Tian Keyu, a former intern at ByteDance, recently made his first public appearance on a podcast to address the previously reported incident involving a 'poisoning' attack on a large model at ByteDance. After departing from ByteDance, Tian went on to establish a world model company, successfully raising nearly $30 million in funding and achieving a post-investment valuation of $200 million.

In late 2024, Tian was dismissed and subsequently sued by ByteDance for allegedly tampering with code to disrupt model training. The court of first instance ruled that he should pay $500,000 in compensation and forfeit all his internship wages. The case centered on a breach of contract, with no criminal charges filed. Shortly after losing the case, Tian, as the lead author, won the Best Paper Award at NeurIPS 2024 for his VAR paper, marking the first time a Chinese team had received this prestigious accolade.

On the podcast, Tian explained that his dissatisfaction stemmed from a new intern who, without any intervention from leadership, arbitrarily stopped others' programs. In response, Tian used GPT to craft a program that selectively halted the other intern's programs. This action, contrary to online rumors of a two-month-long 'poisoning' or offensive-defensive maneuver, lasted approximately ten days and occurred five times in total. He expressed regret to ByteDance, acknowledged the court's decision, and stated that he was not the villain depicted in the rumors. Tian also believed that the attention garnered had some positive effects on his startup and that he had grown personally from the experience.

Following the podcast's release, Tian's former teammates issued a detailed statement spanning tens of thousands of words, highlighting discrepancies between his claims and the actual facts, and calling on relevant parties to verify the information carefully. Currently, Tian has not yet responded to these allegations.

ByteDance previously clarified that online claims regarding the incident affecting 8,000 GPUs and causing millions of dollars in losses were significantly exaggerated. In reality, the incident only impacted internal research tasks within the commercial technology team and did not affect online services or major large model projects.