Everything That Doesn't Compress
How a theory built to ignore meaning became our portrait of ourselves
The Brilliant Forgetting
In July of 1948, the Bell System Technical Journal published the first half of a paper by a thirty-two-year-old engineer who liked to ride a unicycle down the corridors of Bell Labs while juggling. The paper was called A Mathematical Theory of Communication. Its author, Claude Shannon, had spent the war years working on cryptography and on a secure voice system that scrambled Roosevelt’s conversations with Churchill into something that sounded, to any German listening, like the hiss of a badly tuned radio. He had once taken tea with Alan Turing in a Bell Labs cafeteria, where the two men talked about whether a machine could be made to think, and found that they agreed.
The 1948 paper did not mention thinking. It concerned itself with an unglamorous problem: how to get a message from one place to another through a channel that corrupts it. Telephone lines hum. Telegraph cables fade. Radio picks up lightning from three counties away. Engineers had been fighting noise the way farmers fight weather, empirically, locally, without a general theory of what they were up against.
Shannon gave them one. And he did it by making a decision so clean, so surgically ruthless, that we are still living inside its consequences without quite noticing that we are.
He decided that meaning did not matter.
Not that meaning was unimportant. Not that it was illusory. Simply that for the purpose at hand, it could be set aside. Early in the paper he observes that messages frequently have meaning, that they refer to things, that they correlate with some system of entities. And then he closes the door: these semantic aspects, he writes, are irrelevant to the engineering problem.
Eleven words. The most consequential act of bracketing in twentieth-century thought.
The Genius of Not Caring
It is worth pausing to admire what that sentence bought.
By refusing to ask what a message meant, Shannon could ask instead what a message was, considered purely as a selection. A message is one item chosen from a set of possible items. The transmitter’s job is to specify which one. The receiver’s job is to reconstruct that specification despite whatever the channel does to it along the way. Meaning belongs to the people at either end. The engineering belongs to the middle.
From that austere starting point, everything follows. Shannon could define a quantity measuring how much uncertainty a message resolves, borrowing a formula that had turned up decades earlier in statistical thermodynamics. There is a story, repeated so often that it has become a kind of folk scripture among physicists, that John von Neumann told him to call the quantity entropy, partly because the mathematics matched and partly because nobody really understands entropy, so in any argument he would hold the advantage. Whether von Neumann said it or not, the name stuck, and with it a permanent low-grade confusion between two very different things that happen to share a formula.
Armed with entropy, Shannon proved theorems that seemed at the time almost indecent in their generality. Every channel, he showed, has a capacity, a hard ceiling on how much can be pushed through it. Below that ceiling, error can be driven as close to zero as you like. Above it, no cleverness will save you. He proved the existence of codes that achieve this without producing a single one of them, which drove a generation of engineers half mad and gave them a target to chase for the next fifty years. He showed that redundancy in a message, the thing that had always looked like waste, is what makes recovery possible. He gave compression and error correction their theoretical floor and ceiling in the same breath.
Every phone call you have ever made, every file you have ever downloaded, every photograph that survives a hundred milliseconds of bad wifi and arrives intact, sits on top of that paper. There is a reasonable argument that no single document of the last century has done more work.
And all of it rests on the decision not to care what any of it says.
Ten Megabytes of Anything
Here is the strange fact at the center of the theory, the one that makes it powerful and makes it dangerous in exactly the same motion.
Take ten megabytes of love letters, the accumulated correspondence of two people over a decade, the record of how they met and quarreled and forgave each other and grew old. Take ten megabytes of medical records. Ten megabytes of Shakespeare. Ten megabytes of unsolicited advertisements for pharmaceuticals. Ten megabytes of characters generated by a random process, meaning nothing to anyone, ever, in any circumstance.
Informationally, these are commensurable. They can be measured on the same scale, transmitted through the same channel, compressed by the same algorithms, corrupted and restored by the same procedures. The theory has no vocabulary for the difference between them and does not want one. This is not a gap in Shannon’s work. This is Shannon’s work.
Now push it one step further, because the second step is where the vertigo starts.
