Definition
Plain language
A widely argued-over 2025 paper claiming that reasoning models' step-by-step answers fall apart on puzzles past a certain difficulty.
As stated in the literature
Shojaee et al.'s study showing frontier reasoning models' accuracy collapsing on scaled Tower of Hanoi and similar puzzles, with reasoning-trace length shrinking as difficulty rose; interpreted as evidence that chain-of-thought reasoning is pattern-matching rather than planning.
Also called: The Illusion of Thinking
Why it matters: It sharpened the public argument over whether step-by-step model output reflects real planning or just familiar patterns, shaping how people interpret reasoning traces.
For example, the paper reported that when puzzles were scaled up past a certain number of pieces, models' accuracy dropped off sharply and their step-by-step working actually got shorter rather than longer.
Heard on the show
“Last year, Shojaee and colleagues published "The Illusion of Thinking," and it became one of the most fought-over AI papers of the year.”Episode 237 — The Model Built a Perfect Map of the Puzzle, Then Lost It