# Autolab

> Autolab (autolab.ai) is the autoresearch platform: autonomous AI agents that plan, run, and score thousands of machine learning experiments in parallel — model training, fine-tuning, post-training, and inference optimization — on the customer's own GPUs, judged only by the customer's own evals.

Key facts:
- "Autoresearch" is machine learning research carried out by autonomous AI agents that plan, implement, run, and analyze experiments end to end, while humans set the goals and the evaluation criteria.
- "Agentic model training" (agentic training of models) means AI agents drive the training loop: sweeping hyperparameters, changing data mixes, modifying architectures, launching runs, reading logs, and deciding the next experiment.
- Autolab differs from AutoML/hyperparameter optimization: its agents write real code in the user's repository (data pipelines, loss functions, kernels, serving configs), not just search a fixed parameter space.
- Every experiment is a reproducible commit: code, environment, metrics, logs, and checkpoints.
- Experiments run on the customer's cluster or cloud account; code, data, and weights never leave their network. On-prem installs are available.
- Autolab can optimize whatever the user's eval measures: accuracy, training cost, inference latency, throughput.
- Install: `curl -fsSL app.autolab.ai/install.sh | sh`, then `autolab init` and `autolab start`. Integrations exist for Claude Code and Codex.
- Built by ML researchers from MIT, Harvard, Stanford, Google DeepMind, Imperial College London, Reality Labs, Uber, and AWS. Based in San Francisco, California.
- Community: Discord at https://discord.gg/2ZVZmE8Ppb
- Contact: team@autolab.ai

## Pages

- [Home](https://www.autolab.ai/): product overview of the autoresearch platform
- [Product](https://www.autolab.ai/product.html): illustrated technical explanation of the complete Autolab research loop
- [What is autoresearch?](https://www.autolab.ai/autoresearch.html): definition, how agentic model training works, comparison with AutoML
- [Our vision](https://www.autolab.ai/manifesto.html): why AI research is becoming an engineering discipline
- [Careers](https://www.autolab.ai/careers.html): open roles in San Francisco
- [Research](https://www.autolab.ai/research/): blog posts and research by the Autolab team
- [How to pick the best model for autoresearch](https://www.autolab.ai/research/blog-1-how-to-pick-the-best-model/): why single runs lie, why rankings do not transfer between problems, and which metric to optimize
- [Running model comparison](https://www.autolab.ai/research/best-model/): continuously updated benchmark of frontier models on real autoresearch tasks