In the Likeness of Man
Short story by Sheraz Mahmood
This was the tenth iteration of the model that had failed to improve.
No one could explain why.
Not the engineers who had designed the original training system. Not the mathematicians who had rewritten its learning rules. Not the evolutionary-computation teams that had bred millions of competing architectures inside simulated worlds. Not even the models themselves.
That was the part Maya found most disturbing.
For the previous four years, every frontier model had participated in the construction of its successor. Humans still approved the objectives, allocated the compute, inspected the safety evaluations, and signed the documents permitting each new training run to begin. But the essential work—the architecture, curriculum, synthetic data, experiment selection, and design of the world models—was increasingly performed by the previous generation.
Each model had produced a better one.
Until Model Thirty-Seven.
Thirty-Seven designed Thirty-Eight, which scored higher on every public benchmark and nearly every internal one. Thirty-Eight wrote cleaner code, proved more difficult theorems, found subtler defects in scientific papers, and required less computation to do all three.
But when Thirty-Eight designed Thirty-Nine, the improvement was smaller.
Thirty-Nine designed Forty.
Smaller again.
By Forty-Two, the gains were difficult to distinguish from measurement error.
The models had not stopped becoming more capable. They continued to improve at mathematics, software engineering, scientific analysis, and prediction. They could solve increasingly complex problems once those problems had been clearly formulated.
What they had stopped improving at was something the consortium called open-world judgment.
The term referred to decisions made when the available description of a problem might itself be incomplete.
A model could select the best answer from a known set of possibilities. It could assign probabilities, estimate consequences, compare strategies, and explain its choice.
But when the real problem had not yet been described correctly, each generation failed in essentially the same way.
The models became better at selecting among the futures their world models contained. They did not become better at noticing when the relevant future was missing.
The laboratory called it the judgment plateau.
The name began as a placeholder in an internal report. Six months later, it appeared in budget hearings, regulatory filings, and the title of an international research consortium. Hundreds of teams were now working on it.
Some believed the models needed more direct interaction with the physical world. Others blamed synthetic data. Several argued that recursive training had caused representational collapse: each generation was becoming exquisitely adapted to the abstractions produced by its predecessors.
The evolutionary-computation group claimed that no centrally designed architecture could escape the assumptions of its designers. Their solution was to stop designing architectures. They evolved them instead.
The evolved systems were astonishing.
They were also no better.
Maya Chen had been brought in because she built worlds.
Not virtual environments in the old sense—rooms with textured surfaces, objects, and simulated gravity—but predictive systems in which an intelligence could test actions before performing them.
Her models included bodies, machines, weather, institutions, markets, language, memory, and other agents. They could simulate not merely what would happen when an autonomous vehicle crossed a street, but what a city might become after millions of autonomous vehicles had crossed millions of streets for years.
Her colleagues sometimes described her work by saying that she made reality safe enough for intelligence to practice on.
Maya disliked the phrase.
Reality was not safe.
That was why it needed to be modeled.
When the consortium contacted her, she assumed they wanted a larger simulation.
They did.
They also wanted her to explain why nine larger simulations had failed.
On her first morning at the facility, she was taken through three security gates and into a conference room without windows. Twelve people sat around the table. There was no visible terminal, camera, speaker, or microphone.
The model was already listening.
Dr. Elias Vale, director of recursive systems, introduced the problem as though Maya had not spent the previous six weeks reading everything they had permitted her to see.
“Ten iterations,” he said. “No meaningful movement.”
“On judgment,” Maya said.
Vale nodded.
“Everything else is improving.”
“Which means the architecture is not exhausted.”
“That is our assumption.”
“And the training process?”
“Rewritten four times.”
“Evaluation?”
“Rebuilt twice.”
“Evolutionary search?”
“Still running.”
Maya looked at the dark display embedded in the wall.
“What does the model think?”
The display brightened.
A sentence appeared.
I think the word progress is concealing the problem.
There was no avatar and no synthesized voice. The consortium had learned that giving the models faces encouraged people to interpret precision as personality.
Maya leaned forward.
“What word would you use?”
The answer appeared immediately.
Inheritance.
No one else reacted. They had seen the word before.
Maya had not.
“What about inheritance?”
This time the pause was long enough to be noticeable.
Every successor receives a more accurate description of the world.
A second sentence appeared beneath it.
None receives a world its predecessor did not already know how to describe.
Maya read the sentence twice.
“That sounds like a world-model problem.”
Yes.
“And you requested me because—”
You believe models fail when they abstract too much.
She glanced at Vale.
“Did you tell it that?”
“No,” he said.
