“Once the tools flatten the barrier, the mountaintop becomes clearer.”
A 200-person film crew could soon be just 20 people.
That is the prediction from Huang Zongdi, AIGC visual lead at Tashan Culture, where more than 80% of revenue is To B and To G work, and where 60–70% of projects this year already involve AI.
His company has not laid anyone off, but he sees layoffs across the industry. Fewer people, the same amount of work — exaggerating slightly, those 20 people now process the information density that used to absorb 200.
Two years ago, that sentence would have sounded like “creators are being pushed out.” After sitting through this roundtable, I heard the opposite. AI is splitting the content industry in two: the execution layer is getting cheaper, while both ends are getting more expensive — the end that figures out what to actually make, and the end that dares to sign off on the final result.
That single judgment ties together what all four speakers said.
Qi Maocheng is the closest thing to a “solo operator” in this group. He got into AIGC through Disco Diffusion in 2022, quit his job in 2024, and now makes shorts alone, with a lot of AI. Friends call him an “AI-native director.”
His first take was counterintuitive. In 2024, the three things that mattered most for a short were script, storyboard, and editing. This year, the visuals matter less.
Seedance 2.0 arrived earlier this year, and 2.5 in the second half. Model taste, motion, and physical simulation all got strong. Where prompts used to pile up, a few lines now get a usable result.
So the bottleneck moved. Script and editing are once again the two lifelines of a piece.
His attitude toward “catching up” on traditional craft is even more interesting. In the early days of AI creation, having never trained as a director was an advantage — you could break convention and use non-obvious tricks to work around tool weakness.
Now that the tools are mature, traditional directing skills are catching up and collecting their dues: colliding with writers, pushing character psychology, building arcs, blocking movement, refining the edit — none of it goes away.
He is not just talking. In July this year, he joined a master filmmaker’s workshop and made the top 20 out of 400+ submissions, but not the top 10. His own read: script understanding was too thin, and he froze on stage in front of investors.
After that he put it plainly: a solo person can make small-scale work, but longer, more polished pieces still need a team.
Technology handed entry tickets to more people. It also raised the ceiling.
Huang Zongdi’s view is colder. He trained as a director, now runs AIGC visuals, and the company follows the project; personal artistic temperament takes a back seat.
His description of how organizations are changing is unglamorous but accurate. More and more companies look like factories: foremen, quality inspectors, people “tightening screws” on a line. You generate the image, he retouches it, someone else watches delivery.
Creativity obviously matters, but when enterprise and government clients impose requirements, specs, compliance, and deadlines, people easily become a bolt.
For directors trying to make the transition, he draws two conditions: a young mindset that accepts new things, and an understanding of editing or storytelling. TVC directors who are strong on visuals and weak on story used to get by fine; in the AI era they will struggle.
That sounds bleak. Yang Xiaofeng turns it into a production algorithm.
Huanzhou AI started with children’s animation. The original motivation was blunt: an episode used to cost tens of thousands of yuan; they wanted to bring it down to a few thousand. After making ten or twenty animations, and with the rise of short vertical dramas, their internal tooling matured. Only then did they open Huanzhou AI to outside creators.
Their target is not cinematic quality. It is letting three to six people produce a show in three to six days, reliably delivering content at “70–80 points.”
That 70–80 number is the point. Yang has broken it down: automation alone tops out at 50–60; hand-crafted work reaches 70–80; to reach 90, the director, editor, and art lead all have to be excellent.
A company has to pay people. It cannot let one person spend a month or half a year on a single piece.
The real dividing line for AI-native companies is not whether you can use the model. It is whether you can admit that, for now, you are not chasing 90 points.
It is not romantic. It is commercially clear. Aiming at a perfect score, a solo author can keep polishing alone. Aiming at scale, a company has to make “above average” repeatable.
Yin Xueyuan takes it one step further. His company, Xingzhe AI, has several hundred people and builds AI foundations, serving games, education, tourism, and automotive.
His read: organizations cannot iterate as fast as the technology does. Traditional role-by-task job design is being blown open.
Video is the clearest example. The director will not disappear — someone has to be accountable for the result. But many middle steps will be merged by models, and one new task will grow up: organizing all the models and intelligent capabilities around you.
The one who can arrange image, music, video, and copy models into a team that delivers what the director asked for gets the entry ticket to the new role.
He even coins a term: OPG, One Person Group. One person is a group, and that group owns one outcome. OPC is about daring to open; OPG is about being able to orchestrate. The former is courage. The latter is skill.
