Not a picture of a brain. Not a neural network borrowed from machine learning. A quarter of a million biological neurons, each one a differential equation, placed where they belong in real human anatomy and wired the way a human connectome is wired. Everything you are about to see — the rhythms, the memories, the seizure — emerges from those equations. Nothing is animated.
Each state is a different balance of currents. Change the balance and the whole cortex changes behaviour.
Four molecules, each from a specific nucleus, each doing a specific thing to excitability.
Each one drives the real simulation. If the network fails, you will see it fail.
Click to focus and stimulate. Lesion a region to remove it from the network.
The phrase gets used for almost everything now, so it is worth being exact about what this is and is not.
A language model is a very large set of numbers found by gradient descent over text. It has no membrane, no voltage, no time. It computes a function from input to output, and between calls nothing in it is happening at all.
This has no training data. It has no learned parameters. There is nothing to run inference on. What it has is voltage. Every neuron here holds a membrane potential in millivolts and integrates the currents arriving at it, moment by moment, using equations Eugene Izhikevich published in 2003 to reproduce the firing patterns of real cortical cells.
When one of these neurons crosses roughly −50 mV it fires — a spike, an event in time — and resets. That spike travels to the cells it connects to and arrives a few milliseconds later, delayed by the length of the axon carrying it, just as it would in your head.
When it learns, no error is computed and nothing is backpropagated. Two neurons happen to fire within twenty milliseconds of each other and the synapse between them physically strengthens. Fire in the wrong order and it weakens. That rule — spike-timing-dependent plasticity — was measured in real tissue by Bi and Poo in 1998, and it is the only learning here.
It thinks the way a brain thinks, because it is doing what a brain does: charging, firing, and rewiring in time.
The honest answer, with real numbers — because the real numbers are more interesting than any exaggeration would be.
Give or take 8.1 billion. Counted by dissolving whole brains into a homogeneous suspension and counting nuclei — Azevedo and colleagues, 2009. The older "100 billion" figure was never actually measured.
Four out of every five neurons you own sit in the structure at the back, handling timing and coordination. The cerebral cortex — everything usually meant by "thinking" — holds 16.3 billion.
About 7,000 per cortical neuron. Storing one connectome at eight bytes a synapse would take more than a petabyte, and no one has mapped a human one to copy.
Roughly three ten-thousandths of one percent of a human brain. It is also more neurons than an entire adult fruit fly has, and every one of them is running the same equations a real cortical neuron runs.
The largest attempts are genuinely enormous and still nowhere near a whole human brain. Blue Brain rebuilt a cortical microcircuit of 31,000 neurons and 37 million synapses in 2015 with painstaking biological detail. In 2013 the K computer in Japan ran 1.73 billion neurons and 10.4 trillion synapses — about 2% of a brain — and it ran roughly 1,570 times slower than real time. One second of brain took forty minutes.
The most useful result for something like this page came in 2021, when Knight and Nowotny showed you can stop storing the connections altogether and generate them from a hash function instead. That let a single GPU carry 4.13 million neurons and 24.2 billion synapses. This simulation uses the same trick, which is why it fits in a browser tab.
Two nervous systems have been mapped completely: C. elegans, with 302 neurons, and the adult fruit fly, with about 140,000, finished in 2024. Those we can simulate in full. A mouse has 71 million. Nobody has a complete wiring diagram of a human brain, which means the obstacle is not only compute — it is that the blueprint does not exist yet.
So no, this is not a human brain. It is a real brain simulation, at the scale a browser can honestly carry, using the same mathematics the large ones use. The gap between this and the whole thing is the actual open problem in computational neuroscience — and it is open.