gonka-ai/gonka

What it solves

It addresses the high cost and monopolistic nature of centralized cloud providers by providing a decentralized AI infrastructure for model training and inference. It aims to ensure that computational resources are used efficiently for actual AI workloads rather than being wasted on blockchain security.

How it works

The system uses a novel consensus mechanism called "Proof of Work 2.0," where nodes compete in "Sprints" using transformer-based models to solve AI tasks. A node's success in these races determines its voting weight, which in turn controls its share of AI workloads and its responsibility for validating others' results. To maintain trust without excessive overhead, the platform employs randomized task verification and a reputation system that rewards honest behavior and penalizes cheating. For training, it utilizes DiLoCo's periodic synchronization to enable geo-distributed training across independent hardware providers.

Who it’s for

• Developers who want to run LLM training and inference on a decentralized network.
• Hosts (Hardware Providers) who wish to contribute their computational resources to earn rewards.

Highlights

• Proof of Work 2.0: Aligns blockchain consensus with actual AI workloads using transformer-based models.
• Randomized Task Verification: Reduces verification overhead to 1–10% while maintaining probabilistic security.
• Geo-Distributed Training: Leverages DiLoCo for efficient, decentralized model training with minimal communication overhead.
• Reputation-Based Scrutiny: Dynamically adjusts verification frequency based on a node's history of honest contributions.