I built a Proof-of-Training L1 blockchain where miners train a shared MinGRU neural network instead of computing SHA-256 hashes. Every block makes the model smarter.

Key differences from Bitcoin: - Miners compete on validation loss improvement, not hash targets - Each block contains a verifiable model checkpoint - The network produces a publicly available AI model as a byproduct - MinGRU architecture is ~388x more parameter-efficient than Transformers

Technical details: - 65K lines of C++20, builds on Windows/Linux/macOS - 4 GPU backends: CUDA, Vulkan, Metal, CPU fallback - Lightning Network for instant payments - Ed25519 signatures, Keccak-256d hashing, Bech32 addresses - Continuous model growth: d_model starts at 384, grows +2 per block - Supply cap: 21M coins, 10-minute blocks, supply-based halvings

The model grows with the network. Early blocks train a small 26M param model. As more miners join over years, it scales toward billions of parameters — and by then, phones will be powerful enough to run it locally. No API keys, no subscriptions, no censorship.

Genesis message: "OpenAI burns $14B in 2026, adds ads to ChatGPT as last resort"

Whitepaper: https://kristian5013.github.io/resonancenet/ResonanceNet_Whi...