Creating state-of-the-art AI systems, for example ChatGPT, needs extensive technical capabilities due to their costliness of creating and operating. Whilst some open source attempts have aimed to reproduce confidential, closed-source tools invented by organisations such as Alphabet’s DeepMind and OpenAI, they often face obstructions as a result of a lack of funds and knowledge in the subject.
In an attempt to prevent this, the group known as EleutherAI have put together a nonprofit foundation. This organization declared that a research institution which doesn’t generate a profit, the EleutherAI Institute, will be established. Financing was supplied from Hugging Face and Stability AI, Nat Friedman from former GitHub CEO, Lambda Labs and Canva.
Stella Biderman, an AI researcher at Booz Allen Hamilton and part of EleutherAI Institute’s management, explained via email to TechCrunch that formalizing the organization permits them to hire a full-time staff and focus on more difficult, in-depth initiatives than just a volunteer group could manage. When it relates to having a non-profit, it is an easy decision due to their emphasis on research and open source projects.
EleutherAI originated a few years ago as a collection of developers committed to free AI analysis. Its originators —Connor Leahy, Leo Gao and Sid Black— coded the required scripts and accumulated the data essential for the formation of a near-identical machine learning algorithm to OpenAI’s GPT-3 for natural language generation, which was getting plenty of media exposure at the moment.
EleutherAI has collected and made freely available The Pile, a selection of data sets that can be employed to teach GPT-3-like models to finish text, generate code, and more. It has also granted permission for the usage of several models through the Apache 2.0 license, such as GPT-J and GPT-NeoX, language models which began to generate a fresh wave of startups for a period of time.
EleutherAI primarily trained its models using the TPU Research Cloud, a Google Cloud initiative which advocates the dissemination of research outcomes. Additionally, the United States-based cryptocurrency miner and cloud practices provider CoreWeave contributed computing power to EleutherAI in exchange for models that can be implemented and hosted by its clients.
EleutherAI underwent rapid development. Currently, more than twenty frequent contributors of the organization have devoted their time to research, and over the last one and a half years, numerous joint works have been created by the members of EleutherAI, including 28 academic papers, various taught models, and ten released codebases.
EleutherAI had initially wanted to put out a model with the same amount of parameters as GPT-3, but due to technical and financial difficulties, those plans had to be dropped. Parameters measure the complexity of the model when tackling a task, such as developing text. It is stated that…
In the year 2022, EleutherAI became familiar with Stability AI, the recently funded startup responsible for the image-producing AI program Stable Diffusion. Along with other partners, it contributed to the original model of Stable Diffusion. Since then, Stability AI has donated a portion of its Amazon Web Services cluster for EleutherAI’s language model study.
When Hugging Face made contact with EleutherAI, conversations began among the nonprofit personnel, according to Biderman. Members of EleutherAI had been taking part in the BigScience task that aimed to create and make public a model similar to GPT-3 during the course of a year.
Biderman stated that EleutherAI has mainly concentrated on substantial language models which are structurally comparable to ChatGPT in the past and they most likely will keep doing so. Aside from training considerable language models, they are looking forward to allocating more resources to ethical, interpretable and attuned work.
One may question if firms like Stability AI and Hugging Face, which are funded by significant venture capital, could affect EleutherAI’s research outcomes. There could be grounds for this assumption, as research has found a link between donations to non-profits and their capacity to speak up about proposed rulemaking from the government.
Biderman insists that the EleutherAI Foundation will stay independent and does not perceive an issue with the donor base up to now.
Biderman declared that they do not make models for businesses. He added that being funded by various tech companies actually bolsters their independence rather than if only one company was providing all the money.
The EleutherAI Foundation will have to take extreme caution to make sure it doesn’t become bankrupt. The OpenAI company is a notable example, as they had to switch to a “capped-profit” model after being founded as a not-for-profit in the year 2015 to enable their continuous research.
Generally, the effects of nonprofit organizations investing in artificial intelligence research have been varied.
The late Microsoft co-founder Paul Allen founded The Allen Institute for AI (AI2) with the goal of making scientific advancements in the areas of AI and machine learning. The U.K. government has also funded The Alan Turing Institute, a research institute for data science and machine learning. Additional smaller and up-and-coming initiatives include the AI startup Cohere’s Cohere For AI and Timnit Gebru’s Distributed AI Research, a network for distributed research.
However, Eric Schmidt, the former Google chairman, has his own Artificial Intelligence research fund, which is worth more than $125 million. This stirred up new arguments when Politico reported that Schmidt plays an inordinate amount of influence on the White House’s Office of Science and Technology Policy.
It is uncertain what the EleutherAI Foundation will eventually become, however it is likely it will develop in a constructive way in the future.