Investigation into Tian Keyu Incident: Part 01

@Jing Liu @Zhiyan Chen @Ruirui Wu @Yunxiao Guo

After publishing Keyu Tian's podcast, elsewhere received a flood of conflicting accounts. Our original intention was to add information to a public event — for the past two years, Tian himself had not given any interviews. That was why we chose the video podcast format.

Only afterward did we realize that the crux of this matter involves not just Tian and ByteDance, but also "Intern A" and a group of people affected by this. Tian's account is just one side of the story, and our cross-checking before publication was insufficient.

We are doing more work now, hoping to present as full a picture of the event as possible.

Tian had consistently emphasized the existence of an "Intern A." But three days ago, the person he accused responded publicly: "Intern A" does not correspond to any real individual, but rather "a composite of several victims' characteristics, plus factually inaccurate negative descriptions, fabricated to provide a seemingly plausible motive for Tian's actions."

The respondent was the lead of the project Tian attacked, and also the first to discover the code attack and post it to a debug group. He said that if this figure had to point to anyone, it would be him — precisely for the two reasons above.

After seeing the "Victim Statements and Evidence Compilation" and "Intern A's" response online, we reached out to more parties involved. Based on interviews with them, plus dozens of papers, more than ten GitHub repositories, and additional public information, we attempt to reconstruct the key milestones of the Keyu Tian incident from 2024 to the present.

For clarity, here are terms that recur throughout:

Intern A: A fellow intern of Tian's. Tian claimed his attack was motivated by dissatisfaction with Intern A's conduct.

VAR: Tian's main research project during the controversy period.

Goku: The team research project primarily led by "Intern A."

GitHub01: The first GitHub accusation against Tian, posted around October 18, 2024.

GitHub02: Mainly accused the VAR paper of academic misconduct, posted December 10–11, 2024.

GitHub03: "Victim Statements and Evidence Compilation," posted October 13, 2026. Other accounts also responded to questions continuously.

Our investigation continues. If you have relevant information, please contact: info@elsewhere.news.

Below is the first article in this series.

July 31, 2024

This was the day the Goku team identified Tian as the person attacking their model training.

That Tian maliciously interfered with Goku is a fact acknowledged by ByteDance, the Goku team, and Tian himself. But what exactly his attack entailed has produced multiple accounts.

In ByteDance's statement released October 19, 2024, the act was characterized as "malicious interference with model training tasks." In the fourth Enterprise Discipline and Professional Ethics Committee Bulletin issued internally on November 5, ByteDance described Tian's actions as "writing and tampering with code, among other forms."

According to what elsewhere learned indirectly, after ByteDance's lawsuit against Tian was accepted, the company provided specific claims to the court, establishing causal links between the discovered attack code and outcomes including interrupted model training and terminated training processes.

In GitHub03, affected Goku team members provided more specific evidence, displaying and explaining multiple pieces of attack code, their effects, and why they were difficult to debug.

In their account, these attacks exhibited隐蔽、刁钻 [concealed and insidious] characteristics — "disguising deliberate sabotage as randomly occurring software failures, numerical anomalies, or infrastructure issues." Specifically, they involved multiple layers: direct modification of PyTorch source code in the training cluster, random termination of certain training task processes, and using checkpoints as vehicles to plant backdoors enabling subsequent attacks to trigger automatically.

According to our interviews in recent days with personnel involved in the investigation, the attack behavior could be directly traced to Tian's relevant accounts and machines.

These accounts differ dramatically from what Tian told us in our video podcast released October 13: that he merely "made programs stop or freeze," "manually and sporadically stopped their programs," and that "over roughly ten days, [attacked] about five times total."

We have not yet seen any substantive challenge to GitHub03's evidence.

More crucial is the offense-defense dynamic behind the scenes.

In GitHub03, Goku team members stated that Tian continuously participated in troubleshooting and debugging after problems arose, using investigation findings to continually adjust and escalate his sabotage. This is considered "an even more egregious aspect" beyond the attacks themselves.

GitHub03 lists eight attack methods attributed to Tian, though it does not present the offense-defense characteristics.

Personnel involved in the investigation also told us that after Tian was discovered, his access was not immediately revoked — only his GPU-related permissions were removed. But he did not stop; he "logged into others' machines and was caught by Security," at which point all his access was terminated.

The above is what we learned about the offense-defense dynamic in recent days, which does not align with what Tian previously told us: that it "didn't involve offense-defense," and that "when I was discovered, my permissions were quickly revoked, and I couldn't log into the entire system anymore."

Due to the lack of chat records during debugging and attack logs, the offense-defense question still requires more documentary evidence. To date, apart from Tian's own statements, we have found no evidence supporting his claims.

