srush/GPU-Puzzles ? reverse-engineered prompt

Build me an interactive notebook that teaches beginner CUDA through a series of small puzzles, like the GPU Puzzles project.

I want it to start from very simple examples, like mapping over an array, adding two arrays, handling bounds checks, working with 2D inputs, broadcasting, and then moving into blocks, shared memory, reductions, and other core GPU ideas. The experience should feel hands on, with each puzzle showing a short explanation, a place to fill in the kernel, and a way to test the answer right away against the expected result. Use Python code that runs like CUDA, so it’s easy to experiment with, and make it friendly for someone using Colab with a GPU turned on.

Please include the helper code and any visuals needed so the puzzles are easy to follow, and make sure the notebook can be run top to bottom without much setup. If you need to look up current docs online, go ahead.