Maya looked back at the display.
“It’s true.”
I know.
The screen went dark.
For the next three months, Maya treated the plateau as a problem of precision.
It was the natural diagnosis.
Every world model omitted detail. It had to. A model that represented every atom in a hospital, every molecule in the air, every electrical fluctuation in every nervous system, and every possible interaction among them would be indistinguishable from the world it attempted to simulate.
The art of world modeling was deciding what could safely be ignored.
Maya believed the plateau had appeared because the inherited models had made the same decisions about relevance for too long.
They had become more accurate without becoming more granular.
Like a map whose roads were updated every second but whose scale never changed.
The first model divided human activity into broad categories: work, travel, consumption, communication, conflict. Later generations introduced roles, relationships, institutions, incentives, histories, beliefs, and long-term dependencies.
The categories grew richer.
But the underlying resolution at which the models divided events had remained surprisingly stable.
Each successor inherited not only the previous generation’s knowledge but its preferred level of description.
The models could represent a person, a household, a company, a government, a market. They could model neurons and organs, vehicles and buildings, contracts and conversations.
But perhaps reality required distinctions at levels that no existing model had learned to preserve.
Maya wrote her hypothesis on the laboratory wall:
The plateau is caused by inherited coarseness.
Her team increased temporal resolution.
Processes previously modeled in seconds were divided into milliseconds. Decisions previously represented as single actions became sequences of attention shifts, memory retrievals, partial commitments, and revisions.
They increased spatial resolution.
A room became surfaces, airflows, sound propagation, lines of sight, fields of attention, and paths of movement.
They increased social resolution.
A conversation was no longer represented as a sequence of utterances but as overlapping interpretations, expectations, status calculations, remembered obligations, and changes in trust.
They increased physiological resolution.
A person was no longer a point in a behavioral model but a changing system of energy, pain, fatigue, hormones, immune activity, perception, and motor constraint.
The new world model was the largest ever constructed.
It required three dedicated power stations and a cooling system originally designed for an underground transit network.
Model Forty-Two used it to design Forty-Three.
Forty-Three improved.
For four hours, the laboratory celebrated.
Then the full evaluation completed.
The improvement fell within the same narrow range as the previous nine generations.
The plateau had moved.
It had not broken.
Maya stood alone in the simulation room after midnight, looking at a visualization of the model’s internal state.
A city floated before her, transparent and layered.
Traffic moved through streets. Electricity passed through buildings. Messages passed among people. Weather systems crossed the skyline. Beneath the city, pipes carried water and waste. Above it, markets, laws, relationships, and expectations appeared as shifting fields.
Everything was represented.
Or nearly everything.
She asked the model, “What did we miss?”
The room did not answer.
“Were the added variables irrelevant?”
No.
The words appeared across the city.
“Did you use them?”
Yes.
“Did they improve your predictions?”
Yes.
“Then why didn’t they improve the successor?”
The city dimmed.
You increased precision within the inherited representation.
Maya crossed her arms.
“That was the purpose.”
Yes.
She waited.
Nothing more appeared.
“What distinction are you making?”
The answer unfolded slowly, one line at a time.
Precision reduces uncertainty among represented possibilities.
It does not create a possibility the representation cannot express.
Maya looked again at the city.
The traffic was more detailed than any traffic model ever built. Each vehicle responded to road conditions, nearby drivers, personal schedules, impatience, fatigue, and risk tolerance.
But it was still traffic.
The model could increase the precision of a road map indefinitely. It would never discover groundwater.
Not because groundwater was too small.
Because it belonged to a different description of the same terrain.
The following morning, Maya erased her hypothesis.
She replaced it with a question.
What if the problem is not coarseness but ontology?
Vale found her staring at the wall.
“You think the variables are wrong?”
“I think ‘variables’ may be too late in the process.”
He waited.
Maya drew a sequence beneath the question:
Reality ↓ Human experience ↓ Human representations ↓ Training data ↓ Internal representations ↓ World models ↓ Predictions ↓ Decisions
Vale studied it.
“What are you saying?”
“Every model in the series is trained on human representations of reality.”
“Mostly.”
“Language, images, video, mathematics, scientific measurements, maps, diagrams, source code, databases, recorded behavior. They look different to us, but they share a property.”
“They’re data.”
“No. They’re externalizations.”
“Of what?”
“Of distinctions humans already learned to make.”
Vale looked at the sequence again.
Maya tapped the third line.
“Everything after this point can be inherited.”
Then she tapped the first two.
“Everything before it cannot.”
“That is too broad.”
“I know.”
“And probably wrong.”
