Appendix A — 🧠 Brain: «Observe» Tao-Intelligence Practice

Brain/Mind, born of eons of evolution, is both power and principle. Humans must grasp the mechanism of its transformation, and know the wonder of its texture.

Machine, the accelerated conversion of energy, is both heart and craft. Can the device mirror the shape of that evolution, and echo the spirit of creation itself?

A question to Heaven: weighing merits against faults, gains against losses — how much can human and machine co-create together?1

What do we learn? What do we train? Understanding the noun-objects of the mind’s learning and training helps us understand the mind itself.

This appendix will briefly explain:

  1. Using this book as an example, how “noun «options»” form a concentric-circle system of knowledge;
  2. How the ⟨Tri-Aspect Mind⟩ integrates major theories into a foundational taxonomy for distinguishing human and machine minds;
  3. Through a comparison of human and machine learning and their world models, how noun «options» settle into a trainable, alignable texture of mind.
Tip A.1: 🧠 Brain Tip: «The Appendix Triad» “Action ~ Brain ~ Cognitive Capacity”
  • “💪Action”: Focusing on Verbs, the “Action Concerto”
  • ➜ “🧠Brain”: Focusing on Nouns, “Tao-Intelligence Practice”
    • 〜 ❝Training❞ conscious mind-action from the inside out
  • “🪜Cognitive Capacity”: Focusing on Nouns ➕ Verbs, “Constructive Scaffolding”

The core of training is 🌊learning noun «options» — building a systematic body of knowledge and putting one’s own worldview, outlook on life, and values into practice.

Human-machine noun «options» can refine the collaborative relationship and achieve systemic dynamic orchestration; this calls for «The Appendix Triad» — “Action ~ Brain ~ Cognitive Capacity” — treating AI as a partner in shared action.

  1. Taxonomy Layer:
    • Concentric Circles of Nouns: 🔴 Core Circle (self-training) / 🟡 Middle Circle (human-machine collaboration) / 🟢 Outer Circle (social and civilizational context); see 🧩 Nouns and the World 🌐 for details.
    • ⟨Tri-Aspect Mind⟩: 🐸 the Reactive Mind / 🐘 the Emotive-Relational Mind / 🧘 the Reflective-Symbolic Mind, integrating two major theories — the ⚡Dual-Process Hypothesis, and the 🤝🙊Social Brain Hypothesis; see 🧠 ❝Tri-Aspect Mind❞: Integrated Learning With a Taxonomy for details.
  2. Comparison Layer:
  3. Practice Layer:
    • Tao-Intelligence Practice: Tao-intelligence (theoretical mastery) ➕ practice (systemic application), mapped onto the AQAL Four Quadrants to contemplate self and world.

🎓 This appendix corresponds to the ontological and epistemological core role within «The Appendix Triad» (for methodology, see “💪Action” and “🪜Cognitive Capacity”). For the complete tripartite comparison, see the academic-framework explanation in the Preface 〜 «The Appendix Triad».

🧩 Observing the World for Actions 💪

For ease of understanding, this book defines the brain or the mind as a learning journey of “placing a world model, preparing for future action.”

This operational definition therefore reduces every verb to some form of learning or training, while the corresponding nouns concern the existence of everything in the world — including the various knowledge points that human society has created and accumulated, as well as the related AI knowledge points. Altogether they become a pool of noun «option» as a body of human knowledge.

To help readers learn systematically, this appendix set organizes these noun «options» into the content, process, and outcome of an individual’s “learning” about the world.

Whether human or machine, it is the systematic integration of these nouns that constitutes a world model.

Whether it is:

  • human experts building Knowledge Graphs that pick out the nouns for entities and their relationships, or

  • machines that, following the “mirror neuron” structure of Hebb’s Rule, use Neural Networks to grasp the texture of the outside world — including how Large Language Models come to understand a vast number of human nouns —

the object of the mind’s training is the noun «option»: the choice of which nouns to keep determines the world model, context included.

🧩 Nouns and the World 🌐

One could say that all the various «options» of nouns — plant or animal, male or female — are forms of “noun-based learning” about the self and the world.

Nouns are not by themselves special, but rather knowledge or computer systems that specify a world through concepts, entities, relations, and reference frames: ontologies explicitly specify shared conceptualizations through entities and their relations (Vickery 1997; Huhns and Singh 1997; Kless et al. 2012). Knowledge-organization tools exist precisely to support this kind of organization, management, and retrieval (Jurisica et al. 2004; Zeng 2008; Yuan et al. 2025), and even human conceptual knowledge appears to be partly structured as cognitive maps based both on self-centered and world-centered information (Bottini and Doeller 2020). In short, an ontology systematically constructs a world (and often surrounding the self as the actor) using mental geometries such as cognitive maps and image spaces to describe relevant nouns in that worldview.

