What
Is Life?

Preface to What is Life?

Antikythera is a philosophy of technology research program reorienting planetary computation as a philosophical, technological, and geopolitical force.

It takes its name from the first known computer, the Antikythera mechanism, an ancient tool for orientation, navigation, planning, and prediction across planets.

Antikythera understands the future of computation as more than calculation. It is a technology to think with. It is an instrument for epistemological discovery.

Arguably computation was discovered as much as it was invented. It is part of how the universe works, including, as Blaise Agüera y Arcas will show, how life works.

That such a claim could be made is due in no small part to the creative and curious use of our computational tools—or what we might more precisely call artificial computation. With those tools we discover that some otherwise imperceivable building blocks of our reality and of our own flesh are themselves computational. Computation discovers itself through us.

The Antikythera book series, in collaboration with MIT Press, begins with this extraordinary little book that asks perhaps the most complex question of all: what is life?

The answers are surprising not only in themselves but also in how they reframe the question to include also how is life. As it turns out, how and what can be the same thing.

We are proud to launch this initiative by laying a cornerstone with the thoughts of my dear friend, Blaise.

Benjamin Bratton
La Jolla, California
September 2024

Foreword

In late 2023, I set out to write What Is Intelligence?, a short volume on AI to help kick off the Antikythera book series in collaboration with MIT Press. The title was an homage to physicist Erwin Schrödinger’s slim but influential 1944 book, What Is Life?

In retrospect, it seems like no coincidence that my newly formed team at Google, Paradigms of Intelligence, was not only working to extend and improve on existing approaches to AI; we had also caught the Artificial Life bug, as many AI researchers have before. Hence life, artificial and otherwise, became the focus of Part I of the book. Life and intelligence turn out to be deeply intertwined. Both are products of the same fundamental evolutionary forces, and both are (I will argue) computational at heart. It may be misleading even to think about them as distinct phenomena.

Our plan had always been to publish What Is Intelligence? both in print and freely online, in a serialized and media-rich form, beginning with a first installment for Antikythera’s launch event at MIT in October 2024. It seemed natural to begin with Part I. Late that summer, we realized that since What Is Intelligence? had grown a good deal longer than I had originally imagined, Part I could stand alone as a little book in its own right. So we raced to get a few hundred in print. That is what you are now holding. It’s the right size, and covers the right material, to be called—with due deference to the 80-year-old original—What Is Life?

It’s likely that bacteria are themselves the product of such symbiotic events—for instance, between RNA and metabolism. ¹² RNA could replicate without help from proteins, and the metabolic motor could proliferate without help from genes, but when these systems cooperate, they do better. The looping chemical reaction networks in those black smokers can be understood as such an alliance in their own right, a set of reactions which, by virtue of catalyzing each other, can form a more robust, self-sustaining whole.

So in a sense, Darwin may have been right to say that “it is mere rubbish” to think about the origin of life, for life may have had no single origin, but rather, have woven itself together from many separate strands, the oldest of which look like ordinary chemistry. Intelligent design is not required for that weaving to take place; only the incontrovertible logic that sometimes an alliance creates something enduring, and that whatever is enduring … endures.

Often, enduring means both creating and occupying entirely new niches. Hence eukaryotes did not replace bacteria; indeed, they ultimately created many new niches for them. Likewise, the symbiotic emergence of multicellular life—another major evolutionary transition—did not supplant single-celled life. Like an ancient parchment overwritten by generations of scribes, our planet is a palimpsest, its many-layered past still discernible in the present. Even the black smokers, throwbacks to a Hadean sea billions of years ago, are still bubbling away in the depths. The self-catalyzing chemistry of proto-life may still be brewing down there, slowly, on the ocean floor.

The idea that evolution is driven by symbiotic mergers, the earliest of which preceded biology as we know it, has far-reaching implications. One is that the boundary between life and non-life is not well defined; symbiogenesis can involve any process that, one way or another, is self-perpetuating. Evolutionary dynamics are thus more like a physical law than a biological principle. Everything can be subject to evolution, whether we consider it to be “alive” or not.

The symbiogenetic view also renders the idea of distinct species, classified according to a Linnaean taxonomy, somewhat ill-defined—or at best, of limited applicability. Such taxonomies assume that only branching takes place, not merging. Bacteria, which promiscuously transfer genes even between “species,” already challenge this idea. ¹³ Taxonomies break down entirely when we try to apply them to even more fluid situations, like the possible proto-life in black smokers, or microbiomes, or the more complex symbioses we’ll explore later.

Perhaps most importantly, symbiogenesis explains evolution’s arrow of time, as classical Darwinian theory alone cannot. When life branches, specializing to adapt to a niche—like Darwin’s finches with their differently shaped beaks, each optimized for a specific food source—those branches are in general no more complex than their ancestral form. This has led some classical evolutionary theorists to argue that the increasing complexity of life on Earth is an anthropocentric illusion, nothing more than the result of a random meander through genetic possibilities. ¹⁴ Relatedly, one sometimes hears the claim that since all extant species are of equal age—as it seems they must be, since all share a common single-celled ancestor some four billion years ago—no species is more “evolved” than any other.

On its face, that seems reasonable. But as we’ve seen, classical Darwinian theory struggles to explain why life seems to become increasingly complex, or indeed, how life could arise in the first place. In themselves, random changes or mutations can only fine-tune, diversify, and optimize, allowing the expression of one or another variation already latent in the genetic space. (The space of possible finch beaks, for instance.)

When one prokaryote ends up living inside another, though, or multiple cells band together to make a multicellular life form, the resulting composite organism is clearly more complex than its parts. Something genuinely new has arisen. The branching and fine-tuning of classical evolution can now start to operate on a whole different level, over a new space of combinatorial possibilities.

Classical evolution isn’t wrong; it just misses half of the story—the more rapid, more creative half. One could say that evolution’s other half is revolution, and that revolutions occur through symbiosis.

Suggestively, the same is true of technology. In Paleolithic times, a hafted spear was not just a specialized stone point, but something new that arose by combining at least three parts: a stone point, a shaft, and something like sinew to bind them. This, too, opened up a new combinatorial design space. Or, in more recent times, the silicon chip was not just an evolved transistor; it was something new that could be made by putting multiple transistors together on the same die. One doesn’t arrive at chip design merely by playing with the design parameters of an individual transistor.

That’s why evolution progresses from simpler to more complex forms. It’s also why the simpler forms often remain in play even after a successful symbiogenetic event. Standalone stone points, like knives and hand axes, were still being made after the invention of the hafted spear; and, of course, spearheads themselves still count as stone points. ¹⁵ Moreover, no matter how recently any particular stone point was made, we can meaningfully talk about stone points being “older” as a category than spears. Spears had to arise later, because their existence depended on the pre-existence of stone points.

It’s equally meaningful to talk about ancient life forms, like bacteria and archaea, co-existing alongside recent and far more complex ones, like humans—while recognizing that humans are, in a sense, nothing more than complex colonies of bacteria and archaea that have undergone a cascade of symbiotic mergers.

Turing’s most important contribution to computer science came fifteen years earlier, though. After earning a degree in mathematics at Cambridge in 1935, he focused on one of the fundamental outstanding problems of the day: the Entscheidungsproblem (German for “decision problem”), which asked whether there exists an algorithm for determining the validity of an arbitrary mathematical statement. The answer turned out to be “no,” but the way Turing went about proving it ended up being far more important than the result itself. ¹⁷

Turing’s proof required that he define a general procedure for computation. He did so by inventing an imaginary gadget we now call the “Turing Machine.” The Turing Machine consists of a read/write head, which can move left or right along an infinite tape, reading and writing symbols on the tape according to a set of rules specified by a built-in table.

First, Turing showed that such a machine could do any calculation or computation that can be done by hand, given an appropriate table of rules, enough time, and enough tape. He suggested a notation for writing down the table of rules, which he referred to as a machine’s “Standard Description” or “S.D.” Today, we’d call it a “program.”

Then, the really clever part: if the program itself is written on the tape, there exist particular tables of rules that will read the program and perform whatever computation it specifies. Today, these are called “Universal Turing Machines.” In Turing’s more precise language, “It is possible to invent a single machine which can be used to compute any computable sequence. If this machine U is supplied with a tape on the beginning of which is written the S.D. of some computing machine M, then U will compute the same sequence as M.”

In the early 1940s, John von Neumann, a Hungarian American polymath who had already made major contributions to physics and mathematics, turned his attention to computing. He became a key figure in the design of the ENIAC and EDVAC—among the world’s first real-life Universal Turing Machines, now known simply as “computers.” ¹⁸

Turning an imaginary mathematical machine into a working physical one required many further conceptual leaps, in addition to a lot of hard nuts-and-bolts engineering. For instance, over the years, much creativity has gone into figuring out how simple the “Standard Description” of a Universal Turing Machine can get. Only a few instructions are needed. Esoteric language nerds have even figured out how to compute with a single instruction (a so-called OISC or “one-instruction set computer”).