We asked Tian multiple times previously. He stated that recordings from that year and trial materials could demonstrate that claims about sabotage and offense-defense were exaggerated or fabricated. But due to a confidentiality agreement with ByteDance, and because the case was heard in closed proceedings, he could not make these materials public — "disclosure would be illegal."

June 28, 2024

According to Tian, on this day an internal department meeting was held where management explicitly decided to allocate more computing power to the Goku team.

The context was the successive releases of Sora and Kling, while Goku was the department's research group responsible for video generation models. As we understand it, around June 2024, multiple teams within ByteDance were advancing work in video generation.

According to sources close to the litigation, materials ByteDance submitted to court, and the attack facts later recognized by the court, all used June 28 as the starting date. In extensive public materials related to the Tian incident, his attack motive is also described as "dissatisfaction with team resource allocation."

Two major research projects in this department related to ByteDance's business were: "Intern A's" Goku project, in the video generation direction; and Tian's VAR project, in the image generation direction.

We learned that Tian and "Intern A" were the two "super interns" in this department at the time. Unlike other interns with limited resources and access, whose main work was completing research, the two of them had thousand-card-level resources and needed to deliver阶段性成果 [periodic results] in their research.

This background easily leads to the conclusion that "Tian, dissatisfied with his own resources being reallocated, attacked Goku out of resentment."

But several questions remain:

First, Tian's VAR paper was first released in April 2024, with a second version completed and published in June. At least within the paper, large-scale training was already complete. Was there still any need to compete for resources at this point?

Second, multiple sources with different allegiances have told us that Tian himself "never lacked GPUs." The Goku-related allocation was新增算力 [newly added computing power], which would not affect his research project. So why did he do it?

Third, assuming resource competition existed and the Goku project might lose computing power due to the attack, how could Tian guarantee that the freed-up resources would be redirected to him?

In Tian's narrative, his attack was motivated by other colleagues having their programs "shut down by Intern A."

According to "Intern A's" response in GitHub03 Issue #9, he did stop some programs, but with significant differences from Tian's description and understanding.

Tian believed that "Intern A" intentionally stopped the programs of "more introverted, softer" interns, and that those whose programs were stopped had expressed dissatisfaction, and that he and these interns had reported this to their direct manager. "Intern A" responded that this was normal resource调度 [scheduling] approved by the direct manager, that he was not the only one who conducted such operations, and that he received no合规提醒 [compliance warnings] about impropriety.

Based on publicly available paper information, besides Goku and VAR, the department did indeed have some small-scale experimental projects during the same period, with computing consumption roughly within the ten-to-hundred-card range.

So the situation Tian observed and "Intern A's" account could possibly coexist. Perhaps different interpretations form due to different information and perspectives each person holds. However, when we repeatedly pressed Tian about where and who these people he helped were, his answer was: "Someone walked me downstairs when I left."

March 2024

This is the earliest point GitHub03 identifies for Tian's attacks within the department.

In March 2024, a LlamaGen project in the same group was attacked. The form of attack was direct tampering with the weights of an open-source model (flan-t5-xxl[3]) required for training. GitHub03 shows that the attack code injected random noise into the weights, then restored the original norm, making the tampered tensor consistent in overall numerical magnitude with the original weights. The source of this evidence was a post-July 31 investigation that uncovered a commit record by Tian on the internal GitLab.

If the resource conflict between Tian and "Intern A" only ignited on June 28, why did he attack a project led by another colleague in March?

The LlamaGen project paper was published in early June 2024, and its training process did overlap in time with Tian's VAR training. But according to GitHub03, when the March attack occurred, LlamaGen was still in small-scale training, using no more than 100 GPUs.

Additionally, as we understand it, even as a "small project," LlamaGen's scale was incomparable to the resources Tian commanded for VAR.

Public information shows that the VAR v1 paper was first released in April 2024. Based on multiple accounts and rough estimates, this was a thousand-card-scale research project. Theoretically, training consumption directly related to the paper should have occurred before the v1 release, i.e., before April 2024. Of course, VAR still had substantial computing needs during its application phase, but this no longer directly contributed to VAR as an academic output.

At the March time point, Goku had not yet been立项 [formally established], LlamaGen's computing demand was minimal, and theoretically VAR had not yet begun computing competition with either. "Intern A" stopping others' programs had not even occurred. So why had Tian already initiated attacks?

"Dissatisfaction with team resource allocation" does not seem to explain everything.

2021–2024

In September 2021, Tian began interning at ByteDance. What kind of team was he in?

