-
2026-06-07
I built a full LLM pipeline on two desktop machines. The bottleneck was never the disk.
I built a recipe-writing LLM end to end on two desktop machines, data prep, teacher-student distillation, fine-tuning, eval, and serving, and measured where each stage actually got slow. The disk was almost never the bottleneck; the interesting part is what was, and where storage finally takes over. Measured numbers tucked into "for engineers" toggles.
-
2026-06-07
The quiet bottleneck — when storage actually slows your AI down
Storage is usually quiet; the GPU or the network is your real bottleneck. But when storage does take over, it stalls the most expensive hardware you own. A field guide to when it is innocent and when it is guilty: one cool-to-hot gauge per pipeline stage, with the measured numbers tucked into "for engineers" toggles.
-
2026-05-31
The harness, the model, and me
Five things I had to internalize about AI coding agents (Claude Code, Cursor, Cline, Codex, Antigravity, custom SDK builds): the harness/model split, where context goes, why descriptions matter, picking the right primitive, soft vs hard governance.
All posts