There are irreducible requirements, though: the instruction, or instructions, must change the environment in some way that subsequent instructions are able to “see,” and there must be “conditional branching,” meaning that depending on the state of the environment, either one thing or another will happen. In most programming languages, this is expressed using “if/then” statements. When there’s only a single instruction, it must serve both purposes, as with the SUBLEQ language, whose only instruction is “subtract and branch if the result is less than or equal to zero.”

Both Turing and von Neumann were keenly aware of the parallels between computers and brains. Von Neumann’s report on the EDVAC explicitly described the machine’s basic building blocks, its logic gates, as electronic neurons. ¹⁹ Whether or not that analogy held (it did not; neurons are more complex than logic gates), his key insight was that both brains and computers are defined not by their mechanisms, but by what they do—their function, in both the colloquial and mathematical sense.

In real life, though, the brain is not an abstract machine, but part of the body, and the body is part of the physical world. How can one speak in purely computational terms about the function of a living organism, when it must physically grow and reproduce?

Although working independently, Turing and von Neumann both became captivated by the connection between biology and computation toward the end of their lives. ²⁰ Turing did pioneering work in “morphogenesis,” working out how cells could use chemical signals, which he dubbed “morphogens,” to form complex self-organizing patterns—the key to multicellularity. ²¹ Although computers were still too primitive to carry out any detailed simulations of such systems, he showed mathematically how so-called “reaction–diffusion” equations could generate spots, like those on leopards and cows, or govern the growth of tentacles, like those of the freshwater polyp, Hydra. ²²

Around the same time, in a 1951 paper, ²³ von Neumann imagined a machine made of standardized parts, like Lego bricks, paddling around on a reservoir where those parts could be found bobbing on the water. The machine’s job was to gather all of the needed parts and construct another machine like itself. Of course, that’s exactly what a bacterium does to reproduce; in fact it’s what every cell must do to divide, and what every mother must do to give birth.

It’s possible for a trivially simple structure, like a seed crystal, to “reproduce” merely by acting as a template for more of the same stuff to crystallize around it. But a complex machine—one with any internal parts, for example—can’t serve as its own template. And if you are a complex machine, then, on the face of it, manufacturing something just as complex as you yourself are has a whiff of paradox, like lifting yourself up by your own bootstraps. However, von Neumann showed that it is not only possible, but straightforward, using a generalization of the Universal Turing Machine.

He envisioned a “machine A” that would read a tape containing sequential assembly instructions based on a limited catalog of parts, and carry them out, step by step. Then, there would be a “machine B” whose function was to copy the tape—assuming the tape itself was also made of available parts. If instructions for building machines A and B are themselves encoded on the tape, then voilà—you would have a replicator. ²⁴

Instructions for building any additional non-reproductive machinery could also be encoded on the tape, so it would even be possible for a replicator to build something more complex than itself. ²⁵ A seed, or a fertilized egg, illustrates the point. Even more fundamentally, encoding the instructions to build oneself in a form that is itself replicated (the tape) is the key to open-ended evolvability, meaning the ability for evolution to select for an arbitrary design change, and for that change to be inherited by the next generation.

Remarkably, von Neumann described these requirements for an evolvable, self-replicating machine before the discovery of DNA’s structure and function. ²⁶ Nonetheless, he got it exactly right. For life on Earth, DNA is the tape; DNA polymerase, which copies DNA, is “machine B”; and ribosomes, which build proteins by following the sequentially encoded instructions on DNA, are “machine A.” Ribosomes and DNA polymerase are made of proteins whose sequences are, in turn, encoded in our DNA and manufactured by ribosomes. That is how life lifts itself up by its own bootstraps.

Of course, there are meaningful differences between biological computing and the kind of digital computing done by the ENIAC or your smartphone. DNA is subtle and multilayered, including phenomena like epigenetics and gene proximity effects. Cellular DNA is nowhere near the whole story, either. Our bodies contain (and continually swap) countless bacteria and viruses, each running their own code.

Biological computing is massively parallel, decentralized, and noisy. Your cells have somewhere in the neighborhood of 300 quintillion ribosomes, all working at the same time. Each of these exquisitely complex floating protein factories is, in effect, a tiny computer—albeit a stochastic one, meaning not entirely predictable. The movements of hinged components, the capture and release of smaller molecules, and the manipulation of chemical bonds are all individually random, reversible, and inexact, driven this way and that by constant thermal buffeting. Only a statistical asymmetry favors one direction over another, with clever origami moves tending to “lock in” certain steps such that a next step becomes likely to happen. This differs greatly from the operation of logic gates in a computer, which are irreversible, ²⁷ and designed to be ninety-nine point many-nines percent reliable and reproducible.

Biological computing is computing, nonetheless. And its use of randomness is a feature, not a bug. In fact, many classic algorithms in computer science also require randomness (albeit for different reasons), which may explain why Turing insisted that the Ferranti Mark I, an early computer he helped to design in 1951, include a random number instruction. ²⁸ Randomness is thus a small but important conceptual extension to the original Turing Machine, though any computer can simulate it by calculating deterministic but random-looking or “pseudorandom” numbers.

Parallelism, too, is increasingly fundamental to computing today. Modern AI, for instance, depends on both massive parallelism and randomness—as in the parallelized “stochastic gradient descent” (SGD) algorithm, used for training most of today’s neural nets, the “temperature” setting used in chatbots to introduce a degree of randomness into their output, and the parallelism of Graphics Processing Units (GPUs), which power most AI in data centers.

Traditional digital computing, which relies on the centralized, sequential execution of instructions, was a product of technological constraints. The first computers needed to carry out long calculations using as few parts as possible. Originally, those parts were flaky, expensive vacuum tubes, which had a tendency to burn out and needed frequent replacement by hand. The natural design, then, was a minimal “Central Processing Unit” (CPU) operating on sequences of bits ferried back and forth from an external memory. This has come to be known as the “von Neumann architecture.”

Turing and von Neumann were both aware that computing could be done by other means, though. Turing’s model of morphogenesis was a biologically inspired form of massively parallel, distributed computation. So was his earlier concept of an “unorganized machine,” a randomly connected neural net modeled after the brain of an infant. ²⁹ These were visions of what computing without a central processor could look like—and what it does look like, in living systems.

Von Neumann also began exploring massively parallel approaches to computation as far back as the 1940s. In discussions with Polish mathematician Stanisław Ulam at Los Alamos, he conceived the idea of “cellular automata,” pixel-like grids of simple computational units, all obeying the same rule, and all altering their states simultaneously by communicating only with their immediate neighbors. With characteristic bravura, von Neumann went so far as to design, on paper, the key components of a self-reproducing cellular automaton, including a horizontal line of cells comprising a “tape” and blocks of cellular “circuitry” implementing machines A and B.

Designing a cellular automaton is far harder than ordinary programming, because every cell or “pixel” is simultaneously altering its own state and its environment. When that kind of parallelism operates on many scales at once, and is combined with randomness and subtle feedback effects, as in biological computation, it becomes even harder to reason about, “program,” or “debug.”

Nonetheless, we should keep in mind what these two pioneers understood so clearly: computing doesn’t have to be done with a central processor, logic gates, binary arithmetic, or sequential programs. One can compute in infinitely many ways. Turing and his successors have shown that they are all equivalent, one of the greatest accomplishments of theoretical computer science.

This “platform independence” or “multiple realizability” means that any computer can emulate any other one. If the computers are of different designs, though, the emulation may run s … l … o … w … l … y. For instance, von Neumann’s self-reproducing cellular automaton has never been physically built—though that would be fun to see! It was only emulated for the first time in 1994, nearly half a century after he designed it. ³⁰

It couldn’t have happened much earlier. Serious processing power is required for a serial computer to loop through the automaton’s 6,329 cells over the 63 billion time steps required for the automaton to complete its reproductive cycle. Onscreen, it worked as advertised: a pixelated two-dimensional Rube Goldberg machine, squatting astride a 145,315 cell–long instruction tape trailing off to the right, pumping information out of the tape and reaching out with a “writing arm” to slowly print a working clone of itself just above and to the right of the original.

It’s just as possible, though similarly slow, for a serial computer to emulate a neural network, heir to Turing’s “unorganized machine.” That’s why it wasn’t practical to run really big neural nets like those in Transformer-based chatbots until recently, thanks to ongoing progress in the miniaturization, speed, and parallelism of digital computers.

A few collaborators and I set out to explore this question in late 2023. ³² Our first experiments made use of an esoteric programming language invented thirty years earlier by a Swiss physics student and amateur juggler, Urban Müller. I’m afraid he called this language … Brainfuck. Please direct all naming feedback his way.

However, the shoe fits; it is a beast to program in. Here, for instance, is a Brainfuck program that prints “helloworld”—and good luck making any sense of it:

++++++[−>+++++<]>−[>[++++>]++++[<]>−]>>>>.>+.<<…<−.<+++.>.+++.>.>>−.