In his view, the people AI-era companies most need have four abilities: know what you actually want; organize models and workflows; have the taste to pick one version out of ten; and dare to sign off and own the result.
Those four almost map onto the other three speakers’ pain points: Qi Maocheng is stuck on script and expression, Huang Zongdi on accuracy and narrative logic, Yang Xiaofeng on choosing a single user segment.
Everyone is answering the same question: when the model can do a little of everything, which piece does the human own?
Huang draws the hardest line. Two things are non-negotiable: content accuracy (one extra pixel in a client’s logo is unacceptable; Party and government elements must be exact), and narrative logic (you cannot let AI generate frames at random and string them into a slideshow).
Those two, in practice, are what “signing off” means.
My favorite exchange in the roundtable is this example from Yin Xueyuan.
When serving large B2B clients, they receive bizarre requests. A near-retired executive once asked whether they could, on the side, make a personal documentary for him to leave to his grandson. Under old labor costs, that request was impossible — too expensive, too fragmented.
But once AI drives production cost down, the request becomes interesting.
The more productive we become, the more personalized demand dares to surface. A lot of content people “only thought about” can suddenly be worth making.
That also explains why Xingzhe AI’s best-selling line is AI + education.
Yin is careful to distinguish it from test-score cramming. Their focus is aesthetic and cognitive ability. The buyer changes the product: parents who are score-obsessed get a score-boosting tool; parents who care about whole-person education get AI-generated science-film content; schools and governments want something else again.
Tourism follows the same logic. The old immersive installation surrounded visitors with four beautiful screens, but production cost was so high it might not be updated for six months or a year. With AI, content can be generated in real time, and the experience shifts from “watching” to “playing” — from video toward game.
Generation only boils the water. What is valuable is knowing which pot sits under it.
When money came up, the industry turned out to be less radical than expected.
Yin says video pricing still follows the old model: 8,000 yuan per minute, 80,000 yuan per minute for premium. Clients now ask: you use AI, why is the price the same? There is no clean answer yet.
Token billing shows the problem. Tokens measure consumption, not outcomes. Someone can burn tokens every day and ship nothing — and still pay the bill.
B2B clients will eventually ask the harder question: are they paying for generation attempts, or for the thing that actually works?
In some industries, outcome-based pricing already computes. In AI + automotive, for instance, automated testing used to cost a known amount per test case per person per day. Switching to AI, the discount is quantifiable.
Video has not gotten there yet, but the direction is visible.
Internal profit sharing is also shifting. People who deliver projects one-off collect a production fee. People who turn a project into a reusable workflow, data asset, or sample library sit closer to a “middle platform” and may share long-term IP revenue. People who talk to clients and translate vague asks into executable tasks still bill by project.
The unspoken line: a capability that can be reused is only now eligible for long-term upside.
Yang Xiaofeng builds tooling the same way: internal workflow first, then external distribution. Domestic revenue concentrates on Hongguo and Douyin; overseas, on TikTok. The main channels can account for 80% of revenue.
Huang Zongdi serving B2B does the same. Clients want low cost, short time, and good content — an impossible triangle.
His solution is not to roll the dice more. It is to design the choice architecture in advance, guiding the client into decidable options. One roll that passes, or two with a forced pick, and the client leaves feeling it was their own decision.
That is the most expensive craft in AI cost-cutting: making the client believe they chose this version.
The opening name-checked OPC, and even the more extreme NPC — as if companies were sliding from “one person” to “no person.”
After this roundtable, the human position looked clearer, not weaker. AI handles the middle — faster, cheaper, denser. People are pushed back toward the harder parts: defining the ask, organizing the models, making the choice, owning the consequence.
Once the tools flatten the barrier, the mountaintop becomes clearer.
Moderator: Zhou Shen, Associate Professor, Department of Science Communication, USTC; Executive Dean, USTC AI Cultural Tourism Integration Research Institute.
Speakers: - Qi Maocheng, AIGC director, multi-platform creator - Huang Zongdi, AIGC Visual Lead, Tashan Culture - Yang Xiaofeng, Founder, Xingmen Yuedong (Huanzhou AI) - Yin Xueyuan, Founder & CEO, Xingzhe AI
Zhou Shen: AI is moving fast, and “OPC” — One Person Company — keeps coming up. At WAIC this year I even heard a new term, NPC, or Non-Person Company. People seem to be gradually exiting the frame. Today we want to ask where the human subject sits in these new production relationships.
Qi Maocheng: I got into AIGC early, through Google’s Disco Diffusion model in 2022. In 2024 I left my company. OPC was not yet a buzzword, but I was already making AI video alone as an independent creator.
