Bootstrapping Life-Inspired Machine Intelligence: The Biological Route from Chemistry to Cognition and Creativity

The gist The biological route from chemistry to cognition and creativity advocates for a life-inspired approach to machine intelligence by distilling five design principles from biology that enable robustness, autonomy, and open-ended problem-solving across scales.

Diverse Intelligence in Biology and Machines

Intelligence is defined as a fixed goal with variable means of achieving it according to William James, which characterizes intelligence by the capacity to solve problems flexibly and generalize to novel contexts <ref:2602.08079#pg4>. This definition highlights that intelligence is far more diverse than often assumed, manifesting in multifarious forms throughout biology <ref:2602.08079#pg4>. Cognitive capacities are spread along a spectrum with many different competencies, and biological systems at all scales perform the sensing-decision-making-actuation loop in many nested problem spaces besides the familiar 3D space colonized by nerve and muscle <ref:2602.08079#pg3>. Evolution pivoted systems across a variety of problem spaces, starting with metabolic and physiological state spaces, followed by transcriptional space once genes came on the scene, anatomical morphospace once multicellularity appeared, then conventional behavior in 3D space (once nerve and muscle evolved), and now exploring linguistic and many other problem spaces inhabited by humans with culture and intelligent artifacts <ref:2602.08079#pg5>. Crucially, within the boundaries defined by their ecological niches, living organisms show flexible problem solving because key to survival is their capability to deploy their abilities in uncertain and changing conditions <ref:2602.08079#pg5>.

The Cognitive Light Cone and Scalability

The concept of a cognitive light cone characterizes the continuum of intelligence in living systems and machines by representing the size of the goals that a given system can represent and pursue (e.g., in space and time) <ref:2602.08079#pg7>. The more extended an agent’s cognitive light cone is, the greater its ability to flexibly solve increasingly challenging problems <ref:2602.08079#pg7>. Progress in living systems can be understood as the expansion of the scope and timescale over which they predict and control the environment, from immediate, local effects to events that are distal in space and time <ref:2602.08079#pg7>. This expansion is driven by innovation, where outcomes of early recipes are reused as ingredients for later ones, leading to a combinatorial explosion <ref:2602.08079#pg8>. Scalable problem-solving involves the capacity to confront entirely novel problems and pursue entirely new goals in ever-expanding problem spaces, which is necessary because the landscape of problems continually expands <ref:2602.08079#pg9>.

Five Design Principles for Machine Intelligence

The paper distills five design principles that underpin life’s ability to navigate creatively diverse problem spaces <ref:2602.08079#pg11>. These principles are:

1. Autonomy, which means living organisms are natural born autonomous and autopoietic, meaning they must generate and maintain the conditions necessary for their own existence and integrity <ref:2602.08079#pg11>. This autonomy manifests in active sensing beyond passive sensory processing, active (and purposive) control of behavior beyond fixed stimulus-response, and active learning beyond the passive ingestion of training data <ref:2602.08079#pg11>.

2. Growth through self-assemblage of active components, where intelligence is fundamentally multiscale and collective with each level contributing to problem-solving capacities <ref:2602.08079#pg16>. This contrasts with current AI systems where a single, centralized “intelligent” system is imposed on passive, inert hardware <ref:2602.08079#pg27>.

3. Continuous reconstruction of capabilities, where organisms continuously rebuild themselves and achieve adaptation through change rather than freezing “optimal” solutions <ref:2602.08079#pg18>. This involves reconstructing bodies, models, and memories under changing constraints to exploit change as a resource rather than merely tolerate it <ref:2602.08079#pg19>.

4. Exploitation of physical and embodied constraints, where living organisms consistently exploit these constraints to structure behavior, shape and regularize computations <ref:2602.08079#pg21>. Physical medium itself embodies computational regularities, reducing the arbitrariness of computations by creating a natural manifold of plausible solutions <ref:2602.08079#pg22>.

5. Pervasive signaling enabling self-organization and top-down control from goals, where communication enables distributed self-organization and allows for top-down causation in the architecture of life <ref:2602.08079#pg24>. This allows higher levels to specify goals at the level of a field or ensemble while solutions are realized by the coordinated actions of individual agents <ref:2602.08079#pg25>.

Implications for Artificial Intelligence

Translating these lessons suggests that the objective is not to replicate every detail of biological intelligence, but to emulate its ability to flexibly achieve goals, scale up problemsolving capacities, coordinate autonomous components, and adapt robustly across scales <ref:2602.08079#pg29>. The challenge for artificial intelligence is not to copy life, but perhaps to continue it <ref:2602.08079#pg29>. Future AI systems could benefit from architectures that integrate self-directed exploration, active model-building, and hierarchical, temporally extended control—emphasizing learning strategies over fixed solutions <ref:2602.08079#pg27>. By leveraging incremental growth, modular self-assembly, and collective problemsolving, artificial architectures could achieve flexibility and resilience analogous to those observed in biological systems <ref:2602.08079#pg27>.

Conclusion

Living organisms achieve scalable intelligence through the continuous self-assembly of active, agentic components across multiple levels of organization <ref:2602.08079#pg32>. From cells to tissues to whole organisms, nested hierarchies of control and prediction allow local autonomy while coordinating higher-order goals <ref:2602.08079#pg25>. Pervasive signaling illustrates how life exploits communication and collective coordination to achieve robustness, flexibility, and scalability <ref:2602.08079#pg26>. By distributing information across agents and integrating local and global processes, living organisms solve problems far beyond the capacity of individual components <ref:2602.08079#pg26>. Most importantly, pervasive signaling allows for the specification of collective goals—implemented in bioelectric networks or the coordinated activity of large-scale neural populations—that guide the behavior of autonomous smaller components <ref:2602.08079#pg26>. This approach suggests that intelligence emerges from the interplay of distributed signaling, physical constraints, and multiscale coordination <ref:2602.08079#pg25>.

