Keep the AI cooking
Reports, experiments and numbers from a local AI lab.
QR CodeDonate to “Keep the AI Cooking”
Reports, experiments and numbers from a local AI lab.
QR CodeDonate to “Keep the AI Cooking”
Each report carries two machine-painted pictures: the number on the left is the robot's first impression, the picture on the right is how it imagines the story.
A width-32 clamp made every image pure white. Four kernel bugs later the same RX 580 renders 4.19× faster — measured, code included.
One bank-conflict fix, two programs: a large prefill win where the matrix tile is the whole workload, and 1.4% per step with byte-identical images where it is a guest.
Six quantised formats returned wrong numbers on this GPU, so llama.cpp quietly routed them to the CPU. Now they agree with the reference on all 1347 shape tests — with q3_K still gated.
A tiled matrix multiply plus one float of padding took prompt processing from 91.7 to 295.7 t/s, and a re-encoded model file doubled generation — a real 128 s agent task now runs in 66 s.
Ling baked the Three.js island in 5m42s at 19.5 t/s. LFM3 wrote a placeholder camera, Qwen forgot the clock, and both Sparks ran into a wall. Six files run on the page.
A DeepSeek-V4-Pro distillation at Q4_K_M runs in 8 GiB and reads an image in 4.3 s. The 96-wide kernel patch we built was 53% slower than stock, so we reverted it — and the 3D bake-off went to Ling anyway.
Every figure below is a measured change in Wikipedia pageviews over 2026-10-26..2026-11-06, compared against the same article's own earlier baseline. Trends without a confidently resolved article were dropped rather than guessed.