README
🧠 Brave New World: On-Chain Reinforcement Learning
A synthesis of blockchain-native AI training, consensus learning, and the emerging frontier of on-chain reinforcement learning — built on the convergence of Solana, decentralized data, and open-weight models.
📚 Origins
This article synthesizes and extends:
- Blockchain & AI — The Convergence of Blockchain & AI, Bitcoin × Solana
- Decentralized AI Training Using Blockchain — Architecture, federated learning, key components
- A Brave New World — Consensus Learning — Flare Research's CL paper, gossip protocol, blockchain-native ML
And builds toward:
- DeepSolana — The first open-weight Solana-native language model, the reference model for ORL fine-tuning
What is Brave New World?
Brave New World is a living research document and roadmap for On-Chain Reinforcement Learning (ORL) on Solana — a framework where AI agents learn, improve, and are rewarded entirely on decentralized infrastructure, with every training step recorded, verified, and incentivized by a public ledger.
It is not a single product. It is a direction.
The Core Idea
Agent observes on-chain state
→ takes action
→ receives reward (smart contract)
→ transition written to on-chain replay buffer
→ aggregator updates shared policy
→ participants rewarded proportional to contribution
→ repeat
The blockchain does not just store the model. It is the model's teacher.
Why Solana?
| Property | Value for ORL |
|---|---|
| 400ms block times | Near-real-time environment steps recorded on-chain |
| < $0.001 tx cost | Economically viable to log millions of training steps |
| Smart contracts | Transparent, programmable reward functions |
| Compressed NFTs | Cheap versioned model checkpoints at scale |
DeepSolana — The Reference Model
DeepSolana is the first open-weight model in this lineage:
- Trained on Solana transaction data, protocol docs, and on-chain events
- Distributed via Ollama for local inference
- Foundation for ORL experiments on Solana's live data stream
ollama run 8bit/DeepSolana
12-Month Roadmap
Q3 2026 — Foundations
- DeepSolana v1 fine-tuned on Jupiter transaction dataset
- On-chain replay buffer prototype using Solana accounts
- First consensus learning testnet: 3–5 nodes, gossip protocol
Q4 2026 — Incentive Layer
- Token-gated participation (staking to contribute training steps)
- Smart-contract reward oracle: DeFi-native reward signals
- Byzantine-fault-tolerant aggregation with slashing
Q1 2027 — Scale
- 50+ node consensus learning network
- Compressed checkpoint storage (cNFTs for model versioning)
- Cross-chain reward signals: Ethereum and Bitcoin bridged to Solana ORL agents
Q2 2027 — Open Ecosystem
- Public ORL API: any developer can define a reward function and spawn a training run
- DeepSolana v2: fine-tuned via ORL on six months of live Solana data
- Integration with Bittensor subnets for cross-network model evaluation
Consensus Learning — The Bridge
Between today's federated AI training and tomorrow's full ORL lies Consensus Learning (CL), introduced by Flare Research:
- Participants never share raw data or model weights — only predictions
- A gossip protocol aggregates predictions into a consensus output
- Blockchain records every contribution and enforces reward distribution
- Byzantine-resilient: malicious nodes are filtered by confidence-weighted aggregation
ORL extends CL into the temporal domain — agents learning from actions over time, not just predictions on static datasets.
Related Work
| Project | Relationship |
|---|---|
| Bittensor | Incentivized subnet architecture for AI inference |
| FLock.io | Federated fine-tuning with on-chain rewards |
| Ritual | AI coprocessor for smart contracts |
| solanaclawd | Sovereign AI agent runtime on Solana |
| DeepSolana | Open-weight Solana-native base model |
Source Article
Full article: docs/brave-new-world.md
Original research: Blockchain & AI GitBook
Built on OpenClawd · Solana · MIT License
🦞 Drift in ambiguity. Beach before harm. Earn before survival. Truth before strangers.
