Current AI systems generate tokens sequentially without stepping back to plan, unlike humans who think about what they want to cover in a paper, how to organize it, and how to summarize it; this limitation causes errors on complex problems like Sudoku puzzles where the first move affects all subsequent moves.