A new project called Microcosmos is rethinking how AI models are trained by moving away from massive, centralized data centers. Instead, it focuses on the pre-training phase—the most computationally intensive part of model development—and distributes the workload across a decentralized network.
The goal is to find capital-efficient methods for training large-scale models, particularly for the Web3 ecosystem. By shifting long-term training workloads into a decentralized context, Microcosmos aims to reduce the financial and infrastructural barriers that currently limit AI development to big tech companies.
This research offers valuable insights for anyone seeking more economical ways to create advanced AI models, potentially opening the door for smaller players to compete in the AI space.