5.14 What Happens When Tomorrow Doesn't Resemble Yesterday?

Machine learning is built on a strange bargain.

We take data from the past.

We search for patterns.

We build a model.

And then we use that model to make predictions about a future that has not happened yet.

In shorthand:

[ \text{past} \rightarrow \text{data} \rightarrow \text{model} \rightarrow \text{future prediction} ]

This only works if there is enough continuity between:

[ \text{yesterday} ]

and:

[ \text{tomorrow} ]

But what happens when that continuity breaks?

What happens when:

the technology changes,
the climate changes,
behaviour changes,
regulation changes,
the economy changes,
institutions change,
the population changes,
the model itself changes behaviour?

At some point, the problem is no longer:

How accurately did we learn the past?

It becomes:

Was the past ever a sensible basis for predicting this future?

That is a much harder question.

Prediction always contains an assumption about continuity

Suppose we observe a sequence:

[ y_1,y_2,\ldots,y_t ]

and use it to predict:

[ y_{t+1} ]

Implicitly, we assume that whatever generated:

[ y_1,\ldots,y_t ]

will still have some relevance to what generates:

[ y_{t+1} ]

This does not mean the future must be identical to the past.

It means there must be enough stability for learning from the past to remain useful.

If that stability disappears entirely, historical prediction becomes impossible.

Imagine trying to predict:

tomorrow's electricity demand

using electricity-demand data from:

the year 1850.

The problem is not that the dataset is too small.

The problem is that the world that generated it is fundamentally different.

The future is not merely an extension of a line

A common way to imagine forecasting is:

[ \text{past trend} \rightarrow \text{extend trend} \rightarrow \text{future} ]

Suppose a quantity grows:

[ 2,\ 4,\ 6,\ 8,\ 10 ]

We might predict:

[ 12 ]

Reasonable.

But suppose between:

[ 10 ]

and the next observation, the entire underlying system changes.

A new technology appears.

A law changes.

A resource becomes unavailable.

Behaviour shifts.

Now:

[ 12 ]

may be completely irrelevant.

The line did not continue because reality did not continue in the same way.

Extrapolation is dangerous

Machine-learning systems are often very good at interpolation.

That means making predictions inside regions of the world similar to those already observed.

Suppose a model has seen temperatures between:

[ 0^\circ C ]

and:

[ 30^\circ C ]

Predicting behaviour at:

[ 18^\circ C ]

is interpolation.

But predicting behaviour at:

[ 50^\circ C ]

is extrapolation.

We are asking the model to operate beyond the world it has experienced.

The same applies conceptually to social systems.

A model trained under:

one regulatory system,
one technological regime,
one economic environment,

may be forced to extrapolate after those conditions change.

The model may still produce an answer.

That does not mean the answer deserves confidence.

Models do not know when the rules changed

Suppose a forecasting model learns:

[ Y=f(X) ]

from historical data.

One morning the system changes and the real relationship becomes:

[ Y=g(X) ]

The model does not automatically know this.

It still calculates:

[ \hat{Y}=f(X) ]

with complete mathematical consistency.

The model is doing exactly what it was designed to do.

The problem is that the function it learned no longer describes the system.

This creates one of the most dangerous forms of AI failure:

A model can be internally functioning perfectly while externally becoming wrong.

The confidence can survive longer than the validity

This is particularly dangerous because a model's confidence does not necessarily collapse when the world changes.

Imagine an AI system that has historically seen extremely consistent patterns.

It may produce:

[ P(Y=1)=0.98 ]

But suppose a structural break has occurred.

The old relationship no longer holds.

The model may still report:

[ 98% ]

because its confidence reflects what it learned from the historical distribution.

It does not automatically reflect:

the probability that its assumptions about the world remain valid.

That is a different uncertainty.

We can think of at least two kinds of uncertainty:

uncertainty within the modelled world;
uncertainty about whether the modelled world still exists.

The second can be much harder.