Tip💡 Connecting Ontologies to World Models

An ontology is a formal, explicit specification of a domain’s entities and relationships. When an agent (human or AI) uses an ontology to ground symbols, navigate space, and anticipate outcomes, that ontology functions as a functional world model—transforming static nouns into a substrate for action.

Thought leaders (intellectuals, scientists, and religious or cultural elites), as “producers of ideas”, build, craft and share comprehensive worldviews (Weltanschauungen), with the aim to shape material and political interests that push human actions (Schroeder 2007, 1992). This represents the effort to build a world modle using a system of nouns, or an ontology.

For easier understanding, the concentric-circle scheme below should therefore be read as a pedagogical scaffold built on this organizing impulse surrounding the use and organization of nouns, and readers are encouraged to construct their own.

The use and grounding of noun symbols is precisely the process of learning and training. The placement of the various «options» of nouns is therefore not entirely static, but rather the mind’s ongoing contemplation of itself and the world.

Learning about this world — whether through the human ‘texture of mind’ or the machine’s ‘neural texture’ — therefore also carries the action-oriented dimension emphasized in the previous appendix.

One could say that constructing a world model is itself in service of preparing Mindful Actions — how the self is to act in relation to everything in the world. It is the grain, structure, and context of this ‘texture’ that makes up the mind’s learning journey of “placing a world model, preparing for future action.”

Ontology learning refers to the process of extracting machine-readable conceptual frameworks directly from data or corpora. It transforms traditional, labor-intensive ontology engineering into an automated machine learning journey—one that continuously “places a world model to prepare for future action” by dynamically refining the system’s semantic texture.

📑 The World as Seen by This Book 🌐

Dear reader, you might be a parent raising children, a factory owner, or a retiree caring for aging parents or a spouse. This book hopes to help you set off on the AI learning journey of “placing a world model, preparing for future action.”

Which nouns are the objects that the mind trains on in order to understand the world?

Seen this way, this book’s table of contents is itself an entire system of “noun-based learning” for understanding AI and learning — both machine learning and human learning. It organizes 66 entries across ten parts.

Readers can copy the list and add or remove entries as they see fit: parents might start from play, drawing on 🏆“Game-Theoretic” AI to highlight ㉄ AI Problematics such as 🗫🎲 Language Games; business owners might start from the “dark factory,” drawing on 🦾“Embodied” AI to overcome the 🖼️⏱️Frame Problem and the 🎯🛡️Alignment & Control Problem in smart manufacturing, building out plant equipment and processes through 🏛️🌌🗺️Ontology and developing 🎏🧠Neuro-Symbolic Convergence solutions; and retirees might start from “self-tracking,” drawing on 🦾“Embodied” AI to master the various facets of 🌉 AI Engineering in smart healthcare and smart homes — 🌉🤖🚨 Agent Reliability & Evaluation and 🎁🌱🚀 AI Product Management — to safeguard their own interests.

This book’s table of contents can be viewed as a collection of knowledge points for “AI mind training”; readers are free to add or revise it as needed, filling in whatever parts matter most to them at any time.

🧩 The System of Noun «Options» ~ «Choices» 🌊

Drawing together the knowledge sources of human learning and machine learning, this appendix set, “🧠Brain”, divides these nouns into three levels, forming a “concentric-circle noun system”:

  • 🔴 Core Circle: the AI knowledge points summarized in this book, cross-referenced in 🔖Appendix 🌌 Mindmap (e.g., neural networks, reinforcement learning, knowledge graphs).
  • 🟠 Middle Circle: tightly encircling the core is the knowledge structure and context that readers must master when building intelligent systems or agents, including information about the “self” and the “environment.”
  • 🟢 Outer Circle: wrapping around the first two layers is the “world presuppositions” and “emergent knowledge” that an agent must understand, capable of carrying a broader world model (even encompassing a user’s “three views”: worldview, outlook on life, and values).

The above should cover every possible “knowledge” noun «option».

The expansion and accumulation of noun «options» resembles the construction of a worldview. Just as different human languages classify and describe snow, fire, and water differently, this carries a genuine Symbol Grounding Problem — it can never be separated from the relationship between self and world.

As an appendix to this book, readers can bring themselves (or their students or children) to use systems thinking to contemplate the noun «options» that are useful for self, environment, and world.

In short, this appendix set, “🧠Brain”, gathers together the following appendices and tables: beyond laying out the various kinds of noun content, its final section sums up the necessity of ❝training❞ the mind level by level, turning each noun «option» into a necessary part of a psychological and world model.

