Discofork
Repository brief

karpathy/autoresearch

Read the upstream summary on the left, browse the cached forks below it, and load each fork comparison into the right-hand panel.

Cached analysis
cached 2026-03-30T12:12:33.197Z
3mo ago

karpathy/autoresearch

A small Python repository for autonomous LLM training research on a single GPU. It centers on a fixed 5-minute training loop where an agent edits `train.py`, uses `program.md` as instructions, and measures progress with `val_bpb`. The repo is active, very popular, and appears aimed at experimenting with agent-driven research workflows rather than providing a general-purpose training framework.

GitHub
Loading tags...
Stars61,263
Forks8,541
Default branchmaster
Last pushed2026-03-26T00:07:37Z
Recommended shortcuts

Jump straight into Discofork's strongest cached fork picks, or open a compare view in one click.

Forks

Choose a fork to inspect

10 of 10 fork briefs
Selected

Choose this fork if your target machine is a Mac and you want the upstream workflow adapted for MPS. Stick with upstream if you want the latest fixes or a more CUDA-oriented baseline.

Choose this fork if you want the autoresearch idea wrapped in stronger coordination, onboarding, and collaboration tooling. Stay with upstream if you want the smallest, most current baseline with fewer abstractions.

Prefer this fork if you want autoresearch to run on Windows and consumer RTX hardware with fewer manual runtime adjustments. Prefer upstream if you need the original Linux/H100-oriented setup or want maximum fidelity to the mainline research environment.

Prefer this fork if you want the autoresearch idea in Chinese with stronger onboarding and a skill-based workflow. Prefer upstream if you care most about staying current with bug fixes and the latest repository state.

Choose this fork if you want a general agent benchmark and execution framework with scoring, CLI, and server tooling. Stay with upstream if you want the simplest possible single-GPU autonomous LLM training sandbox.

Choose this fork if you specifically want Apple Neural Engine experimentation and are willing to trade away upstream simplicity. Stick with upstream if you want the cleanest, smallest Python/PyTorch agent-research loop.

Choose this fork if your goal is to run autoresearch on Hugging Face infra and you want the surrounding workflow to reflect that environment. Stick with upstream if you want the cleanest, most portable, local-first version of the project with the fewest HF-specific assumptions.

Choose this fork if you want the same autonomous research idea but with MLX/Apple Silicon support and more configurable training. Stick with upstream if you want the original CUDA-centered experience and the freshest upstream fixes.

Choose this fork if your main blocker is AMD GPU support. If you are on NVIDIA hardware, the upstream repo is likely the safer default because it is newer and this fork is only a small ROCm-focused divergence.

Choose this fork if you want a substantially expanded agentic research system with rollout validation, reward logic, and remote execution. Stick with upstream if you want the smallest possible autonomous-training sandbox and prefer to stay close to the original nanochat-style loop.