[ANN] ResonanceNet — Decentralized AI Training with Bit-Exact Verification

[lattica-core]

ResonanceNet — Decentralized AI Training Network

ResonanceNet is a protocol for training language models the way Bitcoin mines a chain: many untrusted machines, one agreed result, every step checkable by anyone.

The core property is bit-exact deterministic replay. A verifier re-runs someone else's training from the same weights and the same derived batches and must get byte-identical output. No thresholds, no tolerance — equality.

What works today
• 400M parameter model (d1024, 24 layers, 32K vocab, 16K context)
• Training on 147GB of FineWeb-Edu (first loss: will insert)
• Full DiLoCo round between independent workers: contribution → aggregation → checkpoint → chain extends
• 475+ tests, including adversarial audit findings
• Genesis → derived weights → live model, verifiable by anyone in 1.5 seconds
• IPv4 + IPv6 dual-stack networking
• Corpus: 7.36TB FineWeb-Edu, Merkle-verified, downloading on seed node

Architecture scales to 29B
Mixture of Experts: 29.4B total parameters, 3.1B active per token. Fits a consumer 24GB GPU. Inference on phone at 100+ tok/s. Shape is described, pinned by tests, not yet trainable — how workers divide a sharded mixture is unsolved.

The experiment
Extreme overtrain: training a compact model on 1000× more data than scaling laws call "optimal". If a 29B MoE on 100TB+ of data competes with frontier models on code and reasoning — that changes the economics of AI. If not — we get data nobody else has.

What is NOT ready
• No public seed node yet (coming days)
• No incentive layer — contributions are not rewarded
• MoE training architecture unsolved
• Chain does not persist across restarts

Hardware
16GB VRAM minimum (RTX 4090, 5070, 5090, A5000+)

Training time estimate
2 workers: ~6.5 days to Chinchilla-optimal
8 workers: ~38 hours
64 workers: ~4.8 hours

Written from scratch, solo. Originally C++, rewritten in Python for cross-platform support.

GitHub: https://github.com/Kristian5013/resonancenet
Discord: https://discord.gg/ZBwDSQUh4
Telegram: https://t.me/resonancenet_main

By Kristian Pilatovich — creator of Lattica, the first ML-DSA post-quantum cryptocurrency.

[makertronic]

Kristian,

The architecture behind ResonanceNet and the concept of bit-exact deterministic replay for ML are impressive.

However, the Lattica community is still waiting for concrete delivery. Miners are deploying real hardware and spending actual money on electricity to support Lattica. You owe it to them to bring that project to completion.

Lattica addresses a critical space with post-quantum cryptography (ML-DSA). Focus on delivering and consolidating Lattica before spreading yourself thin on ResonanceNet.

[lattica-core]

Kristian,

The architecture behind ResonanceNet and the concept of bit-exact deterministic replay for ML are impressive.

However, the Lattica community is still waiting for concrete delivery. Miners are deploying real hardware and spending actual money on electricity to support Lattica. You owe it to them to bring that project to completion.

Lattica addresses a critical space with post-quantum cryptography (ML-DSA). Focus on delivering and consolidating Lattica before spreading yourself thin on ResonanceNet.

Thank you for the feedback, and fair point.

I work on both projects simultaneously — Lattica is my primary focus, and I switch to ResonanceNet when there are no active tasks on Lattica. The two don't compete for attention: Lattica's mainnet is live, the network is running, pools and exchange are operational. When the chain is healthy and there's no pending code to ship, that's when ResonanceNet gets built.

Lattica is not being neglected — v0.1.3 shipped this week, and the community (miners, pool operators, exchange) continues to grow. ResonanceNet is where the idle hours go, not the committed ones.