AI News

Investigation into a Zhejiang Sophomore's AI Startup Making 600,000 Yuan in Two Months: Code Highly Similar to Open Source Project Infinite Canvas, Models Called via Third-Party Relay

HuxiuUpdated Nov 11 21:41

AI summary

A Zhejiang sophomore claimed that the AI video platform "Benzhu" generated 600,000 yuan in revenue and 20,000 users in two months, but an investigation found that its homepage HTML is almost line-by-line identical to the GitHub open source project Infinite Canvas, and the model configuration structure is also the same, retaining the original project's naming. The code shows that it calls models such as GPT Image, Gemini, Grok, and Sora 2, with requests first forwarded through a third-party relay station; the only open source model, Flux, is marked as a test, and no evidence of self-developed post-training or short-drama-specific training was found. The website was based on the AGPL-3.0 license at the time, and no public disclosure of modified source code or attribution to the original author was seen.

Body · AI translation

[[**A sophomore at a Zhejiang university claimed on television that his AI startup made 600,000 yuan in two months, but an investigation found that his product's code is highly similar to an open-source project, that it mainly relies on a third-party relay to call overseas models, and that the claims of "independent development" and in-house optimization are questionable.**

**Key points**

**1. The 600,000 yuan in revenue and 20,000 users lack verifiable evidence**

In an interview with Hangzhou TV, he claimed that the AI video platform had revenue of 600,000 yuan and 20,000 users in two months, and that generating one second of video cost only 5 fen; however, the website subsequently became inaccessible, and the investigation could only rely on server and code information from before the maintenance.

**2. The so-called technological breakthrough is mainly a relay calling third-party models**

Clues to models such as GPT Image, Gemini, Grok, and Sora 2 appear in the code, some of which are packaged as an "all-purpose model"; the only open-source model seen is Flux, and it is marked as being in testing, with no evidence found of in-house post-training or training specifically for short dramas, and requests are first forwarded through a third-party relay station.

**3. The project's code and assets rely heavily on open-source projects and external resources**

The homepage HTML of the website and of the GitHub open-source project Infinite Canvas is almost identical line by line, the model configuration structure is also the same, and the original project's naming is retained; the shot-language assets directly reference Melies resources, and the prompt library comes from multiple GitHub repositories. Under the AGPL-3.0 license in effect at the time, no public modification of the source code or attribution to the original author was seen.

This article is from the WeChat public account: Shan Sen Nan, author: Shan Sen Nan, original title: "The Truth Behind 'A Zhejiang Sophomore Made 600,000 Yuan from an AI Startup in Two Months'"

This time, Hangzhou TV may have been taught a lesson by a sophomore.

Over the past few days, I kept coming across a short video that can be called a new-generation "meritocracy" promotional film. In the video, a sophomore at a university in Zhejiang is interviewed by Hangzhou TV and claims that he made 600,000 yuan in just two months through an AI startup.

What is even more impressive is that, according to him, almost the entire project was completed independently by himself; he founded the company alone, and even a work that won second prize at Milan Design Week used their software.

After this interview was released on November 21, it quickly trended on Weibo and attracted a great deal of attention. I believe that millions of parents across the country, after seeing this news, could not help saying to their children: "Look at what this college student is doing, and then look at you!"

For this product, which is said to have been almost entirely independently developed, part of the code is highly similar to an open-source project on GitHub. And the main AI API it calls actually comes from a third-party relay station.

After I shared these findings with good friends, someone immediately gave a sharp comment:

"There are thousands of people running API relay station businesses, but this is the first time I've seen someone go on television and put themselves in the spotlight 🤣."

"A sophomore, two months, an AI startup, made 600,000 yuan"

First, a sophomore at a university in Zhejiang was interviewed by Hangzhou TV's comprehensive channel. The interview began circulating on the evening of October 21, and by the 22nd it had trended and was widely reposted.

After all, a sophomore making 600,000 yuan in two months through an AI startup is indeed an eye-catching gimmick.

