CrowdGPT launches a distributed crowdsourced training platform
Understand this story
CrowdGPT launches a distributed crowdsourced training platform that lets a large number of independent GPUs collaboratively train the same language model. Its process has four steps: the server assigns each device a 10MB shard of tokenized text data and the current global model weights, the client downloads the weights from HuggingFace and trains locally, and then only the model delta is sent back, which the server aggregates. CrowdGPT says the dataset is fully open and anyone can contribute data; model weights are automatically pushed to HuggingFace for public audit, use, and testing.
AI overview based on source reports · Updated 16 min ago
Coverage timeline
Follow the coverage to explore different perspectives.
- Hacker News AICrowdGPT – AI Owned by Everyone
CrowdGPT describes how it coordinates independent GPUs to train a single language model: the server assigns each connected device a 10MB data shard and the current global weights, the client trains locally and computes a delta, and the server validates submissions for NaN/Inf values, size matching the architecture, and acceptable delta norms to prevent poisoning. Valid updates are averaged into the global model periodically, and weights are pushed to HuggingFace for public auditing and testing.
Activity for this story
Not enough continuous observations to show a trend yet.