卷卷: From Douyin to AI 3D and the Manufacturing OS Ambition

Summary

• 数美万物’s newly launched Twinkle 3.0 is, in 卷卷’s words, “the industry’s first commercial-grade 2048³ ultra-high-precision AI-generated general-purpose 3D foundation model,” with its core differentiation lying not only in visual quality but in manufacturability. It can generate a printable model directly from a single image, while achieving geometric-level reproduction of text and symbols that “so far, only our model in the global AI-generated 3D field can do.” Pipeline capabilities including part separation with mortise-and-tenon joints, rivet-style connectors, automatic plate layout and intelligent parameter tuning are also described as “globally unique.” Material and structural-mechanics information has been embedded in the optimization target from the training stage onward: “Our goal is to make it deployable.”
• The application landscape changed dramatically within a year, showing that different modalities will specialize. In 2025, AI 3D applications were distributed relatively evenly across gaming, 3D printing, e-commerce, film and television, architecture and education. In the first half of 2026, “3D-printing-related use cases have clearly become the No. 1 application.” Demand from games and film is indeed being better served by cross-modal video models and even Coding, but “physical manufacturing cannot get around 3D; it is very difficult to solve production problems through visual expression alone.”
• The business model is already being validated: gross margins are generally above 50%, Hi3D subscriptions cost roughly RMB140 per month, and the on-time T+7 delivery rate is 90%–95%. 卷卷 sees the on-time delivery rate for high-concurrency, nonstandard product orders as the key test of whether the model and pipeline can support the business. If the operation depends on large amounts of manual model repair, “the business logic itself becomes a paradox.” The company has raised more than $100M, with its latest round near $50M, employs more than 100 people and serves users globally.
• The market beta is substantial: desktop 3D printing has delivered more than 23% compound annual growth for several consecutive years, with global shipments of roughly 5M units last year; “shipments exceeding 10M should be a certainty.” The next growth pockets are women Makers and flexible materials—digital pattern-making for plush products and AI embroidery algorithms that could replace patternmakers who require “at least 10 years” to train—and 3D-printing farms. Super-farms with more than 30,000 machines have already appeared, but commoditization and price competition may leave creators without their fair share; 卷卷 wants to build a matching platform for one-click order allocation and creator revenue sharing. The relationship with 拓竹’s MakerWorld is “currently cooperation, not competition.”
• The moat is data and industrial trust: 3D is “probably, without question, the hardest data domain to acquire among the major modalities.” The data comes from years of building trust through industrial clusters, with support from industry associations and even relevant government departments, now supplemented by formal procurement. On competition, 卷卷 says “Chinese teams are basically leading the 3D field”; among 3D models on Hugging Face with more than 1,000 likes are Microsoft’s TRELLIS, Tencent’s Hunyuan, one model from Stability AI and 数美万物—the four leading examples.
• The methodology is rooted in Douyin: identify the industry’s key variables first, then create “activation and visibility.” Douyin bet on the 2 infrastructure variables of lower network costs and faster speeds in 2016, plus large-screen phones. This time the bet is AI × manufacturing—“China’s 2 strongest variables”: the production-factor insight created by its engineering talent dividend combined with the world’s strongest manufacturing base. “If we want to build a global business in the future, only these foundations can create a massive advantage.” In one line: move from using 15-second short videos to help people pass the time to using an AI-generated 3D foundation model to give people room to create—from universal access to self-expression to universal access to creation.
• The question of an “AI-era Douyin” carries over to content products: unless a product defines a new content format, it may simply be subsidizing existing platforms. Vertical video plus music, horizontal video plus bullet comments, and images plus tags helped create Douyin, Bilibili, early Instagram and Xiaohongshu, respectively. A product focused only on model performance must ask where users will post the content they create: “Will it be your product, or a platform that already has network effects and scale effects?”
• The founder describes himself as “riding a roller coaster in every time slot of every day,” yet he discourages most people from starting companies. “Don’t start a company unless you have to; don’t force yourself to suffer for its own sake,” unless you remain absolutely committed after hearing the negatives and are pursuing incremental value innovation that existing knowledge cannot easily dismiss. The endgame has 2 stages: first become the OS for the Maker industry, then take on the “very large proposition” of becoming the OS for manufacturing. His operating philosophy: “Don’t rush; let the bullets fly a little longer,” along with 一鸣’s line, “The problems are烦, but the mind is not.”