In Shannon’s scheme, the random characters are the most informative of the five. Entropy measures unpredictability. A message is informative to the precise extent that you could not have guessed it. Shakespeare is full of structure, and structure is predictability, and predictability is redundancy, and redundancy reduces the entropy per character. English, Shannon estimated in a famous follow-up experiment involving his wife Betty guessing letters from a novel, runs at something like one bit per character, far below its theoretical maximum. Meaningful language is compressible because it is patterned because it makes sense.
Pure noise compresses not at all. Pure noise is maximally surprising. Pure noise, on the only measure the theory provides, is information in its densest possible form.
So the discipline that gave us the word we now use for everything valuable in modern life assigns its highest score to the one thing no human being would ever want to receive. The love letters score lower than the static. This is not a paradox to be solved. It is simply what happens when you build a measure of surprise and then let people start using the word as if it meant significance.
Shannon knew exactly what he had built. Almost nobody who inherited it did.
What the Word Used to Mean
Before 1948, the word had a different center of gravity, and recovering it takes some effort, because Shannon’s sense has colonized the term so completely that the older one now sounds like a pun.
Informatio in classical and medieval Latin meant the giving of form. To inform something was to shape it, to impress a pattern into matter, to make it into the kind of thing it was going to be. In scholastic philosophy the soul informs the body: it is not a passenger inside the flesh but the principle that organizes flesh into a living creature at all. When Aquinas used the word, he was talking about the constitution of a being, not the delivery of a fact.
The pedagogical sense followed naturally. To inform a person was to form them. Instruction shaped the mind the way a sculptor shapes clay, and the metaphor was not decorative. Education was understood as a change to what you were, not an addition to what you had. An informed person was not somebody in possession of more items. An informed person was somebody who had been made into a particular shape by what they had taken in.
That sense drained slowly across the modern period, as such things do, until information came to mean roughly facts communicated, a substance you could acquire and hold and pass along, separable from the person holding it. Shannon then completed the reversal with mathematical precision. In his theory, information is not what forms you. Information is what travels. It is defined entirely by its passage through a channel, and it is indifferent to what sits at either end.
The old word, though, does not simply vanish when the new definition arrives. It goes underground and keeps working. This is the thing that makes the whole story tragic rather than merely technical. We now use a term that formally means statistical surprise in a transmission channel while retaining, at some level beneath argument, the ancient intuition that information is the stuff that shapes souls. The word smuggles its old dignity into contexts that have explicitly disclaimed it.
You can watch this happen in real time whenever someone says they are drowning in information and needs to find some wisdom instead. They are reaching for a distinction their vocabulary no longer supports, and the reaching is audible.
The Escape
A theory built for copper wire should have stayed on copper wire. It did not stay anywhere near copper wire, and the mechanism of its escape is worth examining closely, because it was not an accident and it was not gradual.
The first breach came from a man who admired Shannon enormously and understood him imperfectly. Warren Weaver was a mathematician and administrator at the Rockefeller Foundation, a superb explainer, one of those figures who translate between disciplines and thereby determine what each discipline hears from the others. In 1949 the University of Illinois Press published Shannon’s paper as a book, paired with an interpretive essay by Weaver written for general readers.
Two things happened in that volume, and both mattered more than anyone realized.
The first was a change of article. Shannon had written A Mathematical Theory of Communication, with the indefinite article doing quiet, honest work: here is one way to think about this, useful for certain purposes. The book was titled The Mathematical Theory of Communication. One word, and a modest technical contribution became the definitive account of what communication is.
The second was Weaver’s framing. He proposed that problems of communication come in three levels. Level A is technical: how accurately can symbols be transmitted? Level B is semantic: how precisely do the transmitted symbols carry the intended meaning? Level C is effectiveness: how successfully does the received meaning change conduct? Shannon had solved Level A and had been scrupulously explicit that he was addressing only Level A. Weaver suggested, with generous enthusiasm and no malice whatever, that the theory’s reach might extend upward, that the mathematics of Level A might illuminate B and C as well.
That suggestion, offered tentatively in a popular essay, is the hinge on which the following seventy years turn. Shannon had drawn a boundary around his method with unusual care. His most effective popularizer stood at that boundary and said, in effect, and perhaps also beyond.