In the interview, the student said that his AI video generation platform, Benzhu, already had 20,000 users and revenue exceeding 600,000 yuan. Moreover, the price was very cheap: generating one second of AI video cost only 5 fen.

He also mentioned that the work "Jingwei Fills the Sea," which won second prize at Milan Design Week, was also made using their platform.

During the interview, the student appeared quite confident. He said that generating one second of AI video in the industry generally costs 1 yuan, while his platform, by compressing costs, could lower the price to 5 to 8 fen per second, or even as low as 4 fen.

After all, according to him, this amounts to directly reducing the price by more than 90%. If he could truly achieve this through his own technology, that would indeed be quite astonishing 🤯.

Moreover, according to his account, from registering the company to developing the product, almost the entire entrepreneurial process was completed independently by him. Even more exaggeratedly, subsequent media reports claimed that the platform was mainly optimized based on open-source models and had even been specially trained for short drama creation.

If all these claims are true, then this sophomore is indeed quite impressive, and I suggest ByteDance Seed recruit him.

Unfortunately, when I actually investigated this product, I found that many of the above claims contain considerable exaggeration. Some of the content is basically just young people shooting their mouths off.

Next, let us look at them one by one.

What is the truth?

I took a look, and one possible reason is that the website was updated recently, causing old files to become invalid. For example, the yTxsb7ic that appears in the error message is an identifier in the file name. After the website is updated, if the old file is deleted, the browser may report an error because it cannot find the corresponding file.

So I suspect that the website administrator may have urgently carried out a round of "maintenance" after seeing public opinion online 😅.

Fortunately, however, I conducted the investigation before the website maintenance, so I was still able to find its server information. The website uses a Tencent Cloud server, with the IP address 124.xxx.xxx.xx, and the server is located in Shanghai.

The entire website is a typical Vibe Coding-style AI video generation canvas tool, and from the interface to the functions, obvious traces of AI-assisted development can be seen.

Actually, there are already quite a few AI video creation agents of this kind now, and products like Lovart and Liblib TV all offer similar features.

Optimized based on open-source models, specially trained? False

Actually, the method is very simple: on a Windows computer press F12, and on a Mac press Command+Option+I, and you can open the browser's developer tools. Then search for "Grok," and you can find some related code and API clues.

As it turned out, I really did find quite a few model names, and some of them looked quite familiar 🤣.

After all, the foundation models from various companies are already very strong now. For many ordinary users, whether the video generation result is good mainly depends on the model itself. Of course, the website's workflow and interaction design also have value, but what everyone cares about most is still which models can be used, how good the results are, and whether the price is cheap.

And on this website, I found that it actually involves quite a lot of models. Besides the Seedream series, I also found related names or call clues for models such as GPT Image, Gemini, and Grok.

What's even more interesting is that some of these models are packaged as "all-purpose models" on the Benzhu website, but judging from the code clues, what is behind them may still be calls to these familiar third-party models 😅.

I also used Opus 5.5 to help sort through the code and model information in the Benzhu website. For example, for the "All-Purpose Model 2.5" labeled on the interface, clues related to Image 2.5 and Sunburst can be found in the code.

In addition, I even found related information such as Sora 2 and the Grok image generation model.

As for how these models are actually connected, there are even more interesting findings later 🤣.

After all, if you directly call the official APIs of vendors such as OpenAI, xAI, and Google, the cost is not low.

So I carefully looked through the code again and found that requests for these models are actually first sent to a third-party relay station, and then forwarded by the relay station.

This is very interesting.

To put it plainly, this AI video website is almost a "relay station of a relay station." A large number of the main models behind it are actually connected through the aggregated API service of another relay station.

Let's continue. I also checked the credit prices of the various models on the website and found them to be surprisingly cheap.

For example, generating one image with GPT Image 2 costs only 18 credits to start, and Image 2.5 also costs only about 40 credits.

I checked around, and the only open-source model that can be found on the entire website is Flux, and it is even labeled as a "test model" in the backend.

