The past is all we have

Everything we know about the world comes from the past.

That might sound like a philosophical statement, but it has a very practical consequence for data, prediction and artificial intelligence:

Every piece of data describes something that has already happened.

A thermometer reading tells us what the temperature was when the measurement was taken. A photograph records light that reached a camera in the past. A bank transaction records an exchange that has already occurred. A sensor on a robot reports something it has already observed.

Even the words you are reading now describe information that existed before you read them.

There is no dataset containing observations of the future.

There cannot be.

The future has not happened yet.

Data is a record of the past

Imagine we measure the outside temperature every hour:

Time	Temperature
09:00	12°C
10:00	13°C
11:00	15°C
12:00	16°C

At 12:01, every value in this dataset belongs to the past.

Some observations are only one minute old. Others are several hours old. But none tells us directly what the temperature will be at 13:00.

We might reasonably predict that it will be 17°C.

But notice what has happened.

16°C at 12:00 is an observation.

17°C at 13:00 is a prediction.

Those are fundamentally different kinds of information.

One describes something we observed.

The other describes something we believe might happen.

Much of artificial intelligence exists in the gap between the two.

The world comes first

It is useful to distinguish between an event and the data describing that event.

Suppose a car passes a sensor at 10:42:15.

The sequence might look something like this:

EVENT → MEASUREMENT → DATA → STORAGE → PROCESSING

The car passes.

A sensor detects it.

The measurement is converted into data.

The data is transmitted somewhere.

A computer processes it.

Only then can another system use that information.

The event itself always comes first.

This gives us an important principle:

Data is not the world. Data is a record produced by observing the world.

That distinction will become increasingly important throughout this course.

Even "real-time" data comes from the past

We frequently describe systems as operating in real time.

A live traffic map shows real-time congestion.

A power-system operator monitors real-time electricity demand.

A financial terminal displays real-time prices.

A robot processes a real-time camera feed.

But "real time" does not mean that the system has somehow escaped the distinction between past and present.

There is always some delay.

Consider a camera connected to an artificial intelligence system.

Light reaches the camera sensor.

The sensor measures it.

The image is converted into digital information.

That information is transferred into memory.

The model processes the image.

The result is returned.

By the time the AI responds to what it has seen, the world has already moved on.

Usually this delay is tiny.

Sometimes it matters enormously.

For a system recommending a film, a delay of a few seconds may be irrelevant.

For a self-driving vehicle travelling towards a pedestrian, even a small delay can matter.

Data has an age

We can therefore think about data as having an age.

Some data might be:

milliseconds old,
seconds old,
hours old,
years old,
centuries old.

This matters because the usefulness of information depends partly on how quickly the system being observed changes.

Knowing someone's date of birth from ten years ago is still useful because it does not change.

Knowing the location of a moving vehicle from ten minutes ago may be almost useless.

Knowing yesterday's electricity demand can help us understand patterns, but it cannot tell us exactly how much electricity is being consumed now.

The faster a system changes, the more important the age of our observations becomes.

A dataset is a memory

Another useful way to think about data is as memory.

A dataset allows information about previous states of the world to persist.

Without memory, every moment would disappear as soon as it occurred.

Writing was an early technology for storing information about the past.

Books extended that memory.

Photography allowed us to record visual information.

Audio recording captured sound.

Scientific instruments recorded measurements.

Computers made it possible to store, retrieve and process enormous quantities of these observations.

The internet connected many of those stores of information.

Modern artificial intelligence can learn patterns across them.

In that sense, the extraordinary capabilities of modern AI depend upon something very simple:

humanity has become exceptionally good at recording the past.

But the past is incomplete

There is another problem.

We do not record everything.

Imagine trying to create a dataset describing everything happening in a city at this exact moment.

You might measure:

traffic,
temperature,
electricity consumption,
air pollution,
mobile-phone locations,
public-transport movements,
financial transactions,
hospital admissions,
water consumption,
conversations,
weather,
noise.

But even this enormous dataset would represent only a tiny fraction of everything happening.

What were people thinking?

Why did somebody choose one route rather than another?

Was somebody about to leave their house?

