Content Industry's Three Deaths and Creators' Rebirth in the AI Era
Summary
庄明浩 frames AI’s impact on the content industry as the “death of raw material, the death of workflow and the death of copyright”—a simultaneous reconstruction at the technology, product and industry, and capital layers. The shift was already under way before ChatGPT launched on November 30, 2022: Sequoia had mapped the generative-AI roadmap across text, code, image, speech, video, 3D and gaming on September 19, 2022, while Stable Diffusion 1.5 had already opened the Pandora’s box of image generation. The core point is not that a particular application suddenly took off, but that the basic unit, organizational model and ownership structure of content production are failing at the same time.
Multimodal models are turning expensive, static, manually built content assets into instantly generated, interactive environments. From GPT-4o, Nano Banana and Image Two to Kling 3 and Seedance 2.0, the contest has moved beyond whether models can generate at all to subject consistency, native 4K, audio-video synchronization, text rendering, long-video stability and dynamic generation; world models push the frontier into rendering, physics simulation and planning. If a coffee cup in a 3D game no longer needs hours of modeling but can be generated, rotated and pushed at will, 庄明浩’s summary is: “Material things are no longer needed; the conceptual is enough”（“唯物的事情不需要了，唯心就可以了”）.
The short-drama market, with AI shorts now a major participant, shows how workflow reconstruction can create a new and brutally competitive market larger than film in less than a year. The figures cited in the talk put last year’s Chinese film box office at more than RMB50B, or roughly RMB140M a day, versus RMB350M a day for short dramas, potentially rising to RMB400M; live-action shorts account for about RMB200M and AI shorts about RMB150M. Hongguo has also captured nearly 5% of total Chinese mobile-internet usage time. It took only about 9–10 months to move from memes and PowerPoint-style motion comics to animated comics, photorealistic comics and AI/live-action hybrids, replacing the old film-and-TV process with a real-time loop of “make the asset, test paid distribution, revise based on conversion.”
ByteDance is both critical infrastructure for the AI short-drama boom and a distribution bottleneck that creators struggle to escape. Model companies buy Seedance from ByteDance, short-drama studios use the tools to produce content, then publish it back to Douyin and Hongguo while continuing to buy traffic from ByteDance; final income depends on the platform’s revenue split, which 庄明浩 says may average only 3%–8%. The loop gives Seedance vast, second-by-second conversion feedback that is difficult for rivals to replicate, but it also pushes supply rapidly to the limit: “Eleven short dramas made,” potentially yielding just “RMB9.60.”
The “death of copyright” does not mean copyright disappears; it means the boundaries and prices of training, distribution, real-time generation and voice-personality rights must be reset. Leading model companies have expanded from buying training rights to publication, live-streaming and real-time distribution rights. Music exposes the problem most directly: a deceased singer’s recordings may belong to a company, but whether voice rights belong to the family—and how to price them—remains unresolved. AI music uploads are expected to exceed 50% soon, while actual plays account for only 1%–3%; capital has nevertheless pushed ElevenLabs to an $11B valuation and invested $400M in Suno in June 2026 at a $5.4B valuation.
As models begin helping build the next generation of models, the crisis for content creators is shifting from fear of unemployment to doubt about their own taste and creativity. Anthropic’s When AI Builds Itself uses an animation in which models are repeatedly fed back into the training process until “humans disappear.” 庄明浩 also said Fable 5 was banned by the US government three days after launch, marking the first time the capability curve retreated because it was “too dangerous,” while leaving open the possibility that it “might be released again in a couple of days.” One leading AI creator put the feeling more bluntly: “I became the person running behind, gasping for breath, with the finish line moving farther away,” only to realize, “Maybe I really wasn’t that interesting after all.”
The creator’s rebirth does not come from beating AI again, but from personally supplying the entirely new human vision that technology, distribution and capital cannot generate automatically. After falling 0:3 against AlphaGo, 李世石 used desperate intuition to play move 78, winning one game at odds estimated at “one in 10,000”; humans have never beaten AI at Go again. That led him to ask whether it can still be called human progress when even the strongest players are copying AI. 庄明浩 ultimately places the residual value of humans in narrative, aesthetics, trust and intuition, closing with a friend’s verdict: “How could AI be as interesting as you?”（“AI 哪有你有趣啊”）.
Deep dive
1. Generative AI Had Been Defined as an Industry Well Before ChatGPT
“Where are the artists we were promised?” was originally a complaint from podcaster 仲卿 about singer 唐汉霄 spending years as the vendor-side contractor making film background music for clients. 庄明浩 brought the line into the AI context to ask where content creators fit after models, clients and platforms have all reorganized the production process.
Sequoia published Generative AI: Creating a New World on September 19, 2022, more than two months before ChatGPT, built on GPT-3.5, launched on November 30. Stable Diffusion 1.5 had also made open-source image models usable by that point. For 庄明浩, the starting point of this cycle is therefore generative AI itself, not merely the large models that later became its focus.
