Chapter 1 Self-Reproducing Machines: The Evolution of an Idea
“Within the next century we will likely witness the introduction on earth of living organisms originally designed in large part by humans, but with the capability to reproduce and evolve just as natural organisms do. This promises to be a singular and profound historical event—probably the most significant since the emergence of human beings.”
— J. Doyne Farmer and Alletta d’A. Belin, Artificial Life: The Coming Evolution, 1991 (Farmer & Belin, 1991, p. 816)
In the chapters that follow, we explore the early history of thought about machines that can reproduce and evolve. Unless you work in one of a small set of rather specialised academic or engineering disciplines, you may not have come across much discussion of these ideas before, at least not beyond the realm of sci-fi films and novels. We believe that will soon change as the work whose origin we describe here progresses.
1.1 Two Central Questions
There are two underlying questions that have provided the central motivation for all of the work that we cover:
Is it possible to design robots and other machines that can reproduce and evolve just like biological organisms do?
And, if so, what are the implications: for the machines, for ourselves, for our environment, and for the future of life on Earth and elsewhere?
Out of these two questions spring many others. Might this be a route by which we could create machines whose capabilities go beyond the rather narrow focus of today’s artificial intelligence (AI) systems, and which automatically evolve towards a more powerful and wide-ranging artificial general intelligence (AGI)? In contrast to today’s AI systems, might evolution and natural selection instil inner desires and purpose in these machines, as it has done in the biological realm? More futuristically, could a spaceship that can build copies of itself from raw materials scavenged from asteroids and other planets be a route by which we could travel immense distances to explore and colonise worlds in other solar systems or even other galaxies? And what of the economic disruption that might be caused here on Earth by a nanoscale manufacturing plant that could autonomously build more copies not just of its output produce but of the manufacturing plant itself, all at an exponentially increasing rate?
From even a cursory consideration of these questions, it is clear that technologies like these could potentially pose serious threats to our environment, and even to the future of the human race itself. And yet, perhaps instilling machines with the power of reproduction and evolution might be the most promising approach to ensure the survival of intelligent life in the far future of the Earth and elsewhere in the universe?
1.2 The Promise of Self-Reproducing Machines
“Oh my goodness! Shut me down! Machines making machines? How perverse!”
— C-3PO, Star Wars: Episode II Attack of the Clones, 2002 (Lucas & Hales, 2002)
“We realized that the true problem, the true difficulty, and where the greatest potential is—is building the machine that makes the machine.”
— Elon Musk, Tesla Annual Shareholder Meeting, 2016 (Musk, 2016)
“I knew that baby meant we are more than just slaves. If a baby can come from one of us, we are our own masters.”
— Freysa (a Nexus-8 replicant), Blade Runner 2049, 2017 (Fancher & Green, 2017)
At the heart of all of these questions is the concept of a self-reproducing machine (or self-replicator for short). Is it conceivable that we might be able to design and build such machines in the near future? Take a moment to look at any modern gadget you own: your mobile phone, your television, your bicycle, your toaster. Think about all the components from which it is made. The chances are, every single component has been manufactured by another machine. The idea of machines making machines is so commonplace that we rarely stop to give it a second thought. Likewise, the raw materials required for each component have been mined and processed by other machines. The transportation of raw materials to factories, and of manufactured goods to warehouses and stores, is achieved by yet more machines. At every step, the role of humans is becoming increasingly redundant. Recently developed technologies ranging from 3D printing to self-driving vehicles are enabling every aspect of the manufacturing process to become more and more automated.
Could we someday reach the point where the entire manufacturing process is completely automated, from the mining of raw materials to the delivery of a new gizmo to your front door? What are the limits of what could be manufactured by completely autonomous machines? Given where we are at present, it doesn’t seem too far a leap to imagine a machine that could manufacture a complete copy of itself, if provided with the necessary parts. If you find this hard to imagine, how about a group of machines: could we design a large number of different machines that, between them, collectively build a copy of every machine in the group?
If such a feat were achievable, how about finessing the design so that the machine also collects and processes the raw materials required to build its offspring? Although there are doubtless major engineering challenges to be overcome to realise this, there are no obvious reasons why this could not be accomplished in theory. An autonomous system like this would be a completely self-reproducing machine. Just like a biological organism, in appropriate conditions it would be able to produce offspring of its own kind.
