Did Claude increase bugs in rsync?

Original Article ↗ Hacker News Discussion ↗

Scope of the rsync / Claude issue

  • Several commenters note that rsync saw a huge spike in commits and rapid security hardening, driven by a flood of (often AI-generated) security reports.
  • Some users hit serious regressions (e.g., broken incremental backups, high CPU usage) and pinned rsync to older versions.
  • Others point out that regressions and severe bugs also occurred in pre-LLM releases and that rsync has long been a complex, hard‑to‑test C codebase.

Statistical analysis of bugs

  • The linked article analyzes “bugs per 10 commits,” later extended with an LLM-based severity score, across many releases.
  • Main claim: the two Claude-attributed releases fall well within historical variation and don’t look unusually buggy, even when weighting by severity.
  • Critics argue:
    • The dataset for “Claude releases” is just two points; the study is underpowered.
    • Bugs/commit hides severity and masks a large spike in commit volume.
    • Non-significant p-values show “no evidence of difference,” not “evidence of no difference.”
  • Supporters reply that this still exceeds the rigor behind the original “Claude broke rsync” accusations, which were mostly anecdotes.

Code quality and concrete regressions

  • Example discussed: a change that effectively turned many allocations into calloc, later reverted after memory/CPU regressions. Maintainer says the decision was theirs, not Claude’s, though an AI co-author tag appears.
  • Some see this and similar changes as emblematic of “AI slop” and excessive churn; others attribute it to rushed security work under heavy report load, regardless of tools used.

LLMs in development and attribution

  • Many developers say LLMs significantly boost productivity, especially for boilerplate, tests, and refactoring, but produce code that must be reviewed like a junior’s work.
  • Reviewers complain AI-generated PRs are large, noisy, and hard to reason about, increasing review burden.
  • Strong debate over commit attribution:
    • One side wants explicit LLM markers for risk, review style, and copyright provenance.
    • Others argue the human committer is responsible regardless of tools; undisclosed LLM use is not inherently “fraud.”

Reaction to AI-written prose and community behavior

  • The article’s initial LLM-written prose triggered strong backlash; many described it as “AI slop” and said this alone reduced their trust.
  • The author later rewrote the prose by hand; some then found it much more readable, others remained skeptical because of the original choice and visible AI tooling.
  • Multiple comments lament the toxicity of the outrage, personal attacks on maintainers, and growing entitlement in OSS users.