02: What Would Digital Life Mean?

Suppose life were possible in software.

What would we actually be looking for?

The honest answer is that we do not know yet, and that is the reason for this book.

We could certainly build something that looks alive: a graphical creature, an animated agent, a personality-driven assistant. That is not the experiment. The harder question is whether a computational system can produce, through its own operation, properties that justify some of the language we associate with living things.

Produce them, rather than declare them.

That sounds simple until we try to say what those properties are.

We reach immediately for a familiar vocabulary — not a list of requirements, just the properties biology has taught us to notice:

structure
persistence
response
adaptation
repair
memory
reproduction
inheritance
evolution

Those words are useful. They are also a trap. Because the easiest way to build something that looks like digital life is to put the vocabulary of life directly into the program. We can write:

class Metabolism:
    ...

class Memory:
    ...

class Reproduction:
    ...

class Organism:
    ...

and within an afternoon produce software that is unmistakably biologically inspired.

What we would not know is whether any of those biological words describe properties of the system rather than decisions made by its programmer.

Calling a method reproduce() does not establish reproduction.

Calling a number energy does not establish metabolism.

Writing state to disk does not establish memory. Copying a configuration does not establish inheritance. Running an optimizer does not establish evolution.

The names are ours.

The evidence has to come from the system.

That is where this book begins.

The Cargo Cult of Life

A cargo cult copies the visible form of something without reproducing the mechanism that made the original work.

Artificial life has an obvious version of this problem. We can borrow the vocabulary of biology — cell, organism, energy, fitness, birth, death, memory, inheritance, evolution — and then build software structures carrying those names. The result may be complicated. It may be beautiful. It may even be useful.

But the terminology proves nothing.

So we will work under a stricter rule.

We do not get to claim a life-like property merely because we implemented something with the same name.

Persistence has to be measured.

Regeneration requires damage: disturb the system, then measure what returns.

Learning requires experience to produce a measurable change in performance.

Inheritance means that something useful survives into a continuation or successor.

Reproduction demands more than visual resemblance between an earlier structure and a later one. There must be evidence that the earlier organization played a causal role in producing the later one.

Evolution imposes the hardest standard. We need to identify what varies, what persists, what is selected, and what changes across generations or through time.

Naming these properties is easy. Demonstrating them is the work.

But There Is a Second Trap

Avoiding biological vocabulary is not enough, because there is a mistake available in the opposite direction.

We could decide in advance what life must contain:

boundary
metabolism
death
reproduction
inheritance
selection

and then spend the rest of the book manufacturing those properties one at a time.

That would be better than naming classes after them. It would still smuggle in an assumption:

that digital life must solve the same problems, in the same way, that biological life does.

Why should it?

Biological organisms exist under very particular constraints. They occupy physical bodies. Matter is expensive to rearrange. Copying an organism is difficult and slow. Bodies degrade. Individuals die. Acquired knowledge is only partially transferable to a successor. Lineages branch and generally do not merge.

A computational substrate has a different set of possibilities and a different set of costs. Copies can be nearly free. State can be checkpointed and restored. A process can move between machines. Two lineages can fork and later merge. Acquired information can, at least in principle, be transferred directly. One organization can be distributed across many locations. A digital system might simply grow, and never reproduce at all.

None of this means computation is unconstrained. Quite the opposite. It means that if digital life is possible, its important constraints may be computational rather than biological — and we should not decide in advance what those constraints will be.

So this book will not open with a checklist called REQUIREMENTS FOR LIFE.

We do not know the requirements.

We do not even know whether “requirements” is the right way to think about the problem. A biological checklist could make us blind to phenomena for which biology gives us no ready analogue.

Discovering what survives is the experiment.

Biology Is Evidence, Not Specification

For centuries people watched birds and tried to understand flight.

Birds have feathers, wings, hollow bones, flight muscles, flapping motion. All of that was worth studying, and studying it was how the problem became tractable at all. But successful aircraft did not require us to manufacture an artificial bird. The important discoveries were underneath the anatomy:

lift
drag
thrust
stability
control

Birds were evidence that flight was possible. They were not the engineering specification.

Life may be similar.

Biology gives us one extraordinarily successful implementation, and the only one we can currently examine. But an organism is a solution shaped by the medium it is made from, and many of its most familiar features may be consequences of that medium rather than universal requirements for whatever forms organized persistence might take elsewhere.

