Merriam-Webster named "slop" its 2025 Word of the Year. This essay examines what distinguishes useful AI-generated content from hollow output, drawing on Clement Greenberg's kitsch theory, the MINT Lab's taxonomy, and the author's daily experience shipping real work with AI tools. The test is simpler than you'd expect: does the text commit to a claim that could be wrong?
A deep dive into how the Freerouting autorouter actually works: A* search on free-space polygons, rip-up-and-reroute optimization, and the architectural limits that make copper pours impossible. Includes the mathematics of A* and why it appears everywhere from game pathfinding to PCB layout.
A walkthrough of batch-processing 20 years of photos (51,414 JPEGs and Nikon RAW files) on an AMD Strix Halo APU. Covers EXIF-based orientation correction, horizon straightening with OpenCV, AI captioning with BLIP, and cataloging everything in SQLite. Includes the EXIF parsing bugs that cost a day of debugging.
Part one of a two-part series on redesigning an Arduino Giga R1 level-shifter shield using Claude Code and open-source EDA tools. Covers replacing TXB0108 auto-sensing translators with SN74LVC8T245 driven shifters, generating PCB layouts with Python, autorouting with Freerouting, and producing Gerber files, all driven from a terminal with AI assistance.
Qwen3-TTS benchmarked across three very different machines: Apple M3 Max, AMD Radeon 8060S, and NVIDIA Tesla P40. The results reveal how much hardware architecture matters for autoregressive TTS, why short benchmarks lie, and what "local inference" actually costs in time and electricity when you are generating hours of audio on your own hardware.
A practical guide to building an AI inference server from retired enterprise GPUs. Four Tesla P40s, 96GB of VRAM, $2,500 all-in, and the benchmarks to prove it was worth it.