The upside of Brainfuck is its utter minimalism. It’s not quite a single-instruction language, like SUBLEQ, but as you can see, it includes only a handful of operations. Like a Turing Machine, it specifies a read/write head that can step left (the “<” instruction) or right (the “>” instruction) along a tape. The “+” and “−” instructions increment and decrement the byte at the current position on the tape. ³³ The “,” and “.” instructions input a byte from the console, or output a byte to it (you can count ten “.” instructions in the code above, one to print each letter of “helloworld”). Finally, the “[” and “]” instructions implement looping: “[” will skip forward to its matching “]” if the byte at the current position is zero, and “]” will jump back to its matching “[” if the byte is nonzero. That’s it!

It’s hard to believe that Brainfuck could be used to fully implement, say, the Windows operating system, but—it is “Turing complete.” Here that means: given enough time and memory (that is, a long enough tape), it can emulate any other computer and compute anything that can be computed.

In our version, which we call bff, there’s a “soup” containing thousands of tapes, each of which includes both code and data. This is key: in “classic” Brainfuck, the code is separate from the tape, whereas in bff, we wanted the code to be able to modify itself. That can only happen if the code itself is on the tape, as Turing originally envisioned.

Bff tapes are of fixed length—64 bytes, just one byte longer than the cryptic “helloworld” program above. They start off filled with random bytes. Then, they interact at random, over and over. In an interaction, two randomly selected tapes are stuck end to end, and this combined 128 byte–long tape is run, potentially modifying itself. The 64 byte–long halves are then pulled back apart and dropped back into the soup. Once in a while, a byte value is randomized, as cosmic rays do to DNA.

But after millions of interactions, something magical happens: the tapes begin to reproduce! As they spawn copies of themselves and each other, randomness quickly gives way to complex order. The amount of computation taking place in each interaction skyrockets, since—remember—reproduction requires computation. Two of Brainfuck’s seven instructions (“[” and “]”) are dedicated to conditional branching, and define loops in the code; reproduction requires at least one such loop (“copy bytes until done”), causing the number of instructions executed in an interaction to climb into the thousands.

The code is no longer random, but obviously purposive, in the sense that its function can be analyzed and reverse-engineered. An unlucky mutation can break it, rendering it unable to reproduce. Over time, the code evolves clever strategies to increase its robustness to such damage.

This emergence of function and purpose is just like what we see in organic life at every scale; it’s why we’re able to talk about the function of the circulatory system, a kidney, or a mitochondrion, and how they can “fail”—even though nobody designed these systems.

Let’s pause and take stock of why this is so remarkable.

On an intuitive level, one doesn’t expect function or purposiveness to emerge spontaneously. To be sure, we’ve known for a long time that a modest degree of order can emerge from initially random conditions; for instance, the lapping of waves can approximately sort the sand on a beach, creating a gradient from fine to coarse. But if we were to begin with sand on a beach subject to random wave action, and came back after a few hours to find a poem written there, or the sand grains fused into a complex electronic circuit, we would assume someone was messing with us.

The extreme improbability of complex order arising spontaneously is generally understood to follow from thermodynamics, the branch of physics concerned with the statistical behavior of matter subject to random thermal fluctuations—that is, of all matter, since above absolute zero, everything is subject to such randomness. Matter subject to random forces is supposed to become more random, not less. Yet by growing, reproducing, evolving, and indeed by existing at all, life seems to violate this principle.

The violation is only apparent, for life requires an input of free energy, allowing the forces of entropy to be kept at bay. Still, the seemingly spontaneous emergence and “complexification” of living systems has appeared to be, if not strictly disallowed by physical laws, at least unexplained by them. That’s why the great physicist Erwin Schrödinger (1887–1961) wrote, in an influential little book he published in 1944 entitled What Is Life?, ³⁵

“[L]iving matter, while not eluding the ‘laws of physics’ as established up to date, is likely to involve ‘other laws of physics’ hitherto unknown, which, however, once they have been revealed, will form just as integral a part of this science as the former.”

The Second Law of thermodynamics holds that any closed system will rise in entropy over time, becoming increasingly disordered. A hand-powered lawn mower, for example, starts off as a beautifully polished machine with sharp helical blades, round wheels, and toothed gears, all coupled together on smoothly rotating bearings. If left out in the elements, the bearings will seize up, the blades will dull, and oxidation will set in. After enough time, only a heap of rust will remain.

Similarly, if you were to take a dead bacterium (which, though a lot smaller, is far more complicated than a push mower) and drop it into a beaker of water, its cell membrane would eventually degrade, its various parts would spill out, and after a while only simple molecules would remain, dispersed uniformly throughout the beaker.

The Second Law gives time its arrow, because the fundamental laws of physics in our universe are very nearly time-reversible. ³⁷ Strange, but true: Newton’s equations (classical dynamics), Maxwell’s equations (electromagnetism), Schrödinger’s equations (quantum physics), Einstein’s equations (special and general relativity)—all of these physical laws would work the same way if time ran in reverse. The difference between the past and the future is statistical.

Theoretically, nothing prevents the exact series of collisions from happening that would result in all of the energy being transferred to a single ball while leaving the others arrayed in a perfect triangle; but statistically, it’s vanishingly unlikely for such an ordered state to arise out of disorder—unless, perhaps, some mastermind set the balls in motion just so.

Although key thermodynamic concepts weren’t developed until the nineteenth century, an Enlightenment-era belief in a God who set the universe in motion just so arises intuitively from the apparent impossibility of order arising spontaneously out of disorder. ³⁹ In a mechanical, billiard-table universe where the laws of physics are inviolable and the laws of statistics seem to inexorably degrade any pre-existing order over time, it seems absurd that anything as complicated as life could arise spontaneously without some supernatural agent acting as “prime mover.” Only a God with exquisite foresight could have “initialized” the Big Bang such that, in the Earth’s oceans billions of years ago, simple organic molecules floating around apparently at random could coalesce into a working bacterium—an improbability many, many orders of magnitude less likely than fifteen out of sixteen whizzing billiard balls spontaneously coalescing into an unmoving triangle.

The billiard ball universe I’ve just described may seem abstract or arbitrary, but nineteenth-century theorists like Boltzmann had become interested in this problem for the most practical of reasons: it was the physics behind steam power, hence the entire Industrial Revolution. Engineering had preceded theory, as it often does.

Thomas Newcomen (1664–1729), an English inventor and Baptist lay preacher, devised the first practical fuel-burning engine in 1712. It was based on a heating-and-cooling cycle. First, steam from a boiler was allowed to flow into a cylindrical chamber, raising a piston; then, the steam valve closed, and a second valve opened, injecting a jet of cold water, causing the steam to condense and pull the piston back down. As the piston rose and fell, it rocked a giant beam back and forth, which, in Newcomen’s original design, was used to pump water out of flooded mines (which, in turn, supplied the coal these engines would soon be consuming so voraciously).

Scottish inventor and entrepreneur James Watt (1736–1819) greatly improved the steam engine design in 1776, making it a practical replacement for human and animal power across a wide range of applications. This was when the Industrial Revolution really got underway; for the first time, human-made machines began to metabolize on a large scale, “eating” complex organic molecules to perform mechanical work.

Soon, far more energy would be flowing through this artificial metabolism than through the Krebs cycle in our own bodies. ⁴⁰ It was, in the broad sense I’ve been using in this book, a major evolutionary transition: a symbiogenetic event between humans and machines, like earlier symbioses between humans and draft animals. Machine metabolism allowed human society to explode from its pre-industrial scale (about one billion people in 1800, most of whom lived in extreme poverty) to its scale today (eight billion, most of whom no longer live in poverty). ⁴¹ In the process, we’ve become nearly as dependent on machines for our continued existence as they are on us.

However, even as coal-powered engines transformed the Victorian landscape—both figuratively and literally, for the pollution was dire—nobody understood them at a deep level. What was heat, and how could it be converted into physical work? For a time, the leading theory held that heat was a kind of invisible, weightless fluid, “caloric,” that could flow spookily into and through other matter.

By combining Newtonian physics, statistical calculations, and experimental tests, Boltzmann and his colleagues figured out what was really going on. Their conceptual setup was a three-dimensional version of the billiard table, in which the balls were gas molecules whizzing around in a pressurized chamber, bouncing off its walls. Calculating the average effect of all the bouncing led to the Ideal Gas Law, which established theoretical relationships between the pressure, volume, and temperature of a gas. The theory closely matched observations by experimentalists and engineers. ⁴²

Volume is straightforward (that’s just the size of the chamber), but the idea that pressure is the aggregate force on the chamber’s walls as molecules bounce off them, and that temperature is the average kinetic energy ⁴³ of those whizzing molecules, was a profound insight. There was no need for any mysterious “caloric” fluid; heat was just motion on a microscopic scale. And the tendency toward a random distribution of molecular positions and momenta explains why, if you open a valve between two chambers containing gas at different pressures and/or temperatures, those pressures and temperatures will quickly equalize.

Before moving beyond classical thermodynamics, let’s add a bit more realism to our billiard-ball universe. We know that balls bouncing around a billiard table don’t actually go on bouncing forever. They encounter friction, slowing down as they roll. And when they bounce off each other, the collisions are slightly “inelastic,” meaning that after the collision, they’re moving a bit slower than before. After a little while, they stop rolling.