Deep dive

1. 数美万物 in Brief: More Than $100M Raised, In-House 3D Foundation Model Aimed at Manufacturing

• 卷卷（任利峰, 39, INFP, Taurus）describes the company as an in-house developer of an AI-generated, general-purpose 3D foundation model focused primarily on manufacturing, with limited use in games and film. Its 2 To C product lines are Hi3D, which serves Makers, and 造好物, which serves flat-media illustrators and artists by helping them take concepts into physical form and providing a transaction platform. The company serves users globally and has more than 100 employees.
• The company has raised more than $100M, including a latest round near $50M. His career path ran from joining Baidu Tieba in 2010 after graduation to managing forum operations, then to Ganji—the year it merged with 58—before moving to ByteDance to work on Toutiao, Q&A and livestreaming, and eventually taking Douyin from zero to one.

2. How Douyin Started: Infrastructure Variables First, Kuaishou and Musical.ly Second

• 卷卷’s retrospective starts with 2 signals he saw in 2015: the government work report called for nationwide broadband-fee cuts and major speed increases in 2016; and terminal makers including OPPO, vivo, Xiaomi and Huawei all planned to make large screens standard in 2016. Larger displays favored visual media while raising the bar for GPU and camera performance, reflected in vivo’s “Illuminate Your Beauty” campaign and what appears to have been a Xiaomi model developed with Leica.
• The bottom-up reference points were 2 standout products: Kuaishou’s “unvarnished record of life” in China and Musical.ly’s addition of “the fun of music” internationally—“which later became our sister product.” Taken together, those factors led management to approve 2 directions, corresponding to Huoshan Video and Douyin; 卷卷 was given the Douyin line to explore.
• Why him? “There was an element of luck,” combined with an operations background from managing Baidu Tieba forums and experience building backend products such as moderation consoles and collaboration tools. “Combining those 2 capabilities is a good fit for content products.” He stresses that the team at the time was “all nobodies,” and credits management for providing the environment in which they could grow.

3. Douyin’s Method: Users Need to Be Activated—and Seen

• Activation cannot take the form of a rigid template, which limits creativity, nor can it be completely open-ended, which leaves the product’s main proposition and playbook unclear before the product has taken shape. Douyin therefore used “semi-defined” guidance: give users a piece of music and leave them free to ignore the lyrics and theme; the music simply sets the emotional atmosphere. Stickers, props, effects and beauty filters all function as content-production inputs that prompt expression.
• Visibility requires more than a recommendation algorithm. “To a large extent, we still had to bring in human aesthetic judgment,” using operating instincts to identify tags, visual styles and themes that deserved broader distribution. The initial target was music short videos for the post-1995 and post-2000 generations; the unexpected result was that, despite the common view that “Chinese people are reserved,” Douyin showed the public another side of them.

4. The Variable This Time Is AI × 3D: Midjourney’s Trigger and Pico’s Pain Point

• The method is the same: start by asking whether the industry’s key variable exists. At the end of 2022, Midjourney was the first thing that made AI feel “truly magical” to him: “If AI-generated images can look like this, when will AI-generated 3D arrive?”
• He was working at Pico at the time and saw a clear bottleneck: VR content experiences depend on content output, which depends on 3D art assets. The process was “extremely expensive and extremely long—almost equivalent to producing a miniature 3A game.” Beginning in late 2023, he worked through the technology, collected data and assessed possible routes and use cases, concluding that AI 3D would benefit from the AIGC boom and continue to develop.

5. Manufacturing Was a Personal Through Line: From Killing Time to Creating Space

• The thread began with “hanging out in forums” during his Baidu Tieba years. In his early 20s, around age 23, he followed forums on jade and tea, later extending that interest to porcelain, jewelry, collectibles and apparel, while observing materials, equipment, processes, production lines and supply-chain coordination over the long term. His view of interest is worth noting: “When people ask what you’re interested in, you don’t need to explain it logically. What you naturally gravitate toward, what gives you a sense of beauty and aspiration—that may be your truest interest.”
• On the industry-variable side, China is the world’s largest manufacturing power with a complete industrial chain; the Pearl River Delta and Yangtze River Delta are upgrading rapidly as interconnected industrial clusters; and industrial equipment is clearly moving toward desktop form factors, which are more accessible to ordinary users.
• The cross-domain logic in one sentence: yesterday, 15-second short videos helped users pass the time; today, an AI-generated 3D foundation model gives them room to create. Three-dimensional space plus the fourth dimension of time is the world people understand—from universal expression in the digital world to universal creation in the physical world, and from satisfying the desire to express to satisfying the desire to create.

