## ResonanceNet: Decentralized AI Training via Proof-of-Training Blockchain - Technical Deep-Dive for YouTube Creators
Executive Technical Summary
ResonanceNet introduces a novel approach to AI model training by leveraging a decentralized blockchain. This experimental Layer 1 blockchain employs a Proof-of-Training (PoT) consensus mechanism, where "miners" compete to improve a shared neural network model's validation loss. Each block contains a verifiable model checkpoint, effectively accumulating training progress on the blockchain. The key shift lies in distributing AI model training and inference, moving away from centralized API-driven services. For YouTube creators, this signals a potential future where AI-powered content creation tools could be trained and distributed in a decentralized manner, impacting content generation, moderation, and potentially rights management. The immediate weight is moderate, as the technology is nascent, but the long-term implications for decentralized content creation are significant. The potential for AI models to be trained on creator-sourced data and then deployed for content generation, moderation, and rights management presents a paradigm shift from current centralized AI services.
Structural Deep-Dive
Proof-of-Training (PoT) Mechanism
ResonanceNet's core innovation is its Proof-of-Training (PoT) consensus. Instead of traditional hashing, miners train a MinGRU neural network. A block is considered valid only if the new model checkpoint improves the validation loss compared to the previous block. This ensures that every block contributes to the model's intelligence.
Verifiable Model Checkpoints
Each block includes a Keccak-256d hash of the model checkpoint. This allows any node in the network to independently verify the training progress and the integrity of the model. This is crucial for ensuring trust and transparency in the decentralized training process.
Continuous Model Growth
The network autonomously adjusts the model's architecture (layers, dimensions) at predetermined block heights. This allows the model to evolve and improve its capabilities over time without centralized intervention.
UTXO Expiry
To prevent state bloat, ResonanceNet implements a UTXO expiry mechanism. Coins expire if the network's validation loss improves by 10x from their creation point. Expired funds are returned to mining rewards. Sending coins to oneself resets the timer, allowing for continued use of the UTXO.
Architecture and Layering
ResonanceNet employs a strict layered architecture (L0-L7) to maintain modularity and separation of concerns. Each layer depends only on lower layers, ensuring stability and maintainability.
- L0 (rnet_util): Logging, filesystem, hex encoding, serialization.
- L1 (rnet_crypto): Keccak-256d, Ed25519, BIP32/39 HD keys, AES-256, ChaCha20.
- L2 (rnet_primitives, rnet_script, rnet_consensus): Blocks, transactions, script interpreter, consensus rules.
- L3 (rnet_chain, rnet_mempool): UTXO set with expiry tracking, skip-list indexed mempool.
- L4 (rnet_training, rnet_wallet): MinGRU training loop, HD wallet with coin selection.
- L5 (rnet_net, rnet_rpc, rnet_miner): P2P networking, JSON-RPC server, PoT mining coordinator.
- L6 (rnet_node, rnet_interfaces): Node lifecycle, interface abstractions.
- L7 (rnet_gpu, rnet_inference, rnet_lightning): GPU backends, model inference, payment channels.