Summary

Meshy.AI’s revenue has grown 14x over the past 12 months, putting ARR close to $40M while monthly growth remains around 20%, but 胡渊鸣 has decided to bet the company’s next curve on AI-native games. Its first small game, 《黑箱》, is scheduled to launch on Steam within months, followed by medium- and then large-scale titles; his goal is to gradually turn the company from a generative AI business into a games-led company over 3–5 years — “at least build a company as successful as Nintendo.”
胡渊鸣 sets 2 non-negotiable tests for an AI-native game: it must not work without AI, and AI must make the core gameplay genuinely more fun. Replacing NPC dialogue with AI is merely a presentation-layer reskin; 《黑箱》 instead uses AI to generate attack skills, weapon combinations and world interactions, creating surprise through “the unexpected” while preserving player control through “the reasonable,” then lowering the barrier to entry with natural language. Games are a winner-take-most market: “There is no point generating 10,000 60-point games”; the goal must be “using AI to directly make a 99-point game.”
Meshy found PMF not by continuing to productize its existing graphics capabilities, but by shifting from “what can I build?” to “what will users actually pay for?” The commercial ceiling for graphics tools was limited: one user said they would not pay for rendering software but might pay RMB5 for the 3D asset library inside it. That pushed the team from tools to model libraries, and then toward 3D generation using the paradigm represented by ChatGPT and Stable Diffusion. The initial output was so rough that “only horror games could use it,” but the direction finally matched a real customer need.
胡渊鸣 believes the long-run default for game world models will not be pure pixel-based neural networks, but a hybrid world model combining generative AI with classical graphics. Existing video-diffusion world models may require more than 1,000 TFLOPS, while the iPhone 17 has roughly 30–40 TOPS and only a few TOPS can be sustained in a game; the gap remains 2 orders of magnitude. Moore’s law and battery technology are unlikely to close it. Classical engines can cheaply guarantee 60 fps, spatial consistency and physical rules, while AI generates variation — much as vibe coding has models generate code before inexpensive executors run it.
AI can replace work for which training data exists, but it is not yet good at unexplored directions, out-of-paradigm imagination or creative judgment. The organizational value of pure managers may therefore decline: everyone will need to produce directly or become an AI manager. But when tackling something no one has ever done, people still need “a flash of insight,” logical reasoning, random guesses or a “leap of faith.” 吴茗 asked whether intuition could ultimately be solved by a world model; 胡渊鸣 left the question open, noting that long-term prediction is highly chaotic and current AI still lacks the ability to “think out of the box.”
The commercial premise for AI for fun is that productivity gains will shift scarce resources from labor toward time, experiences and spiritual fulfillment. 胡渊鸣 defines freedom as “the right to say no to things you don’t like” and expects people to work less in exchange for more experiences; generative AI could deliver something close to infinite experience within a finite lifetime. The risk is as real as the one depicted in 《WALL·E》, so genuinely valuable content cannot be an addiction and monetization machine; it must use taste to find the intersection of mass acceptance, creator expression and commercial success.
For founders, the key variables are not pedigree but motivation, the courage to face facts and the resilience to “climb back out after falling into a hole.” About 18 months into entrepreneurship, after 2 consecutive product commercialization failures, 胡渊鸣 briefly considered returning investors’ money; a third pivot continued only because someone said, “Sure, give it a try. I support you,” eventually leading to Meshy. 吴茗 puts the greatest weight on motivation; 胡渊鸣 emphasizes: “If someone in the world makes this work, but it isn’t me, I’ll regret it,” and reduces the rest of his life to “science, art and fun.”

