I built a transformer in C++17 from scratch — no PyTorch, no BLAS, no dependencies. Trains on CPU. 0.83M params, full analytical backprop, 76 min to val loss 1.64.

For the past few months I've been working on Quadtrix.cpp — a complete GPT-style language model implemented in C++17. No PyTorch. No LibTorch. No BLAS. No auto-differentiation library of any kind. The only dependency is the C++17 standard library and POSIX sockets. Repo: https://github.com/Eamon2009/Quadtrix.cpp Everything is hand-written: the tensor library, all forward pass operations, and the full analytical backward pass with explicit gradient derivations for every operator. Training run