How can that be? At a microscopic level, the laws of physics are reversible. Momentum is supposed to be conserved. And the amount of matter and energy also remains constant, whether we run time forward or in reverse. That’s the First Law of thermodynamics!

Zooming in will reveal that, on a real billiard table, balls aren’t the smallest elements that bump against one other. Each billiard ball consists of lots of vibrating molecules bound together, and collisions between these individual molecules are what really cause the balls to bounce off each other—or, for that matter, cause a ball to roll across the felt rather than falling right through it. ⁴⁴ In each case, momentum is transferred between molecules. Every time this happens, the distribution of molecular momenta becomes a bit more random, that is, a bit less correlated with which ball the molecule happens to be in, or indeed whether it is in a ball at all, or in the felt underneath.

In the most random distribution of molecular velocities, there would be no more correlation between the velocities of two molecules in one ball than in the velocities of molecules in different balls. Every ball would be imperceptibly jiggling in place, with each of its constituent molecules contributing minutely to the dance. We call that random jiggling “heat.” When all correlated motion has been converted into uniformly distributed heat, we’ve reached a stable equilibrium.

While the balls are still rolling, the correlations between the velocities of their molecules are by no means all equal; the distribution is far from random. This is not a stable equilibrium. Hence, the inevitability of friction and the inelasticity of collisions are statistical phenomena—just more symptoms of the inexorable Second Law.

Going forward, then, we can zoom back out and imagine once more that the billiard balls are indivisible particles, not bound collections of molecules. In that case, all collisions would have to be elastic, all motion frictionless, and disordered, randomly colliding balls would be the equilibrium.

Once a system has reached equilibrium, it will stay that way forever—an end state Lord Kelvin called “heat death.” ⁴⁵ This seemingly trivial observation has some profound consequences. One is that the arrow of time will lose its meaning; any two consecutive moments A, B could just as likely have been ordered B, A. Nothing, therefore, can be said to be a cause, versus an effect.

Relatedly, no work can be done. If the system were out of equilibrium—for instance, if all of the whizzing balls were on one side of the billiard table—then we could put a movable barrier down between the empty and occupied sides, attached to a loaded crankshaft. As they bounce around, the balls would then nudge the barrier, doing work. What I’ve just described is, of course, a piston, like that of a steam engine.

But if the balls were equally likely to be anywhere, then no matter how fast they whizz and bounce, there’s nowhere to put the barrier that would result in any net force. The piston wouldn’t move because it would be buffeted equally from all sides. This idea can be generalized: work can only be done by a system in disequilibrium, for instance, when the pressure or temperature is high in one place, and low in another. That’s why Newcomen’s engine needed both hot steam and cold water.

I’ve already used the term free energy, but now we can define it. The free energy of a system is the amount of work it can be made to do. Far from equilibrium, when the entropy is low, much of the kinetic energy in the billiard balls is “free”; it can be used to move pistons, raise weights, produce electric currents, carry out computations, or drive metabolic processes. But at equilibrium, the entropy is maximized, and the free energy is zero. This insight into the relationship between energy, entropy, and work lies at the heart of thermodynamics—and life.

They will even be evolving. If the beaker is sufficiently large—the size of a planet, for instance—and we wait a few billion years, then eventually beings as complicated as us may be in there, along with cities, advanced technologies, and perhaps plans to colonize the next beaker.

None of these processes can occur without free energy. For us, it comes, ultimately, from the sun. Thermodynamics tells us that even if the Second Law appears to be violated locally, it still holds when we zoom out. Order created in one place comes at the expense of increased disorder elsewhere. Hence, pollution, the finite lifetime of the sun, and the eventual heat death of the universe.

What concerns us here isn’t this big picture, but its apparent local violations, and the way they seem to become increasingly transgressive over time. The puzzle isn’t only that bacteria exist, but that the more time passes, the more complex life on Earth seems to become: from prokaryotes to eukaryotes; from eukaryotes to multicellular animals; from simple multicellular animals to ones with nervous systems; from brainy animals to complex societies; from horses and plows to space travel and AI.

Is there any general principle behind that complexification process, a kind of “however” or “yes, and” to the dismal Second Law? And could it account not only for evolution and complexification, but also for abiogenesis?

Yes, and yes. Bff can offer us a highly simplified model system for understanding that principle, just as an idealized billiard table gives us a model for understanding basic thermodynamics.

Replicators arise in bff because an entity that reproduces is more “dynamically stable” than one that doesn’t. In other words, if we start with one tape that can reproduce and one that can’t, then at some later time we’re likely to find many copies of the one that can reproduce, but we’re unlikely to find the other at all, because it will have been degraded by noise or overwritten.

Addy Pross, a professor emeritus of chemistry at Ben Gurion University of the Negev, describes the same phenomenon using the bulkier phrase “dynamic kinetic stability” (DKS). ⁴⁶ I’ll drop “kinetic,” since the idea also applies beyond Pross’s field of “chemical kinetics” (describing the rates at which chemical reactions take place). In bff, for example, dynamic stability can just as well apply to programs or program fragments.

As Pross points out, a population of molecules capable of replicating can be more stable than even the hardiest of passive materials. A passive object may be fragile, like a soap bubble, or robust, like a stone sculpture. The sculpture might endure for longer, but, in the end, it’s still ephemeral. Every encounter it has with anything else in the world will cause its composition or structure to degrade, its individual identity to blur. For a sculpture, it’s all downhill. That’s the Second Law at work, as usual.

A self-reproducing molecule—like the DNA inside a living bacterium—is another matter. It is thermodynamically fragile, especially if we consider its identity to consist not only of a general structure but of a long sequence of specific nucleotides. However, its pattern is not just robust, but “antifragile.” ⁴⁷ As long as DNA is able to reproduce—an inherently dynamic process—that pattern can last, essentially, forever. A bit of environmental stress or adversity can even help DNA maintain or improve its functionality. This is how order overcomes disorder.

In fact, Darwinian selection is equivalent to the Second Law, once we expand our notion of stability to include populations of replicators. Through a thermodynamic lens, Darwin’s central observation was that a more effective replicator is more stable than a less effective one. As Pross puts it,

“[M]atter […] tends to become transformed […] from less stable to more stable forms. […] [T]hat is what chemical kinetics and thermodynamics is all about […]. And what is the central law that governs such transformations? The Second Law. […] In both [the static and kinetic] worlds chemical systems tend to become transformed into more stable ones […]—thermodynamic stability in the ‘regular’ chemical world, dynamic kinetic stability in the replicator world.” ⁴⁸

As a chemist, Pross is sensitive to the close relationships between energy, entropy, and stability, whether static or dynamic. However, he does not explicitly make a connection to the theory of computing.

It now seems clear that by unifying thermodynamics with the theory of computation, we should be able to understand life as the predictable outcome of a statistical process, rather than regarding it uneasily as technically permitted, yet mysterious. Our artificial life experiments demonstrate that, when computation is possible, it will be a “dynamical attractor,” since replicating entities are more dynamically stable than non-replicating ones; and, as von Neumann showed, replicators are inherently computational.

Bff has no concept of energy, but in our universe, replicators require an energy source. This is because, in general, computation involves irreversible steps—otherwise known as causes and effects—and thus, computing consumes free energy. That’s why the chips in our computers draw power and generate heat when they run. (And why my computer heats up when it runs bff.) Life must draw power and generate heat for the same reason: it is inherently computational.

Where did all this complexity come from? It certainly doesn’t look like the result of simple Darwinian selection operating on the random text generated by a proverbial million monkeys typing on a million typewriters. ⁴⁹ In fact, such complexity emerges even with zero random mutation—that is, given only the initial randomness in the soup, which works out to fewer bytes than the text of this short book. Hardly a million monkeys—and far too few random bytes to contain more than a few consecutive instructions, let alone a whole working program.

The answer recalls Lynn Margulis’s great insight: the central role of symbiosis in evolution, rather than random mutation and selection. When we look carefully at the quiescent period before tapes begin replicating, we notice a steady rise in the amount of computation taking place. We are observing the rapid emergence of imperfect replicators—very short bits of code that, in one way or another, have some nonzero probability of generating more code. Even if the code produced is not like the original, it’s still code, and only code can produce more code; non-code can’t produce anything!

Thus, a selection process is at work from the very beginning, wherein code begets code. This inherently creative, self-catalyzing process is far more important than random mutation in generating novelty. When bits of proliferating code combine to form a replicator, it’s a symbiotic event: by working together, these bits of code generate more code than they could separately, and the code they generate will in turn produce more code that does the same, eventually leading to whole-tape replication and an exponential takeoff.

However, evolution’s creative work is not done yet. After the takeoff of a fully functional tape replicator, we often see yet further symbiotic events. From a classical Darwinian standpoint, this seems puzzling, since there should be no reason for further evolution to take place once whole tapes are replicating reliably. How could “fitness” possibly improve further, once a tape copies itself in its entirety every time it interacts with another tape?

