Viewing LLM inference as a séance: how much of a model's output comes from the dead

NYCounihan · x · 2026-10-06

Developer NYCounihan offers a playful perspective: if AI is a next-token predictor whose weights come from humanity's corpus, a chunk written by people now deceased, then inference can be seen as speaking with the dead. He notes that if training data could be traced to weights, you could measure the share of inference coming from dead people—high for archaic words like "hornswaggle", low for Python code.

Original post →

Viewing LLM inference as a séance: how much of a model's output comes from the dead

More from AGI Musings