A field guide to the bandwidth of human communication

The Funnel and the Flood

Every communication technology in history widened the channel by orders of magnitude. The human receiver never got faster. So the bottleneck did not disappear. It moved into the mind, and AI is the first technology aimed at that bottleneck instead of the channel.

The Hook 00

Seventeen years to say a gigabyte

Take one gigabyte. Not a large file. Say it out loud, every bit, at the speed a human mouth actually moves. The clock you would need is not measured in hours or days.

Human speech carries information at about 39 bits per secondEstablished, a rate that holds steady across 17 languages even though some race through syllables and some crawl. At that rate, one gigabyte (8 x 10^9 bits (1 GB, decimal)) takes the better part of a decade to pronounce.

One gigabyte, spoken aloud, at the human rate
TimeRateEffort
about 6.5 years 39 bits per second 24 hours a day
about 9.8 years 39 bits per second 16 waking hours a day
about 19.5 years 39 bits per second 8 hours a day
about 25 years 10 bits per second (contested) 24 hours a day

The answer moves by a factor of four depending on which rate and how many hours a day you assume. That dependence is the entire point. Derived

Say a gigabyte. The number is real. So is its softness.

Move the assumptions and the answer lurches. Sixteen waking hours a day gives about 9.8 years. The full clock gives about 6.5 years. Eight hours a day gives about 19.5 years. Accept the contested claim that conscious thought runs nearer about 10 bits per secondContested and the same gigabyte stretches to about 25 years. The famous figure is real. It is also soft, and the softness is the subject, not a flaw in it.

Hold that next to the wire in the wall, which moves a gigabyte in well under a second. We built a firehose for a thimble. The channel was never the slow part of communication. You are.

This page reads at your speed

It runs on a reader feed called nix.baud. Scroll slowly and you get the whole verbose story. Speed up and the text below you summarizes to match your pace. Slam the scroll and it drops to headlines. Slow down and it all comes back. The gist is honestly lossy: it dropped most of the bits and will never tell you which, which is the deal you make every time you skim anything. You can also set it by hand.

nix.baud reader
Auto. Scroll slowly for the whole story, faster for the gist. The page matches your speed.

One question falls out of the number. If the receiver is this slow, and has always been this slow, how did any of this get built? How did a species that cannot take in a gigabyte in under six years run a civilization on the exchange of ideas? The answer is a five thousand year climb, and it starts with a wink.

The Climb 01

The ladder

Start at the bottom of the ladder, where a raised eyebrow says more than a paragraph, and climb. Every rung widens the channel. By the top, the numbers stop meaning anything a body can feel.

human receiver, flat Speech Modem Ethernet Fiber

The channel climbed about twelve orders of magnitude. The receiver did not move. Established

The ladder. Everything got faster except the thing it was for.

Co-present speech and body language run near about 39 bits per second, full duplex, with a feedback loop so tight you rewrite a sentence mid-air when an eyebrow rises. The first commercial computer modem, the Bell 101 in 1958, managed 110 bits per secondEstablished, barely three times the mouth. Then the climb steepens.

Dial-up crawled up rung by rung, the screech of each handshake a little faster, until it topped out near 56,000 bits per second. Ethernet arrived and jumped to 10 to 100 GbpsEstablished. A single strand of optical fiber now carries tens of terabits per second bits per secondEstablished, with laboratory records reaching for a petabit. The printing press did something different and just as large: it left per-link bandwidth alone and exploded fan-out, one mind to many, a scribe-month compressed into a press-day.

Stack it all up and the channel grew by about twelve orders of magnitudeEstablished, from the wink to the fiber. This is the triumph the whole story is built to enjoy first. Everything got faster, wider, cheaper, more glorious. Hold onto that feeling. The next chapter splits the word that has been carrying it.

The Climb 02

The second race

The word faster has been doing two jobs. How much arrives per second is one race. How long until the first bit lands, and a reply lets you correct course, is a different one. They are independent, and conflating them hides half the story.

A station wagon full of hard drives on the highway has absurd bandwidth and catastrophic latency (paraphrasing Andrew Tanenbaum). A telegraph operator had pathetic bandwidth and, for the era, miraculous latency. For all of history before 1844 the top speed of an idea equaled the top speed of a horse or a ship, so communication and transportation were the same thing.

The first Morse line, Baltimore to Washington in 1844Established, cut them apart. The telegraph divorced communication from transportation. Information became faster than matter for the first time in history. Latency across an ocean fell from weeks to minutes while bandwidth stayed a few words a minute. Within a generation the transcontinental telegraph put the Pony Express, peak transportation-based communication at about 10 days coast to coast coast to coast, out of business in days.

