[holg]

Rust-only deep learning is the right direction. Backend-agnostic, no Python required, autodiff built in, runs on embedded - that's a stack that actually deploys cleanly. Curious how training performance compares to PyTorch on real workloads.