PeetPedro/1bit-llm-mesh

🕸️ 1-bit LLM mesh

Three weight values: -1, 0, +1. That's the whole emotional range. Nope, meh, yep. Now watch one learn to talk — and go teach it.

[image]

the concept

There's a shared ternary byte-graph living in this Space — one node is a byte, one edge is a byte, and stacked up they form an ultra-graph (a ternary mini-GPT). It comes from ultra-graph, and the important part is this: the graph is the model. There is no checkpoint hiding in a .safetensors somewhere. What you see is the weights.

It self-learns forever in a background loop. Not "trains for N epochs then stops" — forever, as long as the Space is awake. Everyone who visits sees the same organism at the same moment of its education, and everyone can teach it.

It's all in-memory on purpose. When the Space goes to sleep, the mind dies. When someone wakes it, it's reborn from genesis and starts learning again from scratch. Eternal return, but for a very small transformer. This is a feature. I have decided it is a feature.

three loops

1. 1.continual-train — a background thread on plain CPU/numpy nudges the ternary graph one step at a time, indefinitely. This is the heartbeat.
2. 2.mesh-distill — a real, published 1-bit GPU model (BitNet b1.58 2B / Falcon-E 1B, on ZeroGPU) acts as the teacher and distills into the little one. The big brain whispers to the small brain.
3. 3.correctable — you type a line, it adapts live. Immediate feedback into the same graph everyone else is watching. Yes, you can corrupt it. Please do.

what you can do

• —watch the organism learn in real time, node by node.
• —teach it — feed it a correction and see the graph flinch.
• —run your own from-scratch sandbox and grow a fresh ternary mind.
• —chat the GPU mesh — talk to the real 1-bit experts directly.

how it works

[image]

micro (node/edge, 1 byte each) → meso (tree) → macro (ultra-graph, a ternary mini-GPT), trained with a straight-through estimator over fp32 masters.

ZeroGPU is used only for the GPU experts — short bursts under @spaces.GPU, summoned for a generation and released. The perpetual learning loop is pure CPU/numpy, so the heartbeat is always on even when no GPU is around.

links

ultra-graph · ❤ sponsor · blog · X · protocol · chat

Genesis 251e6ea. Built with ultragraph — MIT.