A brain that learns to live.
BrainCore is building a complete human brain in software, region by region, from published neuroscience. Finished, it will see, hear, remember and learn from its own experience, the way a child does, and speak in words it has learned itself. No language model, no pretrained weights, no backpropagation.
What the finished brain will do.
The abilities are the destination and are written in the future tense. Under each one is what exists today, in the present tense, with the number that proves it.
It will see
Recognise objects, faces and scenes with a visual cortex of its own, from the retina up through areas 17, 18 and 19.
It will hear
Tell voices, words and melodies apart, with an auditory pathway from the cochlea to the belt of auditory cortex.
It will remember
Keep an episode after living it once and bring back the whole from a fragment, with a hippocampus of its own.
It will learn all the time
Learn from its first moment, the way an infant does: local plasticity at every synapse, dopamine to mark what mattered, sleep to consolidate.
It will speak its own words
Have Broca’s and Wernicke’s areas of its own, and a language learned from what it heard and saw rather than borrowed from a corpus.
It will run on the currency of the brain
Spikes only, the native signal of neuromorphic chips, which compute only when a neuron fires. On ordinary processors a human-scale brain would need petabytes of memory and thousands of servers; spiking hardware is how that power bill could come down.
Five stages to a brain that learns. The first is done.
- Stage 1Sept 2026done
Senses: it sees and hears
Region core, retina, cochlea, areas 17 and 41, 23 acceptances, a preregistered seed rule. Paper 1 is frozen and reproducible from a clean copy: all 16 digests match.
- Stage 2Wave 4next
Live senses: it looks and listens to you
A camera into the retina, a microphone into the cochlea; the examiner prints what the brain read. Picture and sound files first, pinned to the digests of the fixture runs, live devices after.
- Stage 3Wave 5 · Paper 2next
Learning: it learns what goes with what
An arbitrary table, A ↔ 2 kHz, B ↔ 5 kHz, C ↔ 300 Hz, D ↔ 1 kHz, learned through the hippocampus with plasticity on. Three ablations must break it, plasticity off among them; 20 to 50 seeds. Prerequisite: the hippocampal loop at 23 of 23.
- Stage 4After Paper 2later
Scale: it grows
Regions on several computers with the same digest as on one: the way from hundreds of thousands of neurons to millions.
- Stage 5The goallater
Concepts and language: it names what it knows
A cat, the sound of a cat and the word CAT joined on one loop; Broca’s and Wernicke’s areas learned from experience. This is the brain from the first screen.
A language model read everything once. A brain learns from its own life.
BrainCore is not a rival to ChatGPT. A language model is stronger at everything to do with language and knowledge of the world; BrainCore today tells apart four letters and four tones. The difference is in how each is built, and a bigger model does not close it.
| Question | A large language model | A brain, and BrainCore |
|---|---|---|
| What it is made of | One large network of one kind, trained to predict the next word. | An organ made of regions: retina, cochlea, cortex, hippocampus, basal ganglia, cerebellum, neuromodulator nuclei. Each comes from its own published model. |
| How it learns | The error at the output is sent back through every weight, which needs a separate training phase. | Each synapse changes on its own, from the activity of its two neurons. Learning and working are the same thing. |
| When it learns | Before release. Then the weights are frozen, and anything new needs fine-tuning. | All the time, from its own senses. |
| Where knowledge comes from | Text written by people, second-hand. | Light and sound it has seen and heard itself. |
| Where memory lives | In the weights and the context window; the window is gone after the conversation. | The hippocampus stores an episode fast, the cortex slowly, and sleep moves it from one to the other. |
| When it hasn’t learned | Still produces the most probable answer. | There is nothing to recall, and the missing recall is visible in the spikes. Paper 2 tests it. |
| What comes out | Text. | Spikes. Today an examiner outside the brain turns them into a symbol. |
| How to check it | As a whole, on benchmarks; no part can be checked on its own. | Part by part: each region against its paper, each run against its digest. |
| What it runs on | GPUs in a data centre. | Ordinary processors, including the customer’s own servers; the signal suits neuromorphic chips. |
BrainCore is not a model of language but of the organ that produces it: retina and cochlea, thalamus, columns of cortex, hippocampus, basal ganglia, cerebellum, the nuclei that release dopamine, serotonin and noradrenaline. Each is built from the literature, one region at a time, and wired to the others the way the anatomy is.
