A 16,200-prompt land-or-water test for LLMs
A post describes giving an LLM latitude and longitude coordinates as text, then plotting its answers as an image.
TLDR
A post describes asking an LLM “Land or Water?” 16,200 times using latitude and longitude coordinates, then plotting the answers as an image. Its author suggests the models know the answers from “compressing the internet.” The quoted post links to an image it calls “results for all claudes.”
Combined views
166.2K
1 Source, first seen ago
