EleutherAI is a nonprofit AI research group that grew out of a Discord server opened in July 2020. Volunteers there built The Pile, a training dataset released on December 31, 2020, and then the GPT-Neo, GPT-J and GPT-NeoX-20B language models under the Apache 2.0 license. It remains one of the clearest examples of how an online community can do serious open research, and it is open to newcomers. This article explains its history, how it works and what to know before joining.
How it started
The group began in July 2020 as a chat server under the working name LibreAI; Wikipedia gives July 3 and July 7 for different aspects of the start, so we say only that it began that month. The founders named in public records are Connor Leahy, Leo Gao and Sid Black. The final name refers to the Greek word eleutheria, meaning liberty, a nod to the goal of keeping powerful language models from being controlled by a few companies.
Rather than raising money, the early members used Google’s TPU Research Cloud for compute. By early 2021 they accepted GPU cluster support from CoreWeave and SpellML. In early 2023 the community incorporated as the EleutherAI Institute, run by Stella Biderman, Curtis Huebner and Shivanshu Purohit.
What the community built
- The Pile. An 886 GB collection of text from many sources, released on December 31, 2020, described in an arXiv paper. See our explainer on The Pile dataset.
- GPT-Neo. Released March 21, 2021 in 125 million, 1.3 billion and 2.7 billion parameter versions.
- GPT-J-6B. A 6-billion-parameter model from 2021.
- GPT-NeoX-20B. A 20-billion-parameter model released on February 10, 2022.
The models were released under the Apache 2.0 license, which permits commercial use. At the time, few organizations offered downloadable models of this size, so they became baselines for researchers and startups. The group also maintains tools and model suites for research, including the Pythia models and an evaluation harness widely used for benchmarking.
A shift toward interpretability and alignment
From 2023 the group put more weight on interpretability, alignment and scientific research. A 2024 investigation by Proof News and Wired found that The Pile included subtitles from more than 170,000 YouTube videos across over 48,000 channels, which raised questions about consent and licensing in training data. In June 2025, EleutherAI and collaborators released Common Pile v0.1, a dataset limited to works whose licenses permit AI training, and trained two models on it. The group also worked with the UK’s AI Security Institute on research indicating that filtering certain concepts from training data can preserve performance while reducing the model’s ability to provide harmful information.
Why it matters
EleutherAI shows three things. First, volunteer communities can produce infrastructure that labs then rely on. Second, openness brings scrutiny: the YouTube finding shows that open datasets can be audited, which closed ones cannot. Third, the field’s norms are changing, from “scrape everything” toward documented, licensed data, and community groups have pushed on that shift.
How to take part
The main hub is the group’s Discord server, linked from its website. Our guide to asking good questions in AI communities applies directly: read the channel descriptions, search before posting and share what you tried. If you want to contribute code or evaluations, see how to contribute to open-source AI, and compare EleutherAI with other groups in our list of AI communities to join.
What to watch next
- New releases built on Common Pile and licensed or public-domain text.
- Interpretability and data-filtering results published with safety institutes.
- How the group is funded as compute costs rise.
- Whether other communities adopt its documentation and licensing practices.
Key takeaways
- EleutherAI started as a Discord server in July 2020 and became a nonprofit institute in early 2023.
- The Pile (886 GB) and models up to GPT-NeoX-20B were released openly.
- The group’s focus moved toward interpretability and alignment research from 2023.
- Common Pile v0.1, released in June 2025, uses only works with licenses that allow AI training.
- Anyone can join the community; preparation and clear questions help.
Frequently asked questions
What is EleutherAI?
It is a nonprofit AI research group that began as a Discord community in 2020 and releases open datasets, models and research tools.
What is The Pile?
The Pile is an 886 GB text dataset released on December 31, 2020 to train large language models. It combines many sources and has been used by researchers worldwide.
Can I join EleutherAI?
Yes. Its community is open through its public Discord server, which the official website links to.
