in the world, while a small-scale wind-tunnel simulation remains feasible.
He observes that fundamental breakthroughs in purely numerical methods such as finite element, finite volume, MPM and mass-spring have become less frequent in recent years. Rendering also appears to be entering a mature phase, repeatedly combining MCMC, path sampling and adaptivity.
But he rejects the idea that simulation has become useless. CAE software such as Simcenter and Fluent is still used for aircraft and rockets, and CFD remains closely linked to wind tunnels; aircraft mainly interact with air, while robots manipulating vegetables and plates must handle far more complex contact and material behavior.

18. If He Rebuilt a Simulator, Learning Would Lead and Physical Laws Would Retreat to the Bottom Layer

A new system would preserve elements that are harder to violate: a cup on a table will not disappear in the next second, two objects cannot occupy the same position, and the camera will still obey perspective projection. Higher-level physical laws and material behavior could be data-driven.
Given enough time, he expects ray tracing to be replaced by neural graphics. With a DLSS-like method, a 4K output might natively render only at 1K, or sample just one-eighth or one-thirty-second of a pixel, before a learning system fills in the rest.
Asked whether Sora amounts to solving 3D, he answered that “to a certain extent, it can.” The long-term direction will certainly be learning-led, but graphics will remain as inductive bias, synthetic data or learnable modules—and earlier graphics technology helped make Sora possible in the first place.

19. Two Different Gaps Separate the Real World, the Observable World and the Model World

The first gap comes from the observer’s limitations: Hu still does not know what was “really” happening to the peach tree when nobody was watching it. Each person sees only a tiny fraction of the world.
The second gap comes from insufficient compute and data: the reality humans can observe and the reality physics simulation can cover are “a hundred thousand miles apart.” No matter how accurate CFD is, it cannot eliminate the need for a wind tunnel before an aircraft takes off.
Discussing GPT-4, he sees a model’s ability to mimic part of human cognition as a miracle. LLMs live in a Platonic cave formed by their training data and loss function, but humans may observe even less data: “Are we really stronger than LLMs? Not necessarily.”
He does not believe AGI must have Newton’s or Lagrange’s equations written into it in advance, but it must be able to use tools: access a computer, command line and Python, and write finite-element code when needed. Fixed next-token compute cannot adapt to every problem; tool use gives difficult tokens additional computation.

20. Robotics Matters Greatly to Him, but Not Yet Enough to Make Him Enter the Field Personally

Hu acknowledges that robotics could generate enormous impact, but he does not yet know what he could contribute. More importantly, his level of interest is not high enough.
His constraint has not changed: only genuine interest can produce excellent work. If robotics someday sparks enough curiosity, he may move in that direction; until then, he will not force a bet simply because the sector is hot.

21. Entrepreneurship Was Not a Long-Term Plan, but a Direct Response to Independent Creation and New Experiences

During his PhD, Hu watched large numbers of interviews with entrepreneurs including Elon Musk, Lei Jun and Wang Chuanfu. Near graduation, the basic options were to join academia or commercialize his PhD work. He ultimately founded Taichi Graphics, later renamed Meshy because it had more overseas customers.
He never seriously considered a long-term career at a large company because he “couldn’t stand someone telling me what to do.” His MIT adviser in the later years met with him for roughly half an hour each week, offering encouragement, connections and guidance without dictating a technical route.
That also explains his blind spot at the start of the company: before becoming CEO, his most professional experience consisted only of internships at NVIDIA, Adobe and Microsoft. “Or put another way, before I became CEO, I was an intern.”

22. Gaming Needs Cost Reduction, but 3D Asset Generation Has Not Reached One-Shot Delivery

Meta is currently Meshy’s largest customer because a large company’s API consumption can exceed that of a single game studio; game customers remain important. Tairan He noted that Hu had previously written that roughly 50% of AAA game-production costs are tied to 3D assets.
Hu describes gaming as a market nearing $300B but in contraction: layoffs, price increases and remakes of old games are happening simultaneously; TikTok and other short-video platforms are taking time away; after PUBG, there have been few new gameplay innovations at the same level, and after free-to-play, few new business models.
Production costs continue to rise. GTA-like projects can take 10 years; a two- or three-hour film can sell for $30–$40, while a game offering roughly 20 hours of experience often sells for a similar price. Pricing power and content costs are badly misaligned.
Meshy can help with prototyping and asset production, but Hu admits that current outputs often require ZBrush or other post-processing. They remain some distance from a Studio saying “wow, I love this” immediately after generation; articulation is also a later-stage capability.