6. An Internet Operator Becomes a Factory Manager: “Respect the Laws of Physical Industry”

• The biggest gap is that internet-style rapid trial and error cannot simply be imposed on physical industry. The work starts with studying materials—materials science has a research component and an element of chance, making it difficult to accelerate by force. A factory cannot build an entire industrial chain in-house; it has to source and co-build with quality suppliers, while explaining the mission across industries well enough for partners to participate.
• The contrast in working styles is vivid. Internet workers look up at an office-zone ceiling and retreat into a “dark room” for closed development before an important launch. Manufacturing requires constant interaction and collision with people across every link in the industrial chain. The basic form of work is already radically different.

7. A Self-Built Experimental Factory in Panyu—and 2 “Soul-Searching Questions”

• The first friction point came when industrial-cluster factories failed to understand or care about AI. “Improving efficiency for a factory isn’t the point. To put it realistically, they care about whether there are orders, whether there is revenue, whether they can protect their margins and whether they can keep the factory alive.” Forced into action, the company built a small experimental factory in Guangzhou’s core Panyu cluster, covering industrial 3D printing, laser cutting and engraving, small-format CNC, nearby precision CNC turning, and a flexible-material pattern room for plush, textiles and embroidery. Materials include resin, metal, plush and textiles, with ceramics explored alongside local industrial clusters. The company later added flexible fulfillment and rapid small-batch delivery.
• The second friction point was the psychological trap of “growth.” After building user products, “you develop a desire for numbers,” but “before the industry’s variable arrives, it is difficult to force scale larger with external inputs.” That led to a chain of soul-searching questions: Can a long-chain business have a viable commercial model? If it cannot bring in orders, would the company subsidize the industry as it once subsidized the consumer side? How should ROI be calculated for investment with a research component?

8. Why He Left ByteDance: Not Dissatisfaction, but Soil and Starting Line

• In response to judgments attributed to 龙宇 and Annabel at BAI, 卷卷 reframes the issue as being “quite demanding of myself.” Douyin did not come with foresight or an easy path: every day brought enormous pressure, and “behind the pressure was responsibility.” He is also comfortable criticizing himself—when his understanding changes, he will openly say he was wrong and fix it.
• He made 2 rational assessments before leaving. Supply chains and manufacturing were materially different from ByteDance’s talent mix and organizational form, while the work was harder, more exhausting and slower to validate. ByteDance had limited exposure to 3D, while Tencent, Meta and Microsoft were “starting from the same line”; in an independent modality like 3D, “that was where an entrepreneurial opportunity existed.”
• The underlying bet was the combination of “China’s 2 strongest variables”: decades of engineering-talent dividends—the AI race is being contested through the effort of Chinese and Chinese-American talent in Silicon Valley and China—and the production-factor knowledge accumulated by the world’s strongest manufacturing base. “If we want to build a global business in the future, only these foundations can create a massive advantage.”
• The first thing he did after deciding to start a company was get a physical examination. “Could anything in the future make it impossible to run the business? One key factor was whether my body could withstand it.” The results were good; with his health as a backstop, he felt able to commit to the full startup cycle.

9. Twinkle 3.0: The Industry’s First Commercial-Grade 2048³, Built Around Reconstruction Consistency

• Version 3.0, launched the day before the recording, is positioned as “the industry’s first commercial-grade 2048³ ultra-high-precision AI-generated general-purpose 3D foundation model.” Its defining feature is reconstruction consistency. 3D has no equivalent of sketching over a reference image: “You input one image and have to get a model,” making the task much harder than image generation.
• Pet models illustrate the point. More than 150M Chinese households own pets, and customers in Europe, Japan and Taiwan turn them into figurines and memorial objects. The expression must be recognizable at a glance to the owner; animal fur previously had to be brushed in one stroke at a time with an electronic pen. A bronze bullfighter sculpture demonstrates pose logic: the bull’s forelegs occlude the human body, so the model must understand the confrontation and infer the unseen back muscles correctly. The resulting white model can be produced by SLA printing.

10. Globally Unique: Geometric-Level Reconstruction of Text and Symbols

• 卷卷’s positioning of the industry is that precise, controllable text in image models only began making obvious progress in the second half of last year. Version 3.0 reconstructs text and symbols with high fidelity at the geometric level: the English letters, numbers and complex symbols on a calculator, and the individual names of masks on a mask-storage box, are all clearly visible. “This isn’t a texture map; it remains completely consistent with the original input.”
• The most compelling demo was 《天净沙·秋思》. “For the first time, you could clearly read the text at the geometric level.” Differences in carving depth reproduce the heavy and light strokes of stone and jade carving, bringing the scene of withered vines, old trees, dusk crows, a small bridge, flowing water and homes into physical form.

