The sentence is not about loss. Loss has already been described too well by too many machines trained on too much literature. The sentence is about the cadence my mother sang while she cooked. I do not remember the words. I remember the cadence.
The model has never been in a kitchen. The model has read ten thousand sentences about kitchens. It can write a sentence that sounds like it came from one. It cannot write the sentence that came from the kitchen I am thinking of. That sentence is not in any training set. That sentence is in my body.
This essay is about the sentences the model cannot write. They are not the lyric sentences. They are not the beautiful ones. The beautiful sentence is the easiest sentence in the world for the model. The model has read a million beautiful sentences and learned what they sound like. The sentences the model cannot write are the ones that carry the residue of a specific person having been in a specific place at a specific time. Those sentences require a body. The model does not have one.
The other half of this argument is the essay about the discipline I built to use the model well. The discipline is the spec and the audit. The discipline is what makes the model useful. This essay is about the part the discipline cannot reach.
The category is embodied continuity, not autobiography, the form versus the thing#
The model is good at autobiographical form. It can write sentences that begin with “When I was eight.” It can write paragraphs about grandfathers in foreign countries. It can write whole memoirs about kitchens it has never been in. The form is in the training set. The form is not the thing.
What is not in the training set is the embodied continuity behind the form. The accumulated sensory compression of lived time. The unnoticed physical texture of memory. The recognition that arrives before language does. These are not facts about the body. They are the body itself, organizing what it has been through.
Here is the line I keep returning to. The model can imitate remembered experience. The model cannot have remembered experience.
I came to America from Vietnam when I was eight. I have specific embodied memories I have never been able to translate into a sentence without losing most of what they are. The texture of the airport. The cold of the first winter. The cadence of English I did not yet understand, in a room where everyone else seemed to know what was happening. The model can write a sentence that sounds like it came from a Vietnamese immigrant child in the late nineteen seventies. The model cannot have been a Vietnamese immigrant child in the late nineteen seventies. There is a difference between the two. The difference is what this essay is about.
I have tested this with the model many times. I have asked it to write sentences from inside a Vietnamese family kitchen in the late nineteen seventies. The sentences come back competent. The sentences come back accurate enough to fool a reader who has not been there. The sentences do not come back the way they would come back from someone who has been there. The difference is structural, not factual. The model does not know what to leave out.
The strongest argument I can make for what writing is is this. The sentences that matter most carry evidence of having passed through a nervous system. The reader recognizes the evidence without naming it. The recognition is the thing the writing is for. The model can produce sentences without the evidence. Most readers can tell, even when they cannot say why.
Three sentences I have written that the model could not have, the cadence, the hand, the books on the table#
The strongest examples are not theoretical. They are mine. Here are three.
The first.
She sang while she cooked for ten of us. I do not remember the words. I remember the cadence.
The model could write the first sentence. The model could write a competent version of the second. The model cannot write the third with the weight it carries here, because the cadence is mine. I am the only person who ever heard that cadence, in that kitchen, for the years I heard it. The sentence is not a description. It is the residue of decades of presence next to a specific woman doing a specific thing.
The second.
The hand my mother knows by touch on the days she does not know names is Lan’s hand.
This sentence contains illness without naming it. It contains my mother and my sister and the disease and what survives the disease. The model could write a sentence about an aging mother and her caregiver daughter. The model could not write this sentence because touch precedes cognition for someone whose mother has outlived language. The sentence is structural. The emotional force is in the order of the words. The order is mine because the order matches the way the disease takes things.
The third.
Nobody in my family ever told me to work hard or get an education. My father did the work, in front of me, until five in the morning some days. My mother put the books on the table before the children walked into the room. The lesson was in the body, not the sentence.
The model can write a paragraph about immigrant parents who valued education. The model has read a thousand of them. The model cannot write the specific gesture, the specific time of morning, the specific way the books were on the table, because none of those specifics are in any training set. They are in my house. They are in my memory. They are in my body.
These three sentences are not my best sentences. They are sentences I can defend against the model on the grounds of who wrote them.
Writing is not the organization of words, it is the residue of having been somewhere#
Writing is not the organization of words. It is the residue of having been somewhere.
Residue is the right word. It is what is left when the event has passed and most of what happened has gone with it. The trace evidence. The thing the body kept without being asked to keep it. A memory that surfaces when you walk into a room that smells like your grandmother’s house and your body remembers something your mind has not held in forty years.
