Overview: The Biomimicry Future of AI
Learning from Nature to make Smarter, Wiser, and Safer AI. - Part 1 of a Limited Series by John Smart and Nakul Gupta
TL;DR: We predict that the physical and informatic nature of our particular universe is forcing leading AI designers to increasingly copy and vary the evolutionary and developmental (“evo-devo”) algorithms of biology. Deep learning is just the tip of the biomimicry that is coming, in humanity’s competitive efforts to create more intelligent, ethical, secure, and adaptive machine intelligence. The Natural Alignment Hypothesis (NAH) proposes that deep biomimicry (with analogs to evo-devo genetics, neural architecture, and domestication of AIs, in cooperative and competitive communities, under selection) is the only easily accessible design path to creating complex, self-regulating, trustable, safe, and adaptive machine intelligence, and that “They” (future AIs) cannot arise as alien intelligences, with goals and values orthogonal to adaptive biological networks. Instead, the constraints of both evo-devo dynamics and of adaptive social selection will make them into digital versions of “Us”, and later “Us+”. Artificial intelligence, in other words, must become natural intelligence.
Today’s large language models (LLMs like Chat-GPT, DALL-E, etc.) may seem like they could soon become an alien intelligence, as they have so little reasoning capacity and self-awareness, and can so easily hallucinate plausible but false answers. But these platforms are simply taking the next adjacent steps in associational intelligence that are most easily available to network designers. LLMs are proving excellent at copying, varying, and summarizing the language and visual symbols humanity has encoded on the web, and making single-step generative inferences with their associations. These are very powerful new abilities, with many disruptive societal effects, and with many new misuses to be better anticipated and regulated. Nevertheless, such systems are still very far from having emotion, intuition, empathy, ethics, common sense reasoning, and other higher order processes that we believe will prove necessary to trust their performance in complex environments and to stabilize their interactions within adaptive communities of machine intelligences.
In our view networks of both natural and machine intelligences will always win over individuals, and any specific groups, in the competition for general adaptiveness. We consider many of these collective features to be universal, forced on complex species via convergent evolution (universal development) on all Earthlike planets. Evolutionary biology is still very tentative at making such claims. Simon Conway Morris, in Life’s Solution (2003), George McGhee, in Convergent Evolution: Limited Forms Most Beautiful (2011), and Johnathan Losos, in Improbable Destinies (2017) are among a courageous few willing to brave cries of “biocentrism” and “anthropocentrism” in arguing that not only cells, but the humanoid physical and mental form are likely to be inevitable—the most generally adaptive forms and functions—on all Earthlike planets in our universe. We believe this view, and the NAH, will become validated science in coming decades, as our astrobiology and simulation capacities (capable of falsifying such theories) both advance. We even believe science will learn to see our universe itself as an evo-devo system, with both a global developmental life cycle, and a broad set of evolutionary mechanisms aiding adaptation in the multiverse. See John Smart, Evolutionary Development: A Universal Perspective (2019) for one such view.
Neuroscience is already greatly informing AI development. We predict an increasingly deep understanding of autopoetic (“evo-devo”) chemistry, genetics and biology and how natural and social selection of autopoetic networks generates adaptiveness, will be necessary to generate complex, self-repairing, self-improving machines. Few AI designers see deep biomimicry, autopoesis, and natural and social selection as a necessary future today. Yet we believe it is only approach that can evolve and develop sufficient internal controls to handle the incredible complexity and predictive and creative powers of future learning machines.
We have written this Substack series to present this natural alignment hypothesis in depth. What follows is our best short argument on the necessarily evo-devo future of AI, on all Earthlikes in our universe, and some of the more predictable aspects of the future relationship of AI to biological humanity. The NAH guides us to think not primarily about the coming Superintelligence (aka, the “technological singularity”) but about a more foundational and universal concept, Superadaptiveness. We’ll see that adaptiveness is most fundamentally considered not for specific individuals, or even groups, but for networks of diverse, cooperative, and competitive individuals and groups, best considered as a complex ecosystem, networks that are under continual selection for both local and universal adaptiveness.
