ml-explore/mlx ? reverse-engineered prompt

Build me a lightweight array framework for machine learning on Apple silicon that feels like NumPy to use, with a Python API and a matching C++ core.

I want arrays, basic math operations, lazy evaluation, automatic differentiation, and simple support for running work on CPU and GPU. It should be easy to use for research, with a small nn style layer and optimizers API if that fits naturally. Keep the design clean and simple, and make it so arrays live in shared memory where possible instead of constantly copying data around.

Please include a few examples and tests so I can see it working, and make the Python package installable from source. If you need current build or platform details, look up the latest docs online and follow the project’s intended setup.