Author: elseswhere
First, an apology. For certain reasons, we have not yet finished editing the podcast video with Tian Keyu. It should be released tomorrow morning.
Before the end of tonight, we will first turn part of the interview into a dialogue and post it here.
We joked earlier that elsewhere's podcast is a bit like "Art Life." If we use that analogy, then this episode with Tian Keyu is probably more like "Legal Report."
An intern rumored to have "poisoned" a large model was fired by the company and repeatedly caught in a vortex of online public opinion, and the offer he received was also revoked. And at the most desperate moment of his life, he received the NeurIPS 2024 Best Paper. After that, he decided to start a business and also secured investment from a USD fund.
All the elements are too full.
One development is: ByteDance's lawsuit against him has reached a first-instance judgment. The general situation we understand is: the court ruled that Tian Keyu should pay 500,000 in compensation. But ByteDance then filed an appeal, and the second instance is still ongoing.
In the eyes of many people, Tian Keyu is a big villain. Some even call him Lao Tian—with the "Lao" from prison.
In July of this year, we recorded this video with Tian Keyu. In it, he gave a complete account of what he saw and the causes and consequences, including why he did it, what exactly he did, what he admits, and what he believes he has been stigmatized for. In addition, it also includes his earlier education and academic choices, which can help us better understand Tian Keyu as a person.
It needs to be explained that this interview took place more than three months ago, and some of the statements in it may differ somewhat from the current situation and market context, such as many things related to his entrepreneurship in the world model direction.
Given that the litigation is still in the second instance, some content in the original interview that might cause controversy has been concealed.
The following text contains some key information from the interview. But we still suggest everyone wait until tomorrow morning. Listen to the full audio version on the Xiaoyuzhou App, as well as the full video version on platforms such as bilibili and Video Account.
Compared with business news, this is more like social or legal news. It involves several parties, including, besides Tian Keyu, ByteDance, another intern he mentioned, and people from various sides.
Everyone will have their own truth that they see and believe. The following is Tian Keyu's version.
What kind of mistake this was
elsewhere: Why, after two years, are you only now willing to come out and formally talk about this?
Tian Keyu: At the time, public opinion was huge. I even found a blogger on Weibo, wanting to explain the actual situation. But soon both the company and the school advised me not to speak. And I thought it was better to wait for a legal result.
elsewhere: Has there been one recently?
Tian Keyu: There is something of a preliminary result. At the time, ByteDance mainly had two claims: one was an economic claim for more than 8 million; the other was about goodwill (requiring me to publicly apologize). The court's current first-instance result is a discretionary compensation of about 500,000, with the others rejected. The second instance is still being heard.
Elsewhere: Is there a logical relationship between 500,000 and that more than 8 million? Why was compensation awarded?
Tian Keyu: There is no logical relationship. The 8 million was ByteDance's claim, and they also claimed at the same time that I should apologize. The amount of 500,000 happens to equal my three years of internship salary, plus some legal expenses from the company's lawsuit.
I have never denied my behavior. I also sincerely apologized in front of the judge. I think the final compensation is very fair, and I accept this amount.
elsewhere: Let's get straight to the point. First, tell us about this matter.
Tian Keyu: Then I'll start from the beginning. I started interning at ByteDance in 2021. In 2024, a new intern came, and later I felt the whole team was not peaceful. At first, he had some soft forms of occupation—like taking a seat, occupying resources that originally belonged to others, and later he would directly stop other programs.
Actually, he never stopped mine, but stopped those of softer people with more introverted personalities. We reported this to the leader, but it was not resolved. Then I used my way—which was of course wrong—I entered the other person's program and made it stuck or shut it down. In this way, his excess resources would be reclaimed, and others could use them normally.
elsewhere: At the time, there was also a widely circulated version, roughly saying: during your internship, in order to seize cluster resources, you stabbed colleagues and the company in the back, and for at least two months continuously launched malicious attacks on cluster code, including but not limited to xxxxx (listing five charges in total).
Tian Keyu: I must say, a great many of the descriptions here are fabricated. In fact, I personally never lacked GPUs. The biggest motive for doing this was also that I really could not stand seeing A bully other people in our group like that.
First, this was not poisoning, nor did it involve attack and defense. I had GPT help me write a program to stop or freeze the program.
Second, I manually and selectively stopped the other person's program. There was absolutely no such thing as lasting two months.
elsewhere: Only automation counts as continuous. Does manual mean it does not count as continuous?
