You Are a Computer from the Dial-Up Era — The Brain's Biggest Unsolved Mystery
> Paper: The Unbearable Slowness of Being: Why do we live at 10 bits/s? > Authors: Jieyu Zheng, Markus Meister (Caltech) > Published: Neuron, 2025 > DOI: 10.1016/j.neuron.2024.11.008 > arXiv: 2408.10234
1. A Story First: Twenty Questions
Imagine this game. I say: "Think of something. I'll guess it using yes/no questions."
"Is it an animal?" "Is it a mammal?" "Is it domesticated?" ...
After twenty questions, I guess it. What does that mean?
Each yes/no question transmits 1 bit of information. Twenty questions = 20 bits, answered within seconds. So your thinking speed is roughly 10 bits/s.
That's the starting point of the entire paper. Not a fancy brain-imaging experiment — a centuries-old guessing game. Zheng and Meister used this simple question to crack open a giant fissure in neuroscience.
2. How Big Is That Fissure?
Three numbers:
| Metric | Value | Meaning | |------|------|------| | Sensory input | ~10⁹ bits/s | What your eyes, ears, and skin receive per second | | Behavioral output | ~10 bits/s | The speed at which you type, speak, think | | Gap | 10⁸ | A hundred million times |
The paper offers a vivid analogy: it's like drinking Hoover Dam through a straw. The dam's flow is a hundred million times the rate of sipping. Your brain — a supercomputer of 86 billion neurons and 10¹⁴ synapses — filters just 10 bits/s from the sensory torrent to drive behavior.
The authors call this ratio the "Sifting Number": Si = 10⁸. They suggest it may be "the largest unexplained number in neuroscience."
3. Is This Number Solid?
Skepticism first. Can one number define human cognition? Let's check the evidence.
Typists: 10 bits/s
A professional typist manages 120 words/min. But English entropy is only ~1 bit/character (due to heavy redundancy and structure). So:
> 120 words/min × 5 chars/word × 1 bit/char ≈ 10 bits/s
Ask typists to type random character sequences and speed collapses — they rely on language's predictive structure, not finger dexterity.
Speedcubing blindfolded: 11.8 bits/s
A perfect measurement of pure perception speed. Competitors observe the cube for 5.5 seconds, memorize every piece, then solve blindfolded. The cube has 4.3×10¹⁶ permutations ≈ 2⁶⁵. So the information rate = 65 bits / 5.5s ≈ 12 bits/s.
Note: perception time and motor time are roughly equal — the perception bottleneck and motor bottleneck are matched. It's not slow fingers; it's a slow brain.
Memory sport world champions: 5-18 bits/s
- 5-minute binary memorization: world record 1467 digits, rate 4.9 bits/s
- Speed card memorization: a full shuffled deck in 12.74 s, rate 17.7 bits/s
- Optic nerve fibers reliably signal specific visual features within 1 ms precision
- Deep in macaque cortex, a single neuron discriminates two visual stimuli as well as the whole monkey
- Retina: 1 million output channels
- Primary visual cortex: 10,000 hypercolumns, each ~100,000 neurons
- A "psychological refractory period" appears under dual-task competition
- Even pure thought pursues one strand at a time
- Wang's (2008) decision-circuit model reproduces much of macaque psychophysics with 2,000 integrate-and-fire neurons. A square millimeter of cortex could hold hundreds of such modules.
- AlexNet image recognition: 650,000 neurons (a few square mm of cortex) extract 10 bits from a 1.2-million-bit image. Why not run many in parallel?
- A fly, with fewer than 200,000 neurons, performs visual flight control, odor navigation, social behavior, and mating. The human prefrontal cortex could run 5,000 fly brains. Why not do several things at once?
- Zheng, J., & Meister, M. (2024). The unbearable slowness of being: Why do we live at 10 bits/s? *Neuron*, 113(2), 192-204. https://doi.org/10.1016/j.neuron.2024.11.008
- arXiv preprint: 2408.10234 [q-bio.NC]
Key finding: extending working-memory demands from 10 s to 300 s barely changes the rate. Memory is not the limiting factor.
More data (Appendix A)
| Activity | Information rate | |------|---------| | English listening | ~13 bits/s | | English reading | 28-45 bits/s | | Professional StarCraft players | ~10 bits/s | | Tetris S-level players | ~7 bits/s | | Choice reaction experiments | ~5 bits/s |
Across nearly a century of measurements, from labs to esports, everything converges to ~10 bits/s. That's not a coincidence.
4. How Strong Is the Hardware?
If you hear "the brain is slow," your first thought might be: neurons themselves must be feeble. Biological circuits are slower than silicon, after all.
Wrong.
Single neuron: 270 bits/s
A single cone cell in the human eye converts light to membrane voltage at ~270 bits/s. Six million cones = 1.6 Gbit/s per eye.
