A nearly twenty-year-old video decoding bug in AMD's Linux graphics driver is fixed, and the contributor who found it says OpenAI's Codex did much of the legwork. The patch shipped in AMD's kernel graphics pull request on November 12 and is queued for Linux 7.5.
The hardware dates from 2007. The bug survived this long because almost nobody still writes code for it.
It is also a clean test case for what AI coding tools are actually good at: unglamorous code that is unfamiliar, poorly documented and owned by nobody.
What was actually broken
Ted Bullock, a first-time contributor, hit corrupted video frames on an iMac7,1 running an RV610 based mobile GPU, PCI ID 1002:94c8. Hardware playback through the Unified Video Decoder produced broken frames.
The symptom was specific and reproducible:
- Decoder reading compressed video from system memory while the GPU was drawing: frames came out corrupted.
- Same input moved into VRAM: frames decoded correctly.
- Graphics draws disabled entirely: corruption stopped.
That points at contention between the 3D engine and the decoder when UVD fetches its buffers over the system memory path. These early chips do not handle that well.
The radeon driver already forces UVD buffers into VRAM for some other GPUs. Bullock's patch extends that rule to the RV6xx family.
The fix and the hardware it covers
Five chip variants share the same decoder path and legacy R600 firmware, so the patch covers all of them:
- RV610
- RV620
- RV630
- RV635
- RV670
That is the Radeon HD 2000 and HD 3000 generation. Cards that stopped receiving active maintenance from AMD years ago.
Bullock was direct about the limits of his testing. He had the RV610 only, on Debian, and said someone with the other chips would need to confirm the behavior. Maintainers took the patch anyway, listed in the pull request as a plain "R6xx UVD fix" under radeon.
He was equally direct about the tooling. Codex helped with the investigation, the testing and the patch itself.
Why a 2007 bug is a 2026 story
Bugs on hardware this old usually sit unfixed. The engineers who know the driver internals have moved on, and the people who still own the cards rarely know where to start in a kernel tree.
An assistant that can read an unfamiliar driver, propose experiments and narrow a fault down to one buffer placement decision lowers that barrier a lot. Not by writing the whole fix, but by making the first hour of the investigation cheap.
Nothing about the review path changed. Mailing list, maintainer review, merged into the pull request sent on November 12. The model did not skip the process, and it did not sign off on its own work.
The rest of that pull request is ordinary modern hardware work:
- amdgpu: DCN 6 DML2 fix, VCE fixes, GPU reset and SR-IOV updates
- amdkfd: SVM and CRIU fixes, trap handler work, dma-buf handling
- drm: EDID parsing for HDMI 2.2 FRL rates
The Linux 7.5 merge window opens around the end of the year, so the fix should reach mainline in early 2027. New RDNA 5 (GFX13) driver code was held back for a later batch.
For the small crowd still running an HD 2400, HD 2600 or HD 3870 as a media box, or reviving a mid-2007 iMac with a modern distribution, accelerated playback should finally behave.
The takeaway is quieter than any benchmark chart. Open source has a long tail of bugs that are too old to be anyone's job, and this one took a patient owner plus an agent that could read driver code.
What it does not prove is that an agent can maintain legacy drivers on its own. Bullock still had to notice the corruption, form a hypothesis and argue the case on a public list.
He just did not have to spend a week learning an unfamiliar driver to get to the first experiment.
Comments