Summary

ChatCut is an early-stage bet with initial paid traction but opaque commercial validation: 10 people, a roughly RMB10M seed round just completed, and revenue and profit undisclosed. 李凯文 spent 10 years directing commercials and documentaries, and his 2024 short film The Misfits was shortlisted for the Golden Horse Awards; his founder-market fit comes not from chasing the AI wave, but from living through the generational gap between ChatGPT and Premiere. “If God hasn’t designated someone yet, I’ve raised my hand.”
The product is not betting on video generation, but on turning “transcription–content orchestration–timeline mapping” into an editing system callable in natural language. The first version had no AI at all: users edited video the way they edit Word documents. ChatCut is still building the editor and AI in parallel, because a Premiere plugin would impose too many constraints. 李凯文’s trade-off: “AI can get to 80%, while leaving people room to edit,” which is better than reaching 90% while locking users in.
The real market inflection came when ChatCut moved from professional editors to people who want to express themselves but cannot edit, shifting the value metric from a 1.3x productivity gain to taking users straight to an 80/100. Professional workflows are hard to dislodge, and ChatCut cannot rebuild a full Premiere; teachers, real-estate professionals and others, by contrast, can already turn lesson plans, talking-head footage and field recordings directly into content. “You wake up when the tires touch the ground.”
ChatCut believes style cannot be brute-fit from 100,000 “raw material–finished video” samples; it has to encode a documentary editor’s habits, sequencing and checklists into an Agent workflow. Coding has bugs as reinforcement-learning feedback, while whether a film is good is subjective and admits multiple answers. 李凯文 argues that AI should first learn when to revisit footage, find a sound bite and add a transition, then let the details accumulate into style. “It isn’t a problem of getting from point A to point B.”
Gemini 2.5’s 1M—and even 2M—token context window means long-form video need not start with RAG for now, while Gemini 3.0 makes code-generated MG animation viable. Context and multimodality remain technical bottlenecks: current models still cannot understand long video at edit-ready precision. Verifiable instructions can create value first, but whether training a model can solve the deeper problems remains uncertain. “I wouldn’t dare say… whether this is a problem a trained model can solve.”
The disagreement over generative video defines the product boundary: generation can be a creative method, but it cannot replace people or real-world footage. 李凯文 believes even rough, real human storytelling has a vitality that generated imagery lacks. ChatCut therefore generates no pixels: it edits real footage and uses code to make MG animation. The team calls its AI “the most grounded AI in the world,” rather than a model with a style recognizable at a glance.
The current wedge comes from documentary experience, timing and organizational design. Antler’s early funding and Lovable credits let 李凯文 catch the opening window when AI coding was just getting started. The team’s “alchemist” culture makes room for unplanned experiments while the platform compounds workflow data. The endgame is clear enough; the execution path can stay open: “When you can hear the music, whatever you play is right.”

Deep dive

1. ChatCut Compresses 10 Years of Directing into “Editing Video by Chat”

李凯文, born in 1992, describes ChatCut as “editing video by chat.” The company now has 10 people, has just completed a roughly RMB10M seed round, and says revenue and profit are “not something I can really disclose.”
The product idea came from a sharp technological time warp: after using ChatGPT, he opened Premiere to edit a film and felt he had “traveled between the future world and the ancient world.” That was when he thought of editing video through conversation.
He does not attribute his startup edge to engineering ability, but to 10 years of observing ordinary people’s pain points through creative work: “If God hasn’t designated someone yet, I’ve raised my hand.” Solving simple problems for ordinary people requires “simple but deep” insight into life.

2. Risk-Taking Wasn’t Learned After Founding; It Started with Bootleg CDs in Anshan

As a teenager, 李凯文 was intensely curious about the outside world. His father would return from business trips with bootleg CDs bought by weight. With no internet and no ready-made music reviews, he had to build his own judgment system “with his own two eyes and two ears,” which became the foundation for independently assessing new things.
In middle school, he started selling bootleg CDs, carrying whole bags of records on spring outings and passing earphones to classmates for auditions. When he got a first pressing of a Blink-182 album that had been copied unusually thin, he played it over the phone and pressed the buyer: “Isn’t it great? Will you pay RMB50? It’ll be gone tomorrow.” The buyer immediately reserved 2 copies.
From proactively researching and persuading Kent School to let him attend high school in the United States, to entering the Rhode Island School of Design, his attitude toward unfamiliar environments was always “let’s just try it.” Borrowing Kurt Vonnegut’s phrase, he calls it “Dancing Lessons from God”—being willing to accept the steps fate hands you.