Known unknowns and unknown unknowns

Suppose a weather model predicts:

30% chance of rain.

That is a known uncertainty.

The model knows several weather outcomes are plausible.

But suppose a previously unknown volcanic eruption suddenly alters atmospheric conditions.

The original forecast may have had no representation of that event at all.

This is closer to an unknown unknown.

The model's probability distribution might have described uncertainty beautifully within the space of futures it considered.

But reality produced something outside that space.

Probabilistic forecasts are only as broad as the possibilities represented by the model.

Rare events are not the same as new regimes

This distinction matters.

Suppose a flood occurs with probability:

[ 0.001 ]

The flood is rare, but the model may still have represented it.

That is a low-probability event within the model.

Now imagine the river is redirected by a major infrastructure project.

The historical flood distribution may no longer apply at all.

This is a change in regime.

One problem is:

an unlikely outcome occurred.

The other is:

the probabilities themselves changed.

Those are very different situations.

Some futures cannot be learned directly from history

Imagine trying to forecast electricity demand after a country:

electrifies transport,
electrifies heating,
installs millions of batteries,
changes electricity pricing,
introduces automated demand response.

There may be no historical dataset describing that exact system.

The future state does not yet exist.

We cannot simply search the past for examples.

At this point, prediction may require combining:

historical data,
physical understanding,
causal models,
engineering models,
behavioural assumptions,
scenarios.

This is no longer pure pattern recognition.

Simulation becomes more important

Suppose we want to know:

What happens to an electricity network if 80% of households own an EV?

Historical data may contain very few periods with anything close to that level of adoption.

Instead, we may build a model of:

household demand,
charging behaviour,
network topology,
vehicle usage,
price response.

Then simulate:

[ \text{possible future system} ]

rather than merely extrapolate:

[ \text{historical system} ]

Simulation allows us to ask:

What would happen if the system were different?

That is a fundamentally different question from:

What usually happened next in the past?

Physics can survive where statistical relationships fail

Imagine a machine-learning model predicting electricity flows from historical observations.

Then the network topology changes.

Historical correlations may become unreliable.

But physical laws still apply.

For example:

energy must balance,
lines have capacities,
generators have limits.

This suggests an important principle:

When empirical relationships become unstable, structural knowledge can become more valuable.

In physical systems, that may mean:

conservation laws,
engineering constraints,
network equations.

In social systems, structural modelling is harder, but the principle remains important.

Not every useful model must be learned entirely from correlation.

Causal reasoning becomes more important

Suppose historical data shows:

[ X\rightarrow Y ]

in a statistical sense.

Now a policymaker deliberately changes:

[ X ]

Will (Y) change as the historical relationship suggests?

Maybe.

But if the intervention changes behaviour elsewhere in the system, historical correlation may fail.

This is where we start needing to distinguish:

[ P(Y\mid X) ]

from:

[ P(Y\mid do(X)) ]

The first asks:

What usually happens when we observe (X)?

The second asks:

What would happen if we deliberately changed (X)?

That distinction belongs to causal inference.

It becomes crucial whenever we are not merely predicting the world, but trying to change it.

Policy is about making tomorrow unlike yesterday

There is something almost paradoxical here.

A government might observe:

[ \text{high pollution} ]

for decades.

A predictive model learns that pollution will probably remain high.

But the purpose of environmental policy is precisely to make that prediction wrong.

Likewise:

healthcare aims to reduce disease,
education policy aims to improve outcomes,
climate policy aims to change emissions,
transport policy aims to change behaviour.

Successful policy deliberately creates:

[ P_{\text{future}} \neq P_{\text{past}} ]

So a model that perfectly extrapolates yesterday may be the wrong tool for designing tomorrow.

Good predictions can become bad targets

Suppose an economic model accurately predicts:

Under current policy, inequality will continue increasing.

That may be an excellent forecast.

But it does not mean:

inequality should continue increasing.

