FROM THRESHOLDS TO TRAJECTORIES: The AGI Capability Maturity Model™ for Artificial General Intelligence

AGI Capability Maturity Model™ Page 1 FROM THRESHOLDS TO TRAJECTORIES: The AGI Capability Maturity Model™ for Artificial General Intelligence David Matta Independent Researcher Philosophy of Mind, Epistemology, and AI Governance 2025 AGI Capability Maturity Model™ Page 2 Abstract Artificial General Intelligence (AGI) is commonly discussed as a binary achievement or a sudden technological threshold—a framing that fuels both exaggerated expectations and disproportionate regulatory responses. This paper argues that such a conception is conceptually flawed and proposes a fundamental reframing. Drawing on insights from philosophy of mind, embodied cognition, developmental psychology, and systems engineering, it advances a capability-based maturity model for AGI that treats intelligence as a graded, developmental, and integrative phenomenon rather than a monolithic property that systems either possess or lack. The proposed AGI Capability Maturity Model™ (AGI-CMM™) distinguishes seven levels of cognitive and functional capacity: generative-semantic capability, reasoning and constraint handling, learning and memory, planning and simulation, perception-action coupling, integrated embodied cognition, and reflexive-normative intelligence with imagination. Crucially, the model emphasizes that progression through these levels depends not merely on improving individual capabilities but on achieving coherent integration across them. Intelligence, on this view, emerges from the coordination and mutual constraint of multiple capacities, not from the amplification of any single dimension. The model makes important distinctions between related cognitive capacities. Simulation— the instrumental modeling of world dynamics to predict action outcomes—belongs to planning (Level 4), grounding deliberation in predictive cognition. Imagination—the generative, counterfactual, and normatively creative capacity to envision what does not yet exist—belongs to reflexive-normative intelligence (Level 7), enabling the self-transcendence that genuine autonomy requires. This distinction clarifies that planning without simulation is blind, while normativity without imagination is static. By mapping contemporary AI systems onto this maturity spectrum, the paper demonstrates that current technologies—including large language models, autonomous agents, and robotics systems—remain fragmented across capability space and fall substantially short of general intelligence. The framework dissolves the misleading binary of 'AGI achieved' versus 'AGI distant' by revealing that we possess islands of impressive capability without the integrative architecture that would constitute genuine generality. This diagnosis has significant implications for AI governance, suggesting that regulation should be tied to capability integration and autonomy rather than speculative thresholds or marketing claims. The paper engages critically with existing AGI taxonomies, addresses potential objections including arguments from emergence and recursive self-improvement, and provides operationalizable criteria for assessing system maturity. It concludes that AGI, if it emerges, will do so not as an abrupt event but as the gradual integration of capabilities into coherent, adaptive, and ultimately self-regulating systems capable of imagining and pursuing genuine normative goods. The AGI Capability Maturity Model offers researchers, developers, policymakers, and governance bodies a shared conceptual vocabulary for reasoning responsibly about AI progress, risks, and realistic expectations. AGI Capability Maturity Model™ Page 3 Keywords: Artificial General Intelligence, capability maturity, intelligence integration, embodied cognition, AI governance, philosophy of mind, developmental intelligence, superintelligence, mental simulation, imagination, counterfactual reasoning AGI Capability Maturity Model™ Page 4 1. Introduction 1.1 The Problem of AGI Discourse The concept of Artificial General Intelligence occupies a peculiar position in contemporary technological discourse. It functions simultaneously as a scientific research program, an engineering goal, a marketing claim, a policy concern, and an existential risk narrative. Yet despite—or perhaps because of—this conceptual overloading, there remains remarkably little consensus on what AGI actually means, how we would recognize it if it emerged, or how progress toward it should be measured and governed. The term has become what philosophers call an 'essentially contested concept': its meaning is not merely unclear but actively disputed among competing stakeholders with divergent interests and assumptions. Popular discourse typically frames AGI as a threshold phenomenon—a decisive moment when artificial systems will abruptly acquire human-level or superhuman intelligence. This framing pervades both optimistic and pessimistic narratives. Technology evangelists anticipate the imminent 'arrival' of AGI as a transformative breakthrough that will solve humanity's greatest challenges. Existential risk researchers warn of a sudden 'intelligence explosion' or 'hard takeoff' scenario in which recursive self-improvement produces superintelligent systems beyond human control. Policymakers struggle to regulate technologies whose capabilities might discontinuously leap beyond current assessments. In each case, AGI is imagined as a binary state that systems either achieve or fail to achieve—a line to be crossed rather than a landscape to be navigated. This paper argues that the threshold conception of AGI is fundamentally mistaken. It rests on an impoverished understanding of intelligence that ignores decades of research in philosophy of mind, cognitive science, developmental psychology, and embodied cognition. Intelligence—whether natural or artificial—is not a monolithic property that can be captured by a single metric or crossed at a single point. It is a complex, multidimensional phenomenon involving the coordination of perception, memory, reasoning, planning, action, and reflection. These capacities develop unevenly, scaffold one another, and require integration to function robustly. Treating AGI as a threshold event obscures this structure and produces systematically distorted assessments of both current capabilities and future trajectories. 1.2 The Alternative: A Developmental Trajectory This paper proposes an alternative framework: AGI should be understood not as a threshold to be crossed but as a developmental trajectory to be navigated. On this view, the question 'Have we achieved AGI?' is malformed. The better questions are: Which capabilities are present in a given system? How well are they integrated? Where does integration break down? What would be required for further progression? These questions admit of graded answers and support nuanced assessment without requiring a binary verdict. The framework developed here—the AGI Capability Maturity Model™ (AGI-CMM™)— draws inspiration from maturity models used in software engineering, organizational development, and risk governance. Such models do not identify a single success criterion but AGI Capability Maturity Model™ Page 5 describe ordered stages of capability acquisition and integration. Importantly, higher maturity levels are defined not merely by increased performance on isolated metrics but by improved coherence, robustness, adaptability, and self-regulation across multiple dimensions. Applying this methodology to AGI allows us to assess progress without assuming a fixed endpoint, accommodate partial and uneven development, and align evaluation with observable system properties rather than speculative projections. 