Sequoia’s framework already covered text, code, image, speech, video, 3D and gaming, and the industry’s evolution over the past several years has broadly followed those modalities. 庄明浩 uses that progression to split the impact into the technology-level “death of raw material,” the product-and-industry-level “death of workflow,” and the capital-level “death of copyright,” before turning to how creators might be reborn.
2. Multimodal Competition Has Moved Beyond Whether Models Can Generate
影视飓风 once used an extremely short prompt to generate a cyberpunk train sequence. The flaws were still visible in isolation, but when dozens of similar clips were tiled across the screen, the comments shifted toward a sense of unease. In follow-up tests, even visual-effects specialists struggled to decide which footage should be made with effects and which should be shot with real people.
庄明浩 uses the evolution of image models to describe humanity as “losing every battle”: AI-generated images were already winning competitions in 2022; GPT-4o triggered a Ghibli-style wave in March 2025; Nano Banana generated an image of Miyamoto Musashi imitating a manga artist’s work in August, with no one on Xiaohongshu questioning its authenticity; and by this April, Image Two could turn a single sentence and an EVA image into a game screenshot.
Kling is working toward unified multimodal input and output, storyboard-level audio-video synchronization, subject consistency, native 4K and team collaboration. Seedance 2.0 is targeting physical realism, long-video stability, text rendering, multilingual and cultural adaptation, prompt precision and dynamic generation. The competitive question is no longer whether a single frame is correct, but whether an entire work can remain coherent and controllable.
China’s market includes giants and startups such as ByteDance, Kling, Alibaba, Kunlun and Hailuo. 庄明浩 believes all of the top 20 could be strong; in the US, Google DeepMind is the main remaining player still pushing ahead. His view is that the market is large enough for later entrants to capture a share even if the leaders establish a clear advantage.
3. World Models Are Trying to Eliminate Static Assets, Not Just Render Better Images
In a dynamic-generation demo from Seedance 2.0, a man initially stood outside a vegetable market. When the user instructed him to walk toward the potato stall, the scene continued generating as he moved. Content shifted from being completed in advance and played back to evolving in real time in response to commands, pushing video models toward world models.
Citing 李飞飞’s taxonomy, 庄明浩 says products currently grouped under the world-model label fall into at least 3 categories. A renderer makes the world visible and operational; simulation models physical rules such as “a round object rolls away, while a square one moves only once”; and planning explains why embodied-intelligence and autonomous-driving companies are also building world models.
If all 3 layers mature, a coffee cup in a 3D game will no longer need to be modeled face by face over several hours and then separately assigned rules for rotation and movement. That is the meaning of the “death of raw material”: objects no longer exist in advance but are generated on demand together with their state and rules. “Material things are no longer needed; the conceptual is enough”（“唯物的事情不需要了，唯心就可以了”）.
4. AI Short Dramas Overturned the Film-and-TV Production Cycle in 10 Months
Hongguo’s App Store screenshots focus almost entirely on view counts, clicks and heat scores; titles on its home page can show 2B views. 庄明浩 uses this to stress that short dramas are not a scaled-down version of traditional film and television, but a business in which real-time data organizes both production and distribution.
The industry began scaling around 2022 with Wushuang-style paid live-action short dramas, shifted to free live-action short dramas in 2024, and moved into AI animated comics in 2025 because video models were not yet suited to live action. Once Seedance 2.0 and Kling 3 appeared this year, AI dramas with a live-action feel began multiplying within months.
In only about 9–10 months, the format moved from static memes and gag comics to PowerPoint-style motion comics, then to genuinely animated comics, photorealistic comics and finally AI/live-action hybrids. Each improvement in model capability rewrote the production methods of the previous stage.
The figures cited by 庄明浩 put last year’s film box office at more than RMB50B, or roughly RMB140M per day, versus RMB350M per day for short dramas, potentially rising to RMB400M. Live-action shorts account for about RMB200M and AI shorts about RMB150M. Based on 1B daily active users spending 10 hours a day on their phones, Douyin and WeChat each capture nearly 20% of usage time, while Hongguo has reached nearly 5%. The numbers speak for themselves.
5. The Platform Loop Builds Model Moats While Driving Creator Economics to the Floor
Jiuzhou expanded from roughly 1,500 employees in October 2025 to 4,000 eight months later, while cutting half of its live-action production capacity and redirecting resources to AI creation and compute. More than 75% of its employees joined within the past 6 months, and new roles such as “card-pullers” emerged.
The workflow at leading companies now looks like this: once an asset is finished, a media buyer throws it into Douyin’s “time furnace” to test conversion; strong results trigger more spending, while weak results lead to immediate revisions. The cycle repeats every day. 庄明浩’s “death of workflow” does not mean efficiency tools entering the old process; it means the traditional film-and-TV production rhythm has been reordered by test data.
That echoes 纳德拉’s view at Microsoft Build that future B2B organizations may contain only full-stack generalists, people responsible for infra and model tuning, and frontend deployment engineers: “There is no middle state anymore.” Short dramas are simply the industry moving fastest and exposing this organizational model first.