Single cortical neuron: ~10 bits/s
Even at average firing rates of a few Hz, one cortical neuron transmits ~10 bits per second — equal to the entire human information throughput.
Optic nerve: 100 Mbit/s
The retina's output layer (ganglion cells): 1 million axons at ~50 Hz average under strong stimulation. Capacity = 1M × 50 spikes/s × 2 bits/spike ≈ 100 Mbit/s.
Note: this is already after ~10× compression inside the retina. From photoreceptors to optic nerve, information has been compressed by an order of magnitude.
So the question isn't "neurons are too slow" but "what are 86 billion fast processors collaborating on?"
5. Dispelling Three Illusions
Faced with this paradox, many instinctively push back. The paper dismantles each objection with data.
Illusion 1: "I have a photographic memory"
If photographic memory existed, memory sport champions would be swept aside by it. In reality, even world champions run at 10 bits/s.
The famous Stromeyer experiment (1970) claimed a subject could achieve random-dot stereo fusion from monocular images shown on different days — implying 100 bits/s — but it was never replicated.
Illusion 2: "My visual field is full of rich detail"
Your peripheral vision feels sharp, right? Fix your eyes on this word, then check how many letters you can clearly identify on either side. The answer is usually: 2-3 letters.
This is called "subjective inflation." Your visual field feels uniformly clear because you can always move your eyes to "check" any location. In fact, resolution collapses outside the fovea. Retinal circuits are deliberately designed this way.
> The famous "invisible gorilla" experiment: when counting basketball passes, about 50% of viewers never see a person in a gorilla suit walk through the scene.
Illusion 3: "Unconscious processing handles huge amounts of information"
Visual development does learn statistics from the environment. But stripe-rearing experiments show: kittens raised among only vertical stripes do reorganize their visual cortex — but the required information is only 2 bits (compressing orientation range from 180° to 40°).
Natural image statistics are translation-invariant — frequency distributions in the lower-left match the upper-right. So the brain can cover the entire visual field convolutionally (like a CNN), learning a few "templates" and replicating them, not learning each point independently.
6. Five Striking Implications
Implication 1: Your brain's lifetime input fits on a USB stick
Upper bound on synaptic storage: ~50 TB (10¹⁴ synapses × 5 bits each). But how does information enter the brain? Two channels:
Innate (genome): ~0.8 GB Learned (senses): even at memory-champion speed, 24/7 for 100 years, only ~4 GB
Total < 5 GB — a keychain USB stick holds everything you can acquire in a lifetime.
This means the brain's 50 TB capacity far exceeds actual needs (four orders of magnitude of redundancy), leaving huge room for inefficient storage algorithms, massive redundancy, and biological learning-rule constraints.
Implication 2: Teenagers finding "reality boring" has a basis
The ecological niche our ancestors chose was a slow-changing world. 10 bits/s is only needed in the worst case; most of the time the environment is more leisurely.
That's why teenagers say "reality is broken" and turn to fast-paced video games; earlier generations sought thrills in skiing and mountain biking. Everyday tasks are unbearably slow for them.
> "Pushing themselves to the cognitive throughput limit is a rewarding experience all by itself."
Implication 3: Flies have the same problem
Measurements show a tethered fly's flight-control throughput is 0.62 bits/s, while a single visual neuron (H1) runs at 100× that rate.
This paradox may not be uniquely human. Different species may survive in their niches at "just enough" information rates.
Implication 4: Musk's Neuralink bandwidth problem
Elon Musk said: "We have a bandwidth problem. You can't communicate through your fingers — it's too slow."
This paper's prediction: Musk's brain-to-computer communication rate would also be ~10 bits/s. The bundle of Neuralink electrodes could effectively be replaced by a phone — telephone data rates were designed to match human language, which in turn matches the speed of perception and cognition.
The authors call this the "Musk illusion" — a generalization of subjective inflation: because we feel we could execute any of 2¹⁰ possible actions in the next second, we think we could execute them simultaneously. In reality, they're serial.
Implication 5: Brain-computer interfaces took a wrong turn
Decades ago, blind patients received retinal electrode arrays attempting to feed raw camera images directly into the optic nerve (Gbit/s scale). Result: all patients remained legally blind. Companies collapsed; patients were left with abandoned hardware in their eyes.
Why? The human brain only ever extracts ~10 bits/s from a visual scene. The right approach: have a computer translate the scene into speech in real time — "who is where doing what" — which became practical by 2018.
Similarly, the most advanced motor-cortex BCI results: handwriting decoding at 90 chars/min (1.5 bits/s), speech decoding at 62 words/min (half a typist). But most paralyzed patients can hear and speak — voice commands are already the most natural brain-computer interface.
7. Four Explanations, None Fully Right
Explanation 1: Neurons are inherently inefficient?