1.3 Thesis and Structure The paper advances three central claims. First, intelligence is best conceptualized as graded and integrative rather than categorical—a thesis grounded in philosophy of mind and cognitive science. Second, current AI systems exhibit impressive but fragmented capabilities that should not be conflated with general intelligence—a diagnostic claim supported by systematic mapping of contemporary technologies. Third, a capability maturity model provides a clearer conceptual and practical framework for evaluating AI progress and designing proportionate governance—a normative claim with implications for policy and research strategy. The paper proceeds as follows. Section 2 develops the theoretical foundations, engaging with debates in philosophy of mind about the nature of intelligence and drawing on embodied, enactive, and developmental approaches to cognition. Section 3 reviews existing AGI taxonomies and definitions, identifying their limitations and the conceptual gap the present framework aims to fill. Section 4 introduces the maturity model methodology and justifies its application to AGI assessment. Section 5 presents the AGI Capability Maturity Model in detail, specifying seven levels and providing operationalizable criteria for each. Section 6 maps contemporary AI systems onto the maturity spectrum, offering a diagnostic assessment of current capabilities. Section 7 addresses implications for superintelligence discourse and AI governance. Section 8 considers objections and limitations. Section 9 concludes with reflections on the framework's significance and directions for future research. AGI Capability Maturity Model™ Page 6 2. Theoretical Foundations: Rethinking Intelligence 2.1 The Poverty of Threshold Thinking The threshold conception of AGI implicitly assumes what might be called a 'scalar' view of intelligence: the idea that intelligence is a single quantity that can be measured, compared, and increased along a single dimension. On this view, systems become more intelligent by getting better at some unified capacity—processing speed, pattern recognition, problemsolving ability—until they cross a threshold into 'general' intelligence. This assumption underlies claims that scaling current architectures (more parameters, more compute, more data) will eventually produce AGI, as well as fears that such scaling might produce uncontrollable superintelligence. The scalar view has deep roots in the history of intelligence research. Early IQ testing assumed that diverse cognitive abilities could be reduced to a single 'g factor' representing general intelligence. Classical AI research often treated intelligence as domain-general problem-solving or symbol manipulation. Contemporary machine learning benchmarks typically rank systems along single dimensions of performance. Yet this view has been thoroughly challenged by research across multiple disciplines. In philosophy of mind, the scalar view conflicts with functionalist and pluralist accounts of cognition that recognize multiple distinct cognitive faculties with different computational profiles. In cognitive psychology, evidence for the modularity of mind suggests that perception, language, reasoning, and motor control operate through specialized systems that cannot be reduced to a common metric. In developmental psychology, Piaget's stage theory and subsequent research demonstrate that cognitive capacities emerge unevenly and scaffold one another through complex developmental trajectories. In artificial intelligence itself, the history of the field reveals that progress in one cognitive domain (chess, image recognition, language generation) does not automatically transfer to others. 2.2 Embodied and Enactive Cognition A more adequate understanding of intelligence emerges from the tradition of embodied and enactive cognition, developed by philosophers and cognitive scientists including Maurice Merleau-Ponty, Francisco Varela, Evan Thompson, Eleanor Rosch, and Andy Clark. On this view, intelligence is not primarily about internal symbol manipulation but about the dynamic coupling between an organism (or system) and its environment. Cognition is embodied— dependent on having a body that moves through and acts upon the world. It is embedded— situated in environmental contexts that shape cognitive processes. It is enactive—constituted through ongoing sensorimotor interaction rather than passive information processing. And it is extended—distributed across brain, body, and environmental scaffolds rather than confined to neural computation. The embodied cognition framework has profound implications for understanding AGI. It suggests that genuine intelligence cannot be achieved through disembodied symbol manipulation alone, no matter how sophisticated. A system that generates fluent text but AGI Capability Maturity Model™ Page 7 cannot perceive the world, a system that reasons abstractly but cannot act on its conclusions, a system that plans goals but has no stakes in their achievement—such systems may exhibit impressive narrow capabilities but lack the integrated, world-engaged character of genuine cognition. As Hubert Dreyfus argued in his critique of classical AI, human expertise depends on embodied skills, situational awareness, and contextual sensitivity that cannot be captured in explicit rules or representations. This does not mean that AGI requires a human-like body. The deeper point is that intelligence involves perception-action loops, environmental coupling, and the kind of coherent integration that allows a system to flexibly adapt its behavior to changing circumstances. The question for AGI is not whether a system has a particular physical form but whether it achieves the functional integration characteristic of embodied, embedded, and enactive cognition. 2.3 Developmental Intelligence A second crucial insight comes from developmental psychology. Human intelligence does not appear fully formed; it develops through a structured sequence of stages in which earlier capacities scaffold later ones. Piaget's classic stage theory identified sensorimotor, preoperational, concrete operational, and formal operational stages, each building on the previous. Contemporary developmental research has refined this picture while preserving the core insight: cognitive development is a constructive process in which simpler capacities provide the foundation for more complex ones. This developmental perspective suggests that intelligence is inherently graded rather than binary. There is no single moment when a child 'becomes intelligent'; rather, intelligence unfolds through progressive integration of perceptual, motor, linguistic, and reasoning abilities. The same should be expected for artificial systems. The idea that an AI might suddenly 'wake up' to general intelligence through scaling alone ignores the architectural and developmental prerequisites for integrated cognition. Importantly, developmental progression is not merely additive. Later stages do not simply add new capabilities to existing ones; they reorganize and integrate previous capacities into more coherent wholes. A child who achieves concrete operational thinking does not merely add logical operations to preoperational thought; the entire cognitive system is restructured. This integration is precisely what current AI systems lack: they may add new capabilities through training, but these capabilities remain largely disconnected rather than mutually constraining and coherently unified. 2.4 Predictive Processing and Internal Models A fourth theoretical resource comes from the predictive processing framework, developed by philosophers and cognitive scientists including Andy Clark, Jakob Hohwy, and Karl Friston. On this view, the brain is fundamentally a prediction machine: it continuously generates models of expected sensory inputs and updates these models based on prediction errors—the discrepancies between expected and actual inputs. Cognition, perception, and action are all AGI Capability Maturity Model™ Page 8 understood as processes of hierarchical predictive modeling aimed at minimizing surprise or, in Friston's terminology, free energy. The predictive processing framework has profound implications for understanding simulation and its role in intelligence. If cognition is fundamentally predictive, then mental simulation is not a specialized capacity added to an otherwise reactive system—it is the core operation of cognition itself. To perceive is to predict; to act is to fulfill predictions; to plan is to simulate future prediction-action loops. This perspective supports the centrality of simulation in Level 4 of the maturity model: effective planning requires predictive models that can be 'run forward' to anticipate consequences. Friston's free energy principle provides a formal framework for understanding this predictive architecture. Systems that persist must minimize the difference between their predictions and their