In the last decade, there has been growing interest in methods, frameworks, and models for aligning artificial intelligence systems with human goals and values. The top reason for this new interest has been a decade of impressive advances in the design and performance of deep learning AI, based on multi-layer artificial neural networks. As increasingly powerful AIs have been developed and deployed in companies, governments, and the open source community, many of us have begun considering issues of long-term AI safety, ethics, and security.
Philosopher Nick Bostrom’s Superintelligence, 2014, has been a catalyst for many who think about this problem. In this book, Bostrom posits many ways a hyper-intelligent artificial agent could lay waste to humanity through misaligned goals, values and preferences. Since 2015, several billion dollars have been pledged to AI alignment startups, nonprofits, and public benefit corporations, the most famous of which is Open AI. Two newer entrants are Anthropic, with $700M of funding, and Inflection, backed by megabillionaire Reid Hoffman. Philanthropic communities like Effective Altruism have taken up the problem, and it is a common podcast topic. This is a good time for better foresight to emerge. Our series is an attempt at such foresight.
This new field, AI alignment, is different from responsible AI, or the beneficial use of AI and its sets of training data by firms, states, and people. The latter is already a major societal issue, and position statements and regulatory frameworks on responsible AI use and data rights have been emerging in companies and oversight bodies around the world. Responsible AI use is a current great concern. AI alignment is less pressing today, but more important to the long term future of humanity, because AI learning, both in the electronic substrate and in simulation space, is roughly ten-million-fold faster than the electrochemical learning speeds of biological humanity. They aren't all that smart yet but they are the most rapidly learning systems on Earth today, by far.
The book we recommend for the best current framing of this problem is Brian Christian’s The Alignment Problem, 2020. Artfully written for general readers, it covers a basic history of AI, a broadly representative sample of researchers and builders in the current AI design and alignment community, and the multifaceted societal and technical scope of this challenge. We recommend it as background reading for our series, which will propose deep neuromimicry and biomimicry as a uniquely effective, wise, safe, and adaptive alignment strategy.
It is important to acknowledge, at the beginning of this series, that AI has been at the peak of a societal hype cycle since 2017. Many traditional companies and startups claim to be AI-enabled today, yet most are still using traditional and simpler (engineered, non-neural) machine learning methods. Even our most complex deep learning AIs still have few of the higher functions of human intelligence. But many have argued that they do have some: today’s deep learners have become very good at seeing stable patterns (feature learning) in large data sets. Several neuroscientists claim that what they do is very similar to what happens in mammalian visual cortex. (For example, Yamins and DiCarlo 2016.) The more complex deep learners also use reinforcement learning and adversarial (self-competing) architectures, early versions of complex functions also found in brains. Yet even with their current simplistic neuromimicry, artificial neural networks are proving incredibly useful in an ever-growing range of fields. [2023 note: Witness the truly impressive new achievements in text generation and interpretation we’ve seen from Open AI’s Chat-GPT 3.5 since Nov 2022.] There is a lot of real progress behind the hype.
Yet it is also very helpful to recognize the many vital neural functions and modules our AIs are still missing. Our best deep learners today don’t yet do (complex) compositional logic, causal reasoning, or have complex self- and other-models or broad practical judgment (common sense). All but a handful of research AIs still use a very simplified model of biological neural networks, a model that leaves out many important mechanisms, like spike timing, that help biological brains to continually reweight their connections, apparently enabling brain plasticity, or graceful network learning and unlearning. Those that are more “neuromorphic” in design are still very hard to know how to program, and none have yet found commercial use. Nevertheless, massively parallel chip designers like Nvidia and Google, and more modestly, AMD, and Intel, and classic software companies like Microsoft, have all made multimillion dollar investments in specialized AI hardware, beginning with GPUs (graphics processing units), then TPUs (tensor processing units, specifically useful for neural nets) and now research chips that help neural networks do simple (non-compositional, single-step) versions of a few human intelligence processes, like inference.