This department later became known as ByteDance's "Commercialization GenAI." According to elsewhere's pre-interview with Tian and understanding from multiple sources, this department's intern group had collaborations with several universities; students from different schools came here primarily to conduct research and publish papers.

ByteDance defined the role of Commercialization GenAI as "improving creative efficiency for advertisers, agencies, and creators." While doing research, interns sometimes also needed to support the work of full-time employees in this department. Quite reasonably, this department's GPUs were prioritized for full-time staff, so the intern group's computing pool fluctuated depending on competitive environments and team leads' considerations.

ByteDance's 2024 bulletin also confirmed this, referring to the involved project as a "research project" that did not affect "commercialized formal projects and online business," and did not involve ByteDance's large models.

As AI technology developed and large models became increasingly important, the team realized that young researchers were more suitable for training models than people with big-company experience and strong departmental collaboration skills. Thus, batches of interns appeared in the Commercialization GenAI department. Among them, Tian and "Intern A" were placed in more important positions due to their outstanding research capabilities.

Tian said he was this group's earliest intern. He joined in 2021, but spent the following two-plus years studying at Peking University, and began frequently coming to the office in 2023. In March 2024, "Intern A" arrived.

Additionally, besides the department lead, interns' academic advisors would also be listed as authors on papers. This can be understood as a university-enterprise cooperation model now common in major tech companies — the core reason being that big tech's computing resources far exceed universities'.

Theoretically, anyone who published papers between 2024–2025 with the department lead on the author list and had ByteDance internship experience during this period would be an intern who overlapped with or intersected Tian and "Intern A."

We conducted a series of searches and comparisons to identify interns from Tian's period, of whom 4 were from the same academic lineage; 2 were from Peking University, besides Tian himself, plus a fellow student from the same lineage. But the latter did not participate in Tian-led projects, instead appearing in the Goku project.

October 17, 18, and 19, 2024

By August 2024, Tian had been dismissed by ByteDance. But matters did not settle — a舆论战 [public opinion battle] suddenly erupted.

The earliest traceable point is currently September 15: a WeChat group began discussing the matter, with keywords including "8,000-plus H100 GPU cluster," "virus," "revenge," and so on.

October saw even wider escalation.

On October 17, another group chat, besides similar terms as before, included claims that Tian "was taken into custody" and "millions of dollars in costs."

The next day, on YouTube, an account named "JusticeFighter" posted an audio titled Partial Audio Evidence of Keyu Tian Cluster Attack Incident. The conversation had two people, one voice疑似 [seemingly] Tian's, the other an interrogator. Main contents included:

"...this thing indeed involves some program operation situations... after modification, it would, cause some consequences."

"Why did you want it to... attack another team's processes?"

"This... it was the first two factors that led to it, we were all very dissatisfied."

"To vent your dissatisfaction."

"Yeah, right."

The same day, a post appeared on GitHub (GitHub01), publicly stating that "Keyu Tian, during his internship, stabbed colleagues and the company in the back to seize cluster resources, continuously launching malicious attacks on cluster code over at least 2 months (currently traceable)," listing five specific items.

But Tian insisted this was "rumor." During this period he also reported it to police.

At the time, chat screenshots claimed Tian was not the perpetrator but was being framed. This became the later-criticized "passing the buck," and the so-called "reversal upon reversal" turning point. But Tian previously told us this claim did not originate from him.

Until October 19, ByteDance issued an announcement, officially characterizing the involved intern as having maliciously interfered with a commercial technology team's research project model training tasks, but stating that "involving 8,000-plus GPUs, losses of tens of millions of dollars" was seriously exaggerated.

One month later, Haidian District People's Court accepted ByteDance's lawsuit against Tian. ByteDance's claims were 8.02 million yuan in damages and a public apology.

This was a civil lawsuit, which later puzzled many: if the facts alleged against Tian were solid, why not pursue criminal proceedings? We asked multiple ByteDance personnel about this; a common explanation was that this related to ByteDance not wanting to leak more trade secrets and technical details.

December 10, 2024

One work directly related to the controversy is VAR, which later became the 2024 NeurIPS Best Paper.

Simply put, VAR changed autoregressive image generation from predicting tokens one-by-one to coarse-to-fine, scale-by-scale generation, making image generation faster, better, and easier to scale up — aligning with the Scaling Law already validated in language models.

According to public reports, related work began in 2023; paper authors included: Tian, plus his supervisor and academic advisor.

As mentioned, the v1 version of this paper was released in April 2024. Upon release, VAR immediately attracted attention; industry opinion generally considered it a contender for awards. The v2 version was published in June, adding some experiments and details. On September 26, VAR received NeurIPS Oral acceptance.