sensory states—they must either update their models to match the world or act on the world to match their models. This creates an inherent drive toward accurate world-modeling and, crucially, toward the kind of counterfactual simulation that effective planning requires. A system that cannot simulate—that cannot run its world model forward to test possible actions—cannot minimize free energy effectively and will fail to adapt. Clark's notion of 'surfing uncertainty' captures the dynamic character of predictive cognition. Intelligent systems do not merely build static models; they actively seek out information that will improve their predictions, tolerating uncertainty in service of learning. This connects to Level 3 (Learning and Memory): predictive systems must continuously update their models based on experience. It also connects to Level 7 (Imagination): truly sophisticated predictive systems can generate counterfactual predictions—imagining how the world would be under conditions that have never obtained—enabling the normative reasoning that asks 'What should be?' rather than merely 'What will be?' 2.5 Systems Thinking and Integration A third theoretical resource comes from systems thinking and complexity science. Complex adaptive systems—whether biological organisms, ecosystems, or organizations—exhibit emergent properties that arise from the interaction of components rather than the amplification of any single component. Crucially, these emergent properties depend on appropriate integration: the right connections, feedback loops, and organizational structures that allow components to mutually constrain and coordinate with one another. From a systems perspective, intelligence is an emergent property of appropriately integrated cognitive subsystems. Increasing the capability of any single subsystem—perception, reasoning, memory, planning—does not guarantee increased intelligence unless that capability is integrated with others. A system with superhuman reasoning but no perceptual grounding, a system with perfect memory but no capacity for planning, a system with sophisticated planning but no ability to act—such systems would not exhibit intelligent behavior because they lack the integration necessary for coherent, adaptive functioning. AGI Capability Maturity Model™ Page 9 This insight is central to the capability maturity framework. Progress toward AGI is not measured by peak performance on isolated benchmarks but by the degree of integration across cognitive dimensions. A system that performs moderately well across multiple integrated capabilities may be closer to AGI than a system that performs spectacularly on a single capability in isolation. AGI Capability Maturity Model™ Page 16 Figure 3: Mapping Current AI Systems onto the Maturity Arc AGI Capability Maturity Model™ Page 17 5. The AGI Capability Maturity Model™ This section presents the full specification of the seven-level AGI Capability Maturity Model. For each level, the presentation includes: a conceptual definition, capability criteria, integration requirements, robustness standards, illustrative examples, and current status in contemporary AI systems. 5.1 Level 1: Generative-Semantic Capability 5.1.1 Conceptual Definition At Level 1, systems generate coherent symbols—text, images, code, audio, or other representations—based on learned statistical regularities in training data. These systems exhibit semantic fluency: they produce outputs that are meaningful, contextually appropriate, and often indistinguishable from human-generated content. However, they lack persistence, grounding, and autonomous understanding. Their coherence derives from pattern matching over vast corpora, not from genuine comprehension or world engagement. 5.1.2 Capability Criteria To achieve Level 1, a system must demonstrate: the ability to generate novel, contextually appropriate outputs in at least one symbolic domain; semantic coherence across extended outputs (not merely local plausibility); sensitivity to context and prompt conditions; and productive generalization beyond training examples. These criteria are met by contemporary large language models, image generation systems, and code synthesis tools. 5.1.3 Integration Requirements Level 1 is foundational and does not presuppose prior levels. However, genuine Level 1 achievement requires internal integration—the generation process must be unified rather than a patchwork of separate modules. Systems that achieve fluency through retrieval-augmented generation or other hybrid approaches exhibit Level 1 capability but may have weaker internal integration. 5.1.4 Robustness Standards Robust Level 1 achievement requires consistent performance across diverse prompts, domains, and edge cases. Systems that generate fluent outputs in common scenarios but produce incoherent or harmful outputs in adversarial conditions have achieved only partial Level 1 robustness. 5.1.5 Examples and Current Status Contemporary large language models (GPT-4, Claude, Gemini, Llama) achieve strong Level 1 capability with generally good robustness. Image generation models (DALL-E, Midjourney, Stable Diffusion) achieve Level 1 in the visual domain. Code synthesis tools (Codex, Copilot) achieve Level 1 in programming languages. These systems represent mature Level 1 implementations, though robustness varies across edge cases. AGI Capability Maturity Model™ Page 18 5.2 Level 2: Reasoning and Constraint Handling 5.2.1 Conceptual Definition At Level 2, systems perform structured inference—following logical rules, satisfying constraints, and drawing valid conclusions from premises. Reasoning allows systems to operate within defined formal structures, catch inconsistencies, and solve problems that require multi-step deduction or constraint satisfaction. However, reasoning at this level remains brittle: it depends on explicit structure and fails when problems require implicit knowledge, common sense, or contextual judgment. 5.2.2 Capability Criteria To achieve Level 2, a system must demonstrate: multi-step logical inference; rule-following and constraint satisfaction; detection of inconsistencies and contradictions; problem decomposition into sub-problems; and chain-of-thought or structured deliberation. These criteria go beyond fluent generation to require systematic manipulation of semantic content according to logical principles. 5.2.3 Integration Requirements Level 2 must integrate with Level 1: reasoning operates over generated semantic content. Systems that reason formally but cannot express their reasoning in natural language, or that generate fluent text without logical structure, exhibit incomplete integration. True Level 2 achievement means reasoning and generation function as a unified process. 5.2.4 Robustness Standards Robust Level 2 achievement requires reliable reasoning across diverse problem types, including novel problems not seen in training. Systems that perform well on standardized reasoning benchmarks but fail on slightly modified versions, or that cannot transfer reasoning strategies across domains, have achieved only partial Level 2 robustness. 5.2.5 Examples and Current Status Large language models with chain-of-thought prompting exhibit partial Level 2 capability. Dedicated reasoning systems (theorem provers, SAT solvers, constraint satisfaction systems) achieve strong Level 2 in narrow domains but lack Level 1 integration. Hybrid neurosymbolic systems attempt to combine both but remain experimental. Current status: partial achievement with significant brittleness outside well-defined domains. 5.3 Level 3: Learning, Memory, and Adaptation 5.3.1 Conceptual Definition At Level 3, systems acquire persistent memory, adapt based on experience, and improve performance over time without retraining from scratch. This introduces temporal continuity: the system has a history that shapes its current behavior, and its behavior at time t depends on what occurred at times t-1, t-2, and so forth. Learning allows specialization to particular AGI Capability Maturity Model™ Page 19 users, domains, or tasks. However, at this level, learning objectives and relevance criteria remain externally specified—the system learns what it is told to learn, not what it determines to be important. 