A new NameRank experiment is sparking discussion about what large language models actually remember. The results suggest that well known software, libraries, and public projects are often easier for AI systems to identify than the people who created them, even when those creators are closely tied to the work.
The findings have prompted developers and researchers to question how modern AI models absorb knowledge from the web and whether personal recognition depends less on expertise than on how frequently a person's name appears across public datasets.
What the NameRank experiment found
The experiment compared how 37 large language models responded when asked about specific creators and their projects. One of the clearest examples involved FSRS, a spaced repetition scheduling algorithm developed by Jarrett Ye.
According to the test:
- 31 out of 37 models recognized FSRS.
- Only 14 out of 37 identified Jarrett Ye by name.
- When the prompt included the phrase "the creator of FSRS," recognition increased to 28 models.
The same pattern appeared elsewhere. Ye also founded Thoughts Memo, a translation community that has translated roughly 2,000 articles and built an audience of about 140,000 followers on Zhihu. Despite that reach, only 10 models recognized the organization, and none correctly identified Ye as its founder.
Why artifacts often outperform people
The results are not especially surprising to many AI researchers. Software names, libraries, frameworks, and products tend to appear repeatedly in documentation, tutorials, GitHub repositories, issue trackers, and technical discussions. The creators behind them may receive only occasional mentions.
One commenter compared this to everyday human memory. Many programmers can instantly recognize the names of hundreds of software projects, but far fewer could accurately name the original authors. The same effect exists outside technology. Actors often become household names, while audiences struggle to remember every film they have appeared in. In other cases, a famous movie may outlast public recognition of its director or writers.
This suggests that frequency of mention, rather than importance, heavily shapes what language models learn during training.
Training data may explain the gap
Several developers connected the findings to the sources used for AI pretraining. Public datasets such as Common Crawl, open source repositories, technical documentation, and community forums expose models to artifacts repeatedly, while personal websites or independent blogs may appear less often.
One developer noted that although their website allows unrestricted crawling through permissive robots.txt settings and a machine readable CC0 license, AI systems still appeared to know their decade old contributions to projects like WebKit and GStreamer better than their more recent writing.
Another observation was that public code hosted on GitHub may receive stronger representation through datasets such as The Stack, while privately hosted writing may remain relatively invisible if it lacks backlinks or broad web distribution.
Recognition is not only about expertise
The discussion also highlighted an important statistical point. If researchers intentionally begin with a highly recognizable project, it is reasonable to expect the associated individual to be less widely known. One cited example compared a reinforcement learning library with its sole creator, where the software achieved a much higher recognition score than its author.
This reflects a broader pattern in knowledge systems. Products can accumulate references across countless documents, while the creator's name appears only in launch posts, interviews, or project histories.
For professionals hoping AI systems recognize their work, simply publishing high quality content may not be enough. Visibility across widely crawled and frequently referenced websites appears to matter just as much.
What this means for AI and the open web
The conversation raises broader questions about how future AI models should ingest information. Some participants argued that AI companies could offer an opt in system allowing authors to explicitly request inclusion of their public writing in future training datasets. Others pointed out that discoverability on the open web still depends on traditional signals such as backlinks, citations, and public accessibility.
As language models become a common way to search for information about people and technology, recognition may increasingly depend on how knowledge spreads across the internet rather than who originally created it. Future benchmarking efforts like NameRank could help track whether newer models become better at connecting influential work with the people behind it.