The Impossible Prompt That's Easy for Humans (teodordyakov.github.io)

Researchers and hobbyists have highlighted an “impossible prompt” that exposes residual blind spots in modern LLMs and multimodal generators: draw an image with two seven-pointed, two eight-pointed and two nine-pointed stars where every star is connected to every other star except its same-sized partner, and none of the connection lines may intersect. The prompt is trivially solvable by a human with pen-and-paper (the author solved it in about a minute), but major models — ChatGPT, Grok, Gemini and Nano Banana Pro — consistently produced incorrect outputs or visually plausible but topologically invalid images. The significance is practical and diagnostic: this task combines discrete counting, planar-graph embedding and spatial/topological constraints, revealing that despite broad advances, current models still struggle with tightly constrained combinatorial geometry and verifiable spatial correctness. Technically, the prompt reduces to finding a non-crossing drawing of a specific constrained graph (a 6-node graph with forbidden same-size edges) — a solvable planar-embedding problem for humans, but one that generators often mishandle. The failure mode suggests value in benchmarks that mix symbolic graph constraints and visual rendering, and argues for integrating explicit graph algorithms, neuro-symbolic reasoning or drawing tools into multimodal systems to guarantee discrete, verifiable constraints in generated images.