So the task of this book is not:

Build a digital animal.

It is:

Discover the dynamics of digital life.

Which mechanisms actually matter? Which properties arise on their own? And which of the things we assume are fundamental turn out to be solutions biology found for the specific problem of living in the physical world?

We should not answer any of that in advance.

Do Not Trust What Looks Alive

There is a further problem, and it is not in the software. It is in us.

A complicated-looking simulation can be mesmerizing. We can generate a beautiful animation and immediately feel that something profound is happening. Maybe it is. Maybe it is not. Random noise can look complicated. A short-lived transient can look extraordinary in the moment just before it disappears. A periodic system can look chaotic if we watch too small a window. A pattern can resemble an organism while possessing no property beyond its geometry.

Surprise is cheap.

And humans are extremely good at turning motion into nouns.

We see a shape move, and start saying it travelled. We see one shape become two, and start saying it reproduced. We see several structures move together, and start saying they flocked. The upgrade happens in under a second, and it happens before any measurement has been taken.

Those words may eventually be justified. The animation does not justify them.

So the book insists throughout on a distinction that is easy to state and surprisingly hard to maintain:

The phenomenon and our explanation of the phenomenon are not the same thing.

In a new substrate, even the definition has to be earned. Before we call something memory, reproduction or individuality, we need to say what observation would count as evidence for it and what would count against it. That operational definition is part of the experiment, not something the vocabulary supplies for free.

Which gives us a second rule:

Look first. Then try to destroy your own interpretation.

Even Our Nouns Will Have to Earn Their Place

This applies to the small words as well as the impressive ones.

Suppose a pattern persists while moving through a lattice. Is it an object? Perhaps. Suppose the visible pattern disappears and an equivalent pattern reappears elsewhere. Is it the same object? Perhaps not. Suppose two disconnected regions participate in one continuing causal process — are they two things, or one distributed thing? Suppose two structures look identical but have entirely independent histories. Are they the same kind of object? Yes. Are they the same individual? Probably not.

We will begin with simple geometric definitions, because they are measurable and because we have to begin somewhere. But we should hold them loosely.

The visible boundary of a pattern may not be the true boundary of a digital individual.

Persistence creates the same problem. A program can persist indefinitely by doing nothing, so duration alone tells us very little.

Reproduction deserves similar suspicion. Biology makes it look unavoidable because organisms die and lineages continue through successors. A digital process might simply continue. Reproduction may turn out to matter — but the substrate has to show us why.

Evolution raises the same problem. Digital inheritance could include acquired state, learned parameters, external memory or modified code.

So the broader question is:

How can history produce cumulative, measurable changes in what a computational system can do?

Biological evolution is one answer.

We do not yet know the others.

The visible object may not be the important unit. History may matter without anything we would recognize as memory. Reproduction may not be fundamental at all.

Those are questions for the experiments. We should not settle them by choosing the right vocabulary.

How This Book Will Find Out

That leaves a practical problem. If we cannot trust the vocabulary and cannot assume the requirements, how do we know what to look for?

We start with what we can observe. The visual system gives us hypotheses. The experiment decides what survives.

So the same cycle runs through the entire book:

SEE SOMETHING
↓
NAME THE HYPOTHESIS
↓
DEFINE WHAT WOULD COUNT AS EVIDENCE
↓
MEASURE IT
↓
BUILD A CONTROL
↓
LOOK FOR A CONFOUND
↓
BUILD A BETTER CONTROL
↓
KEEP ONLY WHAT SURVIVES

Later chapters will return to this sequence again and again, because the procedure is part of the instrument.

The questions will change. The procedure does not.

Three parts of the cycle do most of the work.

Intervention. Correlation is easy to obtain and hard to interpret. If we think a mechanism produces a capability, the strongest available move is to remove or disrupt that mechanism and measure what changes. If the capability does not depend on the mechanism, the explanation was wrong, however satisfying it looked.

Controls. A number alone means very little. A recovery score of 0.87 is impressive or unremarkable depending entirely on what an undamaged system scores, what a random process scores, and what a frozen copy scores. So every control in this book exists because there is a specific alternative explanation we are trying to eliminate — and the first control is often not good enough.