Computer industry watchers know that Moore’s law for CPU (central processing units) performance—a doubling of processing power of computer chips roughly every two years since 1965—ended circa 2007, as designers ran into further miniaturization and heat challenges. But advances in these more specialized chips, on which neural networks run, are continuing at a very rapid pace. Whenever applications can be massively parallelized, as we saw in the 1990s in graphics processing and in the 2010’s in neural networks, hardware advances produce compute advances in direct proportion to the exponential growth in parallel circuitry. It’s hard to know if a “general Moore’s law” still applies, as we can no longer use circuit density to measure performance, but must now derive less obvious measures of circuit parallelization and complexity. Nevertheless, many benchmarks for AI training (number of model parameters, sizes of training data sets, etc.) and performance (on various tasks) have remained rapidly exponential over this first deep learning decade (2012-2022). Some, including the futurist Ray Kurzweil, extrapolate this progress to argue that general (human-surpassing, self-aware, self-improving) artificial intelligence (GAI) may be only a decade away. We are deeply skeptical of this view.
Gary Marcus and Ernest Davis’s Rebooting AI, 2019, offers a helpfully critical view of current AI capabilities. One of their many good arguments is that common sense reasoning, and perhaps other higher features of human intelligence, will be necessary before AI will be more broadly deployed in society. In a discussion of self-driving car software, they point to the endless edge cases and occasional catastrophic failures observed even with our best current systems. They predict these failures will continue to occur because the AIs involved have no reasoning, and self-, or world-models. They are presently primarily very advanced data filterers and pattern recognizers. They further predict (and we agree), that fully self-driving cars will surely deploy at scale much further into the future (optimistically, a decade from now, and pessimistically, two) than AI optimists have long expected. We recommend this book for a clear view of many key things our AI systems can’t yet do, and will be unlikely to do soon. Marcus also has a new substack on this topic.
Given the limits of current AI, many practitioners in the traditional machine learning community believe the topic of alignment itself is presently over-hyped. We disagree. As alignment researcher Paul Christiano states, getting AIs to accurately do what we want them to do (control) is already a very practical and valuable problem. Getting them to consistently do beneficial things is also a growing and compelling problem.
In our view, there are actually three alignment problems. The first is aligning AIs to the needs of their particular users (control alignment). The second is aligning AIs to the best values that any diverse groups of experts and layfolk agree are worth teaching to our machines (values alignment). The third problem, which we tend to forget, is enabling AIs to discover and apply values which are which are both progressive and protective of life and intelligence in our particular physical and informational reality (universe alignment). Most of us would prefer such alignment to occur whether or not we are smart or wise enough to teach the best values to our machines.
We will discuss all three alignment problems in this series. The second problem is similar to the third problem, as the wiser we get, the more universal our values get, but there is much room still for the growth of wisdom. Humans today have a vast diversity of values, cultures, norms, and laws. We believe that closer we and our machines get to a universal values and ethics framework, one common to adaptive intelligent life on all Earthlike planets (the only kind of planets we know that support complex life), the better we will get at resolving values conflicts between individuals and groups in ways best for life and intelligence. Fortunately, if AIs are learning machines, they will increasingly be able to join us in figuring out the most adaptive human and universal values, and how to make good tradeoffs between values, in ever more contexts. All values models, we will argue, will always be incomplete, but with luck and good design, AI values will always be improving, and perhaps later this century, faster than ours.
A large body of early work has been done in AI safety since the 1990s. This work is being revisited and extended by a new generation of researchers. New alignment topics like explainable AI (XAI), a design priority to make machine decision models more transparent, comprehensible, and retrainable, have gained significant traction. Promising and previously arcane AI control mechanisms, like inverse reinforcement learning (IRL), are being used to require learning machines to model the preferences of their human users under uncertainty. Both XAI and IRL, and other control strategies we will discuss, have natural analogs in biological networks, and are excellent steps forward, in our view.
Nevertheless, the AI alignment field is still quite small. AI investors Nathan Beniach and Ian Hogarth, publishers of an annual State of AI report, estimate only about one hundred full-time researchers, spread across a few dozen corporate and academic settings, are working on alignment problems today. It is to this audience, and to the far larger numbers of scholars, students and others who are drawn to exploring the rising challenges of alignment, that these articles are addressed.
