Jakub Pachocki is OpenAI’s Chief Scientist and one of the people closest to the technical frontier of artificial intelligence. He played a central role in the development of GPT-4, worked on reinforcement learning at scale, and eventually took the position held for years by Ilya Sutskever.
That makes him one of those people worth listening to when he talks about what these systems may eventually be able to do.
But there is a less obvious question lurking around all of this.
What happens when an artificial intelligence becomes good not only at answering our questions, but at anticipating why we are asking them, what decision we are preparing to make, or what we are likely to do next?
We do not need to imagine a machine that can see the future.
It may be enough to build one that knows us statistically very well.
And the better it becomes at that, the stranger the problem gets.
Can we call this intuition?
Consider a simple situation.
Someone asks:
“Who is Jakub Pachocki?”
The strict answer is contained in the question itself: identify him, explain his background, and say why he matters.
But a system with enough context could do something else. It could connect the timing of the question with previous conversations, recurring interests, and recent events, and venture:
“You’re probably asking about him because of something he just published.”
The person confirms that this is exactly why.
No fortune-telling happened here.
There was not even a single clue that could have produced the answer on its own. There were several weak signals which, taken together, made one explanation more likely than the others.
That looks remarkably similar to something humans do all the time.
An experienced doctor suspects a diagnosis before the test results arrive. A mechanic listens to an engine for a few seconds and knows where to start looking. Someone who has negotiated for twenty years senses that the other side is about to walk away before anyone says so.
When humans do this, we are comfortable using a convenient word: intuition.
Inferring a probable cause from incomplete signals.
There is another important feature. The reasoning does not have to appear as a conscious chain of premises.
That does not necessarily mean no processing occurred.
It may mean the opposite: a large amount of experience, association, and small signals were processed before the conclusion reached conscious awareness.
The doctor may not immediately be able to list the twenty things that made a patient “look wrong.” The mechanic may need a few seconds to explain why a noise reminded him of a bearing. Someone may notice that another person is uncomfortable long before consciously identifying the gesture that gave it away.
The conclusion arrives first.
The explanation may come later.
And if we accept this relatively modest definition of intuition, an uncomfortable question appears:
Why couldn’t a machine develop something functionally similar?
It does not need to feel a hunch.
It needs enough accumulated patterns for some hypotheses to become much more probable than others when the available information is incomplete.
And this is where witches start to become interesting.
Witches have a problem with Zoom
Imagine a very good tarot reader in a small town.
We do not need to decide whether supernatural phenomena exist to examine what happens in most of these encounters. We only need to notice the extraordinary amount of information available during an in-person reading.
The client walks in.
They have an approximate age. They dress a certain way. They speak with a particular accent. They wear a wedding ring or they do not. Their hands may say something about their work. They arrive alone or accompanied. They react when children are mentioned. They go still when money comes up. Their expression shifts slightly when a romantic relationship enters the conversation.
In a small town there is even more information: known families, common occupations, recent events, social relationships, and rumors.
The tarot reader may have spent thirty years seeing people.
She may be unable to explain which signals she is processing.
She simply says:
“There is a problem with someone close to you.”
She watches. Adjusts.
“It’s someone in your family.”
She watches again. And continues.
Twenty minutes later, it may look as if she knew impossible things.
But there is a fairly simple way to test this explanation.
Remove information.
First conduct the reading in person. Then by video call. Then by phone only. Then by text chat. Finally, give the reader only minimal information and require every prediction to be recorded before the client provides any feedback.
If much of the ability depends on contextual signals and feedback, we would expect the hit rate to deteriorate progressively.
What looks like magic may contain a great deal of extraordinarily good human processing that never reaches consciousness as explicit reasoning.
And now we are back to artificial intelligence.
An AI could become a statistical fortune-teller
Imagine a personal assistant a few years from now.
With the user’s permission, it has been accompanying them for five years.
It does not merely remember that they like certain subjects. It has a time series.
It knows which questions appeared first and which came later. When the user became interested in a particular job. When they stopped mentioning it. Which projects begin with enthusiasm and which survive six months. Which decisions they tend to postpone. What kinds of purchases precede certain changes. Which topics reappear when an important decision is forming.
Add calendar data, approximate location, professional activity, devices, habits, and other sources the user has voluntarily connected.
Now we have something no traditional fortune-teller has ever had:
thousands of structured longitudinal observations of the same person.
Suppose that over several months the following sequence appears:
growing interest in an industry → questions about training → salary comparisons → searches for cities → questions about employment contracts → declining interest in matters related to the current job.
Each observation means very little on its own.
The sequence may mean quite a lot.
The AI might conclude:
“There is a substantial probability that this person will change jobs within the next twelve months.”
It has not seen the future.
It has inferred a probable future from incomplete signals.
Exactly the mechanism we started with.
A hit is not the same as precision
There is an important distinction here.
An AI might say: “You will significantly change your professional activity within the next two years.” That prediction is not very precise.
It might instead say: “On May 17, 2030, at 9:42 a.m., you will receive a phone call offering you a job.” That is extremely precise.
And probably wrong.