Before the divorce, the gap between sending and arriving had a cost you could count in bodies. The Battle of New Orleans was fought on January 8, 1815, two weeks after the Treaty of Ghent was signed on December 24, 1814. The peace had been agreed. The news had not crossed the Atlantic.

Latency had a body count.

Geostationary round trip 480 to 600 msEstablished
Low Earth orbit latency near 25 to 50 msEstablished
The same path, three media. Lower is faster.
Vacuumcthe wall you cannot cross
Microwave in airabout 99.97 percent of cair is faster than glass
Light in fiberabout two thirds of csilica index near near 1.47
Geostationary480 to 600 mslatency you can hear
Low Earth orbitnear 25 to 50 msfly lower, wait less
You can buy more bandwidth forever. You cannot buy your way under the speed of light.

Bandwidth kept climbing without limit. Latency hit a wall and stopped. Light in glass travels at only about two thirds of its vacuum speed, because silica slows it. So high-frequency traders laid a straighter fiber to shave milliseconds, then abandoned it for microwave through air, where the index of refraction is nearly one and the beam runs faster than any signal in glass.

Air is faster than glass.

Starlink is the same move in orbit: fly the satellites low and the round trip shrinks where the high ones could not. Latency became a product, bought by the microsecond and priced against physics itself. You can buy more bandwidth forever. You cannot buy your way under the speed of light.

The Climb 03

The turn

Run the two numbers together and the triumph curdles. The channel climbed by twelve orders of magnitude. The receiver, across the same span, did not move.

Speech still lands near about 39 bits per second. Conscious throughput sits somewhere around about 10 bits per secondContested. Working memory holds about about 4 chunksEstablished and has held that for as long as there have been humans to test. A flat line, five thousand years long, drawn straight under an exponential.

So the obvious story, that communication kept getting faster, is wrong in a specific and important way. The pipe got faster. The thing the pipe was built to fill did not. The bottleneck was never abolished. It was relocated.

We spent five thousand years upgrading the road and never touched the truck.

The bottleneck did not disappear. It moved into the back of the skull. It did not disappear when the telegraph beat the horse, or when fiber beat the telegraph. Each upgrade moved it further from the wire and deeper into the receiver, until the only slow component left was the one no engineer could respec. We built a firehose for a thimble.

Which means the rest of this story cannot be told as one line on one chart. To see what happens next you have to put two minds side by side, because the mind on the other end of the AI is built on the opposite plan from yours. The page splits here.

The Mirror 04

Memory

The first place the two minds diverge is memory, and they diverge completely. One keeps almost nothing and survives on it. The other keeps everything and pays for it.

Human

Gist, not transcript

Working memory holds about about 4 chunksEstablished, a number that has not changed in any recorded population. The classic figure was 7 plus or minus 2 items, and the modern consensus is lower. Long-term memory is reconstructive: you keep the gist and confabulate the rest on recall. Forgetting is not a bug here. It is the feature that makes a finite head survivable.

Machine

Perfect recall, hard edge

The context window holds 1 million to 2 million tokensEstablished at the frontier, and every token in it is equally present. Token one and token one million are recalled with the same fidelity, right up to a hard cliff at the edge of the window. Inside, nothing is lost. Outside, nothing exists.

Cheap because lossy

Throwing almost everything away is what keeps the human cheap to run. The cocktail-party filter drops every conversation but one, and you never miss the bits it discarded because you never knew they were there.

Expensive because complete

Keeping everything is what makes the machine expensive. Its context grew roughly 200,000x in about seven yearsEstablished, a curve that mirrors the channel climb but compressed into years, and every extra token is another thing that must be held, in full, at cost.

AI context window

One flat tier. Token one and token 1 million to 2 million cost the same to hold, then a hard cliff at the edge. No hierarchy, which is exactly why it is expensive.

Same staircase, two materials. The AI skipped the stairs.

Humans had to invent writing to gain persistence. AI is born with perfect persistence and has to invent forgetting.

The Mirror 05

The price of context

Memory is what you keep. Context is what it costs to hold the whole exchange in mind while it is happening. The field almost never measures that standing cost, which is exactly where the two minds are most unalike.

Human

Paid in regulation

The expensive resource is not recall, it is regulation: allocating attention, inhibiting the wrong reply, tracking the topic, and modeling your context on top of holding its own. The universal turn gap is about about 0 to 200 msEstablished, shorter than the over 600 ms it takes to plan a word, which proves the listener is predicting the end of your sentence and pre-loading a reply while you still speak. All of that runs on a brain drawing about about 20 wattsEstablished.

