Built on the live OrbNet network

A sovereign AI you own,
that no lab can switch off.

Hylon turns a million user owned devices and a community owned compute cluster into one self improving AI. Private by construction. Strongest where the frontier labs cannot reach. Powered by ORB.

Run a nodeRead the whitepaper
1M
Target device nodes
4
Verifiable gates
5
Layer architecture
ORB
One network token
Why Hylon can win

Frontier labs share three blind spots that no budget removes

Their strength is concentrated in one dimension, raw compute and scale. Hylon fights on the ground those labs cannot structurally contest, then compounds toward the frontier.

Blind spot 01

No access to your private context

A centralized lab cannot ethically or legally centralize your full private life. It meets you as a stranger every session. Hylon's Personal AI learns your context on your own device, where the raw data never leaves.

Blind spot 02

Absent where the world lives

Across censored and underserved markets the leading assistants are blocked, throttled, or barely fluent in the local language. Hylon is already inside these markets through OrbNet, in Farsi, Arabic, Russian, Turkish and more.

Blind spot 03

Owned by shareholders, not users

The value users create accrues to a lab's shareholders. Hylon is owned and governed by its network, with the people who run it earning ORB for verified work.

The honest line

Out competing every centralized lab is the destination. You do not win head on tomorrow; you earn the frontier fight by winning the wedge first, then compound the data, revenue and treasury it generates. Out data now, out compute later.

The unbuyable moat

A data flywheel money cannot replicate

A million Personal AIs learn from real user tasks in languages and markets no Western lab can collect. Raw data never leaves the device. Only privacy preserving learned signal is aggregated, through federated learning with differential privacy and secure aggregation. The stronger global model ships back to every device, and the loop deepens.

A competitor can raise more money every year. It cannot legally or physically assemble this corpus.

The loop
  1. 01Personal AIs learn your context on device
  2. 02Real task signal reinforcement from what you actually do
  3. 03Private aggregation differential privacy, secure aggregation
  4. 04Stronger model ships back to every device
Decentralized compute

Two tiers, decentralized by ownership, not by pretending phones can train

Bandwidth was solved by low communication training. Real runs have trained across continents on ordinary broadband. The mountain is aggregate compute, which is a treasury question.

Tier 1

Training

Datacenter grade GPU clusters, some owned by the DAO treasury and some added by staked operators, coordinated over ordinary internet with low communication training. Bandwidth is a solved problem; the real limits are memory and raw compute.

Tier 2

A million devices

Phones, PCs and Macs from the OrbNet base serve inference, personalize on device, score reinforcement signal, and contribute data. Every contribution is metered and verified, never rewarded for idle uptime.

Champion vs challenger
Champion
the live model, serving the network
Challenger
a new candidate from a validated recipe
Auto eval + safety
beat by a margin, zero regressions
Release
canary to full, auto rollback, on chain
Self improvement

An AI that improves and releases itself, within immutable guardrails

An automated research loop surveys the literature, proposes recipes, runs cheap experiments, and trains the next model. Cheap experiments find the recipe; expensive runs execute only a validated one. When a challenger provably beats the live model and clears every safety check, it is promoted to a new named version automatically.

Level zero foundations, truth, safety, honesty and integrity, cannot be self modified, and a safety council holds the final release gate. The ambition is maximal; the blast radius is bounded.

Defensibility

Five assets no competitor holds at once

01A live paying user base and revenue, where most networks launch with neither.
02Distribution and trust in markets the major AI platforms cannot serve, plus the censorship resistant technology to reach them.
03Production token and settlement rails with compute and AI rewards already built in.
04A cross platform native pipeline that puts an AI onto every device.
05A private, real task, multilingual data flywheel that money cannot buy at any centralized lab.
Roadmap

Every claim is tied to a gate we can prove

We claim a capability only once its gate is crossed. No dated promises of general intelligence.

Gate 0
~6 months
Sovereign stack live

Personal AI, compute, ORB and the DAO ship as one product on a top open model, self hosted, private, uncensored.

Gate 1
12 to 18 months
Wedge dominance

The continually trained model beats all open models in wedge domains and wins blind preference against the frontier on personal tasks.

Gate 2
Year 2
Self improving

A DAO owned cluster trains the model and the champion challenger pipeline promotes new versions automatically, under a safety gate.

Gate 3+
Year 3+
General frontier

Conditional on compute and validated recipes crossing an explicit threshold. Claimed only once it is crossed.

See the full roadmap

Own a piece of the network

Contribute compute, data or bandwidth from your devices and earn ORB. Or read the full technical whitepaper.

Run a nodeRead the whitepaper