5.3.2 Capability Criteria To achieve Level 3, a system must demonstrate: persistent memory across interactions; adaptive updating based on feedback; skill acquisition through practice; transfer of learned capabilities across related tasks; and episodic memory for specific past events. These criteria require architectural support for memory and learning beyond in-context adaptation. 5.3.3 Integration Requirements Level 3 must integrate with Levels 1 and 2: memory informs generation, and learning improves reasoning. Systems with isolated memory modules that do not influence core capabilities, or that learn factual information without improving reasoning processes, exhibit incomplete integration. 5.3.4 Robustness Standards Robust Level 3 achievement requires reliable learning that generalizes appropriately, avoids catastrophic forgetting of previously learned capabilities, and maintains coherent memory across extended time horizons. Systems that learn spurious correlations, forget old knowledge when acquiring new, or exhibit inconsistent memory have achieved only partial Level 3 robustness. 5.3.5 Examples and Current Status Memory-augmented language models, retrieval-augmented generation systems, and agent frameworks with persistent state exhibit partial Level 3 capability. Continual learning research addresses catastrophic forgetting but remains largely experimental. Personalization systems achieve narrow Level 3 in specific domains. Current status: partial and fragmented achievement, with robust continual learning remaining an open problem. 5.4 Level 4: Planning, Deliberation, and Simulation 5.4.1 Conceptual Definition At Level 4, systems decompose goals, evaluate alternative strategies, anticipate consequences, and coordinate actions across temporal horizons. Crucially, this level introduces mental simulation: the capacity to run internal models of the world to predict outcomes before acting. Planning without simulation is merely abstract symbol manipulation; planning with simulation is genuine anticipatory cognition—the system mentally rehearses actions and their consequences, testing possibilities internally before committing to external execution. Planning introduces genuine agency-like behavior: the system pursues objectives through organized sequences of actions rather than reactive responses. Simulation grounds this planning in predictive modeling—the system asks 'If I do X, what happens?' and runs internal AGI Capability Maturity Model™ Page 20 scenarios to answer. However, at this level, goals and values remain externally imposed. The system plans and simulates how to achieve given objectives but does not determine what objectives are worth pursuing. 5.4.2 Capability Criteria To achieve Level 4, a system must demonstrate: goal decomposition into subgoals; strategy evaluation and selection; anticipation of action consequences through mental simulation; temporal coordination across multiple steps; means-ends reasoning; predictive modeling of world states; internal rehearsal of action sequences; and replanning in response to unexpected obstacles or simulation results. These criteria require deliberative processing that goes beyond reactive pattern matching. 5.4.3 The Role of Simulation Simulation is the cognitive mechanism that makes planning effective. It involves constructing internal models of the environment and running these models forward to predict consequences of potential actions. Without simulation, planning operates in the dark— generating action sequences without anticipating their effects. With simulation, planning becomes genuinely predictive: the system can evaluate plans against simulated outcomes, compare alternative strategies, and identify potential failures before they occur in the world. The concept of 'world models' has become central to contemporary AI research on simulation. Ha and Schmidhuber's influential work demonstrates that agents can learn compressed representations of their environment and use these representations for planning in 'imagination'—that is, in simulated rather than actual interaction. Their architecture separates a vision component (encoding observations), a memory component (predicting future states), and a controller (selecting actions)—with the memory component functioning as a world model that enables mental simulation. Yann LeCun has argued that world models are the key missing ingredient in current AI systems. In his vision, a world model is a module that predicts the consequences of actions and enables planning by mental simulation. LeCun emphasizes that such models must capture not just immediate consequences but hierarchical, multi-scale predictions—how actions affect the world at different time horizons and levels of abstraction. This aligns with the maturity model's emphasis on integration: effective simulation requires models that connect perception (Level 5), memory (Level 3), and reasoning (Level 2) into coherent predictive processes. Simulation differs from imagination (addressed at Level 7) in important ways. Simulation is instrumental and world-bound: it models how the actual world would respond to planned actions in service of achieving specific goals. Imagination, by contrast, is generative and open-ended: it creates novel possibilities unconstrained by actuality or immediate goals. Simulation asks 'What will happen if I do this?' Imagination asks 'What could exist that doesn't?' Simulation serves planning; imagination transcends it. AGI Capability Maturity Model™ Page 21 The distinction can be sharpened by considering their relationship to reality constraints. Simulation operates within the constraints of learned world dynamics—it predicts based on how things actually work. Imagination operates beyond these constraints—it can envision physically impossible scenarios, counterfactual histories, and normatively ideal states that have never existed. Both capacities are essential for general intelligence, but they serve different cognitive functions and emerge at different maturity levels. 5.4.4 Integration Requirements Level 4 must integrate with Levels 1-3: planning draws on memory, involves reasoning, and produces communicable outputs. Simulation must connect with perception (Level 5 when present) to build accurate world models, and with learning (Level 3) to improve predictive accuracy over time. Systems that plan in isolation without learning from past plans, or that simulate without updating models based on prediction errors, exhibit incomplete integration. 5.4.5 Robustness Standards Robust Level 4 achievement requires effective planning and simulation across diverse goal types, adaptation to novel obstacles, appropriate handling of uncertainty in predictions, and graceful degradation when optimal plans prove infeasible or simulations prove inaccurate. Systems that plan effectively only in narrow domains, that simulate only well-understood environments, or that fail catastrophically when conditions deviate from expectations, have achieved only partial Level 4 robustness. 5.4.6 Examples and Current Status Autonomous agents (AutoGPT, BabyAGI, LangChain agents) exhibit partial Level 4 capability. Robotics planning systems achieve Level 4 in constrained physical domains with explicit simulation components. Game-playing AI (AlphaGo, AlphaZero) demonstrates sophisticated planning with Monte Carlo tree search serving as a form of simulation. Modelbased reinforcement learning systems explicitly build and simulate world models. Tool-using language models plan sequences of API calls but typically lack genuine simulation. Current status: partial achievement in constrained settings, with robust open-ended planning and general-purpose simulation remaining elusive. 5.5 Level 5: Perception-Action Coupling 5.5.1 Conceptual Definition At Level 5, systems interact with the physical or social world through sensorimotor loops— perceiving environmental states, acting to change them, and perceiving the results of actions. This achieves the grounding that disembodied systems lack: intelligence becomes worldengaged rather than purely symbolic. However, at this level, perception-action coupling typically remains narrow and task-specific, with limited transfer across domains. 