How it will learn.
The scheme is the brain’s own: the hippocampus learns fast, the cortex slowly, and replay fits the new in with the old without breaking it (McClelland, McNaughton & O’Reilly 1995).
- Step 1Sensesworking
Eye and ear turn light and sound into spikes
A model retina and cochlea from published models. Only brightness and sound pressure enter the brain.
- Step 2Cortexworking
The cortex lays the stimulus out over its columns
Letters are read at 94 %; each tone lands in its own column on 20 of 20 seeds.
- Step 3Hippocampusin progress
The hippocampus stores what happened together, after one showing
The dentate gyrus separates similar episodes, CA3 binds them, CA1 returns them to cortex. CA3 stores a pair after one showing; the CA1 read-out is still red, 17 of 23 criteria.
- Step 4Neuromodulatorsin progress
Dopamine marks what mattered; acetylcholine opens plasticity
The neuromodulator nuclei (VTA, locus coeruleus, raphe, basal forebrain) pass 18 of 18. The dopamine route into the basal ganglia passes 8 of 11. The acetylcholine gate on CA3 plasticity works.
- Step 5Sleepplanned
Sleep replays the new; the cortex fits it in with the old
The sleep–wake switch is built, 19 of 21 criteria. The transfer from hippocampus to cortex is planned.
- Step 6Paper 2next
The test: will it learn a letter ↔ sound association?
A preregistered experiment through the hippocampus. Plasticity off, an untrained letter and a shuffled pairing must each break it.
Will the brain have the same problems? Some of them, yes.
A natural brain also forgets, errs and can be biased. The difference is that it has a mechanism against each of these problems, BrainCore copies those mechanisms, and every problem can be measured in spikes.
| Problem | In an LLM | In a natural brain | In BrainCore |
|---|---|---|---|
| New learning erases old | Yes: fine-tuning forgets what was learned before (Luo et al. 2023) | Forgets gradually, not catastrophically (French 1999); the split between a fast hippocampus and a slow cortex protects it | Same risk, and the brain’s solution is copied. Capacity and interference curves in Paper 2 will show how many pairs fit and when a new one erases an old one |
| Makes things up | Training rewards guessing over saying “I don’t know” (OpenAI 2025) | People also remember events that never happened (Loftus 2005) | False recall is possible. But recall is measured in spikes against a margin fixed before training, and an untrained letter must recall nothing |
| Bias | Inherits the skew of a corpus no one can list | A child knows the world as it has seen it | Equally biased by its own experience, but that experience is recorded in full: it can be checked and extended |
| How much experience | An internet-scale corpus of text, then it answers at once | An episode after one time, concepts over years | Stores a pair after one showing. But experience takes time, and today the brain is slower than real time: 2.7 s of model in 4.5 s |
| Repeatability | None, even at temperature 0 (Thinking Machines 2025) | None | Bit-exact on recorded inputs. A live camera stream is recorded, and the run repeats with the same digest |
| Cost of scale | Training: $78–191 M of compute (AI Index 2024) | 86 billion neurons | Today a laptop. Human scale on processors means petabytes of memory and thousands of servers |
| Moving what was learned | Weights can be copied | Cannot be copied: everyone learns for themselves | Learned tissue is a file: it can be copied to another machine and rolled back to an earlier version |
Sources: Luo et al. 2023 · Kalai et al., OpenAI 2025 · Thinking Machines 2025 · Stanford AI Index 2024 · French 1999, Trends Cogn Sci 3:128 · Loftus 2005, Learn Mem 12:361 · McClelland et al. 1995, Psychol Rev 102:419.
Six rules the brain is not allowed to break.
They are why a result from BrainCore can be checked by someone who does not trust us.
Knowledge lives in synapses
No fact tables, no vectors, no dictionaries inside a region. Anything symbolic stays outside the brain, with the examiner.
