Decentralized AI

Projects take several approaches. Some coordinate open model training across many contributors' machines, others run inference on distributed nodes, and others build markets where model developers compete and are rewarded according to how useful validators judge their outputs. Blockchains handle payments, record contributions and rewards, and sometimes anchor proofs about which model produced a result.

The motivation is a reaction to the concentration of AI in a few large firms that control data, compute and the most capable models. Supporters argue that open, permissionless networks can reduce censorship, share rewards with contributors and make model behaviour more auditable. The sector also attracts investors through tokens tied to these networks.

The gap between vision and reality is wide. The leading models are trained on huge centralised clusters, and distributed training over the internet remains far less efficient, though research is progressing. Incentive systems can be gamed by participants who copy outputs or collude, and evaluating model quality on-chain is hard. A token and a whitepaper do not make a network decentralized if the team controls the infrastructure that does the work.