A startup called Prime Intellect is betting it can build an AI forum that that combines the industrial power of big private-sector ventures with the public ideals of academic research. On Tuesday, the firm announced it had raised a $5.5 million seed round that will be used to launch a decentralized platform on which researchers can collaborate and build open-source AI models. The platform, currently live in beta, aggregates surplus global compute power to build AI models using distributed training across clusters.
“The goal is to create a new wave of AI progress that benefits scientific progress and community-driven AI models,” cofounder Vincent Weisser told Fortune. Weisser sees a divide when it comes to AI research between the academics in universities lacking computing capacity and the Silicon Valley giants working behind closed doors. “There’s no middle ground.”
“The challenge for AI developers and startups is they have to decide: What are the cheapest ones? Do they have the chips I need? And we solve that,” Weisser said. Distributed Global and CoinFund led the raise, closed a few weeks ago, with participation from Compound, Collab+Currency, and Juan Benet with Protocol Labs. Weisser told Fortune he hopes the platform will be ready to launch to the public in a matter of weeks.
By sourcing computing power from across the world, Prime Intellect hopes to offer an affordable solution to AI startups and researchers who lack industrial-size data centers. The hope is to form distributed training frameworks, whereby models are built in a collaborative manner across global, heterogeneous hardware. This will form a decentralized protocol for collective ownership of AI models “that benefits all of humanity,” the company said, making large-scale AI development more accessible and not beholden to Big Tech.
“Anyone will be able to contribute compute, capital, and code to collaboratively train open AI models and share in their ownership and benefits. We envision this enabling a new wave of open AI progress, from community-driven language models to scientific breakthroughs,” the company said in a statement.
The Prime Intellect protocol will offer the infrastructure and building blocks for decentralized AI development by rewarding participants for contributing compute, code, and capital, and enables collective governance of the model. The company currently is locating cloud GPUs (electronic circuits that can perform mathematical calculations at high speed) that allow users to train across each node.
When asked “Why now?” Weisser referred to Gemini, the latest AI model from Google’s DeepMind. Google was unable to build it with one compute class and a centralized cluster, and it was instead trained across different classes, he explained. “There’s not enough compute in one place. You need to use all the computers available to not only get the cheapest compute but also to actually get to these advanced models,” he added.
Weisser envisions the platform operating like a marketplace for GPUs, with users contributing reviews for each compute provider, rating speed and reliability “like finding a flight or hotel.”
By building on a decentralized platform, the data and code used to train a model will be transparent. In an increasingly saturated and competitive market, could the ease with which fellow researchers can access one’s work dissuade potential participants fearful of losing a comparative advantage?
“There’s a lot of people that ideologically want to create open AI models, and they’re looking for new ways to get compute, right?” Weisser replied. “They don’t get them from their universities or commercialize them in a startup. So, they’re looking for a third model to do so—and that’s what we want to enable through our platform.”