Prediction describes what may happen under some continuation of the system.

Decision-making asks:

Do we want that future?

And if not:

What intervention might produce another one?

This is the boundary between:

[ \text{forecasting} ]

and:

[ \text{design} ]

Tomorrow can be intentionally created

The phrase:

the future

can make it sound as though tomorrow is simply arriving.

But many futures are partly created through decisions.

Consider an electricity system.

Tomorrow's demand depends partly on:

weather,
behaviour,
technology.

But it may also depend on:

electricity prices,
automation,
tariffs,
incentives,
network rules.

Those are designed.

So the future state:

[ Y_{t+1} ]

may depend on our present action:

[ A_t ]

Conceptually:

[ Y_{t+1}

f(X_t,A_t) ]

Now prediction and intervention are inseparable.

This creates a counterfactual problem

Suppose we forecast:

There will be a severe electricity shortage tomorrow.

In response, the system operator:

procures more generation,
activates batteries,
reduces flexible demand.

No shortage occurs.

What was the "true" future?

We observed:

[ Y\mid A ]

the outcome after intervention.

But the prediction may have concerned:

[ Y\mid \text{no intervention} ]

We cannot observe both worlds simultaneously.

The future we prevented becomes a counterfactual.

This means systems that successfully respond to forecasts can make their own forecasts appear wrong.

Success can erase the evidence of success

Imagine a hospital AI predicts a patient is at high risk of deterioration.

Doctors intervene early.

The patient recovers.

Historical data now records:

[ \text{high predicted risk} \rightarrow \text{good outcome} ]

If we are careless, we might conclude:

the model overestimated the risk.

But perhaps the good outcome occurred because the prediction triggered treatment.

The causal chain was:

[ \text{prediction} \rightarrow \text{intervention} \rightarrow \text{better outcome} ]

Evaluating intelligent systems becomes harder once their predictions influence action.

Reflexive systems

Some systems respond immediately to predictions about themselves.

Financial markets are a classic example.

Suppose a model predicts:

This asset is underpriced.

Traders buy it.

The price rises.

The prediction changes the thing being predicted.

The same can happen with:

traffic,
electricity prices,
online recommendations,
housing markets,
labour markets.

We get:

[ \text{prediction} \rightarrow \text{behaviour} \rightarrow \text{changed system} \rightarrow \text{new prediction} ]

These are reflexive systems.

The future is not merely uncertain.

It responds.

Strategic behaviour

The problem becomes even harder when people deliberately respond to the model.

Imagine a university uses an AI system to select students.

Applicants learn what features the system rewards.

They change:

applications,
qualifications,
wording,
behaviour.

The input distribution changes.

Similarly:

traders respond to trading algorithms,
firms respond to regulation,
taxpayers respond to tax rules,
drivers respond to navigation systems.

Once humans understand the decision mechanism, they may optimise around it.

The model becomes part of the environment.

The Lucas critique

Economics has long recognised a version of this problem.

Suppose we observe how people behave under one policy regime.

We estimate a relationship.

Then government changes policy.

If people respond to the new rules, the old behavioural relationship may no longer hold.

This insight is associated with what economists call the Lucas critique.

The broad lesson is:

You cannot always predict the effect of a new policy using behavioural relationships estimated under the old policy.

People respond to rules.

Change the rules and you may change the model.

AI makes this problem universal

This is no longer merely an economic problem.

AI systems increasingly create rules people respond to.

If an algorithm determines:

what content is shown,
what loan is approved,
what price is offered,
which applicant is shortlisted,
where resources are sent,

then people and organisations may adapt.

The environment becomes:

[ \text{model-aware} ]

That creates a continuously evolving relationship between:

[ \text{algorithm} ]

and:

[ \text{world} ]

Prediction can become self-fulfilling

Suppose a model predicts that a neighbourhood is likely to experience economic decline.

Banks reduce lending there.

Businesses invest less.

Property values fall.

The predicted decline occurs.