The questions posed at the start of the chapter allude to some of the many profound applications of self-replicator technology. What might be the ultimate outcome of all of this, for the machines, for the environment, and for us?
1.3 Diverging Visions in the Early History of the Idea
These kinds of questions, and, indeed, the very idea of a self-reproducing machine, might seem like very modern conceptions. In fact, they have captured the imagination of scientists, philosophers, writers and the general public for hundreds of years.
One of our primary purposes in writing this book has been to explore the very early history of these ideas.
Our search has taken us back as far as the late 1600s, as the full implications of René Descartes’ views of animals as machines began to be explored and debated. For the first couple of centuries that followed, most of the discussion was centred on the question of whether it was possible to design a machine that could reproduce by building a copy of itself. Throughout this book we refer to machines that possess this basic capacity for building a faithful copy of themselves as standard self-replicators, or standard-replicators for short.
In the mid-nineteenth century, a pivotal development in the history of the subject was triggered by the publication of Darwin’s On the Origin of Species in 1859 (C. Darwin, 1859). Within a decade of its appearance, we find multiple extended discussions of the possibility of machines that can not only reproduce, but can also evolve by the natural selection of heritable variations to become better adapted, smarter and more complex over time. We refer to these kinds of machines as evolvable self-replicators, or evo-replicators for short. The early 1900s saw increasing speculation on these ideas, both by scientists and by sci-fi authors, and the 1950s saw the first implementations of simple self-reproducing systems in hardware and software.
The Darwinian vision of a machine that could reproduce and evolve like a biological species continued to inform a significant strand of work on self-reproducing machines as we move from the 1950s to more recent developments. This work embraced the possibility of heritable changes or mutations occurring in a self-replicator’s offspring, these being the source of variety upon which Darwinian natural selection acts. Practical applications of evo-replicator technology include its use as a potential route for the automatic creation of advanced AI systems of far greater power than could be designed by humans, and also as an experimental tool by which we might better understand the conditions that have led to the evolution of intelligent biological life on Earth.
At the same time, as people began to think more seriously about the practical realisation of self-reproducing machines in the 1940s and 1950s, we also begin to see the emergence of a third distinct direction of work. This line of research was based on the insight that, given the right design, a self-replicator could quite easily be directed to produce specific goods and materials for us, in addition to reproducing itself. Thus, this work focused much more on the potential of physical self-reproducing machines as general-purpose manufacturing systems that could be deployed cost-effectively in inaccessible locations on Earth or further afield. The machines could then be remotely directed to produce a wide variety of specific outputs—think of them as glorified 3D printers—without requiring human maintenance, and with the potential to further replicate their activities elsewhere. We will refer to these kinds of systems as manufacturing self-replicators, or alternatively as maker-replicators for short.
The key advantage of a maker-replicator’s capacity for self-reproduction is that its creators would (in theory) only have to produce one of them initially, which could then automatically ramp up its activities in situ by producing more copies of itself in its target location. Thus, initial manufacturing and deployment costs are reduced by virtue of requiring only a single machine to commence the process. Furthermore, ensuring the long-term reliability of a single machine in a remote location becomes less of a crucial issue if the machine is able to make further copies of itself before it suffers a failure.
Another feature of the technology, which may be regarded as a benefit or a curse depending upon one’s perspective and goals, is that it has the potential—given sufficient raw materials as input—to produce output at an ever accelerating rate. As a self-replicator builds more copies of itself, and the copies build more copies of themselves, the total number of machines could increase from the initial one, to two, then four, eight, sixteen …, the population doubling each time. From a human perspective, having paid the one-off cost of producing the first machine, we would have a process where the rate of output would increase exponentially over time! Everything we know about economics would be turned on its head.
Practical uses of this technology might include the economic large-scale production of valuable resources, for example converting sea water into fresh water, or building devices to capture solar energy. It might also be applied to geoengineering projects to tackle global warming. As we’ll see in later chapters, some authors have proposed using maker-replicators to mine and process valuable resources on asteroids subsequently to be transported back to Earth, or even to terraform other planets prior to colonisation by humans.
