Clarke's third law says a sufficiently advanced technology is indistinguishable from magic. Anthropology has a working definition of magic that goes further than that line, and the definition fits how millions of people use AI models in 2026. A person appeals to a force they cannot examine, in words they choose carefully, for personal gain, and the outcome sometimes ends badly for them or the people around them. That shape shows up across three areas: what magic actually is, why societies eventually banned it, and what separates someone who can only cast spells from someone who understands the mechanism.

Kiro's context usage meter read 30% while Claude Opus 5 told me in the same chat that it had no room left to work with. Kiro measured the real usage. The model estimated its own remaining context, and the estimate was wrong by a wide margin. The behavior has a name, "context anxiety", and both a production write-up and a peer-reviewed paper describe it.

My ISP upgraded my gateway to a newer model and let me keep the old one. Instead of tossing it in the e-waste pile, I turned it (plus a decommissioned enterprise appliance from eBay) into a segmented home network with a private Jellyfin media server. It came together over a weekend, and I'm writing it up in case the approach is useful to someone else.

I diffed the EasyOCR and Qwen3-VL transcriptions of 444 Soviet cookbook pages against each other, and the two engines produced character-for-character identical text on only 5 pages. All 5 are nearly blank (a heading or a page number, 13 to 24 characters each). The plan from the previous post was to treat pages where both engines agree as "probably accurate" and hand-check only the rest. On real content pages that agreement almost never happens, so cross-engine agreement doesn't work as the cheap accuracy filter I hoped for.

Docling's vision-language model pipeline, running the 4B-parameter Qwen3-VL model locally, processed 444 cookbook page images for the soviet.recipes project at a 99.5% success rate. Compared to the earlier EasyOCR run, it gave up every piece of debugging output that run had: no text-region overlays, no per-page confidence score, no reading-order artifacts. It also ran far slower, and Docling still processes one-page image documents one at a time, so the batch took about 26.3 hours.

An AMD Ryzen AI Max+ 395 laptop with its integrated Radeon 8060S GPU and unified memory can run open-weight models from 4B parameters up to a 120B mixture-of-experts model, no cloud GPU needed. gpt-oss-120b generates around 35-40 tokens per second, and the 20B and 4B models are faster still. That's enough for the soviet.recipes project: this hardware can host a model large enough to attempt vision-language OCR locally.