Hi, I'm David W. Rose
I tinker with machines. Some of them think. My projects range from coding agents and LLM benchmarks to maps, computer vision, personal infrastructure, and vehicle electronics. I publish the useful results and strange side quests as cipher982.
Check what data leaks from your browser to websites
Using GPT-2 and StyleGAN to generate fake wines
Check out samples of my code
Latest Writing
-
AEO for a Personal Website: What I Changed After Auditing 13 Sites
July 20, 2026
I audited a month of self-hosted analytics across 13 sites to see whether AI answer engines were already sending traffic. They were. Here is the practical AEO work that survived contact with the data: clean identity signals, accurate sitemaps, citable original evidence, and a deliberately unglamorous llms.txt experiment.
-
A 5 GB Docker Image for a 7.4 MB Model
June 13, 2026
A deploy hung for fifteen minutes. The cause was 3.9 GB of NVIDIA CUDA wheels being installed for a model that runs on CPU. This is the side quest from the HDR project: why PyTorch ships a CUDA runtime you don't ask for, what inference actually needs, and what I learned porting GMNet to tinygrad by hand and proving the outputs identical to a fraction of a pixel.
-
Where the Branch Commits
May 14, 2026
When a tiny prompt edit flips an LLM's greedy output onto a different token, where in the forward pass is that flip actually committed? Across 82 cases on 8 open models, every case has a findable residual-stream position where patching from the unedited run restores the original token. Part 3 of the Nearby Prompts series.