We must consider that since the instructions for whole-tape replication don’t occupy all 64 bytes, there’s extra space on the tape that could be dedicated to … anything. That’s the point of von Neumann–style replication—it allows for open-ended evolution precisely because the tape can contain additional information, beyond the code needed for replication itself. ⁵²

Any extra replicated bytes could, of course, be random—just passive, purposeless information cargo hitchhiking from one generation to the next. But if these bytes contain instructions, those instructions can run. And if they can run, they can replicate themselves, too. Thus, the symbiogenetic process can continue to operate, creating additional replicators within an already replicating tape. Sometimes these sub-replicators even produce multiple copies of themselves in a single interaction.

Sub-replicators can interact with their host in many ways. They can “kill” the host by damaging its replication code, which is generally catastrophic for the sub-replicator, as it thereby destroys the environment within which it can run. Sub-replicators can be neutral, leaving the host’s replication machinery alone. Or, they can be symbiotic, for instance by conferring resistance to mutational damage via redundant copying of the host’s code. The overall tendency is toward symbiosis, since that is the most dynamically stable.

Over time, code colonizes a large proportion of the 64 bytes. Code is more dynamically stable than non-code, and its dynamic stability increases through symbiosis with yet more code—in particular, when code fragments find ways to work in functional tandem.

In a way, symbiosis is the very essence of functionality. When we talk about a kidney’s function only making sense in context, we mean that it is in symbiosis with other functions—like those of the liver (breaking ammonia down into urea), the heart (pumping blood), and so on. Each of these functions is purposive precisely because its inputs are the outputs of others, its outputs are inputs to others, and thus they form a network of dynamically stable cycles.

The same is true of larger, planetary-scale interrelationships. The “purpose” of plants, from the perspective of animal life, is to produce oxygen and sugar, which we breathe and eat. The “purpose” of animals, from a plant’s perspective, is to turn the oxygen back into carbon dioxide, and provide compost and pollination. Our growing understanding of life as a self-reinforcing dynamical process boils down not to things, but to networks of mutually beneficial relationships. At every scale, life is an ecology of functions.

Because functions can be expressed computationally, we could also say that life is code, and code is life. Individual computational instructions are the irreducible quanta of life—the minimal replicating set of entities, however immaterial and abstract they may seem, that come together to form bigger, more stable, and more complex replicators, in ever-ascending symbiotic cascades.

In the toy universe of bff, the elementary instructions are the seven special characters “<>+−,[]”. On the primordial sea floor, geothermally-driven chemical reactions that could catalyze further chemical reactions likely played the same role. Under other conditions, on another planet, or in another universe, many different elementary interactions could do the same—as long as they are Turing complete, enabling them to make the leap from autocatalytic sets to true replication.

As we’ve seen, some of bff’s predictions are quite different from those of classical Darwinian theory:

1. Bff suggests that symbiogenesis is a more important driver of evolutionary innovation than random mutation.
2. Since symbiogenesis must involve combinations of pre-existing dynamically stable entities, we should expect complex replicating entities to emerge after (and be made of) simpler ones.
3. As a result, zooming in on sub-replicators within a larger replicator should allow us to peer back in evolutionary time. ⁵³

These phenomena should sound familiar! They’re all consistent with Lynn Margulis’s observations, which flew in the face of twentieth-century biological orthodoxy, but are now, if not mainstream, at least gaining respectability.

But wait, there’s more:

4. Due to the instability of the imperfect replicators leading up to the first true replicator, we should expect this first true replicator to be a historical “event horizon,” becoming the template for what follows and erasing independent traces of what came before.

This is what we see on Earth too. Although the first chemical steps toward proto-life may still be taking place in environments like black smokers, we don’t see the missing links. Where are the kind of imperfect replicators that fused together to form the simplest life forms today?

We likely don’t see them on their own because, as in bff, their instability made them ephemeral, quickly displaced or absorbed by the first stably replicating cell—which might already have been recognizably kin to today’s bacteria and archaea. (Evolutionary biologists call this the “Last Universal Common Ancestor” or LUCA.) Still, we can assume that many of the component parts of the most ancient surviving life forms—like the reverse Krebs cycle and RNA replication—are fossilized fragments of earlier imperfect replicators.

In the same vein:

5. Evolved code should not only include instructions for replicating itself as a whole, but also be rife with sub-sequences that contain instructions for independently replicating themselves.
6. If symbiosis among these parts generated the novelty driving evolution of the whole, we should see evidence in the genome of many “broken” or incomplete sub-replicators.
7. Code that evolved through such hierarchical symbiotic replication should also contain many sequences that are repetitive, or are copies of other parts.

We can find all of these features in our own genetic code—and they don’t correspond to what we would expect to see if our genomes had evolved primarily through mutation and selection.

When the first complete human genome sequence was published in 2000, some surprises were in store. ⁵⁴ One was the astonishingly high proportion of so-called “junk DNA” that doesn’t code for proteins: about 98%. How did it get there? Is it actually doing anything? We now understand that some of this “junk” is involved in gene regulation, but most is still of unknown or uncertain function. Close inspection reveals, though, that much of our code, whether “junk” or otherwise, is … viral.

The reproductive cycle of viruses is an often-told story. When traveling between cells, a virus looks like a bit of DNA or RNA packaged into a protein “envelope” capable of binding to a target cell and injecting its genetic payload. Once inside the cell, the viral code hijacks cellular resources to begin replicating, both copying its genetic material and manufacturing its envelope proteins. These assemble into more viruses. Maniacally cranking out viral proteins, the cell may eventually burst, releasing a flood of virus particles to repeat the cycle.

“Retroviruses” like HIV (the Human Immunodeficiency Virus, which causes AIDS) include additional “reverse transcription” machinery for permanently incorporating their genetic material into the cell’s DNA. This makes them a lot harder for the host to clear. The key enzyme, “reverse transcriptase,” wasn’t characterized until 1970, ⁵⁵ but Barbara McClintock (1902–1992), working at Cold Spring Harbor Lab on Long Island, had made an earlier discovery that turned out to be closely related.

While studying the genetics of maize plants in the late 1940s and early ’50s, McClintock found mobile segments of DNA that seemed to be able to cut and paste themselves into different locations on the genome, producing differently mottled color patterns in corn kernels. ⁵⁶ Although many of her colleagues were initially skeptical about these “transposons” or “jumping genes,” McClintock eventually won the Nobel Prize for this work.

A much wider zoo of “transposable elements” has since been discovered, some of which don’t just “cut and paste,” but “copy and paste” themselves; they are, in effect, replicators within our DNA. In fact, some of them are fully fledged retroviruses, including instructions not only for splicing their code into a host cell’s genome, but also for envelope proteins, allowing that code to venture out into the world in search of other cells to infect.

Retroviruses are unsettlingly … intimate. HIV specifically targets immune cells, but if a retrovirus infects an egg or sperm cell, it can insert its code into an organism’s germ line, becoming a permanent part not only of that cell, but of an entire species. In a 2006 publication in the prestigious journal Nature, researchers at the University of Queensland in Australia reported catching a retrovirus in this act of becoming “endogenous” for the first time. ⁵⁷ An epidemic of leukemia in koalas had first been traced to a retrovirus, then this retrovirus was found to have invaded the koala germ line over the previous century, causing the disease to become heritable.

Does this mean the end of koalas? Probably not.

The human genome is rife with signs that the same process has taken place many times in the history of our own species. At least 8% of our genome consists of endogenized retroviruses, the remnants of such retroviral invasions. ⁵⁸ Remember, 8% is several times more DNA than codes for our “own” genes!

We usually think of viruses as mere disease agents, opportunistically parasitizing our cells to reproduce, since they have no reproductive machinery of their own. The reality may be very different, though. As many of us know from our recent experience with COVID, ⁵⁹ viruses (and pathogens in general) tend to evolve into forms less lethal to their hosts over time, for obvious reasons—a dead host is a dead end for the pathogen, too. There’s mounting evidence that the relationship between virus and host can go even further, to become symbiotic. Retroviral code in our genome, for instance, has become fundamental to the formation of the placenta, the immune system, cell differentiation, and brain function. ⁶⁰

Moreover, endogenous retroviruses are only the tip of an even larger iceberg. Nearly half of our genome consists of transposable elements of one kind or another. Some were probably once retroviruses, or vice versa. Fully sequenced genomes are often rife with lengthy stretches of highly repetitive sequences, from long and complex to mere alternations of two or three symbols—evidence of sub-replicators running amok.

Our genomes, in other words, are not only reproduced as a whole, but include working reproductive sub-sequences at many scales. Some are new and can still cause disease; others are older and may have lost the ability to make viral envelopes or the other machinery needed to become infectious; and yet others have integrated themselves so deeply into our own code that they are no longer distinct. Some of this code is even serving critical functions in our bodies. It’s obvious that, at least by this last stage of the process, complete symbiosis has been established: neither the host nor the sub-replicator could survive, let alone reproduce, on its own.

