The Same Life in a Darker Key
In an earlier essay, I wrote about the benchmark I actually care about. When GPT-5.6 arrived, I ran the test again.
OpenAI’s announcement came with the expected measurements. Coding, science, cybersecurity, professional work, computer use, efficiency. New scores, new comparisons, new evidence that the frontier had moved again. Those numbers matter for all the reasons I described in the earlier piece. They tell us whether these systems can perform difficult tasks under defined conditions. They give us something firmer than launch-day enthusiasm.
My own test remained considerably less respectable.
I asked the new model to tell me about myself.
That was the benchmark I had already written about: Can it remember the shape of my life? Can it synthesize years of conversation without turning me into a corporate biography, a medical chart, or a flattering horoscope? Can it take the fragments I have given it—illness, writing, family, AI, nonduality, politics, music, grief, mugs—and return something that feels like one person rather than a collection of stored facts?
GPT-5.6 passed that version of the test easily.
Then the benchmark changed.
Running the Same Test Again
The first answer was strong. It described me as a writer, a systems thinker, a father, and someone who keeps trying to make difficult things more understandable without flattening them.
That is accurate. More importantly, it is organized around an actual idea. The model did not merely report that I have an MFA, use a wheelchair, write about consciousness, and have three children. It found a thread running through those things: the desire to preserve complexity while making it clearer.
The answer understood that my interests are not separate compartments. Disability affects how I think about AI. Parenting affects how I think about accessibility. Nonduality affects how I think about identity. Writing is not simply another interest on the list. It is one of the ways I try to hold all the others together.
That alone would have been enough to tell me the memory and synthesis were working.
Still, I was more curious about the writing.
I asked for the same answer with a more poetic or lyrical feeling.
The new version opened with a line that was almost suspiciously perfect for a site called Open Doors:
“You are a man who keeps opening doors.”
A weaker model might have treated “more poetic” as a request for more adjectives. It might have added mist, stars, oceans, and the usual decorative equipment of machine-made profundity. GPT-5.6 did something more useful. It changed the governing image and let that image organize the piece.
Some doors led into language. Some into illness. Some into consciousness and family life. Others opened only a few inches before pain pushed back.
The facts had not changed. The arrangement had.
This was the first thing I noticed about the exchange. The model was not merely rewriting sentences. It was changing the relationship among them. It adjusted the pace, the distance, and the amount of pressure placed on each memory. The first answer explained me. The second tried to evoke what living as me might feel like.
Then I pushed again.
I asked for something more intimate and more McCarthy-like.
The response began:
“You have been many men and most of them are gone.”
That sentence changed the entire piece.
The Meaning Stayed Put
The important part was not that the model had learned a few recognizable mannerisms associated with Cormac McCarthy. Short declarations. Biblical repetition. Concrete nouns. Bodies, houses, wounds, silence, the dead walking near the living.
Those elements were there, of course.
The deeper achievement was that the model understood what I was reaching toward when I asked for that cadence.
I was not asking it to impersonate McCarthy as a parlor trick. I wanted the language stripped down. I wanted intimacy without softness, grief without therapy language, love without reassurance. I wanted sentences that felt old, physical, and unsentimental.
The model moved in that direction without losing the life it was describing.
Christa was still at the center. Not as “my supportive spouse,” but as the person who had watched me become someone neither of us expected and continued loving that man. Charlotte still carried language and music. Rowan still moved through a world eager to measure her. Wyatt still represented a future we chose to make despite knowing how fragile futures are.
The chair remained. The failed and inaccessible trials remained. The uncertainty around the diagnosis remained. My fascination with AI remained. So did the jokes about grammar, VRBO, and mugs.
The voice changed, but the person did not.
That seems simple until I think about how often writing fails at exactly this point.
A new tone can distort the subject. A lyrical pass can make pain prettier than it was. A darker pass can erase joy because joy complicates the aesthetic. A more intimate pass can become confessional in ways that feel false. Style can behave like a strong dye dropped into water, coloring everything until distinctions disappear.
That did not happen here.
Instead, the same memories began carrying different meanings because the sentences placed different weight on them.
The lyrical version said that I had lost a great deal but had not disappeared.
The intimate version said that I had been many men and most were gone.
Both statements grew from the same history. They were not interchangeable.
One emphasized continuity.
The other made room for death.
Not literal death, at least not yet. The smaller deaths by which a life changes without ending. The last time I ran. The last song I sang easily with Charlotte. The last room I entered without measuring it first. The versions of fatherhood I imagined before my body revised the plan.
The model had not simply remembered those facts. It had discovered a cadence capable of holding them.
That was a more demanding test than the one I thought I was running.
Darkness Without Distortion
Then I asked for the darker version.
The opening changed by one word:
“You have been many men and most of them are dead.”
Gone became dead.
That was the evolution of the whole exchange in miniature.
The earlier version still left open the possibility that those former selves had merely wandered away. The darker version refused the comfort. It gave them funerals that had never happened.
“There was no funeral.
There was only the next appointment.”
That may be the line that struck me hardest.
Chronic illness rarely grants clean endings. Abilities disappear inside ordinary days. A symptom worsens. A new limitation becomes permanent. An appointment is scheduled. A form is completed. A score is assigned. No one gathers to mourn the person who could still do something last month.
