AB Kuai Dong

热爱研究、热爱家庭、热爱分享，2016 年入行从业，现居东京。

0Following

1.6Kfollowers

Feed

Recently, ByteDance employees have been almost exploding with anger. Previously, Tian Keyu, who maliciously tampered with code and attacked internal projects at ByteDance, has recently surged back into the trending topics. The reason is that Bloomberg reported these days that after starting his own business, he confirmed receiving nearly 30 million USD in funding, including from well-known domestic institutions like IDG. But the real trigger was that in a podcast, he defended himself as the bullied party. He claimed that when he was interning at ByteDance, a new intern joined the team with admin privileges, who overused computing power and even directly stopped other interns' training tasks. He reported this to the leadership, but nothing was resolved. So he took matters into his own hands and had AI help him write a small program that manually blocked the other party's project runs 5 times over about 10 days to free up extra computing power for others to use. He said he was not short of computing power, and his own tasks were never stopped; he just couldn't stand seeing someone being bullied and denied the online rumors of poisoning or framing others. However, right after this podcast aired, the ByteDance colleague circle exploded. Many ByteDance employees who were in that team at the time, as well as colleagues who helped investigate the issue, came forward to speak. They accused Tian Keyu of almost shutting down several ByteDance projects directly. It was not just 5 times in 10 days as he described, but starting from March 2024, lasting several months, affecting multiple projects (including LlamaGen text-to-image and the later Goku video generation project). ByteDance employees in Beijing, Shanghai, and Seattle took turns investigating but couldn't find the problem. To prevent his malicious actions from being discovered, Tian Keyu not only manually stopped tasks but also altered model weights in shared directories, replaced PyTorch files in the system using sudo, then reverted the modification times, and changed random seeds, etc. This disguised the incident as random failures so no one could tell it was sabotage. At the same time, they also said Tian Keyu was a habitual fraudster within ByteDance. The VAR project he was responsible for used thousands of training cards, but it was later found that it actually ran the open-source PixArt model, not a model he trained himself. To prevent others from detecting this by comparing with the same prompts, he secretly added instructions before the prompts. The model he truly trained performed much worse. As the incident escalates, the number of ByteDance employees publicly accusing Tian Keyu is growing, and this event has become a hot topic internally at ByteDance recently. Many employees joked that if they had known, they would have also sabotaged and gone out to get funding from naive VCs.