Huang Zongdi: We are Tashan Culture’s AIGC visual team. We started around the same time as Qi. At first we thought AI was a gimmick, because our work is commercial and the company cannot afford pure artistic self-expression — we always follow the project. By this year, 60–70% of our work involves AI.
Yang Xiaofeng: We are based in Chengdu. We started as a one-stop tool in 2024, because we had seven or eight years of children’s animation and wanted to bring an episode from tens of thousands of yuan down to a few thousand. After making ten or twenty animations and with the rise of short vertical dramas, we became one of the first shops producing them. Once internal tooling matured, we opened Huanzhou AI to outside creators.
Yin Xueyuan: We are Xingzhe AI, a B2B AI foundation company. We started with AI for Game — content creation tools for game studios. We then turned multimodal and interactive capabilities into an AIGC OS platform used in education, cultural tourism, and automotive. We are the ones building the arena, not just fighting in it.
Zhou Shen: Once AI is inside the workflow, how does the organization change?
Qi Maocheng: I have mostly been a solo operator, one person plus a lot of AI — close to OPC. I have not gone through a large crew production. People like me are called “AI-native directors”: we had no systematic film training before AI.
My observation: in 2024, script, storyboard, and editing were the three pillars. This year, with Seedance 2.0 and 2.5, model taste, motion, and physics all jumped. The two things that now reconstruct narrative are script and editing. Visuals matter less.
Zhou Shen: Do you still need traditional directing training?
Qi: Early on, lacking traditional training was an advantage. We could jump out of convention. But as tools mature, traditional skills matter again: working with writers, deriving character psychology, building arcs, blocking movement, editing.
I am still catching up. Before AIGC I wrote film reviews and short fiction. Then tools let me generate frames and tell stories with them. After Seedance, I realized I need to study directing for much longer. I used to work in architecture, where creators often do their best work in their fifties or sixties. I think the same applies here.
Huang Zongdi: Our business is 80%+ To B and To G, with some TVC and little Douyin vertical drama. Three observations.
First, the creative ecosystem changed. I trained as a director; I no longer direct. AI lets talented artists enter the industry — that is both pressure and motivation for formally trained directors.
Second, from a production standpoint, teams really can shrink. We have not laid off, but many companies have. The previous session’s director said a 200-person crew might become 20. I think that is plausible. But those 20 people now process what 200 used to — information density goes up. That is where the phrase “work animal” comes from.
Third, our company increasingly looks like a factory: foremen, QC, line workers. You generate, he retouches. It is almost unavoidable, especially with enterprise and government clients who impose specific requirements.
Zhou Shen: What distinguishes directors who transition successfully?
Huang: Successful ones have a young mindset — not necessarily young age, but open to new things. They also understand editing or storytelling. TVC directors who are visually strong but weak on story used to do fine; in the AI era that gap is fatal. Directors who can write and tell stories transition more easily.
Yang Xiaofeng: We used to only do distribution and outsourcing. We had no director, no content team. After AI came, I built the product alone, then let internal and external users try it. Only then did we start hiring directors, then “card-pull” operators and editors.
The loop is: production feedback, engineering fixes the tool, more feedback. That lets us scale.
We are not chasing cinematic quality. We cannot afford top-tier art direction. Our target is three to six people producing one show in three to six days, at above-average quality. We have built our own writers’ room and keep iterating with data.
Our production team is now forty to fifty people. We do domestic and overseas short drama, and are looking at AI interactive drama.
Automation alone reaches 50–60 points. Hand work reaches 70–80. To reach 90, director, editor, and art all have to be excellent. We produce at 70–80.
Zhou Shen: Does this show up in tool design?
Yang: I am a product manager and the first user. I test from script through production before deployment. If the team can ship fast, we launch; otherwise we keep iterating. We now iterate by day, sometimes by hour.
Yin Xueyuan: We have hundreds of people serving large B2B and G clients. The organization cannot iterate as fast as the model.
Traditional job design splits work by task: director, editor, writer. As models get stronger, those boundaries blur. Two tasks may be merged by a model. So we ask: which tasks merge, and what new tasks appear?
In video, someone still has to own the result — usually the director. But a brand-new task appears: organizing all the models around you. Image, music, video, copy — whoever orchestrates them best becomes a new role.
We may move toward OPG: One Person Group. One person is a group, and that group owns one outcome.
Zhou Shen: Who are you most missing?
Yin: Four abilities.