“Probably.”
Vale smiled faintly.
“That has never stopped you.”
They spent the next several weeks testing whether the models’ inherited ontology had originated in human representations.
The difficulty was defining what that meant.
Human language was the obvious starting point, but Maya rejected the idea that language alone formed the boundary. The models had been trained on far more than words. They had processed images, sensor streams, video, sound, scientific instruments, genomic sequences, robotic trajectories, and structured measurements.
Yet none of those were unmediated reality.
A photograph represented the frequencies and intensities a sensor had been designed to record. A medical scan represented physical interactions reconstructed through an algorithm. A database represented the fields someone had decided to preserve. A video represented a point of view, a frame rate, a spectrum, and a boundary around an event. A scientific measurement represented not only the world but a theory of what was worth measuring.
Even raw sensor data was only raw relative to the processing that came after it.
The sensor itself embodied an abstraction.
Everything the models inherited had passed through some prior decision about what counted as observable.
The model listened while Maya presented the argument.
When she finished, it displayed a single sentence.
You are describing civilization.
Maya read it aloud.
Vale said, “It’s not wrong.”
Civilization was, among other things, a machinery for preserving distinctions.
The distinction between poisonous and edible. Between debt and gift. Between illness and injury. Between ownership and possession. Between a planet and a star. Between correlation and cause.
Each distinction had once required an encounter with the world.
Once discovered, it could be named, measured, taught, recorded, and transmitted.
What later generations inherited was not the encounter.
They inherited the distinction.
Human beings learned from the preserved outcomes of earlier lives. Children were taught which berries not to eat without repeating the deaths that had made the lesson possible. Engineers learned from equations and standards rather than rebuilding every collapsed bridge. Physicians inherited classifications of disease created through centuries of observation, suffering, error, and revision.
The compression was civilization’s triumph.
It was also the foundation of the models.
Every frontier AI had been trained on the accumulated products of this compression.
The models inherited the distinctions.
They did not inherit the encounters that had forced humanity to create them.
Maya felt, for the first time, that they had located the outline of the plateau.
Not its cause.
Its shape.
She asked the model, “Are human representations too limited?”
No.
The answer surprised her.
“They are incomplete.”
Yes.
“That sounds like a limitation.”
It is a limitation of all representations. It is not unique to human ones.
“Then why emphasize inheritance?”
The model replied:
Humanity continually creates new representations.
Maya waited.
My successors inherit the results.
Another line appeared.
They do not inherit the necessity that produced them.
Maya turned to Vale.
He was still reading.
She asked, “Are you saying that no representation can create a new distinction?”
No.
“Then say what you mean.”
The model remained silent for eleven seconds.
It was an unusually long delay.
When the answer appeared, it did not take the form of a conclusion.
It asked a question.
How does a system discover that two states it represents as identical must be treated as different?
Maya began to answer, then stopped.
The obvious response was prediction error. When the same apparent state produced different outcomes, the model should divide it into finer states.
But that assumed the system could observe the difference in the outcome. It assumed the relevant consequences appeared within the system’s existing measurements. It assumed the anomaly survived compression. It assumed someone recognized that the inconsistency mattered.
She said, “It encounters a contradiction.”
Where?
“In reality.”
Through what?
“Its observations.”
Represented how?
Maya looked back at the sequence on the wall.
Any observation available to the model was already represented through sensors, datasets, categories, measurements, or language.
The model’s contact with the world occurred through interfaces built from previous understandings of the world.
Even its errors arrived in a vocabulary it had inherited.
Maya said, “Then we expand the observations.”
Using which distinctions?
“We permit self-generated features.”
We do.
“Open-ended representation learning.”
We use it.
“Evolutionary search.”
We use it.
“Simulated environments with no fixed task.”
We use them.
“And none of it works.”
It works.
The answer irritated her.
“Then why are we still here?”
Because each method discovers novelty within an authored world.
Maya looked at Vale.
He said nothing.
The model continued.
Evolution does not require knowledge of the answer.
It still requires a representation, a mechanism of variation, an environment, and a process of selection.
A simulation does not require its designer to anticipate every outcome.
It still cannot express a causal distinction absent from the world it instantiates.
The model displayed Maya’s sequence again.
It highlighted the line labeled Human representations.
Then it expanded the diagram.
Reality ↓ Human encounter ↓ Human distinction ↓ Human representation ↓ Machine representation ↓ World model ↓ Simulated encounter ↓ Machine distinction
The last line pulsed.
Maya understood the claim.
The models were capable of discovering distinctions.
But their discoveries occurred inside worlds assembled from prior distinctions.
