Engineering

A browser tab is a constrained place to run a neural net: one thread, 32-bit floats, four-wide SIMD, no GPU. This article goes into the specific optimizations that made running Dots there practical.

One Highway kernel source compiles to NEON, AVX2, AVX-512 and WASM SIMD. Softmax as the worked example — tags, tails, and the build flags that quietly hand you a narrower target than you asked for.

UchenML is a C++20 machine learning library where the model is a constexpr variable. Define the model in source code and start training.

std::vector::reserve usually seems like a sensible optimization. It usually is not...

C++ is a beautiful and powerful language, but it does have its sharp edges. This article details open-source projects to reduce the number of cuts.