An elephant through a straw
A 284B mixture-of-experts model, a 24GB 4090, and one NVMe drive. For an evening it ran at 12.9 tok/s, which turned out to be the model writing the word "dekameters" 108 times because loops are cache-friendly. Also: a 2-bit copy of 11,008 experts, a net that guesses the router two layers early, and five dead ends with the numbers that killed them.
Qwen plays Smash
A 1.7B LLM in Melee's per-frame control loop. 30ms CUDA-graph scoring, an Nsight trace of where the time really goes, 6,144 imagined futures per decision, a critic that loses to a single constant, and a baseline comparison the LLM does not win. With a scrubbable match player.
One launch per token
A from-scratch fused decode megakernel that runs a whole transformer step in one CUDA-graph launch and beats vLLM's batch-1 latency on a 4090, for a dense model and a 64-expert MoE. Plus two H100s and an honest negative result.
Don't read the haystack
Triton flash-decode kernels at 94% of my 4090's memory bandwidth, then 16x past them with page-sparse attention, because the frontier of inference is refusing to read.
Why the answer is D
I inserted a wrong answer into a reasoning model's prompt and probed its residual stream to find out when it actually decides. Spoiler: sometimes before it starts "thinking."
A dial for a language model's mind
A single direction added to an LLM's activations dials a behavior up and down like a knob. Extracted on one 4090, running live in your browser. Drag the slider.
A stupid amount of work for a date
Rebuilding the lantern scene from Tangled as a walkable 3D world you can stand inside. Gaussian splats, a 480p ceiling, and one genuinely cursed NCCL error.