So, as for the "optimization based on open-source models" previously reported by the media, and even "model training specifically for short drama creation," at least from the code I checked, I found no corresponding evidence.

As for the unique technologies we initially imagined, such as post-training open-source models ourselves and greatly reducing generation costs, no relevant signs have been seen so far.

In the end, these technological breakthroughs may be more something we imagined ourselves after watching the interview 🤣.

Self-taught, independent development? Basically false

First, I used Opus 5.5 to analyze the code of this website. Judging from the project structure, I suspect it is very likely a Vibe Coding project. After all, with AI-assisted development now, building such a website in two months is not particularly difficult.

What's more coincidental is that on GitHub there happens to be a classic open-source project with more than 8,400 stars called "Infinite-canvas."

So I took the code of the two projects and compared them.

The result was that the similarity was astonishingly high. The homepage HTML is almost identical line by line, and even the code structure of the model configuration is highly consistent 😅.

Second, the open-source project Infinite Canvas later switched to the MIT license. This license allows commercial use and secondary development, but requires retaining the original author's copyright and license notice.

However, on the homepage of this student's website, I did not see any attribution to the original project. By the time I wanted to check further, the website had already entered maintenance status and could not be logged into normally 😅.

But a large number of traces of the original project remain in the code.

For example, many places are still called Infinite Canvas, and even the local database name and configuration storage keys continue to use the original project's naming.

So, as for the "independent development" mentioned in the interview, to what extent it was actually independent is indeed worth questioning.

Then I discovered another very interesting feature in Benzhu called "camera language." This page collects 424 shooting techniques, divided into 13 categories, somewhat like the popular video creation Skills now. In the interview, this was a key promotional point:

But when I checked the code, I found that these techniques are actually just the images and videos accompanying the open-source project, directly referencing resource addresses at asset.melies.co.

So, for an open-source website, even loading images and videos directly from someone else's server is still a bit immature 🤣.

What's even more interesting is that there is also a prompt library in the Benzhu website, and the content of this prompt library actually comes from another 8 GitHub repositories (the open-source version comes with 5), and a real-time scraping mechanism is also designed in the code.

For example, ZeroLu/awesome-gpt-image, YouMind-OpenLab/awesome-seedance-2-prompts, and so on.

However, you can't exactly say this student made no changes of his own. The original Infinite Canvas canvas had 6 types of nodes, while this website's code has two more: one is the "Director's Stage," corresponding to the 3D director's stage feature.

The other is actually "Snake," which I discovered when I had Uncle Opus review the code—even the AI was baffled.

My personal guess is that this may have been a test feature added during Vibe Coding and was never cleaned up afterward.

Below, let's sort out the timeline of this "open-source wrapper."

I carefully compared the TV interview, the update times of the open-source project, and the launch time of this sophomore student's website, and found the whole timeline very interesting.

First, the open-source project Infinite Canvas ("Infinite Canvas") on GitHub was released on May 18, 2026, using the AGPL-3.0 license at the time, and was only changed to MIT around August.

As for this sophomore student's website, the relevant features had already been added by July 9.

Interestingly, Infinite Canvas updated its canvas node plugin system on July 15, but this student's website does not have that part of the functionality.

So I suspect he most likely developed based on the open-source project code as it was between July 9 and 15.

The problem then arises: at that time, Infinite Canvas was still using the AGPL-3.0 license. Under the requirements of this license, if a modified program provides services to users over a network, it typically needs to provide the corresponding source code to users.

But so far I have not found any information about this student publicly releasing the modified source code, nor have I seen clear attribution to the original author. If the code source and license timing can both be confirmed, then it is worth further asking whether the open-source agreement was complied with 😅.

Looking further ahead, the website domain was registered on July 24, 2026. By October 23, the homepage file still showed modification records.

Putting these times together makes the whole story much clearer.

At this point, let's go back to the "two months of independent development" mentioned in the interview.

The two-month timeframe does indeed match up. But the words "independent development" need to be discussed carefully.