23. Meshy Began with an Almost Brutal Chain of First-Principles Reasoning

In the second half of 2022, the team saw that ChatGPT could generate language and Stable Diffusion could generate images. They also saw users paying for 3D models on Sketchfab, while the market had no product that could directly generate 3D models.
The logic became: “If people are already paying for 3D models, and we can generate them at close to zero cost, then we are rich.” This was not a product derived from passion for games, but from existing willingness to pay.
The business model resembles OpenAI’s, centered on subscriptions and APIs, with Studio accounts as an additional product. The program’s opening mentioned more than 4 million users and annual revenue growth of more than tenfold; in the discussion, Hu added roughly 2.5 million monthly visits, revenue growth of about 20% month over month and growth of more than tenfold over the past 12 months.

24. 3D’s Niche Status Happens to Form Meshy’s Strategic Moat

Hu says Meshy’s traffic is roughly equal to that of the second- and third-place players combined, with more than 50% share in developed markets such as the US and about 55% in the US. That makes it the clear category leader while remaining a market where large companies may not be willing to deploy major resources.
An important consumer use case is 3D printing. In the US, 3D printer shipments grow about 20% a year, and many buyers own machines but lack models. A generated family portrait may still be judged “too ugly,” but once reconstruction quality matures, ordinary households could become users.
The fact that even most WeChat users cannot open OBJ or GLB files shows that 3D is not yet a mass medium. Hu instead believes the market is “just large enough for a startup to take root”: a market as large as video would confront giants directly, while a smaller one could not support a company.

25. “AI for Fun” Tries to Expand Meshy from an Asset Tool into a Mass-Market Entertainment Platform

Hu believes that fully developing 3D generation alone would be enough to build a public-company-scale business, but he is “not particularly satisfied” with that because he wants to build a company like NVIDIA rather than remain a single-purpose tool.
The broader mission is to use multimodal AI to create fun. Today that means games and 3D printing; tomorrow it could mean other content formats, perhaps without relying on 3D at all. “AI for fun” is therefore closer to the company’s long-term definition than Meshy’s current product.
He typically spends about 20% of his time thinking about the future. In his vision, AGI may mean many people no longer need to work, while generative AI lets ordinary people experience Antarctica, ancient Babylon or another life, or even revisit the branch where they went to FAIR—making finite lives richer.

26. Taichi’s First 18–24 Months Failed Mainly Because It Had No Cash-Generating Model

The team added support for AMD GPUs and a JavaScript backend and built documentation and a community, but commercialization ultimately centered on outsourcing simulator development for clients. The work had low margins and was difficult to standardize, while China’s enterprise sales environment was particularly challenging.
Hu gradually concluded that a founder like himself, who dislikes serving large customers over long periods, was better suited to starting with a consumer-facing, subscription-based product that could be delivered in a standardized way through an API. Finding product-market fit was already hard enough without adding the sales model he was least suited to.
He describes the company flywheel as follows: good technology creates good products, good products create revenue, revenue attracts good people, and good people create better technology. Technology that cannot be commercialized becomes “a fireworks show—once it is over, there is nothing left.”
The buzzwords of 2021—metaverse, open source and infrastructure—also made him realize how difficult it is for founders to remain completely insulated from social trends. They therefore need to reassess independently.

27. The Shift from Taichi to Meshy Recreated Intel’s “Fire Yourself First” Decision Experiment

Hu retells the conversation between Andy Grove and Gordon Moore during Japan’s DRAM challenge: since the board would eventually fire them anyway, they might as well walk out of the office first, then return as new CEOs and decide what the company should do. The answer was to shift toward CPUs.
Taichi faced the same logic. Continuing to invest without seeing a commercial future for the core business was not viable, so he chose to “fire myself.” After starting a company, “every three to six months you have to become a completely new person; you have to reinvent yourself every three to six months.”
The pivot encountered real resistance. Some people with no interest in generative 3D left, and the team cut another project before going all in on Meshy. Those who stayed later believed the decision was right, while Taichi’s spark was not completely extinguished.
Meshy also did not appear from nowhere. The team first used Taichi for Gaussian Splatting and NeRF, then saw the possibility of generating assets. In terms of capability, however, it was still a restart: nobody on the original team had trained a large model, although their graphics background let them transfer mathematical and engineering skills quickly.