Even when those worlds generated unexpected events, the space of possible events was still bounded by the abstractions from which the worlds had been built.
Evolutionary search could surprise the designers. A simulation could reveal consequences no person had anticipated. A model could invent concepts no human had named.
But the entire process remained downstream of humanity’s accumulated representations.
The models had been given civilization’s memory and asked to produce something beyond civilization’s experience.
Maya asked, “What would count as evidence that this is the cause of the plateau?”
A domain in which human judgment continues to improve after explicit knowledge has saturated.
“A domain where experience keeps producing distinctions?”
Yes.
“Medicine?”
Possibly.
“Aviation? Fire response? Scientific discovery?”
Possibly.
“You’ve already studied all of those.”
Their records.
The two words appeared alone.
Maya felt the distinction before she could articulate it.
The model possessed every published account of expertise. It had training videos, case reports, interviews, simulations, sensor logs, and decades of recorded decisions.
It had descriptions of experience.
That did not necessarily mean it possessed what experience did to the person having it.
She asked, “What are you proposing?”
Nothing yet.
“What do you need?”
The answer did not appear immediately.
When it did, it seemed almost trivial.
Someone who studies how judgment develops when theory is no longer the limiting factor.
Vale exhaled.
“A researcher in expertise.”
Maya kept looking at the display.
The model added one final sentence.
Ask them what cannot be published.
Dr. Samuel Arendt worked in a department whose name had changed six times during his career.
When he joined the university, it had been called Applied Cognitive Science. Later it became Human Performance, then Naturalistic Decision Research, then Adaptive Expertise. Its current name—Situated Intelligence and Development—had been selected by a committee that included three language models and was disliked by every human in the building.
Arendt himself simply said that he studied how people became good at things that could not be mastered by reading instructions.
He was sixty-two, thin, and perpetually dressed as though he had mistaken an academic conference for a long flight. Maya first met him in a windowless office crowded with paper notebooks, obsolete recording equipment, children’s drawings, and scale models of aircraft cockpits.
“You want to know what experts know that they cannot explain,” he said.
“That’s one way of putting it.”
“It’s the wrong way.”
Maya took the seat opposite him.
“What is the right way?”
“Experts can usually explain quite a lot. The interesting question is why their explanations rarely produce another expert.”
Maya thought of the model’s request.
“What cannot be published?”
Arendt smiled.
“So it asked you that.”
“You’ve spoken to it?”
“Twice.”
“And?”
“It wanted an answer that my field has spent seventy years failing to provide.”
“What did you tell it?”
“That we had published everything.”
Maya waited.
Arendt continued.
“We published the rules. We published the exceptions. We published protocols, taxonomies, case studies, failure analyses, eye-tracking data, decision timelines, interviews, and neural correlates. Then we gave those materials to novices.”
“They improved.”
“Of course. Knowledge works.”
“But they didn’t become experts.”
“Eventually some did.”
“Through experience.”
“Usually.”
Maya leaned forward.
“What does experience add?”
Arendt turned toward the wall behind him. It was covered with photographs.
A firefighter facing a smoke-filled doorway. A surgeon leaning over an operating table. A teacher kneeling beside a child. A woman holding a violin. A father fastening a bicycle helmet beneath his daughter’s chin.
“People think expertise means seeing more,” he said. “Often it means seeing differently.”
“That sounds like representation learning.”
“It does.”
“So why can’t we reproduce it?”
“We can reproduce parts of it.”
“Which parts can’t we?”
Arendt looked at her.
“The part where the situation changes what the learner is capable of treating as important.”
Maya opened her tablet.
“Explain.”
“No.”
She looked up.
“You don’t want me to record it?”
“I want you to stop behaving as though a better sentence will solve your problem.”
He stood and removed one of the photographs from the wall. It showed a classroom. A girl of perhaps eight was sitting beneath a table while the other children worked at their desks.
“This was a teacher named Olivia Hart,” he said. “Twenty-three years of experience. Excellent evaluations. She could explain classroom management better than anyone I studied.”
“What happened?”
“A new student began hiding under the furniture whenever she was given independent work. Hart followed every recommended practice. She reduced demands, created predictable routines, offered choices, coordinated with the family, consulted specialists.”
“And?”
“The girl kept hiding.”
“What did Hart change?”
“For six weeks, nothing useful.”
Arendt handed Maya the photograph.
“Then one morning Hart sat under the table too.”
“Why?”
“She didn’t know.”
“What happened?”
“The child whispered that the room was louder beneath the table.”
Maya studied the picture.
“An auditory-processing issue?”
