A New Way to Train Big AI Models in Web3

If you follow crypto and artificial intelligence, you know that merging the two sounds great on paper. Web3 apps generate tons of useful data, but training massive AI models on a blockchain is surprisingly tough. I recently came across a study that looks into this exact problem. It turns out that basic user hardware often struggles to handle huge models across a decentralized network.

To tackle this issue, researchers created a new algorithm called DICE for blockchain-powered federated learning. They wanted to figure out how to let Web3 participants customize big pretrained models without running out of computational power or memory. Their approach uses a dynamic progressive learning scheme and a model slice verification scheme so people without full model capacity can still join in. They also introduced a tree-structured committee consensus protocol and an efficiency-driven incentive scheme to help the system scale to many data holders while staying safe. The researchers tested their framework using theoretical analysis along with experiments on vision and language models. Their results show that DICE provides an efficient, effective and secure way to customize large models in Web3.

Curious about the details? Look up DICE: Decentralized Incentive-driven Federated Learning for Efficient & Scalable Customization of Pretrained Models in Web3 at arxiv.org/abs/2610.39132. The authors are Hong Lin, Zhongle Xie and Lidan Shou, with several coauthors.

This is not financial advice. The information provided is for educational and informational purposes only.

The cover image was generated by AI.