28. Early Meshy Was Fit Only for Horror Games; It Now Enters AA and AAA Prototyping Workflows

The first product went live after about 8 hours and attracted roughly 1,000 users. The output was so poor that during user interviews Hu would say, “Whatever you do, don’t tell anyone Meshy was made by Hu Yuanming.” He was not willing to publicly claim it until Meshy 1 and Meshy 2.
In 2023, he judged that the technology had solved only 10% of the problem and hoped to solve 90% by 2025. His answer today is roughly 50%. The reason is not that progress was slow, but that “we may have solved the 90% we imagined then, only to discover later that the ceiling was 200%.”
The initial model generated one texture on each of the front, back, left and right sides, giving characters four faces and making it “usable only for horror games.” Today, AA studios and even some AAA studios use Meshy for prototypes, but overall it still covers only about 5% of the niche’s use cases.

29. Customer Interviews Pointed More Directly to Meshy Than the Team’s Existing Technical Moat

During the Taichi period, many customers explicitly said: “I wouldn’t pay for your software, but if you could sell me all the models inside it, I would pay for them.” This evidence of willingness to pay mattered more than whether the original team was good at diffusion.
Hu was not concerned that the team lacked experience with text-to-image or text-to-3D, because nobody was doing this at the time. Meshy was the first publicly accessible product to launch.
He therefore came to see market size as a dynamic variable. For an adaptable organization, the boundary is determined not only by today’s 3D TAM, but also by whether it can keep assembling smart people, identifying new demand and entering adjacent markets.

30. A CEO’s First Lesson Is Not Management Technique, but Admitting He Has No Idea How to Run a Company

The main pressure at the beginning was “not knowing what a normal company should look like.” Hu admits that he made a humble mistake: he did not know how much he did not know.
His advantage was the speed of correction: “When I discover I’ve done something wrong, I immediately admit I’m an idiot and change it.” After seeing many founders 20 years older than him, including former large-company executives, struggle in the same way, he realized the problem was not unique to someone going directly from a PhD to CEO.
He believes CEOs can be trained. The key is learning quickly and improving one’s understanding. A PhD can teach humility, hard work, presentation and judgment about frontier technology, but not markets, users or mass-producing products; those must be learned inside a real company.

31. A Good CEO Does Not Aim to Be Liked, but to “Get Things Done and Win”

Hu once wanted everyone to see him as nice, generous and a good friend. Later he realized that “if the goal of work is to make everyone like you, eventually everyone will hate you.”
Tough decisions include killing popular projects, letting go of colleagues who contribute little but are well liked, and choosing only one direction when users A and B want conflicting things. Trying to do both often means the company fails and both sides complain.
He sees conflict avoidance as a form of selfishness. If a CEO refuses to solve problems in order to be liked by a small group, he is not taking responsibility for all employees. “Whether everyone likes me is a private matter; whether the company succeeds is a public matter.”
His assessments of Elon Musk and Jensen Huang serve the same standard: direct reports may find them hard to work with, but the company’s success benefits a broader group of employees. Meshy, however, still believes in giving people room to make mistakes and improve.

32. New Managers Often Become Too Nice by Misapplying Their Experience as Strong Individual Contributors

People who are promoted were often self-driven, excellent individual contributors who needed little pushing. They therefore assume the loose management they received will work for every subordinate. But the very fact that they were promoted shows that their peers do not all share the same level of initiative and problem-solving ability.
A new manager who puts “the team likes me” first will struggle to set higher standards. Hu uses his own “history of blood and tears” to explain that some of the company’s current problems came from indecision six months earlier.
He does not equate clarity with talking more. New managers often try to dominate the conversation, outputting a great deal while the other person absorbs only 10%. Hu’s own shift was realizing that “speaking is not important; listening is.”

33. Recruiting, Strategy and Hands-On Work Make Up the CEO’s 40/30/30

Hu roughly divides his CEO time into 40% recruiting, 30% thinking about future roadmaps and strategy, and 30% close involvement in daily business. The best people create the best products, users and revenue; that revenue is then invested in R&D, forming a growth chain.
The first traits he looks for are hungry and humble. The former comes from not yet having proved oneself and wanting to create impact; the latter requires facing reality, accepting feedback, and continually rejecting and updating one’s own assumptions.
Smart does not mean showing off technical ability, but resourceful problem solving: if compute is insufficient, change the optimizer, reduce floating-point precision or use a hack to improve efficiency. Startups need some management, but rely more heavily on strong individual contributors.
Clarity means whether one’s thinking is genuinely clear. If a top performer says something completely incomprehensible, “nine times out of ten the problem is not you.” Hu’s own mistake was being smart at solutions while neglecting strategy—spending too much time on how and too little on why.

