learn

Brain vs. computer

From Matter to Thought

How does physical stuff become a thought, in a brain and in a computer?

This page climbs the ladder one rung at a time: the smallest switches, the signals, memory, and finally what a "thought" might be. At each rung it compares the two, and says how good the comparison really is.

Brain

Living cells, water, salts. About 86 billion neurons.[R1]

Computer / AI

Silicon and metal. An H100 AI chip holds 80 billion transistors.[R4][R11]

Small grey tags like open the source. Boxes marked What this leaves out say where a demo simplifies.

Part 1

Building blocks: transistor vs. neuron

Computer

The transistor

A tiny on/off switch. Wire a few together and you get a rule (a gate); wire billions and you get a chip.

on/off 1 transistor a switch AND a few a gate (a rule) 80 billion a chip (H100)
80 billiontransistors in an H100 [R4][R11]
1.755 GHzclock: billions of lockstep steps/sec [R14]
Read the details
  • Basic unit: the transistor. A tiny switch controlled by electricity: on or off. A modern AI chip (Nvidia H100) contains 80 billion transistors.[R4][R11]
  • Electricity: electrons move through metal wires. High voltage = 1, low voltage = 0 (a "bit").
  • Logic gates: a few transistors wired together make a rule:
    • AND: output is on only if both inputs are on
    • OR: output is on if either input is on
    • NOT: output is the opposite of the input
    Everything a computer does (math, video, AI) is built from huge numbers of these simple rules.
  • Clock: the whole chip runs in lockstep to a central clock, taking billions of steps per second. NVIDIA lists the PCIe version of the H100 at up to 1.755 GHz (1,755 million steps per second).[R14]
  • AI on top: an artificial neural network is software that imitates simplified neurons. Each "neuron" multiplies its inputs by learned numbers called weights, adds them up, and passes the result on. So AI is brain-inspired math running on logic gates.
Brain

The neuron

A living cell that collects thousands of inputs and decides whether to fire. Far more than a switch.

dendritesinputs in cell body axonoutput cable → terminals
86 billionneurons [R1]
10¹⁴–10¹⁵synapses (connections) [R2]
5–8 layersof AI network to mimic one neuron [R3]
Read the details
  • Basic unit: the neuron, a living cell. The adult human brain has about 86 billion neurons (86.1 ± 8.1 billion), plus a similar number of non-neuronal support cells (about 84.6 billion).[R1]
  • Parts of a neuron:
    • Dendrites: input branches that receive signals
    • Cell body (soma): where inputs are combined
    • Axon: the output cable
    • Axon terminals: the endpoints that pass the signal to the next neuron
  • Connections: roughly 10¹⁴–10¹⁵ synapses (100 trillion to 1 quadrillion).[R2]
  • A neuron is not a logic gate. It combines thousands of inputs in complex ways. A 2021 study found that mimicking one realistic cortical neuron took an artificial deep network 5–8 layers deep.[R3]
Go deeper: how transistors work, how gates are wired, and NAND
M2

Logic gate playground

See that "computing" bottoms out in simple on/off rules.

This demo needs JavaScript.

M3

Neuron vs. artificial neuron, same inputs

See what the two share (inputs × strength, a threshold) and where they differ.

This demo needs JavaScript.

What this leaves out Real neurons combine inputs far more richly, closer to a 5–8 layer network,[R3] and the timing of inputs matters. The millivolt nudge per input here is a toy scale; −70 mV rest and −55 mV threshold are textbook values that vary by neuron type.

Part 2

Signals: electrical vs. chemical

Inside a neuron: electrical

Charged atoms push the voltage up or down. Cross the threshold and the neuron fires a full spike; otherwise nothing.

rest ≈ −70 mV threshold ≈ −55 mV spike: all-or-nothing small nudges: no spike
Textbook values. Information is in when and how often spikes happen.
1–100 m/sspike speed along the axon
few hundred/smaximum firing rate
Read the details
  • The "electricity" isn't electrons in a wire. It's charged atoms (ions), mainly sodium and potassium, flowing through tiny protein gates in the cell wall called ion channels.
  • A resting neuron sits at about −70 millivolts (textbook value).
  • Inputs push the voltage up or down. If it crosses a threshold (about −55 mV, textbook value), the neuron fires a spike (action potential).
  • A spike is all-or-nothing: it either fires fully or not at all. Information is carried in when and how often spikes happen.
  • Spikes travel down the axon at about 1–100 meters per second, and a neuron fires at most a few hundred times per second (textbook values).