No. Single-neuron transmission is startlingly precise:
The supposed "population noise" may just be variables experimenters can't control (like an itchy monkey).
Explanation 2: Redundant design?
No. Many fly neuron types exist as single left/right copies. Primate retina has one neuron per cell type per visual field location, with almost no overlap. A small cortical stroke can wipe out a corresponding region of the visual field — no built-in redundancy for the most common brain injury.
Explanation 3: Serial processing (attention bottleneck)?
Partially right, but incomplete.
The peripheral nervous system is massively parallel:
But central processing is strictly serial:
The question is: why? Chess masters evaluate candidate moves one by one. Even with parallel evaluation, only one move can be played.
The paper offers an evolutionary angle: the brain's primordial function is odor navigation, as in jellyfish. An organism occupies one location, senses one local environment, executes one move. There's no need to process multiple paths in parallel — only one path corresponds to current reality. Human thought can be seen as "navigation in abstract spaces," inheriting the one-line-at-a-time constraint.
But the authors concede: "This still leaves the question of how the constraint is embodied in neural architecture."
Explanation 4: Complexity bottleneck?
Psychology describes seriality with metaphors — "single-channel operation," "attention bottleneck," "limited processing resources" — but these are labels, not explanations, unless you specify what neural resource is limited.
Moreover, the neural hardware demands of relevant cognitive functions aren't high:
No viable proposal currently explains what neural mechanism enforces single-threaded operation.
8. Outer Brain vs Inner Brain
The paper's core framework:
| | Outer Brain | Inner Brain | |---|---|---| | Connections | Sensory input + motor output | Behavior control | | Dimensionality | High (millions of receptors/muscle fibers) | Low (a few bits) | | Rate | Gbit/s | 10 bits/s | | Parallelism | Massively parallel | Serial | | Examples | Retina, primary visual cortex | Prefrontal cortex |
Big Problem 1: How does the inner brain listen to the outer brain?
The slow inner brain listens to the outer brain's sensory torrent like drinking Hoover Dam through a straw (the paper's own words).
The matching must occur in stages. Retinal output, for example, is already compressed over 10×, keeping only features useful downstream. But given a Sifting Number of 10⁸, many more orders of magnitude of visual compression await understanding.
The superior colliculus may be one further bottleneck: its upper layers have fine receptive fields (retina-like), its lower layers premotor signals reflecting behavioral decisions — distilling high-dimensional sensory input into a map of behavioral targets.
Big Problem 2: Do the two sides run on different neural principles?
Hoover Dam's generators and beer bottles aren't made from the same materials. Yet the brain seems to use the same materials everywhere: neurons, synapses, glia.
Primary visual cortex (V1) and prefrontal cortex (PFC) contain comparable neuron counts. V1 is described as high-dimensional parallel processing (10,000 hypercolumns each handling a patch of visual field), while PFC research reduces it to 2-3 "latent variables."
How can regions with nearly identical cell types and architectures run so differently?
The paper proposes a radical speculation: maybe we've been studying the inner brain the wrong way.
Imagine a counterfactual history: if our understanding of V1 rested only on low-dimensional stimuli ("train monkeys to stare at one periodically rotating spiral propeller, record thousands of neurons, find two principal components explain most of the variance, publish") — our understanding of visual cortex today would be entirely wrong.
Real V1 was revealed through high-dimensional fine stimuli and single-neuron analysis, exposing its receptive fields, functional anatomy, and circuit mechanisms.
PFC has as many neurons as V1. Perhaps it also contains thousands of small modules, each dedicated to a specific microtask, like V1's hypercolumns? It's speculative — but the only cure for ignorance is new kinds of experiments.
9. Questions Feynman Would Ask
> "Are you sure 10 bits/s isn't an artifact of measurement methods?"
The paper's answer: a century of measurements, different fields, different methods, all converging. If it were an artifact, every measurer would have made the same mistake — unlikely.
> "Then why doesn't parallelism work?"
The paper gives no answer. That's the heart of it — posing the question, not solving it. The authors call this possibly "the largest unexplained number in neuroscience."
> "Is 10 bits/s enough?"
Yes. In the niches our ancestors chose, the world was slow enough. Most of the time the environment was leisurely. 10 bits/s is needed only in the worst case.
But that also means: if the environment speeds up, we're out. The paper notes coldly: once the last human driver retires, road infrastructure can be upgraded for machine cognition (kbits/s scale). Humans should then avoid those niches — just as snails should avoid highways.
10. One-Sentence Summary
> Your sensory system takes in a billion bits per second, but everything you will ever think, say, or do outputs at 10 bits per second. Eighty-six billion neurons, 10¹⁴ synapses — the most complex known object in the universe — ultimately produce the throughput of a 1980s modem. > > Why? We don't know.