Tian Keyu: Uh... it does count as continuous. But not in the day-and-night sense described.
elsewhere: How long was it actually, and how many times?
Tian Keyu: In about ten days, about five times in total.
elsewhere: Some say that when you were discovered, you were still continuing this behavior, a bit like playing cat and mouse with the company.
Tian Keyu: Actually, when I was discovered, my permissions were quickly revoked. I could no longer log into the entire system, including Feishu.
elsewhere: There is also a claim that after the incident you said someone else exploited a loophole and shifted the blame onto you, and you even called the police saying you had been slandered.
Tian Keyu: I must say on this point, "saying I framed someone else"—there was absolutely none of that, and it was also the statement I least understood at the time and felt most strongly about. Daring to do it but not daring to own it is not my value system.
I did call the police. Actually, it was because the company came to me and asked whether I knew who posted those posts. I said I did not know, and wanted to see whether the police could investigate who posted them. After confirming with the company, I went to report it.
elsewhere: How do you assess the impact of this matter on ByteDance and on other people? Some say that because of your personal behavior, the research and paper progress of other interns were seriously affected.
Tian Keyu: I guess my superiors must have been implicated to some extent because of this, and I feel very guilty. Also there was public opinion.
Beyond that, I can be certain that my behavior did not affect anyone else's research or papers.
I admit that several times I did have some impact on A's experiments. But I absolutely never touched other people's experiments. There are also rumors that my Best Paper was completed by seizing resources in this matter. That is pure rumor.
elsewhere: Why did this matter eventually escalate into a legal lawsuit?
Tian Keyu: The lawsuit was filed by (ByteDance) at the end of November 2024, and that public statement was issued in October. I think by today I can quite understand the company: internally, ByteDance has always been strict in discipline. Mine was a kind of violation, and with such huge public opinion, the company definitely had to handle it.
elsewhere: Some people call you "Lao Tian."
Tian Keyu: That is already one of the milder terms among the many used to describe me.
elsewhere: What is the term you find least acceptable?
Tian Keyu: "Stabbing colleagues in the back and framing others." Many factual exaggerations or fabrications, I think, are okay; the truth will eventually come out. But the descriptions in the original post that are completely opposite to my character and integrity—I find those very hard to understand.
elsewhere: The overall feeling from listening so far is: you acknowledge that you made a mistake, but believe it was a mistake that was later excessively magnified.
Tian Keyu: I acknowledge this behavior of mine. First, it was definitely a violation of discipline; second, it did indeed waste everyone's time. For example, at the end of July that year, we had an internal meeting, with many people on the Feishu conference. I said a lot that time and explained the whole context. At the end I said: I solemnly apologize to everyone attending this meeting. I think it had a certain impact both on the research level and on the company's operating efficiency.
"Using violence to counter violence"
elsewhere: Many previous interpretations were that your own compute was affected, but you emphasize that in fact it was not.
Tian Keyu: My own GPUs were never reduced.
elsewhere: Other people were affected. What does that have to do with you?
Tian Keyu: There is one detail I remember very clearly. One night, another person and I were the last to leave. After I submitted many tasks, I saw that he had not submitted a single one. I asked him if he was not going to do it anymore. He was just scrolling on his phone there, and he said that even if he submitted, A would shut it down anyway.
elsewhere: Is the stopping behavior of this A you mentioned similar to your behavior?Tian Keyu: To be honest, it wasn't an act, because he had admin privileges, so he was more aboveboard about it.
elsewhere: Even if others were affected, you were just one of the interns.
Tian Keyu: I was the intern who stayed in that group the longest. I think I had a mindset that when I saw this team become like this and the leader at the time didn't actually step up to solve it, I genuinely wanted to do something. Of course, in hindsight, I should have used a more mature way to solve the problem. At the time, I was indeed very impulsive.
elsewhere: But from another perspective, could it be said that your ego was also relatively big? Everyone else didn't take a stance, but you insisted on doing something.
Tian Keyu: I think it wasn't that they themselves had nothing to say. When everyone talked privately at the time, the language was already very intense.
elsewhere: Is there a reason that you felt your so-called big-brother status was being challenged?
Tian Keyu: But a big brother wouldn't bully others. Which big brother would bully other people?
elsewhere: Did the people you "protected" ever express anything to you afterward?
