Tian Keyu speaks out on ByteDance 'poisoning' for the first time, says program was halted only 5 times in 10 days; the other side retaliates by posting code, also accuses him of faking model demo
Beating AI News Flash: Tian Keyu, who was dismissed for sabotaging ByteDance's model training, responded in detail for the first time to the controversy from two years ago on the podcast "elsewhere." He said that at the time, a new intern frequently used excessive computing power and even halted other colleagues' experiments. After reporting this to his superiors without result, 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 5 times over about 10 days, but denied attacking the training cluster for months, stressed that he was not trying to compete for computing power, and acknowledged that his approach was inappropriate.
However, a rebuttal document titled "Intern A's Perspective Is Here" gives a completely different account. The document lists multiple code records allegedly from inside ByteDance, accusing Tian Keyu of starting his sabotage in March 2024, lasting for months and involving multiple projects. He first mixed random noise into the model parameters of the LlamaGen project, and then tampered with the underlying PyTorch code used by the video generation project Goku, causing incorrect model training results and random program interruptions. These failures were extremely difficult to reproduce, and there was even a case in which 100 machines ran the same program, only one malfunctioned, and it returned to normal after a restart. Chinese and U.S. teams took turns investigating for a long time because of this, and only at the end of July that year did they discover the abnormal code. During this period, Tian Keyu would also adjust his attack methods based on his colleagues' investigation progress.
More serious is another allegation of research integrity. The document claims that Tian Keyu was then responsible for a VAR text-to-image project trained on more than a thousand GPUs, but the results shown to colleagues for testing actually came from the already open-sourced PixArt model. He also secretly added "City" before the user-input Prompt, making it difficult for colleagues to discover the connection between the two models through the same prompt words. Previously, testers had found that city buildings often appeared inexplicably in generated images, but they had never been able to find the reason. A later team inspection found that the real VAR model was far from reaching the displayed effect. At the end of the document, several researchers on related projects are also listed, saying they are willing to testify about the parts they personally experienced.
ByteDance confirmed as early as 2024 that Tian Keyu had maliciously interfered with the training of internal research projects by writing and tampering with code, then dismissed him and sued for 8 million yuan in compensation plus reasonable expenses. Tian Keyu revealed this time that the first-instance court awarded about 500,000 yuan in compensation, while the other claims were rejected, but ByteDance has already appealed and the case is still in the second instance. He said he accepts the compensation and apologizes for his actions, but insists that the outside world has seriously exaggerated what he did.
