MAIN ARTICLE
A new profession in the model race
Fangcun AI describes itself as a safety company jointly incubated by Tsinghua University’s College of AI and Institute for Interdisciplinary Information Sciences. Based near the university in Haidian, an established center of academic and technology research, it sits within an unusually dense talent network. South China Morning Post reported on November 18 that the company launched in January, employs more than 40 people and intends to reach roughly 100 by the end of next year. That expansion is striking in a Chinese market where independent commercial evaluations are still taking shape.
Model releases are usually celebrated for reasoning scores, coding performance and linguistic fluency. Yet those benchmarks tell the public relatively little about harmful instructions, persuasive manipulation, fabricated information or the behavior of AI agents given permission to operate tools. Independent evaluation requires a different discipline: clearly defined threats, repeatable test procedures, transparent limitations and the freedom to report disappointing findings even when a developer would prefer silence.
Chief executive Luke Li Xiaojian has emphasized the importance of disclosure. Speaking to AFP in a report published November 14, he suggested that DeepSeek was doing more work on safety alignment than it publicly explained. That assessment should not be mistaken for a declaration that the system is safe. It points instead to a gap between safeguards implemented by developers and evidence outsiders can actually examine.















