Autoresearch
Let an AI agent run autonomous LLM training experiments overnight on your GPU — Linux/NVIDIA or Apple Silicon, no babysitting required.
Media
What this seed does
This seed installs *autoresearch*, a compact language-model training loop purpose-built for an AI agent to run and improve autonomously. You point an agent at this SEED, it detects your platform (Linux + NVIDIA or Apple Silicon), installs the right variant, downloads training data, and sets up the environment — then walks away and lets experiments run.
Once installed, the agent reads a program.md runbook and iterates on the training script in a loop: edit, commit, run a ~5-minute training pass, check the validation metric, advance if it improved, repeat. The idea is to leave it running overnight and come back to findings.
The seed is platform-aware out of the box: on Linux it installs the upstream version with CUDA; on Apple Silicon Macs (≥16 GB RAM) it installs a Metal/MPS fork with no CUDA requirement. Owners of sub-H100 GPUs get tested batch-size settings to avoid out-of-memory errors on cards from 11 GB to 24 GB. It also serves as the canonical worked example of the SEED convention format.
When to use it
- You want to let an AI agent run LLM training experiments overnight and report back what it found in the morning.
- You have an NVIDIA GPU sitting idle and want to put it to work iterating on a language model training loop autonomously.
- You're on an Apple Silicon Mac with plenty of RAM and want to run LLM experiments without needing a Linux server or CUDA setup.
- You're learning the SEED convention and want a well-structured, real-world reference example to study or fork.
- You have a sub-H100 GPU (e.g. RTX 3090 or 4090) and need tested batch-size guidance so training doesn't run out of memory.
Source
Version history
1 releaseInitial Seed Release.
Dependencies
3 required · 3 optional- SeedCUDA dependency (cuda/SEED.md)optional· link ›· Required on Linux + NVIDIA path only; in-repo sub-SEED at cuda/SEED.md
- StateDiscrete NVIDIA GPU (Linux)optional· Required on Linux x86_64 path only
- StateApple Silicon with ≥16 GB unified memory (macOS)optional· Required on macOS / Apple Silicon path only; training peaks at ~14 GB under MPS
- State~5 GB free disk in $HOMErequired· Needed for venv, model cache, and data shards on both platforms
Contributors
2 contributors
Comments
0 commentsYou need to be signed in with GitHub to comment.
No comments yet — be the first to share how this seed worked for you.
Similar seeds
Run the Hermes AI agent locally in Docker with a browser dashboard, ChatGPT login, and files you can edit directly from your computer.
by @plow-pbcBoth install AI agent frameworks locally with platform detection and automated setup for hands-off operation.
Give your Hermes AI agent a persistent, searchable knowledge graph it can read and write without any image rebuild.
by @plow-pbcBoth are agent-tooling seeds that set up persistent infrastructure for AI agents to run autonomously without rebuilds.
Give this recipe to an AI coding agent and it builds Almanac — a Figma-style design-review app with pin comments — from scratch.
by @maryldyerBoth are SEED convention examples where AI agents build or configure complex software from a specification runbook.
A local, no-login teleprompter: control your script from your laptop while your phone shows the words at the camera — built by an AI agent from a single spec file.
by @delattre1Both demonstrate AI agents autonomously building functional applications from specs and iterating without manual intervention.