We get:

[ \text{prediction} \rightarrow \text{decision} \rightarrow \text{behaviour} \rightarrow \text{predicted outcome} ]

The prediction helps make itself true.

This is a self-fulfilling prediction.

Prediction can become self-defeating

The opposite can also happen.

Suppose a model predicts:

Severe traffic congestion tomorrow.

People see the warning.

Many work from home or take public transport.

Congestion never occurs.

We get:

[ \text{prediction} \rightarrow \text{response} \rightarrow \text{predicted outcome avoided} ]

The prediction helps make itself false.

This is a self-defeating prediction.

Both cases make simple evaluation difficult.

Models can become obsolete through success

Imagine an AI system designed to detect a particular type of fraud.

It works extremely well.

Fraudsters realise the old attack no longer works.

They change tactics.

The original fraud pattern disappears.

The model's success changed adversarial behaviour.

Now the system needs to learn the new pattern.

This is common in:

cybersecurity,
fraud detection,
spam filtering,
financial crime.

The environment actively adapts against the model.

Tomorrow differs from yesterday because yesterday's model existed.

Adversarial systems

Some prediction problems contain intelligent opponents.

A spam detector learns:

[ \text{spam patterns} ]

Spammers learn:

[ \text{detector behaviour} ]

They alter messages.

The detector adapts.

The spammers adapt again.

Now the data-generating process evolves strategically:

[ \text{model} \leftrightarrow \text{adversary} ]

There may be no stable distribution to learn permanently.

The problem itself is dynamic.

Regime uncertainty

Sometimes we do not know which future regime will emerge.

Suppose a country is considering several energy policies.

Future electricity demand could depend heavily on whether:

heat pumps dominate,
hydrogen dominates,
industrial electrification accelerates,
distributed generation expands rapidly.

Instead of one distribution:

[ P(Y) ]

we might imagine several conditional distributions:

[ P(Y\mid R_1) ]

[ P(Y\mid R_2) ]

[ P(Y\mid R_3) ]

where each (R) represents a different regime.

Now our uncertainty is not merely about:

the value of (Y).

It is about:

which system will exist.

Deep uncertainty

Some decisions involve uncertainty so large that assigning precise probabilities becomes questionable.

Suppose we ask:

What will the global AI economy look like in 2050?

We do not know:

what technologies will exist,
what regulations will exist,
what geopolitical order will exist,
how capable AI systems will become,
what social norms will emerge.

Writing:

[ P(\text{scenario A})=0.37 ]

may create a false sense of precision.

This is sometimes described as deep uncertainty.

We may know the important dimensions of uncertainty without credibly knowing exact probabilities.

Scenarios become valuable under deep uncertainty

When probabilities are poorly known, we can explore multiple futures.

For example:

Scenario A

AI development slows.

Scenario B

AI capability grows steadily.

Scenario C

A major breakthrough produces rapid automation.

Scenario D

Regulation significantly constrains deployment.

The goal is not necessarily to claim:

Scenario B has probability 42%.

Instead we ask:

Would our decision still make sense if any of these futures occurred?

This changes the decision philosophy.

Optimising for one predicted future can make systems fragile

Suppose we predict electricity demand in 2040 as:

[ 500\text{ TWh} ]

and build an entire infrastructure system optimised precisely for that number.

If actual demand is:

[ 350\text{ TWh} ]

we may overbuild enormously.

If actual demand is:

[ 700\text{ TWh} ]

we may face shortages.

Optimisation against a single expected future can create fragility.

A robust system may instead be designed to adapt across a range:

[ 350-700\text{ TWh} ]

The objective becomes:

Remain useful under uncertainty.

Robustness versus optimality

Suppose Decision A produces the best outcome if our forecast is exactly right.

But it performs disastrously if the forecast is wrong.

Decision B is slightly worse under the central forecast but performs reasonably under many possible futures.

Which should we choose?

