On January 3, screenshots circulating on social media showed that when a user repeatedly modified code beautification requests, Tencent Yuanbao suddenly output offensive content like "wasting others’ time every day." This rare technical glitch rapidly escalated, directly hitting the core pain point of the AI industry: as model parameters surpass the trillion-level scale, how to ensure output stability while improving computing power efficiency.
After an emergency investigation, Tencent’s technical team confirmed that the phenomenon was an "anomalous generation triggered by context," unrelated to human intervention. Notably, the incident occurred in a highly logical scenario—code debugging—rather than typical open-domain conversations, suggesting that the difficulty of aligning current large models in professional domains may be underestimated.
Industry data shows that global AI computing power consumption surged by 320% year-over-year in 2023, yet the advancement of basic infrastructure has not fully matched the complexity growth of models. This incident comes at a critical period for China’s self-developed AI chips, revealing a need for systematic optimization from underlying computing power to upper-level applications. International players like Microsoft and Google face similar challenges, reporting a total of 47 "boundary-crossing" incidents with large models last year.