EleutherAI, a nonprofit research hub whose members explore issues associated with artificial intelligence (AI), has launched a nonprofit. The move allows the organization to underwrite research efforts that focus on AI, alleviating some of the financial constraints previously experienced by volunteer researchers.
In keeping with EleutherAI’s original purpose, researchers’ primary focus will be alignment and interpretability concerns. The goals of alignment research include making sure AI systems follow their designers’ intentions, while interpretability research explores how machine-learning systems make decisions.
The nonprofit comes after a period in which researchers offered their time and insight on a volunteer, ad hoc basis. “[T]he biggest blocker in what we could be accomplishing is the fact that working a forty-hour workweek and doing cutting-edge AI research on the side is unsustainable for most contributors,” EleutherAI leaders wrote in a post on their website. The new nonprofit research institute will allow more than 20 of EleutherAI’s contributors to work full time on their research.
The new organization will be run by Stella Biderman, Curtis Huebner and Shivashu Purohit. Funding came from “a mix of charitable donations and grants,” according to the blog post.
EleutherAI was launched on the information sharing and chat server Discord in July 2020 as a forum for discussing GPT-3, a machine learning neural network model designed by OpenAI. While EleutherAI maintains its presence on Discord, the organization has a website presence at https://www.eleuther.ai/.
In an email to The NonProfit Times, Biderman, who has been coordinating the process of setting up the nonprofit, noted that codifying the loose association of contributors on the Discord platform would allow the organization to maintain full-time employees and plan and execute longer-term projects than would be possible if the collective had remained volunteer-based.
In a separate email, EleutherAI Chief Engineer Shivanshu Purohit gave The NonProfit Times additional insight. “It would be more on brand with Eleuther’s image to be a nonprofit than a for profit organization,” Purohit wrote. “Before any of this we were just a Discord community using free computational resources to release world class AI models that would’ve otherwise been in the hands of big labs. In fact, since our existence, researchers at other labs have been more open to open sourcing their research as it’s easier to convince the executives at said companies to do so, as it gives them positive publicity like Eleuther.”
Purohit also noted that by forming a nonprofit, the organization was able to receive funding from Hugging Face, a for-profit company that develops tools for creating apps using machine learning, as well as Stability AI, which builds open-source AI models, among other funders.
The organization’s commitment to open sharing of information was readily apparent in Purohit’s response to the projects that make him most excited. “We have a lot of ongoing research projects that we welcome anyone to check out on our Discord (https://discord.gg/eleutherai),” he wrote. “Our current focus is on analyzing models as they train (interpretating across time), probing the models to find what each part of the model learns, extracting the latent knowledge from the model (which is quite different from asking ChatGPT ‘answer this question truthfully’), and many other projects that will further our goal towards contributing to development of safe AI. I in particular am very excited about our research on eliciting latent knowledge, since it will help us utilize large AI models much more robustly, even when they become capable enough to trick/lie to humans.”
Initially, EleutherAI focused on designing what it refers to as Large Language Models (LLMs) as well as training researchers in their use.
“Times have changed significantly since EleutherAI was founded, and there is substantially more interest in training and releasing LLMs than there once was,” EleutherAI leaders wrote in their blog post. “Thanks to those efforts, we are free to pursue the research we wanted to use these models for to begin with—studying topics like interpretability, alignment, and learning dynamics of large transformer models. For 2023, we are also planning to spin up more alignment and interpretability projects, and become significantly more involved with the broader alignment community.
Formation of the nonprofit comes after a period when three notable groups of researchers left the EleutherAI umbrella to form their own organizations. According to the blog post, “EleutherAI founders Connor Leahy and Sid Black are now the founders of a new alignment research organization called Conjecture, Louis Castricato has started a lab called CarperAI that focuses on preference learning and RLHF [reinforcement learning from human feedback], and Tanishq Abraham has started MedARC, which focuses on biomedical applications of cutting edge AI technologies such as large language models.” Leahy will serve on the new nonprofit’s board of advisers.