Between neurons: mostly chemical

  1. Spike arrivesat the axon terminal
  2. Chemicals releasedneurotransmitters
  3. Cross the gapand bind to receptors
  4. Voltage nudgedglutamate ↑ toward firing, GABA ↓ away

The chemical step is slower but adjustable. That adjustable strength is where learning and memory live.

Read the details
  • The synapse is the tiny gap between one neuron's axon terminal and the next neuron's dendrite.
  • Step by step:
    1. A spike arrives at the axon terminal.
    2. The terminal releases chemical messengers called neurotransmitters.
    3. They cross the gap and bind to receptors on the next neuron.
    4. This opens ion channels and nudges the next neuron's voltage:
      • Excitatory (e.g., glutamate) pushes it toward firing.
      • Inhibitory (e.g., GABA) pushes it away from firing.
  • The chemical step is slower but adjustable. A synapse's strength can change, and that adjustability is where learning and memory live.
  • Electrical synapses (gap junctions) also exist. They're direct, fast, and less flexible, and they're the minority.
  • Broadcast chemicals (neuromodulators like dopamine and serotonin) wash over whole regions and change how those regions behave, like a global settings change. Hormones act even more slowly, through the blood.
M4

Synapse close-up

Electrical inside, chemical between, and learning = changing synapse strength.

This demo needs JavaScript.

What this leaves out Sizes, timings and response heights are illustrative, not measured. The dopamine wash here simply turns the response up, as a stand-in for "changing how a region behaves". "Practice" makes the synapse stronger in fixed steps; real strengthening and weakening follow richer rules.

Speed, scale and power

> 1 million×faster signals in board wiring than the fastest nerve [R12]
86 billionneurons, no clock, possibly <1% strongly active at once [R1][R13]
16,896cores on an H100, all in lockstep [R11]
20 W vs 700 Wa brain vs. one H100 [R4][R11]
Read the details
  • Electrical signals in circuit-board wiring travel at about half the speed of light, about 15 cm (6 inches) per nanosecond.[R12] That's more than a million times faster than the fastest nerve signal (~100 m/s).
    • What travels that fast is the electrical wave, not the electrons themselves.[R12]
  • Chip switches flip billions of times per second; a neuron fires at most a few hundred times per second.
  • The brain makes up for its slowness with massive parallelism: billions of neurons, each working independently, with no central clock.
    • But not all at once: firing is expensive, so only a small fraction of neurons is strongly active at any moment, possibly fewer than 1% in the cortex (an estimate).[R13]
  • An H100 has 16,896 processing cores (CUDA cores), all marching to one clock.[R11]
  • Power: the brain runs on about 20 watts (textbook value). One H100 is rated at 700 watts.[R4][R11]
M5

Speed, scale and power

A slow but massively parallel, efficient brain vs. a fast, serial-leaning, power-hungry chip.

This demo needs JavaScript.

What this leaves out The <1% figure is an energy-based estimate,[R13] not a measurement. The dot grids are not to scale: each dot stands for a huge number of neurons or cores.

Part 3

Memory: where each one stores what it knows

Computer / AI
  • Computer: memory is stored separately from the processor (in RAM or on disk) and fetched by address. Recall is exact: the same address returns the same bits.
  • AI models: what a model "knows" is stored in its weights. Concepts are stored as overlapping patterns of numbers, so related ideas activate each other.
Brain
  • Memory is stored in the strengths of synapses, in the same place where processing happens.
  • Learning = strengthening or weakening synapses.
  • Recall = reactivating a pattern of neurons. A partial cue (a smell, a shade of blue) can reactivate part of a stored pattern, and the network fills in the rest.
  • Memories are rebuilt each time they're recalled, not played back like a file. That's why they drift and why false memories happen.
Key contrast In a brain, the hardware is the software, and it rewires itself while running. In a computer, the hardware is fixed and the software changes. Most current AI models are also frozen after training; they don't keep learning from use.
M6

Two kinds of memory

Exact address lookup vs. reconstructive pattern completion.

This demo needs JavaScript.

What this leaves out The brain-style side is a toy model of the idea (a tiny Hopfield-style network), not a model of how real brain circuits store memories.

Part 4

What is a "thought"?