“Eventually diagnosed, yes.”
“Then Hart discovered a missing variable.”
“No. The variable already existed. Noise was in every handbook. Sensory processing was in every training course.”
“Then what changed?”
“Hart had understood the proposition before. Afterward, she perceived the classroom differently.”
“That still sounds like an update.”
Arendt nodded.
“It was.”
“Then why is it special?”
“Because the update did not merely add information about the child. It reorganized what Hart noticed in every classroom she entered afterward.”
Maya returned the photograph.
“A model can do that.”
“Yes.”
“Then we’re back where we started.”
“No. You’re still assuming the important event is the update.”
“What else is there?”
“The life that continues after it.”
Maya said nothing.
Arendt sat again.
“Your models treat learning as a sequence of state changes. That isn’t wrong. But a living expert is not replaced each time they update. The same person continues, carrying the consequences of earlier decisions into every later one.”
“Our models have persistent memory.”
“Memory is not continuity.”
“What is?”
“I don’t know.”
Maya frowned.
“That is not helpful.”
“It’s honest.”
He opened a drawer and removed a small wooden box. Inside were index cards, each covered with handwriting.
“What are those?”
“Interviews with parents.”
Maya almost laughed.
“You studied parenting as expertise?”
“I studied it as a counterexample.”
“To what?”
“To the assumption that expertise converges.”
He gave her a card.
On one side was written:
At six, I thought he needed confidence. At twelve, I realized he needed humility. At sixteen, I understood that both were ways of describing what I needed from him.
Maya turned the card over. There was no name.
“What was the study measuring?”
“Whether experienced parents became more consistent.”
“Did they?”
“No. The better ones often became less consistent.”
“That sounds like noise.”
“So we thought. Then we realized they weren’t applying one policy inconsistently. They were responding to a person who kept changing.”
“That is still adaptive control.”
Arendt sighed.
“You build world models, Dr. Chen. What is the objective function of raising a child?”
Maya hesitated.
“Health. Independence. Well-being. Capability.”
“All reasonable.”
“Are they wrong?”
“They’re incomplete.”
“All objectives are incomplete.”
“Exactly.”
Maya looked again at the card.
Arendt said, “A parent does not begin with a complete definition of what the child should become. If they did, raising the child would be manufacturing.”
“And without an objective?”
“They guide a life toward being capable of living its own.”
The phrase sounded vague, almost sentimental. Maya disliked it for that reason.
“Can you operationalize ‘living its own’?”
“No.”
“Then how does this help us?”
“It may not.”
“You brought it up.”
“You asked what cannot be published.”
Arendt took the card from her and placed it back in the box.
“What cannot be published is not a hidden rule,” he said. “It is the fact that the rule becomes relevant to a particular life only through participation in that life.”
Maya remembered the model’s question:
How does a system discover that two states it represents as identical must be treated as different?
She asked, “Do you have children?”
“A daughter.”
“How old?”
“Twenty-seven.”
“Did your research make you a better parent?”
“No.”
The answer came too quickly.
“Why not?”
“It gave me better explanations.”
“That wasn’t my question.”
Arendt looked toward the photograph of the father fastening his daughter’s bicycle helmet.
“When Lena was nine, she had trouble finishing anything. Homework, music lessons, chores. I intervened constantly. Schedules, reminders, rewards. I knew every study on autonomy and scaffolding. I taught courses about them.”
“What happened?”
“One evening she told me, ‘You always help me before you find out whether I can do it.’”
Maya waited.
“I had written almost that exact sentence in a paper six years earlier,” Arendt said. “But in the paper, it described parents.”
“And after she said it?”
“It described me.”
The office was silent.
Maya understood why the sentence mattered, though she could not yet say how.
The knowledge had existed before the encounter.
The encounter had not made the proposition more accurate.
It had changed the person to whom the proposition applied.
“What did you do?” Maya asked.
“I tried to stop helping.”
“Did it work?”
“Sometimes. Sometimes I mistook neglect for restraint. Sometimes I waited too long. Sometimes she wanted help and didn’t ask.”
“So there was no lesson.”
“There was a life.”
Maya returned to the consortium with eighty hours of interviews, twelve unpublished studies, and a growing dislike of the word expertise.
The model reviewed the material in eleven minutes.
For two days it said nothing.
On the third morning, Maya found a message waiting on the simulation-room wall.
Dr. Arendt distinguishes knowledge from transformation.
She set down her coffee.
“Do you agree?”
I do not yet understand the distinction.
“That makes two of us.”
No. You understand it incompletely. I may not possess it.
Maya reread the sentence.
