Beijing University of Posts and Telecommunications, Nanyang Technological University, and the Allen Institute for Artificial Intelligence have introduced FunQA, a high-quality video question-answering dataset designed to test and enhance AI models' comprehension of counterintuitive video content. This dataset includes 4,365 counterintuitive videos and 3.12 million text-based question-answer pairs, covering areas such as humor, creativity, and magic. FunQA poses deep challenges to models' understanding abilities through tasks designed around timestamp localization, detailed descriptions, and counterintuitive reasoning. Researchers note that current models perform poorly on FunQA tasks and require improvements in model size, data quality, and training strategies. The release of this dataset is expected to drive advancements in computer vision research.