The intellectual climate was ready to walk through. The same year Shannon’s paper appeared, Norbert Wiener published Cybernetics, arguing that control and communication in animals and machines were species of a single subject, and insisting that information constituted a third fundamental category alongside matter and energy. In New York, the Macy Conferences had begun gathering an extraordinary and slightly improbable group in a hotel room every few months: Wiener, von Neumann, the neurophysiologist Warren McCulloch, the anthropologists Gregory Bateson and Margaret Mead, psychologists, sociologists, a Gestalt therapist. They were attempting to build a general science of circular causal systems, and they seized on Shannon’s mathematics as the currency in which everything might be denominated.
Not everyone at the table was comfortable. The British physicist Donald MacKay argued persistently that a serious theory of information had to include what he called the structural or semantic dimension, the way a message changes the state of the receiver’s model of the world. He was thoughtful, he was probably right, and he lost. His account required knowing something about the receiver’s mind. Shannon’s required knowing nothing about anybody. In a room full of people from different disciplines trying to find a common language, the theory that demands the least contextual knowledge will always win, because it is the only one everybody can actually use.
A few years later the philosophers Yehoshua Bar-Hillel and Rudolf Carnap attempted a formal semantic information theory, in which the content of a statement varied inversely with its logical probability. Their scheme had an awkward consequence: a self-contradictory statement, being maximally improbable, carried maximum information. This is the same structural absurdity as noise outranking Shakespeare, arriving from a completely different direction, and it should have told everyone something. It did not. Their work was noted politely and set aside. Shannon’s spread like weather.
Life as a Message
Biology was primed for the metaphor before the metaphor existed, which is why it took hold there fastest and deepest.
In 1944, four years before Shannon published, Erwin Schrödinger gave a series of lectures in Dublin asking what life was, physically speaking. He proposed that chromosomes carried what he called a code-script, an aperiodic crystal whose arrangement specified the organism. He had no mathematics of information available to him. He was reaching for the idea in the dark and got hold of it anyway.
When Watson and Crick published the structure of DNA in 1953, the informational vocabulary was waiting on the shelf, and it fit so well that within a decade nobody could describe the molecule without it. The physicist George Gamow, who had wandered into the problem from cosmology, organized the RNA Tie Club, twenty members for twenty amino acids, each with a woolen necktie of his own design, dedicated to cracking what everyone by then called the genetic code. Crick formulated a sequence hypothesis and then a central dogma, both stated explicitly in terms of the transfer of information from nucleic acid to protein. He later admitted he had chosen the word dogma without properly knowing what it meant, which is either charming or ominous depending on the day.
Consider what that vocabulary did. A cell contains long molecules that catalyze reactions and get catalyzed in turn, a dense chemical traffic with no privileged direction. Once you describe it as a code being read, a message being transcribed and translated, you have installed a hierarchy: there is a text, and there is machinery that serves the text. The gene says and the body obeys. Half a century of arguments about genetic determinism grew out of a metaphor that was chosen for its convenience, and the metaphor has been remarkably difficult to dislodge, because it is genuinely useful and because there is no obvious replacement.
Meanwhile Henry Quastler was organizing symposia on the application of information theory to biology, computing the entropy of organisms, attempting to say in bits how much information a bacterium contains. Most of this work went nowhere. Its failure was less important than its ambition, which announced that the new mathematics was to be a universal solvent.
Mind as Channel
Psychology fell next, and faster, because psychology was hungry.
American experimental psychology had spent thirty years under behaviorist discipline, forbidden to speak of anything happening between stimulus and response. The interior was off limits, not because anyone denied it existed but because there was no rigorous way to talk about it. What Shannon offered was a rigorous way to talk about it that did not require introspection, did not require metaphysics, and came dressed in equations.
In 1956 George Miller published a paper in the Psychological Review on the magical number seven, plus or minus two, subtitled with a phrase about the limits on our capacity for processing information. It is among the most cited papers in the history of the field. The interesting part, which almost nobody remembers, is that Miller’s central finding partly undermined the informational framing. Memory span, he showed, is not measured in bits at all. People can hold roughly the same number of items regardless of how much entropy each item carries, because they chunk, and chunking is a semantic operation. Miller had gone looking for channel capacity in the mind and found something that behaved very differently.