34. A Good Market Must Be Large, Non-Consensus and Matched to the Team’s Advantage

Compared with Cursor and Lovable, Meshy’s growth is not the fastest. Hu acknowledges that choosing the market is one of a CEO’s most important jobs: if you pick the right one, you can “make money lying down.” But he also warns that “anything gained through luck must eventually be paid back through ability.”
Market size answers whether a business is worth pursuing; non-consensus answers whether there is still an opportunity; competitive advantage answers whether a large company can easily take the market away. Interest determines whether one wants to do it, while timing determines whether the window has already closed.
Tairan He compared Meshy with Unitree and Pop Mart. Hu responded that today’s differences in scale do not necessarily determine future outcomes. Staying at the table, retaining the team and maintaining a growing business create the chance to catch the next good card.
Skills can be learned, but should not be completely disconnected from existing accumulation. Hu believes his technical ability is even “somewhat excessive.” A CEO ultimately has to add fundraising, recruiting, customer and commercial capabilities; no level of CUDA or PTX expertise can become an organizational advantage without time to use it directly.

35. “Respect Facts” Is Closer to Meshy’s Cultural Foundation Than Any Narrative About Effort

The foundation of the company culture is respect facts: good is good, bad is bad, and data-free claims should not be used to exaggerate performance. When model quality is subjective, the team randomly displays old and new versions on the left or right so evaluators do not know which is newer.
Blind evaluation sometimes shows that the old version is better, meaning the team spent a long time making no progress. Hu therefore rejects the self-comforting ideas that every effort earns a return and that hard work is always rewarded: “Most effort in the world yields no return.”
The brutal part is that the team still has to complete all 100 attempts, because it does not know in advance which one will work. But it must also acknowledge that the other 99 may have no value; effort cannot substitute for results.
A “losing culture” includes internal friction, blame-shifting, disguising reality, failure to conduct retrospectives and an inability to criticize the CEO publicly. A few days or roughly a week before the interview, two colleagues had pulled him into a hallway to point out management problems. He felt his “heart tighten,” but still admitted after listening that there were things he had failed to do well.

36. Stage Two Is Not Doing Stage One Faster, but Moving from Answering How to Defining Why

Stage one has a clear goal: someone else sets the problem, and the individual executes well. Stage two has no predefined objective; one must set the problem for oneself and the organization. Undergraduate students acquire knowledge while graduate students find problems; ICs complete tasks while managers define goals; employees use platforms while founders create them. It is the same transition.
The more successful stage one is, the more painful stage two can become. Obedience, efficiency and mastery of one skill are no longer enough; in some cases, even the efficiency metric has not yet been defined. Hu compares it to a boss entering a second health bar: continue using stage-one tactics and you will fail quickly.
His concrete advice for PhDs is to first publish three or four strong conference papers and accumulate feedback. Once in stage two, ask about citations, adoption, whether the work forms a coherent system and whether it changes another field, rather than mechanically turning three papers into six.
His role models include Frédo Durand, Kaiming He, Elon Musk, Lei Jun, Jensen Huang, Shigeru Miyamoto and Gunpei Yokoi. What they share is not a common path, but the ability to keep reshaping themselves through real-world feedback in their respective fields.

37. Innovation Means Most Attempts Fail; Real Failure Is Never Trying Again

Hu no longer sees a single failure as final: “As long as I’m still alive, I don’t think I’ve failed.” A company going badly or an experiment failing is only a temporary result. Real failure is to stop innovating and become extremely conservative.
He estimates that if a person works from age 20 to 70 and each major undertaking takes five years, they can attempt at most 10 major things. If all 10 succeed, the risk was probably too low for it to be true innovation; a reasonable expectation might be five successes.
His current self-assessment is “0.5 plus 0.5.” Taichi completed the first 0.5, while Taichi 2.0 still needs four or five years. Meshy has passed PMF and entered the stages from 1 to 10 and 10 to 100; reaching an IPO would complete the other 0.5.