Key issues in maker-replicator development include the engineering, economic and safety challenges in designing this kind of system. In contrast to evo-replicator development, those working on maker-replicators usually view the potential of evolution in their machines—and hence the possibility that the human designers might lose control of them as the machines develop unexpected abilities—as a danger that should be avoided through the design of appropriate safeguards. Having said that, as we’ll see in Sect. 5.1.1, the development of the first significant theoretical work on self-replicators, by John von Neumann, addressed the design of machines that could both manufacture a wide variety of products and evolve to become more complex over time—that is, von Neumann’s work was about evolvable manufacturing self-replicators (evo-maker-replicators).
There has also been a more limited amount of research on software maker-replicators. This work seeks to produce software self-replicators that have a general capacity to perform other specified computational tasks as well as reproduction. As with their hardware counterparts, a motivation for developing software maker-replicators is to create systems that can perform with high reliability in environments where they cannot be easily maintained by human operators. At first glance it might seem that safety issues associated with software maker-replicators are less critical than those of their hardware counterparts. However, as we discuss in Chap. 7, we should not underestimate their potential to cause serious harm not just online but also in the real world.
Following the first implementations in the 1950s, research in recent decades has seen many developments in all of these different flavours of self-replicator technology. The focus of work has shifted from speculation and science fiction to detailed studies of the design and implementation of self-reproducing machines in hardware and in software. During this time, work in designing physical self-reproducing machines in hardware has been largely concerned with maker-replicator systems, while software implementations mostly focus on evo-replicator systems. In both hardware and software we also see a significant amount of work on standard-replicators as a foundation upon which to progress towards the other two kinds of replicator. As we’ll see later on, work on hardware maker-replicator systems is mostly still at the conceptual and prototype stage. On the other hand, work on software evo-replicator systems is more advanced, with an active and growing group of researchers working on improving the evolutionary potential of their implementations.
1.4 Relevance Today
In recent years, the idea of self-reproducing and evolving machines has been overshadowed in the media by impressive breakthroughs in other areas of AI and machine learning with more immediate practical relevance. However, the long-term implications of self-replicator technology are potentially far more transformational.
We have already alluded to some possible applications of self-reproducing machines in the preceding pages. The introduction of real-world maker-replicator technology has the potential to revolutionise the production of materials on Earth, to provide an economical means of mining the resources of other moons and planets, and even to act as a route by which humankind—or our technological offspring—may explore and colonise worlds beyond our own solar system. In short, this technology could profoundly reshape our society, our relationship with the environment, and our place in the universe.
While most of the early work we describe in the following chapters considered large-scale (i.e. approximately human-scale or larger) versions of this technology, more recent developments in molecular-level systems such as nano-bots provide an alternative medium in which physical maker-replicators could be instantiated. It is likely that research on systems at this scale will produce significant results before work at larger scales. At the same time, guarding against the risk of a runaway exponential self-replication process is more challenging at the smaller scale. If not carefully managed, physical self-replicators at both small and large scales bring with them substantial risks of causing catastrophic damage to the environment and existential threats to biological species, including our own.
As we show in the chapters that follow, it appears that the barriers to building physical maker-replicator systems are chiefly technical and economic rather than theoretical. The truly transformative potential of self-reproducing technology for commercial and sociological goals—and the potential financial returns that could be captured by a first-mover in the field—mean that we must assume that, sooner or later, some research group or corporate organisation will be successful in manufacturing an operational physical self-reproducing system. While this is unlikely to happen in the near-term (especially for larger-scale systems), we expect significant progress in this area over a time frame of several decades. It is unclear whether, in the long run, this will be a positive or negative development for humanity. One thing that is clear, however, is that only by thinking carefully about these machines and their associated risks and implications will we be able to guide their future, and ours.
The development of software evo-replicator technology is the area of most relevance in the short-term, not least because systems already exist that implement evolutionary self-replicators in software. This technology holds the promise of surpassing current commercial AI, by evolving systems with capabilities that go far beyond what they were originally designed to do. Some see evo-replicators as the most likely route by which to achieve human-level, or even super-human-level, artificial general intelligence (AGI). Research on instilling software evo-replicator systems with the ongoing creative power of biological evolution is currently a central focus of research within the academic field of Artificial Life (ALife); this quest for open-ended evolution has recently been identified as a “grand challenge” for the field (Stanley et al., 2017). Although software-based evo-replicators don’t present the same level of danger of environmental damage as posed by physical maker-replicators, they nevertheless have the potential to cause havoc in the online world if not properly managed. Recent trends in computer viruses that are designed specifically to cause damage to real-world infrastructure also demonstrate the dangers of underestimating the potential negative effects of software-based evo-replicators.