Deep dive

1. Games convinced 胡渊鸣 that the simplest rules can generate the richest worlds

胡渊鸣 entered the world of computing through Pac-Man, 《仙剑奇侠传》, Command & Conquer and Age of Empires 2. When his father’s students played networked games in the lab, he would even help “keep watch.” What drew him in was not just playing: assembly instructions offered little more than addition, subtraction, multiplication and division, yet could render worlds full of color.
In 3rd or 4th grade, he used RPG Maker XP to turn his classmates into NPCs, then emailed the RPG to the entire class. He moved from BASIC and Visual Basic 6.0 to Ruby and C++. His parents, both teachers, strongly opposed the idea that “a student would not only play games, but make games and influence even more people.” The creative thread, however, never stopped.
A gold medal in high-school competitions earned him direct admission to Tsinghua and the Yao Class, followed by a PhD at MIT researching computer graphics, physics simulation and GPU compilers. The path looked like a series of upgrades, but the underlying question never changed: “How do you describe the richest world with the simplest rules?”
At Tsinghua, he once built a game for his wife’s birthday. About 20,000 people played it within 7 days. Walking into a classroom and seeing students ignoring the lecture to play his work, he felt for the first time the achievement of creating a world that other people could enter.

2. From graphics to transformer, the primitive changes but the urge to create does not

Traditional rendering constructs worlds by simulating light transport and basic physical laws. AI-native games may instead use transformer, attention and large numbers of examples as primitives, learning from data “what mechanisms are fun and what makes players happy.” 胡渊鸣 sees both as starting from basic representations and composing them into complex worlds.
The shift to data-driven systems is that some experiences no longer need to be encoded as explicit quantitative rules. Models can build an understanding of gameplay, pacing and feedback from large bodies of text and examples, making fixed game mechanics more dynamic. But the final test remains the player’s experience.
His ultimate benchmark for “fun” is intensely visceral: “You can see a smile on the player’s face while they’re playing, and then I can hear their laughter.” 吴茗 immediately pressed him on the gap: laughter can only be observed after launch, while development still needs a system that can be optimized and iterated.

3. You can usually tell during development whether a game is fun, because every great-game feedback loop is designed

After nearly 30 years of playing games, 胡渊鸣 has built a player model inside himself. The kill-loot-level-up loop in Diablo and the multiple-path puzzles in The Legend of Zelda lead him to ask “why is this fun?” rather than simply record that he enjoyed playing.
He uses 宫本茂’s example of a rock: a director might ask the development team, “Why is there a rock here?” If the answer is that someone placed it there casually, that is not enough. Every element in a game must serve exploration, challenge, narrative or feedback; nothing should occupy the player’s attention without a reason.
A good game needs plot and an art style, but also a positive feedback loop between challenge and reward, keeping players on an appropriate “happiness curve.” The difficulty cannot become excessively frustrating, but it must keep delivering achievement and the desire to explore. A developer who truly understands players does not need to wait until the game is finished to know whether the direction works.

4. An AI-native game is one that cannot exist without AI

胡渊鸣 defines it as 2 intersecting circles. The first condition is: “The game is unplayable without AI.” The second is: “Using AI genuinely makes the game more fun.” Only the intersection is a valuable AI-native game.
Replacing NPCs in a traditional game with AI dialogue does not satisfy the first condition: remove the AI and the game still runs normally. Starting with a cool technology and then “using a hammer to look for a nail” also does not guarantee value for players; it may even sacrifice what was already fun about the traditional game.
吴茗’s challenge to the current market is worth preserving: many teams claim AI can generate infinite worlds and dungeons, but it remains difficult to find a truly breakout native AI game. 胡渊鸣’s answer is not to generate more content, but to first identify gameplay that has already been validated and then ask whether AI can remove its existing constraints.
Games are an extremely top-heavy content market: “There is no point generating 10,000 60-point games, because everyone will always play the 90-point game.” If AI is used to make a 50-point product, technical novelty provides no compensation. The only acceptable target is “using AI to directly make a 99-point game.”

5. 《黑箱》 turns unpredictability directly into the core gameplay

In 《黑箱》, AI generates the player’s attack skills, weapon combinations and interactions with the world. Remove the AI and the game becomes “very, very boring.” AI is therefore not a decorative layer for art or dialogue; it is indispensable to the gameplay loop.
Every time players combine weapons, they expect an outcome that is “unexpected, yet reasonable.” The unexpected part delivers surprise; the reasonable part means the result is still influenced by the input, giving players participation and control. AI’s value is to keep producing different outcomes without allowing difference to collapse into pure randomness.
The 3rd element is accessibility. Large language models are trained on natural-language corpora, so players do not need to master a complex editor; they can express intent intuitively. As 胡渊鸣 puts it, “everyone from 0 to 99” could find enjoyment in it.
The team plans to launch a small-scale game on Steam within months, using real feedback to accumulate experience and resources. If it succeeds, the next step will be a medium-sized game, followed by larger titles. 胡渊鸣 estimates that this progression will take 3–5 years.