38. Taichi 2.0 Is Looking for 50 Times the Impact of Taichi 1.0 in the AI Era

Hu’s “big fish” is not publishing a few more graphics papers, but making Taichi 2.0 infrastructure for Transformers, LLMs, robotics or hybrid simulators, with more than 50 times the impact of 1.0.
He says projects such as NVIDIA Warp, Triton and Modulus all have fewer stars than Taichi, showing that the user base remains. What is missing is a sufficiently large application domain, and the combination of simulation and neural networks may provide the next window.
Meshy plans to hire two or three GPU compiler or graphics specialists to revive the project, while continuing to recruit AI researchers, machine-learning engineers and compiler engineers across models, data, evaluation and performance optimization.

39. Meshy Uses High Pay and Real Ownership to Compete for Global Technical Talent

Hu says interns are paid on a global scale, with top annual compensation reaching RMB1M or $140,000. He personally works with interns on products and research; full-time packages are higher.
The team is distributed across Silicon Valley, Beijing, Shanghai and Shenzhen, but has shifted from extensive work from home to at least four days in the office each week because face-to-face discussion is more efficient.
One company goal is to “make Meshy a great place to work.” Some employees who have been there for three or four years consider it the best company of their 10-plus-year careers. At large companies they might once have written a single shader; now they can operate 512 GPUs and take on more important roles as the business expands.

40. An IPO Is a Three- or Four-Year Goal, Not the Ultimate Purpose of Entrepreneurship

Hu treats an IPO in three or four years as an attainable goal and believes it is achievable under favorable conditions. But an exit is only a means, because the company must still answer “what happens after the IPO?” and “if we cannot IPO, would we stop?”
The more fundamental goal is to build products users love, give employees financial returns, career growth and pride, and help everyone on the team become a better version of themselves. After listing, the company would gain the resources to pursue “the next one.”
Culture is not the goal either: “A company’s culture always serves winning.” Effort, responsibility, a pursuit of excellence, taking risks and embracing change are mechanisms for producing good people and products. If a culture makes the company bound to lose, the CEO must abandon it.

41. Hands-On Work Is the Technical CEO’s Last Safeguard Against Losing Touch with Reality

Hu still writes code, spending roughly two or three hours on it last week. This week he helped colleagues debug, read code and fix bugs, saying such work might account for about 10% of his time that week.
He worries that if he stops writing entirely, decisions will drift into “yin-yang and the five elements” style empty talk. Hands-on work need not always mean coding; it can mean designing products, speaking with suppliers or understanding technology deeply. But it must involve first-hand contact.
His minimum standard for “actually understanding technology” is specific: a leader of the model team should ideally have trained a GAN and CNN personally and understand optimizer dynamics; a compiler leader should at least have written assembly and a compiler. Otherwise it is difficult for the team to learn much from the leader.

42. He Preferred PhD Life After Changing Advisers, but Chose the Entrepreneurial State That Was Better for Growth

Hu admits that, subjectively, he preferred the PhD life after changing advisers, when he only had to take care of himself and could focus on coding. But “what you like and what benefits you are two different things”; staying permanently in a comfortable environment means giving up many opportunities to grow.
Being CEO is more exhausting: the company has to win, employees have to eat, investors want growth, customers want delivery, and he must handle organizational problems he is not good at. Yet every month or two he can see personality flaws and cognitive blind spots being forcibly corrected.
Money accounts for only about 5%–10% of his entrepreneurial motivation. Even after buying a home, one still has to answer “then what?” Doing meaningful work matters more than the wealth figure. He summarizes the cost and reward in one line: “In peacetime, nothing trains a person more than entrepreneurship.”

43. His Final Advice to Young People Is to Make Sure They Can Bear Failure, Then Take Action

For people considering a PhD, he recommends first finding something they genuinely love rather than enduring years of pain for a degree. He had reproduced dozens of SIGGRAPH papers as an undergraduate, so he knew early that he would enjoy the path from reproduction to original research.
For people considering entrepreneurship, the first step is to get ready: parents should be healthy, the family stable, and the individual able to bear failure and accept an AI startup’s near-7×24-hour demands. Once those conditions are in place, there is no need to have the product direction and every relevant experience figured out in advance.
What matters is stepping out: “Many things can be learned.” Action reveals mistakes, and mistakes can be used to become a better version of oneself quickly. His thinking at the time was that graduating from the PhD in three and a half years meant that even if entrepreneurship cost him two years, he would graduate in five and a half years—“so what?”
When Tairan He asked him to leave a time capsule for 2025, Hu did not summarize technology, fundraising or revenue. He left only two words: “Courage.”

GPT-5 and Agents Breakdown – w/ OpenAI Researchers Isa Fulford & Christina Kim Li Yifan’s 11 Years Building Hesai—and Where Opportunity Comes From