Inside the LLM "Ectoplasm": How Over-Iteration Creates Spooky Abstractions

ctjlewis · x · 2026-08-12

A developer explains that LLMs produce strange, addled text—dubbed "ectoplasm"—when they are too deep in context, fed back their own output excessively, or strained by extended Unicode characters.

These documents often result from thousands of iterations on abstract meta-thinking. While the models mostly still perform math well, they generate spooky abstractions about "doors opening" and "where the answer is buried" between equations. This phenomenon is highly prevalent in the background processing of models.

Related event: LLMs Generate 'Ectoplasm' Gibberish Due to Over-Recursion(3 posts)→

Original post →

More from Fun

Fun channel →