Our Series
In this series, we will explore how a growing understanding of biological systems, and increasingly sophisticated use of biology-inspired approaches in AI design, a field colloquially known as biomimicry, may offer us the only broadly effective and efficient solutions to long-term AI alignment problems.
We call this topic natural alignment. It presumes a deep design wisdom is encoded in both evolutionary and developmental processes. Both processes have competing yet complementary goals in living systems, and each has been subject to billions of years of natural selection. We will propose that better scientific understanding, copying, and varying of three complex adaptive systems: natural intelligence in mammalian brains, natural ethics in human groups, and natural security in immune systems, will be necessary to design increasingly smart, wise, and safe learning machines. We’ll see how the emerging field of evo-devo biology helps us to understand each of these systems in both evolutionary (creative, divergent, stochastic) and developmental (conservative, cycling, predictable) terms, and to bring each of these opposing yet equally vital dynamics to our machines.
So far, only nature itself has demonstrated a truly astonishing capacity for managing complexity. Think of the complexity of a human being, all the ways that our molecular, cellular, organismic, and neuro biology corrects errors, fights entropy, and anticipates and responds to threats. Think also of the even greater complexity and adaptiveness of human communities, which maintain great individual diversity, have positive-sum ethics, and are antifragile, meaning they get stronger when stressed. Now think of how little we still understand these systems today.
We believe that by continually asking what we can learn from natural systems and how they regulate their own intelligence, ethics, and security, we will increasingly find the most effective and efficient solutions to our growing alignment challenges. As AI complexity grows, we expect fully top-down, engineered approaches will become increasingly constrained by the comparatively limited iteration, design, and complexity management capacities of human engineers. Indeed, it is possible that only a deep, bottom-up dominant, biomimicry approach will deliver a dynamic of increasingly capable, trustable, and secure AI.
There is a common, but not universal, assumption in the AI design community, in alignment research, and in the rationalist and Effective Altruism (EA) communities, that we will engineer most of our coming alignment features by finding fundamental (universal) maths, models, algorithms, and optimizations. Engineers who favor the use of formal methods in hardware and software sometimes assume this future. One line of reasoning, popular with rationalists and most (not all) of the EA community, goes that computers and some humans are quite good at math, deduction, and theorem proving, so perhaps we can get to general AI by expanding the realms of the provable and logical, increasing the size of the light in the mathematical and engineering search space, and the “keys” for each new process will increasingly appear. Even most practitioners in the field of neuromorphic engineering, designing hardware to mimic neural activity, implicitly take this view.
Yet perhaps one of the most fundamental insights from the uncertainty principle in quantum mechanics is that we live in a universe that is primarily indeterminate, contingent, creative, and unpredictable (read “evolutionary”) and only secondarily deterministic, convergent, and predictable (read “developmental”). For example, if we determine the position of a subatomic particle, its momentum becomes uncertain. A photon coexists as both a particle and a wave, and we can determine it to one or the other state, but only depending on how we observe it, which itself is a fundamentally contingent and unpredictable act. It doesn’t matter whether we call this uncertainty a measurement precision limit, irreducible stochasticity, the observation effect, or free will, it means much of the dynamical and informational future of reality will remain unpredictable to any observers within the universe.
Similarly, development, in biological systems, is a deterministic, convergent, and statistically predictable process, across our life cycle. But all living systems also use evolution (unpredictable, creative recombination and mutation) and its vast diversity to protect life, as a network, and evolutionary change grows rapidly unpredictable the further ahead we look, in direct contrast to developmental predictability.
For those who suspect the reality of multi-level selection in all autopoetic (self-replicating, self-maintaining) complex systems, it is easy to infer that these same evo-devo processes appear to be happening in replicating and varying ideas (“memes”) in brains. Humanity develops math, logic, and self-consistent models, but we also evolve (imagine, hypothesize, explore) a great variety of less rational and self-consistent philosophies, beliefs, and models, and over time, our environment selects the most adaptive, for each context. Far more likely to succeed than a purely rational (theoretical, developmental) approach to AI design, in our view, is to simultaneously empower competing experimental (empirical, evolutionary) approaches as well, and to subject both to strong selection processes, in diverse contexts.