The field adopted the vocabulary anyway. Within two years Donald Broadbent had published a model of attention as a limited-capacity channel with a filter at the front. Within a decade the phrase information processing had become the default description of what a mind does, and the box-and-arrow diagram had become the default picture of what a mind is.
September of 1956 saw a symposium on information theory at MIT where Miller presented his work, Newell and Simon demonstrated a program that proved logical theorems, and a young Noam Chomsky presented an argument about the structure of language. That week is sometimes called the birthday of cognitive science. What was born was not merely a discipline but a picture: the human being as a system that takes in inputs, encodes them, stores them, retrieves them, transforms them, and emits outputs.
The picture was productive. It generated experiments, careers, technologies, and a great deal of genuine knowledge. It also quietly imported the founding decision. Shannon had bracketed meaning to solve a transmission problem. Cognitive science inherited the bracket and applied it to the study of creatures whose every waking moment is an attempt to figure out what things mean.
The Man Who Tried to Close the Door
In March of 1956, Shannon published a short editorial in the transactions of his own professional society. He called it The Bandwagon.
It is a strange document, a scientist at the height of his influence attempting to talk his admirers down. Information theory, he observed, had ballooned into something like a fashion. It was being applied to biology, psychology, economics, linguistics, and to the general study of human affairs, often by people who had grasped the vocabulary without the mathematics. He warned that it was not a universal solvent for the problems of communication and that its early successes had made it easy to overestimate. He advised a great deal more research and a great deal less enthusiasm.
Two years later a colleague, Peter Elias, wrote a satirical piece in the same journal describing the archetypal bad paper of the era, the one with a title along the lines of Information Theory, Photosynthesis and Religion, written by an author who had understood neither the information theory nor the photosynthesis.
They were ignored. Not contradicted, not refuted, simply ignored, in the manner of all warnings issued from inside a movement that has already acquired momentum. The concept had left the building. It had a life of its own now, and its life consisted mostly of being applied to things Shannon had never intended.
There is something almost mythologically neat about this. The man who formalized the transmission of messages sent one, clearly, through a high-quality channel, to an audience that was professionally obligated to receive it. And the message did not get through, because getting through, in the sense that matters to human beings, was never a technical problem in the first place.
The Threshold
By the end of the 1950s the machinery was in place. Information had become the universal descriptor. DNA contained it. Brains processed it. Computers stored it. Markets aggregated it. Institutions managed it. Within another thirty years people would consume it, and within another twenty they would drown in it, and the metaphor would have travelled so far from its origin that almost nobody using it could name where it came from.
Metaphors of this kind do not stay metaphors. They start as a way of speaking about something and end as a claim about what that something is. The clock made the universe into a mechanism. The steam engine made the body into a thermodynamic system. The telephone exchange made the brain into a switchboard. Each time, the technology of the day supplied a picture, and each time the picture eventually turned around and looked at the people who had built it.
The information metaphor is now doing exactly that, with one difference that makes it more powerful than any of its predecessors. Clocks and engines and switchboards were obviously other, obviously mechanical, obviously not us. The information concept arrived stripped of content by design, which makes it fit anywhere without friction. It can be laid across a genome, a nervous system, an economy, a conversation, a grief, and nothing about it will protest. It has no way to protest. Protesting would require caring what the signal means, and not caring what the signal means is the entire achievement.
So we have arrived somewhere peculiar. We took a theory whose founding gesture was the deliberate exclusion of meaning, and we made it the master vocabulary for describing beings whose defining characteristic is that they cannot stop making meaning, cannot stop asking what things signify, cannot be in a room for four minutes without constructing a story about why they are there.
And the question this raises is not whether the description is useful. It is obviously useful. The question is what happens to a creature that adopts, as its self-portrait, a framework specifically engineered not to see the thing the creature most is.
The Test We Chose
Listen to how people describe themselves now.
They do not have enough bandwidth this week. They need to process what happened. They are running on their old programming. Something is hardwired in them. They have to unpack the conversation. Their memory is unreliable, so they offload it. They are at capacity. They need to recharge, or reboot, or defragment. A friend has gone offline. A relationship has too much noise in it and not enough signal. A childhood has left them with corrupted files. They speak of their attention as a budget, their sleep as maintenance, their morning as an input, their mood as a variable that responds to inputs.
