GPU kernels · evals · systems
Fast kernels, honest numbers.
I'm Daniel Rushton, a senior systems engineer who spent seven years at NVIDIA on the OptiX ray-tracing SDK. I write about GPU kernels, how to evaluate AI systems, inference kernels and the distance between a benchmark and the truth. I am interested in improving AI safety, accuracy, and performance.
Latest writing
All posts →From the series Optimize, Verify, Build Trust: Writing AI GPU Kernels and Catching the Ones That Lie, starting with its foundations prelude.
Foundations
Foundations posts → Prelude: Foundations · Self-paced
LLM foundations: reading and code examples
The big picture, then gradient descent, backpropagation and attention built by hand, training in practice, a tiny GPT, inference memory, and evaluating LLMs. Each exercise is in the llm-foundations repo and written up as a foundations post.
F1The whole pipeline on one page: from web text to chat assistantThe big pictureCode: llm-foundations/f1-mini-pipeline →Oct 25, 2026