3. Advertising Brought Skill and Status—and Exposed the Ceiling of a No-Leverage Business

After joining Vice China in 2016, he realized content was difficult to sustain through traffic-sharing arrangements with iQiyi, Youku or Tencent, so he actively moved into commercial projects. Budgets showed him what could be filmed “when money isn’t a problem”; citing projects for Burger King and Adidas, he says brand credits create industry credibility.
Advertising also demanded creativity, client communication, actor direction and coordination across multiple layers of clients and agencies. Once the technical growth and vanity wore off, he admitted he might not be suited to compete head-on with people who excel at that entire combination: “Everyone is a different species.”
A director’s day rate once reached RMB100K, but the business was still fundamentally “trading time for money,” with no leverage. From 2022 to 2023, he moved near Dianshan Lake for a full year, stopped taking advertising work, and began meditating and exercising, allowing a film, product or app to grow from a blank slate.
During that year he wrote several short films and made The Misfits, later shortlisted for the 2024 Golden Horse Awards, while also beginning to conceive ChatCut. More important than leaving central Shanghai was escaping the fixed identity of a “Shanghai advertising documentary director” and gaining the composure to redefine himself.

4. Directors and CEOs Are Both Always Day One, but a Company Cannot Run Like a Film Set

李凯文 sees directors and founders as starting by forming a vision, then attracting capital, crew and actors to realize it together. Thick skin, resilience, repeated persuasion and “taking complete ownership of the vision” are highly similar across the 2 careers.
Every film has to go from zero to one—“Always day one, it’s always day one”—but once a film is finished, it stays fixed. The day an app launches is merely the starting point for shaping the product. He had to move from delivering a work to continuously iterating a system.
A film set is like a SWAT team that selects rather than develops people; execution can even require decisions in a military-operation mode. A startup, by contrast, must invest in people over the long term. He rejects the idea that dictatorship builds a good organization, believing the combined force of 10 energized people far exceeds that of a single dictatorial genius.
He describes his management role as “the person who lays the road”: first align the team’s understanding, then smooth the asphalt. A product partner with big-tech methodology complements his instincts; goals are negotiated jointly rather than handed down unilaterally.

5. The First Money Tested Persistence; Only Later Did the Story Become a VC Story

During the Golden Horse Film Festival, 李凯文 met Eric, a partner at Singaporean VC firm Antler. Eric first gave him a small amount of money to see whether he could make something and whether he had the persistence to keep going. More importantly, Eric looked for “why he could succeed,” rather than cataloging reasons he would fail.
At the time, he was not preparing to build a scalable, VC-backed AI startup. He only wanted to find the first user, then the first 10, then the first 100. He cited the small-but-beautiful idea of “1,000 true fans, each paying $30 a year,” and believes scarce resources forced him to learn genuine product validation.
ChatCut’s first version only handled speech transcription and text-based editing: users changed the text and the video followed, with no AI orchestration yet. By April, AI editing had begun to take shape; the full concept gradually came to include transcription, content rearrangement and timeline mapping.
Once a large number of paying users appeared before a market even existed, he concluded that the next slope required genuine VC support. He returned to China and met investors continuously for a month, describing the process as undergoing “deep reinforcement learning” like GPT. Fundraising was not a gradual approach to the finish line; until the deal closed, it always looked impossible.
His connection with ZhenFund investor 刘源 began with film: the investor actually watched a movie 李凯文 recommended at 2 a.m., then tried to build a long-term model of the founder. 李凯文’s fundraising advice is not to train a perfect pitch, but to find people with chemistry who are willing to stay on the same boat for the long term.

6. Natural-Language Editing Isn’t a Future Invention; It’s Putting WeChat Instructions on the Timeline

柯基 suggested that natural-language editing might never happen. 李凯文 responded that it has always existed: clients already tell editors over WeChat, “Put this shot here, that one there.” The editor simply used to execute the instructions.
For an entire interview, a user could ask ChatCut to remove several topics, move a good line to the front, then add music and MG animation, letting it produce a rough cut first. Intent-level commands are where chat has an edge over dragging clips on a timeline.
His concession is equally clear: if the task is moving a single frame, dragging it manually may be faster than talking for half a minute, and AI may move the wrong frame. Editing will inevitably be human–machine interaction, so ChatCut keeps both the timeline and text editing rather than pursuing 100% dependence on natural language.
Premiere, Final Cut, DaVinci and Jianying already share roughly 99% of the same editor form factor: “The wheel is already round.” ChatCut does not need a superficial revolution, but it still has to rebuild parts of the underlying system because a Premiere plugin would impose too many limits.

