Pluralis Research is developing a decentralized approach to training large artificial intelligence models. Its core innovation is Protocol Learning, a method that allows multiple participants to contribute to training a foundation model without any single party ever possessing a complete copy of it. The company's work focuses on technical challenges including model-parallel training and operating over low-bandwidth networks.
The company has demonstrated the viability of its approach by successfully training an 8-billion parameter LLaMA model to a performance level matching centralized training. This achievement was accomplished using devices distributed across four different locations. Pluralis Research is positioning this technology to enable community-owned protocols for frontier-scale AI development, presenting an alternative to centralized models of AI training.
The company has secured $7.6 million in seed funding to advance its research and development.