Research on large AI systems requires substantial compute, sophisticated infrastructure, and specialized knowledge. These resources are concentrated in a handful of corporations and universities, and the models, data, and methods behind much of the resulting work are unavailable to researchers outside them. We do not believe that a small, elite circle should have a monopoly on the ability to do influential research.
EleutherAI works to widen that circle. We release models, datasets, code, and methods so other researchers can inspect, reproduce, and extend our work, and we do that work in the open, where anyone can follow along and contribute. But public artifacts are not enough: researchers also need compute, mentorship, and the chance to lead work of their own. Our community reaches across more than thirty countries to train aspiring researchers, provide resources to those without them, and grow them into independent researchers in their own right.
We also help independent research communities build capacity of their own. Through shared compute and collaboration, new groups can pursue their own questions and publish openly. The goal is not to bring every promising group inside EleutherAI, it is for more researchers and organizations to set their own agendas, contribute work others can build on, and stand on their own. Several communities featured below began with that support.
Every year EleutherAI holds SOAR, a 5-week fully online research internship program. We invite people with little research experience to contribute to open science under the mentorship of experienced researchers from our staff and broader community.
EvalEval Coalition
How AI is measured shapes what gets built and believed. EvalEval is a global coalition of researchers, co-founded by EleutherAI, Hugging Face and the University of Edinburgh, working to rebuild evaluation as shared open infrastructure: common ways to report results, honest accounting of where benchmarks break down, and standards that range from the benchmark all the way to policy.
Multi-Agent Security Initiative
The prominence of AI agents is growing rapidly, but work on multi-agent security and safety is lagging severely behind. Founded by Ben Hagag and Michael Mulet, MASI is focused on multi-agent security as a field of its own by studying how the presence of multiple interacting agents changes the AI safety and security landscape.
Data Provenance Initiative
Modern AI is built on data whose origins, ownership, and consent are mostly undocumented. The Data Provenance Initiative is a volunteer collective of researchers that audits that foundation at scale: tracing training datasets from their origins through creation and reuse; documenting sources, licenses, creators, languages, and geographies; tracking how web consent is changing; and mapping who actually holds power in the open-model ecosystem.
UniverseTBD
UniverseTBD is a collection of researchers with backgrounds in AI and across the sciences working to bring recent advancements in AI to bear on problems of interest to astrophysics researchers. They put a heavy focus on making these advances more accessible to scientists, building the datasets and tools that open discovery to more people, alongside agents that help researchers search the literature and generate new hypotheses.
We helped these communities get started, and they have since grown into independent organizations with work and communities of their own.
EleutherAI members organize technical discussions around active areas of AI research, from close readings of recent papers to presentations of work in progress.