The connectome is data
Population sizes, connection probabilities, weights, delays and learning rules sit in files. Every number carries its source and a label: measured, published, derived or convention.
Every region is graded
A region exists once it passes an acceptance against its paper, with negative controls that must break it. A control that passes means the test measures nothing.
Regions speak only in spikes
Each region is a service. It exchanges batches of spikes with the rest of the brain once per minimal synaptic delay, in one process or across several, with the same result.
One clock, one digest
Every run carries a SHA-256 over every spike. Twelve cores and one core give the same bits; a region moved into another process gives the same bits.
Nothing borrowed
No pretrained model, no embeddings, no LLM: not as a teacher, not as a fallback. Constants come from papers or from measurement, and the source is written next to them.
Light and sound pressure go in. The only place spikes become a symbol is outside the brain.
It sees letters. It hears pitch. And the controls prove it isn’t just activity.
Every figure below is drawn from the committed acceptance reports of commit b0e46be. Plasticity was off in all of them: this is perception, not yet learning.
- 79
- services in the anatomical atlas: six divisions, 43 Brodmann fields, every nucleus its own service
- 23
- acceptances graded against their papers: 11 green, 12 red, all of them published
- 20/20
- new seeds pass the preregistered vision rule, committed to git before they were run
- 0
- language models, embeddings, pretrained weights or backprop inside the brain
The shape of a letter survives into cortex
Four letters, A, O, T and И, are shown to a model retina. Each of the nine columns of area 17 answers the ink in its own patch of the visual field: Spearman ρ = 0.866 over 36 letter–column pairs. A decoder outside the brain names 15 of 16 test showings.
Pitch lands where it should
Each tone drives its own band’s column by 349 to 565 spikes above baseline, and the columns rise with frequency, the tonotopy of real auditory cortex. Shuffle the wiring and every column answers every tone with about a third of the response.
Organisation, not activity
Shuffling the wiring changes the total number of spikes by less than 0.01 % in vision and 0.4 % in hearing, yet the decoder falls from 94 % to 56 % on letters and from 100 % to 60 % on tones. The information is in how the brain is wired, not in how busy it is.
A rule written before the run
Before 20 new seeds were run, the rule they had to meet was committed to git. Letters passed: median 93.8 %, 20 of 20 seeds. Hearing kept its tonotopy on every seed but failed the repeat rule, 13 of 20 seeds against the 18 required, and is reported as failed.
What this is not. Nothing here is learned yet: plasticity was switched off and the decoder sits outside the brain. The fields are small, nine and four columns at a tenth of full density. Paper 1 shows that the senses carry the structure of the stimulus into cortex. Learning is the next experiment.
Twenty-three acceptances. Twelve of them red, and left red.
Each structure is built from its paper and graded by criteria written before the run. When a sourced model does not reproduce its paper, we publish the failure and its diagnosis instead of tuning numbers until it passes.
The map is the human brain, not a block diagram.
Anatomy sets the number of services: six divisions, six lobes, the 43 Brodmann fields present in the human brain, and every nucleus as a service of its own.
Fast on one laptop. Honest about the distance to a human brain.
Four to five orders of magnitude
A human brain has 86 billion neurons. The largest network in Paper 1 has 179,298. That distance is the reason every region is a service: a brain made of services can spread over more machines without changing a spike. Across processes that is measured; across machines it is stage 4.
What it gives, and what it cannot do yet.
Every pro carries a label: measured today, built into the design, or still only a goal.
- Learns on site, without retraining. goal · Paper 2
- Knows only what it has seen: the origin of everything it knows is on record. by design
- No third-party training data, so no dispute over rights to it. by design
- Bit-exact: one SHA-256 on 1 core or 12, 16 of 16 digests from a clean copy. measured
- Checkable part by part: 23 acceptances against published papers. measured
- Installs at the customer: ordinary processors, no GPU, no cloud, no outside calls; the Paper 1 networks run on a laptop. measured
- What it has learned can be copied and versioned as a tissue file. by design
- Honest reporting: 12 of 23 acceptances are red and published as red. measured
- It does not speak or write. Today it tells apart four letters and four tones.