张月光 says the commercial loop sits almost entirely inside the ByteDance ecosystem: tool companies buy Seedance at high prices, then buy traffic to acquire customers; clients produce short dramas, publish them to Douyin and Hongguo, and continue buying ByteDance traffic, while payback depends on the platform revenue split, which may average only 3%–8%. The huge volume of content and second-by-second conversion data gives Seedance feedback that other models struggle to replicate, but one creator uploaded 11 short dramas and made only “RMB9.60”—a reminder that a booming market does not mean broad profitability.
6. Copyright Deals Are Expanding from Training Licenses to Real-Time Generation
GPT-4o’s Ghibli-style images spread at scale without any copyright payment to 宫崎骏. 庄明浩 specifically clarifies that 宫崎骏 did not comment on the latest GPT-4o wave; in a 2021 documentary, however, he saw early AI images and criticized their creators for “having no idea what human suffering actually is.”
Books, television, films and video games all use copyright as a core part of their business models, so the problem did not begin with generative video. Faced with an issue that is fundamentally about money, leading model companies have chosen to “solve it with money,” expanding purchases from initial training rights to publication rights in 2023 and 2024, then to live-streaming, real-time distribution and ancillary rights around generated content.
The assets being purchased are led by news text, followed by music, audio and voice, then images, video, Wiki, books and Reddit. Publishers and media companies are therefore caught in a practical dilemma: models may undermine their existing businesses, yet are willing to pay high prices for their data. “If it were you, would you sell?”
Music exposes the boundary problem most completely. Someone on Spotify used a deceased singer’s voice to release an AI song; the original company owns the copyright to the old recordings, but there is no obvious answer as to whether voice rights belong to the company, the family or another party. AI songs are expected to account for more than 50% of uploads soon, while their share of plays is only 1%–3%—supply has exploded, but listeners still mostly choose old songs.
7. Capital Is Pricing the New Rights Structure in Advance
ElevenLabs’ valuation has climbed to $11B over the past several years, making it the leading AI voice company. Suno raised $400M in June 2026 at a $5.4B valuation, after reporting $300M in ARR in February. Copyright uncertainty has not stopped capital from treating voice and music generation as independent, multibillion-dollar markets.
庄明浩’s view is that this cannot be stopped: “If iQIYI doesn’t do it today, another company will.” The statement conveys high conviction in the diffusion of the technology without promising that most participants will make money.
Copyright is called “dead” not because paid rights will disappear, but because training, distribution, live-streaming, real-time generation and voice rights are all forcing a reset of the boundaries and allocation mechanisms around copyright.
8. As Models Approach Self-Evolution, the People Who Understand AI Best Lose Their Sense of Value First
In When AI Builds Itself, Anthropic lays out a recursive path: humans first train a model, then put the model back into the training process to help build the next generation, repeating the cycle until the diagram ends with “humans disappear.” 庄明浩’s reaction to the animation on the company’s website was that it was “terrifying.”
He also said that Fable 5, launched by Claude, was “absurdly powerful” and was banned by the US government 3 days after launch because it was dangerous, marking the first backward move in the AI capability curve. But he does not view that as a stable inflection point: “We don’t know how long this rollback can last. It might be released again in a couple of days.”
AI blogger 卡兹克 wrote that the “bottleneck used to be on the AI side,” allowing him to extract clear value by wringing the model dry. Now, even after pushing himself to the limit, he has become “the person running behind, gasping for breath, with the finish line moving farther away.” The most painful conclusion is not that his job will be replaced, but: “Maybe I really wasn’t that interesting after all.”
9. The Creator’s Rebirth Is Not About Defeating AI, but Defining the Human Vision AI Cannot Generate Automatically
In 2016, AlphaGo beat 李世石 4:1. The first game reflected underestimation; after the machine played an incomprehensible “divine move” in the second, underestimation gave way to doubt; the rapid collapse in the third left him with what felt like a shattered faith. In the fourth game, move 78 came from desperation, intuition and impulse, with the winning move later estimated at odds of only 1 in 10,000. It delivered humanity its only victory.
Humans have never beaten AI at Go since, while Master went 60:0. Ten years later, when 李世石 appeared alongside T1’s Faker, he said AI could be a good thing if it could offer guidance like a teacher and lower the barrier to entry. But he immediately asked whether “we can still call it human progress” when even the world’s top professional players are copying AI’s moves.
李世石 said he may be “the last generation to learn Go as an art.” From an artistic perspective, winning and losing are merely byproducts of creating a perfect work. 庄明浩 connects that idea to 杜尚’s urinal: modern art has long ceased to be a simple contest of technique and is instead about “narrative, concepts and subversion.” Even though AI and Notebook LLM can generate a complete report almost instantly, he still builds his PPT by hand.
The Fourth Pillar: The Renaissance in the AI Era argues that distribution disruption, the revival of dormant knowledge, and fluid but concentrated pools of patronage capital will happen naturally. Only an “entirely new human vision” will not generate itself. Media has already moved from summaries, institutions and glossy packaging toward secrets, live, founders and eccentric personalities. 庄明浩 leaves narrative, aesthetics, trust and intuition to humans, answering with a friend’s line: “How could AI be as interesting as you?”（“AI 哪有你有趣啊”）.