The "runaway" intern from ByteDance has raised 200 million yuan.
Foreign media reports say Tian Keyu, a former ByteDance intern, has launched his startup focused on "world models", raising nearly 30 million U.S. dollars in financing, with a post-money valuation reaching 200 million U.S. dollars (equivalent to about 1.4 billion yuan).
You may not know who Tian Keyu is, but you have most likely heard rumors about him. In 2024, during his internship, Tian Keyu was sued by ByteDance in court for attacking the company's internal model training cases. At that time, ByteDance requested the court to order Tian Keyu to compensate the company for 8 million yuan in tort losses plus 20,000 yuan in reasonable expenses, and make a public apology.
Later, the two sides held their own versions of the story, and no public progress was made until Tian Keyu started his startup recently, which brought the old incident back to public attention.
Strictly speaking, this financing is not new news. About three months ago, some people had already disclosed information about Tian Keyu's financing. At that time, we contacted relevant investors, who said they could not respond to the matter, but revealed that the case had been ruled, and Tian Keyu would give the public an explanation. Eventually, the protagonist of the news accepted an interview with Bloomberg, which did bring some new information.
The outcome of this case is that the court ordered Tian Keyu to pay ByteDance 500,000 yuan for breach of contract, and all his salary was confiscated at the same time.
The ruling only defines his act as "breach of contract" rather than a criminal offense, which means the nature of the act is not extremely serious. Previous rumors may have been exaggerated, or there may have been various behind-the-scenes mediations. The details are unknown. In any case, in my opinion, in a civilized society ruled by law, having a "previous record" does not mean that one cannot start a business. But whether investors mind this episode and whether it will affect their judgment on this person's character is entirely a matter of different opinions.
This wave of AI entrepreneurship is particularly special, and ethical discussions triggered by the dangerous side of powerful technologies are an important topic in the overseas AI circle. If a person with blemishes in professional ethics who once used technology to "fight violence with violence" really masters the key to opening the door to the future, where will he lead humanity? Questions like these are at least worthy of careful consideration.
Judging from the results, investors are still willing to cast their votes of approval. At present, the name of Tian Keyu's product and company has not been announced to the public, and it is still in stealth mode. This means that investors are betting more on their professional judgment of this individual. In terms of resume, as a Peking University doctoral student, Tian Keyu has published many papers in top academic journals, and won the NeurIPS Best Paper Award as the first author, which fits the standard "little genius" profile among this wave of AI entrepreneurs.
The world model track Tian Keyu chose is undoubtedly a huge hit this year. Not long ago, the lab of Li Fei-Fei, known as the top figure in the world model field, was acquired by AMD for 8.2 billion U.S. dollars, equivalent to more than 50 billion yuan. Lisa Su emphasized that one of the reasons for the acquisition is to obtain the world's top talents gathered by Li Fei-Fei, which also shows how eager tech giants are for top talents.
The AI industry is still extremely hot. For people who master scarce technologies and have top research achievements, the technical value itself can be a ticket to the game. Other things may have to be put aside for the time being.
The Intern Who Was Sued by ByteDance for 8 Million Yuan
The name Tian Keyu was first known to many people in a rather undignified way.
In August 2024, ByteDance terminated Tian Keyu's internship. According to ByteDance, from June to July 2024, Tian Keyu, dissatisfied with the team's computing power resource allocation, interfered with the model training of the company's internal research projects by modifying codes and other methods, which affected a research project of the commercial technology team. ByteDance then dismissed him and referred the relevant situation to his school for handling.
Tian Keyu later denied ByteDance's allegations and even called the police claiming he was defamed. In November of the same year, ByteDance filed a lawsuit against Tian Keyu at Haidian District People's Court of Beijing, demanding that he bear 8 million yuan in tort losses and 20,000 yuan in rights protection expenses, and make a public apology.
According to some controversial details disclosed online, Tian Keyu exploited the security risk existing in Hugging Face's model checkpoint loading process. To put it simply, a checkpoint is a phased file during the model training process. When the model reloads these files, it needs to go through deserialization to convert the content in the files into executable objects for the program. If malicious code is implanted in the file itself, the training machine that loads it may be further controlled, thus affecting model training and even leading to consequences such as data leakage.