6. Meshy’s rapid growth gives its second curve time and resources

Meshy’s revenue grew 14x over the past 12 months and is now approaching $40M ARR, while monthly growth remains around 20% at this scale. 胡渊鸣 recalls that the team once thought even $10M ARR would be difficult; by continuing to build, it eventually crossed that threshold.
The core business makes games an affordable second curve rather than an act of escape or an all-in gamble. 胡渊鸣 sees it as a reward to himself: the first curve has already “flowered and borne fruit,” so he can now search for the intersection of what he loves, what he can do and what the world will need in the future.
His long-term organizational goal is clear: start with small projects to lower the learning cost, then roll experience, talent and cash flow into larger works, gradually turning the company from a generative AI business into a games-led company rather than starting with a large-scale production.

7. 3 pivots rewrote the product question from “what can I do?” to “what will users buy?”

Tai Chi Graphics built industry influence but exposed the commercial ceiling of graphics infrastructure. 胡渊鸣 observed that even successful businesses such as Unity and Unreal Engine capture only a limited share of the value created by games; Epic Games earns most of its revenue from Fortnite, not Unreal Engine. The engine itself is not an ideal business model.
As the team tried other graphics products, it also found that many use cases were contracting. The real turning point came from one user: “If you sell us this software, we won’t pay you. But if you sell me the 3D assets inside the software, I might pay you 5 yuan.”
That sentence moved the team from rendering software to a 3D model library, then to the next question: traditional model libraries depend on manual uploads or trading, so could 3D models be generated cheaply in the same way that ChatGPT generates text and Stable Diffusion generates images? At the time, Meshy was the first publicly accessible product of its kind to launch.
The initial generations were highly abstract. 胡渊鸣 joked that “only horror games could use them.” But the more important change had already occurred: the product was no longer starting from existing technical capabilities and searching for use cases. It was starting with empathy and asking what users actually wanted. “No demand” is what kills most startups.

8. A technical CEO does not need to remake himself into a different kind of person

Before starting a company, 胡渊鸣’s only professional identity was intern: “I was intern before I was CEO.” He had to learn management and business, but even how an ordinary employee works each day. Every generation of his family had been teachers, and his parents “would rather see me eat chalk dust than go into business.”
He recounts the lesson Lisa Su gave him: a large number of MIT PhDs work for Harvard MBAs, and that arrangement is not necessarily rational. A technical CEO’s advantage is genuinely understanding technology; business knowledge can be learned later. The key is not denying one’s weaknesses, but building an operating logic that fits.
He also redesigned severe introversion into a way of working. Writing a public account is like building a RAG for himself: people search his articles first and only come to speak with him when something remains unclear, reducing the social cost of explaining the same thing repeatedly. After recording the podcast, he expects to need “5 hours of not talking” to recover.
When raising capital, he gives investors the data transparently rather than polishing it: “I am what I am. Come see for yourself.” This may not give him an advantage in every numerical comparison, but it means he does not have to spend energy performing and attracts investors who value candid communication.

9. Empathy needs boundaries; founders must take care of themselves before they can energize a team

In the first stage of entrepreneurship, 胡渊鸣 cared too much about whether employees were happy and was prone to placing internal sentiment ahead of customer needs. He later realized that a company exists to create customer value; otherwise, the organization can drift into a “charity,” losing its mission and positioning.
In the second stage, he swung to the opposite extreme: to hit targets, he could skip meals, sleep and “torture and abuse” himself in various ways. Now in the third stage, he retains his concern for the team while recognizing that a founder who chronically depletes his body will ultimately damage the company.
His summary is: “A founder always loves himself first, and only the surplus love can go to the team.” A CEO’s low energy transmits directly through the organization. Empathy helps with understanding users and employees, but when facing a tough decision, one must “make the call decisively” rather than let perspective-taking become endless internal churn.
Finding an AI-native game is the product of this new balance. It is both a potential second growth curve for the company and creative work he genuinely wants to pursue. The high-growth phase is not fully relaxed, but passion itself can replenish energy instead of merely consuming it.

