Research at EleutherAI

We make AI systems legible, reproducible, and safer by training and releasing what the rest of the field needs to study them: models, datasets, evaluations, and tools.

Open Source AI

Science

  • Interpretability Over Time

    How model behavior and internal structure emerge over the course of training, and what shaped them.

  • Evaluation

    What benchmarks measure, when they saturate, and how evaluation shapes our picture of model capability.

  • AI for Science

    Models and data for scientific work, mathematical reasoning, and methods for checking the results.

Safety

  • Open-Weight Safety

    Safeguards that remain useful when anyone can download, inspect, and fine-tune the model.

  • Behavioral Safety

    When a model's answer reflects what it knows, and when it reflects what it has learned is safe or rewarded to say.

  • Privacy and Security

    Memorization, extraction, and the security properties of open systems.