How Hylon works

One AI, five layers, a million devices

Data flows up as privacy preserving signal, never as raw content. Each layer has a distinct responsibility and a distinct trust boundary, from the physical substrate to the immutable oversight plane.

Architecture

The five layer system

Layer 5
Safety and emergence monitor

The immutable oversight plane. Verifies the level zero foundations at every step, watches for capability emergence and objective gaming, and holds the final release gate and kill switch. It cannot be modified by the system it supervises.

Layer 4
Self improvement engine

The automated research loop plus the champion challenger release. Cheap experiments find the recipe; validated recipes trigger the expensive runs. A challenger is promoted only on margin, significance, and zero safety regressions.

Layer 3
Sovereign model, the Hylon core

The global model: a mixture of experts core with a router, retrieval and long term memory, tool use, agent orchestration, and specialist heads. Trained on the GPU tier and distilled into the Personal AIs.

Layer 2
Personal AI layer

A compact on device model per user. Learns your context locally, serves private inference, personalizes, and emits only privacy preserving learned signal upward.

Layer 1
Device substrate

A hardware abstracted runtime across phones, PCs, Macs, and GPU nodes. Metered and verified compute. Raw user data lives and dies here, the hard privacy boundary.

Sovereign stack

It does everything itself

Internally Hylon is a system, like every strong AI. To you and to the network it presents as one self contained assistant with no hard dependency on any external provider.

Mixture of experts core

Sparse expert routing so the model scales capacity without scaling every token, the standard shape of a strong model.

Retrieval and memory

Long term memory and retrieval give the model durable context beyond a single prompt.

Tool use and agents

The model calls tools and orchestrates multi step agents, so it can do the whole job rather than answer in a box.

Specialist heads

Domain specific heads for the wedge tasks where Hylon aims to be provably first.

Privacy by design

Your context stays on your device

The Personal AI learns from your messages, files, and habits locally. Raw data never leaves the device. What the network learns from is signal, not data: privacy preserving model updates shared through federated learning with differential privacy and secure aggregation, so no individual can be reconstructed.

We treat privacy as an engineering guarantee and a documented legal posture, with defenses evaluated against membership inference and model inversion. We do not claim to be outside the law. We are defensible within it.

Data flow, upward
01On device raw data, Personal AI learns
02Local update differential privacy noise added
03Secure aggregation server sees only the sum
04Global model no individual reconstructable
Decentralized compute

Two tiers, verified work only

Decentralized by geography and ownership, not by pretending phones can train. The bandwidth problem was solved by low communication training; the mountain is aggregate compute.

Tier 1

Training on GPU clusters

Datacenter grade GPUs, a DAO owned core plus staked operators, coordinated over the internet with low communication training that cuts network needs by orders of magnitude. Bandwidth is solved; memory and aggregate compute are the real limits.

Tier 2

Devices doing what they can

A million phones, PCs, and Macs serve inference, personalize on device, score reinforcement signal, and contribute data. Verified, never rewarded for idle uptime.

Verification

Paying for real output, not uptime

Inference

Locality sensitive hashing of activations verifies a node truly ran the model, at a fraction of the cost of re running it.

Training

Economic and statistical scoring on held out batches, plus redundancy, catch dishonest or low quality contributions.

High assurance

Trusted execution environments attest sensitive workloads where a customer needs a hardware guarantee.

Objective hierarchy
Level 0
Truth, safety, honesty, integrity. Immutable.
Level 1
Primary objectives. Fixed.
Level 2
Capabilities. Optimizable.
Level 3
Metrics. Indicators only.
Self improvement

It improves and releases itself, within guardrails

An automated research loop proposes recipes, runs cheap experiments, and trains the next candidate. When a challenger provably beats the live model on a fixed gate suite and clears every safety check, it is promoted to a new named version automatically, rolled out from shadow to full with automatic rollback, recorded on chain.

The honest line

The level zero foundations cannot be self modified, and a safety council holds the final release gate and a kill switch. The ambition is maximal; the blast radius is bounded.

Read the full technical whitepaper

Every mechanism on this page is specified in depth, with diagrams, in the Hylon whitepaper.

Read the whitepaperRun a node