7. AI’s Value Lies at the “Thinking Layer,” Not the UI-Control Layer

李凯文 wants AI to understand what an actual editing assistant does, rather than manufacture so-called innovation on the page. The editor and AI must advance in parallel; the team will not build a complete wheel first and wait for the day when it can announce, “Now it’s AI’s turn.”
He compares current AI to an elementary-school student. The next step is not making it better at obeying “do exactly what I point to,” but sending it through high school and university so that it can see footage, propose a plan and even cut a version above the user’s current level.
First, there is a class of tasks that can be marked right or wrong: cutting off the end of a sentence, or failing to remove filler words after being instructed to do so, are clear errors. Once the basics are correct, however, pacing, judgment and personal style no longer have a single answer, shifting the product problem from engineering execution to aesthetic judgment.

8. Style Comes from Process and Habit; Agents Need to Learn Sequence, Not Finished Answers

柯基 points out that coding can establish a reinforcement-learning loop through bugs, while whether a video “looks good” is highly subjective. 李凯文 admits that “nobody has the answer” on stylization and rejects treating the path from raw material A to finished video B as the only thing to brute-force fit.
When editing projects for Vice or Discovery, he does not begin by imposing his own ego. He follows a process: no matter how much footage there is, start with the interviews; no matter how long the interviews are, start by making a Word document. A stable process naturally produces a stable style. “It isn’t a problem of getting from point A to point B.”
If the team simply uses 100,000 raw-to-finished samples and “wins through brute force,” he believes it can “go off the rails.” What actually determines macro style may be how interviews are screened early, when footage is revisited, what sound bite is sought when a story gap appears, and how a transition is filled.
The next step is to break a real editor’s workflow into different Agents and let them operate across contexts and footage. Much of the sequencing, checklists and human common sense must be defined by people because this knowledge has not been sufficiently internalized in LLM training data.

9. After the “Tires Hit the Ground,” the Target User Shifted from Editors to People Who Cannot Edit

柯基’s challenge was that editing steps are tightly coupled, so professional users will be extremely cautious about changing their workflow. 李凯文 confirmed that reality fully validated the point: after initially targeting professional editors, the team found it hard to disrupt their habits and impossible to build a complete Premiere plus AI.
That pushed ChatCut toward people who were not editors but had always wanted to make content. Phones have made shooting easy enough; editing has become the bottleneck keeping people from ever getting started. The team wants to remove that barrier outright.
李凯文’s value comparison is stark: taking a new user from unable to edit to an 80/100 is more valuable than making a professional editor 1.3x more efficient. “You wake up when the tires touch the ground” became his summary of the target shift.
Early examples include teachers turning courseware or outlines into MG animation, pairing it with their own voiceover to build an account, and real-estate professionals completing videos themselves. Because ChatCut only touches editing, users’ subject matter varies widely and their work shows no uniform “AI signature.”

10. Generation Can Be Source Material, but It Cannot Replace Reality Itself

柯基 proposed another path to serve the same type of user: the user provides only an idea, while the product generates visuals and automatically produces the finished video, bypassing both shooting and editing. 李凯文 believes the 2 routes reflect different judgments about why video exists.
His core view is that video is first a medium for human-to-human communication. A rough person speaking directly to an audience can still possess a vitality that professional polish cannot conceal; if content has no connection to the real world, it is hard to generate the same emotional force.
He does not reject generation, explicitly calling it “one creative method, one artistic technique.” But believing generation can replace shooting wholesale is “a wrong judgment about the world.” AI can replace many steps, but not the core connection between people and reality.
The team compares ChatCut’s AI to a “blue-collar worker,” calling it “the most grounded AI in the world.” 李凯文 observes that the real-world ways people consume video on planes, subways and Douyin have not been rewritten by generative models. A niche product such as Frame.io, which only solves review, can still become what he calls a multi-billion-dollar company.

11. Gemini 2.5 and 3.0 Loosened the Context and MG-Animation Constraints

In the early days of Claude or GPT, a 30-minute video could hit context limits, making RAG and a vector database the team’s likely first response. With Gemini 2.5 offering 1M and even 2M tokens, the problem can at least be deferred for now.
Gemini 3.0 made the MG-animation feature genuinely viable. ChatCut “doesn’t generate any pixels”; it generates code to make the animation, more like creating a PPT for the user than generating a fixed image.
Multimodality does not yet have a major impact on the product, but 李凯文 stresses that it is unavoidable. Existing models have only a rough understanding of audio, language and images, not yet the precision required for editing; in the future, multimodality will be used for both analysis and generation.
The team will not connect an API to every new model just to “show that we’re working hard.” A model is worth building on only when the team adds value tied to the editing workflow. Context and mature multimodality remain the key breakthroughs required before natural-language editing can fully land.
The absence of a massive open-source corpus of “editing code” does not necessarily mean editing is harder than coding. 李凯文 divides capability into verifiable and subjective parts: accurately removing verbal slips, filler words or changing a video to vertical already creates enormous value. The platform can continue accumulating this information and tuning prompts, but he “wouldn’t dare say” whether this is something model training can solve.