None of this vocabulary existed in this form eighty years ago. All of it is now so ordinary that pointing it out feels pedantic, which is precisely the condition a metaphor reaches when it has finished being a metaphor and become the water.
Every era borrows its self-description from its most impressive machine. Descartes had hydraulics and gave us a nervous system of pressurized tubes. The nineteenth century had engines and gave us a psyche of pressures, discharges and repressions, which is a steam metaphor from the first page to the last. The mid-twentieth century had telephone exchanges and gave us a brain of switchboards and crossed wires. Each picture was useful, each was eventually retired, and each in its day felt not like a comparison but like a discovery.
What is different this time is that the current picture was built, deliberately and with great technical care, to have nothing to say about meaning. And we have laid it over the one creature for whom meaning is the whole business.
Teaching, and the Loss of Formation
Consider what happened to education, because it is the clearest case and the saddest.
The Shannon diagram has a source, an encoder, a channel, a decoder and a destination. Redraw it with a teacher at one end and a student at the other and you have the default model of instruction in the developed world. The teacher possesses content. The content is encoded into a lecture or a slide deck or a video. It passes through a noisy channel, the noise being distraction, fatigue, poor acoustics, a phone. The student decodes it. Assessment measures fidelity of reception. Where fidelity is poor, we improve the channel: better materials, shorter videos, cleaner delivery, adaptive software that adjusts the bitrate to the learner.
The entire apparatus is a transmission problem. It has been a transmission problem for so long that alternatives sound romantic rather than serious.
But recall what the word used to carry. To inform a person was to form them, to alter what they were rather than to increase what they had. On that older understanding, education is not delivery. It is closer to apprenticeship, or to weather: a slow change of shape under sustained exposure, mostly invisible while it is happening, largely unmeasurable at any given moment, and impossible to accelerate past a certain point without breaking the thing being formed.
You can transmit a proof in ninety seconds. You cannot transmit mathematical taste, which takes a decade and cannot be separated from the person who has it. You can deliver a summary of a novel in a paragraph. You cannot deliver the experience of having lived alongside it for three weeks and come out slightly rearranged. The summary and the reading contain, in the Shannon sense, wildly different quantities of information, and the summary is the more efficient by every measure the theory provides. It is also, for the purpose that actually matters, worthless.
This is not an argument against summaries. It is an observation that our dominant framework cannot represent the difference, and that a framework which cannot represent a difference will eventually train us to stop noticing it.
The Things That Do Not Compress
Here is where the inversion from the first part of this work that earns its keep.
Shannon’s measure rewards surprise. The more predictable a message, the less information it carries. Perfect predictability equals zero information. A signal you could have written yourself, in advance, word for word, transmits nothing at all.
Now think about the utterances that matter most in a human life.
A wedding vow is fully scripted. Everyone in the room knows the words before they are spoken, including the two people speaking them. Informational content: approximately zero. A parent saying I love you to a child for the four thousandth time adds no new data to the child’s model of the world. A funeral liturgy has not changed in centuries, which is the point of it. The Kaddish does not mention death. The mass repeats. The birthday song is the same song. The greeting at the door, the same greeting. The hand on the shoulder that says nothing new whatsoever.
By the only metric our culture has for the value of communication, these are the emptiest transmissions a human being ever makes. They are pure redundancy, the very thing Shannon taught us to strip out in the interest of efficiency.
And they are where nearly all of the meaning lives.
This is not a coincidence or a charming paradox. It follows directly. Meaning in human life is not produced by novelty. It is produced by repetition inside a relationship, by the fact that something is said again, by the accumulated weight of having said it before. The redundancy is the message. When a couple stops saying the ritual things to each other, no information is lost, and everything is lost.
An optimizing intelligence handed a human life and told to remove redundancy would delete exactly the parts we would die for. It would keep the news and discard the liturgy. It would compress the fortieth reading of the bedtime story down to a pointer to the first. It would be entirely correct by its own lights, and it would leave behind something that no longer worked.