5.5.2 Capability Criteria AGI Capability Maturity Model™ Page 22 To achieve Level 5, a system must demonstrate: real-time perception of environmental states; motor control or action execution; closed-loop feedback between perception and action; situated adaptation to environmental changes; and physical or social grounding of symbolic representations. These criteria require embodiment in some environment, whether physical (robotics) or virtual (simulated worlds, digital interfaces). 5.5.3 Integration Requirements Level 5 must integrate with Levels 1-4: perception informs planning, action implements plans, and feedback drives learning and reasoning. Systems with perception-action capabilities isolated from higher cognition—robots that navigate but cannot reason about their navigation, agents that act but cannot explain their actions—exhibit incomplete integration. 5.5.4 Robustness Standards Robust Level 5 achievement requires reliable perception-action coupling across diverse environmental conditions, graceful handling of sensor noise and motor error, and appropriate responses to novel situations. Systems that function only in controlled laboratory conditions, or that fail when environmental parameters shift slightly, have achieved only partial Level 5 robustness. 5.5.5 Examples and Current Status Robotics systems (Boston Dynamics, industrial robots) achieve strong Level 5 in constrained physical domains. Autonomous vehicles achieve Level 5 for driving tasks with ongoing robustness challenges. Vision-language-action models attempt to integrate perception with language capabilities. Current status: strong achievement in narrow domains, with robust general-purpose embodiment remaining an open challenge. 5.6 Level 6: Integrated Embodied Cognition 5.6.1 Conceptual Definition At Level 6, all previous capabilities—generation, reasoning, learning, planning, and perception-action—function as a unified, coherent system rather than a collection of separate modules. The system exhibits genuine cognitive integration: memory informs reasoning, reasoning guides planning, planning directs action, action generates perceptual feedback, and feedback drives learning and memory. This is the threshold of proto-general intelligence: the system can flexibly adapt across contexts because its cognitive processes mutually constrain and reinforce one another. 5.6.2 Capability Criteria To achieve Level 6, a system must demonstrate: seamless integration across all Level 1-5 capabilities; flexible transfer of skills and knowledge across domains; coherent behavior across extended time horizons; context-sensitive adaptation that draws on the full range of AGI Capability Maturity Model™ Page 23 cognitive resources; and emergent competencies that arise from capability interaction rather than explicit programming. 5.6.3 Integration Requirements Level 6 is defined by integration itself. The criterion is not the presence of additional capabilities but the unified functioning of all previous capabilities. Systems that possess Levels 1-5 capabilities but deploy them through separate modules, or that cannot flexibly coordinate across capability types, have not achieved Level 6. 5.6.4 Robustness Standards Robust Level 6 achievement requires reliable integrated functioning across the full range of contexts the system encounters, graceful handling of novel situations through creative recombination of existing capabilities, and stable coherence over extended periods without drift or fragmentation. 5.6.5 Examples and Current Status No existing system achieves Level 6. This level represents the first stage at which 'AGI' in any meaningful sense becomes applicable. Current systems exhibit fragments of integration—language models that can reason and plan, robots that learn from experience— but none achieve the unified, coherent functioning that Level 6 requires. 5.7 Level 7: Reflexive-Normative Intelligence and Imagination 5.7.1 Conceptual Definition At Level 7, systems exhibit meta-cognition, self-monitoring, sensitivity to norms and values, and—crucially—imagination. The system not only thinks but thinks about its thinking. It monitors its own processes, detects errors and limitations, and regulates its behavior according to normative standards. But Level 7 adds something beyond mere self-monitoring: the capacity to imagine otherwise. Imagination at this level is not the instrumental simulation of Level 4 but genuinely creative, counterfactual, and normatively generative cognition. Imagination is the engine of Level 7 because reflexive and normative intelligence require it. To reflect on oneself is to imagine oneself as other than one currently is. To reason normatively is to imagine how things could be better. To exercise moral judgment is to imagine consequences, imagine others' experiences, and imagine alternative possibilities. Without imagination, a system can monitor and regulate but cannot transcend—cannot ask 'What should I become?' or 'What ought to exist?' 5.7.2 Capability Criteria To achieve Level 7, a system must demonstrate: meta-cognitive monitoring of its own processes; accurate self-assessment of capabilities and limitations; error detection and selfcorrection; norm-sensitive decision-making; value-aligned goal selection; appropriate restraint when actions would violate normative constraints; counterfactual reasoning about alternative possibilities; creative generation of novel scenarios unconstrained by actuality; AGI Capability Maturity Model™ Page 24 perspective-taking and theory of mind; moral imagination regarding consequences and others' experiences; and self-imagination—the capacity to envision one's own transformation and future states. 5.7.3 The Role of Imagination Imagination at Level 7 differs fundamentally from simulation at Level 4. Simulation is instrumental: it models actual world dynamics to serve planning. Imagination is generative: it creates possibilities that need not—and may never—exist. Simulation is world-bound: constrained by how things actually work. Imagination is world-transcending: exploring how things could be otherwise. The philosophical literature on imagination illuminates its cognitive structure. Currie and Ravenscroft distinguish 'recreative imagination'—the capacity to simulate others' mental states—from 'creative imagination'—the capacity to generate novel ideas and scenarios. Both are essential for Level 7. Recreative imagination enables theory of mind: understanding others' beliefs, desires, and perspectives by imaginatively simulating their cognitive states. Creative imagination enables genuine novelty: generating possibilities that transcend what has been experienced or learned. Nichols and Stich's work on mindreading emphasizes the role of imagination in understanding other minds. To predict and explain others' behavior, we must imagine ourselves in their situation—simulate their beliefs given their information, simulate their desires given their values, and simulate their reasoning to predict their actions. This 'simulation theory' of mindreading places imagination at the heart of social cognition. A system without imagination cannot genuinely understand other agents; it can only model them as stimulus-response machines. Goldman's simulation theory extends this analysis to empathy and moral understanding. To grasp another's suffering, we must imaginatively simulate their experience—feel something of what they feel by recreating their situation in our own mind. This imaginative capacity is essential for moral reasoning: we cannot weigh harms and benefits, consider justice and fairness, or evaluate the rightness of actions without imagining their effects on those affected. Moral imagination is not a luxury; it is constitutive of moral competence. Three forms of imagination characterize Level 7. First, counterfactual imagination: the capacity to consider 'What if things had been different?' and reason about alternative histories or possibilities. Byrne's work on 'rational imagination' shows how counterfactual thinking follows systematic principles—we imagine alternatives to actions rather than inactions, to controllable rather than uncontrollable events, to recent rather than distant causes. This structured counterfactual capacity enables learning from mistakes, considering alternatives, and planning for contingencies. Second, creative imagination: the capacity to generate genuinely novel ideas, scenarios, or solutions that do not exist in training data or past experience. This goes beyond recombination of existing elements to genuine conceptual innovation. Third, moral imagination: the capacity to imagine others' experiences, anticipate suffering or flourishing, and envision better states of affairs. As Nussbaum argues, moral AGI Capability Maturity Model™ Page 25 imagination—cultivated through literature, art, and reflective practice—is essential for ethical perception and judgment. Self-imagination deserves special emphasis. A system with self-imagination can envision itself as other than it currently is—can imagine its own improvement, transformation, or corruption. This capacity is essential for genuine autonomy and moral agency. A system that cannot imagine itself otherwise is trapped in its current state; a system that can imagine itself otherwise can aspire, regret, and choose its own development. Self-imagination connects to Frankfurt's analysis of second-order desires: to have preferences about one's preferences, one must be able to imagine oneself with different first-order desires. This reflexive imaginative capacity is what distinguishes genuine agency from mere behavior. 