There is no agreed scientific definition. This page uses a working definition:

Working definition (not a settled one) A thought is a temporary pattern of activity across many units that represents something (an object, idea, or plan) and can trigger the next pattern, an action, or a memory.
Brain

Millions of neurons firing in a particular pattern.

AI

A pattern of numbers moving through the network, called an activation vector.

What's unknown Both process information, but only brains are known to have experiences. Whether an AI's pattern is felt in any sense is an open question. This page doesn't take a side.

Part 5 · The centerpiece

The model: from matter to thought

Ten rungs, from atoms at the bottom to thought at the top. Tap a rung to compare the two sides and see how good the analogy is.

M1

The ladder

Hold the whole model in mind at once, and know where the analogy holds and where it breaks.

This demo needs JavaScript. The same content is in the table below.

Analogy strength: Good Moderate Loose Weak

The ladder as a table
#StageBrainComputer / AIHow good is the analogy?
1MatterLiving cells, water, saltsSilicon, metal— (just different materials)
2What carries chargeIons (sodium, potassium) crossing cell wallsElectrons in wiresModerate: both are moving charge, by very different mechanisms
3Smallest switchIon channel (protein gate)TransistorGood: both are gates that open or close based on voltage
4Basic signalSpike: all-or-nothing, meaning carried in timing and rateBit: 0 or 1Loose: spikes carry information in timing; bits in fixed levels
5Processing unitNeuron, which combines thousands of inputsLogic gate; in AI, a software "neuron" built from many gatesWeak: one real neuron ≈ a 5–8 layer artificial network [R3]
6ConnectionSynapse: chemical, adjustable strengthFixed wire on a chip; adjustable weight in AIGood for AI weights, poor for chip wires
7MemorySynapse strengths, in the same place as processingSeparate RAM/disk; weights in AILoose
8CircuitA group of neurons doing one job (e.g., detecting an edge)Circuits like adders; in AI, layers that detect featuresLoose
9SystemBrain regions working togetherChip + software → modelLoose
10ThoughtA pattern of firing across neuronsA pattern of numbers (activation vector)Loose: same idea of "a pattern that represents something"; whether it's experienced is unknown
Common mistake

"A neuron is like a terminal" is wrong. In the brain, a terminal is just the end of an axon, the neuron's output point.

The closest pairings are:

  • ion channel ↔ transistor
  • synapse ↔ AI weight
  • neuron ↔ a small processor, or a mini neural network of its own

Key differences at a glance

BrainComputer / AI
Signal speed~1–100 m/s~Half light speed in board wiring (~15 cm/ns) [R12]
Switching rateUp to a few hundred spikes/sec per neuronBillions of steps/sec (up to 1.755 GHz on an H100 PCIe) [R14]
Parallelism~86 billion independent neurons [R1], but only a small fraction (possibly <1%) strongly active at once [R13]16,896 cores on an H100, in lockstep [R11]
TimingNo central clockCentral clock, lockstep
Power~20 W~700 W per H100 [R4][R11]
Memory locationSame place as processing (synapses)Separate from processing
LearningRewires constantlyMost current AI is fixed after training

Part 6

Mapping the brain

A connectome is a complete wiring diagram: every neuron, every branch, every synapse.

How it's made: the H01 project (Harvard + Google)

  1. Tissueremoved in epilepsy surgery, preserved, stained
  2. Slice~5,300 slices, each 30 nm thick
  3. Image4 nm electron microscope, ~225 million images
  4. TraceAI follows every cell in 3D, finds synapses
  5. Proofreadpeople fix the errors

Steps from Google Research's H01 write-up.[R5]

57,000cells in 1 mm³ [R6]
150 millionsynapses [R6]
1.4 PBof images [R5]
1 millionthof a human brain's volume [R5]
Read the details

How it's made (H01, Harvard + Google):[R5]

  1. A small piece of human brain tissue, removed during epilepsy surgery, is preserved and stained.
  2. It's cut into about 5,300 slices, each 30 nanometers thick.
  3. Each slice is imaged at 4-nanometer resolution with an electron microscope, about 225 million images in total.
  4. AI traces every cell through the stack of slices in 3D, and finds the synapses.
  5. People proofread the errors.

Result: about 57,000 cells and 150 million synapses in one cubic millimeter,[R6] stored as 1.4 petabytes of images. That's about one-millionth of a human brain's volume.[R5]

M7

Slice-through

Feel why mapping is hard (scale, data, proofreading) and why a map alone isn't a simulation.

