LLMs Can Get "Brain Rot": A Pilot Study on Twitter/X
Researchers tested if continually training large language models (LLMs) on junk text from Twitter/X causes lasting cognitive decline. Using metrics like engagement and semantic quality, they found that exposure to junk data significantly reduced LLMs' reasoning, long-context understanding, and safety capabilities. The study suggests that the social popularity of content is a stronger driver of this 'brain rot' effect than just its length.