Macula Mesh Use Build Macula for… Docs Contact

Macula · sectors · All Macula for… pages

Macula for Training AI

Macula lets you train small neural networks for control, robotics and classification on GPUs and CPUs that households and businesses rent out, reached over the Macula Mesh, with no open port on any machine and no cloud provider in the data path. Neuroevolution sends candidate networks out and gets scores back, so the work runs on ordinary machines. When private data scores the networks, each data owner breeds its own population on its own machines and data, and only the best genomes move between sites over the mesh. We are looking for partners to build it with.

Without Macula: rent cloud GPUs and copy your data to them

Teaching a network to steer a robot, run a control loop or play a game well takes thousands of trial runs. That usually means renting GPUs from a cloud provider, while capable hardware in homes and offices sits idle most of the day.

billed by the hour

Cloud GPU rental bills by the hour, even for small networks

You pay for a cloud provider's GPUs by the hour, even when the networks are small enough to fit on a single ordinary computer.

data leaves its owner

Cloud training needs your data copied to the provider

Sensor logs, plant recordings and customer data get copied to wherever the training runs. Many owners are not allowed to do that, or will not.

strangers' code

GPU rental marketplaces run strangers' containers on your machine

A gaming PC, an office workstation or a small server could do useful work overnight. On a GPU rental marketplace anyone who pays can run any program on your machine.

Neuroevolution needs no fast links between machines

Neuroevolution trains a population of networks by trying each one, keeping the best and breeding the next generation from them. That makes it very different from the gradient training behind large language models, which keeps many GPUs in constant conversation over fast links that no home uplink can match. So this is not a way to train an LLM. It suits the small, specialised networks that control, decide and classify.

independent work

Each network is tried on its own

A worker needs a network's genome, the compact description the network is built from, and the task, nothing more. It runs the trial locally and returns a score.

small messages

Only genomes and scores travel

Every step of a simulation stays on the worker. What crosses the network is a compact genome going out and a number coming back.

fault tolerant

A worker that drops out costs one trial

If a machine goes offline mid-run, its network is simply tried again elsewhere. The population carries on.

any hardware

GPUs and plain CPUs both count

Small networks run well on ordinary processors, and machines of different speeds can work on the same population side by side.

With Macula: workers dial out, and only members of their realm can call them

The Macula Mesh is a network of stations. Stations link up to form the mesh, keep its phonebook (a Kademlia DHT: which node offers what, and where to reach it), and relay calls between the parties a realm admits. There is one mesh and many Macula Realms. Each realm is an identity space on the mesh with its own key: its operator admits anything that can be given an identity, such as people, organisations, services, machines and devices, gives each a certificate signed with that key, and can revoke it. Anyone can run a station, as many as they like; the more stations, the stronger the mesh. Macula runs a number of seed stations and a community realm; a company or sector body can run its own realm under a Macula licence. A worker machine joins a realm and offers one capability, evaluating a network; a trainer in the same realm calls it. Trainer and workers only make outbound connections; none of them accepts an incoming one. It is the same shape as Macula for Running AI, with training work instead of chat.

The trainer keeps the population and decides what to breed. It sends each candidate's genome to a worker its realm admitted, the worker runs the trial on its own GPU or CPU and returns a score, and the trainer breeds the next generation from the best. In this drawing the task is a simulator, so nothing about it is secret and any rented machine can run it. Private data trains differently, as the section after next explains.

With Macula: a marketplace inside a realm, where owners rent out machines and companies pay for training

Macula for Training AI is a marketplace inside a realm: owners of spare machines offer them, companies that need networks trained pay for the evaluations, and the realm has admitted both sides.

offer

Households and businesses rent out capacity

A family or a firm runs a worker on hardware it already owns and decides when it works and how much. The worker only makes outbound connections, so renting out capacity never means opening a port on the home network.

request

Industry brings the training jobs

A company that needs a controller, a policy or a compact classifier brings the task and its trainer, and pays for the evaluations its population uses. The machine's owner is paid for that work.

admitted

Both sides hold a certificate from the marketplace's realm

That realm's operator admitted every owner and buyer and can revoke either. A worker runs one thing only: scoring a network.

With Macula, private data trains as islands that swap their best networks

An island is one site's whole training loop: a local trainer that holds the population, local workers on its own machines, and the site's own data.

When a simulator scores the networks, nothing is private, and the trainer can rent workers anywhere on the mesh. When private data scores them, the data cannot leave its site, so the whole loop stays there: each site that holds data runs its own island. Islands still learn from each other; from time to time one sends copies of its best genomes to the others and takes theirs in, then breeds them into its own population and scores them against its own data. Only genomes cross the mesh, never a record of the data. Sites that cannot share their data still breed from each other's best networks.

Every island runs the same loop: its trainer sends genomes to its own workers, they score them against the site's own data, and the scores come back. The arrows to the mesh are outbound connections; what travels over them is a copy of an island's best genomes, never a record of its data.

real data

Each site trains on its own data

Plant recordings, sensor logs or customer data score networks only on the machines of the site that holds them. Nothing but genomes leaves the site.

migration

Islands swap their best genomes

Each island sends copies of its best networks to the others and breeds the ones it receives into its own population. A network that does well on one site's data gets tried on every other site's.

synthetic data

Simulators can go anywhere

When the task is a simulator or generated data, nothing about it is secret, and the trials can run on rented household capacity at whatever scale the marketplace offers.

trained models

A finished network can change hands

A genome that does well on one site's data is a small, finished model. Its owner can list it in the realm, and a buyer can score it against its own data before paying: the seller never sees that data, and the buyer does not buy blind.

Where it fits

Any task where a small network makes decisions and a program can score how well it did. An evolved network can be exported to ONNX, the open format that inference runtimes read, and served the way the Running AI page describes.

Control policiesRoboticsControl loops in industrial plantsGame AICompact classifiers for edge devicesSearching for a network's shape

Build it with us

Partners bring the machines, the training tasks and the sites; we bring the station software, seed stations, the evolution engine, which is open source, and a place in Macula's community realm or a realm of your own under a Macula licence. We are looking for partners in three roles.

rent out capacity

Households and businesses with machines

You own GPUs or spare CPUs that sit idle part of the day. Help us shape how a worker is offered, scheduled and rewarded, on hardware you already have.

bring a task

Companies with a network to train

You need a controller, a policy or a compact classifier, and you have a simulator or data that can score it. Bring the task, and keep your data where it is.

pilot or research

Pilot sites and research partners

A plant, a lab or a university group that wants to run trainings across the mesh with us, measure them, and publish what they find.

Evolve your first network

Want to try the engine before we talk? The two libraries are on Hex; you need Erlang/OTP 28 with rebar3, and a Rust toolchain for the network library's native speed-up.

%% rebar.config {deps, [ {faber_tweann, "~> 2.0"}, {faber_neuroevolution, "~> 1.2"} ]}.

You write one evaluator module that scores a network on your task; the library breeds the population around it. The quick start walks through it.