This demo needs JavaScript.

What this leaves out The counters grow evenly with the slider; real imaging, tracing and proofreading don't progress this neatly. The cell shapes are invented for illustration, not H01 data.
Records so far (as of October 2026)

The complete central nervous system of a male fruit fly, brain plus nerve cord, with about 166,700 neurons. It was published in Cell on 3 September 2026 by Janelia, Cambridge, the MRC Laboratory of Molecular Biology, and Google.[R10]

A whole mouse connectome is considered a 10–15 year goal, and the most expensive step is still human proofreading.[R9]

Why a map isn't enough The worm C. elegans has 302 neurons, fully mapped in 1986,[R7] and still can't be fully simulated. A map shows what connects to what, not how each connection behaves, and a static map can't learn.[R8]

Part 7

Self-check

Two quick ways to lock it in. Your progress is saved on this device only.

M8

Match and quiz

Retain the model: the right pairings, and why.

This demo needs JavaScript.

What this page simplifies

  1. Neuron as a "summing unit" is simplified. Real neurons compute much more (about a 5–8 layer network).[R3]
  2. Synapse ↔ AI weight is a functional analogy, not a physical one.
  3. Ion channel ↔ transistor: both are voltage-controlled gates, but by different mechanisms (proteins and ions vs. semiconductors and electrons).
  4. "Thought" definition is a working definition, not a settled scientific one.
  5. Textbook values (−70 mV rest, about −55 mV threshold, 1–100 m/s, a few hundred spikes/sec, about 20 W) are typical figures. Real values vary by neuron type.
  6. Chip signal speed (about half light speed) is for circuit-board wiring. It's the electrical wave that moves that fast, not the electrons.[R12]
  7. Sparse firing (<1%) is an energy-budget estimate for the cortex, not a measurement.[R13]
  8. Memory demo (M6) is a toy illustration of pattern completion, not a model of how real brain circuits store memories.

Sources

  1. R1Azevedo et al. (2009). Equal numbers of neuronal and nonneuronal cells make the human brain an isometrically scaled-up primate brain. J. Comp. Neurol. 513:532–541. doi.org/10.1002/cne.21974
  2. R2Carlsmith, J. (2020). How Much Computational Power Does It Take to Match the Human Brain? Open Philanthropy. openphilanthropy.org
  3. R3Beniaguev, Segev & London (2021). Single cortical neurons as deep artificial neural networks. Neuron 109(17). doi.org/10.1016/j.neuron.2021.07.002
  4. R4Tom's Hardware. Nvidia Hopper H100 SXM5 pictured (80B transistors, 700 W). tomshardware.com
  5. R5Google Research (2021). A Browsable Petascale Reconstruction of the Human Cortex (H01). research.google
  6. R6ScienceAlert. Amazingly detailed images reveal a single cubic millimeter of human brain in 3D. sciencealert.com
  7. R7White et al. (1986). The Structure of the Nervous System of the Nematode C. elegans (summary via OpenWorm). openworm.org
  8. R8Asterisk Magazine. We Can, Must, and Will Simulate Nematode Brains. asteriskmag.com
  9. R9Wellcome. Scaling up connectomics. wellcome.org
  10. R10Berg, S. et al. (2026). Sexual dimorphism in the complete Drosophila male central nervous system connectome. Cell 189(18):5504–5526. Published 3 Sept 2026. doi.org/10.1016/j.cell.2026.08.015 (press release)
  11. R11NVIDIA Technical Blog. NVIDIA Hopper Architecture In-Depth (H100 SXM5: 80 billion transistors, 16,896 FP32 CUDA cores, 700 W TDP). developer.nvidia.com
  12. R12Siemens EDA blog (2010). Fundamentals of SI (Part 2): Transmission lines (signals travel ~6 inches/ns on a circuit board, about half light speed). blogs.sw.siemens.com
  13. R13Lennie, P. (2003). The cost of cortical computation. Current Biology 13(6):493–497. doi.org/10.1016/s0960-9822(03)00135-0
  14. R14NVIDIA. H100 PCIe GPU Product Brief, PB-11133-001_v02 (30 Nov 2022): base 1,125 MHz, boost 1,755 MHz. Third-party-hosted copy. openzeka.com (PDF)

Textbook values (resting potential, threshold, conduction speed, firing rate, brain power ≈ 20 W) are standard neuroscience figures with no single linked source here.