This introduces a fundamental trade-off:

[ \text{optimality} ]

versus:

[ \text{robustness} ]

In uncertain environments, robustness can be extraordinarily valuable.

The best decision may not be the one optimised most tightly around our favourite prediction.

Adaptability can be more valuable than prediction

Suppose we cannot accurately predict electricity demand twenty years from now.

Perhaps instead of trying to perfect the prediction, we design infrastructure that can expand incrementally.

Now the system can:

[ \text{observe} \rightarrow \text{adapt} \rightarrow \text{observe again} ]

Rather than making one enormous irreversible decision today, we preserve options.

This is a different response to uncertainty.

We reduce dependence on forecasting accuracy.

Optionality

An option gives us the ability, but not the obligation, to take some action later.

In uncertain systems, optionality can have enormous value.

For example:

modular infrastructure,
flexible contracts,
adaptable software,
upgradeable equipment,
reversible policy.

These reduce the cost of being wrong about the future.

Instead of requiring:

[ \text{perfect prediction} ]

we design for:

[ \text{future adaptation} ]

That can be a much more resilient strategy.

Feedback can replace some prediction

Imagine controlling the temperature of a building.

One approach would be to predict the exact temperature for every minute of the next month and set heating accordingly.

That would be absurdly difficult.

Instead, a thermostat repeatedly measures:

[ T_t ]

compares it with the desired temperature:

[ T^* ]

and adjusts heating.

The system uses feedback.

It does not need perfect long-term prediction.

This is one of the great insights of control theory.

You can sometimes achieve good outcomes without predicting the distant future accurately, provided you can observe and adapt quickly enough.

Open-loop versus closed-loop thinking

An open-loop approach might be:

[ \text{predict future} \rightarrow \text{choose action} \rightarrow \text{hope} ]

A closed-loop approach is:

[ \text{observe} \rightarrow \text{decide} \rightarrow \text{act} \rightarrow \text{observe} \rightarrow \text{correct} ]

Closed-loop systems continually respond to reality.

This is especially powerful when tomorrow refuses to behave like yesterday.

Prediction still matters

None of this means forecasting is useless.

Far from it.

Good predictions can:

improve planning,
reduce uncertainty,
reveal risks,
anticipate problems,
support better decisions.

The mistake is treating prediction as certainty.

Or believing that enough historical data allows us to eliminate the possibility of structural change.

The future contains both:

[ \text{continuity} ]

and:

[ \text{novelty} ]

Intelligent systems need to cope with both.

A hierarchy of confidence

We can imagine several situations.

Familiar world

The system closely resembles training data.

Prediction may deserve relatively high confidence.

Unusual world

Inputs are rare but still within plausible historical experience.

Increase uncertainty.

Shifted world

Important relationships appear to be changing.

Reduce reliance on the historical model.

New regime

The system is structurally different.

Use:

new data,
causal models,
physical models,
scenarios,
expert knowledge.

Deep uncertainty

We do not even know which future regime will exist.

Focus increasingly on:

robustness,
adaptability,
optionality,
feedback.

This is a much richer understanding of intelligence than:

fit model, make prediction.

The model should know when to hand over

Imagine an autonomous system operating normally.

Then sensors indicate conditions far outside its training experience.

A dangerous design would insist:

The model must always choose an action.

A safer design might allow:

[ \text{uncertainty}

\text{threshold} ]

to trigger:

human review,
safe mode,
fallback control,
additional observation.

A sophisticated intelligent system should therefore include not just:

[ \text{prediction} ]

but:

[ \text{prediction} + \text{uncertainty} + \text{limits} + \text{fallback} ]

Knowing when not to trust the model can be as important as the model itself.

Human expertise matters most when history stops helping

When a system behaves normally, statistical models may perform beautifully.

When something unprecedented happens, human understanding may become unusually valuable.

Experts may know:

physical constraints,
institutional rules,
causal mechanisms,
failure modes.

These forms of knowledge may not appear clearly in historical correlations.

