The Decentralised AI Vision
The Decentralised AI Vision
Hylon today is a hybrid support assistant. Its long-term vision, set out in the OrbAI technical specification, is far bigger: a privacy-first, community-owned intelligence built on a global network of user-run nodes, where the model learns from patterns and never from your raw data.
A Roadmap, Not Today's Product
This page describes where OrbAI is heading. Everything here is a forward-looking plan. What ships in the app today is the hybrid on-device plus cloud assistant described in the overview and the Using Hylon guide. The decentralised network, token rewards, and self-improvement engine are vision, not current features.
The Big Idea
Today's largest AI systems are trained in central data centers on data harvested from the open web. OrbAI proposes a fundamentally different design: a network of up to one million user-owned compute nodes, where each node runs a Personal AI that learns locally from its owner's data. The global OrbAI model then learns from the network of Personal AIs, never from raw user data.
The result is intelligence that is distributed instead of centralized, and privacy that is structural instead of promised. OrbAI stays two layers removed from any personal content: your data informs your Personal AI, your Personal AI shares only what it has learned, and OrbAI learns from those learnings.
Distributed, Not Centralized
No central GPU cluster. Compute comes from many user-owned OrbMesh nodes, coordinated by a lightweight cloud layer rather than a giant data center.
Private by Architecture
OrbAI never touches raw data. It learns from Personal AIs that have already learned from their users, keeping the global model two layers removed from anyone's content.
Community-Owned
The people who provide compute and connectivity share in the value the network creates, instead of one company owning all of it.
The Two-Layer Privacy Model
Privacy in OrbAI is not a policy bolted on afterward. It is built into how data moves through the system.
Your Data Stays on Your Device
Raw data, such as messages, documents, and browsing history, never leaves your device. It is read only by your own Personal AI, running on your own hardware under standard local app permissions.
Your Personal AI Learns Locally
The Personal AI processes your data on-device and turns it into learned patterns and model weights. The original content is never transmitted.
Only Learnings Are Shared
The Personal AI shares learned weights with the network, never the source data. Learned weights cannot reconstruct the original content.
OrbAI Learns From the Network
The global OrbAI model learns from one million Personal AIs at once. It is analogous to learning from a well-read tutor rather than reading every book yourself.
Why Two Layers Matter
Because OrbAI only ever interacts with AIs that have already learned, and never with raw content, the architecture is designed to keep the global model permanently separated from personal data, by construction rather than by trust.
How Learning Happens Without Data Leaving
The network trains using federated learning: devices contribute what they learned, never what they saw.
Federated Learning
Each Personal AI trains locally. Only weight updates leave the device, never data, and every update is encrypted before transmission.
Efficient by Design
Only significant updates are transmitted, and they are compressed to dramatically reduce the bandwidth needed to keep the network learning together.
Hierarchical Aggregation
Nearby nodes aggregate into local clusters, clusters roll up into regional coordinators, and regions synchronize periodically, with rotating coordinators so there is no single point of failure.
Resilient Coordination
Asynchronous updates let fast nodes contribute without waiting for slow ones, and peer-to-peer gossip lets the network heal automatically after a split.
Defense in Depth
On top of the two-layer design, the plan layers added protections: differential-privacy noise on shared learnings, end-to-end encrypted communication between AIs, secure enclaves for sensitive processing, and an open-source privacy layer that anyone can verify.
Safety That Cannot Be Edited Away
A self-improving system needs guardrails that the system itself cannot remove. OrbAI's design centers on an Objective Hierarchy whose foundation is immutable.
Five Immutable Foundations
At the base sit Truth, Consistency, Honesty, Safety, and Integrity. These are hardcoded and can never be changed by the system, no matter how capable it becomes.
Bounded Self-Improvement
A self-improvement engine can extend the system's capabilities, but every proposed change is checked against the hierarchy and tested before it is deployed, or rolled back.
Human Oversight
Significant changes require human approval. The oversight checkpoints, the hierarchy structure itself, and the verification requirements are all off-limits to self-modification.
Gaming Detection
If surface metrics improve while the core objectives do not, the system flags it as suspicious, so progress cannot be faked by optimizing the wrong target.
Earn by Powering the Network
The decentralised vision is also an economy. OrbToken is designed to bootstrap the network by rewarding the people who run it.
Run a Node, Earn Rewards
OrbMesh devices are designed to host a Personal AI and contribute compute to the network. Node operators are the largest planned share of token rewards, turning spare capacity into a decentralised data center.
A Fixed, Deflationary Supply
The plan sets a fixed total supply with the largest allocation reserved for the million node operators, plus a small burn on transactions so the token is deflationary over time.
The Flywheel
More nodes make the AI better, a better AI attracts more users, more users raise the token's value, and higher value attracts still more nodes.
Utility Across the Ecosystem
Tokens are planned to power AI usage, developer API access, enterprise licensing, premium features, and priority compute, all within one community-owned network.
Why This Approach
This is how OrbVPN intends to build powerful AI without harvesting personal data: keep raw data on the owner's device, learn only from patterns, reward the people who provide the infrastructure, and bind the whole system to safety rules that cannot be edited away.
From Today's Hylon to Tomorrow's Network
You do not have to wait for the full vision to benefit from OrbAI. The journey is incremental, and each step is genuinely useful on its own.
Today: Hybrid Assistant
An on-device model plus a local knowledge base answers common questions instantly and privately, with a cloud engine and human agents for everything else. This ships in the app now.
Next: Smarter On-Device Understanding
The local knowledge base is already built to store vector embeddings, the groundwork for on-device semantic search that understands meaning, not just keywords.
Then: Personal AI on OrbMesh
As OrbMesh nodes roll out, the plan is for each to run a Personal AI that learns locally for its owner, the foundation layer of the decentralised network.
Eventually: The Global OrbAI
With a network of Personal AIs in place, the global OrbAI model learns from the network, never from raw data, within the immutable safety foundations.
Honest About the Timeline
We are deliberately clear about what is real today and what is roadmap. The assistant you use now is the hybrid Hylon. The decentralised network, Personal AIs, and token rewards described here are the direction of travel, and we will keep these docs honest as each piece ships.
Privacy-First AI, Owned by Its Community
OrbAI's vision is intelligence without data harvesting: your data stays yours, the network learns from patterns, and the people who power it share in what it builds.