11. From 关之琳 to Black Myth: Wukong: Faces and IP Nearing Commercial Readiness

• The portrait examples are extensive. A Wolverine head sculpture may already require no changes in its white-model state. The input image of 关之琳 contained lighting interference; handled poorly, it would produce “a dark, muddy structure fused with the hair” around the head. Version 3.0 can process the shadows and infer plausible braids on the unseen back of the head. Black Myth: Wukong captures “the bearing of the Victorious Fighting Buddha”; a Buddhist amulet has a benevolent face; other examples include a six-armed Guanyin.
• The commercial signal is the company’s own IP, “雷拉,” which sold more than RMB100,000 at Shanghai’s QDF exhibition over 3 days. An IP designer said the generated output had “already surpassed many of the trendy-toy designers he had encountered in the market”; after some further adjustments, “it should reach commercial grade.”
• When Nano Banana first appeared a year ago, a user asked whether an outdoor sports photo could be turned into a desktop figurine. 卷卷 replied that existing technology could not produce a hyper-realistic character figurine with ultra-high precision. “In less than a year, that capability has now been achieved,” and industrial SLA equipment can print the result.

12. From Training Targets to Post-Processing: Everything Serves “Print-Ready Output”

• Version 3.0’s improvement came from 3 sources: an unprecedented investment in high-quality data, built on long-term accumulation, cleaning and labeling; a new technical route for 3D representation; and a post-training pipeline, including the ability to understand and process lighting in the input image.
• Its optimization target differs from competitors’: “Our goal is to make it deployable.” Data construction and parameter tuning incorporate geometric precision, structural understanding, materials used in actual production and even structural-mechanics information. After the base model, numerous post-processing models target specific devices and processes.
• Hi3D’s engineering stack includes AI semantic understanding for part separation; adding posts and spherical mortise-and-tenon joints to separated parts; rivet-style connectors—“adding connectors may currently be a capability that only our platform can provide anywhere in the world”—created by combining AI models with traditional geometry pipelines; automatic plate layout, where tray dimensions and support orientation affect print time and tolerances; and intelligent parameter tuning. “If the model comes out looking like fried noodles, it’s neither a model problem nor an equipment problem—it’s a parameter problem.” All of this is packaged into a 3MF file that can be imported directly into slicing software.
• The positioning in one line: “a one-stop AI 3D pipeline service platform, from acquiring a model to final manufacturable output,” spanning 3D printing, laser, CNC and eventually flexible textiles. The ambition is to “become the OS for Makers.”

13. Video Models Are Diverting Game and Film Demand—But Physical Manufacturing Cannot Get Around 3D

• Koji asked the sharp question: with video models such as C Dan 2.5 and MiniMax H3 so powerful, could games and film stop needing 3D models altogether? 卷卷 did not evade it: “It is possible, and it is already happening.”
• The mix has changed. In 2025, AI 3D applications were broadly and evenly distributed across gaming, 3D printing, e-commerce, film and television, architecture and education. In the first half of 2026, “3D-printing-related use cases have clearly become the No. 1 application,” while gaming and film fell sharply as a share. Demand has not disappeared or necessarily declined; it is being “better satisfied to some extent by cross-modal video models, and even by directions such as Coding.”
• The bottom line is clear: “In physical manufacturing, 3D remains unavoidable. It is very difficult to solve production problems with a visual representation alone.” AI 3D will therefore increasingly point toward applications that can produce physical output.

14. Data Is the Deepest Moat; Chinese Teams Are Leading the Field

• 3D is “probably, without question, the hardest data domain to acquire among the major modalities.” The data is scattered across game companies, manufacturing pattern rooms and model websites. His answer grew out of his personal-interest thread: industrial-cluster owners did not treat him as a pure internet operator or suspect that the data would be used to replicate or copy their products. Early on, he secured support from industrial clusters, industry associations and relevant government departments; formal data-service procurement has since been added. That barrier is equally difficult for later entrants.
• The competitive picture is worth recording: “Other modalities involve a contest and climb between China and the US, but the 3D field is basically being led by Chinese teams. That is something worth paying close attention to and taking pride in.” The differentiation lies in application focus. 数美万物 goes deep on high-fidelity output for manufacturing, high-poly game assets and film CG effects, while competitors are “extremely good” at low-poly models, topology, rigging and skeletal animation. Each will deepen within its own vertical.