In his 2009 book Virolution, ⁶¹ author and physician Frank Ryan goes further, arguing that viruses may have been symbionts all along. Plague viruses like HIV don’t come from nowhere; they are species jumpers. We know that HIV was originally SIV, the Simian Immunodeficiency Virus, variants of which are endemic to Old World primates including African green monkeys, sooty mangabeys, mandrills, and chimpanzees. Yet many of these viruses don’t sicken their original hosts. ⁶²

We also know that if a virus finds itself inside an organism whose physiology is too different from that of its original host, it can’t gain purchase—a lucky thing for you, if you’ve ever swallowed a mouthful of seawater, which likely contained about a billion virus particles! (Most would have targeted single-celled marine life.) The greatest danger seems to come from viruses adapted to a different but closely related species. Perhaps, when it kills, such a virus is doing its job: wiping out rivals who have invaded the original host’s territory.

Viruses could, in other words, work like an out-of-body immune system. Within our bodies, our immune systems seek out and destroy cells that are recognized as “not-us.” Outside our bodies, “our” viruses could be similarly seeking out and destroying whole animals who are recognized as “not-us.” Once a virus becomes endemic to a new population, though—even if it initially kills many—it will differentiate, co-adapt, and perhaps eventually go native.

One could consider bacterial pathogens through a similar lens; hence the well-documented plagues of smallpox that European colonists brought to the Americas, decimating Native populations, and the virulent (though less deadly) syphilis epidemic believed to have been brought back to Europe by Columbus. ⁶³ Unlike bacteria, though, retroviruses don’t just take up residence in our environment or in our bodies, but fuse into our very genomes, becoming inextricably part of us—especially when they alter the germ line.

By sheer volume, our DNA appears to be made primarily of the layered remnants of many such past fusions. It seems clear that transposable elements and “endogenous viral elements” or EVEs have done much more editing of our genome in recent evolutionary history than mutation has. ⁶⁴ And they have been at this for a long, long time. Based on the best available evidence, viruses are at least as old as the Last Universal Common Ancestor, if not older. ⁶⁵

This ongoing ecology of mobile, self-reproducing, and even infectious genes causes the supposed “tree of life” to continue entangling with itself in the oddest ways. Fully a quarter of the cow genome, for instance, consists of copies of the retrotransposon BovB. BovB appears to have leapt many times between species to create a phylogenetic tree of its own, a “bizarre parallel universe where cows are more closely related to snakes than to elephants, and where one gecko is more closely related to horses than to other lizards.” ⁶⁶ These are instances of “horizontal gene transfer” (HGT), long associated with bacteria, but clearly more universal. ⁶⁷

Whether a bit of new DNA is acting in the moment as friend, foe, or somewhere in between is really just a matter of where on the symbiotic trajectory it falls. When brand new, it’s unlikely to be friendly. But if it has persisted for a long time, and especially if it has become endogenous, its dynamic stability implies that it will have become non-lethal or even friendly due to combined evolutionary pressures on host, invader, and “host plus invader” as a “holobiont” or symbiotically fused entity. When some gene or other functionality in the fused code proves valuable to the larger whole, that functionality will be conserved, even as the sub-replicator’s reproductive capability degrades and eventually vanishes. Perhaps that is how new genetic functionality arises in general.

In this light, random point mutation can even be seen as something like a minimal abiogenesis event, creating a tiny parasitic “life form” within the genome. On its own, it would of course be unable to reproduce, since it possesses no independent reproductive machinery; ⁶⁸ then again, neither does a virus. So, like a virus, a mutation can reproduce—using the host’s resources, which will copy it along with everything else. When introduced into the genome of a sophisticated existing organism, such a tiny, random entity is unlikely to be friendly, though, as with any larger invasion, it will either kill its host (and itself, in the bargain) or achieve dynamic stability, either by being neutral or (occasionally) helpful.

Hence, just as replicator thermodynamics encompasses classical thermodynamics as a special case, evolution via symbiogenesis can encompass classical Darwinian theory as a special case. Point mutation is unlikely to be the main driver of evolution once life has taken off because it’s so much weaker and slower on its own than higher-order symbiogenesis. Unlike point mutation, a chunk of code that has already circulated, jumping around in the genome or even between species, isn’t random. It necessarily includes real functionality. And at least in some settings, that functionality has been under evolutionary pressure to help, or at least not kill, its host. Evolution picks up steam over time, with major evolutionary transitions becoming more frequent precisely because increasingly high orders of symbiogenesis become possible.

French-American mathematician Benoit Mandelbrot (1924–2010) is the figure most associated with fractals, though they were first described by English meteorologist Lewis Fry Richardson (1881–1953). In the aftermath of World War II, Richardson, a Quaker, had been trying to build a mathematical model to predict the likelihood of conflicts between countries. One of the parameters in his model was the length of their common border. He was perplexed to find, though, that certain borders didn’t appear to have agreed-upon lengths. The Portuguese, for instance, reported that their border with Spain was 613 miles long, but the Spanish claimed that it was 754 miles long. This seemed like a weirdly large discrepancy for what ought to have been an easily surveyed distance!

Living organisms exhibit obvious fractal properties in certain of their structures, such as tree branches, bronchial tubes, and circulatory systems. In structures like these, fractal geometry over some range of scales offers an elegant design solution to a biophysical problem. For instance, blood carrying oxygen and nutrients needs to reach every cell in the body, which means that oxygenated blood has to be pumped through a branching network of tubes, and these tubes—which are, in Euclidean terms, one-dimensional—have to repeatedly subdivide to “fill” three-dimensional space. Hence, we have fractally branching (and then merging) networks of arteries, arterioles, capillaries, venules, and veins.

By working out the design properties of such fractal networks and their implications, theoretical physicist Geoffrey West and his collaborators at the Santa Fe Institute have figured out many fundamental “scaling laws” in biology. ⁷⁰ One of the most well-known results from this body of work explains why every mammalian heart will beat about 1.5−2 billion times over a lifetime, whether it belongs to an Etruscan shrew (weighing a twentieth of an ounce, and living a year or so), or a blue whale (weighing over 300,000 pounds, and living about a century).

Self-similarity in general yields “power law” scaling relationships, which look like straight lines when plotted on logarithmic axes. ⁷¹ Body mass, for instance, has power-law relationships with metabolic rate, the total length of the circulatory system, and lifespan. ⁷² Combining various power-law relationships to calculate the number of heartbeats in a lifetime causes all the scaling variables to cancel out beautifully, yielding the famous 1.5−2 billion beat constant. While the theory is not exact (humans, for example, live somewhat longer than predicted), it does a remarkable job of parsimoniously explaining a wide range of phenomena.

Similarly, although genomes are always replicated, and are made out of other genomes that are themselves replicated, every whole is different from its parts, just as an integrated circuit is different from a transistor. How, then, would one go about quantifying this multifractal-like property in a genome?

The answer is closely related to data compression. Consider bff. In its initial state, the tapes consist of random bytes; no repetitive structure exists. If we were reading through the tapes one byte at a time, the next byte would have an equal probability of assuming any value, from 0 to 255, independent of any of the previous bytes. Therefore, the sequence is entirely unpredictable, which means that if you were to try compressing the soup with file compression software, like ZIP, the compressor wouldn’t be able to squeeze the file size at all. (This is, harkening back to thermodynamics, a working definition of a completely disordered state.)

Data are redundant—that is, compressible—precisely to the degree that they are predictable, or, equivalently, exhibit order. Compression algorithms like those in ZIP work by looking for patterns in the data stream so far and exploiting those patterns to encode subsequent data using as few bits as possible. The usual trick is to incrementally build a dictionary or “phrase book” out of the source material, and to substitute sections of the text for pointers into this dictionary whenever the pointer can be encoded in fewer bits than the original text. The dictionary is, in effect, a predictive model, and since the compressor builds that model incrementally as it compresses, the decompressor is able to incrementally reconstruct an identical copy of the model as it decompresses.

Absent any prior knowledge about how to model a long sequence, describing it in terms of previously encountered parts of itself is a good strategy. The full text of this book ZIPs down to about 35% of its original size as a raw text string, since some letters occur more often than others, there are a lot of repeated words, and those words often combine into stock phrases. ⁷⁸

When bff undergoes its sharp transition to whole-tape replication, the tapes suddenly become highly compressible; they squeeze down to just 5% of their original size. Interestingly, many regions of the human genome are highly compressible too. The reason these data streams compress so well is that, at any given point in the sequence, one can often make a pretty good guess as to what the next letter will be, simply by looking for the longest previous sequence that matches the current context.

Such matches arise due to replication. In the human genome, we’re not talking about replication at the level of the individual, but, rather, at the level of transposable elements (or their fossil remnants) within our genome. Similarly, even a single bff tape is highly compressible, because it formed through a symbiotic fusion of smaller replicators … and these replicators are themselves likely to be either copies of each other or related, having formed out of yet smaller imperfect replicators … which are themselves either copies or related.

Like fractals, sequences of symbols constructed by repeatedly copying sub-sequences of all lengths exhibit power-law scaling—in this case, between the lengths of repeated sequences and their frequency. For instance, a replicated sequence twice the length of another might be found half as often—which is pretty often! (These statistics would be dramatically different for random sequences, where increasing the sequence length would cause a much faster exponential decline in expected frequency. ⁷⁹ ) Sequences constructed according to a fixed self-copying rule are, in a sense, infinitely compressible, for they contain only as much information as it would take to write down the recipe, no matter how long the sequence.