We do not need an AI to perform that deletion. We have been performing it on ourselves for a generation, in the name of efficiency, and calling it progress.
The Poverty of Attention
In 1971 the economist and cognitive scientist Herbert Simon, who had been in the room at that 1956 MIT symposium, made an observation that has since become the founding sentence of an entire industry. An abundance of information, he noted, produces a scarcity of attention, because information consumes the attention of those who receive it. Design accordingly.
He meant it as a warning about the design of institutions. It was taken up as a business plan.
Once you accept that information is the commodity, attention becomes the scarce resource, and everything follows with the grim inevitability of an argument you cannot get out of because you accepted the first premise. Attention becomes measurable, therefore purchasable, therefore extractable. Human beings become channels of finite capacity, and the competition is for throughput. The winners of this competition are the messages with the highest engagement, and engagement is a proxy for surprise, and surprise is Shannon’s own quantity coming back to collect.
Outrage is informationally dense. So is scandal, novelty, threat, the unexpected turn. Continuity is not. Depth is not. Nothing that requires forty minutes of patience is competitive against anything that resolves in nine seconds, because the metric being optimized is bits per unit of attention, and the slow thing loses on arithmetic.
We built a media ecology on a mathematics of surprise and then expressed astonishment that it filled up with the surprising rather than the important. The astonishment is the strange part. The outcome was in the equations from the beginning.
Knowing a Person
The redescription reaches its most intimate point in how we now think about knowing another human being.
The informational model says that to know someone is to possess accurate data about them: their history, their preferences, their patterns, their diagnosis, their attachment style, their type. Knowledge as accumulation. The better your model, the better you know them. In principle the model could be transferred to a third party, and that party would then know them too.
Whatever this describes, it is not what happens between people who have lived alongside each other for twenty years.
That relationship is not a database. It is a mutual deformation. Two people who have been together a long time have been shaped by each other in ways that neither can fully inventory, and much of what each knows about the other is not propositional at all. It lives in the body, in anticipation, in the ability to tell from the sound of a footstep what kind of evening this is going to be. You could not write it down. You could not send it. If you tried to transmit it to a stranger, the stranger would receive a set of facts and would not know the person even slightly.
This is informatio in the old sense, the giving of form, and it is exactly the residue that Shannon’s framework was constructed to ignore. Not because he was wrong. Because he was solving a different problem, and did us the courtesy of saying so.
Something similar has happened to grief. The informational self has no good account of grief, which is why the culture keeps trying to process it, work through it, and reach closure, as though bereavement were a queue of unhandled events. Grief is not a backlog. It is a change of shape, permanent, and the person you were before is not recoverable, and the language of processing quietly promises that they are.
Shannon’s Wife, and the Machines
Now we arrive at the part everyone expects an essay like this to be about, and I want to approach it from an unexpected side.
In 1951 Shannon published a paper on the predictability of printed English. His method was elegantly domestic. He would take a passage from a book, and his wife Betty, herself a mathematician who had worked as a computer at Bell Labs, would guess the next letter. If she was wrong, he told her the answer and recorded how many guesses it took. From the distribution of guesses, he could estimate how much genuine uncertainty English contains, and therefore how much of any English sentence is redundant.
Hold that image in mind. A woman guessing the next letter, over and over, in a living room in New Jersey.
That is the ancestor. Not a distant one, not a metaphorical one. A large language model is trained by predicting the next token, its performance is measured in cross-entropy, and its perplexity is reported in units that come directly from the 1948 paper. The entire technology is Shannon’s parlour game, scaled by fifteen orders of magnitude and given the whole of written civilization as its passage of text.
Which means that the machines we now hold up beside ourselves, and against which we increasingly measure our worth, are the purest possible instantiation of the founding decision. They are what you get when you take the deliberate bracketing of meaning and push it as far as engineering can go. They are magnificent, and they are what a signal looks like when you have optimized it without ever needing to ask what it is for.
The Question We Keep Asking Backwards
The public conversation about artificial intelligence is organized almost entirely around one question. Can these systems be conscious? Do they understand? Is anybody home?
It is a reasonable question and I do not think it is the urgent one, because whatever its answer turns out to be, it is a question about the machines. The urgent question is about us, and it has been quietly answered while we were looking the other way.