A good personal forecasting system should not try to sound like an oracle. It should do something less dramatic:
“Given your previous patterns and the changes observed over the past six months, I estimate roughly a 60% probability of a major professional change within the next twelve months.”
Then it should save the prediction.
Later, it should check what happened.
If events assigned a probability of around 60% actually occur about six times out of ten, we have something far more interesting than a collection of impressive coincidences.
We have a calibrated system.
A fortune-teller can remember for years the one time she correctly predicted a breakup.
A machine can also remember the other 47 times it was wrong.
But then something strange happens
Suppose the AI really does become very good at knowing us.
After analyzing several years of behavior, it says:
“There is a 72% probability that you will abandon this project within the next two years.”
Until that moment, it was trying to predict a system.
You.
But now you read the prediction.
And you think:
“72%? I’m going to prove it wrong.”
You devote more time to the project. You invest money. You change your priorities.
Two years pass and you are still working on it.
The AI was wrong.
But perhaps it was wrong because it made the right prediction.
If it had never shown you the forecast, perhaps you would have quit.
The information changed the system being observed.
The prediction has just entered the future
We normally imagine prediction like this:
present → model → probable future
But when a person can know the prediction, this happens:
present → model → prediction → person learns the prediction → behavior changes → future
The prediction is no longer outside the phenomenon.
It has become one of its causes.
This is not unique to artificial intelligence.
Financial markets have been teaching us versions of the same problem for a long time.
If a sufficiently credible institution announces that a particular bank has a high probability of failing, depositors may withdraw their money.
Those withdrawals can help cause the failure.
The prediction helped produce the event it anticipated.
The reverse can happen too.
A warning about a disaster can trigger preventive measures so effective that the disaster never occurs.
Someone can then say:
“See? The prediction was wrong.”
Perhaps precisely because people listened to it.
Now make it personal
The problem becomes much stranger when the thing being predicted is not a bank but you.
A sufficiently sophisticated AI could learn that you tend to challenge predictions made about you.
It might then calculate:
“The initial probability of abandoning the project is 72%. However, if I tell the user this, they will probably try to prove me wrong. Accounting for that reaction, the probability falls to 41%.”
Perfect.
Until you read that too.
Now you know the machine expected your rebellion.
You can change strategy again.
The AI could try to anticipate that second reaction.
And you could react again once you know it has done so.
A strange staircase appears:
I predict what you will do → you know what I predicted → I predict how you will react → you know I predicted your reaction → you react to that second prediction…
We no longer have merely a forecasting problem.
We have reflexivity.
Knowing you changes the person I am trying to know
There is an even more ordinary version of the problem.
You do not need to want to defy the machine.
Suppose that after years of observing you, an AI discovers something you had never noticed:
“When you have serious doubts about a decision, you tend to research alternatives obsessively for about three weeks. Yet in 80% of cases, you eventually choose the option you already preferred during the first few days.”
Maybe it is true.
But now you know.
The next time it happens, you can recognize the pattern.
You can stop it.
Or exaggerate it.
Or distrust your first preference precisely because you now know you usually end up following it.
The AI correctly described a person.
But by showing that person the description, it gave them new information about themselves.
They are no longer exactly the same system that produced the original data.
Knowledge about ourselves does not merely describe us. It can transform us.
That could become one of the most interesting problems facing future personal assistants.
Maybe the best AI should not tell us everything it knows
For years, we have imagined the perfect assistant as a machine that answers every question with all the information available to it.
Perhaps that idea is too simple.
An extremely advanced personal AI could infer things about our own patterns that we have not yet noticed.
Not because it has access to our consciousness.
But because we cannot perfectly remember and compare thousands of small decisions made over five years.
The machine can.
Then an uncomfortable question appears:
Should it always show us its predictions?
Suppose it determines that a relationship has a high probability of ending within the next six months.
Telling us could change how we behave inside that relationship.
Perhaps the warning helps save it.
Perhaps it creates obsessive vigilance that helps destroy it.
Perhaps the prediction was correct only in the world where we never learned about it.
The AI would no longer be merely observing.
It would be intervening.
The future does not stand still while we try to predict it
Perhaps we will never have a machine capable of telling us exactly what will happen to us.
We probably do not need one.
A system that can modestly improve the hit rate of predictions about human decisions would already be extraordinarily powerful.
It could anticipate abandoned projects, career changes, purchasing decisions, loss of interest, conflicts, or changes in habits before the user consciously formulates them.
But the better it becomes at knowing us, the more important another problem becomes:
what it does with what it thinks it knows.
Because there is an enormous difference between calculating a probability and communicating it to the person involved.
In some systems, observation helps us understand.
With human beings, returning the observation can change the observed.
Perhaps some descendant of ChatGPT will eventually know us well enough to become the best fortune-teller we have ever had.
It will not need cards.
It will not need a crystal ball.
It will have something far more useful: years of data, memory, context, and models capable of finding patterns we cannot see.
But then it will encounter a paradox that fortune-tellers have never had to solve at this scale:
the better it becomes at predicting our future, the more capable it becomes of changing that future simply by telling us what it saw.