Tian Keyu: At the time, I was quickly fired. But there was one person who was usually very introverted, didn't talk much, and didn't interact with everyone much either, but when I left, he walked me downstairs. I was quite touched at the time.
elsewhere: I saw an online post from your undergraduate days. At the time, your self-evaluation was: not aloof, mild-tempered. Why would a mild-tempered person have this kind of behavior?
Tian Keyu: I think being mild-tempered and having principles can coexist. At the time, I felt that the abuser's behavior had crossed a certain bottom line.
elsewhere: "Abuser"?
Tian Keyu: Yes, I would use such a harsh word. Of course, looking back, I was actually using violence to counter violence. So today I deeply feel that my behavior was very problematic.
elsewhere: For Tian Keyu, was this an accidental event? Or is this just the kind of thing you would do?
Tian Keyu: If you stretch it across a lifetime, some things are inevitable, and some are accidental. I want to call that behavior at the time "naive justice," but it was indeed very immature. For example, if I encounter something while walking on the street, I might handle it in a relatively impulsive way.
Best Paper, Eternal Night, a lifeline
elsewhere: And then came the very dramatic scene: a few days after ByteDance issued its notice, you received that year's NeurIPS Best Paper.
Tian Keyu: Yes. This is the award I most wanted to win during my student days. But at the time, I was facing enormous public opinion pressure and litigation, and my career was already completely bleak, so my first reaction upon seeing this wasn't much joy either.
My first thought at the time was, at least I can still find a job, at least I won't have nothing to eat. It felt like I had grabbed a lifeline.
elsewhere: I heard a large model company had originally given you an offer, and later canceled it?
Tian Keyu: I remember it very clearly. I talked with their co-founder for five or six hours in one go. I felt this was exactly the place I wanted to come to.
elsewhere: So during that period, your sky was all gray?
Tian Keyu: I was afraid I would have to leave this industry forever. For me, this might not just be gray. It was as if the sky was gone, and I had entered an eternal night.
elsewhere: Did you ever have extreme thoughts?
Tian Keyu: For a moment. One day I was alone in the dormitory. I thought about how I had transferred majors to enter this field because I truly loved it, but I might have to leave because of this, and I didn't know what I could do for the next few decades. There was a window in the dormitory, and outside was the street. For a moment, I had that kind of thought.
elsewhere: What pulled you back to reason?
Tian Keyu: My parents both took leave and came to Beijing, and some other family members also came to find me. The company of family was very important to me.
elsewhere: Throughout this whole matter, to whom do you want to say sorry?
Tian Keyu: Even now, I still have very deep feelings for ByteDance. I once wanted to achieve great things at ByteDance after graduation. Every time I think about how things turned out like this with ByteDance, I feel very regretful.
I have always been full of apology toward the company. After all, it was my improper behavior that indirectly triggered many later things. These public opinion storms must have cost the company a lot of resources to deal with. I am very sorry. That also includes the company's lawsuit against me. They also spent money on legal fees and other costs. So I have also repeatedly emphasized that I accept the compensation ordered by the court. I should pay the price for my mistake.
What I have always told the company is not "I was not wrong," but "I am not the great villain in the rumors."
elsewhere: Is there anyone you want to thank?
Tian Keyu: I am very grateful to many people, including ByteDance. I think ByteDance was very good not only in terms of resource support for interns, but also in the atmosphere for communication. At the time, my leader also helped me secure a lot of resources.
elsewhere: Are you a person who has a sense of danger?
Tian Keyu: I generally imagine quite a lot. But back then, I think any normal person, leader, or company could not have predicted the later outcome of this matter. For example, such huge public opinion, such an exaggerated lawsuit, and even A fabricating that our Best Paper was obtained through competition. This is already beyond the range that world models can predict.
The once youngest intern
elsewhere: Before you changed majors, you studied materials science. Why did you choose that major?
Tian Keyu: I was very interested in chemistry in middle school. At the time, I would even run to chemical reagent factories, because some things were not available in our middle school laboratory. For example, there was an experiment called the iodine clock experiment: after you mix three things together, there is a cyclical process, every 8 seconds it changes from blue to yellow, and then back to blue. It was so magical, like a magic trick.
elsewhere: Then why did you switch midway to study software engineering?
