SUB-01 is a computer model of brain activity, built from one person's public MRI scans. It drives an invented character who lives in a small room in Tokyo. It is a model, not a mind.
1What this is
A model of large-scale brain activity built from one person's open MRI data (OpenNeuro ds000031): the wiring between brain regions from diffusion MRI, the shape of the cortex, and resting-state recordings. The model runs continuously, and its activity drives an invented character: when he sleeps, what he reads, how his pulse moves.
It is a model, not a mind. It does not understand, feel or think. It is not alive and it is not conscious. When the site says that "he" reads, sleeps or remembers, that is shorthand for what the simulation and the code around it do.
2What it is not
- It is not the scanned person. The data are public, and the site never depicts or speaks as that person or uses them to promote anything.
- It does not show the scanned person's thoughts, memories or dreams. The model cannot read any of those from MRI data.
- The character, his room, his day and his life are fiction. His food orders are simulated. His diary lines and dream images are generated by AI from the model's state.
3Numbers
Every number below comes from a file in the project's repository. Negative results stay on this page.
A slow-layer model that uses only the wiring and one fitted number produces activity whose correlation pattern matches the recordings at r ≈ 0.21 (0.206), on sessions never used for fitting. It beats randomised brains that keep each region's total connection strength (p = 0.01). Against randomised brains that keep the spatial layout it is only borderline (p = 0.059). Those randomised brains already reach r ≈ 0.16–0.17, so most of the fit comes from the brain's geometry, not the specific wiring. The two halves of the recordings agree with each other at r = 0.96, so the model is far from that ceiling.
docs/real_wiring_report.mdAn earlier version reported r = 0.40. That number came from replaying the person's own recorded activity as input, not from the wiring. A model with no coupling between regions at all scores 0.45 the same way. So it is not shown as a result.
docs/real_wiring_report.md, analysis/null_test.py, docs/HANDOFF.mdA sudden kick to one region, repeated 60 times per state, and a measure of how complex and widespread the response is (Casali et al. 2013). All three values are below the 0.31 clinical cut-off, because the pulse stays local: the struck region responds, and the rest of the brain barely does. The ordering (awake ≈ REM > deep sleep) is the human one.
What it does not show. PCI is validated on human brains. A model scoring high would not be conscious, and a model scoring low says nothing about experience either: it says this model's fast layer is too damped and weakly coupled to carry a perturbation across the brain.
docs/pci/README.mdA network of one million simulated spiking neurons was trained on Pong. It does not learn: in the exam it scored 17.5 %, against 25 % for the same network with every learned change removed. Pong is off by default and labelled "not learning yet" where shown.
docs/spiking_report.md, docs/HANDOFF.mdThe same one-million-neuron network stores the day's top chat messages and readings as patterns. Shown half of each pattern a simulated day later, after a compressed night with REM replay, it completed all 6 stored items better than never-stored lures. This is a single run (RTX A5000, 2026-10-01), not a statistic. The chance of a false "recalled" is about 1 in 7 per item in this test. Recall here means a pattern is completed in simulated neurons; it is not remembering in the human sense.
docs/spiking/gpu_reading/results.json, docs/spiking_report.md, docs/HANDOFF.mdThe daily test runs once a day on a rented GPU when it is switched on. A summary of its results over time
will go here, from /api/memory.
4What it does each day
- Sleeps. A published sleep-wake circuit (Phillips & Robinson 2007) decides when the
model is awake, in NREM or in REM sleep, on the Tokyo clock. Undisturbed, it falls asleep around 23:56 and
wakes around 07:54. A diary entry is written each morning from images rendered during REM.
source:
docs/twin/sleep.md - Reads and browses. It follows a chain of Wikipedia articles and browses the open web. The only thing it ever types is a filtered search query into a site's own search box; it never logs in, signs up, posts or buys anything, and a guard blocks unsafe sites. It may look at crypto and stock pages, read-only: it never connects a wallet or clicks buy, sell or swap, and the screen says "nothing here is financial advice". What it reads becomes a stimulus to the model.