By contrast, in a complex replicating system like bff—let alone the genome—the way any particular set of replicating sub-sequences combine to form a larger replicating sequence will always be specific. It will depend on the functional particulars of how those parts come together to make a working whole. Therefore there is novelty, or information, in each such combination. However, that information will be much less than would be needed to specify an arbitrary symbol sequence of the same length.

When we look for this property in bff and in human genome data, that’s just what we find! Not only do both compress very well; they also compress better and better the longer the stream of data gets. For instance, a single bff tape compresses somewhat, but the whole soup compresses better, and as the size of the soup grows, the compression ratio continues to improve—though no matter how much has already been seen, the remainder is never fully predictable.

Similarly, while a single stretch of a human genome already compresses well, the whole genome compresses better. If you compress it in the statistical context of many other human genomes, so much of the information becomes redundant that a compressed version of your genome would be small enough to send as an email attachment. ⁸⁰ The pattern continues if we zoom yet farther out to consider genomes across species. Hence the famous statistic that the human genome is 98% identical to the chimpanzee genome … and 60% similar to that of the fruit fly!

While common ancestry alone can explain many of the similarities across species (though not all, as BovB illustrates), such classical Darwinian arguments can’t explain why the genome of a single individual is also so repetitive. That only makes sense when we begin to consider genomes to be made out of smaller genomes.

DNA and bff tapes, in other words, are both systems that evolve through what we could call “multifractal symbiosis.” Our genomes are aggregates of cooperating replicators, all the way down. That’s why they are so compressible. But it’s also why wonderful new kinds of complexity emerge at every scale—and therefore why, in biology, we can always learn something new by zooming in or out.

Embodiment

There’s a profound yet subtle relationship between the multifractal properties of our bodies and the multifractal properties of the genome.

The relationship is subtle because the genetic code is by no means a one-to-one blueprint explicitly representing the body’s final shape. One might imagine, for instance, that because a snake has hundreds of ribs, it might have a stretch of DNA coding for a vertebra and rib pair, which might have replicated itself a few hundred times in the snake’s genome. Could snakes be the result of a retrotransposon coding for ribs run wild, like BovB?

Not exactly. So-called “Hox” genes, shared widely among animals, control overall body plan, and they “execute” their program using a combination of gene regulation and the kind of distributed computation Turing described in his work on morphogenesis. During embryonic development, ribs begin as a chemically controlled spatial oscillation, like the stripes of a tiger. This mechanism is vastly more powerful, general, and evolvable than merely replicating some “build a rib” code. It allows for code reuse, just as a programmer would do: invoking the same “build a rib” function a hundred times instead of copying and pasting it a hundred times. Thus, tweaking rib flexibility or curvature involves making one change, not a hundred.

Evolution can produce real programs that reuse code because life is computational. Remember that if you’re a replicator of the kind von Neumann described, you need to contain instructions for building yourself, including a machine B to copy those instructions, and a machine A, the “universal constructor,” to follow them. The universal constructor is the engine not only of reproduction, but also of growth, development, healing, and maintenance. It created your circulatory system and skin during embryonic development; it also heals your skin and replaces the torn capillaries underneath when you get a cut. As long as you are alive, your body will be continuously constructing and reconstructing itself. The code never stops running.

And since a universal constructor is a kind of Universal Turing Machine, its computations, or equivalently mathematical functions, are inherently “compositional”: complex functions (or programs) can be made out of sub-functions, which can in turn be made out of sub-sub functions … until, at bottom, everything could be expressed in terms of the elementary instructions of a Turing machine.

Functions can even be made out of themselves. In computer science, this is known as “recursion.” The Fibonacci sequence, for example—1, 1, 2, 3, 5, 8, 13, 21, and so on, where each number is the sum of the previous two—is most easily expressed as the recursive function f(n)=f(n−1)+f(n−2) for numbers greater than 1, with f(0)=f(1)=1. We find Fibonacci numbers in biology all the time, as in the spiral pattern of a pine cone, the head of a sunflower, and the chambers of a nautilus. This is a clue that recursive computation takes place in the construction of pine trees, sunflowers, and nautiluses. ⁸¹

The smoking gun, though, is the growth of self-similar structures in the body. Compositionality and recursion are what allow genetic code to build a many-tiered fractal like the circulatory system. Code for building ribs or nautilus segments could, in theory, just be copied a couple of hundred times, but that’s not possible for arteries, veins, or capillaries, where the number of segments and branches becomes astronomically large. The blood vessel branching code must get reused. And with relatively minor tweaks, it must be possible for the parameters of that code to be adjusted so that the branching stops at the appropriate size for a shrew, or goes on to construct something the size of a blue whale, since shrews and whales are, in the grand scheme of things, close relatives.

From an evolutionary perspective, compositionality implies a hierarchical network of symbiotic relationships. The bacteria that became mitochondria and the archaea that engulfed them each started with the code necessary to reproduce themselves. When they merged into a eukaryote, they still needed to retain the instructions for their own reproduction, but they also needed to evolve additional functionality for mutual regulation. When eukaryotes became multicellular, the same needed to happen again; our cells still know how to reproduce individually, and, indeed, cellular reproduction is a fundamental operation involved in the larger-scale growth and reproduction of a whole animal.

What is true of evolution is also true of development. ⁸² Blood vessels, for instance, aren’t just Euclidean line segments, but tubes made of layers of smooth muscle cells. Each of those cells contains a complement of organelles, and each mitochondrion in each of those cells contains its own membranes and loop of bacterial DNA. A living organism is a compositional structure par excellence. It could only be built computationally, through the composition of many functions. And life could only have evolved as the hierarchical composition of those functions—that is, through symbiogenesis.

They are certainly purposive, or we couldn’t talk about them being broken or buggy. But hardware and software are, in general, unable to reproduce, grow, heal, or evolve on their own, because engineers learned long ago that self-modifying code (like bff, or DNA) is hard to understand and debug. ⁸⁴ Thus, phones don’t make baby phones. Apps don’t write new versions of themselves.

And yet: there are more phones in the world this year than last year; apps acquire new features, become obsolete, and eventually reach end-of-life, replaced by new ones; and AI models are improving from month to month. Electronic components and computer code also exhibit the same kind of compositionality we’ve seen in bff and DNA. It certainly looks as if technology is reproducing and evolving! Debating its aliveness is thus a bit like the debate over whether viruses (which also can’t reproduce on their own) are alive.

If we zoom out, though, putting technology and humans in the frame together, we can see that this larger, symbiotic “us” is certainly reproducing, growing, and evolving. The emergence of technology, and the mutually beneficial—if sometimes fraught—relationship between people and tech is nothing more or less than our own most recent major evolutionary transition. Technology, then, is not distinct from nature or biology, but merely its most recent evolutionary layer.

And what about that age-old inanimate stuff—rocks and rivers, mountains and beaches, clouds and storms? Water molecules in themselves are clearly not capable of general computation, and yet, in the context of the hydrologic cycle, clouds, rainstorms, and rivers certainly serve critical ecological functions, and are profoundly shaped by life. Likewise, our planet’s metal and sand get shaped into steam engines and computer chips. All of these are part of that grand network of interdependency we call Earth. Why do we draw boundaries around certain networks of functions and insist that they are “alive,” while the surrounding functions are not?

This way lies vitalism, a view espoused in various forms by a long line of philosophers from Posidonius of Apameia (circa 135–51 BCE, and undoubtedly reflecting a much older tradition) to Henri Bergson (1859–1941). Some modern thinkers, too, defend the vitalist position, such as Jane Bennett:

“The quarantines of matter and life encourage us to ignore the vitality of matter and the lively powers of material formations […]. By ‘vitality’ I mean the capacity of things—edibles, commodities, storms, metals—not only to impede or block the will and designs of humans but also to act as quasi agents or forces with trajectories […] or tendencies of their own. [… Our] analyses of political events might change if we gave the force of things more due.” ⁸⁵

We resist such ideas because we tend to reserve the notion of agency only for ourselves. The idea of agency in a molecule or a storm, let alone an abstraction like money, seems especially far-fetched. We also tend to think in terms of a hierarchy in which “we” (for whatever value of “we”) are at the top, and agency must surely diminish for anything “lower”—a view reminiscent of the medieval Great Chain of Being. When we (hesitantly) extend “agency” to the nonhuman, we tend to do so only for things that act obviously, individually, and on fast timescales, rather than in the aggregate and on slower, more evolutionary ones. It might be time to re-examine these ideas more holistically.

My purpose here is not to follow in Bennett’s footsteps—though I do find her project worth taking seriously. Language is, necessarily, imprecise, no matter what definitions we pick. This doesn’t mean that words are useless, though. When our language can become more rigorous and scientifically grounded, and when we use it to describe patterns across a wide range of phenomena, we can start to see through ideological thickets.

I hope I have explained both clearly and rigorously how the phenomena that give rise to the complexifying dynamics of life apply much more broadly than to the entities we normally think of as “alive” or “agential.” Accordingly, we could expand our definitions of these existing words, or adopt new ones, or do a bit of each. Personally, I would find some broadening of the old everyday words helpful.