10. AI is rewriting organizations, and the value of pure managers may decline

胡渊鸣 studied Andy Grove’s High Output Management and Only the Paranoid Survive, along with Pat Lencioni’s work on obstacles to team collaboration. But he feels that “the entire world changed again after the Spring Festival,” and that which parts of old management wisdom still hold needs to be revalidated.
He had initially planned to return to China and hire the best people for certain roles. Then, on a flight, he realized that with Claude Code, GPT Codex and agent orchestration, AI might handle these tasks reasonably well — and keep thinking for nearly 24 hours a day. Once again, the only certainty is “change itself.”
His view of AI-era organizations is therefore more radical: companies may need fewer pure managers, and everyone will need either something they can produce themselves or the ability to become an important AI manager in the future.
吴茗 pushed the question toward the moat: when OpenAI, Anthropic, 豆包 and 通义千问 release stronger or even free models, what can Meshy retain? 胡渊鸣’s answer is not that Meshy will always lead in model capability, but that it must find new needs and new works that training data cannot directly cover.

11. Work with data will eventually be absorbed by models; a future without data still requires creation

胡渊鸣’s view is explicit: “Everything for which training data exists can be handled by AI.” As long as a dataset exists, a model can be trained to solve the corresponding problem. What remains scarce are things no one has ever done and therefore have no corresponding training data.
Exploring such new directions may require “a flash of insight,” or perhaps logical reasoning, random guesses or an emotional “leap of faith” to predict the future. To “think out of the box” means forming an imagination beyond the boundaries of the existing paradigm. He admits that current AI “does not really have” this ability yet.
吴茗’s counterpoint was not dismissed: intuition may simply be a choice made when data is insufficient. If a large world simulator existed, it might contain mathematical structures behind the process and eventually turn intuition into a problem that can be modeled and searched. 胡渊鸣 accepts that possibility, while noting that long-term world prediction is highly chaotic and technically very difficult.

12. Intuition does not guarantee correctness, but it reveals what you really want

When choosing between A and B, 胡渊鸣 imagines tossing a coin. If heads means A and tails means B, which side would he want to land? Handing the choice temporarily to a random-number generator can expose the outcome he actually hopes for. That is his approach to “following my heart.”
He does not claim that intuition is always right. Entrepreneurship is a continuing collision with reality: discover that the choice was wrong, learn the lesson and make a better adjustment next time. The judgment itself is a benchmark that can only be calibrated through repeated feedback.
His actual method is closer to this: “It’s not that I make the correct choice. I make the choice first, and then I work hard to prove it correct.” That contains the patience of learning to love whatever one does, while acknowledging that execution must turn post-decision possibility into reality.
Making games is where rationality and emotion overlap. Rationally, he believes Meshy’s accumulated capability in training multimodal models can make established gameplay more fun. Emotionally, it is “if I don’t make games now, I’ll be old”; he wants to move from an operator handling personnel matters and reports every day toward becoming a creator and artist.

13. A world model fundamentally predicts the future from historical states, but the long-term future is highly chaotic

胡渊鸣’s broad definition of a world model is: given known history, predict the future. Under a deterministic view of space-time, if one knows the configuration of every particle in the world at t=0, 1, 2…n−1, one tries to infer the new state at t=n.
By this definition, physicists were studying world models 200 years ago. In today’s industry context, the term more commonly refers to using a video diffusion model or video autoregressive model for next-frame prediction, primarily for robotics, embodied intelligence and games.
Such models could theoretically simulate decisions as well: predict what the world would look like after taking A, B or C, then use those predictions to assist the choice. But 胡渊鸣 retains a key limitation — the real world is highly chaotic, prediction itself is difficult, and short-term frame prediction cannot easily be extrapolated into a reliable long-term future.

