Five pending PRs on autoresearch, five repos quietly accumulating commits, and a licensing question that nobody answered.
A contributor adds capability-aware flash attention switching to run training on cheaper cards.
#268 from Nidhicodes adds Turing GPU (T4/RTX) compatibility via capability-aware FA3 fallback with fp16 and GradScaler support. The change addresses a real friction point: not everyone has H100s. Meanwhile, #265 saves checkpoints before eval to survive crashes, and #264 guards against zero grad accumulation steps.
Community contributors chip at progress plots, gradients, and contribution guidelines.
#263 fixes undefined 'best' variable in progress plots. #258 proposes a minimal CONTRIBUTING.md focused on research context. The PR queue suggests active tinkering but no merges this week.
The original BLS treemap spawned India, UK, and EU variants as the job market visualization empire quietly expands.
@karpathy's jobs project has fractured into specialized regional instances: india-jobs-ai-exposure for NCO occupation groups, jobs-uk for ONS data, and AI-ISCO for EU ESCO standards. Each repo shows five commits this week. An open issue credits an Indian adaptation built on the original concept. The archive fork karpathy-jobs-archive preserves the original before deletion.
The open issues reveal ambitions: pluggable optimization tracks, snapshot-diff scoring, and data-centric research.
Someone asked whether autoresearch code can be reused without a license and got no answer. The backlog includes Auto-Research Tracks (pluggable idle optimization), a JudgeModel layer for snapshot-diff scoring and binary rubric verification, and expanded harness generation for pattern and security exploration. No timelines yet.
| autoresearch |
|
48,252 |
| jobs |
|
1,023 |
| india-jobs-ai-exposure |
|
1 |
| AI-ISCO |
|
1 |
Your GitHub week, turned into something worth reading.
Generate your dispatch →