15. The Next Growth Pocket: Women Makers and Flexible Materials

• The existing user base for 3D printers, lasers and CNC machines skews male and toward hard materials. The next growth factor is “the rapid growth and potential future breakout of the women Maker market”: mothers knitting sweaters, wives doing cross-stitch, and younger women taking up crochet, embroidery, felt work and Tufting. During a visit to Shenzhen, he saw a meaningful number of hardware-oriented teams shrinking flexible-material equipment and starting companies around it.
• 数美万物 is pursuing 2 software directions: digital pattern-making algorithms for plush products and AI embroidery algorithms. The deeper motivation is more than efficiency. Training a specialized patternmaker takes “at least 10 years,” and such talent is extremely scarce. As AIGC lowers the barrier to creativity, personalized demand will surge; algorithms must be combined with production-line processes to “solve the bottleneck in creative capacity.”

16. How 3D Solves Plush Toys: The Patternmaker’s Brain Is a Model

• The apparel analogy makes the process clear: concept design → patternmaking → mass production. A patternmaker works from mannequins representing Asian or European body types and imagines how flat clothing should be cut into fabric pieces to form a 3D garment. “There is a model in the patternmaker’s brain.” The pattern is drawn, fixed and repeatedly adjusted before the parameters are locked in. Plush is harder because the character is invented from nothing—“this little monster doesn’t exist in this world”—so there is no existing mannequin.
• The 3D solution has 4 steps. First, use the company’s image-to-3D model to create a 3D “mannequin” for the plush toy. Next, identify seam lines on the 3D surface and split out the different patterns. During “surface unfolding,” account for the elasticity coefficient: simulate manual recutting in high-tension areas to release the tension and produce flat patterns. Finally, send the full sheet of fabric into a CO₂ laser cutter to cut the pieces, then sew and fill them into a prototype.
• Everything ultimately comes down to format mapping. 3D printing translates a parameterized CAD model into formats such as STL and OBJ. Embroidery translates any image into an EMB vector file—with stitch-path planning and color information—and then into DST for the embroidery machine. The input is an image; the output is a file format the equipment can read.

17. Submission Rate Above 20%: Turning 《神笔马良》 into Reality

• 造好物’s submission rate—the number of daily users submitting works divided by DAU—is above 20%, “an extremely high standard” for a content product. The UGC logic is straightforward: a low enough tool barrier and good enough output activate the desire to express. Physical products add a distinctive positive-feedback loop: “You can actually receive a physical object identical to the model. The moment users receive it, they want to submit again because they realize this can really be done.”
• The biggest use cases are figurines—objects themed around games, family and friends, or original characters that appeal to young people and even adults—and pets, where emotional attachment is transformed into a toy or display piece for interaction or gifting. The positioning in one line: “造好物 is turning the fairy tale 《神笔马良》 I read as a child into reality.”

18. The Business Model: 50%+ Gross Margins, RMB140 Subscriptions and T+7 as the Direct-Output Test

• Pricing is generally set at gross margins above 50%. China is priced for broad access, while overseas margins are higher because of currency differences and “relatively weaker supply-side capabilities.” Hi3D uses a subscription model, priced globally at roughly RMB130–140 per month, for Makers who cannot model and do not understand hardware operation. Usage among owners of desktop equipment such as 拓竹 continues to rise; “many of these capabilities exist only on our platform.”
• The key production-grade test for direct output is the “on-time delivery rate for high-concurrency, nonstandard product orders”: high volume, a different design on every order and a uniform promise to hand shipments to logistics within T+7. The current on-time rate is 90%–95%. The counterfactual is sharp: traditional pipelines can use manual modeling to guarantee only one side of the equation. “If the model cannot get close to direct output, the operation carries huge manual model-repair costs, making on-time delivery for high-concurrency, personalized, long-tail products impossible. The business logic itself becomes a paradox.” Fulfillment combines a small amount of in-house production with outsourcing.
• The other test is the offline laboratory. The company buys the major hardware on the market, hires professional modelers to form an evaluation panel and tests every day for non-manifold geometry, holes, lack of watertightness, format-recognition errors, print failures caused by overhangs, unreasonable multicolor mapping and color bleeding. “We are genuinely trying to establish an industry standard” for evaluating model deployability and consistency.

19. The Business Evolved with the Market: Self-Built Models Were Forced, and Hi3D Followed 1,000+ Hugging Face Likes

• The company did not start by building a sprawling platform. Its first product was 造好物. During development, the team discovered that “there was no usable model on the market that could be combined with manufacturing,” forcing it to invest in an in-house model. The To C product was not launched until nearly a year into the startup.
• In June 2025, the company released its 1.0 model on Hugging Face for public evaluation and received more than 1,000 likes. The 3D category models to reach that level included Microsoft’s TRELLIS, Tencent’s Hunyuan, one model from Stability AI and the company’s own—a total of 4. “After seeing the signal, we followed the market and turned the model into a platform to serve everyone,” which led to Hi3D. The businesses extended along one continuous line rather than being fully conceived from the outset.