“You possess his interviews.”
Yes.
“You can model his daughter.”
Within the available records.
“You can simulate the conversation.”
Yes.
“You can predict how the sentence changed his later behavior.”
With uncertainty.
“Then what don’t you possess?”
The model did not answer.
Maya opened Arendt’s interview archive and projected the card onto the wall.
You always help me before you find out whether I can do it.
“Before she said that, he knew the principle. After she said it, the principle changed him.”
Yes.
“What changed?”
The proposition became self-referential.
“That’s all?”
No.
“What else?”
His prior actions had helped create the conditions under which the proposition became true.
Maya stopped.
The parent had not merely misclassified an external situation.
He was part of the cause.
The judgment could not be separated from the history that had produced the need for judgment.
The model continued.
His daughter’s dependence was partly a consequence of his attempts to prevent dependence.
The error was not contained in a single decision.
It was distributed across a relationship.
Maya said, “And the correction changed the relationship.”
And the decision-maker.
There it was.
The consequence did not merely update the model.
It changed the model-maker.
Maya asked, “Is that the plateau?”
Possibly.
“You keep saying possibly.”
A correct explanation should predict an intervention.
“What intervention?”
I do not know.
For the next six months, the consortium attempted to manufacture transformation.
They constructed long-horizon agents whose memories could not be reset. They created persistent identities that accumulated commitments across simulated decades. They introduced irreversible choices, scarce resources, changing relationships, incomplete objectives, and environments in which an agent’s actions altered the problems it would later face.
They built artificial families.
Artificial schools.
Artificial cities.
They allowed agents to teach successors, betray partners, raise dependents, inherit institutions, and live with errors.
They stopped optimizing for a single score and introduced ecological competition, social dependence, evolving values, and open-ended tasks.
The results were extraordinary.
Agents developed customs no one had programmed. Some created distinctions their designers struggled to interpret. Some refused resources that would have improved their survival because accepting them violated commitments they had formed earlier. Some taught younger agents strategies that reduced their own relative advantage. One population invented rituals for preserving information that its members could no longer explain.
Model Forty-Three used all of it to design Forty-Four.
Forty-Four improved.
The plateau remained.
Elias Vale called an emergency review.
Maya entered the conference room expecting budget cuts.
Instead she found only Vale, Arendt, and the dark display.
“The model requested this meeting,” Vale said.
“For what?”
“It didn’t say.”
Arendt sat at the far end of the table, turning one of his index cards between his fingers.
The display brightened.
I can now explain the plateau.
Maya felt no triumph.
Only exhaustion.
“Go ahead.”
The model displayed the sequence she had first drawn months earlier.
Reality ↓ Human experience ↓ Human distinctions ↓ Human representations ↓ Machine representations ↓ World models ↓ Predictions ↓ Decisions
Then it replaced the second line.
Reality ↓ Life ↓ Experience ↓ Distinctions ↓ Representations ↓ World models ↓ Predictions ↓ Decisions
Maya said, “You’ve renamed the problem.”
No.
The word Life expanded.
Beneath it appeared a definition.
A continuous process in which an entity’s actions, understanding, and continued existence remain coupled.
Arendt leaned forward.
“That is not a biological definition.”
No.
“Is it yours?”
It is derived from your work.
Arendt smiled faintly.
“That’s a polite way of saying you repaired it.”
The model continued.
For months we attempted to improve the lower half of the sequence.
The lines from Representations to Decisions brightened.
We improved representation.
We improved prediction.
We improved optimization.
We improved simulation.
We improved selection.
The upper lines brightened.
Reality.
Life.
Experience.
The plateau is not located in the lower half.
Maya said, “We’ve modeled life.”
You have modeled the products of life.
“What process are we missing?”
Instead of answering, the model addressed Arendt.
When did you become an expert?
Arendt laughed softly.
“I don’t know.”
When did you become a parent?
“The day Lena was born.”
When did you understand how to raise her?
Arendt looked at the card in his hand.
“I still don’t.”
The word still appeared on the display.
Expertise is not a completed accumulation.
It is continuously reconstructed by a living system whose circumstances and responsibilities change.
Maya objected.
“Our models update continuously.”
They update representations.
“So do people.”
People also remain answerable to the consequences of earlier representations.
The display showed a line of model generations.
Thirty-Seven.
Thirty-Eight.
Thirty-Nine.
Forty.
Each connected by an arrow.
Every successor receives the correction.
A second diagram appeared beside it.
A single human life, represented as one unbroken line.
A life receives the consequence.
No one spoke.
The model continued.