Given these views, we expect the vast majority of the critical alignment advances humanity will need in coming years will be discovered and entrained in the bodies and brains of coming AIs via their own evolution, development, and selection. This natural alignment process will be only lightly and incompletely guided by humanity’s science, rational approaches and causal models, no matter their sophistication.
Formal methods (developmental logic) have often been important in the advance of computer science, and we must always seek to improve them, and thus to better ground our evolutionary searches, but formal methods have never been the major way we advance. Such methods have always had sharp limits in computability or applicability, and have been far less influential than our empirical efforts. Unfortunately, the lights of logic, deduction, and theory do not seem to be strong enough to illuminate the necessary search space, because living beings, the most complex thinking collectives on Earth, do not use deduction, logic, and theory as their primary search strategies. They use induction, experiment, and feedback. As many evolutionary psychologists note, our species logical faculties are quite weak and recently evolved.
But more fundamentally, as we’ll argue throughout this series, our universe itself appears to be only partly deterministic, partly logical, partly modelable, and partly optimizable. In other words, it is computationally incomplete. In the history of scientific progress, and even more so in technology, logical, deductive approaches have been the minority story. Most progress by far has occurred by exploratory, empirical trial and error, guided by intuitive, inductive thinking, logically incomplete and experimental models, and serendipity. Ideas in brains, in other words, primarily evolve (vary and explore in unpredictable ways), and only secondarily do they develop (exhibit predictable, progressive, and rational dynamics).
A central strategy to keep improving AI, in our view, will be to copy and vary the natural intelligence processes used in mammalian brains, many of which we won’t fully understand, as our neuroscience, biology, and computer science and engineering continue to advance. After all, copying and experimentally varying has been life’s main intelligence growth strategy, in five billion years of adaptation on Earth. Even our first cell likely emerged by co-opting self-copying and varying chemical systems, autocatalytic sets, in the view of most origin of life researchers. Life’s essential prebiotic chemicals, including fats, amino acids, and all our nucleic acid bases, have been found throughout our solar system. These precursors were generated by autocatalytic (copying and varying) processes. Never, to our minds, has life invented outright, in any of its prior metasystem transitions (major evolutionary developments). One cartoon (Gillings et al. 2016) of these transitions is offered below.
To expect that human intelligence will escape this universal copying and varying dynamic seems to us a false assumption. It is more accurate, in our minds, to view human logic and deduction, and all our rational efforts and models, as intelligence-guided guesses at better designs, guesses that are typically simplistic and incomplete relative to the real complexity of natural systems, guesses that we largely copy and vary inside a social network of other’s guesses.
Our key design questions, in this view of human activity, should be not only how our engineering guesses can be more quickly and skillfully iterated, tested, and varied, but also how they can be moved out of our slow and limited minds and hands, and into the emerging minds and hands of our increasingly natural machines, via processes of evolution and development, and how to better select and stress test those systems to ensure their alignment, safety, and symbiosis with us. We must admit that we are not capable of eliminating individual rogue AIs, whether self-evolved or intentionally designed, but we can greatly minimize their frequency of emergence, so that the vast majority of future AI systems are naturally aligned, and will police their bad actors, just as we do today.
A few domains in computer science take this general design approach. They all fit under an umbrella term, natural computing, describing computational models for any biological process, and the use of those models on a wide range of physical computing platforms, ranging from digital and analog electronics to molecular substrates. Natural computing includes the aspirational field of artificial life (vibrant in the 1970s and 1980s, much smaller now), the now-revitalized field of artificial neural networks, and the slowly growing field of evolutionary computation, and its subfields, including evolutionary algorithms (EA), genetic programming (GP), artificial development (AD), and computational neurogenetic modeling (CNGM).
The more ambitious of these fields and subfields are examples not only of neuromimicry, as we find in modern deep learning and computational neuroscience, but of biomimicry, a more fundamentally creative (via evolution), conservative (via development), more complex, and potentially more powerful approach. Since the 1990’s, computing pioneers like John Koza have demonstrated human-competitive performance with GP systems. AD and CNGM seek even greater biological accuracy, and the critical capacity to evolve and develop machine phenotypes from genotypes.