This kind of security risk is not a completely unfamiliar technology for AI engineers, but the focus of the dispute in this conflict is not "whether this vulnerability exists", but whether Tian Keyu used this technology to interfere with other people's training tasks, and how this act should be defined.
Recently, Tian Keyu also gave an explanation in an interview, saying that the team's computing resources were limited at that time, and he thought another intern was occupying resources that should have been allocated to other researchers, so he chose to stop the other party's project. He admitted that this way of handling was improper, and described his behavior at that time as a kind of "lynching". "What I did at that time was no different from fighting violence with violence," he said. "If I could do it all over again with today's perspective, I would definitely take a wiser approach."
Foreign media cited the written judgment they saw, saying that the court finally ordered Tian Keyu to pay ByteDance 500,000 yuan for breach of contract, and all his salary was confiscated at the same time.
After this incident, Tian Keyu's whereabouts once became a topic of discussion. Among all possibilities, starting a business is almost what most people think is the most likely outcome. It has to be admitted that his academic achievements are very outstanding. He graduated from the School of Software, Beihang University with a bachelor's degree in 2022, studied under Professor Wang Liwei, with research directions in deep learning optimization and algorithms. He served as a reviewer for NeurIPS, ICML and ICLR, and has published 4 papers in top conferences and journals, including 2 first-author papers on NeurIPS.
Last summer, Tian Keyu graduated from Peking University with a master's degree, and is now a doctoral candidate at Peking University. In July this year, news came out that Tian Keyu had started his startup in the "world model" field. The project was incubated and led for investment by Meng Xing, a partner at 5Y Capital, with total financing of tens of millions of U.S. dollars and a valuation of about 200 million U.S. dollars.
30 Million U.S. Dollars for an Admission Ticket
Although his departure from ByteDance was unpleasant, the achievements he made during his internship at ByteDance paved the way for his subsequent startup.
During his internship at ByteDance, he co-authored the paper *Visual Autoregressive Modeling* with ByteDance as the first author, which proposed to let the model generate images layer by layer according to different scales, instead of processing pixels one by one or visual tokens one by one like traditional models. The paper argues that this method can improve the speed and efficiency of image generation.
This paper won him the NeurIPS 2024 Best Paper Award, which is regarded as a major breakthrough in the way AI systems generate images, and its value is widely recognized in the industry. Now Tian Keyu's startup direction is also carried out on the basis of this paper, that is, trying to fundamentally change the way models understand and generate visual content.
This startup currently has only 10 employees, including other former ByteDance employees. Tian Keyu revealed that the team has already developed a system similar to an AI "dictionary", containing 200,000 symbols that ordinary people cannot read, to help the model process videos. He believes that human language is difficult to describe a video completely and accurately, so the team hopes to first build a set of its own language for AI, and then train it with 100 million hours of video.
According to his statement, this method is expected to reduce the cost of video generation by an order of magnitude, that is, the cost of generating 1 second of video will be reduced to at least one tenth of the original. The team plans to demonstrate its technical capabilities through on-site events before the full model is officially released in 2027. At this stage, they are more focused on whether the technology itself can work, and have not put their main energy on consumer-grade products and business models for the time being.
However, this sum of money also puts Tian Keyu in an extremely fiercely competitive position. The world model track has extremely high requirements for the quality of talents. Facing him are top scientists such as Li Fei-Fei and Yann LeCun, as well as a large number of young researchers with top universities, large tech companies and lab resumes who are being competed for by capital and startups.
The fact that capital is willing to pay a premium for technical talent does not mean that some controversies can be ignored, let alone that he can easily achieve "great results". More and more voices have pointed out that "the world model is a basket where everything can be put in". For the startups we have contacted, there is still no conclusion on which route among 3D generation, video understanding, robotics, spatial intelligence and many other directions can succeed. There is also no answer as to whether it will be the next round of AI infrastructure, or enter a new adjustment cycle after a round of capital frenzy.
For Tian Keyu, 30 million U.S. dollars is more of an admission ticket. Next, he still needs to prove that besides technology, he can be a qualified entrepreneur. Building a team, handling relationships with core employees, turning a technical concept into a truly verifiable product, and making timely adjustments when the technical route deviates, these are all capabilities that cannot be replaced by the label of "genius".
This article is from the WeChat Official Account "Dong 40th Street Capital" (ID: DsstCapital), written by Wei Xianghui, and published with authorization from 36Kr.