First, knowing what you actually want. Models can do a lot, but what do you want them to do? We once got a request from a near-retirement executive to make a personal documentary for his grandson. Under old labor costs we could not touch it. Extreme productivity releases extremely personal demand.
Second, organizing models and workflows. How do you arrange the pipeline to maximize the chance the output gets picked?
Third, taste. When the model gives you ten versions, which one do you choose?
Fourth, accountability. Someone has to say: this one, we ship it, and I own it.
Yin: We cannot say one vertical is universally best. Our best seller is AI + education. Not test-score cramming — aesthetic and cognitive development.
Cultural tourism also has huge upside. The old immersive installation surrounds you with four beautiful screens, but production cost means it is rarely updated for six months to a year. With AI, content can be generated in real time, and the experience shifts from watching to playing — from video toward game.
Zhou Shen: Young people hear “AI-native director” and think they can become solo creators. What is real, and what is fantasy?
Qi: The barrier and organizational cost are now low. Anyone with compute can make a short.
But two warnings. First, generating frames does not make you a director. Script and editing are older crafts and matter more than generation itself.
Second, a solo person can start, but does not have to stay solo. In July I entered a master filmmaker’s workshop and made top 20 out of 400+, but not top 10. The top 10 get to develop a feature. I missed because my script understanding was thin and I froze on stage as my own producer.
After that I realized: solo work is fine for small pieces, but longer, more polished work needs a team. If you just want to play, go for it. If you want a long career, you need organization and collaboration.
Zhou Shen: Which visual choices do you hold the line on, and where do you let the model drive?
Huang: Most of our work is B2B/G. One memorable project was L’Oréal. They wanted to show an AI Agent that did not yet exist, and asked us to “film” it for a Paris HQ presentation, with only a one-week deadline.
ChatGPT did a lot of the early scripting. I joked that ChatGPT was a 5,000-yuan copywriter. When content is abstract or slogan-like, I let the model run first, then humans edit.
Two principles are non-negotiable. First, accuracy: a misrendered logo, an extra pixel, a wrong Party or government symbol — these must be fixed by humans until they are exact. Second, narrative logic: you cannot let AI generate frames at random and string them together. Editing and direction own that.
Yang: There are many short-drama tools — LibTV, Jimeng — backed by hundreds or thousands of people. We cannot satisfy everyone. We choose a segment, find who will pay, and serve them.
We prioritize internal workflow first. Users who want a professional, scaled pipeline follow our path. Users who want full freedom are not our segment, and we will not build for them.
Yin: Industry pricing is still traditional: 8,000 yuan per minute, 80,000 for premium. Clients now ask why AI did not lower the price.
Token billing measures consumption, not outcomes. People burn tokens and ship nothing.
For B2B, I expect clients will increasingly pay based on the result they receive.
Internally we split contribution three ways. Project delivery earns a one-off fee. People who build reusable workflows, data assets, and sample libraries sit closer to the middle platform and may share long-term IP upside. People who translate client asks into executable briefs still bill by project.
Outcome-based pricing already works in AI + automotive: automated test cases had a known cost per case per day; with AI, the discount is quantifiable. Video has not gotten there, but the direction is visible.
Zhou Shen: AIGC self-media traffic is terrible. How do you adjust?
Qi: It is hard to answer; my own traffic is also poor. I am not anxious. My goal is to go deeper as a creator, so short-term traffic does not matter that much.
Zhou Shen: Has Hollywood’s industrial model collapsed?
Huang: Not yet. The unions now allow vertical dramas; everything blends. Directors like Nolan still hold up. Taste and industrial process are still there.
Zhou Shen: Can a novelist become an AI-native director?
Qi: That is close to my own path. Starting from text is an advantage — novelists know how to write a story. The hard part is learning the tools, but that is learnable.
Zhou Shen: Which AI + education direction wins?
Yin: No universal answer. Extreme productivity releases extreme personalization. Find who pays — parent, school, or government — then build for what they care about. Score-obsessed parents get score tools; whole-person parents get AI science content.
Zhou Shen: How do you expand distribution?
Yang: Focus on the main channels. Domestic: Hongguo, Douyin. Overseas: TikTok. Main channels can account for 80% of revenue.
Zhou Shen: What is the common problem B2B clients bring?
Huang: Low cost, short time, good content — the impossible triangle. The job is producing with a director’s brain. AI and card-pulling are secondary. If you fully understand the ask, one pull may pass; otherwise two pulls with a forced either/or. You design the choice architecture ahead of time, and guide the client into it. “Trick” in quotes. That is the core B2B/BG skill.