Contributors
6 commits
Similar repos
README
🧠 Brave New World: On-Chain Reinforcement Learning
A synthesis of blockchain-native AI training, consensus learning, and the emerging frontier of on-chain reinforcement learning — built on the convergence of Solana, decentralized data, and open-weight models.
📚 Origins
This article synthesizes and extends:
- Blockchain & AI — The Convergence of Blockchain & AI, Bitcoin × Solana
- Decentralized AI Training Using Blockchain — Architecture, federated learning, key components
- A Brave New World — Consensus Learning — Flare Research's CL paper, gossip protocol, blockchain-native ML
And builds toward:
- DeepSolana — The first open-weight Solana-native language model, the reference model for ORL fine-tuning
What is Brave New World?
Brave New World is a living research document and roadmap for On-Chain Reinforcement Learning (ORL) on Solana — a framework where AI agents learn, improve, and are rewarded entirely on decentralized infrastructure, with every training step recorded, verified, and incentivized by a public ledger.
It is not a single product. It is a direction.
The Core Idea
Agent observes on-chain state
→ takes action
→ receives reward (smart contract)
→ transition written to on-chain replay buffer
→ aggregator updates shared policy
→ participants rewarded proportional to contribution
→ repeat
The blockchain does not just store the model. It is the model's teacher.
Why Solana?
| Property | Value for ORL |
|---|---|
| 400ms block times | Near-real-time environment steps recorded on-chain |
| < $0.001 tx cost | Economically viable to log millions of training steps |
| Smart contracts | Transparent, programmable reward functions |
| Compressed NFTs | Cheap versioned model checkpoints at scale |
DeepSolana — The Reference Model
DeepSolana is the first open-weight model in this lineage:
- Trained on Solana transaction data, protocol docs, and on-chain events
- Distributed via Ollama for local inference
- Foundation for ORL experiments on Solana's live data stream
ollama run 8bit/DeepSolana
12-Month Roadmap
Q3 2026 — Foundations
- DeepSolana v1 fine-tuned on Jupiter transaction dataset
- On-chain replay buffer prototype using Solana accounts
- First consensus learning testnet: 3–5 nodes, gossip protocol
Q4 2026 — Incentive Layer
- Token-gated participation (staking to contribute training steps)
- Smart-contract reward oracle: DeFi-native reward signals
- Byzantine-fault-tolerant aggregation with slashing
Q1 2027 — Scale
- 50+ node consensus learning network
- Compressed checkpoint storage (cNFTs for model versioning)
- Cross-chain reward signals: Ethereum and Bitcoin bridged to Solana ORL agents
Q2 2027 — Open Ecosystem
- Public ORL API: any developer can define a reward function and spawn a training run
- DeepSolana v2: fine-tuned via ORL on six months of live Solana data
- Integration with Bittensor subnets for cross-network model evaluation
Consensus Learning — The Bridge
Between today's federated AI training and tomorrow's full ORL lies Consensus Learning (CL), introduced by Flare Research:
- Participants never share raw data or model weights — only predictions
- A gossip protocol aggregates predictions into a consensus output
- Blockchain records every contribution and enforces reward distribution
- Byzantine-resilient: malicious nodes are filtered by confidence-weighted aggregation
ORL extends CL into the temporal domain — agents learning from actions over time, not just predictions on static datasets.
Related Work
| Project | Relationship |
|---|---|
| Bittensor | Incentivized subnet architecture for AI inference |
| FLock.io | Federated fine-tuning with on-chain rewards |
| Ritual | AI coprocessor for smart contracts |
| solanaclawd | Sovereign AI agent runtime on Solana |
| DeepSolana | Open-weight Solana-native base model |
Source Article
Full article: docs/brave-new-world.md
Original research: Blockchain & AI GitBook
Built on OpenClawd · Solana · MIT License
🦞 Drift in ambiguity. Beach before harm. Earn before survival. Truth before strangers.
Similar repos
Linked in READMEs
Contributors
6 commits