14. Games need a real-time interactive hybrid world model, not a pure pixel model

Pure pixel-based world models spend substantial compute relearning rules that traditional engines have already solved. As 胡渊鸣 points out, after the camera turns 360 degrees, a person who was sitting in front of it should still be sitting in the same place. Video diffusion has to work extremely hard to preserve that consistency; a traditional engine naturally retains object and spatial state.
He uses a deliberately playful estimate to illustrate the waste: in a 10B-parameter model, “maybe 5B” of the parameters are learning the physics of the world. Classical graphics can implement these constraints more cheaply and deterministically; there is no need for a neural network to guess them again on every frame.
Meshy’s envisioned hybrid world model combines AI with classical graphics. The engine guarantees 60 fps, viewpoint consistency and basic physics, while Generative AI generates content and interaction changes so that players get a different experience each time they enter.
The path also has a practical commercial advantage. The team can ship a game, acquire users and resources at every stage, then gradually increase the AI component. Compared with waiting for a pure transformer world model to mature, “it has a market at every moment.”

15. The compute gap and vibe coding both point to hybrid as the default, not a transition

吴茗 presented the strongest counterargument: if compute performance improves by 3 or more orders of magnitude in the future, would hybrid become merely an intermediate state? 胡渊鸣’s answer is that “it is the normal state,” and he makes the case from both edge compute and software execution.
A current video-diffusion world model may require more than 1,000 TFLOPS. The iPhone 17 may have roughly 30–40 TOPS, while a game may be able to sustain only a few TOPS before the phone overheats out of control. The gap is roughly 2 orders of magnitude. Moore’s law “has basically stopped,” and batteries are unlikely to make a sudden leap.
He does not rule out further advances in neural-network architecture and representation, but says crossing 2 orders of magnitude at once remains “quite difficult.” The result is a structural division of labor imposed by compute constraints: deterministic rules go to a low-cost engine, while generative capability stays with AI.
vibe coding has already demonstrated the same mechanism. A large language model generates code first, then inexpensive execution layers such as x86 assembly and NVPTX run it, rather than having the transformer perform all the computation directly. A world model may likewise generate mechanisms or intermediate representations first, then let traditional systems execute them. The difference is that the software has become “the most complex form of software” — a game.

16. The scarce resource AI for fun is chasing is the time released by productivity

胡渊鸣 reduces happiness, abundance and freedom to one sentence: “I have the right to say no to things I don’t like.” People may not be able to escape every responsibility, but higher productivity makes it more possible to refuse unwanted labor and allocate a finite life to deliberately chosen experiences.
He uses the evolution of the workweek to illustrate the direction. Agricultural societies required people to remain on call and might demand 7 days of farming a week. The Industrial Revolution brought the weekend; Ford then helped establish a 40-hour week with 5 8-hour days. Some European countries are continuing to explore a 4.5-day week. Meshy’s work-from-home Wednesday is also intended to increase flexibility, not reduce the team’s commitment.
If the trend continues, people will have more leisure time and seek “to experience nearly infinite experience within a finite life.” 胡渊鸣 believes AI for work has largely converged, leaving limited room in another round of similar productivity tools. The next prize is the supply of spiritual enjoyment and experiences.
High-risk experiences such as wingsuit flying are not realistic for a founder, but generative AI can help someone experience “what wingsuit flying feels like.” The vision for AI for fun is not simply to kill time, but to give each person a richer set of experiences, fulfillment and value within a limited lifetime.

17. Taste determines whether AI-released time leads to works or to 《WALL·E》

吴茗 invoked 《WALL·E》 as a warning: people who no longer need to work lie in hovering chairs, comfortable but empty and sad. 胡渊鸣 believes the answer is high-quality, substantive content rather than maximizing time spent and the impulse to pay.
A game can be “a machine that makes players addicted,” driving constant in-game spending. Or, like many Nintendo works, it can “treat you like a child” and provide a purer form of happiness. Both types of content can make money, but the value they create and their spiritual consequences are entirely different.
Taste is therefore one of the creator’s most important judgments. 胡渊鸣 believes there is a sweet spot where the mass audience can accept the work and the creator can still express something. 宫崎骏, 宫本茂, 久石让, Nolan and Cameron represent the possibility that commercial scale and personal taste can coexist.
He uses 庵野秀明’s 《新世纪福音战士》 as an example. The later part of the TV version resorted to stream-of-consciousness expression because of budget constraints, including unusually long stretches of still images, and the initial response was not especially strong. The work ultimately found its audience. Truly good AI for fun should make people feel “elevated,” not merely anesthetized.

