What If Life Is Just Computation… With Better PR?
A field guide to Blaise Agüera y Arcas’ What is Life? Evolution as Computation — and why it feels uncomfortably relevant to AI right now
I finished Blaise Agüera y Arcas’ book, What is Life? Evolution as Computation.
And here’s the neat part: this “book” is also the first chapter of a larger project called What is Intelligence, which is available online.
Matryoshkas, everywhere. Books inside books. Ideas inside ideas. Like nature, but with more footnotes.
This one is more popular-science than hard science, and that’s not an insult. It’s what makes it useful. The author takes the core results from his artificial-life work (the one we discussed recently), then uses the extra breathing room of a book to do what papers rarely allow: step back, speculate responsibly, and connect dots that usually live in different academic zip codes.
And the central dot is this:
Life and computation aren’t just metaphorically related. They may be structurally the same kind of thing.
Which sounds like a philosophy seminar until it suddenly doesn’t.
The book starts where it should: before life had a user manual
The opening stretches back to abiogenesis — the messy prequel where chemistry is trying to become biology without knowing what biology is.
There’s a detour through metabolic pathways, including the idea that early chemical cycles (like variants of the reverse Krebs cycle) could have produced useful organics before anything resembling modern “life” existed.
The point isn’t to win a debate about the One True Origin Story.
It’s to set up a theme that returns again and again:
life is what happens when matter finds a way to keep processing information without falling apart.
Mutations are “updates.” Symbiogenesis is a platform shift.
Here’s where the book gets properly interesting: symbiogenesis.
Classic Darwinian evolution is often sold like this:
mutation generates variation
selection filters it
repeat until you get peacocks, poison frogs, or JavaScript frameworks
That’s fine. But it mostly describes incremental optimization.
Symbiogenesis, in this telling, is something else: it opens new combinatorial space. It’s not “tune the parameters.” It’s “merge two systems and suddenly new capabilities exist.”
Mutation can fine-tune. Symbiogenesis can invent whole new “modules.”
If you work in tech, you already know this pattern. It’s the difference between:
polishing a feature
and acquiring a company that gives you an entirely new product category overnight
In evolutionary terms, symbiogenesis is the moment when nature stops updating… and starts integrating.
Why computer science keeps sneaking into biology
Agüera y Arcas leans into a historical thread I enjoyed: early giants of computing didn’t treat biology as “that other thing.”
They were fascinated by it.
Turing wandered toward morphogenesis and reaction-diffusion patterns.
Von Neumann explored self-replicating automata and the idea of a universal constructor.
The book also drops a fun detail: Turing apparently pushed for a random-number instruction on the Ferranti Mark I — a small historical flourish with a big conceptual message.
Because randomness isn’t a bug in life.
It’s a feature.
Artificial life, but in the most hostile neighborhood possible
The heart of this book is the artificial-life work: replicators emerging in a minimal computational environment (the same line of research you referenced).
In the book version, the story gets extra visual support — more figures, clearer “how it grows” intuition — including charts that show:
step-like jumps in computational activity
changes in how much “coding material” exists over time
And here comes the line that quietly steals the whole show:
replicators appear because they’re dynamically stable.
Not “strong.” Not “well-protected.” Stable.
A passive object can be “robust,” yet still fragile in the deeper sense: it can only endure. It can’t recreate itself when the environment chews it up.
A replicating pattern is different. It’s anti-fragile: the environment can destroy instances, but the pattern persists.
That’s why DNA feels immortal even though every organism is not.
Life doesn’t beat entropy by being indestructible.
It beats entropy by being copyable.
“Isn’t complexity illegal?” No. You’re just looking at the wrong accounting.
At some point everyone raises the same objection:
“How can complexity increase if the second law of thermodynamics is a thing?”
The book’s answer is essentially: the second law isn’t being violated. We’re just mixing levels.
In a population of replicators:
“more efficient replication” maps to “more stability”
selection becomes a thermodynamic story about attractors, not a moral story about progress
In that framing, Darwinian selection starts to look like a cousin of the second law, not its enemy.
And once you accept that, the core claim lands:
If a system allows computation, replicators can become an attractor.
Not guaranteed everywhere. Not magic. But plausible as a predictable outcome of statistical processes.
Life, then, isn’t a miracle.
It’s what tends to happen when the physics allows a certain kind of information processing to persist.
The AI analogy: thermodynamics all over again
One of the smartest moves in the book is the comparison to thermodynamics as a field.
Thermodynamics existed as practical engineering long before it had a crisp theoretical backbone. People built engines. Then, later, we learned how to explain what those engines were actually doing.
Modern AI feels like that.
We have systems that work. Often disturbingly well. Sometimes embarrassingly badly. And our best tool for understanding them is still: try things, measure, iterate.
Artificial life (in a minimal computational environment) could be the “billiard balls” of this era: a simplified model that won’t capture everything, but might capture enough to reveal the shape of the underlying laws.
Not because it’s cute.
Because it’s controllable.
The punchline: symbiogenesis is the real accelerant
The book’s closing thrust circles back to symbiogenesis, now inside the artificial-life story:
Replicators don’t just appear fully formed. They can assemble from smaller pieces through “symbiotic events.” Once a robust replicator exists, evolvability changes. Now, anything that doesn’t break copying can be inherited, recombined, selected.
That’s when the curve bends.
That’s when you get something that looks suspiciously like an exponential takeoff inside the soup.
If you’re watching AI agents, tool-use, memory systems, and model composition right now and thinking, “This feels like modules snapping together,” you’re not hallucinating.
You’re noticing a pattern that might be older than biology itself.
Closing thought
The most unsettling idea in What is Life? is not “life is computation.”
It’s this:
Life may be the default outcome when computation is allowed to run in the right conditions for long enough.
Which reframes a lot of things.
It reframes biology as a kind of engineering that nature stumbled into.
And it reframes AI as something less like a product category and more like… a new habitat.
If this topic grabs you, reply and tell me what you want next: a straight summary of the “computational life” results, or the bigger question it implies — what counts as “alive” when the substrate is software?