- An AI reads and remembers for him; his brain model chooses. A language model (Claude Haiku) writes a short note on each page he reads and keeps the notes as a searchable memory that fades with time. Each Tokyo morning it proposes two or three day plans, and the brain model votes among them the way it votes among links; hunger and sleepiness still come first. While he sits at the computer it may write a short chat reply, tagged "his reply (AI-written)", from those notes. The replies never pretend to be human, never mention the coin and never give financial advice. Notes, plans and replies share one AI budget with the browsing options: at most $2.00 per Tokyo day. Over it he rests: no notes, yesterday's plan, no replies.
- Gets chat messages as stimuli. Messages are filtered, then turned into input to the brain networks their words relate to. Some are replayed during REM sleep.
- Reacts to the coin's trades. Public trades move the model's simulated neuromodulators
(dopamine and noradrenaline levels), and noradrenaline moves its simulated pulse. These rules are a design
metaphor, not a measurement of anyone's brain chemistry.
source:
docs/twin/neurochem.md,docs/twin/body.md
5Costs
What it costs to run, per day. To be filled from the cost ledger once it exists.
| item | per day |
|---|---|
| Eye images (the frames on /world and the dream images) | placeholder |
| AI options while browsing, page notes, day plans and chat replies (one ledger, capped at $2.00) | placeholder |
| GPU memory test (rented GPU, at most 1 hour a day) | placeholder |
| Server | placeholder |
One-off GPU validation runs: the first cost about $0.01 (docs/spiking/results.md).
The reading-memory run behind the 6/6 figure cost about $0.033, and that whole task about $0.036 (PR #9;
results in docs/spiking/gpu_reading/).
6Credits and licences
- MRI data: the MyConnectome project: Poldrack RA et al. (2015), Long-term neural and physiological phenotyping of a single human, Nature Communications 6:8885 (doi:10.1038/ncomms9885). Shared on OpenNeuro as ds000031 under the PDDL (public domain). With thanks to the project for making these data public.
- Receptor densities: Zilles K & Palomero-Gallagher N (2017), Multiple transmitter
receptors in regions and layers of the human cerebral cortex, Frontiers in Neuroanatomy 11:78,
CC BY 4.0. These are group data from other people, not the scanned person's receptors.
(
twin/data/RECEPTORS_SOURCES.md) - Maps and routes: © OpenStreetMap contributors
(ODbL). The map is drawn from one static extract of his neighbourhood
(
web/tokyo_area.json,room_build/osm_area.py), not from map tiles. - Articles: Wikipedia, CC BY-SA. Article titles and extracts shown on the site belong to their Wikipedia authors.
- 3D room (
web/room/room_clean.glb): modelled for this project and released under CC0, with Poly Haven textures and props (CC0 1.0). The poster text is set in Noto Sans JP (SIL Open Font License 1.1); only rendered pixels are used. Each asset and its authors are listed inweb/room/CREDITS.md. - 3D character (
web/room/boy*.glb,web/room/hand_phone.glb): made for this project with Higgsfield (image concepts turned into 3D with Meshy, then rigged and animated). placeholderNo licence is stated for these files; the terms they are used under are to be confirmed. - Neuron shapes (the close-ups and the whole-brain neuron view on /brain,
web/neurons_morph.json): real reconstructed neurons from NeuroMorpho.Org (RRID:SCR_002145), licensed CC BY 4.0, simplified to at most 300 segments a cell (40 in the whole-brain view, drawn ~6× enlarged) and rotated. They are illustrative: the simulated network uses point neurons, and none of these cells is from the scanned person. Tecuatl C, Ljungquist B, Ascoli GA (2024), FASEB Bioadv 6(7):207-221, and the original papers:- pyramidal and basket cells, human neocortex (NMO_107146, NMO_107145): Molnár G et al. (2016), eLife 5:e18167;
- medium spiny neuron, rat nucleus accumbens (NMO_10032): Spiga S et al. (2005), Eur J Neurosci 22:2332;
- thalamic relay cell, rat VPM (NMO_115363): Iavarone E et al. (2019), PLoS Comput Biol 15:e1006753;
- CA1 pyramidal cell, rat hippocampus (NMO_00222): Ishizuka N, Cowan WM, Amaral DG (1995), J Comp Neurol 362:17;
- principal cell, mouse basolateral amygdala (NMO_139102): Hsu TT, Huang TN, Hsueh YP (2020), Front Mol Neurosci 13:47;
- layer-5 pyramidal cell, human anterior temporal cortex, dendrites only (NMO_159993; the deep pyramids of the zoom's
cortical column,
web/column_morph.json): Moradi Chameh H et al. (2021), Nat Commun 12:2497.