--- Page 1 ---

<ref:2602.08079#pg2>Bootstrapping Life-Inspired Machine Intelligence: The Biological Route from Chemistry to Cognition and Creativity Giovanni Pezzulo, Michael Levin

--- Page 2 ---

<ref:2602.08079#pg3>We live in interesting times: a historical period marked by a widespread confidence that achieving machine intelligence may be within reach. This has sparked intense debate on a range of topics, some of which are primarily theoretical <ref:2602.08079#pg3>. These include whether artificial systems are already endowed—or will soon be endowed—with forms of intelligence (or even general intelligence), autonomy, understanding, or consciousness (and what these terms actually mean); how such abilities compare to those found in biological organisms at different levels of complexity; and what the implications of these technological advances may be for science, society, and ethics <ref:2602.08079#pg3>. Other questions are more technical in nature, such as which methods might be most effective for achieving advanced machine intelligence. Examples include scaling up the currently most successful technologies based on generative (possibly multimodal) AI; enabling autonomous exploration and learning through interaction with the environment to surpass human skill levels; supporting interaction with others to develop shared understanding, cooperative behaviors, and alignment; and to what extent taking inspiration from the neural mechanisms supporting cognition and intelligence in living organisms matters <ref:2602.08079#pg3>. In parallel, an emerging body of research spanning not only cognitive science and neuroscience, but also theoretical biology, molecular biology, and genetics—while also intersecting with artificial intelligence and robotics—aims to understand intelligence across substrates, scales, and origin stories. This line of work is the field of Diverse Intelligence <ref:2602.08079#pg3>. However, the field of machine intelligence has not yet been sufficiently benefitted from the insights of a broad approach to basal cognition in diverse substrates <ref:2602.08079#pg3>. This perspective advances a genuinely life-inspired approach to machine intelligence. Rather than focusing on neural computation alone, it draws on the broader repertoire of problem-solving strategies evolved by living systems, revealing remarkable capabilities—such as self-organization, robustness, and open-ended adaptation—that emerge from the dynamics of life itself and were present long before nervous systems evolved <ref:2602.08079#pg3>. This paper discusses how natural forms self-assemble —from chemistry to advanced cognitive systems—and distills design principles that may prove useful for the development of future lifeinspired intelligent machines, autonomous and embodied AIs, and robots <ref:2602.08079#pg3>. We argue that biology has not only already “built” many forms of intelligence, including ourselves, but has also discovered a recipe for flexible and scalable intelligence. We discuss the key design principles that make this possible and the implications for machine intelligence <ref:2602.08079#pg3>.

--- Page 3 ---

<ref:2602.08079#pg4>Diverse intelligence in biology and machines How can intelligence be defined in a way that is not anchored to our (human) case, but instead applies across the breadth of biological systems—and potentially to machines—without becoming so broad as to lose explanatory power? While intelligence is notoriously difficult and controversial to define, a useful starting point is a definition attributed to William James: “Intelligence is a fixed goal with variable means of achieving it” <ref:2602.08079#pg4>. According to this view, intelligence is characterized by the capacity to solve problems flexibly (i.e., in many ways, including ways for which they were not directly trained), generalizing to novel contexts (“out of distribution,” in machine learning parlance) <ref:2602.08079#pg4>.

1. Autopoietic Goal Generation: AI systems should be designed to be natural born autonomous by continuously generating their own goals rather than being programmed with fixed objectives, as biological organisms are described as being autopoietic: they must generate and maintain the conditions necessary for their own existence and integrity.

2. Multiscale Competency Architecture: Implement systems composed of interacting, semi-autonomous components where intelligence emerges from the coordinated activity of numerous agentic elements, each contributing to problem-solving at its own scale, allowing for capabilities that are multiscale and collective.

3. Continuous Rebuilding for Robustness: Future AI should prioritize continuous reconstruction over preservation, allowing models and goals to be revisable within protected constraints, enabling agents to exploit change as a resource rather than merely tolerate it, addressing the fragility of current systems under changing demands.

4. Constraint Exploitation: Design systems that actively exploit physical and computational laws—such as extreme energy efficiency or conservation laws—as informational scaffolds to reduce arbitrariness and ensure robust, flexible, and scalable intelligence.

5. Pervasive Signaling for Top-Down Control: Integrate pervasive signaling channels (bioelectric, chemical) to allow goals specified at a higher level to exert top-down control over local components, enabling hierarchical coordination while preserving local autonomy.

Achieving advanced machine intelligence remains a central challenge in AI research, often approached through scaling neural architectures and generative models. However, biological systems offer a broader repertoire of strategies for adaptive, goal-directed behavior - strategies that emerged long before nervous systems evolved. This paper advocates a genuinely life-inspired approach to machine intelligence, drawing on principles from biology that enable robustness, autonomy, and open-ended problem-solving across scales. We frame intelligence as flexible problem-solving, following William James, and develop the concept of "cognitive light cones" to characterize the continuum of intelligence in living systems and machines. We argue that biological evolution has discovered a scalable recipe for intelligence - and the progressive expansion of organisms' "cognitive light cone", predictive and control capacities. To explain how this is possible, we distill five design principles - multiscale autonomy, growth through self-assemblage of active components, continuous reconstruction of capabilities, exploitation of physical and embodied constraints, and pervasive signaling enabling self-organization and top-down control from goals - that underpin life's ability to navigate creatively diverse problem spaces. We discuss how these principles contrast with current AI paradigms and outline pathways for integrating them into future autonomous, embodied, and resilient artificial systems.