Given the potential significance of all of the lines of research outlined above, it is vital that these developments are accompanied by a proper consideration of the possible risks and benefits involved. An appropriate starting point for this endeavour is a careful consideration of what has been achieved to date, and what are the motivations and goals of those involved. Our review of the early history of the subject, as set out in the following chapters, is our contribution to constructing a firm foundation for these considerations.
1.5 A Note on Scope and Terminology
Before proceeding, some clarification is required on the scope of our review, on how this book differs from other reviews, and on our use of terminology.
As we describe in Chap. 5, the towering figure in the theory of self-reproducing machines is the Hungarian-American polymath John von Neumann, whose work on the topic in the late 1940s and early 1950s put the subject on a firm theoretical footing. Existing reviews of the subject generally start with von Neumann and concentrate on developments from the 1960s onward. In contrast, our focus in this book is on the earlier and less well-known history of the subject. After discussing the earlier work at length, we also summarise more recent developments and provide an introduction to more detailed reviews. By the end of the book, we therefore aim to have provided a sound overview of the entire history of the field, with pointers to further information about the more recent work.1
This is not a review of mechanical models of living things in general (a topic which has over two thousand years of history (Cave et al., 2018)), nor does it cover the much wider and more general idea of the evolution of technology.2 Our focus is specifically on self-reproducing and evolving machines—we cover both physical machines (e.g. clockwork automata, electromechanical robots, and molecular-scale devices)3 and logical machines (e.g. software programs and abstract automata), but we do not branch deeply into the area of bio-mechanical hybrids, bionics and cyborg technology where the reproductive functions remain predominantly biological.
Regarding our use of terminology, we acknowledge from the outset that the term self-reproduction can be problematic. No system is truly self-reproducing: the process is always the result of an interaction between a suitable structure and a suitable environment, causing the production of further copies of the structure. Many of the authors discussed in the following sections have offered insights into the various issues involved. We highlight these as we proceed and provide further discussion of the topic in Chap. 7.
With that said, we use the term self-reproducing machine to refer to a machine (or manufacturing plant) that, within a defined range of environments, can manufacture a copy of itself after collecting and processing the required raw materials. The term “self-reproducing machine” can be something of a mouthful when used frequently, so we also use the term self-replicator as a slightly shorter synonym.4
As stated above, we use the more specific terms standard-replicators, evo-replicators and maker-replicators to refer to particular flavours of work on self-reproducing machines. The focus of work on maker-replicators (manufacturing self-replicators) is on their ability to manufacture a wide range of products in addition to being able to produce copies of themselves. Those working with maker-replicators are generally very concerned that these machines work in a very controlled manner and are not able to evolve new capabilities. In contrast, the focus of work on evo-replicators (evolvable self-replicators) is very much on their ability to evolve and to acquire capabilities beyond those originally given to them by their human designers.
1.6 Outline of the Rest of the Book
In the chapters that follow we discern three major steps in the intellectual development of thinking about self-reproducing and evolving machines:
The first step involved the introduction of the view that animals can be understood as machines, due in large part to René Descartes in the 1630s–40s. This step, which we discuss in Chap. 2, introduced—implicitly at first, but later more explicitly—the first glimmerings of the idea of machine self-reproduction. The first direct mentions we find of the idea are in ab absurdo arguments against the view of animals as machines, but in the eighteenth and nineteenth centuries we begin to see the idea discussed without necessarily being rejected as obviously absurd. Using the terminology we introduced earlier, the discussion of the subject during this period focused on the possibility of standard self-reproducing machines (standard-replicators).