docs/spiking/morphology_credits.md,analysis/build_morphologies.py) - The brain as a specimen (the "fresh brain", "dissected" and "bundles" views on /brain,
web/anatomy/): the cortex, the cerebellum, the brainstem and the cut surface are his, from the same MRI as the fibres (the cut shows his T1 scan coloured as grey and white matter). The arteries, the optic, trochlear and ophthalmic nerves, the pituitary and a spinal cord stub are an atlas, another adult's body fitted to his brain: BodyParts3D, © The Database Center for Life Science licensed under CC Attribution 4.0 International (CC BY 4.0; dbarchive.biosciencedbc.jp; Mitsuhashi N et al. 2009, Nucleic Acids Res 37:D782). The dural sinuses and large veins are the VENAT venous atlas (Huck J et al. 2019, Brain Struct Funct 224:2467; figshare 10.6084/m9.figshare.7205960, CC BY 4.0), a population average fitted to his brain. Cranial nerves I, VI–XII and the trunks of III and V are not drawn: no openly licensed model of them exists, and nothing is invented. To separate his cerebellum and brainstem the build used the ICBM 152 2009c template (Copyright (C) 1993–2004 Louis Collins, McConnell Brain Imaging Centre, MNI, McGill University; "permission to use, copy, modify, and distribute ... for any purpose and without fee"); nothing of the template is shown. Tissue colours are approximate, read off CC0 photographs. (web/anatomy/CREDITS.md,docs/research/real_brain_sources.md) - Cortical layers of the zoom's column (/brain, zoom in with neurons on): the layer thicknesses L1–L6
are measured from the public layer annotation of H01, a nanoscale reconstruction of about 1 mm³ of
human temporal cortex: Shapson-Coe A et al. (2024), A petavoxel fragment of human
cerebral cortex reconstructed at nanoscale resolution, Science 384:eadk4858 (Google and Harvard), released under
CC BY 4.0; its neuron count (~16 000 per mm³) sets how many
neurons the column stands for. Not from the scanned person.
(
analysis/h01_layers.py,docs/spiking/h01_layers.json) - Real electron-microscopy neurons (the cortex close-up on /brain,
web/neuron_render/h01/): cells, synapses, axon labels and an EM image from H01, a reconstruction of about one cubic millimetre of human temporal cortex by Google and Harvard: Shapson-Coe A, Januszewski M, Berger DR et al. (2024), A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution, Science 384:eadk4858 (doi). Released under CC BY 4.0; the meshes are decimated and a few cells are moved sideways into one field of view (the manifest says which). The glia come from H01's automated segmentation and are not proofread. The activity drawn on these cells is simulated; none of them is from the scanned person. (docs/spiking/morphology_credits.md,web/neuron_render/build/fetch_h01.py) - 3D engine: three.js, MIT licence.
- Sleep model: Phillips & Robinson (2007), J Biol Rhythms 22:167; constants cross-checked in the MIT-licensed R package sleepR.
- Code: placeholderThe repository does not state a licence yet.
7The coin
The site shows a coin's public trades only as stimuli for the model. Nothing on this site is financial advice. There are no promises, no price talk and no calls to buy. The project never holds or moves anyone's funds, and it never asks for a wallet.