Project Overview

[Dingo]

I have detached all my computers from this projects due to their accessing my personal folders on my computer.  I have also dleted the folder created at C:\Users\%USERNAME%\Axiom\contribute
 
Project Summary

Axiom is a distributed AI training platform that uses Hebbian learning — 'neurons that fire together, wire together' — to train a massive neural network across thousands of volunteer computers worldwide.
Unlike traditional backpropagation, Hebbian learning is local and biologically inspired. Each sample strengthens connections between neurons that activate together — no complex gradient chains required.

The 85/15 Split: 85% of learning happens locally on your machine, while 15% comes from gossip synchronization with peers training other experts.

HOW IT WORKS

Axiom trains a 17.8 billion parameter language model using volunteer computing. The model is a Mixture-of-Experts architecture with 420 specialized transformer experts (42.6M parameters each).

Training Pipeline

1. Worker downloads expert weights + seed data
2. Trains locally via backpropagation + SGD
3. Compresses weight delta via 4-bit quantization (90% gradient fidelity)
4. Sends delta to coordinator via HTTP POST
5. Coordinator averages 10 deltas, applies to master weights

Architecture

420 transformer experts (Mixture-of-Experts)
42.6M params each (d=768, 12 heads, 6 layers)
Self-attention + FFN per layer
Gradient clipping (max norm 1.0)
GPU (CUDA) and CPU support

Training Data: 10MB of diverse public domain text (Project Gutenberg) refreshes every 24 hours. Add your own files to %USERPROFILE%\Axiom\contribute\

Applications
The following applications are supported:

Please see their Applications page as they change with ne applications HERE

Connecting to the Project

The projects Home Page is at: https://axiom.heliex.net/

Don't forget to join the BOINC@AUSTRALIA Team following your registration.

FreeDC Statistics

View our Team Stats https://stats.free-dc.org/team/adai/27

All Axiom User Stats: https://stats.free-dc.org/users/adai