Machine

Paid per token

Attention scales as O(n squared)Established, so doubling the context roughly quadruples the work. The model re-reads the entire conversation every turn rather than living through it. Holding a million tokens can cost on the order of 15 GB of KV cache per userEstimated of cache, and prefill can run over two minutesEstimated before the first new token.

Machine: cost rises as n squared

Human: working memory overflows

O(n squared)EstablishedTransformer self attention near 40 percentdegradationEstimatedLiu et al., Lost in the Middle, 2023, and later work about 4chunksEstablishedCowan, 2001

Double the context and the machine's attention work roughly quadruples (O(n squared)). The human's working memory holds about about 4 chunks and overflows almost at once, so the human pays in regulation while the machine pays in compute, money, joules, and time, with lost-in-the-middle risk near near 40 percent degradation at scale.

Two cost curves, opposite currencies, both climbing.

And paying the full price does not even guarantee uptake. Retrieval sags for whatever sits in the middle of a long context, a degradation near near 40 percent degradationEstimated. The whole context-engineering industry, the retrieval and summarizing and sliding windows, is invented forgetting bolted onto a thing that natively forgets nothing.

Both minds get the same discount from the same source. Talk to someone who already shares your context and there is less to regulate and less to send. Shared priors are compression, whether the priors live in a marriage or in a set of weights.

  1. A strangerYou explain caching, embeddings, and retrieval from scratch. Many words.
  2. A colleagueCache the embeddings so retrieval stops re-reading the doc.
  3. An expertCache the embeddings, skip the re-read.
  4. Your spouseA look. They already know.

Same idea, fewer words, because the receiver already holds the context. The conversational rate drops from about 39 bits per second toward about 13 bits per second once shared context is counted.

Shared context is compression. Jargon, a marriage, and embeddings are the same trick.

The Mirror 06

Noise and the broken message

Every channel fights noise, and the bill for reliability always comes due in one of two currencies: bandwidth spent up front, or latency spent on failure. Shannon wrote the equation, C = B log2(1 + S/N), and put a wink and a fiber optic on the same axis.

Human

Forward error correction

Tone and face and gesture are parity bits. You can lose a third of the words in a loud bar and still get the meaning, because the channel is multiplexed and redundant on purpose. Language itself is roughly roughly half redundantEstablished redundant, which is why you cn stll rd ths. When the redundancy runs out, you fall back to asking again.

Machine

Checksums and retransmits

The machine detects corruption with a checksum and fixes it by asking for the packet again, to lossless delivery. Read-back is a checksum. The huh is a retransmit request. The strategies have names on both sides because they are the same strategies.

UDP, fire and forget

UDP, like live speech

the contract says the [lost]begins monday and the payment clears [lost]delivery

A word is lost here and there, but the conversation never freezes. A late word is worthless.

TCP, like a contract

the contract says the work begins monday and the payment clears on delivery (waited, retransmitted, then continued)

Every word arrives in order, but the stream stalls to recover the missing ones. Nothing stale is acceptable.

Pick a protocol by whether stale data is useless or sacred.

Real-time speech runs like UDP: you would rather the listener miss a word than freeze the conversation, because a late word is worthless. A signed contract runs like TCP: acknowledged, ordered, retransmitted, because nothing stale is acceptable. You pick a protocol by whether stale data is useless or sacred.

  • fullyou can still read this even when the letters start to disappear
  • half strippedyu cn stll rd ths evn whn th lttrs strt t dsppr
  • too fary c s r t e w t l s t d

English is roughly half redundant, so you can lose a large fraction of the characters and still recover the meaning. That redundancy is not waste. It is built in error correction.

Half the letters are insurance. You only notice when they run out.

One piece of the protocol is about to matter more than the rest. Low-latency channels are full of tiny acknowledgments: the nod, the mhm, the go on. They tell the sender the message landed. But an acknowledgment is only as good as its authentication, and a nod is trivially forged. Hold that thought. It comes back with a body.

The Mirror 07

The convergence

Set the two tracks against each other and they stop looking like two systems. They look like one system photographed twice, once in positive and once in negative. The columns merge here on purpose.

Line up the costs. The machine has perfect recall and ruinous scaling, cannot forget cheaply, and must read everything. The human has terrible recall and runs almost free, forgets natively and ruthlessly, and must manage everything. One problem is that it cannot forget. The other is that it cannot remember much.