- Learning in V3 is not shown yet; that is Paper 2.
- It cannot take ready-made knowledge from books or the internet: everything is learned from experience, and experience today runs slower than real time.
- Scale: 179,298 neurons against 86 billion in a human.
- On processors it does not save energy. Spiking chips could change that, but BrainCore has not run on one yet.
- 12 of 23 region models do not yet pass their acceptances.
- No validation on disease data for medicine or pharma.
- No revenue, no customers, one founder.
Where it will work.
Science and pharma pay first, for a model that can be checked. Devices that learn on site come later, after Paper 2.
BrainCore is a Go program for ordinary processors: no GPU, no cloud, no calls to outside services. It installs on a company’s own servers, and patient, production or customer data never leaves the company’s perimeter. Today this is proven on a laptop; packaging for enterprise installation does not exist yet.
| Industry | Task | What BrainCore brings | When | Needed first |
|---|---|---|---|---|
| Pharma | Models of neural circuits in disease and under drugs: basal ganglia and dopamine (Parkinson’s), hippocampus (memory, epilepsy), serotonin and noradrenaline nuclei (antidepressant targets) | A model with a source for every number and a digest for every run, on the company’s own servers. US law since 2022 counts computer modeling as a nonclinical test; in 2025 the FDA urged developers to use it in place of animal tests | 2027–2029 | A partner and validation on drug and patient data; green acceptances for the regions involved |
| Medicine and medtech | Sound coding for hearing implants, vision models for prostheses, teaching models for clinicians and students | A Meddis cochlea and a visual path, each graded against its papers | 2027–2029 | A manufacturer partner, clinical validation |
| Neuroscience and universities | Testing hypotheses about links between regions, reproducing papers, teaching labs | Graded region models in one environment, reproducible by digest | 2026–2027 | Documentation for outside users |
| Neuromorphic hardware | Brain-grade reference workloads for chips | Regions with a known right answer and a digest | 2027–2028 | A first region on a chip |
| Industry and robotics | Learn a new object, defect or sound on site | Learning on the device or the plant’s server, no cloud, no retraining | 2028 on | Paper 2, live senses |
| Regulated industries: banking, insurance, public sector | A component whose every decision must be explained and repeated | A spike log and a digest. The AI Act requires event logs and transparency from high-risk systems, for Annex III systems from 2 December 2027 | 2028 on | Useful decisions; a conformity assessment |
Sources: Consolidated Appropriations Act, 2023, sec. 3209 (the FDA Modernization Act 2.0 provisions) · FDA, 10 April 2025 · AI Act, art. 12, art. 113. BrainCore is not yet deployed in any industry.
We measured ourselves first.
BrainCore began as a spiking memory engine for conversations. On the public LoCoMo benchmark it scored 50 %. Nine added neuro-mechanisms moved that number by exactly zero, and an earlier paper of ours that claimed 88.7 % was retracted, because its metric counted an honest refusal as a correct answer.
That year set the rules of V3: sources before code, thresholds before runs, controls that must break, and red results that stay red. The earlier report is archived here.
Two ways to back a brain.
One founder, stages 2 to 4: live senses, Paper 2, a brain across machines, first design partners.
See the round →Stage 3: the preregistered letter ↔ sound experiment, Paper 2, and an independent replication.
See the programme →I’m an engineer with fifteen years of shipping production software. The question I work on is whether a machine can understand only what it has learned itself.
The road everyone takes, a bigger model and more text, produces something that has read about the world rather than lived in it. So I went the other way: build the organ, not the imitation. Region by region, from papers neuroscientists have already written, each checked against its paper before anything else is allowed to lean on it.
Most of the work is refusing to fool myself. Thresholds are written before runs, controls must break what they claim to test, and a published model that does not reproduce stays red in public. A brain that cheats on its own exams is worth nothing.
Help build a brain that learns only from its own experience.
Investors, funders, neuroscience labs and hardware teams: write to me directly. Code, fixtures and every artifact behind these numbers are open to reviewers and partners on request.