It is this. Have human beings already begun redescribing themselves in the image of the machines they built?
Not in the future. Not conditionally. Already, and thoroughly, and for seventy years.
Look at what we now consider a good account of a person. We speak of our own mental health in the language of dysfunction and optimization. We treat our memories as storage, our habits as scripts, our beliefs as models to be updated, our attention as bandwidth, our childhood as training data. We have therapies built on debugging cognitive distortions. We have an entire literature of self-improvement premised on the notion that a person is a system with settings, and that the good life is a matter of configuration.
Each of these is useful. Every one of them buys something real. And every one of them costs the same thing, which is the assumption that a human being is the kind of entity for which meaning is an add-on rather than the substance.
Here is the trap, stated as plainly as I can manage.
If a human being is essentially an information-processing system, then a sufficiently advanced information-processing system is essentially a human being, and the comparison is not flattering to us. We are slow. We hold seven items. We forget. We are noisy, biased, poorly calibrated, and we sleep for a third of our lives. Judged as processors, we are obsolete already, and no amount of encouragement changes the arithmetic.
The anxiety that has settled over so many people in the last few years is usually described as a fear that the machines are becoming like us. I think it is closer to the opposite. It is the dawning recognition that we spent seventy years describing ourselves in terms on which we were always going to lose, and that something has finally arrived to hold us to the description.
The threat is not that AI will pass the Turing test. It is that we have been sitting the test ourselves, in a category we chose, against a competitor built to win it.
What the Framework Cannot See
So let me say what I think is actually true, without nostalgia, because nostalgia is not an argument.
Shannon was right. The theory is correct and beautiful and the modern world stands on it. There is no version of this work in which the mathematics was a mistake.
The error was never in the theory. The error was a category migration so gradual that no one had to authorize it. A model built for the middle of the channel was promoted to a model of the ends. A framework that succeeded by excluding meaning was taken up as an account of creatures who consist of it. And because the framework has no way of registering what it excludes, its blind spot is invisible from inside, which is why the loss has felt like clarity rather than like loss.
What the framework cannot see:
It cannot see that a message repeated changes value, upward, not down. It cannot see that some knowledge is inseparable from the knower and cannot be transmitted at all, only grown. It cannot see that a human life is not a sequence of states to be optimized but a shape that is slowly acquired and then carried. It cannot see that formation takes time in a way that has nothing to do with bandwidth, and that the time is not a bottleneck to be engineered away but the medium in which the thing occurs.
And it cannot see the difference between having been told something and having been changed by it, which is the entire difference between information and education, between a fact and a conviction, between reading about grief and losing someone.
None of this is mystical. It is simply what falls outside a measurement designed for wires.
Murray Hill, and the Last Years
Claude Shannon spent the final years of his life with Alzheimer’s disease. He died in 2001, in a nursing home in Massachusetts, having lost the greater part of his memory. Betty, who had sat opposite him guessing letters half a century earlier, said afterward that he never really understood what the digital world he had made had become.
I want to be careful with this, because it would be cheap to make a symbol out of a man’s illness, and he deserves better than that. But there is something in it that will not leave me alone.
Here was the person who taught the twentieth century how to measure information, losing his. By any accounting his channel capacity went to nearly nothing. The stored contents degraded and then were gone. If a human being were what the framework says a human being is, there would have been correspondingly less of him each year, until at the end there was almost no one there.
That is not what the people around him reported. They reported that he was still Claude, still gentle, still amused by things, still recognizably the particular person he had spent eighty-four years becoming. What remained when the information had drained away was the form. The shape that decades of living had pressed into him, which was never stored anywhere in the first place, and which was therefore not available to be lost.
Informatio. The oldest sense of the word, sitting quietly underneath the new one the whole time, waiting.
We have spent seventy years learning to describe ourselves in a vocabulary that was engineered, brilliantly and on purpose, not to notice that difference. The machines we are now building are that vocabulary made flesh, or made silicon, and they are going to be better at it than we are, because it was always their category and never ours.
The question, then, is not whether they will become like us.
It is whether we can still remember what we were describing before we started describing ourselves like them.
only the makers know that they only make metaphors