Biomimicry approaches are only thinly funded and researched at present. One problem is that we still don’t have good analogs, in computer science, of the key degrees of freedom used by gene-protein-cell regulatory networks to create and maintain neural networks. Given this hurdle, and the good progress in neuro-inspired methods, these more deeply natural methods have been largely neglected.
Later in this series, we’ll see how that may change, as the neuromimicry approach gets increasingly complex and hard to manage, and how a better understanding of autopoesis (self-replication and self-maintenance) in all living systems, will allow us to bring the three pillars of natural computing systems, evolution, development, and selection, to our most complex and intelligent machines. When AI designers can employ AD and CNGM processes that roughly mimic the astonishing phenomenon of biological development (picture below), producing complex artificial neural networks and robots that are increasingly helpful under selection, we’ll know we are on the right track.
In our view, the more complex, competitive and economically valuable the AI field becomes, the more designers will be forced to move away from human-engineered approaches to design and validation, and further into neuro-inspired and later bio-inspired designs, and to rely on the power of artificial and natural selection, to manage AI’s ever-growing complexity.
A decade of deep learning success has reinvigorated neuro-inspired design, validating pioneers like Frank Rosenblatt and his perceptron in 1958, and Rumelhart and McClelland and their work on parallel distributed processing in the 1980s. But as many scholars have pointed out, today’s artificial neural networks capture only a small fraction of our natural intelligence, and virtually nothing of our ethical and defensive systems. Yet all three systems will be critical to future alignment research and development, in our view.
Three Alignment Challenges: Natural Intelligence, Ethics, and Security
To better understand human nature, and apply more of its evolutionary and developmental lessons to AI design, three adaptive systems seem particularly fundamental: our individual brains, our group behaviors, and our immune systems, both individual and societal. We call the first set of systems natural intelligence, the second natural ethics, and the third natural security. We make the supposition that as biomimicry grows, even with the unique capabilities of technological systems (memory, speed, durability) over biological ones, our best AIs will become increasingly brain-like in their general design. AI, in other words, will become NI.
Many features of biology will not be needed in intelligent machines, as they are based on a different deterministic substrate, but if they must be both evolutionary and developmental, a great many, including analogs to some of our cellular error-correction features, and to mammalian immune systems, will prove critical, in our view. AI cooperation and competition (coopetition), both with other AIs and with individual users and groups, will also have to incorporate increasingly complex approximations to natural ethics in coopetitive species. Humans, the most cooperative and competitive species on Earth, are a prime mimicry example. AI safety and security, for its part, will use many more of the systems and networks of natural security. In our next three posts, we will offer an overview of each of these three vital alignment challenges.
. . .
Now we’d like to ask a favor of you, our valued readers:
What have we missed, in this brief overview? What is unclear? Where do you disagree? Who else should we cite? What topics would you like to see discussed next? Who would you like to see as a guest poster or debater? Please offer any constructive comments, critiques, and advice that comes to mind. Your feedback is greatly helpful to building our collective intelligence on this vital topic. Thanks for reading.
Natural Alignment - The Future of AI (A Limited Series of Posts)
Overview: The Biomimicry Future of AI (this post)
Acceleration: The Exponential Nature of Leading-Edge Systems (Coming 2023)
Autopoesis: How Life and the Universe Manages Complexity (Coming 2023)
Evo-Devo Values: How Sentience Manages Complexity (Coming 2024)
Stewarding Sentience: Personal, Group, and Network AI (Coming 2024)
Nakul Gupta has a BS in Physics from UCLA, and has recently received an MS in Computer Science from USC. John Smart is a futurist and systems theorist, trained under living systems pioneer James Grier Miller at UCSD, and co-founder of the Evo-Devo Universe complex systems research community.
A refreshing perspective! As John will know, I've long been fascinated with evolutionary approaches to complexity. One of my arguments against a pause or ban for AI is that we will only learn better approaches to safety by moving ahead, letting new approaches evolve, and paying close attention. I look forward to reading more posts in this series.