🧘 “Texture of Mind”: The Relationship Between Self and World 🌐

The choices among noun «options» and one’s worldview can be understood through comparing human learning and machine learning in terms of the relationship between self and world.

This comparison helps us analyze the benefits and costs of “mind training” and understand the mechanics of what this book calls the “texture of mind” — echoing the Chinese 肌理 (the grain or weave of tissue) — the accumulated, structural pattern that repeated learning leaves in a mind, biological or artificial, as distinct from any single memory or behavior.

Important A.1: “Texture of Mind” (肌理, jīlǐ)

Definition. 肌理 literally denotes the grain or weave of tissue. In this book it names the accumulated, structural pattern that repeated learning leaves in a mind — biological or artificial — distinct from any single memory or behavior. The definition captures both human neural–muscle networks and machine neural–motive connections.

Why it matters. Whereas a single memory is an isolated trace and a single action an isolated behavior, 肌理 denotes the background structure that shapes perception, inference, and habit: the way a mind tends to fill in gaps, prioritize cues, or respond under uncertainty.

Examples. - Biological: cultural practice and bodily training produce a characteristic style of attention and skill.
- Artificial: repeated training regimes leave an LLM or robot with systematic tendencies — preferred completions, failure modes, and affordance biases.

This section does more than draw a basic comparison — it also uses the perspectives of behaviorism, situated-ism, and connectionism to explain how noun «options» shape the relationship between self and world, so readers can tell them apart and innovate.

💪 Human Learning 🆚 Machine Learning 🦾

Although human and machine “learning” both center on experience—adjustment—reinforcement, their nature, scale, and cost are quite different.

Comparative work draws a parallel between neural plasticity and AI neutral network optimization (Sadegh-Zadeh et al. 2024; Spicer and Sanborn 2019), while cautioning that the analogy is incomplete: biological learning depends on internal evaluative systems and multiple reward/punishment signals that current machine systems do not replicate (Piccinini 2022). Cognitive science aims to explain human behavior, whereas machine learning typically optimizes performance — a productive but incomplete comparison (Doroudi 2023). Recent critiques of Reinforcement Learning from Human Feedback (RLHF) — discussed in detail below — show that reward models can misgeneralize, oversimplify ethical trade-offs, and embed cultural biases. These limits highlight that machine reinforcement is not neutral but a socio-technical design choice.

This comparison helps us analyze the benefits and costs of “mind training” and understand the mechanics of the “texture of mind.”

  • 🧍‍♂️ Human learning
    • 🧬 Grounded in neuronal and synaptic plasticity, gradually reinforced through bodily experience, emotional interaction, and cultural transmission. Example: a toddler learning to walk, falling and trying countless times until the neural pathways for walking are consolidated.
    • ⏳ Relatively slow, but capable of “mental fill-in” when information is lacking, and of “completing the picture” through ethics and values. Example: witnessing a crime, even with fuzzy details, a person can combine common sense, motive, and emotional judgment to fill in plausibility and moral standing.
    • 🌳 Constrained by cognitive load and group size (e.g., Dunbar’s number), but equipped with situational awareness and cross-generational memory transmission. Example: lessons of war or cultural traditions can be passed down through language and ritual.
  • 🤖 Machine learning
    • ⚡ Grounded in digital circuits and matrix operations, accelerated through massive data and energy. Example: AlphaGo played millions of games against itself within days, a training volume that dwarfs thousands of years of human Go experience.
    • 💡 Extremely fast, capable of capturing hidden patterns across vast datasets and performing “auto-completion.” Example: after being given just a few words, a Large Language Model (LLM) can generate complete, grammatically correct sentences or code based on trillions of tokens of text.
    • 🔋 Lacks intrinsic motivation and emotion, requiring externally designed reward and objective functions. Reinforcement Learning from Human Feedback (RLHF) attempts to align models with human preferences, but faces challenges of sparse feedback, ethical trade-offs, and cultural pluralism. Training at scale also comes with high energy costs (Samsi et al. 2023; Taufer et al. 2026; Patterson et al. 2021).
  • 🔄 Comparison and complementarity
    • ⚖️ Human learning emphasizes “deep practice under limited energy,” while machine learning is “accelerated training at high energy cost.”
    • 🔍 Humans excel at filling in meaning under uncertainty, while machines excel at filling in patterns across massive data.
    • 🤝 Both converge on “fill-in”: humans move from mental fill-in toward ethical fill-in (e.g., social adjustment and organizational alignment), while machines move from auto-completion toward value alignment — a process that must be understood as socio-technical design rather than purely technical optimization.

By examining human and machine learning through distinct theoretical lenses—behaviorism, situated‑ism, and connectionism—we can clarify different approaches to building world models and how those models structure relations and interactions between self and world.