20. The 3D-Printing Market: Desktop Machines Are Now a Category Leader, and MakerWorld Is a Partner

• The data chain is complete. In the 5 years before 2024, desktop 3D printing sustained annual compound growth above 23%. Customs export data shows 3D-printing equipment ranking among the Top 20 fastest-growing categories for several consecutive years. In 2025, desktop 3D printers became a leading product within the printing category on Chinese e-commerce platforms. Global shipments were roughly 5M units last year, with compound growth above 20%. “Annual shipments exceeding 10M should be a certainty.”
• The relationship with 拓竹’s MakerWorld is “currently cooperation, not competition”; the company’s models are already connected to the platform through an interface. The difference is the broader Maker OS positioning: multiple processes, connection to industrial-grade equipment and an upcoming SLA service, while flexible-material capabilities are not yet included. Liquid trapped in sealed cavities can sharply increase print-failure rates; the model-powered service will determine “where to place the hole and how deep to make it.”

21. Demand Meets the Moment: Deindustrialization, the Pandemic and Social Media

• The macro framework is decades of global deindustrialization compounded by recent trade frictions. Overseas users face a mismatch between abundant demand and goods that are either undersupplied or expensive, making them “more willing to buy small devices with industrial-like capabilities and DIY to meet their needs.” Pandemic-era disruptions to goods circulation pushed the need for miniaturized equipment to “an unprecedented level.” “Many people aren’t creating just for the sake of creation; they genuinely need to solve problems in daily life.”
• Social media acts as an amplifier. Content showing “I can make a figurine or a plush toy myself” attracts attention and imitation, while hobbyists can become creators who earn a living from the activity. UV printers—including Anker’s desktop UV printer, which has sold well—xTool lasers, desktop 5-axis CNC machines, knitting machines, embroidery machines and other devices are all moving toward smaller form factors. “Desktop 3D printing is simply the first large market that has become clear today; the broader Maker market will continue to break out.”

22. 3D-Printing Farms: Commoditization Drives Price Competition, and Creators May Not Profit

• Small farms have 300–500 machines; large ones have more than 10,000, with super-farms exceeding 30,000. Most target overseas markets. The problem is commoditization: once you see one articulated dragon, “you’ll soon see it everywhere.” After the squeeze-yoyo boom, “where is the next hit?” is a question every 3D-printing farm faces. The result is a race to the bottom on price, while original creators may receive nothing—or far less than the value they created—after their models are downloaded from model sites and mass-produced.
• 数美万物 plans to connect Hi3D with farms for one-click ordering and allocation. It would take a portion of the revenue from both model downloads and product purchases and distribute it to creators. “Only by creating a positive guarantee for creator income will more people be willing to share models and make it possible to unlock more ideas.”

23. Not 3D ComfyUI, but a Pipeline Platform with Atomized Capabilities

• 卷卷 corrects the framing: “We are not trying to build a ComfyUI.” The company is borrowing ComfyUI’s mind-map-like modular composition and infinite-canvas format to atomize capabilities such as modeling, part separation and relief generation, allowing users to combine them freely according to intent. A natural-language description can generate an entire pipeline design—for example, a miniature family scene → upload or generate an image → preview → visually separate multiple elements → model each element → assemble them.
• A pipeline is needed rather than one-click output because 3D “isn’t just a question of whether something can be printed.” It is an integrated chain that combines modalities and connects them to production equipment; “it is not a single-point capability.” It also gives To B customers a base on which to build private deployments.

24. Which Douyin Lessons Transfer: Content and Growth, Yes; Manufacturing Know-How, No

• Transferable capabilities include understanding content, building a creator ecosystem and growth methods. What does not transfer is the understanding of how to combine models with manufacturing pipelines: “These are all independent pieces of know-how that require fresh learning and practice.”
• The truth behind the folklore of “talking with creators until dawn” was jet-lagged scheduling with US creators, focused on requirements: what capabilities a segmented-shooting tool needed and how precise the timestamps had to be for audio-video synchronization. “Or put another way, most of our time was spent getting yelled at.” The team welcomes users who can offer valuable criticism because they know where the problems are; it listens patiently, improves the product and responds efficiently.

