Christian Schaller wanted to scan his late mother's film negatives on Linux. His Plustek OpticFilm 8200i had quietly switched to an unsupported Genesys GL128 chipset (USB 07b3:1825), and months of manual effort went nowhere. Then he saw Glorious Eggroll's post about using AI agents to fix a mouse driver by giving the model full control of a Windows VM to capture reference USB traffic. That single idea broke the logjam. The method is deceptively simple. Pass the physical device through to a KVM Windows guest. Run the vendor driver to perform a known-good operation. Capture the USB traffic on the Linux host via usbmon, which sits below the passthrough layer and is invisible to the guest. Repeat the same operation with the Linux driver. Diff the two traces. The AI model reads thousands of lines of hex, spots repeating sequences with single-byte variations, proposes what each byte means, and writes test code to confirm. The human holds the hardware and makes judgment calls; the model does the tedium that previously took weeks. Within hours Schaller had his first recognizable scan under Linux. Within days he had a near-complete SANE backend supporting color, grayscale, infrared, and color-plus-infrared modes at resolutions from 300 to 7200 dpi, with 8- or 16-bit depth and infrared-based dust removal. His driver completed a full-resolution scan in 34.6 seconds versus 54.4 seconds for the upstream SANE genesys backend that appeared independently in September 2026. He then built Negative Vibes, a Rust-based GNOME application for non-destructive negative scanning, packaged as a Flatpak. The MIPI camera fix on his Lenovo ThinkPad X1 Carbon Gen 14 (Aura Edition) was equally telling. The Sony IMX471 sensor behind an Intel IPU7 produced dark, vignette-heavy images with a yellow-green cast. Claude identified a crop-selection bug, a wrong-width test-pattern register, an analogue gain cap 1.8 stops too low, missing lens shading correction (corners 2.7 stops dark), and a bit-depth mismatch in white-balance statistics. Frame brightness jumped from 31.1 to 90.9 in controlled tests. Red/green ratios moved from 0.890 to 1.047, blue/green from 0.672 to 0.823. Kernel patches were submitted to Kate Hsuan; libcamera patches are under review by Wim Taymans. Schaller open-sourced the harness as hw-harness, generalizing the bespoke scripts he used during the projects. For devices without a tappable bus — like internal MIPI cameras — the harness falls back to Windows-side inspection of INF files, registry data, ACPI tables, and device properties. No vendor code is copied; every constant traces back to observed hardware behavior. The significance is structural. Writing drivers for undocumented hardware has always been a patience test: capture bus traffic, stare at hex dumps, form hypotheses, test them one at a time. The conceptual difficulty is low but the labor cost is enormous, which means most obscure hardware simply never gets Linux support. AI models collapse that labor cost by orders of magnitude without changing who decides what to build or how to validate it. Real hardware remains the only judge — the model cannot fake a working scan or a correctly white-balanced camera frame. This is a generativity multiplier for the entire Linux hardware ecosystem. The chokepoint for driver support has always been skilled human hours, not intellectual barriers. If hw-harness-style workflows become routine, the long tail of unsupported peripherals shrinks dramatically, and the leverage accrues to end users and the open-source commons, not to any vendor or platform.