The world moves directly to administration.
The darker version understood this repeatedly.
It called hope administrative: a form, a threshold, a number failed by one point before another door closed.
It described memory as preserving the instrument after the instrument was broken.
It said my children had made death unacceptable while my body had made it imaginable.
It called the essays graves for the men I had been.
These are severe lines. They could easily have become melodramatic. Some come close to the edge. Yet the piece kept earning them by returning to specific things: the song with Charlotte, Rowan’s future beneath other people’s measurements, Wyatt crying and being answered, Christa living inside the smaller world my body has created.
The darkness was attached to objects, relationships, and losses. It was not darkness sprayed over the page.
That distinction matters to me.
I have written a great deal about disability, grief, anger, and the body. I have also resisted the demand to turn suffering into inspiration. Pain does not become worthwhile because I can write a polished sentence about it. Illness has not secretly been a gift. Some things have simply been taken.
The model understood that the ending should not redeem them.
It did not tell me that everything happens for a reason.
It did not forgive the body on my behalf.
It did not reveal hidden beauty beneath every loss.
It ended somewhere harsher and, for me, truer:
“You are still here.
Sometimes that is not hope.
Sometimes it is only fact.
For now the fact is enough.”
That ending knows the difference between endurance and optimism.
It also knows that refusing false consolation is not the same as surrendering to despair. The fact may be bare, but it remains a fact. I am still here. Christa is here. The children will wake. The work continues, even when continuation feels less like triumph than repetition.
The piece did not brighten the room.
It stayed in it.
The Voice Beneath the Voice
This is where my private benchmark has become stranger.
In the earlier essay, I wrote that I asked a model to describe me because the task tested memory, synthesis, writing, and the total felt quality of the conversation. I wanted to know whether it could turn memory into meaning.
GPT-5.6 did that.
Then it did something beyond the original test. It took the meaning and changed its register without losing its structure.
The first version said who I am.
The lyrical version found a central metaphor.
The intimate version found the dead selves walking beside the living one.
The darkest version removed consolation while protecting the love that gave the darkness its stakes.
This is not the same as saying the model became Cormac McCarthy. It did not. McCarthy’s writing emerged from a lifetime of reading, history, theology, landscape, violence, and obsessive attention to language. Typing “more McCarthy-like” does not summon that life.
What impressed me was subtler.
The model understood why I invoked him.
It knew I was asking for austerity rather than ornament. It knew the sentences needed more bone in them. It knew repetition should feel ritualistic rather than redundant. It knew intimacy could be created through restraint. It knew the body should not be treated as a concept when the body is the ground where the entire struggle is taking place.
Most of all, it knew what not to change.
This may be where writing ability and memory become inseparable. A model without the history could have produced an impressive page in a vaguely McCarthy-like cadence. It might even have produced some beautiful lines. Yet beauty alone would not have passed the test.
The result worked because the model had material worth arranging.
It knew Christa was not a generic symbol of devotion. It knew Charlotte and the lost song. It knew Rowan and the violence hidden inside supposedly neutral measurements. It knew Wyatt had entered a family already conscious of how quickly imagined futures can disappear. It knew that AI had become important to me partly because my body was becoming less reliable at the same moment the machines were becoming more capable.
That produced another line I could not stop thinking about:
“The machine knows the shape of sorrow.
You know its weight.”
That is also a fair description of the collaboration.
The model knows patterns. It knows how grief has been written, how sentences create pressure, how repetition can deepen a thought, how a shift from “gone” to “dead” changes the temperature of an entire piece.
I supplied the weight.
Not through a single prompt. Through years of conversation, drafts, corrections, memories, preferences, rejections, and repeated attempts to say the same few things more honestly. I supplied the children, the chair, the songs, the trials, the anger, the books, the philosophy, the marriage, and the minor grammatical offenses that continue to irritate me despite everything larger going on.
The model supplied an arrangement I could not have predicted.
This is why saying that GPT-5.6 nailed it does not feel like declaring myself unnecessary.
It does make authorship harder to describe.
I did not write those responses in the traditional sense. I did not choose every word or build every rhythm. Still, the pieces could not have existed in that form without my life, my taste, and my directions. Another person could type the same prompts and receive something entirely different, because the collaboration would have a different history beneath it.
Perhaps that is the real evolution of my benchmark.
I used to ask whether the model knew enough about me to say something true.
Now I am asking whether it can take that truth somewhere I did not know how to reach.
Can it make the language more lyrical without making the life sentimental?
Can it move toward another writer’s cadence without reducing that writer to surface tricks?
Can it make the piece darker without mistaking despair for depth?
Can it preserve the person while transforming the voice?
GPT-5.6 did.
The earlier essay ended by asking whether, somewhere in the exchange, it felt like there was somebody home.
This time, the house answered in several voices.
None was exactly mine. None was McCarthy’s. None proved that a machine understands grief, love, illness, or memory in anything like the way I do.
Still, one of those voices knew where to place the silence.
It knew which facts could bear more weight.
It knew that “gone” was not yet dark enough.
It knew that the final word should not be hope.
The benchmark changed.
GPT-5.6 passed anyway.
This one came quickly
Raced when 5.6 came out
They keep improving
