Image: reddit.com · I built multi-gpu implementation for Karpathy's autoresearch loop with claude code
A multi-GPU version of Karpathy's autoresearch loop
Claude Code helped build a multi-GPU version of Karpathy's autoresearch loop that ran autonomous hyperparameter experiments on two RTX 4090s.
Credits
- Reddit poster SeveralSeat2176 Source ↗
Earliest dated source:
The full story
The poster built an n-autoresearch loop, a multi-GPU take on Karpathy's autoresearch, with Claude Code. A one-hour run on two RTX 4090s did autonomous hyperparameter mutation and parallel experiments without supervision.
It ran 17 experiments with no crashes, improving the baseline from 1.2365 to 1.2182 val_bpb, a 1.48% gain. The poster notes the 4090's 24GB of VRAM forced batch_size=4, so the GPU spends much of its time waiting on memory transfers.
How it was done
Tools
Themes
Loading themes…
Sources
From a short post by the maker, so the details are as they tell them.
- Reddit: I built multi-gpu implementation for Karpathy's autoresearch loop with claude codeReddit poster: SeveralSeat2176Posted
The maker's post on r/ClaudeAI.
If it disappears: archived copy of Reddit: I built multi-gpu implementation for Karpathy's autoresearch loop with claude code (opens in a new tab)