The second step involved the development of the idea that machines, like animals, might not only be endowed with the capability of self-reproduction but also of evolution (that is, evo-replicators). After the first step had been taken, two further important factors contributed to the realisation of the second step—the climax of the Industrial Revolution in Great Britain in the early nineteenth century, and the publication in 1859 of Darwin’s ideas of evolution by natural selection in On the Origin of Species. Within a decade of the publication of Darwin’s work, we see several authors discuss at length the idea of the evolution of self-reproducing machines. The most significant early work on this topic comes from Samuel Butler, Alfred Marshall and George Eliot, as we discuss in Chap. 3. These works mark the arrival of the concept of evo-replicators in the published literature. Coming as they did at the end of the British Industrial Revolution, these ideas now seemed much less far-fetched, and therefore potentially more frightening. Being a more realistic prospect, this step also led to the development of more thorough discussions of the implications of evo-replicators—in addition to Butler, Marshall and Eliot, John Desmond Bernal in the 1920s provided a deep, scientifically-grounded discussion of how such technology might shape the direction of the far future of humanity. At the same time, evo-replicators became a common theme in early works of science fiction. These early twentieth century developments are covered in Chap. 4.
The third step, discussed in Chap. 5, saw the first serious studies of the design and implementation of practical self-reproducing machines. This step comprised two separate strands. One strand was inspired by Alan Turing’s work on the concept of universal computing machines in the 1930s (Turing, 1937) and involved the development by John von Neumann—starting in the late 1940s—of a theory of universal constructing machines. Although von Neumann was very much interested in the potential for self-reproducing machines to evolve, the universal construction aspects of his theory, which concerned general-purpose manufacturing machines, can be seen as the seed of the concept of a maker-replicator. During the same period, another strand of work was inspired (in part) by developments in molecular genetics and in understanding the process of DNA self-replication. It involved the study of much simpler artificial self-reproducing systems than those considered in the first strand, and it is exemplified by the first instantiations of artificial self-replicators in software (by Nils Aall Barricelli) and in hardware (by Lionel Penrose). This strand represents the first attempts at implementing evo-replicators.
In Chaps. 2–5, we flesh out the detail of these three steps and discuss the people and ideas involved.
These chapters take us up to the early 1960s. The period from that point onward is already fairly well documented—in Chap. 6 we provide an overview of this more recent work and give references to existing reviews. We also mention some very recent work that has not been covered elsewhere.
Having traced the development of these ideas, and the thoughts and motivations of those involved, we end by summarising in Chap. 7 what has been achieved, what key issues remain unresolved, the outlook for future work in this area, and the implications for the future of humanity.
But first we need to go back to the beginning. In the next chapter we investigate where, when and why the concept of self-reproducing machines first arose.
References
We acknowledge that our literature search has been conducted primarily in the English language. While we have spent some time searching for sources in other languages (including French, German, Spanish and Russian), we cannot rule out the possibility of the existence of relevant non-English language work in addition to those that we have found and cover here.↩︎
Excellent coverage of these broader topics can be found elsewhere: for a review of mechanical models of living things, see, e.g., (Riskin, 2016), (Mayor, 2018), and for discussion of the general idea of the evolution of technology, see, e.g., (Basalla, 1988), (Arthur, 2009), (Mesoudi, 2017).↩︎
We use the terms automaton and robot more or less synonymously throughout the book. Both are machines driven by their own internal instructions and source of movement. The term automaton (plural automata) carries with it more of a sense that the device is acting in a rote fashion according to predefined rules; historically, it is the term that has been applied to clockwork models of living beings, and also to the kind of computational cellular automata models employed by John von Neumann (Sect. 5.1.1) and Nils Aall Barricelli (Sect. 5.2.1). Robot is a more recent term (see Sect. 4.1.2) and usually implies a autonomous machine with more intelligence than a simple automaton. While automaton may refer to a machine implemented either in hardware or in software, we reserve the term robot strictly to refer to hardware devices.↩︎
Throughout the book we generally use the terms reproduction and replication synonymously. Within a certain subset of disciplines concerned with self-reproducing machines there is a convention of using the word replication in the case where a perfect copy is produced (so there is no evolution) and reproduction in the case where mutations and other genetic operators might produce variety in the offspring and thereby allow the possibility of evolution (Sipper, 1998). However, within the broader range of sources that we review here, there are several conflicting definitions of the distinction between these terms (e.g. (Dawkins, 1976), (Griesemer, 2001)). Within the context of all of the work we discuss, we believe it is clearer to make the distinction between standard-replicators, evo-replicators and maker-replicators, rather than relying upon the reader to remember technical distinctions between the words reproduction and replication.↩︎