Grassroots Researchers Race to Build an Open Source GPT-3 Alternative
A decentralized group called EleutherAI leads the charge to replicate and freely release a model comparable to OpenAI's massive GPT-3, which remains locked behind a commercial API.
OpenAI reveals GPT-3, a massive 175-billion-parameter language model that excels at writing code, making analogies, and mimicking historical authors. Unlike previous versions, OpenAI keeps this highly sophisticated model closed-source and limits access exclusively through a commercial API and a unique licensing deal with Microsoft.
In response to this restricted access, a grassroots collective known as EleutherAI steps up to recreate the technology. The group focuses on a project called GPT-Neo, aiming to publicly release the code and weights needed to run a model similar to GPT-3 so that the broader research community benefits from the breakthrough.
EleutherAI founders Connor Leahy, Leo Gao, and Sid Black initially start the project as a joke on a machine learning Discord server before turning it into a serious, hardware-intensive endeavor. They rely on massive amounts of computing power and a distributed team of researchers to prove that building powerful AI models remains possible outside of traditional corporate labs.