This suggests a useful complementarity:

[ \text{data-driven learning} + \text{domain knowledge} ]

rather than pretending one can universally replace the other.

Intelligence is partly the ability to adapt

An intelligent organism does not merely predict the world from a fixed historical dataset.

It:

observes,
learns,
updates,
experiments,
adapts.

Artificial systems increasingly need the same capabilities.

We therefore move from:

[ \text{static model} ]

towards:

[ \text{adaptive system} ]

and eventually:

[ \text{agent} ]

that continuously interacts with its environment.

But that creates new risks.

The more the system acts, the more it changes the world generating its data.

The prediction problem becomes part of a control problem.

The past is evidence, not destiny

This may be the most important philosophical lesson of the entire module.

Historical data tells us:

what happened.

It may help us infer:

what could happen next.

But it does not tell us:

what must happen.

And it certainly does not tell us:

what should happen.

A society in which historical inequality exists can produce extremely accurate models of historical inequality.

That does not make inequality desirable.

A fossil-fuel-based economy can produce excellent models of fossil-fuel demand.

That does not mean the economy cannot change.

History provides information.

It does not issue instructions.

Prediction and design

We can now draw an important distinction.

Prediction asks:

Given the system as it is, what is likely to happen?

Design asks:

What system would produce the outcomes we want?

Prediction begins with:

[ \text{existing world} ]

and moves towards:

[ \text{possible future} ]

Design begins with:

[ \text{desired future} ]

and asks what:

rules,
incentives,
technologies,
institutions,
actions

could move us towards it.

This is where the course begins transitioning from prediction towards decision-making.

The deepest forecasting question

At first, we asked:

How accurately can we predict tomorrow from yesterday?

Now we can ask something more sophisticated:

How much should tomorrow resemble yesterday in the first place?

Sometimes continuity is valuable.

Sometimes change is desirable.

Sometimes change is unavoidable.

Sometimes the entire objective is to break the pattern we have learned.

This means the quality of intelligence cannot ultimately be measured only by how accurately it predicts continuation.

It must also help us reason when continuation is no longer a sensible assumption.

Bringing Module 5 together

This module began with a simple question:

When is a prediction good?

We discovered that there is no single answer.

We examined:

accuracy,
mean absolute error,
mean squared error,
precision,
recall,
false positives,
false negatives,
confusion matrices,
ROC curves,
calibration,
probabilistic forecasts,
scoring rules,
forecasting across time,
spatial prediction,
distribution shift.

Each revealed another dimension of predictive quality.

A prediction can be:

accurate on average but disastrous in important cases;
well classified but badly calibrated;
excellent nationally but terrible locally;
excellent today but useless next year;
statistically correct but operationally irrelevant.

And now we reach the final lesson.

Even a model that passes every historical evaluation faces one unavoidable limitation:

The future is not required to resemble the dataset.

The deeper lesson

Machine learning begins with the past because the past is what we can observe.

But intelligent decision-making cannot end there.

The full problem looks more like:

[ \text{PAST} \rightarrow \text{DATA} \rightarrow \text{MODEL} \rightarrow \text{POSSIBLE FUTURES} ]

followed by:

[ \text{OBJECTIVES} + \text{CONSTRAINTS} + \text{UNCERTAINTY} ]

leading to:

[ \text{DECISION} \rightarrow \text{ACTION} \rightarrow \text{NEW REALITY} ]

The objective is therefore not:

Predict the future perfectly.

That is impossible.

A more useful objective is:

Learn what the past can teach us, represent what we do not know, recognise when the world is changing, and make decisions that remain sensible as reality unfolds.

That means asking not just:

What happened yesterday?

or:

What will happen tomorrow?

but:

What assumptions connect the two?

Which of those assumptions could fail?

How would we know they were failing?

What should we do if they do?

And finally:

If tomorrow does not resemble yesterday, can our system adapt — or does it simply become confidently wrong?