12. Unscripted Is ChatCut’s Real Wedge

李凯文 divides video into scripted and unscripted. In film, the script and scene logic are fixed, and editing searches performance and camera angles for the best moment; documentaries, creator content and YouTube must discover content, logic and structure inside unordered footage.
He believes asking AI to make the most delicate judgments in scripted work is currently extremely difficult. If Tony Leung says to Maggie Cheung, “I have a boat ticket. Would you come away with me?” and there are 100 takes, asking AI to select the best performance is “forget it, cannot be done.”
Documentary is the hardest form of unscripted work: enormous footage, an ambiguous objective. Wang Bing shot West of the Tracks for 3 years and cut it into 3 hours, not by executing a preset answer, but by continually confronting the footage until a living structure formed organically.
Yet the same workflow can scale down to ordinary creation. Going skiing in Changbai Mountain and shooting a pile of random clips is also a lower-complexity version of solving how to find rules in disorder. He even treats this podcast episode as a documentary rather than a fundamentally different medium.

13. The Wedge Came from Fine-Grained Judgment—and Not Missing the AI-Coding Window

李凯文 does not think he is operating in a market with no competition; many people place ChatCut in the crowded AI-video category. What is truly scarce is finding a sufficiently nuanced entry point and methodology after choosing a direction, then making users feel a simple “yeah, this works.”
Documentary experience provided the entry point, while timing turned it into a product: LLMs had been around for 1-2 years, AI coding was just beginning, and Antler’s early investment in Lovable gave the team unlimited credits. He realized they could build a web product directly.
He says entering today with the same idea would be “a little late”: the world has not stopped, and other teams have already “thrown a radish into the hole.” A later entrant can still force another radish in, but “life is not gonna be easy for you.”
Even so, he is not a natural-language fundamentalist. ChatCut values the timeline and text editing because “AI getting to 80% while people can edit” is more useful than “AI getting to 90% while people cannot edit.” Control itself is a product capability.

14. The “Alchemist” Culture Turns Unplanned Experiments into Formal Productivity

“Alchemist” is not a job title at ChatCut, but a second identity for engineers, designers and product managers. 李凯文 understands an amateur as someone who does things for love: money is foundational, but real breakthroughs often come from someone pursuing a thing out of love and obsessing over details.
MG animation was not in the PRD or weekly goals at first; a product teammate kept experimenting outside the main task. 2 days after producing a prototype, Gemini 3.0 launched, and the team immediately paused other work to catch the window. He saw it as “an opportunity fate gave us.”
This is not a slight against engineers. It means engineers should also learn to practice alchemy and tinker. If someone only completes assigned work to spec, AI will gradually take that work away; what is genuinely difficult to replace is discovering standards, trying unknown combinations and judging what deserves continuation.
Hiring has therefore shifted from looking only at technical background to evaluating broader interests. Someone intensely obsessed with even a single food may have developed their own high standards; someone with no interest in anything will struggle to make real product judgments.
李凯文 wants to “systematically manufacture” unexpected gains, invoking Antifragile to argue that organizations should actively create favorable surprises because their leverage is far greater than one plus one equaling two. The management challenge is leaving room for trial and error without losing direction.

15. The Clearer the Endgame, the Freer the Team; “Having Something to Do” Is Nine-Tenths of Happiness

柯基 asked whether encouraging surprises means the route to the destination is unclear. 李凯文 answered, “No, I think it’s exactly the opposite.” Because the outcome is clear enough, the team can try multiple paths: “When you can hear the music, whatever you play is right.”
The program also revisited an experience after Jumei went public in 2014 or 2015. The narrator said he “should have been 29” that year, quit with a sum of money and moved into a friend’s lake-view home. By Day 4, the absence of anyone contacting him triggered depression, and he realized that “lying around is not happiness at all.”
李凯文 rates his current happiness at 9/10. He used to envy the wealthy classmates he met at Kent School; now, when introducing ChatCut to them, he feels for the first time that financially free people envy him in return. “Having something to do is your greatest wealth.” He calls ChatCut “this is my life’s work.”
The pain and pleasure of entrepreneurship are not contradictory; the key is whether you believe the work has value. From film sets he learned, “nothing’s supposed to go right, of course it goes wrong, that’s the job.” Obstacles are not interruptions on the way to the work: “If this problem didn’t exist, I have no job.”
The remaining 1 point comes from competitiveness and ego. They may be fuel for efficient decision-making at this stage, but they also prevent complete calm. He has not given up directing; he simply does not need to rush back. Film is a vessel for packaging an understanding of life, while the startup experience is becoming fuel for future work.

The Future of Everything: What CEOs of Circle, CrowdStrike & More See Coming in 2026 The Internet Computer: Caffeine.ai CEO Dominic Williams on Unstoppable, Self-Writing Software