Findings
Synthesis posts. Multiple benchmarks woven together into a single read. The brand promise stays the same here: every claim links to the underlying run; methodology is shown, not promised.
syntheses
How fast can a personal AI setup actually be?
A walk through the three machines that answer when we ask something. The numbers, the caveats, the places we’re still figuring out.
Read the finding →dead ends
Things we tried that didn’t work. Published so the next person can skip the same dead ends — what doesn’t pencil is often more useful than what does.
Gemma 4 MTP on Pascal — a dead end. Use Vulkan instead.
Six attempts to get a 3× speedup from multi-token-prediction on Pascal-class GPUs, all dead ends. The 27–47% real win lives on the Vulkan backend instead.
Read the dead end →More findings + dead ends shipping as we run more campaigns.