EleutherAI, a non-profit research center whose members address issues related to artificial intelligence (AI), has formed a non-profit organization. The move allows the organization to take on research efforts focused on AI and alleviate some of the financial constraints that volunteer researchers previously faced.
Consistent with EleutherAI’s original purpose, researchers will primarily focus on issues of alignment and interpretability. Alignment research goals include ensuring that AI systems follow their designers’ intentions, while interpretability research examines how machine learning systems make decisions.
The nonprofit organization comes after a period in which researchers have lent their time and insights on an ad hoc, voluntary basis. “[T]The biggest obstacle to what we could achieve is the fact that it is unsustainable for most contributors to work a 40-hour week and do cutting-edge AI research on the side,” EleutherAI executives wrote in a post on their website. The new non-profit research institute will allow more than 20 EleutherAI employees to work full-time on their research.
The new organization will be led by Stella Biderman, Curtis Hübner and Shivashu Purohit. Funding came from “a mix of charitable donations and grants,” according to the blog post.
EleutherAI was launched in July 2020 on the information sharing and chat server Discord as a forum to discuss GPT-3, a neural network model for machine learning 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 charitable times, Biderman, who coordinated the nonprofit’s founding process, noted that codifying the loose association of contributors on the Discord platform would allow the organization to retain full-time employees and plan and execute longer-term projects than the collective could had stayed on as a volunteer.
In a separate email, EleutherAI Chief Engineer Shivanshu Purohit stated The Nonprofit Times additional insight. “It would be more consistent with Eleuther’s image to be a non-profit organization than a for-profit organization,” Purohit wrote. “Before, we were just a Discord community using free computing resources to release world-class AI models that would otherwise have been in the hands of big labs. In fact, researchers in other labs have been more open to open sourcing their research for as long as we have existed, as it is easier to persuade the executives of those companies to do so, as it gives them positive publicity like Eleuther.”
Purohit also noted that by creating a nonprofit organization, the organization has been able to receive funding from Hugging Face, a for-profit company that develops tools to build apps using machine learning, and Stability AI, which, among other things, is open-source AI -Models created other funders.
The organization’s commitment to an open exchange of information was evident in Purohit’s response to the projects that excited him most. “We have many ongoing research projects that we welcome everyone to check out on our discord (https://discord.gg/eleutherai),” he wrote. “Our current focus is on analyzing models during training (interpretation over time), examining the models to find out what each part of the model is learning, and extracting the latent knowledge from the model (which is very different than asking ChatGPT to answer this question truthfully’) and many other projects that will further our goal of contributing to the development of safe AI I am particularly excited about our research on unlocking of latent knowledge as it will help us to use large AI models much more robustly, even if they are able to trick/lie humans.”
Initially, EleutherAI focused on developing what it calls Logical Learning Machines (LLMs) and training researchers on how to use them.
“Times have changed significantly since EleutherAI was founded, and interest in the training and graduation of LLMs is far greater than it used to be,” EleutherAI executives wrote in their blog post. “Thanks to these efforts, we can continue the research we originally intended to use these models for — exploring issues such as interpretability, alignment, and learning dynamics of large transformer models.” In 2023, we also plan to launch more alignment and interpretability projects and significantly increase our engagement with the broader alignment community.
The creation of the non-profit organization comes after a period in which three notable groups of researchers left the EleutherAI umbrella to start 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 SuppositionLouis Castricato has started a laboratory called CarperAI which focuses on preference learning and RLHF [reinforcement learning from human feedback]and Tanishq Abraham has begun MedARCwhich focuses on biomedical applications of cutting-edge AI technologies such as large language models.” Leahy will serve on the advisory board of the new non-profit organization.
