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