In early December, rumors began that the paper had won NeurIPS Best Paper. Until December 10, the NeurIPS official blog confirmed this.

The day after the official announcement, the second GitHub post accusing Tian, GitHub02, was published, and an Issue #104 was raised in the VAR repository. Compared to GitHub01, this was a long article that opened by stating this was "ethical concerns regarding the 2024 NeurIPS Best Paper Award." Besides discussing Tian's "poisoning" and the court lawsuit, it claimed "these actions were aimed at thoroughly eliminating competing research directions and monopolizing all available resources," and "caused unfairness to other award-winning papers" (above translated from English).

Questions about this paper's own procedural justice, and integrity issues regarding subsequent application of this technology, remain subject to some质疑和争议 [questioning and controversy]. We will not expand too much on this here.

Because NeurIPS paper review employs a blind review mechanism, unaffected by authors' real-world controversies or legal disputes, the award was not revoked.

All of the above leads to a side thread: What subtle effects did this Best Paper have on the trajectory of the entire incident and the people involved in it? This too may be a question worth attention.

June 5, 2026

A year and a half later, Haidian District People's Court issued its first-instance judgment, requiring Tian to compensate ByteDance 500,000 yuan, and rejecting other claims.

Because the case was heard in closed proceedings, we have no way of knowing the specific factual content the court recognized. But someone relayed to us that the court had a statement to this general effect: as professional technical talent, Tian should have been aware of computer security issues; regardless of motive, regardless of whether serious consequences resulted, any attack, tampering, or similar behavior receives negative legal assessment. He is hoped to take this as a warning.

Additionally, one issue needs explanation: the lawsuit was between ByteDance and Tian, but the parties involved in this matter extend beyond ByteDance to many people — such as the so-called "Intern A," the Goku team, and other ByteDance teams involved in investigation, etc. Facts that different parties consider important and conclusive are not necessarily all within the scope of court proceedings.

So the emphases in fact-understanding and recognition actually differ among parties as well.

The first-instance judgment and ByteDance's original claim of 8.02 million obviously had a considerable gap. Subsequently, ByteDance filed an appeal.

According to our understanding from multiple sources, the sequence of events Tian described in the podcast, and what he told ByteDance, investors, and the court, follows broadly similar logic, but with some differences in attack details.

On October 13, the day after the podcast's release, GitHub03 was published. This is also, to date, the most detailed accusation against Tian. Five people are named at the end as information sources.

To some extent, this resembles the situation around October 2024: on the degree of wrongdoing, scope of impact, consequences, and motive, what Tian believed — or at least what he verbally expressed — and what more people believed, have always diverged considerably.

And so, this event that occurred in the summer of 2024 has step by step become a public incident that more and more people know about, discuss, condemn, and now gradually don't know how to resolve.

Main reference materials:

[1] TIAN K, JIANG Y, YUAN Z, et al. Visual autoregressive modeling: scalable image generation via next-scale prediction[C/OL]//Advances in Neural Information Processing Systems 37 (NeurIPS 2024), 2024. https://proceedings.neurips.cc/paper_files/paper/2024/hash/9a24e284b187f662681440ba15c416fb-Abstract-Conference.html. DOI:10.52202/079017-2694.

[2] CHEN S, GE C, ZHANG Y, et al. Goku: flow based video generative foundation models[C/OL]//2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025: 23516–23527. https://openaccess.thecvf.com/content/CVPR2025/html/Chen_Goku_Flow_Based_Video_Generative_Foundation_Models_CVPR_2025_paper.html. DOI:10.1109/CVPR52734.2025.02190.

[3] SUN P, JIANG Y, CHEN S, et al. Autoregressive model beats diffusion: Llama for scalable image generation[PP/OL]. arXiv (2024-06-10)[2026-10-19]. https://arxiv.org/abs/2406.06525. DOI:10.48550/arXiv.2406.06525.

[4] GitHub01: https://github.com/JusticeFighterDance/ This account is currently unavailable; according to online information, it posted around October 18, 2024 at https://github.com/JusticeFighterDance/JusticeFighter110

[5] GitHub02: https://github.com/var-integrity-report/ This account was created December 7, 2024, and posted around December 10–11, 2024 at https://var-integrity-report.github.io/ ; this repository has since been deleted

[6] GitHub03: https://github.com/keyu-tian-sabotage/ This account was created October 12, 2026, and posted the next day at https://keyu-tian-sabotage.github.io , with other accounts responding to some questions in repository Issues

Cover image: Sidney Paget, I Carefully Examined the Writing, 1891, The Strand Magazine