If a large amount of a project's code comes from an existing GitHub open-source repository, and then AI programming tools are used to add login, billing, and API integration features, that of course still counts as development work.

Moreover, counting from the domain registration on July 24, the website had only existed for about three months by the time the interview spread. Yet in the interview, he already claimed the platform had 20,000 users and revenue exceeding 600,000 yuan.

In just a few months, from developing the website and launching operations to claiming to have earned 600,000 yuan, then accepting a TV interview and trending on Weibo, and finally being questioned over the source of the code and models.

This whole chain of events happened far too fast.

Fast at building products, fast at telling startup stories, fast at trending, and even astonishingly fast at having one's code exposed 🤣.

Then I also checked the website's WHOIS domain registration record and found that a Shanghai cultural communication Co., Ltd. appeared in the registration information.

To protect privacy, I won't disclose the specific name here. Moreover, the WHOIS record only has an English name, and to match it to domestic business registration information, one would have to verify it one by one on Qichacha.

I did find a company with a similar name and whose person in charge shares the same surname as this student, with registered capital of about 1 million yuan.

But there are simply too many companies with the same name and same surname, and there is currently not enough evidence to prove they are the same one, so I will not make any connection judgment here.

After all, people who have set up companies themselves should know that from preparing materials to completing registration, quite a few procedures still have to be gone through. Even registering as a sole proprietor is not necessarily as simple as one imagines 😅.

I originally wanted to continue checking information such as shareholders' subscribed and paid-in capital, but some detailed materials on Qichacha require a paid membership.

But the problem is that many people nowadays love to promote the labels "young age," "independent development," and "making big money" together.

Although it's a joke, it does reflect that people are increasingly distrustful of this kind of startup story 🤣.

Second, there is "independent development," which cannot be played this way.

I think there is nothing shameful about using AI Vibe Coding, and secondary development based on GitHub open-source projects is completely fine as well. Who doesn't use AI to write code these days?

But the problem is, if a large amount of the code comes from someone else's open-source project and you mainly added features like login, billing, and model API integration, then you should explain these things clearly and not present it as "independent development."

Moreover, this project also involves open-source license issues.

Finally, there is the API relay station.

To be honest, I don't want to criticize him simply for using a relay station. There are indeed quite a few developers in China now using third-party API aggregation services.

But if you have already turned it into a paid commercial product, things are different.

For example, has the third-party relay service obtained the relevant authorization? Which servers do the videos, images, and prompts uploaded by users pass through? Will the data be recorded? If the upstream interface suddenly fails, what happens to the credits and services users purchased?

So why is it that a sophomore student, merely by putting out the keywords "AI startup," "two months," and "600,000 yuan," can so easily gain a large amount of attention?

Why, during the media interview, did no one ask a few more questions about where the models came from, who actually developed the technology, and how the 600,000 yuan in revenue was calculated?

Many people will think that students are still students after all, lacking social experience, with relatively simple ideas, and easily influenced by their parents, schools, and the whole social environment.

But if a young person has been raised from childhood on the narrative of "you must be better than others," "the earlier you succeed the better," and "only making money counts as ability," then he may be even more eager to prove himself than adults.

Put positively, this is called ambition and drive.

But if it develops in another direction, it may turn into being eager to start a business, eager to make money, eager to become famous, and even eager to package a project that has not yet been fully validated into a success story.

Of course, I don't know what this student specifically experienced, and I don't want to attribute all these problems to him personally.

Because what is more worth reflecting on than a young person's eagerness to prove himself is why the whole society encourages young people to succeed as early as possible, while few tell them that it's okay to do things a little more slowly and that failure is completely normal.

We probably don't need a "city of impetuousness." AI industry signal channel: Frontier technology. This content is published with the author's authorization; the views represent only the author and do not represent Huxiu's position. If you have objections or complaints about this manuscript, please contact tougao@huxiu.com. Everyone who is changing and wants to change the world is in the Huxiu APP [ ]

Source

Huxiu