That would hardly break new ground. Many traditional, nominally “prescientific” worldviews embrace notions of aliveness, agency, and even personhood that are far broader than the modern Western ones. This seems a likely case of convergent evolution in languages and ideas, motivated by the common need among traditional societies to take symbiosis seriously to secure their own survival, and to flourish. It’s practical as much as it is spiritual: encouraging richer modeling of agency in “nature” enhances a society’s dynamic stability, since all things in “nature,” ourselves included, are so mutually interdependent.

Thus it can be useful to take the view of an animal, plant, or river at times, even if they can’t take ours, the better to care for them—and for ourselves. That, ultimately, is the best reason to consider adopting, or at least adapting, a more inclusive view of the animate. Potawatomi writer and biologist Robin Wall Kimmerer makes this case eloquently in her book Braiding Sweetgrass. ⁸⁶

In agreeing with Kimmerer, I am not encouraging superstition. When scientists castigate animist beliefs as superstitious, they typically appeal to the materialist discoveries of the Enlightenment, which show that the atoms that make up our bodies are no different from the atoms that make up rocks or air. This is true. Atoms are atoms; they all obey the same rules. Hence, the Enlightenment model of a clockwork universe, governed by dynamical laws.

Yet as Schrödinger pointed out in 1944, our understanding of these laws—which he played such a central role in developing—remains incomplete. The laws as they stand do not account for the complex, dynamically stable, symbiotic phenomena that comprise so much of our experience on Earth—indeed, without which there would be no such thing as experience at all. There would be no life, or purpose, or minds, or agency.

As we both embrace scientific rigor and start to figure out those “‘other laws of physics’ hitherto unknown,” ⁸⁷ we should perhaps be less surprised to find the shoe on the other foot. What the poet Dylan Thomas called “the force that through the green fuse drives the flower” ⁸⁸ drives all atoms, not just the ones we presume to be alive.

Acknowledgments

Although this is a small book, it (appropriately enough) nestles into a tangled heterarchy of larger projects, owing its existence to a great many people and several institutions—especially Google, Antikythera, the Berggruen Institute, the Santa Fe Institute, the Mila-Quebec AI Institute, and MIT Press. Many of the people I’ll thank below are affiliated with more than one of these.

As a designed object, this book’s co-creator is James Goggin of Practise, in Auckland. Our third collaboration, What Is Life? further develops an approach to simultaneously physical and digitally native media we’ve established with our first two projects, Ubi Sunt and Who Are We Now? (both published by Hat & Beard Press, with warm thanks to JC Gabel).

Clea Agüera-Arcas, Loraine Agüera-Arcas, and Johan Michalove have helped research, edited painstakingly, offered line-by-line critique, chased down sources and permissions, coded workflows, wrangled spreadsheets, and in short dedicated an immense effort over many months to making the book as beautiful, readable, fair, and accurate as we could. Brian Sholis did the final editing and proofreading. However, I take sole responsibility for any remaining errors or omissions, and all opinions, mistaken or not, are my own.

In collaboration with James Goggin, Minkyoung Kim and Marie Otsuka designed and implemented the web version, which includes a good deal of video and interactive content. The simulations owe a heavy debt to Alex Mordvintsev (also the inventor of neural cellular automata), and his elegant SwissGL library.

Close readings of the text and detailed editorial input, which improved the final result significantly, came from Benjamin Bratton, Stephanie Sherman, Adrienne Fairhall, Anselm Agüera y Arcas, Eliot Agüera y Arcas, David LeBrun, Ben Laurie, Luca Versari, David Wolpert, and Emily French. Warm thanks also to Patricia Churchland, Walter Fontana, N. Katherine Hayles, Reid Hoffman, Michael Levin, Johan Liedgren, Charles Mudede, Addy Pross, Carl Schoonover, Terry Sejnowski, Justin Smith-Ruiu, Dan Sperber, Korina Stark, John Thornhill, Sara Walker, and Ren Weschler. Thank you, Margaret Levi and Bob Kaplan, for connecting me with Papunya Tula Artists in Australia’s Western Desert, and thank you, Spike Mafford, for the beautiful photography. Yalti Napangati’s artwork, which envelopes this book, is the most eloquent expression of symbiogenesis.

At Google, I owe deep thanks to James Manyika and Sundar Pichai for giving my team, Paradigms of Intelligence (PI), the freedom and resources to pursue interdisciplinary basic research. I’m grateful to Demis Hassabis and Yossi Matias for their support of internal collaboration with Google DeepMind and Google Research, respectively, and to Yul Kwon for his deft co-leadership of PI. A big thank you to PI teammates and collaborators: Jyrki Alakuijala, Kenric Allado-McDowell, Travis Beals, Sami Boukortt, Martin Bruse, Iulia Comșa, Yiğit Demirag, Moritz Firsching, Thomas Fischbacher, Alice Guan, Florian Hartmann, Anthony House, Geoff Keeling, Evgenii Kliuchnikov, Seijin Kobayashi, Marcin Kowalczyk, Rachelle Lacroix, Mira Lane, Alison Lentz, Brittni Maekawa, Kaitlin Maile, Delaney McMillen, Alexander Meulemans, Alex Mordvintsev, Eyvind Niklasson, Peter Norvig, Robert Obryk, Ettore Randazzo, Esteban Real, João Sacramento, Mark Sandler, Rif Saurous, Nino Scherrer, Anoop Sinha, Oliver Siy, Winnie Street, Zoltan Szabadka, Luca Versari, Sarah de Haas, and Johannes von Oswald. Visiting faculty James Evans, Guillaume Lajoie, and Blake Richards have been wonderful friends and colleagues, and have all made intellectual contributions to the material presented here.

Time spent at the Santa Fe Institute over the past two years has profoundly influenced my perspective on life, complexity, and intelligence. I have also met a number of new collaborators, fellow travelers, and constructive critics (often a bit of each) there, including Kevin Berger, Sean Carroll, David Chalmers, Ted Chiang, Michael Hersch, Chris Kempes, David Krakauer, John Krakauer, Gary Lupyan, Brice Ménard, Melanie Mitchell, Carlo Rovelli, Zenna Tavares, Geoffrey West, and David Wolpert.

What Is Life? and its mothership What Is Intelligence? are among the first long-form publications of the Antikythera program, under the direction of Benjamin Bratton. Benjamin has instigated many of my major writing projects in recent years, including these books, and has been a loyal friend and frequent collaborator. Warm thanks also to Haley Albert, Nicolay Boyadjiev, Nils Gilman, Emily Knapp, Dawn Nakagawa, Estela Oliva, Tobias Rees, and Claire Webb; my gratitude, also, to Nicolas Berggruen for his intellectual friendship and support, and to the Berggruen Institute.

Finally, warm thanks to Noah Springer of MIT Press, for believing in this project and for helping us bend the rules of traditional academic publishing.

About the author

Blaise Agüera y Arcas is a VP and Fellow at Google, where he is the CTO of Technology & Society and founder of Paradigms of Intelligence (PI). PI is an organization working on basic research in AI and related fields, especially the foundations of neural computing, active inference, sociality, evolution, and Artificial Life. In 2008, Blaise was awarded MIT’s TR35 prize. During his tenure at Google, he has innovated on-device machine learning for Android and Pixel, invented Federated Learning (an approach to decentralized model training that avoids sharing private data), and founded the Artists and Machine Intelligence program. A frequent public speaker, he has given multiple TED talks and keynoted NeurIPS. He has also authored numerous papers, essays, op-eds, and chapters, as well as two previous books, Who Are We Now? and Ubi Sunt. What Is Life? is Part I of the larger book What Is Intelligence?, forthcoming from Antikythera and MIT Press in 2025.

Bibliography

Image credits

Note on the artwork

The front and back endpapers of this book feature details of an artwork by Yalti Napangati, born around 1970 in Australia’s Great Sandy Desert. She is one of the “Pintupi Nine,” a traditional family who remained unaware of European colonization until 1984, when they made contact with relatives near Kiwirkurra. This painting depicts designs associated with the rockhole and soakage water site of Marrapinti, west of Kiwirrkurra. During ancestral times a large group of women gathered at this site during their travels toward the east. While at the site the women made nose bones, also known as marrapinti, which are worn through a hole made in the nasal septum. These nose bones were originally used by both men and women but are now only inserted by the older generation on ceremonial occasions. Upon completion of the ceremonies at Marrapinti the women continued their travels east to Ngaminya and then onto Wilkinkarra (Lake Mackay). The shapes in the painting represent geographical features of the landscape and bush foods the women collected as they traveled.

Note on the type

This book, both print and web editions, is typeset in Exposure, a variable typeface designed between 2019 and 2022 by Federico Parra Barrios at the Atelier National de Recherche Typographique in Nancy, France. Variable fonts are an OpenType technology enabling parameters like weight, slant, and width to vary continuously. Instead of traditional type weight, Exposure uses the new technology to simulate an obsolete one, producing the eroded or blown-out effects of under- or over-exposure in phototypesetting.