Discovering a confound in your own experiment is not a setback. It tells you that the result was carrying more than one possible explanation.

Remove one, and you may finally be able to see what remains.

Bounded claims. Compare:

The system heals.

with a deliberately bounded statement such as:

Under perturbation P in region R, structural measure M recovered to 0.87 within T steps.

The second claim may sound less dramatic, but its edges are visible. We know what was changed, what was measured, and where the conclusion stops.

One consequence needs stating plainly, because it is where investigations usually go wrong.

The standard of evidence does not change when a hypothesis becomes attractive. It does not relax because the result would be exciting, and it does not tighten because the result is inconvenient. A beautiful animation generates a hypothesis. It never establishes an interpretation. If the strongest honest conclusion is smaller than the idea that motivated the experiment, then the smaller conclusion is the result.

Expect Us to Be Wrong

This deserves to be explicit, because it is not a disclaimer. It is a description of what actually happened while the book was being written.

We form hypotheses that fail. We design controls that turn out to be inadequate. We produce measurements that look spectacular and then collapse when a confound appears. Sometimes we discover that a question we had been asking for three chapters was really two different questions wearing one name.

The recurring shape is roughly this:

we expect X
↓
the experiment does not support X
↓
something else appears
↓
we investigate that
↓
the new interpretation may fail too
↓
a smaller or stranger phenomenon survives

That last line is the point.

This is not a book in which every chapter proposes a property and then successfully proves that property exists.

Quite often we went looking for one thing and were shown another.

Sometimes the first explanation failed. Sometimes the second failed too. Occasionally several experiments were needed before we understood what the system had been showing us all along.

That leads to one of the central rules of the investigation:

An explanation can die without the phenomenon dying.

A failed interpretation does not oblige us to throw away the observation that killed it. Sometimes the observation is the discovery.

Separating the measurement that survives from the interpretation that does not is most of the work.

We are not trying to prove that digital life exists.

We are trying to find out which claims remain standing after we attack them.

There Is No Biological Ladder Here

We will meet familiar properties along the way — structure, persistence, identity, damage tolerance, repair, reproduction, inheritance, adaptation, evolution. They are experimental questions, and the order in which we study them is a path through the investigation rather than a ladder toward life.

Nothing earns points toward an ALIVE label.

The Question We Are Actually Asking

At the end of this book we may still decline to say that anything we built is alive.

That would be an acceptable result. It might even be the correct one.

What matters is whether we can say things like:

this structure persists under these conditions

this pattern fails under this perturbation

this organization restores part of its form after this damage

this later structure is causally descended from that earlier one

this earlier state measurably changes a later response

this effect survives this control

this apparent behaviour disappears when this confound is removed

Those are modest sentences. They are also somewhere solid to stand, which is more than most of the vocabulary we started with can offer.

Perhaps that is where the aerodynamics begin: with measurements precise enough to survive before there is a general theory to contain them.

And if enough surprising properties survive enough attempts to explain them away, a different question eventually becomes unavoidable. Not:

Does this look alive?

But:

What kind of thing is actually here?

That is the question this book keeps returning to.

We will repeatedly ask the system to show us something for which we already have a name. Sometimes it will.

Sometimes it will show us something else.

The job is to notice the difference.

Where to Start

To ask any of this, we need a world simple enough that we can see what is happening.

Small computational systems — a lattice, a local rule, a clock — are almost ideal for this, and not because they resemble organisms. They do not. They are useful because they remove excuses. If something interesting happens inside a very large model, the explanation can disappear into millions of parameters and never be found. In a system with a handful of ingredients we can inspect every state, replay every step, flip one bit, run the counterfactual, compare neighbouring rules, and measure exactly what the perturbation did.

When a claim fails there, it has almost nowhere to hide.

That is what makes these systems worth beginning with. They are not models of life.

They are experimental microscopes for emergence.

In the next chapter we will calibrate the microscope.

We will begin with computational systems rich enough to make the temptation almost irresistible: shapes that move, persist, collide and appear to behave like things.

We will let ourselves see the creature.

Then we will start taking assumptions away.

Continuous state. Rich neighbourhoods. Complicated rules. Even the idea that a persistent pattern must be made from the same material from one moment to the next.

By the time almost nothing is left, we will already have encountered the problem that drives the rest of this book:

Computation can produce something worth explaining long before we know what to call it.