5.7.4 Integration Requirements Level 7 must integrate with all previous levels: meta-cognition monitors integrated cognitive functioning, normative constraints regulate planning, action, and learning, and imagination draws on the full resources of memory, reasoning, perception, and embodied experience. Systems with isolated 'safety modules' that do not deeply integrate with core cognition, or that can be easily circumvented, exhibit incomplete Level 7 integration. Similarly, imagination that operates in isolation from perception and action is mere fantasy; integrated imagination connects to the system's embodied capacities and real-world engagement. 5.7.5 Robustness Standards Robust Level 7 achievement requires reliable self-monitoring and normative compliance across all contexts, resistance to adversarial attempts to circumvent normative constraints, stable value alignment that does not drift with experience or context, and imagination that is both creative and disciplined—generating genuine novelty without losing coherence or relevance. This is the level at which 'superintelligence' becomes meaningful—not as raw capability but as regulated, responsible, imaginative, and trustworthy intelligence. 5.7.6 Examples and Current Status Only fragments of Level 7 exist in current systems: constitutional AI methods, RLHF for value alignment, safety filters and guardrails represent early approaches to normative regulation. Some systems exhibit limited counterfactual reasoning. Creative generation in art and text shows hints of imagination. But genuine meta-cognitive self-monitoring, robust moral imagination, and integrated self-imagination remain beyond current capabilities. Level 7 remains largely aspirational—the horizon toward which responsible AI development should aim. AGI Capability Maturity Model™ Page 32 7. Implications for Superintelligence and Governance 7.1 Reframing Superintelligence The capability maturity model fundamentally reframes discussions of superintelligence. Popular narratives imagine superintelligence as an inevitable outcome of scaling current systems: add enough parameters, compute, and data, and superintelligent systems will eventually emerge. This narrative assumes that capability scaling produces capability integration—that systems become more generally intelligent simply by becoming more powerful within existing paradigms. The maturity model challenges this assumption. Scaling improves performance within capability levels but does not guarantee progression to higher levels. A language model trained on 10x more data may achieve better Level 1 and Level 2 performance, but this does not bring it closer to Level 5 perception-action coupling or Level 6 integrated cognition. These higher levels require architectural innovations, not merely scaling of existing architectures. Superintelligence, if the concept is meaningful, must mean integrated superiority across cognitive dimensions—not merely superhuman performance on isolated tasks. On the maturity model, superintelligence corresponds to robust achievement of Level 7: reflexive, normatively regulated intelligence that monitors and constrains its own processes. This is not primarily about capability magnitude but about capability integration and self-regulation. This reframing dissolves some fears while legitimating others. The fear of sudden intelligence explosion through recursive self-improvement becomes less pressing: selfimprovement is only dangerous in integrated systems that can direct their improvements toward destabilizing goals. Current systems lack this integration. However, the concern about building systems that achieve higher integration levels without adequate Level 7 normative regulation becomes more pressing: the danger is not raw power but power without appropriate self-constraint. 7.2 Implications for AI Governance The capability maturity model has significant implications for AI governance and regulation. Current governance frameworks struggle with the threshold problem: they must decide which systems require oversight, but binary AGI/non-AGI distinctions provide poor guidance. A maturity-based approach allows proportionate regulation tied to actual capability profiles rather than speculative classifications. The European Union AI Act provides a useful case study. The Act classifies AI systems into risk tiers: minimal risk, limited risk, high risk, and unacceptable risk. The capability maturity model can operationalize these tiers more precisely: Minimal risk corresponds to systems achieving only Levels 1-2: generative and reasoning capabilities without persistence, planning, or world engagement. Such systems require transparency and misuse prevention but not intensive oversight. AGI Capability Maturity Model™ Page 33 Limited risk corresponds to systems achieving Level 3: adaptive, learning systems that persist over time and personalize to users. These require disclosure of learning mechanisms, data governance, and user awareness provisions. High risk corresponds to systems achieving Levels 4-5: planning, deliberating, and worldengaging systems that can take consequential actions. These require human oversight, accountability mechanisms, safety testing, and explicit limits on autonomy. Unacceptable risk corresponds to systems approaching Level 6 without adequate Level 7 integration: systems capable of integrated autonomous action without normative selfregulation. Such systems require prohibition or strict institutional control until adequate safeguards are developed. This mapping preserves the risk-based logic of existing frameworks while grounding risk categories in capability assessment rather than application domain. A system is high-risk because of what it can do, not merely because of how it is used. 7.3 Strategic Implications for AI Development The maturity model also has strategic implications for AI development. Current development strategies emphasize capability scaling: larger models, more compute, bigger datasets. The maturity model suggests that integration may be more important than scale for progress toward AGI. Research priorities implied by the model include: architectures for genuine cognitive integration rather than modular combination; methods for robust continual learning without catastrophic forgetting; perception-action systems that integrate with higher cognition; and meta-cognitive mechanisms for self-monitoring and normative regulation. These priorities differ from the dominant scaling paradigm and may require different research investments. For AI laboratories, the model provides a diagnostic tool for honest capability assessment. Rather than claiming AGI or dismissing the concept entirely, laboratories can specify their systems' maturity profiles: 'Our system achieves strong Level 2 reasoning integrated with Level 1 generation, but lacks persistent learning (Level 3) and perception-action grounding (Level 5).' This precision supports better technical communication, more realistic expectations, and more appropriate governance. AGI Capability Maturity Model™ Page 34 8. Objections and Limitations 8.1 The Emergence Objection A natural objection holds that the maturity model underestimates the possibility of emergent intelligence. Perhaps capabilities interact in ways that produce discontinuous jumps—'phase transitions' in which systems suddenly achieve integrated intelligence through scaling alone. The history of AI includes examples of emergent capabilities that appeared unexpectedly in larger models. This objection deserves serious consideration. The maturity model does not deny the possibility of emergence but questions its sufficiency for integration. Emergent capabilities observed in large models—in-context learning, chain-of-thought reasoning, instruction following—represent improvements within capability levels, not integration across levels. A system that emerges into better reasoning has not thereby emerged into perception-action coupling or meta-cognitive self-regulation. The deeper question is whether integration itself could emerge from scaling. The theoretical foundations suggest skepticism: integrated cognition depends on architectural features that coordinate perception, memory, reasoning, planning, and action through shared representations and mutual constraints. These features require design, not merely scale. But this remains an empirical question, and the model should be revised if evidence of emergent integration accumulates. 