Tian Keyu responds to ByteDance “poisoning” controversy; rebuttal document continues to accuse him of code tampering and falsifying a model demonstration

AT A GLANCE

Tian Keyu admits that he interfered with ByteDance experiments but denies attacking them for months; the relevant allegations remain disputed.

Article

Tian Keyu speaks for the first time about ByteDance "poisoning," saying 10 days, during which he stopped programs 5 times; the other side released code in rebuttal and also accused him of falsifying a model demonstration

Beating news flash,AI Tian Keyu, who was dismissed for disrupting ByteDance model training, responded in detail for the first time to the controversy from 2 years ago on the podcast “elsewhere.” He said that, at the time, a new intern frequently exceeded allocated computing resources and even halted other colleagues’ experiments. After receiving no response from superiors, he chose to “fight fire with fire.” He admitted to using ChatGPT to write a program and manually interfering with the other party’s experiments 10 times over approximately 5 days, but denied attacking the training cluster for months. He stressed that he was not competing for computing resources and acknowledged that his actions were improper.

However, a rebuttal document titled “The Perspective of Intern A Has Arrived” presented a completely different account. It lists multiple purported internal ByteDance code records and alleges that Tian Keyu’s disruptive conduct began in 2024 year March and continued for months, involving multiple projects. He allegedly first injected random noise into the model parameters of the LlamaGen project, then modified underlying PyTorch code used by the Goku video-generation project, causing model training outputs to fail and programs to terminate randomly. These failures were extremely difficult to reproduce. In one instance, 100 machines ran the same program, but only one malfunctioned and returned to normal after restarting. Teams in China and the United States took turns investigating for an extended period, until the anomalous code was discovered at the end of July. During this period, Tian Keyu allegedly adjusted his methods based on colleagues’ investigation progress.

A further allegation concerning research integrity is more serious. The document alleges that Tian Keyu was responsible for a GPU text-to-image project trained using thousands of VAR, but that the results he presented to colleagues for testing actually came from the already open-sourced PixArt model. He allegedly secretly added “City” before the Prompt entered by users, making it difficult for colleagues to identify the relationship between the two models using identical prompts. Testers had previously noticed that generated images often inexplicably contained urban buildings, but could not determine why. A subsequent team review found that the actual VAR model was far from achieving the demonstrated results. The document also lists several researchers from the relevant projects, saying they are willing to testify about the portions they witnessed firsthand.

Key points

Tian Keyu says he manually interfered with the other party’s experiments 10 times over approximately 5 days and denies attacking the training cluster for months.
The rebuttal document says the alleged disruption began in 2024 year March and involved code or parameters in the LlamaGen and Goku projects.
The document also alleges that Tian Keyu passed off the already open-sourced PixArt model as the results of the VAR project.
ByteDance previously confirmed that Tian Keyu interfered with internal research projects by writing and modifying code, and dismissed him.
Tian Keyu says the first-instance judgment awarded approximately 500000 yuan in damages, and the case remains in second-instance proceedings.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This is a dispute over whether research code was maliciously modified and whether the model demonstration was genuine. Tian Keyu admits to some interference but denies the sustained attacks described in the rebuttal document; the document’s other allegations have not yet been independently verified by evidence.

Why it matters to readers

The incident involves model-training stability, code security, and the credibility of research results, and also shows that internal permissions and project investigations may face risks.

Model training relies on large amounts of code, data, and computing resources. If code or parameters are altered, results may be erroneous, training may be interrupted, and reproduction may even become difficult.

Risks and unknowns

The evidence has not confirmed whether the internal code records listed in the rebuttal document are authentic.
Regarding the allegation that the PixArt model was passed off as VAR results, the evidence provides no independent verification outcome.
It remains unclear how the second-instance court will ultimately determine the amount of damages and the relevant facts.
Tian Keyu and the rebuttal document give markedly different accounts of the duration and scope of the interference.

Tian Keyu responds to ByteDance “poisoning” controversy; rebuttal document continues to accuse him of code tampering and falsifying a model demonstration

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center