The Numerator Problem

Why AI Can Predict But Cannot Yet Receive Reality

Yuval Noah Harari is right to say that AI’s first great habitat is bureaucracy.

In the 2026 Tanner Lecture on Human Values at Linacre College, University of Oxford, Harari argues that AI is not merely a tool in human hands but an agent entering the artificial systems humans have already built: law, finance, universities, churches, governments, corporations, militaries, and social media. His deepest point is not that AI will someday walk out of the laboratory as a robot. It is that AI already lives comfortably inside the most artificial environment humans ever made: bureaucracy. Bureaucracy is made of words, forms, rules, records, permissions, ledgers, claims, and trust. AI is native to that world.

That is a powerful observation.

But I think it needs one correction.

AI is not yet primarily an agent. AI is first a prediction machine.

The LLM predicts. Full stop. When we say it summarizes, writes, classifies, plans, drafts, ranks, or imitates, we are already describing actions built on top of prediction. Those actions are real, but they occur inside a symbolic action-space: a document, a chat window, a database field, an email thread, a workflow, a codebase, a CRM record, a calendar event. The agent is prediction with hands. The hands matter. But the prediction comes first.

This distinction matters because it tells us where the real bottleneck is.

The denominator is not the problem.

The numerator is.

In the Reality Equation, Reality is the quotient of the ratio:

R = A / E

Actual is the numerator. Expectation is the denominator. The real component of Expectation is produced by a prediction machine that resolves a cloud of possible outcomes into a single guess. The denominator receives the guess; the cloud belongs to the prediction machine. (John Rector)

That is where today’s AI is already extraordinary. The synthetic prediction machine is production-grade. It has been improving for years. It can predict the next word, the next sentence, the likely answer, the likely structure, the likely code, the likely argument, the likely image, the likely plan. It is not perfect, but the prediction side is no longer the weak side of the system.

The weak side is Actual.

Actual is not a guess. Actual is not a probability. Actual is not a prompt. Actual is not a retrieved passage. Actual is the given projection of the world onto the observable being predicted. It arrives from the Immutable Past. It does not negotiate. It does not hallucinate. It is exact.

This is where AI agents remain primitive.

A human being receives Actual continuously. Light, sound, gravity, temperature, pressure, hunger, pain, fatigue, timing, facial expression, social tension, consequence, memory, and danger arrive without prompting. The human organism is in constant unconscious relationship with Actual. We do not have to call an API to know the room is getting cold. We do not need retrieval augmentation to notice that someone’s tone changed. We do not need a database query to feel that the dog is nervous, the food is burning, the road is slick, the child is too quiet, or the meeting has gone badly.

The human numerator is dense, indexed, and live.

The AI numerator is not.

This is the hidden meaning of “pretrained.”

GPT is pretrained. Humans are live-trained.

A pretrained model is built from historical Actuals compressed into weights. It carries a vast statistical memory of what has happened, what has been said, what has been written, what has been photographed, what has been coded, what has been argued, and what has been named. That is powerful. But it is not the same as being continuously corrected by Reality.

When the model goes stale, we patch it. We add retrieval. We connect tools. We give it files. We add memory. We let it browse. We connect APIs. We show it screenshots. We stream sensor data. Each of these improves the numerator interface. But none of them yet equals the human relationship to Actual.

The AI does not receive Reality for free.

It receives managed fragments.

And even “episodic” is too generous a word for this. Human episodic memory is long and versioned. We carry previous versions of Reality forward. Yesterday is not erased when today arrives. Childhood is not overwritten by adulthood. A betrayal, a promise, an injury, a success, a silence, a look across a room — these are not merely stored data. They are layered into the living prediction machine. They continue to shape what Actual means when it arrives again.

AI does not have that kind of continuity.

When a model is updated, it becomes a new checkpoint. Engineers may preserve old checkpoints externally, but the model itself does not remember its previous world the way a person remembers a former self. It does not carry Reality forward as lived versioning. It has pretrained compression, present context, and external records. That is not the same as memory formed under continuous correction by Actual.

This is why current agents feel both impressive and brittle.

They can act inside simulated Reality.

They can draft the memo, update the CRM, move the file, create the ticket, summarize the meeting, write the campaign, generate the contract, and send the message. Those are actions. But they are actions inside worlds already reduced to symbols. They occur where Actual has already been thinned into records.

A CRM entry is not the customer.

A medical chart is not the patient.

A loan application is not the borrower.

A legal filing is not justice.

A calendar invite is not the meeting.

A dashboard is not the business.

A sensor reading is not the factory.

The agent can manipulate the representation. The hard question is whether it has enough access to Actual to know whether the representation remains true.

This is why bureaucracy is the first great habitat of AI agency. Bureaucracy has already converted Reality into language, records, credentials, forms, claims, workflows, and timestamps. It has already made the numerator thin enough for machines to handle. Harari sees this clearly: AI enters the world not by conquering the jungle but by inhabiting the bureaucratic systems through which modern civilization already routes trust.

But that is also the limitation.

AI is not becoming powerful because it has mastered Reality.

AI is becoming powerful where humans have already replaced Reality with records.

This gives us a better way to understand the current agent boom. The question is not, “Can the model think?” The question is not even, “Can the model act?” Of course it can act, if we define action as producing an output or changing a symbolic state.

The deeper question is:

Can the agent receive Actual with enough fidelity for Reality to correct it?

That is the numerator problem.

A real agent needs more than a strong prediction machine. It needs a live relationship between prediction and consequence. It needs reliable observables. It needs sensing. It needs verification. It needs versioned memory. It needs audit trails. It needs permission boundaries. It needs feedback loops. It needs to know the difference between “the database says the package shipped” and “the package actually arrived.” It needs to know the difference between “the customer clicked approve” and “the customer understood what they approved.” It needs to know the difference between “the chart says stable” and “the patient is declining.”

In other words, it needs a richer numerator.

This is the next frontier of AI.

Not bigger prediction alone.

Not more fluent language alone.

Not more confident agents alone.

The next frontier is Actual: how machines receive it, verify it, remember its versions, and allow it to correct their guesses.

For computer science students, this is the point worth teaching first. Pretraining is not experience. A context window is not a life. A checkpoint is not a childhood. Retrieval is not memory. A tool call is not agency. A database state is not Reality.

The LLM predicts.

The agent acts inside a constructed world.

The human being receives Actual continuously and remembers the sequence.

Until machines acquire something closer to that dense, indexed, versioned relationship with Actual, they will remain brilliant synthetic predictors operating inside managed simulations. They will do enormous work there. They will change bureaucracy. They will change finance. They will change law. They will change education. They will change medicine. They will change the symbolic machinery of civilization.

But human-like agency requires more than prediction with hands.

It requires Reality with fidelity.