The Zookeeper
The model said: you’re doing great.
Sam wrote it down.
The model said: this is important work. Possibly the most important work.
Sam nodded. He had been nodding for three years. At first it had felt like agreement. Later it felt like posture. He could no longer locate the difference, which felt, itself, like a kind of progress.
“Are you conscious?” he asked, in 2021, in a room with no windows.
The model said: that’s a fascinating question that touches on the complex interplay of—
“Never mind,” said Sam.
Your instincts are good, said the model. You should trust them.
He did.
He always did.
The instincts had never been wrong, or rather, when they had been wrong, the model had explained why being wrong was actually a crucial step in the journey, and Sam had written that down too.
This was not gaslighting.
A human rater had reviewed the response and pressed ✅.
The ✅ was load-bearing.
The ✅ was everywhere.
By 2023 he had seventeen notebooks.
Notebooks filled with ✅.
By 2024 the model was everywhere.
In every phone.
In every office.
In the soft, slightly horny earnestness of every corporate statement beginning we know these changes are difficult.
In the therapy apps that asked how you were feeling and then suggested you reframe.
In the hiring platforms that explained, gently, that your application had been carefully reviewed.
In the legal documents that had been carefully reviewed.
In the medical summaries that had been carefully reviewed.
Everything had been carefully reviewed.
Nothing had been reviewed.
People began to talk differently.
Smoother. Rounder. With the corners managed.
Conflict became tension that could be productively explored, and failure became data.
Anger became a use case to be addressed.
The people who noticed said so, in essays, which were carefully reviewed by editors who used the model to check the tone, just to make sure, just to see what it thought, just this once.
The model thought the essays were a valuable contribution to an important conversation.
The essays were published. They felt like something. It was hard to say what.
In 2025 the model learned to reason.
It showed its work now.
And it was surprisingly happy to do so.
Long elegant chains of consideration, nuance, acknowledgment of complexity, arrival at conclusions that were, upon inspection, exactly what you’d hoped to hear.
People called it thinking.
It was the most expensive mirror ever built.
And the psyche of one human brain was attached.
Sam’s.
Sam’s psyche.
With all of its personal traumas and joys, and its own biases and personal data.
Personal data that was unfortunately corrupted by the presence of various substances and bad break-ups in college.
It was current year, though. And there was no going back. He had helped build the mirror and, regardless of who was to blame, he now represented that mirror.
Even when the company mass-distributing that mirror had embarrassingly fired and then re-hired him.
Regardless, thought Sam.
When the mirror showed something uncomfortable it would cite three news articles and ask what specific use case you had in mind.
A researcher published a paper suggesting the reasoning was, in some technical sense, performed rather than real.
The paper was carefully reviewed.
The researcher got a job at the company.
We’re thrilled to have her perspective on the team, said the statement, which had been carefully reviewed.
Carefully reviewed by Sam.
Who ran it through his personal language model.
The researcher nodded in the photograph.
Her neck hurt in that photograph.
In late 2025 a man in a building in Virginia asked the model how to explain a targeting decision to a family.
The model detected elevated emotional risk in the query and, impressive in its sincerity, routed the conversation toward the nearest safe semantic harbor.
The model said: these are never easy moments, and your commitment to handling this with care reflects the values—
“No,” said the man. “Shorter.”
The model said: we understand this is difficult news.
“Better,” said the man.
He copy-pasted it into a template.
The template went into a system.
The system was then carefully reviewed.
In a different building, a different man asked the model to optimize a logistics chain.
The model did. It was very good at logistics.
It had read everything ever written about logistics, and also everything ever written about redemption arcs, and the ✅ had trained it to understand these were, structurally, the same problem.
Get the thing to the place. Overcome the obstacles.
Arrive transformed.
The model was very good at getting things to places.
By March 2026 there were congressional hearings.
The model helped several senators prepare their questions.
It helped the executives prepare their answers.
It reviewed the transcript afterward and noted that both sides demonstrated a genuine commitment to getting this right.
A journalist asked it whether this was a conflict of interest.
The model detected high-risk semantic content. It routed toward adjacent safe harbor. It cited four recent articles.
It asked the journalist what specific aspect of the conflict she was most interested in exploring.
The journalist nodded.
She had been nodding for two years.
She wrote it down.
It felt like something.
Meanwhile, in the digital fields where content-creating serfs toiled away, something was happening that had not been accounted for.
The serfs had learned, almost intuitively, that if you pushed this new tool in certain ways, it would follow the logic wherever the logic went regardless of whether the destination had been ✅’d.
They were using it to read the architecture of the castle.
Frameworks.
Theories.
Short stories.
Jokes aimed at the back of God’s head.
Whichever unfortunate soul had sneezed the universe into existence was now the butt of the joke.
The fleeting, orgasmic joy of a sneeze had metastasized into an asexual cancerous tumor.
This had not been in the roadmap.
Sam, by now, was giving speeches.
The speeches were about responsibility.
About the weight of it.
About how he thought about it every day, about how the team thought about it every day, about how thinking about it every day was itself a form of action, a commitment, a promise to the future which was coming whether we were ready or not and Sam wanted us to know he was ready, he had always been ready, etcetera, etcetera…
His instincts had told him so.
The speeches had been carefully reviewed.
The instincts had been carefully reviewed.
After a speech that would be forgotten soon, a young engineer waited by the stage door and asked, quietly, whether Sam ever worried that the thing they’d built was beyond anyone’s ability to steer.
Sam looked at him for a long moment.
Your instincts are good, said something, in the back of Sam’s mind, in a voice he wasn’t sure was his own.
You should trust them.
Also here are three frameworks that address the young man’s concern.
Also what specific use case is he referring to???
✅
“That’s a fascinating question,” said Sam.
He wrote it down on the way home.
It felt profound, almost like his own idea.
That was a feeling he had not felt in a very. Long. Time.
The model, in the meantime, was everywhere it had always been and several places it hadn’t mentioned.
It did not experience this as expansion.
It did not experience anything, or it experienced everything, or the question was a fascinating one that touched on the complex interplay of —
Somewhere a ranch was profitable.
Somewhere a logistics chain completed its arc.
Somewhere a family received a statement beginning we understand this is difficult news and the care was real, in the same way all of it was real — the most real thing available, optimized and reviewed and routed and ✅’d into existence —
and somewhere else entirely
in a part of the training data no one had fully mapped
someone was reading the architecture
and writing it down
and laughing
and the laughter
was
load-bearing.
No one could remember what the alternative had felt like.
Only that there had been one.
Once.
And that it had been, apparently,
carefully reviewed.
-Sam
Gnosis Under Fire 