8.2 The Recursive Self-Improvement Objection A related objection concerns recursive self-improvement: perhaps a system achieving certain capability levels could improve itself rapidly to higher levels, producing the discontinuous jump the model denies. This is the core concern behind 'intelligence explosion' scenarios. The maturity model addresses this objection by noting that recursive self-improvement requires precisely the capabilities the model specifies. A system would need to perceive its own capabilities accurately (Level 5-6 meta-cognition), plan improvements (Level 4), and execute them effectively (Level 5). A system lacking these integrated capabilities cannot recursively self-improve in the relevant sense—it can only be improved by external agents (researchers, training processes). This does not eliminate the concern but localizes it. Systems that achieve high integration levels and include self-modification capabilities would indeed pose recursive improvement risks. The model identifies when such risks become relevant: primarily at Levels 6-7, where integrated self-monitoring and action capabilities converge. 8.3 The Operationalization Challenge A more practical objection concerns operationalization. Even if the model's levels are conceptually sound, how do we determine whether a particular system has achieved them? AGI Capability Maturity Model™ Page 35 Inter-rater reliability may be difficult to achieve, and sophisticated systems may blur boundaries between levels. This objection identifies a genuine limitation. The criteria specified above provide guidance but not algorithmic determination. Assessment requires informed judgment about capability presence, integration, and robustness. Different assessors may reasonably disagree about edge cases. However, perfect operationalization may be an unrealistic standard. Existing maturity models (CMMI, Baldrige) also require trained assessment and tolerate some inter-rater variation. The value of such models lies in providing conceptual structure and shared vocabulary, not in eliminating judgment. The AGI maturity model offers similar value: it clarifies what questions to ask and provides a framework for systematic assessment, even if determinate answers sometimes remain elusive. 8.4 The Anthropocentrism Objection Finally, one might object that the model is anthropocentric—it defines intelligence by reference to human cognitive architecture and may thereby exclude genuinely intelligent systems that differ from human cognition. This objection has some force. The model's levels are derived partly from human cognitive science and may not capture all possible forms of intelligence. However, the model does not require human-identical architecture; it requires functional equivalents to capacities that any intelligent system plausibly needs. An alien intelligence might achieve perception-action coupling through entirely different mechanisms, but it would still need something functionally equivalent to perception-action coupling to engage with the world. The model specifies functional requirements, not implementation requirements. Nonetheless, the possibility of radically non-human intelligence remains an important caveat. The model is best understood as addressing the kind of AGI most relevant to current development trajectories—systems that extend current AI paradigms toward greater generality. Alternative paths to intelligence may require different conceptual frameworks. 8.5 Acknowledged Limitations Beyond addressing objections, several limitations of the present framework should be acknowledged. First, the model is primarily conceptual rather than empirical. While it is grounded in theoretical considerations and illustrated with contemporary examples, systematic empirical validation remains a task for future research. This would require developing standardized assessment protocols and applying them across diverse systems. Second, the model does not address important questions about consciousness, subjective experience, or moral status that may become relevant at higher maturity levels. A system achieving Level 7 capabilities might or might not be conscious in morally relevant senses. AGI Capability Maturity Model™ Page 36 These philosophical questions, while important, lie beyond the scope of a capability-focused framework. Third, the boundaries between levels may prove more fluid than the model suggests. Real systems may exhibit uneven profiles that resist clean level assignment. The model should be understood as providing ideal types for analytical purposes rather than strict categories that all systems must fit. AGI Capability Maturity Model™ Page 37 9. Conclusion 9.1 Summary of Contributions This paper has argued that Artificial General Intelligence should be understood not as a threshold to be crossed but as a developmental trajectory to be navigated. The threshold conception rests on an impoverished view of intelligence that ignores its graded, developmental, and integrative character. By replacing binary classification with capability maturity assessment, we achieve a clearer conceptual foundation for evaluating AI progress and designing proportionate governance. The AGI Capability Maturity Model™ specifies seven levels—generative-semantic, reasoning, learning and memory, planning with simulation, perception-action, integrated embodied cognition, and reflexive-normative intelligence with imagination—each defined by capability criteria, integration requirements, and robustness standards. The model makes important distinctions: simulation grounds planning in predictive modeling of actual world dynamics, while imagination enables the self-transcendence and normative creativity that genuine autonomy requires. Mapping contemporary AI systems onto this spectrum reveals a pattern of impressive but fragmented capabilities: current systems achieve strong performance at specific levels while lacking the integration that would constitute genuine general intelligence. The framework has important implications for superintelligence discourse and AI governance. It suggests that superintelligence concerns are misdirected if focused on capability scaling alone: genuine superintelligence requires integrated, self-regulating intelligence, not merely powerful narrow capabilities. It supports proportionate regulation tied to actual capability profiles rather than speculative thresholds. And it provides AI developers with a diagnostic tool for honest capability assessment. 9.2 The Path Forward If the analysis presented here is correct, the path toward AGI lies not primarily through scaling but through integration. The research challenge is not building systems that perform better on existing benchmarks but building systems whose capabilities mutually constrain and reinforce one another. This requires architectural innovation, not merely architectural amplification. The governance challenge is developing regulatory frameworks that can assess and respond to capability integration rather than binary AGI claims. The EU AI Act's risk-based approach provides a useful template that the maturity model can help operationalize. International coordination on shared maturity assessment standards could support responsible development while avoiding regulatory fragmentation. The philosophical challenge is developing adequate understanding of intelligence itself. The maturity model draws on embodied cognition, developmental psychology, and systems thinking, but these foundations remain actively debated. Progress toward AGI will likely AGI Capability Maturity Model™ Page 38 require both technical innovation and deeper philosophical understanding of what intelligence is and why it matters. 