AI has become astonishingly capable in just a few years. But almost all of it runs the same way: you type into a box, your words travel to a data center owned by one of a handful of companies, a model you'll never see produces an answer, and you pay — in money, in data, or both.
The Hidden Cost of Centralized AI
The convenience is real. So is the trade-off. Every prompt you send to a cloud model is a piece of your thinking handed to someone else's server. The model never adapts to you — it's the same frozen system serving millions. And you're renting access to something you can never own or run on your own terms.
The deeper problem is concentration:
- Your data leaves your device — prompts, documents, and context all flow to servers you don't control
- The model isn't yours — you can't inspect it, keep it, or run it offline
- A few companies set the terms — pricing, access, and capabilities are all decided upstream
As AI becomes infrastructure for daily life, that concentration of data and control matters more, not less.
Centralized AI Still Has Its Place
Let's be clear: frontier cloud models are extraordinary, and the teams building them do remarkable engineering. For the hardest reasoning tasks and the very largest models, the cloud is often the right tool.
But not every request needs a trillion-parameter model in a distant data center. A great deal of everyday AI — drafting, summarizing, answering questions about your own files — can run perfectly well on hardware you already own. And when it does, it can be private in a way the cloud fundamentally cannot.
What If the AI Ran on Your Device?
That's the idea behind Centram. Instead of sending your data to a model, Centram runs a model on your device — a local LLM that lives on your machine, answers from your hardware, and learns from how you use it. Your data is the training signal, and it never has to leave.
The result is an AI that adapts to you over time instead of staying frozen, and does it without uploading a single private byte.
Genomes: Sharing What You Learn, Not Your Data
Here's the part that makes it a network instead of an island. When your local model learns something, the change is captured as a compact weight-delta "genome" — a small summary of what was learned, not the data it was learned from.
Those genomes are what peers exchange. You benefit from what other devices have figured out, and they benefit from yours, while everyone's raw data stays home. It's collective intelligence without collective surveillance.
You Become Part of the Infrastructure
When you run a Centram node, you're not just a user — you're a builder. A network of devices can do things no single one can:
- Distributed inference (verified experiment) — a model can be cut into layer shards and run across cooperating processes. Every run so far has been separate processes on a single host; a real two-machine test has not been done.
- Resilience (research) — there is no central server to go down or lock you out
- A compute marketplace (research) — the design calls for nodes to serve inference for others at a price they set. Nothing like it runs today.
- Shared progress (open question) — the hope is that every genome makes the whole network a little smarter. Peers training on the same task do transfer usefully; peers training on different data have not, in the experiments published so far.
Think of it as the opposite of the cloud: instead of everyone renting from one tower, the intelligence lives at the edges and pools together — securely, with no raw data ever changing hands.
The Economics — a design, not a shipped feature
Status: research. This section describes a design that is written down but not running. Nothing here is something you can use today, and none of it is required to run a node.
The design calls for:
- Compute providers who set their own price per inference request
- Gas fees charged per operation type, funding the peers that confirm transactions
- A CenTrust reputation score that discounts those fees for reliable nodes, up to 50%
- Competition between providers as the mechanism that keeps prices fair
Whether peer-to-peer inference on hardware people already own ends up cheaper than renting cloud capacity is an open question — it has not been measured.
Privacy Comes Built In
Your data and the model that learns from it are designed to stay on your device. Anything sensitive at rest — your keys, wallet, and local data — is encrypted with AES-256-GCM using keys derived from your passphrase. The genomes you share are signed weight deltas rather than raw records; formal privacy depends on the differential-privacy settings and validation gates used for each shared genome.
The Bigger Picture
AI is becoming the most important infrastructure of the decade. The real question is who controls it. Centralized providers will keep serving the frontier — and they'll keep doing it well. But there's room for something else alongside them: a network where the intelligence runs on your own hardware and learns from you privately.
Centram is building that network. Whether a network of peers can make each other's models measurably better is the project's central open question, and the results so far are published either way.
Your device. Your AI. Part of something bigger.
Centram is in active development. Read the documentation or see what is and is not working on the home page.