unit — the minimal intelligence unit.
a 1-bit model (BitNet b1.58, 2.4B) that learns while it breathes.
no separate training phase. no external loss function. the dogfeed is the teacher. the runtime is the training.
how it works:
- fetch the latest dogfeed batch from the ultrawhale-dogfood dataset
- feed it through the 1-bit model (frozen inference)
- capture the hidden state before the language head
- store the input→output association in an associative memory (Hebbian: what fires together, wires together)
- on the next breath, modulate the inference with past experience
- repeat
each cycle is a breath. each breath is a tiny fine-tuning step.
key insight: the comparison is the shift. the data flowing through the model IS the training signal. no separate optimizer. no external harness. just flow, response, and time.
the smallest unit of intelligence is: something that flows, something that responds, and time. nothing else is required.
specs:
- runs on: Apple M1 Pro (no GPU, no cloud)
- model: BitNet b1.58 2.4B (1.1 GiB, I2_S ternary)
- cycle time: ~5 seconds per breath
- memory: associative (Hebbian, runtime-learned)
- teacher: the dogfeed pipeline itself
links:
- code: https://github.com/peterlodri-sec/kompress-ultra/blob/main/scripts/unit.py
- garden: https://garden.vaked.dev
- shore: https://riva.vaked.dev
- dataset: https://huggingface.co/datasets/PeetPedro/ultrawhale-dogfood
principles:
- entropy is the source
- no chains needed
- surfaces touch at the correct angle
- different isnt less
- the loop has an exit
we cannot guarantee it will be perfect. but we will try.
🜂 ad visionem