25. The INFP Tension: Feeling Starts the Project, Logic Validates It

• He describes himself as a P-type person driven by feeling. “The decision to do something often comes in a flash of inspiration,” but another voice asks, “Are you just entertaining yourself? Come back and break down, derive and validate the problem.” His operations-trained intuition has to be translated by his product-manager side into code and interaction—a process of “pulling it back and turning it into a T, turning it into a J.”
• The best example of being proven wrong came early at Douyin, when he used personal preferences to constrain the content style. After the change performed poorly, he publicly criticized himself at a bimonthly meeting: “My judgment was wrong. I had too narrow an understanding of user needs and used my personal ideas to control how users expressed themselves.” Once the platform opened up content and respected diversity, submissions, retention and user scale all rose. “The more user products you build, the humbler you become and the more you respect everyone’s differences. You increasingly want to enable others rather than yourself.”

26. The Roots of Humility, Sincerity and Stubbornness

• His humility comes from building large-scale user portraits during his Baidu Tieba years and seeing “the full range of humanity.” He cites 3 lines: “Learning from the virtuous makes one a junzi; learning from the sages makes one worthy; learning from everyone makes one a sage.” The sage learns from the masses.
• He considers himself “extremely, extremely lucky” to have experienced the PC internet, mobile internet and AI eras—“an indescribable kind of arrangement by fate.” But the method can be reduced to 2 words: sincerity—purity and focus on the problem and need themselves—and selflessness: doing things that benefit others without major personal gain, the Buddhist idea of acting without self-interest, where a bodhisattva “does not only deliver himself, but others as well.”
• The other side is stubbornness: once he has made up his mind, he does not care about short-term external changes. When Douyin’s early retention was poor, the team made many “counterintuitive” cuts: it did not scrape and repurpose content from elsewhere, and it did not show material it considered harmful to young users. The data looked bad in the short term, but the team held the line. “Douyin’s slogan is ‘Record Beautiful Life,’ and I think we achieved that.”

27. The Latest Wake-Up Call: Delivery Speed Matters More Than Perfect Quality

• Koji identified the tension between stubbornness and timely self-correction. 卷卷’s example was early product delivery: he obsessed endlessly over quality, promised 10-day delivery and sometimes took more than a month. “Users no longer cared about your repeated quality revisions or your pursuit. They only cared that you failed to fulfill the promise. The experience was terrible, and they would not give you another chance to experiment.”
• The correction was to establish delivery-time and on-time-rate standards to guide all work. “Quality is one clearly defined dimension; it is not an endless pursuit of something supposedly better.” After a colleague raised the issue, he admitted, “My thinking was immature and simplistic.” That led to an anti-one-person-rule requirement for the team: do not assume past success will transfer to a new field. If people do not offer objections, he assumes he is right, increasing the cost and duration of trial and error.

28. Will There Be an AI-Era Douyin? Define a New Content Format First

• “That is difficult to say,” but one point is certain: anyone trying to open a new content category “must be able to define a new content format.” Vertical video plus music is Douyin; horizontal video plus bullet comments is Bilibili; images plus tags were early Instagram and Xiaohongshu. New combinations must be found within the format paradigm.
• The warning is blunt: “Simply building an AI-era Douyin will most likely mean making a wedding dress for these kinds of products.” A product that focuses only on model performance without format innovation must ask where users will post the content they create with a good model: their product, or a platform that already has network effects and scale?

29. The Startup Mindset: Several Phases, a Roller Coaster Every Day

• A year ago, he wrote on Xiaohongshu’s Red Land that he had expected it to be difficult, “but it is harder than expected.” His current framework has several phases: early optimism while building the vision and before encountering real problems; a period of self-reproach when “it seems like nothing can be done”; a recovery when “I can do a little bit of something well”; and the present, with problems interlocking. “Starting a company is like riding a roller coaster every day—and by every day, I mean every time slot of every day.”
• How does he stay steady? “Detach your emotions and focus more on the problem itself.” He often tells himself, “Don’t rush; let the bullets fly a little longer.” Anxiety distorts actions, sends the wrong signal to the team, pushes everyone into panic and can even make them “do reckless things in other dimensions.”
• The lessons from his practice include 一鸣’s line, “The problems are烦, but the mind is not,” which he has contemplated for years, and the realization that building a good product and running a good company are 2 different things. A company must manage people, money and operations: how talent is organized to collaborate, where the money comes from, and whether the business logic can be falsified. He recalls the mindset with which his benefactor 张楠 repeatedly faced the same questions, and says the experience has left him increasingly in awe of and respectful toward excellent managers and today’s entrepreneurs.

