SysadminNews

Linux 7.2-rc7: more fixes do not measure AI accuracy

On this page
  1. The announcement, with the release outcome
  2. Three numbers that must not be confused
  3. Small changes still need relevant testing

The unusually busy seventh candidate is evidence of many accepted changes. It is not a standalone measurement of the accuracy of the AI tools that helped find some of the issues.

Invented review example: 100 reports yield 60 confirmed defects and 40 false alarms; the 60 defects are fixed through 80 commits. Report precision is 60%, not 80%. No actual kernel or AI-tool measurements are represented.
Invented review example: 100 reports yield 60 confirmed defects and 40 false alarms; the 60 defects are fixed through 80 commits. Report precision is 60%, not 80%. No actual kernel or AI-tool measurements are represented. Chart : PeopleAreGeek. Data source.
View full-size image

The announcement, with the release outcome

In his August 9 rc7 message, Linus Torvalds attributes many fixes to review by AI tools. He highlights s390 crypto, the return of Btrfs fixup infrastructure and some netfilter work among many small changes. He saw no reason then to delay absent a serious new problem. The v7.2 tag records the final release on August 16; that date is now an outcome, not a prediction.

Three numbers that must not be confused

Our diagram separates tool reports, validated defects and merged changes. One defect can need several patches; several reports can duplicate one defect; rejected reports never appear in a release’s shortlog. Counting the final commits cannot reconstruct the number of false alarms or missed bugs upstream of them.

For an invented example, imagine 100 tool reports containing 60 confirmed defects and 40 false alarms. If those 60 defects become 80 commits, “80 fixes” still does not mean 80% report precision. The report precision in that example is 60%. Recall cannot be calculated without knowing the defects the tool missed. None of these figures measures a kernel-review tool.

Small changes still need relevant testing

A two-line change can affect a common error path; a larger mechanical change may leave behavior intact. Size is context, not a universal risk score. Read the affected subsystem and reproduce the operation your machine relies on instead of treating a large or small diffstat as a verdict.

The kernel’s regression handling guide tracks where behavior stopped working, including version ranges or a bisected commit. A useful local record pairs the last working kernel with the first failing one, the same configuration, exact trigger and observable failure. That evidence helps maintainers distinguish a new regression from an older defect that broader testing has merely exposed.

September 8: preserve the rc7 announcement and add the final release date; separate AI-assisted discovery, accepted fixes and regression risk.