Artificial intelligence is fast becoming the core infrastructure of how we work, learn and
decide — and it is being built in a shape that concentrates power, cost and control in
remarkably few hands. A handful of firms own the frontier models, the data centres that run
them, and the terms on which everyone else may use them. Every question is metered, logged,
and routed through machines you will never see or govern.
None of this is a law of nature. It is a design choice — and design choices can be remade.
LowAI is the wager that intelligence can run on hardware people already own, on open models
anyone can inspect, with data that never leaves the device. Not merely because that is
cheaper — though it is — but because privacy, autonomy and sustainability should be the
floor a technology stands on, not a premium tier sold back to us.
The argument is structural, economic and ecological at once. Below: the flags worth raising,
the numbers worth seeing, and the openings worth taking.
🏡Tiny-house living
💻Runs on what you own
🌿Low consumption
🛠️Repairable & yours
Structural flags
🚩 Concentration of powerA few companies set the models, the compute and the price of access for everyone else.
🚩 Rented, never ownedTerms, prices and availability can change overnight on tools you don't control.
🚩 Metered surveillancePay-per-token means every prompt becomes a data point about you, stored elsewhere.
🚩 Ecological footprintThe race for ever-larger models scales energy and water use to industrial levels.
🚩 Designed-in exclusionBudgets, connectivity and permissions gate who gets to use AI at all.
LowAI's reply is deliberately modest: keep the model small enough to understand, local enough
to own, and open enough to audit. Frugal AI is not a downgrade — it is AI that answers to you.
Cloud vs Local, by design
Data that leaves your device
Cloud
100%
Local
0%
Works fully offline
Cloud
0%
Local
100%
You can audit the model & code
Cloud
~10%
Local
100%
Recurring cost to keep using it
Cloud
High
Local
~0
Indicative figures
≈1–2%
of the world's electricity used by data centres today — and climbing.
$10s M
to train a single frontier model.
0
tokens, dollars or data leaving your machine with LowAI.
Figures are indicative, drawn from widely reported ranges; LowAI's are by design.
Try it yourself
How AI development got here
For a decade, progress in AI has been read as a single dial: scale. More parameters, more
data, more compute — and, reliably, more capability. That scaling story produced genuine
breakthroughs, but it also concentrated the field around whoever could afford the largest
clusters. The result is an ecosystem optimised for centralisation: capability lives in the
cloud, value accrues to platform owners, and users become a source of data and revenue
rather than people to be served. Smaller, efficient, local models were always possible —
quantisation, distillation and better architectures keep widening what a laptop can do — but
they were under-funded precisely because they don't fit the rental business model. LowAI
treats that neglected branch as the main road.
The opportunity: a better world through LowAI
A frugal, local, open approach is not only a defence — it is an opening. Intelligence that
runs offline reaches the places connectivity doesn't. AI you can audit is AI you can trust
in medicine, law and education. Models you own can be tuned to a person, a village, a
language or a craft without surrendering the data that makes them useful. Energy spent once,
locally, beats energy spent endlessly in a distant data centre. And a community building in
the open compounds: every efficient model, every shared recipe, every reproducible tool
lowers the floor for the next person. The goal is not a smaller AI for its own sake, but a
wider one — owned by many, understood by its users, and gentle on the world it runs in.
🦉
Athenea
Our first prototype: an LLM agent that runs 100% on your machine — chat,
knowledge and RAG, fully offline. No tokens, no external APIs, no cloud.
How Athenea works
One local engine
Open GGUF models run on your own hardware. Nothing leaves your device — no tokens, no cloud.
Your own knowledge
Keyword RAG over your PDFs and EPUBs feeds the engine, with atomic storage in the .mv2 format.
Two tools, one core
From that core branch an Assistant you talk to, and Content you generate from everything it knows.
Frugal & free
It all runs inside one frugal envelope: no key, no subscription, no lock — ~8 GB RAM, open formats you can export anytime.
Assistant
Local conversational agent
Talk to open models running on your own hardware. Persistent memory, tools and knowledge —
all offline, without a single byte leaving your machine.
How it works with your knowledge
RAG over your documents: extracts text from PDF and EPUB.
Incremental ingestion with /learn, /fetch and /ingest.
OpenAI-compatible API server to connect your own apps.
Atomic persistence in the .mv2 format (conversations + knowledge).
Runs on ~8 GB of RAM — no bleeding-edge GPU required.
Modest hardware, long life: the tiny-house logic applied to AI.
Content
Content generation & transformation
Produce and transform content from your own local knowledge: drafts,
summaries and rewrites. Nothing sent to the cloud, no subscriptions.
Workflow
Batch processing of documents and notes.
Reusable templates for recurring tasks.
Generates from the same sources your Assistant already ingested.
Exports to standard formats (Markdown, plain text).
Reproducible, no vendor lock
Pipeline wired to the same .mv2 knowledge base.
Same input → same output: reproducible generation, locally.
Open models and formats: migrate and export whenever you want.
Zero dependence on OpenAI, Google or any third-party API.
Built on Athenea
Implementations
Projects that extend the Athenea core toward new capabilities.
Personalized fine-tuner
on the nety network
Prototype
A personalized assistant that fine-tunes your own Athenea model on your
data, spreading the compute across the nodes of the nety P2P network.
Your AI, trained by and for you.
How it fits
Starts from a local base model in Athenea's catalog.
Tuning is distributed across nety's compute nodes.
The result comes back as a GGUF model that runs offline on your machine.
Your raw data never goes to a central server.
Why it matters
Real personalization without proprietary GPU farms.
Uses the P2P resource market — low cost by design.
Proves the Athenea ⇄ nety bridge: local AI + decentralized network.
Related projects
Distributed computing
nety · Rust + libp2p
Active
The distributed compute layer of nety: a P2P network where nodes
offer and consume resources. The infrastructure that makes the fine-tuner possible.
Pieces of the network
Compute and storage nodes across the P2P mesh.
Resource Manager: tracks CPU, RAM, GPU and disk.
VM Runtime with WASM execution (Wasmtime / WASI).
Tiered storage: local, IPFS, S3 or hybrid.
Market & reputation
Market Protocol: peer-to-peer supply/demand of resources.
Representing Athenea's knowledge as a hypergraph: instead of edges
joining two nodes, hyperedges that connect many at once.
The idea
One hyperedge links question ↔ documents ↔ concepts together.
Models many-to-many relations a flat graph can't capture.
Richer than the linear JSONL index of today's RAG.
Where it goes
Retrieval by hyperedge neighborhood, not just keyword.
Connects sources, concepts and agents in one structure.
A basis to reason about why a document answers a question.
Ephemerall
ephemeral by design
Concept
Data and compute that exist only as long as they are needed, then vanish.
The simple-living instinct applied to memory: don't hoard what you can let go.
The idea
Sessions, context and artifacts with a built-in lifespan.
Nothing persists by accident — keeping is an explicit choice.
Less stored state means less to leak, secure and maintain.
Why it matters
Privacy as a default: what doesn't exist can't be exposed.
Low consumption: ephemeral state is cheap state.
Pairs with Athenea's local engine and nety's transient nodes.