The machine cannot forget and pays for it. The human cannot remember much and pays for it.

Each one's weakness is the other's strength.

Channel Bandwidth

Exploded. No ceiling in sight.

Channel Latency

Collapsed fast, then hit the light wall. Now we fight over microseconds.

Semantic Bandwidth

Flat for five thousand years.

AI aims here
Semantic Latency

Flat for five thousand years.

AI aims here

Two cells solved or wall-bound. Two cells untouched from Socrates to the day before yesterday. AI is the first technology aimed at the flat ones.

The map of the whole argument, in four cells.

That symmetry is not a curiosity. It is the reason the two compose. The AI supplies the wide, perfect, expensive context the human was never issued. The human supplies the ruthless prioritization the AI lacks.

The human is the executive function for the AI's enormous flat context. The AI is the working memory the human was never issued.

Two minds, opposite problems. Put them in a loop and each one covers the other's fatal flaw. When the context is too large for any one mind to regulate, you do the only sensible thing. Hand the firehose to the firehose.

I see you skimming. Good.

It is the rational move. You are throttling input to fit a budget, which is exactly the funnel defending itself. You just proved the point.

The Last Mile 08

The enemy we never beat

Strip the romance off the history and communication was never really a race to move bits. It was a running fight against four enemies at once.

Noise, distance, forgetting, and lies.

Shannon gave us noise. The light-speed wall and the global network gave us distance. Writing and then machines that forget nothing gave us forgetting. Three enemies, beaten or walled. The fourth was always different, because it does not live in the channel. It lives in the question of whether the message is what it claims and came from who it claims.

We beat noise, distance, and forgetting. We never beat lies.

For all of history the medium carried authentication for free. A voice, a face, a handwriting, a wax seal. Every step up the bandwidth ladder quietly stripped that out, trading provenance for reach. The channel does not care whether its payload is true, and the selection pressure runs the wrong way: on Twitter from 2006 to 2017, falsehoods were about 70 percent more likely to be retweetedEstablished and truth took about six times as long as long to reach 1,500 people people, across about 126,000 rumor cascades cascades. The driver was human novelty and emotion, not bots.

AI is the first technology that attacks the funnel and detonates authentication at the same time. You can no longer trust a voice, a face, a writing style, or a video. The pipe got infinitely wide and the last heuristics for who is real went out with the same tide.

Which is where the forged acknowledgment comes back. You leave the meeting certain the idea landed because you saw the nod, the mhm, the scroll to the bottom. But an acknowledgment with no authentication is just a polite shape. The receiver was distracted, or skimming, or being kind.

The nod was spoofed.

The Last Mile 09

The four gates

Suppose the message arrives perfectly and is understood completely. It can still transfer nothing, because understanding was never the last gate.

  1. 01
    Transmission

    The bits arrive intact. Shannon's whole domain. Mostly solved.

  2. 02
    Comprehension

    The meaning lands in another mind. Hard, but not the narrowest gate.

  3. 03
    Acceptance

    The mind agrees to hold the idea as true and act-worthy. Usually the real bottleneck.

  4. 04
    Action

    The idea changes what someone does. Anything short of this transferred nothing useful.

Each gate is narrower than the one before. An idea understood but not adopted has moved zero useful bits. The narrowest gate is almost always acceptance, not comprehension, which means the thing we spent five thousand years optimizing, raw transmission, was rarely the binding constraint.

There is an older engineering discipline aimed exactly at the acceptance gate, and it predates Shannon by millennia. Aristotle called it ethos, pathos, logos. Rhetoric is not decoration on top of the message. It is the technology for the gate that information theory was built to ignore.

The Last Mile 10

The million dollar sentence

Here is the abstraction with a body. A real project, a real chain of people, and a real decision with seven figures on it, made in two to four meetings of about an hour each.

Start with the project and its data. Feed it through the AI, which reads every token. Hand the result to the solutions architect, then to other engineers and peers, then to the Customer, who signs. Each handoff is a hop, and every hop injects noise. No one downstream holds the whole thing, and each person reconstructs it through a different decoder, because the chain shares no common codebook.

Hop 1, Project

the migration must finish before the audit or the contract is void

Hop 5, Customer

the migration must [unclear] before the [unclear] or the contract is [unclear]

Five hops, five decoders, no shared codebook. Meaning decays with every relay. Redundancy is the only thing that gets more of it across.

No one understands it because the chain is longer than the message can survive.

