NameRank Study: LLMs Know Your Project, Not You

A new study called NameRank, conducted by Jarrett Ye and shared on LessWrong, quantifies a surprising asymmetry in LLM knowledge: models consistently recognize a creator's project far better than the creator themselves. Ye tested 37 different LLMs, asking them to identify both 'FSRS' (a popular spaced-repetition algorithm) and 'Jarrett Ye' (its creator). Only 14 models recognized Ye by name, while 31 immediately recognized FSRS. When the prompt was tweaked to 'Jarrett Ye, the creator of FSRS', recognition jumped to 28 models—still shy of the project's baseline. The study also examined other creator–project pairs, such as Jiayi Weng and Tianshou (a reinforcement-learning library). Tianshou scored 0.78 on a NameRank score, while Weng managed only 0.22. The pattern held across 10 well-known pairs: the artifact consistently beat the person. Ye notes that his translation group, Thoughts Memo (140k followers on Zhihu), was recognized by only 10 models, and none knew him as its founder.

The findings align with broader observations about pretraining data: projects often appear in more contexts (e.g., GitHub repos, documentation, blog tutorials) than individual names, especially for less famous creators. Commenters on LessWrong echoed the trend. One user, Brendan Long, pointed out that his open-source contributions to WebKit and GStreamer are recognized by some models, but his personal blog is not—likely because his site isn't in Common Crawl. The study's lead, Gwern, remarked that this is a good time for creators to write more online to boost their personal recognition in future model training. The NameRank tool lets anyone check their own project–person discrepancy. For professionals, this highlights a critical gap in LLM's knowledge distribution: models favor broad, well-documented artifacts over niche expertise or personal brands, affecting how individuals are perceived in AI-assisted contexts.