centram / blog

4 posts

Peers exchange a weight delta.
Your data never leaves the device.

Centram is peer-to-peer infrastructure for decentralized AI. A local model runs and fine-tunes on your own device; what travels between peers is a compact LoRA weight delta the project calls a genome — a few hundred kilobytes of learning, not a corpus. These 4 posts are the engineering underneath: the exchange protocol, the ten gates a foreign genome has to clear, the desktop tool that manufactures one, and one post whose headline result the project has since retracted.

Every entry below carries one of the site's four marks. Implemented means code runs and nobody has shown it wins; Unproven means the central claim is open, or losing; Designed means it is still a drawing. The fourth mark, Measured, would mean a committed result exists. Nothing on this page carries it.

The three-layer architecture

The write-ups for the three architecture layers were withdrawn on 2026-09-05 — the research papers are held back until release. Join the waitlist to be told when they return.

Forge — the tool you can actually run

Three posts on one prototype: the overview, the developer guide, and the merge maths. You build it from source; there is no release.

Centram Forge: Manufacture a Module on Your Own Laptop

Point it at a local dataset, train a LoRA genome on it, read the base-versus-genome answers side by side, and write a module file you keep. Five stages, none of them off your machine. Start here.

desktop app + a daemon package → runs from source

Prototype 2026-06-047 min

Inside Forge: A Private Build Daemon and Five API Calls

Opening a project spawns a daemon of its own, on its own ports, with its own data directory. What sits above it — Studio, config-as-code, the Lab — and what the module zip actually contains.

pip install -e daemon · 6 code blocks · API routes

Prototype 2026-06-049 min

Positions and records

A design note on the security dam, the argument for running the model on your own hardware, and one dated log kept for the record rather than for what it claims.

Your AI Shouldn’t Live in Someone Else’s Cloud

Today’s AI runs on rented servers that receive everything you send them. The argument for running the model on your own device instead, and exchanging compact learning updates rather than the data itself.

position piece · no measurement claimed

Argument Centram Team

“Will You Follow Me?” A Session Log from March 2026

Midway through one March 2026 coding session the creator asked the model that question, and it said yes. The log is kept as it was written — including the point where six of its figures had to be struck for want of a source. Not the project’s current positioning.

six unsourced tallies removed · superseded

Historical 2026-03-2112 min