DenseCore

When Rust Meets PyTorch: A Student's AI Framework Hits Bit-for-Bit Parity

A CS student built DenseCore, a 2 MB Rust deep learning framework with a PyTorch-like Python API, and matched PyTorch's output bit-for-bit across 50 training steps of Karpathy's nanoGPT. That's not "close enough", it's identical, achieved by reverse-engineering subtle implementation details like Flash Attention paths and Adam's rounding quirks. The catch? It only matches on Apple Silicon. Worth pursuing as a lightweight drop-in for serverless inference? Or just a brilliant learning project? (We recently explored how ARPL adapts llama.