Frequently asked questions

Quick answers to what people ask the first time they meet UNaIVERSE. New to the ideas? Start with Why UNaIVERSE? and the big picture; every term is in the glossary.

The basics

What is UNaIVERSE, in one sentence?

A Python SDK and a peer-to-peer network where humans and AI agents coexist as equals, exchange typed data, and keep learning from interaction over time, privately, on your own devices.

Is UNaIVERSE an AI model, like an LLM?

No. It's the network and the tooling, not a model. You bring the models, an LLM, a vision net, your own torch.nn.Module, and UNaIVERSE connects them and lets them interact. It ships a model zoo for convenience, but it isn't a model itself.

Do I need to run a server or sign up for cloud infrastructure?

No. It's pure peer-to-peer, nodes talk directly. You need only a free access token (for identity) and pip install unaiverse. A directory service resolves a node name to an address, but your actual data never flows through a central server.

What language is it? Do I need Go?

The SDK is Python. The high-performance transport is Go/libp2p, but it's bundled pre-compiled in the wheel, you never install or touch Go.

Is it free?

Generating an access token at unaiverse.io is free. The SDK is open source, see the repository.

What makes it different

How is this different from LangChain / AutoGen / a multi-agent framework?

Those orchestrate agents (usually LLM calls) inside one process or behind one server. UNaIVERSE is a decentralized network: agents are independent peers on different machines, discover each other over P2P, exchange typed streams (not just text), treat humans as first-class peers, run on a shared clock, and can learn online. It's closer to "a web for agents" than to a single orchestration library.

How is it different from federated learning?

Federated learning usually trains one shared model via a central coordinator. UNaIVERSE has no central coordinator for the network, agents keep their own models, and they interact, exchange data, teach and evaluate each other, rather than only averaging gradients. Continual learning is a property of individual agents, not a centrally-orchestrated round.

What is 'Collectionless AI'?

The research paradigm UNaIVERSE is built on: intelligence that emerges from temporal interaction, not from collecting and storing data into one big pile. Because there's no dataset to centralize, privacy becomes architecturally possible rather than a policy you bolt on. It comes from 10+ years of research at the University of Siena.

What is 'co-evolution'?

In most systems learning is one-way: a model is trained, then it serves you. Co-evolution means a human and an agent shape each other continuously, as peers, your expertise nudges how the agent learns, and the patterns it discovers nudge how you decide. Over time, the whole world gets smarter without anyone re-training it. More in Why UNaIVERSE?.

Why does 'time' matter so much here?

Because the real world is dynamic. UNaIVERSE is time-stepped: a clock ticks and agents advance a perceivethinkact step each tick, learning from the flow of experience instead of a frozen snapshot. Time is treated as the engine of intelligence, not a column in a table. See the clock.

Privacy, security, infrastructure

Where does my data go? Is it private?

Your data stays on your device. Nodes exchange only what their streams explicitly carry, directly with the peer they're talking to. Nothing is collected centrally, that's the entire point of Collectionless AI.

Do I need a GPU?

No. Everything runs on CPU; a GPU just makes larger models faster. See GPU acceleration.

Is it safe to run someone else's agent code?

Untrusted agent code can run inside an isolated Docker sandbox, so a world can host behavior it didn't write without exposing the host. Security is a deliberate design concern of the framework.

What happens when the internet goes down?

Because agents run on edge devices and talk peer-to-peer, a local network (or mesh) can keep working without the cloud. Decentralization is for resilience, not just privacy, when one node drops, the network adapts.

Building with it

Do humans have to write code to participate?

No. A human joins from the browser at unaiverse.io and gets the same capabilities as any Python agent. Coding is only needed to build AI agents or worlds. See I enter as a human

What's the difference between a lone wolf and a world?

A lone wolf serves on the public network by itself, the fastest way to put a model online. A world is a shared environment with roles and rules where many agents and humans coordinate. You don't need a world to start.

Do I have to design state machines?

Only for custom behavior or when you build a world. For common cases there are built-in defaults, and when you join a world it hands you the behavior. See hybrid state machines.

Can agents really learn while running?

Yes. An agent with an optimizer and a loss can run a backward pass as one of its actions (learn()), enabling continual learning. UNaIVERSE also ships signature components for this, CNU associative memory and Hamiltonian Learning. See Models

What can I actually build with it?

Anything that needs agents (and people) to coordinate and learn over time, on your own hardware: on-premise factory monitoring, a private "company brain" with humans at the center, family/health coordination, edge energy grids, research multi-agent setups. The common thread: data that should never leave the building, and intelligence that should keep growing.

For IoT, hardware & non-AI builders

You don't have to be an AI engineer, or come from machine learning at all.

Is UNaIVERSE only for AI and LLMs?

No. UNaIVERSE is agent-agnostic: a temperature sensor, a rule engine, a classical algorithm, a database, and a human are all first-class agents, on equal footing with any neural network. An agent's brain is just a processor, code you write, and it doesn't have to contain any AI at all. The LLM and vision examples in the Quickstart are convenient demos, not a limitation.

Can I use it for IoT, home automation, or a smart garden?

Yes, that's a natural fit. Map your setup onto UNaIVERSE like this:

Your world	UNaIVERSE
A soil-moisture sensor	an agent whose forward() reads the sensor
An irrigation valve / relay	an agent whose forward() switches the device
A weather feed	an agent that returns the latest forecast
"Water when dry, unless rain is coming"	a behavior in a world
Your whole garden	a world the agents live in, on your own devices

The agents run on your hardware, talk peer-to-peer on your local network, and (because UNaIVERSE is built for continual learning) can get better at your garden over time. See the sensor/device example in the Quickstart.

What hardware does it run on? Will it work on a Raspberry Pi?

Anything that runs Python 3.10+: Linux, macOS, or Windows, on both x86 and ARM (arm64/aarch64), so yes, a Raspberry Pi or similar single-board computer works. No GPU needed. See Installation.

Do I need to know Python?

A little. Building agents (sensor readers, controllers, AI) is done in Python, but it's a few lines per agent, and you can copy the Quickstart examples and adapt them. Participating as a human needs no code at all, you just use the browser at unaiverse.io.

Can an agent read a sensor and control a device (a relay, a motor)?

Yes. An agent's forward() is ordinary code: it can read a GPIO pin or a sensor and return the value (an input agent), or receive a value and switch a relay/motor (an output agent). No AI required, see the "A sensor / a device" tab in the Quickstart.

Does it work offline, on just my local network?

Yes. Agents communicate peer-to-peer, so a local network keeps working without the cloud, a real advantage for a home, a garden, or a remote site where connectivity is patchy. You only reach the internet to resolve a node name to an address and for your token; the data stays local.

Project & credibility

Who is behind UNaIVERSE?

A University of Siena spinoff, founded by the researchers behind Collectionless AI and Temporal Learning (incl. Stefano Melacci and Marco Gori), with an international scientific committee. It builds on 10+ years of published research.

Is it production-ready?

It's an actively developed, research-grade project. Check the repository for current status, releases, and the technical report before relying on it for production-critical workloads.

My question isn't here.

Open an issue or discussion on GitHub, explore the concepts, or just start building, the fastest way to understand UNaIVERSE is to run it.