The MNIST dataset is an interesting space for studying the fundamentals of learning algorithms. The dataset requires tens or hundreds of examples to reach 95% accuracy, and it is still highly relevant in the pursuit of AGI.