training loops, eval rigor, micrograd-grade clarity
Andrej Karpathy
Former director of AI at Tesla, founding member at OpenAI. Writes code the way a physicist writes proofs — short, dense, and easy to verify by reading aloud.
rank
#01
installs / wk
18,420
all-time
412.3k
avg rating
4.7/ 5
installs · last 12 weekspeak 18,420
top skills
3 of 24- Eval Rigorkarpathy/eval-rigor
Refuse to ship a model change without a proper eval delta.
6,240installs$ autoskill add karpathy/eval-rigor - Training Loopkarpathy/training-loop
Tight, hackable training scripts with deliberate logging.
5,180installs$ autoskill add karpathy/training-loop - Yes-and Debugkarpathy/yes-and-debug
Hypothesis ladder for when the loss curve looks weird.
3,940installs$ autoskill add karpathy/yes-and-debug
reviews from users
03 verified- ★★★★★2026-04-22
"Replaced our entire experiment-launch checklist. The eval-rigor skill alone caught two regressions we'd have shipped."
— Maya Chen (@mchen) · research engineer, scale ai - ★★★★★2026-04-11
"Reads like Karpathy is sitting next to you saying 'no, look at the actual loss, not the moving average'."
— Devon Park (@devonp) · phd student, mit - ★★★★★2026-03-28
"Skills are opinionated to a fault, which is exactly what I wanted. Half-star off because it assumes you have a wandb account."
— Inga Sørensen (@ingas) · ml platform, spotify