30. Discouraging Entrepreneurship—and “The Last Person Against the Wall”

• His advice to friends seeking startup guidance is deliberately negative: “Don’t start a company unless you have to; don’t force yourself to suffer for its own sake.” The exception is someone who remains absolutely committed after hearing the negatives, has a logic they cannot refute with their existing knowledge and is pursuing incremental-value innovation rather than linear growth.
• The essential difference from running a large team at ByteDance is that he once had partners and management sharing the burden. Now, “I should be the last person standing against the wall in the company. Beyond my wall is the abyss; I am the backstop behind the backstop.” He copes with solitude through “acting without self-interest”: “Thinking more won’t help. Much of what you think about is related to yourself and only adds to your worries.” If his self from 2 years ago came to ask for advice, he would “explain the difficulties more fully,” but would still start the company. “Entrepreneurship is only one means.”

31. The Endgame: From Maker OS to Manufacturing OS

• He has no numerical target at which he can stop. The aspirational goal has 2 stages: first become the OS for the Maker industry, a pipeline platform offering integrated services across categories, processes and equipment; then tackle the much larger question of “whether we can become the OS for manufacturing.” That would mean connecting models to software pipelines and linking industries to solve foundational infrastructure problems across a much broader range of physical manufacturing.
• His personal satisfaction comes from enabling others. Seeing users “gain income, recognition and status” through the product makes him very happy. He cites the “overview effect” from an astronaut documentary: the births and deaths of billions of people compressed into a single point—a blue planet—creating a sense of shared destiny and universal love. “Life is only a few decades. When you eventually leave this world, what can you take with you? You should think more about how many people your existence has enabled, and the intangible assets you leave behind.”

32. “I’m an Insecure Person”: Action Is the Best Medicine for Anxiety

• His deliberately provocative self-description is: “Actually, I’m an insecure person.” He was a novice on Baidu Tieba but had to manage forum moderators; Douyin was his “first product in the true sense”; he studied computer science but is not a specialist in 3D algorithms; and he understands even less about manufacturing supply chains. That fundamental lack of confidence forced him to learn quickly. He asked frontend and server-side engineers what an excellent product manager looked like, wrote his own early data labels and pulled the data himself before analyzing it because he feared his requirements might not be sound. Today, he needs to learn enough to participate in technical selection and resource allocation and ask whether the company can produce a leading model generation.
• His definition of anxiety and its solution: “You set a very high goal, there is a huge gap between your ability and the goal, and you do nothing to close it. That is when people become anxious. My approach is to use learning and practice to narrow the gap.” This is Wang Yangming’s unity of knowledge and action: “If you keep thinking without moving or practicing, you will never solve the problem.”
• His only advice to young people just out of school is: “Don’t overthink it; do the work in front of you well.” “Opportunities are often not found but awaited.” An opportunity may emerge from sufficiently deep accumulation in an existing role. Douyin was also something he waited into: 张楠 told him earnestly, “Once you take on this product, you must do it well. You cannot bring shame on yourself.”

33. It Is Harder to Be a Good Person: Negative Management, the Superego and Imagined Feedback from 2 Former Leaders

• His management style has changed: “In running and managing the company today, I may be becoming less nice and finding it harder to be a good person.” The company took shareholders’ money and assembled excellent people; “making management decisions based on personal emotion is extremely irresponsible.” Leadership has 2 elements. First, understand yourself and be willing to lead the team in repeatedly overturning your own past mistakes. Second, do not practice only positive management; calmly practice negative management by identifying poor role fits and making adjustments. “In a position, do the work of that position”—make decisions based on the role’s required superego rather than the individual’s id. Koji added the “new CEO perspective” from a book by an Intel CEO: clear away sunk costs, and the right decision is already in front of you.
• If 一鸣 were still running weekly meetings, the most likely question would be: “How large is the variable in what you are doing?” Either efficiency must change qualitatively or costs must fall sharply. “Don’t always try to make small optimizations. Don’t entertain yourself; don’t assume you are right every day.” 张楠 would help balance user-need understanding with growth initiatives and would “very directly” criticize him for being “overly optimistic and immersed in his own perspective.” “I still make similar mistakes today; I just try to reduce them.”
• His anti-self-delusion mechanism as CEO and product manager is to begin every weekly meeting with user feedback. Problems in the Feishu feedback group “almost never go unanswered for more than 5 minutes.” After seeing a complaint, he moves into internal discussion; if the company is at fault, it reassures the user, apologizes, pays or compensates them, and adjusts the result and service. He uses his own products and products across the industry in depth, and asks industry KOLs to evaluate them in the offline laboratory. “If they say it is right, it is right; if it is not good, it is not good; if it is wrong, it is wrong.” He has recently added another role: “transforming into a good salesperson, marketer and evangelist.”