18. Science, art and fun form a creative path measured in decades

Concerts, sunrises, sunsets and natural environments led both speakers to discuss the experience of being present. Even when content can be replicated remotely, people still need to enter a place and step away from the everyday world. 胡渊鸣 is especially drawn to the vibe created by nature and has started using a camera again to train his sense of beauty.
He compresses the rest of his life into 3 keywords: “Science and technology determine what we can do; art and taste determine what we do not do; then fun and vision give us the motivation to keep doing it.” Together, the 3 set the capability boundaries, selection criteria and reasons for sustained investment in a work.
At 30, he does not want to impose a short-term deadline on the combination of art and business. Assuming he lives to 120, the remaining 90 years can be divided into 9 parts. He hopes to produce 1 good work every 10 years, with each better than the last.
Tai Chi pushed him further in science; Meshy taught him how business works; the next stage is meant to fill in art and aesthetics. The camera that had sat unused for years after he started the company has therefore come back into use. Photography is both a personal interest and taste training for the future games business.

19. A founder’s core asset is the intense desire to try again after facing failure

Behind the polished résumé, 胡渊鸣 nearly dropped out of MIT in his first year after clashing with his adviser over research interests and principles. Around 18 months into entrepreneurship, Tai Chi’s commercialization failed, the second product found no revenue, and a family elder died. He seriously considered, “Maybe I should just return the investors’ money.”
Nearly every mistake listed in a YC video matched the company at the time: a team of more than 20 people before PMF, hiring a recruiter too early and failing to speak with users. He even joked, “The aunt selling fruit downstairs seems to have PMF. Why couldn’t I find PMF after getting a PhD?” The pressure forced him to learn about market ceilings, competition, growth, talent and choosing the right track.
By the time he was building the 3rd product, he no longer dared to contact investors proactively, fearing that repeated changes of direction would be seen as evidence of incompetence. The answer on the other end of the phone was: “Sure, give it a try. I support you.” “The whole world got better” in that moment, and the team stayed with him to the present; many members have now been there for 4 or 5 years.
吴茗 summarizes building a company as a “permanent crisis.” 胡渊鸣 says he has gradually come to believe that “as long as I’m here, the company won’t have a problem. I can handle anything.” The bad events have not disappeared; the founder has simply become accustomed to the state.
Every 3–6 months, 胡渊鸣 seems to fire an older version of himself: the insecure version, the one unable to leverage the team, the people-pleaser, the excessively conservative version, and the one focused only on saving money and worrying about everything have left in turn. The current version has chosen to make peace with himself, while acknowledging that he may eventually discover that “too much reconciliation doesn’t work either.”
吴茗 gives motivation the highest weight in a founder, not a genius résumé. A consistently high-performing student who encounters negative feedback in entrepreneurship may blame the environment for mistakes. Those who can truly continue will think: “If someone gets this done, but it isn’t me, I’ll regret it.” 胡渊鸣 agrees with the tension: one must face reality pragmatically without allowing reality to become a prison. “Pessimists are often right; optimists are often successful.”
Entrepreneurship did not give him the ability to see every hole. It gave him confidence that “if I fall into a hole, I will definitely climb back out.” The worst outcome is starting again, and what he has accumulated will make rebuilding faster. That is why 胡渊鸣 says that when the two discuss AI games, they seem never to have considered the option of not doing it. He is willing to invest the rest of his life and at least build a company as successful as Nintendo.

AppLovin CEO: Why Founders Shouldn't Angel Invest & Why the Best Don't Need Mentorship AI-Era Global Marketing with 沈晨岗 of Feishu Shenno Meetsocial