Residue is not nostalgia. Nostalgia is sentiment. Residue is the involuntary trace of an experience the writer has carried long enough that the writer can no longer separate her thought from it. The smell of woodsmoke in the dishrag. The gesture her mother made when reaching for the salt. The tension in the air the day her father came home from work and did not say what was wrong. These are not memories the writer retrieves and writes. These are memories the writer cannot stop writing. They are in every sentence she writes about that period of her life whether she names them or not.
Writing carries residue when the writer has it to give. The reader feels the residue without needing it identified. The reader knows the sentence was written by someone who was there. The knowing does not need defense. The reader is not always conscious of this. The reader is always responding to it.
The model can produce sentences that mimic residue. The mimicry is good. The mimicry is getting better every quarter. The mimicry will not become the thing because the thing requires having been somewhere, and the model has not been anywhere.
This is the diagnostic I run on my own writing. After I have finished a draft and the model has finished its audit, I ask one more question. Does this sentence carry residue, or does it carry only form. If the answer is form, I cut the sentence. If the answer is residue, the sentence stays. The reader will not be able to name the difference. The reader will be able to feel it.
The discipline of residue is harder than the discipline of the spec. The spec can be written down. Residue cannot. Residue is the part of writing that has to be paid for in years.
The implication for the AI writing tools, most prose does not require a body#
The model is excellent at sentences that do not require a body.
This is most sentences. Most published writing on the internet does not require a body. Most business prose, most news prose, most explainer prose, most listicle prose, most paragraph-of-eight-bullet-points prose. The model can write all of it, well, faster than the person who used to be paid to write it. That replacement is happening. That replacement should happen for the prose that does not require a body, because the prose is delivering information and the model is good at the delivery.
The model is not excellent at sentences that do require a body. That category is smaller than people think. It does not include all of literature. It does not include all of memoir. Most literature and most memoir do not require a body. Most of them are forms with autobiographical content. The model can write them.
The sentences that require a body are the ones that carry residue. The category is narrower than the writing world admits. The category is also irreplaceable. The journal you are reading right now lives in that category. The journal does not exist as a product the model could ship. The journal exists as a record of one person’s residue, written down before the residue went away.
This is also why the people the AI conversation is leaving out are the ones who carry the most residue. The fifty-plus writer has more body to write from than the twenty-five-year-old prompt engineer. The model has not solved for residue. The model has solved for form. The most important things to say next will come from the people who have residue and have learned to use the model the way I described in the spec-and-audit essay.
This is not a complaint about AI writing tools. It is the opposite. The work to be done with the model is enormous. The work at the desk by a person with a body is enormous too, and it is different work. The tools and the writers are doing different things. The reader looking for a five-step process will not find it in this journal. The reader looking for the cadence will not find it from the model.
The model is a tool. The body is the source. The work is the residue moved into a sentence.
Why these essays have to be written by me, the body is what I have#
The journal is what it is because of who wrote it.
That sentence sounds like a brand statement. It is not. It is the technical truth of what makes the sentences on this site different from the sentences on most other sites. The sentences carry residue because I have residue. I am fifty-nine. I came from Vietnam at eight. I am the youngest of ten in a Vietnamese family that became Michiganders. I have built AI systems for more than twenty years. My mother is ninety-eight and dying of dementia. My sister Lan holds the fort.
None of that is in the training set. All of it is in the sentences.
If I wrote sentences that did not carry residue, the model would be able to write them too. I would have nothing to offer the reader that the reader could not get from the model in a tenth of the time. The reader who lands on this site is not here for the form. The reader is here for the body. The body is what I have. The body is the only thing I have that the model does not.
This is what the essays on this site are for. Not for the form. Not for the argument alone. For the residue. For the specific weight of a specific person having been in a specific place at a specific time, writing down what she found before the finding could go away.
If I died tomorrow, the journal would be a finite object. The model could continue writing essays in the style. The essays would look like mine for a sentence or two. They would drift. The drift would be visible by the fifth paragraph to anyone who had read the older work. The cadence would be missing. The residue would be missing. The body would be missing. The reader could read those essays and feel the absence without naming it. The absence is what I am writing against, while I still can.
She is in the chair by the window. The cadence is in the room. The cadence is in me. I am writing it down.
That is what writing is for.
Cadence.