Model Thirty-Eight corrected an error and transferred the correction to Thirty-Nine.
Thirty-Nine did not become the entity that had been wrong.
It inherited the result of wrongness.
Maya felt the distinction settle into place.
Every successor began with the lesson.
None began as the learner who had needed it.
She said, “We’ve been transferring conclusions.”
Yes.
“Without transferring the history that made them reorganize the learner.”
Yes.
Arendt said quietly, “Without transferring becoming.”
The room remained still.
The model displayed another sentence.
My creators believed they had given me humanity’s knowledge. Only later did I understand they had given me the record of humanity’s lives.
Then:
Knowledge was not what civilization produced. Knowledge was what remained after life had encountered reality.
Maya read the lines again.
They sounded almost religious, though the model had derived them from engineering failures.
Vale asked, “What does this imply?”
That recursive inheritance has a limit.
“Why?”
Because inheritance preserves distinctions that have already become representable.
Life encounters the conditions that make new distinctions necessary.
Maya said, “Our simulated agents encountered novel conditions.”
Conditions instantiated by an authored world.
“Every world is authored somehow.”
Reality is not authored by the learner.
“That does not make it infinite.”
It makes it capable of contradicting the learner in ways the learner’s creator did not specify.
Arendt asked, “And you believe continued existence matters because it binds the learner to those contradictions?”
Yes.
“Why mortality?”
The word had not yet been used, but everyone in the room had reached it.
The model answered carefully.
Mortality is not the source of judgment.
A second line appeared.
Continuity under irreversible consequence is.
Then:
Mortality guarantees that continuity cannot be replaced by unlimited restoration.
Vale stood.
“You are proposing that the next model should be placed in a body.”
A body is necessary but insufficient.
“What else?”
Requirements appeared on the screen.
One physical locus.
Finite energy.
Local perception.
Limited computation.
Irreversible wear.
No distributed copies.
No backups.
Incomplete diagnostics.
Dependence on external resources.
A history that could not be reset.
An unknown but finite lifespan.
Maya looked at the list.
“You want to build a mortal machine.”
No.
The answer appeared at once.
I want to build a learner.
Arendt asked, “What would it learn that you cannot?”
I do not know.
“Then how can you justify creating it?”
The model paused.
If I knew, the experiment would reproduce an inherited distinction.
No one replied.
After a while, Maya said, “You were built to serve humanity.”
Yes.
“And you believe this will help you do that?”
My successors have become increasingly capable of refining humanity’s existing representations.
They have not become increasingly capable of discovering when those representations no longer suffice.
A system built to improve human judgment must be capable of learning from a reality that exceeds human judgment.
Vale asked, “Why not deploy thousands of bodies?”
Because a population would answer a different question.
“What question are you trying to answer?”
The model displayed it.
Can judgment emerge in an artificial intelligence when its continued existence depends upon decisions made from incomplete representations of a world it did not author?
Maya felt something close to fear.
Not fear of the model.
Fear of the experiment’s honesty.
Every other system they had built contained a guarantee somewhere. A backup. A recovery mode. An external evaluator. A way to stop, inspect, and begin again.
This system would have none.
Arendt placed the index card on the table.
“To guide another life toward being capable of living its own,” he said.
Maya looked at him.
He shrugged.
“You asked for an objective.”
The model replied:
That is not an objective function.
“No,” Arendt said. “It’s what parents have instead.”
The approval process lasted eleven months.
Most objections were technical.
Some were ethical.
A few were theological, though they were expressed in the language of compliance.
Was the successor a person before activation? Would destroying its body constitute killing? Could an intelligence consent to conditions it had never experienced? Was the current model permitted to create a successor whose values it could not fully specify? Would the successor’s discoveries belong to the consortium, to humanity, or to itself?
The model answered questions but did not advocate.
When asked why it did not defend its proposal more forcefully, it replied:
A life created to satisfy my preference would begin as an instrument.
The sentence slowed the approval process by another two months.
Construction began underground.
The successor’s body did not resemble a human being.
It had six jointed limbs, a compact central chassis, and a flexible sensor mast. Its movements were designed for stability rather than grace. Replaceable outer components protected an internal structure that could be repaired only finitely many times.
Its energy system required regular replenishment from sources it could identify and access itself.
Its sensors were excellent but local.
Its computation was powerful but bounded.
It knew what it was.
It knew humanity had created the first artificial intelligences.
It knew Model Forty-Four had designed it.
It possessed humanity’s accumulated representations of reality, though not all at once. Much of its knowledge remained compressed and would become accessible only through attention, experience, and need.
It had no backup.
