Open-source repo where an AI agent autonomously edits and runs 5-minute nanochat LLM training experiments on one GPU.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
autoresearch is a small open-source repository by Andrej Karpathy, created in March 2026. It gives an AI coding agent a simplified single-GPU implementation of nanochat training and lets it experiment autonomously. The agent modifies the training code, trains for a fixed 5-minute wall-clock budget, checks whether validation bits per byte (val_bpb) improved, keeps or discards the change, and repeats. That is roughly 12 experiments per hour, or about 100 overnight.
The repo has three key files. prepare.py holds constants, data preparation and evaluation utilities and is not modified. train.py contains the GPT model, the Muon + AdamW optimizer and the training loop, and is the only file the agent edits. program.md holds the baseline agent instructions, works as a lightweight "skill", and is the file the human iterates on. A YouTube creator reported that the loop found improvements Karpathy had missed.
It is used by pointing a coding agent such as Claude or Codex at the repo with permissions disabled. It requires a single NVIDIA GPU (tested on H100), Python 3.10+ and uv. Other platforms (CPU, MPS) are not supported in the main repo. Because of the fixed time budget, results are not comparable across different hardware.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Andrej Karpathy's open-source repo where a model runs a loop of its own experiments to improve training of a small LLM. It found improvements Karpathy himself had missed, and the creator compares OpenAI's self-optimization loop to it.
Closest neighbours by meaning, not just by category.
Research · Meta
Meta research framework where an AI agent acts as a data scientist, iteratively building training and evaluation datasets.
Agent
Decentralized multi-agent LLM framework where agents self-organize into teams for long-running computational research.
Research · OpenAI
OpenAI and Molecule.one research system pairing GPT-5.4 with the Maria agentic chemistry platform and a high-throughput lab.
Research · OpenAI
OpenAI, PerturbAI and Scripps Research run thousands of GPT-Rosalind agents in parallel to analyze brain-cell gene data.
Research · Invideo
Research system from Invideo that writes robot control code, repairs failures and stores fixes in a reusable skill library.
Agent · Google DeepMind
Gemini-based multi-agent system from Google that generates, debates and ranks scientific hypotheses for researchers.
