An immutable weekly snapshot of the focus published on the current Now page.
01This week
Closing the GPU + LLM pair
Where the kernel atlas meets the runtime atlas. I'm reading CUDA memory traffic patterns and pairing them with vLLM PagedAttention so the next GPU atlas chapter lands with the right model context, not in isolation.
GPU memory bandwidth
PagedAttention
KV cache
Atlas cross-references
02This month
Atlas as a learning loop
Converting the public atlas map from a flat directory into the learning system you now see on the homepage — a study order, three parallel tracks (hardware, serving, training), and a converge step. The /now page is the next piece of that loop.
Learning system
Guiding questions
Investment bar
Evidence flow
03Long-term focus
Industrial AI that operators trust
The through-line from mechanical engineering to AI engineering is "decisions a real person has to act on". Every atlas and every notebook is an attempt to make industrial intelligence legible, reviewable, and operable — not just accurate.