Was a machine beginning to fail?

Did someone change their mind?

The world contains vastly more information than we can observe or store.

So data is not simply the past.

It is a partial record of the past.

What we measure is a choice

This introduces another important idea.

Datasets do not simply appear.

Somebody decides what to measure.

Somebody chooses the sensors.

Somebody defines the categories.

Somebody decides how frequently observations are recorded.

Somebody decides what information is stored and what is discarded.

Consider a dataset about a school.

It might contain:

examination results,
attendance,
age,
class,
teacher,
postcode.

From those variables we might attempt to predict educational outcomes.

But perhaps the dataset does not contain:

whether a student was caring for a family member,
whether they had somewhere quiet to study,
whether they were hungry,
whether they enjoyed the subject,
whether they trusted their teacher,
whether something significant happened at home.

The model can only learn from the information that was recorded.

What is absent from a dataset can matter just as much as what is present.

This will become particularly important when we later examine bias and fairness in artificial intelligence.

Historical data contains history

There is an even deeper consequence.

If data records the past, then historical datasets inevitably contain the structures of the societies that produced them.

Imagine training an AI system using historical employment records.

Those records might contain information about:

who was hired,
who was promoted,
what people were paid,
who became managers,
who left the organisation.

The dataset records what happened.

But what happened historically was not necessarily fair.

If some groups had fewer opportunities in the past, a model may discover that pattern perfectly.

From the perspective of prediction, the model might be doing exactly what it was asked to do.

From the perspective of society, reproducing that pattern into the future might be precisely what we do not want.

This creates one of the fundamental questions of artificial intelligence:

If we train machines on the past, how do we prevent them from simply recreating the past?

We will return to this question throughout the course.

Learning from the past

Despite all these limitations, the past is extraordinarily useful.

Patterns repeat.

The sun rises every morning.

People tend to travel at particular times.

Electricity demand follows daily and seasonal patterns.

Machines often exhibit warning signs before they fail.

Languages contain recurring grammatical and semantic structures.

Human behaviour has regularities.

Images contain recurring visual structures.

These patterns allow us to learn relationships from previous observations and use them to reason about situations we have not yet observed.

This is the basic opportunity behind statistical inference and machine learning.

We observe:

what happened before

and attempt to learn:

what tends to happen under similar conditions.

That learned relationship can then be used to make predictions.

The fundamental problem of prediction

Suppose we have observations up to the present moment:

PAST → PAST → PAST → NOW

Everything to the left of NOW can, in principle, be observed.

Everything to the right cannot:

PAST → PAST → PAST → NOW → ? → ? → ?

That is the fundamental problem.

We have information about what has happened.

We need to make decisions about what has not happened yet.

Prediction attempts to bridge that gap.

We use information from the past, combined with our understanding of the present, to estimate possible futures.

Artificial intelligence can make that process enormously more sophisticated.

But it cannot eliminate the fundamental distinction.

A prediction about the future is not an observation of the future.

It is an inference.

From the past to the future

This gives us the first part of the framework we will use throughout this course:

PAST → DATA → MODEL → PREDICTION

The past produces observations.

Those observations become data.

Models learn relationships from that data.

Those relationships allow us to make predictions about things we have not yet observed.

But prediction is not the end of the story.

Predictions are useful because they help us make decisions.

And decisions change what happens next.

Our complete framework will therefore become:

PAST → DATA → MODEL → PREDICTION → DECISION → ACTION → FUTURE

And then something interesting happens.

The future arrives.

Once it happens, it is no longer the future.

It becomes the past.

It can be measured.

It becomes new data.

And the cycle begins again.

The central idea

Artificial intelligence is sometimes presented as though machines somehow know the future.

They do not.

They have access to records of the past and information arriving from the world now.

From those observations they learn patterns, construct representations and estimate what might happen next.

The distinction matters.

The past can be observed. The future must be inferred.

Understanding that difference is the starting point for understanding data, uncertainty, prediction and artificial intelligence.

In the next lesson, we will turn our attention to the narrow boundary between them:

the present — the moment in which observations become decisions, decisions become actions, and possible futures begin to become real.