Now do the arithmetic honestly. If the project cannot be fully transferred to a peer engineer over weeks of shared work, then a Customer absorbing it in roughly eighty minutes of real payload, with no shared codebook, is not improbable. It is impossible.

The bandwidth math never closes.

So the sale was never a comprehension transfer. The Customer is not buying understanding. They are buying the belief that the understanding exists and is in good hands. The meeting is the channel, small talk is the handshake that establishes trust before payload, and the verbal channel is the highest-trust, lowest-detail pipe there is. Pushing a high-detail concept through the worst channel for detail turns out to be correct, once you accept that it is a trust sale and not a data transfer.

The check is an authentication payment, not a data transfer.

In the room, the asymmetry is exact. The AI read everything and the human kept what mattered. Neither mind had both.

The AI understood most, the human understood best.

The only thing that ever crosses the last mile is trust.

The Close 11

The flood, and what we are optimizing out

One asymmetry held for all of history. Writing was expensive and reading cheap. Explaining was hard and nodding was easy. AI just broke it.

Production cost fell to nearly zero while consumption stayed pinned at the same human funnel it has always been. The history of communication was about widening the pipe. AI is about shrinking the payload. The history of communication tech was about widening the pipe. This is the first technology pointed at the payload instead, and it points a firehose at a thimble that did not get any bigger.

The funnel is firehosed by a pipe that costs nothing to fill.

So the scarce resource flips. It is no longer who can produce, since anyone can produce infinitely. It is who can filter. The human's oldest liability, the ruthless discarding of almost everything, becomes the most valuable function in the system.

Human editorial judgment is the throttle on an infinite tap.

  1. A letter, months Deep, revised, self-contained. You say the whole thing properly because the next reply is a season away. Darwin built a theory in letters.
  2. Email, hours Considered but quicker. A draft, a pause, a send. Some revision survives, some is lost to the clock.
  3. Chat, seconds Fast, reflexive, shallow. Minimum latency, minimum depth. The pause that depth needs is gone.

As turnaround approaches zero, the question is what kind of thinking we are optimizing out of existence.

High latency was a forcing function, not only a tax.

And lower latency is not simply better. Different turnaround times breed different species of thought. A letter that costs a month makes you say the thing properly. A feed that costs a second rewards reflex and punishes the pause that depth needs. The last mile is the mind.

The honest question, the one with no tidy answer: as turnaround approaches zero, what kind of thinking are we optimizing out of existence, and does an idea absorbed in three seconds ever become yours the way one wrestled with for a month does. There is no bow to tie on that. There is just the question, left open on purpose.

Built by the firehose it indicts.

The only honest page about the funnel is one that watches you fall through it in real time.

Subject, instrument, and audience all obey the same law.

nix.baud reader feed

The page that read you back

The whole field guide made one argument: the channel kept getting wider and the receiver never did, so the only honest move left is to stop blasting a fixed payload at a variable reader and start matching the payload to the reader. This page did that to you while you read it.

The problem, one more time

You take in language at roughly about 39 bits per second, and reading is not much faster. A page has no idea how fast you are going, so it ships every reader the same wall of text and lets most of it fall through the funnel. Skimming is you throttling the input by hand because the page would not do it for you.

The solution: meter the prose to the reader

nix.baud is adaptive bitrate streaming, the trick that keeps video smooth on a bad connection, pointed at prose instead of pixels. Every chapter exists at four bandwidths: the full verbose text, a condensed version, a one or two sentence gist, and a single headline. Your scroll speed is the connection speed. The page picks the rendition that fits.

How it works

  • 01 It measures your scroll velocity every frame and smooths it, so a flick does not whipsaw the text.
  • 02 Slow or stopped serves the full story. Faster summarizes. Slamming the scroll drops to headlines. Ease off and it all flows back.
  • 03 It only ever changes the text below where you are reading, and it never collapses a paragraph you are looking at. That is why it adapts without the screen flashing.
  • 04 The full text is always present underneath, so screen readers and a reader with JavaScript off get the entire essay. The gist is labeled lossy on purpose. Nothing is invented.

Watch the instrument in the corner. It shows your reading bandwidth live: how fast you are moving, which rendition you are getting, and how many of the words you let fly past. That last number is the funnel, measured on you, in real time.

Try it. Fling the page, then stop.

The story compresses to keep up, then unfolds when you slow down.

The essay said the only honest page about the funnel is one that watches you fall through it in real time. This is that page. It was built by the firehose it describes, read by a mind it cannot speed up, and it answered the only way left: by handing you exactly as much as you could take, and no more.