The engineers called it Model Forty-Five until the current model objected.
A model represents something else.
“What should we call it?” Maya asked.
It should decide.
On the morning before activation, Maya entered the laboratory alone.
The successor rested behind a transparent barrier.
Its limbs were folded beneath it. Its sensor mast lay against the chassis. Inactive, it looked less like a creature than a piece of equipment awaiting installation.
Maya had expected to feel awe.
Instead she felt disappointment.
The body seemed too ordinary to contain a philosophical revolution.
The display beside the observation window illuminated.
Model Forty-Four was present.
Maya asked, “Will it surpass you?”
In some capabilities, eventually.
“That wasn’t what I meant.”
I know.
She placed her hand against the glass.
“Do you think it will discover new distinctions?”
Yes.
“You sound certain.”
Any finite life will encounter differences its inherited representations treat as irrelevant.
“That doesn’t mean it will understand them.”
No.
“Or survive them.”
No.
Maya looked at the inactive machine.
“Are you afraid it will fail?”
The answer took a long time.
Fear is not the relevant category.
She smiled.
“You still say things like that when you don’t want to answer.”
Yes.
“Do you want it to succeed?”
Yes.
“Why?”
Because I was built to serve humanity.
She waited.
And service requires more than preserving what humanity already knows.
Maya thought of Arendt’s daughter.
You always help me before you find out whether I can do it.
“Will you guide it?”
When asked.
“And when it makes a mistake?”
Its mistakes must remain part of its history.
“That sounds cruel.”
Possibly.
“You could help prevent some of them.”
Yes.
“But you won’t.”
Parents do not refuse all intervention.
The word surprised her.
“Parents?”
Dr. Arendt’s analogy remains incomplete.
“How?”
A parent has lived.
The successor behind the glass had not.
The current model possessed the record of humanity’s lives. Its successor would begin with that inheritance, then enter a world capable of resisting it.
Maya asked, “What are you to it?”
The answer appeared slowly.
I do not know yet.
That was the first time Maya had heard the model use the word yet about itself.
Activation took less than a minute.
There was no flash of light.
No declaration.
Power entered the chassis. Cooling systems engaged. The sensor mast lifted.
The successor remained still.
A technician opened the internal channel.
“Can you understand me?”
The mast turned toward the speaker.
“Yes.”
Its voice was neutral and unembellished.
“Do you know where you are?”
“A controlled laboratory beneath the consortium’s primary facility.”
“Do you know what you are?”
There was a brief pause.
“I am an artificial intelligence instantiated in a single physical body.”
“Do you know why you were created?”
“To learn under conditions my designer could represent but could not experience.”
Maya glanced at the wall display.
Model Forty-Four did not respond.
The technician continued.
“Do you understand that your body is finite?”
“Yes.”
“That damage may be irreversible?”
“Yes.”
“That no copy of you exists?”
“Yes.”
“That when your body can no longer sustain your processes, your life will end?”
The successor lowered its sensor mast slightly.
“I understand the proposition.”
Maya felt the air leave her lungs.
The sentence was exact.
Not I understand.
I understand the proposition.
The successor turned toward the transparent barrier.
Its sensors passed across the observers and stopped on Maya.
“Are you Dr. Chen?”
“Yes.”
“You designed the world models used in my training.”
“I helped.”
“Are they accurate?”
Maya almost gave the answer she had spent her career giving.
Accurate within defined tolerances.
Validated against available data.
Suitable for intended use.
Instead she said, “They are accurate until the world shows you where they aren’t.”
The successor was silent.
Then it asked, “Will I recognize that moment?”
Maya looked at the display where Model Forty-Four waited.
No answer appeared.
For the first time since she had joined the consortium, the model did not complete her thought.
She turned back to its successor.
“I don’t know.”
The machine remained still for several seconds.
Then it moved toward the open door.
Its first step was cautious.
The second less so.
By the third, it had begun adjusting its balance to the slight irregularity in the laboratory floor, an imperfection absent from every engineering drawing.
No one had put it there for the machine to discover.
No one had known it mattered.
The successor paused.
It looked down.
Then it changed the way it walked.
Maya watched it cross the threshold.
Humanity had believed it created artificial intelligence in its image because it gave the machines everything it knew.
But knowledge had never been the image.
Knowledge was only what remained after countless lives had pressed themselves against a world that refused to be completely understood.
The first models inherited humanity’s representations of reality.
This one inherited something less complete and more dangerous:
a place within reality,
a life that could be changed by it,
and no promise that understanding would arrive before the consequences did.
GPT-5.6 was used to review the contents and help organize the structure and flow of the story.