let's be clear
op doesn't provide code, so my effort stays at throwing in a claude session and providing code without ANY review
trust my numbers as much as you usually trust claude code
trust the original numbers as you usually trust no code
yes unfortunately there is no baseline
once you get past the article reporting, original study lives at https://geopjr.dev/blog/democratizing-abandonware
author provides no repo/code, but describes methodology
i run a 95% autonomous claude code session to re-run the experiment
after all, since he pulls pr/repo data for ai slop, I can just pull all data (not just ai slop) in the same time period and compare
https://github.com/rufasterisco/slopware
1
"ai slop prs/repos" starting numbers mostly match (120 to 199)
i only tested prs from github repos, bringing it down from 119 to 116
still, based on those numbers, looks like we are off by a neat 25%
could be different methodology or claude making mistakes
author mentions doing some manual cleanup
│ │ Original │ Ours │
│ Unique repos │ 120 │ 116 │
│ Maintained │ 32 (27%) │ 58 (50%) │
│ Abandoned │ 88 (73%) │ 58 (50%) │
2
given the above, this is what comparison to "baseline" looks like
as said, "baseline" here means PRs in the same period
│ │ AI Slop │ Baseline │
│ GitHub repos │ 116 │ 332 │
│ deleted / 404 │ 11.2% │ 4.8% │
│ ≥3mo stale (surviving)│ 43.7% │ 39.2% │
│ abandoned (total) │ 50.0% │ 42.2% │
│ alive │ 50.0% │ 57.8% │
post "github fetch" data lives in repo
if you want to run from zero, the repo has the scripts
it will download 600+ MB from github