Tian Keyu: After I entered university, I came into contact with programming for the first time, and I found it extremely interesting. When I first started my undergraduate studies, I might spend 10 hours a day doing programming-related things. So actually, before I transferred departments, I had already finished the most core courses in it. Basically, I could finish a specialized course in three to five days by self-studying.
elsewhere: What is the joy of programming?
Tian Keyu: Programming is a kind of creation. After you create something, its operation can create something else. In addition, many things in programming are quantitative, such as efficiency and metrics. The positive feedback brought by these numbers is also very strong. Also, programming has a strong relationship with AI.
elsewhere: Did you play games in university?
Tian Keyu: Programming is much more fun than games. And it has many progress bars. Comparing me playing a game dungeon with programming progress is simply incomparable.
elsewhere: Your very special experience is that starting in your sophomore year, you entered SenseTime for an internship, stayed there for three years, and then went to ByteDance for another three years of internship. Why did you intern so early and for so long?
Tian Keyu: When I graduated from high school, I had already observed some companies up close, so I developed an orientation very early: after graduation, I would definitely enter industry, not become a teacher. At that time, for us, SenseTime was also a kind of white moonlight existence.
elsewhere: You should have been the youngest intern at SenseTime at the time?
Tian Keyu: Yes, I was the youngest. The two years at SenseTime were a bit like doing a mini PhD.
elsewhere: Would such early and extensive internship experience make you socialized earlier?
Tian Keyu: I think a more appropriate expression is "academic maturation." Actually, in my sophomore and junior years, I had already published two first-author papers at NeurIPS.
"Double Buff"
elsewhere: Before talking about your startup, I want to ask a question first: Did the ByteDance incident, in some way, actually make you more visible?
Tian Keyu: In practical terms, yes. It brought an especially large amount of attention, and combined with the Best Paper matter, it was like a double buff.
elsewhere: So if you handle this story well, could it become a kind of life asset?
Tian Keyu: Honestly, this kind of attention or heat is definitely beneficial for entrepreneurship. But the more important asset is the growth from the incident, a kind of baptism.
elsewhere: Why was your first investor Meng Xing, a partner at 5Y Capital?
Tian Keyu: At the beginning of last year, I had very few investors in contact, but I think what was special about Meng Xing was that he had founded a company, served as an executive, and also done investment. Many of his perspectives come from the diversity of his experiences and identities. Another point is that he has very strong judgment about technology.
elsewhere: Do most investors lack judgment about technology?Tian Keyu: Top investors should be quite something. But I think what makes Xing-ge special is that he can grasp some fundamentally principled things.
elsewhere: Actually, Meng Xing has invested in many young entrepreneurs who are quite outliers. Each person has a lot of story, and of course yours is definitely the most dramatic.
Tian Keyu: I've heard a bit about it. Later he brought me to meet Richard (Liu Qin). At the time he said something I didn't quite understand at first, but gradually came to understand. He said I have a certain drive in me. I think I cried at the time, tears welling up in my eyes.
elsewhere: Has anyone ever said that your choice of words is very not post-00s?
Tian Keyu: ...Really? OK, OK.
elsewhere: I heard that you also cried the night you were rejected by a fund.
Tian Keyu: That night I did cry a lot. I really liked this fund. In the end it was an online meeting, and at the end there was a chain of questions about this issue (ByteDance), which made me feel despair. Of course now I also understand that more trust should have been built in the first place, and I also didn't do well enough on this matter.
elsewhere: You feel you need to repeatedly prove yourself.
Tian Keyu: It's exactly that repeated self-justification. Especially before the verdict comes down, everyone will believe the side with a bigger say, right?
My world, my model
elsewhere: How would you define a world model? Are you a world model company?
Tian Keyu: We define it as predicting the future. Today there are too many definitions and too few implementations. We set this direction back in '23, focusing mainly on implementation.
I don't care whether our company gets labeled with this tag. I neither actively seek it nor avoid it like the plague. Anyway, we're just building a multimodal foundation model.
elsewhere: Over the past six months, nearly a hundred startups in China have claimed to be world model companies. What's your differentiation?
Tian Keyu: We are the most pioneering of all the routes today. On world models, we've observed the same pretraining scaling law as with language models. This scaling law that hasn't saturated even at nearly a hundred billion parameters is unique.
elsewhere: In the long term, what can world models achieve?
Tian Keyu: The ultimate form is predict everything. For example, if I take 10 billion years ago, and input the state of matter at the very beginning of the Big Bang into a world model, it can simulate forward 10 billion years. Another intuitive example is like Doctor Strange in Marvel: when he was fighting Thanos, he saw many different futures, and in the end only one succeeded.
elsewhere: Why did you choose world models instead of language models?