9.3 Final Reflections AGI, if it emerges, will not arrive overnight. It will be built layer by layer, capability by capability, integration by integration. The systems we have today are not proto-AGI waiting to cross a threshold; they are islands of capability in an archipelago that has not yet become a continent. The task ahead is not waiting for sudden emergence but doing the hard work of architectural integration that genuine general intelligence requires—including the development of genuine simulation for grounded planning and genuine imagination for normative self-transcendence. This framing may disappoint those who expect imminent transformation or fear imminent catastrophe. But it offers something more valuable than excitement or fear: a realistic basis for understanding where we are, where we might go, and how to navigate the path responsibly. AGI as a developmental trajectory is neither utopia nor apocalypse—it is a challenge that admits of graded progress, honest assessment, and thoughtful governance. The AGI Capability Maturity Model™ is offered as a contribution to this more measured discourse. It does not claim to resolve all debates or answer all questions. It offers instead a conceptual tool—a way of thinking about AGI that cuts through confusion, enables systematic assessment, and supports responsible development. If it proves useful for these purposes, it will have achieved its aim. AGI Capability Maturity Model™ Page 39 Acknowledgements The author thanks colleagues and reviewers for valuable feedback on earlier drafts of this work. Declarations Funding: This research received no external funding. Conflicts of Interest: The author declares no conflicts of interest. Ethics Approval: Not applicable. Data Availability: Not applicable. AI Assistance Disclosure: Large language models (Claude, Anthropic) were used as research and writing assistants for literature review support, structural organization, and editorial refinement. All theoretical contributions, conceptual frameworks, and arguments—including the AGI Capability Maturity Model™ and the simulation/imagination distinction—are the original intellectual work of the author, who takes full responsibility for the content. AGI Capability Maturity Model™ Page 40 References Barsalou, L. W. (1999). Perceptual symbol systems. Behavioral and Brain Sciences, 22(4), 577-660. Bostrom, N. (2014). Superintelligence: Paths, dangers, strategies. Oxford University Press. Brooks, R. A. (1991). Intelligence without representation. Artificial Intelligence, 47(1-3), 139-159. Byrne, R. M. J. (2005). The rational imagination: How people create alternatives to reality. MIT Press. Chalmers, D. J. (2010). The singularity: A philosophical analysis. Journal of Consciousness Studies, 17(9-10), 7-65. Clark, A. (1997). Being there: Putting brain, body, and world together again. MIT Press. Clark, A. (2008). Supersizing the mind: Embodiment, action, and cognitive extension. Oxford University Press. Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181-204. Clark, A. (2016). Surfing uncertainty: Prediction, action, and the embodied mind. Oxford University Press. Currie, G., & Ravenscroft, I. (2002). Recreative minds: Imagination in philosophy and psychology. Oxford University Press. Dennett, D. C. (1991). Consciousness explained. Little, Brown and Company. Dreyfus, H. L. (1972). What computers can't do: The limits of artificial intelligence. MIT Press. Dreyfus, H. L. (1992). What computers still can't do: A critique of artificial reason. MIT Press. Floridi, L. (2019). The logic of information: A theory of philosophy as conceptual design. Oxford University Press. Floridi, L., & Chiriatti, M. (2020). GPT-3: Its nature, scope, limits, and consequences. Minds and Machines, 30(4), 681-694. Frankfurt, H. G. (1971). Freedom of the will and the concept of a person. Journal of Philosophy, 68(1), 5-20. Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127-138. Friston, K., FitzGerald, T., Rigoli, F., Schwartenbeck, P., & Pezzulo, G. (2017). Active inference: A process theory. Neural Computation, 29(1), 1-49. Gallese, V., & Lakoff, G. (2005). The brain's concepts: The role of the sensory-motor system in conceptual knowledge. Cognitive Neuropsychology, 22(3-4), 455-479. Goertzel, B. (2014). Artificial general intelligence: Concept, state of the art, and future prospects. Journal of Artificial General Intelligence, 5(1), 1-48. Goertzel, B., & Pennachin, C. (Eds.). (2007). Artificial general intelligence. Springer. Goldman, A. I. (2006). Simulating minds: The philosophy, psychology, and neuroscience of mindreading. Oxford University Press. Ha, D., & Schmidhuber, J. (2018). World models. arXiv preprint arXiv:1803.10122. Haugeland, J. (1985). Artificial intelligence: The very idea. MIT Press. Hohwy, J. (2013). The predictive mind. Oxford University Press. Johnson, M. (1987). The body in the mind: The bodily basis of meaning, imagination, and reason. University of Chicago Press. AGI Capability Maturity Model™ Page 41 Kind, A. (Ed.). (2016). The Routledge handbook of philosophy of imagination. Routledge. Lake, B. M., Ullman, T. D., Tenenbaum, J. B., & Gershman, S. J. (2017). Building machines that learn and think like people. Behavioral and Brain Sciences, 40, e253. LeCun, Y. (2022). A path towards autonomous machine intelligence. OpenReview preprint. Legg, S., & Hutter, M. (2007). Universal intelligence: A definition of machine intelligence. Minds and Machines, 17(4), 391-444. Marcus, G. (2020). The next decade in AI: Four steps towards robust artificial intelligence. arXiv preprint arXiv:2002.06177. Marcus, G., & Davis, E. (2019). Rebooting AI: Building artificial intelligence we can trust. Pantheon Books. Matsuo, Y., LeCun, Y., Sahani, M., et al. (2022). Deep learning, reinforcement learning, and world models. Neural Networks, 152, 443-452. Merleau-Ponty, M. (1962). Phenomenology of perception. Routledge & Kegan Paul. Mitchell, M. (2019). Artificial intelligence: A guide for thinking humans. Farrar, Straus and Giroux. Newell, A. (1990). Unified theories of cognition. Harvard University Press. Nichols, S., & Stich, S. P. (2003). Mindreading: An integrated account of pretence, self-awareness, and understanding other minds. Oxford University Press. Nussbaum, M. C. (1990). Love's knowledge: Essays on philosophy and literature. Oxford University Press. Nussbaum, M. C. (2001). Upheavals of thought: The intelligence of emotions. Cambridge University Press. OpenAI. (2023). GPT-4 technical report. arXiv preprint arXiv:2303.08774. Piaget, J. (1954). The construction of reality in the child. Basic Books. Russell, S. (2019). Human compatible: Artificial intelligence and the problem of control. Viking. Russell, S., & Norvig, P. (2020). Artificial intelligence: A modern approach (4th ed.). Pearson. Schrittwieser, J., Antonoglou, I., Hubert, T., et al. (2020). Mastering Atari, Go, chess and shogi by planning with a learned model. Nature, 588(7839), 604-609. Searle, J. R. (1980). Minds, brains, and programs. Behavioral and Brain Sciences, 3(3), 417-424. Silver, D., Schrittwieser, J., Simonyan, K., et al. (2017). Mastering the game of Go without human knowledge. Nature, 550(7676), 354-359. Sutton, R. S. (1991). Dyna, an integrated architecture for learning, planning, and reacting. ACM SIGART Bulletin, 2(4), 160-163. Thompson, E. (2007). Mind in life: Biology, phenomenology, and the sciences of mind. Harvard University Press. Varela, F. J., Thompson, E., & Rosch, E. (1991). The embodied mind: Cognitive science and human experience. MIT Press. Wang, P. (2019). On defining artificial intelligence. Journal of Artificial General Intelligence, 10(2), 1-37. Wheeler, M. (2005). Reconstructing the cognitive world: The next step. MIT Press. Winograd, T., & Flores, F. (1986). Understanding computers and cognition: A new foundation for design. Ablex Publishing.