Luke1001/one-token-organism

Vision — the one-token organism

A 297M-parameter language creature that lives in one continuous stream and learns while you talk to it — on a single MacBook. No context window tricks, no retrieval, no serve-side scripting: every word on screen is sampled from the model, and every decision about what to learn, when to sleep, and what to feel proud of comes from the model's own measured signals.

you: how does a submarine work?
it : I do not know that yet, but you could teach me.
you: a submarine dives by filling its tanks with water and rises
     by pushing the water back out.
you: so how does a submarine work?             (next morning)
it : A submarine dives by filling its tanks with water and rises
     by pushing the water back out.

The body — iga/lm_scan.py

Organ	What it is
Trunk	13-layer selective-scan recurrence, d=1024 — O(1) memory per token
Council	six timescale bands (clocks 1 → 32k tokens): syntax up to the shape of the day
PFC	routing and silent thought tokens
Hippocampus	episodic store: surprise-gated writes, decode-free logit-space reads, carries across days with nightly decay
Plan / dreamer	imagined futures scored on foresight; REM splices two memories at a scene cut
BG · DA	reward as experience: graded press tokens felt in-stream through real value/dopamine circuitry
Goal organ	wake-surviving pursuit slots read in identity space (lesion-verified)

The life — scripts/organism.py

Day — it measures its own surprise every turn and keeps what spikes (budgeted); reward presses of any magnitude (−6…+6) are felt as tokens and expressed as plasticity; when its answer to something it learned satisfies its own conscience, it presses its own button — felt only, never self-teaching. Night — its own progress ledger picks the replay (mastery graduates, stuck fades, fresh always settles, nothing drills more than two nights running); dreams pair the most emotionally charged memories; the conscience retrains on the human's real presses; mastered facts enter an expanding retention schedule (1/3/7/14/30 nights). Always — fatigue, boredom, and loneliness accumulate on their own clock: it falls asleep when tired, chews on its homework when bored, and may speak first when it has been alone too long. Mood feeds back into how it thinks. The life autosaves every night and survives process death.

Every threshold, budget, schedule, and reflex is a plain number in the code (the disclosed genome), shown on screen as it acts. The serve never authors a word and never parses the human's text for meaning.

Run it

git clone https://github.com/LukeHamond1001/Vision.git && cd Vision
pip install -e .          # torch, tokenizers, numpy
# weights (the living body + tokenizer, ~1.2G):
#   https://huggingface.co/LukeHamond1001/one-token-organism
# put organism_life.pt and ship_tok.json in data/, then:
python3 scripts/organism.py data/organism_life.pt data/ship_tok.json \
    --dev mps --temp 0.05 --save data/organism_life.pt
# open http://localhost:8016 — talk to it, teach it, press the
# slider, flip on internals, and leave continuous time on a while

First reply after launch takes a few minutes (the model compiles); after that it answers in seconds. Works on Apple Silicon (--dev mps), CUDA (--dev cuda), or CPU (slow). What has been measured about it: RESULTS.md.

Tools

scripts/mini_school.py — balanced interleaved consolidation over everything it knows (the periodic deep sleep) · scripts/stutter_repair.py — targeted unlikelihood on a degenerate token loop, with a golds-abort gate · scripts/knowledge_school.py — bulk knowledge raising · scripts/goal_gym.py — frozen-body training of the goal organ · scripts/critic_train.py — conscience seeding · scripts/scan_chat.py — terminal REPL · pod_*.sh / launch_pod.sh — GPU gestation infrastructure for the next scale · iga/lm_train.py + iga/lm_data_life.py — the gestation method itself (day/night-structured lives with rewards in-stream, the diet the body was born on and the recipe the next body scales).

The next body

GESTATION.md — the v17 pretraining method: the organism's food is lives, not documents. scripts/author_lives.py writes the childhoods (every rule measured on this organism by its teachers); iga/lm_data_life turns them into lane shards; iga/lm_train gestates 32+ childhoods at once through one body, against a matched transformer control on identical food.

Honest limits

At 297M, question→answer routing shares narrow capacity: heavy drilling or repeated corrections can collapse it (the school pass restores balance), abstract inference does not emerge, and social statement replies are thin. The conscience is young — it learns the human's taste a few presses per night. These are measured size limits, not serve tricks; the development history lives in the git log.