Tian Keyu: It can be traced back to the best paper, and even back to before '23, when GPT first came out. Many people around me switched to working on language models. But I had no interest in language models. I wanted more to transfer its success to the image and video domains.
So in '23 I posted a little Zhihu post, roughly two sentences: first, we need to learn an autoregressive model from video at scale; second, we first need to do representation, to make an AI's own language.
elsewhere: You say you're not interested enough in language models. Specifically, what are you not interested in?
Tian Keyu: The deeper reason is that I had always been in the self-supervised field. What attracted me to this field is that without humans teaching it, it can spontaneously learn things. For example, after seeing a large number of images, it can vaguely sense the difference between cats and dogs.
elsewhere: Language models can also self-learn.
Tian Keyu: But language is first of all man-made. I prefer distilling nature, not distilling humans.
elsewhere: Let's bring it closer. What will the first stage of intelligence be like?
Tian Keyu: For example, sealed in a small room, just like this bar we're in now. If I throw a big rock at that wall, how those glass bottles should shatter and how they should fall, it can predict very accurately.
Or say when we want to tell fortunes, I toss three copper coins, and in the instant they're in the air, the world model should also be able to predict it.
elsewhere: Has this already been achieved?
Tian Keyu: It used to be very hard, but at this stage we've already achieved it. We're thinking of creating a scenario and letting everyone experience it on site.
elsewhere: How many people do you need to do these things? Or rather, what do you need most?
Tian Keyu: Doing pretraining doesn't need that many people. By the end of this year our target might be just 20 people (on the team). I think the most important thing is collective intelligence, and a soil that can inspire everyone's free play.
elsewhere: If one day a 24-year-old Tian Keyu appeared in your company, what would you do?
Tian Keyu: From a leader's perspective, I might think about how to handle conflicts, or even how to intervene before conflicts happen, to take preventive measures. Of course, whether as an individual or a founder, in doing anything one still needs to handle it more maturely.
Rabbits, MJ, and tenderness
elsewhere: Some people mock your hairstyle, thinking it's very scene-kid.
Tian Keyu: I've had this hairstyle since freshman year. To this day I still get my hair cut at a place downstairs from the Beihang dormitory building.
elsewhere: Do you like little rabbits? (The night before recording, Tian Keyu sent us a video of his little rabbit)
Tian Keyu: On one hand, rabbits are easy to raise, and sometimes it seems very smart, sometimes it seems to have no memory, and it can't speak, giving people a natural feeling of wanting to cherish it.
elsewhere: What songs do you like to listen to?
Tian Keyu: For ones without lyrics, I like listening to Techno. It gives me a very strong sense of rhythm and maintains my flow state while working.
For ones with lyrics, I like Michael Jackson. He is very talented and innovative, and resilient. He is also a very compassionate person, for example he paid special attention to children. And this compassion is in great contrast with his own life experiences.
elsewhere: Whether it's the little rabbit or MJ, you mentioned tenderness. Do you think you have some self-projection?
Tian Keyu: I actually haven't really realized that. The rabbit is something I wanted to raise since high school. For MJ, I still prefer his dazzling talent.
A solemn oath
elsewhere: If one day AGI is achieved, what would you want to do?
Tian Keyu: If that happens, I would switch to researching hibernation. Then I'd want to hibernate myself for 10,000 years, and then see what the world has become.
elsewhere: What if you wake up and find the world is already gone?
Tian Keyu: Then before that I'd first do one thing: make ASI not do anything that destroys this universe.
elsewhere: Why 10,000 years later?
Tian Keyu: I just want to see what technology and civilization have become by then. If hibernation technology is limited and I can only see for 10 seconds, that's fine too. It's already the greatest meaning.
elsewhere: Can you swear that everything you've said here is true?
Tian Keyu: Of course. About the virus, about the attack and defense, it's all baseless.
elsewhere: Anything else? You can get it all out at once.
Tian Keyu: There are also some things I absolutely cannot accept, such as claiming (the ByteDance incident) was done by someone else. I deeply despise this kind of behavior.
elsewhere: What if you told a lie?
Tian Keyu: My research would suffer setbacks for life. Of course, if it's all true, I hope my research will also go more smoothly in the future.