Review economics
NVIDIA/Megatron-LM
Read from the 100 most recently merged pull requests ·
Pull requests with attributed agent authorship took only 0.7× the reviews submitted of the rest.
≥4%
Attributed
a floor
5
Attributed PRs
95
Other PRs
100
Read
Attributed — the latest 5
#6250 refactor(ci): migrate triage consumer to Cerno
1 reviews · 3h to merge
#6158 Fix/dsv4 cp relax assert offline packed thd
2 reviews · 27h to merge
#6077 Fix GTP full-iteration CUDA-graph capture regression
1 round · 17 reviews · 2.9d to merge
#6041 Rebind MFSDP v2 sharded grads once per gradient reduction
3 reviews · 5.2d to merge
#6104 test: add async scheduling inference coverage
1 reviews · 33h to merge
The rest — the latest 5
#6230 Release MFSDP storage after model deletion
5 reviews · 2.0d to merge
#6292 Revert "Add Agent Compose package and runtime interface skeleton"
1 reviews · 3h to merge
#6247 avoid circular dependency usage between mcore and modelopt- #5644
5 reviews · 21h to merge
#6291 build: AUT-1353 bump transformer-engine to 2.17.1
1 reviews · 1h to merge
#6281 cp: Update DEEPEP_COMMIT in Dockerfile.ci.dev to de0dd1185142c727b9c118a
2 reviews · 4h to merge
How this was measured
4% of merged commits carry agent attribution — a floor, not the share; tools that only complete code inline leave no commit trail, so the unattributed side includes AI-assisted work; 4 PRs came off an agent branch with no commit attribution, so the real share runs higher; reads 100 of 2404 merged PRs — the rest have not been analyzed yet.
Detected: OpenAI Codex, Claude Code.
“Attributed” means a commit carried an agent’s signature — a co-author trailer, an agent commit identity, or an agent bot account. Tools that only complete code inline leave no such mark, so the other column is “rest”, not “human-written”. Full method and its limits
The badge reads this repo’s current report, so it follows the number.
Tell me when this moves
We re-read NVIDIA/Megatron-LM weekly and email only when the number changes materially.
Read your own repos
This one is public. For a private repo, run the same read locally through your own GitHub credentials — nothing is installed and nothing is sent to us.
npx @ambera/review-taxWant this continuously on NVIDIA/Megatron-LM — each pull request paired to the task it came from, and the work graded from its review loop rather than its diff size? Claim this repo in Forge
Browse every repo people have read · Read a different repository