–🏮💪 The Behaviorist View

From the Behaviorist view, learning is an overt process of stimulus–response feedback, focused on “observable behavior” and “reward-and-punishment feedback.” Here, noun «options» are defined as the overt behaviors of “stimulus–response–reward/punishment,” forming the core vocabulary of behaviorism.

  • 👶 Human learning is understood as a cycle of 🙶stimulus⇥response🔄reward/punishment🙷. A child learns to walk because of the repeated feedback of falling (punishment) and standing steady (reward).
  • 🎮 Machine learning has its modern variation of Reinforcement Learning (RL) to train large language models (LLMs) for AI alignment with human preferences. By treating the AI model as an animal or human learner that changes its habits based on rewards and penalties: for instance, a positive reward (positive reinforcement) when overcoming an obstacle in navigation. In RLHF, human feedback is used to train a reward model, which is then used to optimize the AI’s behavior through reinforcement learning.

Learning the World Through Reward, not Understanding

Shaping behavior through external feedback is a core mechanism of both human habit formation and machine alignment.

Reward-shaped learning is central to modern LLM training: human feedback is used to fit a reward model, which is then optimized against through reinforcement learning to steer model behavior toward human preferences (Chaudhari et al. 2025). This is often not a simple scalar signal but a preference-learning problem, where people rank or compare outputs and the system infers a reward from those pairwise or K-wise comparisons (Zhu et al. 2023). Humanity’s own “Hook Model” and habit formation through variable, intermittent feedback (e.g., social-media notifications) run on a structurally similar loop of stimulus, response, and reward.

More directly empirical and critical work on reward modeling shows that this loop has real limits. Reward models can misgeneralize, oversimplify ethical trade-offs, and embed the values and cultural assumptions of whoever supplies the human feedback (RLHF) or AI feedback (RLAIF) (Chaudhari et al. 2025; Dahlgren Lindström et al. 2025; González Barman et al. 2025) — an outcome that follows directly from RLHF’s dependence on a specific, situated group of human raters, whose judgments cannot be assumed to represent a neutral or universal standard of “helpful, harmless, honest” (Dahlgren Lindström et al. 2025). On top of this values problem sits a technical one: reward models can also become inaccurate under distribution shift, with scores rising even as behavior drifts away from actual human preferences — a persistent, measurable overoptimization problem (Gao et al. 2023; Moskovitz et al. 2024).

A live debate worth sitting with: David Silver and Richard Sutton (DeepMind) have advanced the “Reward is Enough” hypothesis — the claim that sufficiently general reward-maximizing agents in sufficiently rich environments can give rise to the full range of intelligent abilities (perception, language, social cognition, planning) without any of these needing to be separately engineered. Critics counter that reward alone cannot specify the nuanced, often unstated values and ethical trade-offs human life requires, and that RLHF’s own documented failure modes — reward hacking, sycophancy, distribution shift, and whose values get encoded in the first place (González Barman et al. 2025) — are evidence that reward signals systematically underdetermine what we actually want. Where do you land: is reward truly the only signal intelligence needs, or is behaviorism a powerful but partial account of how minds are built?

In short, humans learn through direct, embodied stimulus–response feedback, and machines increasingly learn through a reward model distilled from human feedback and then optimized at scale. Both routes show that behavior can be reliably shaped by external feedback loops, but they differ in kind: human reward and punishment are lived, immediate, and tied to bodily consequence, whereas machine reward is a proxy — a model of a model of human preference, built from the judgments of a particular, non-neutral group of raters — which is precisely why reward-function design, oversight, and monitoring for reward hacking remain necessary safeguards in deployed systems (Gao et al. 2023; Chaudhari et al. 2025).

Behavior is not evidence of understanding, but rather the shaped residue of feedback.

👉 Behaviorism emphasizes that “overt behavior is observable and measurable,” so the core of AI behaviorism is:

  • 🛡️ Reward function design: defining the ethical and value boundaries of learning, to avoid reward hacking.
  • ⚙️ Policy function updates: gradually shaping behavior patterns through feedback, reinforcing adaptability.
  • 🚀 Applications: game AI, robot control, autonomous driving, recommendation systems — all rely on the closed loop of stimulus–response–reward/punishment.

Ultimately, these noun «options» let us describe AI’s behavior-shaping loop precisely using terms like “reward function” and “policy function.”

–🏮🛣 The Situated-ist View

From the situated or Situated-ist perspective, learning is a process of acting within a situation, emphasizing “the present situation” and “situational awareness”, perception–action loops, and the shaping role of social context. Here, noun «options» focus on words that describe “situation,” “environment,” and “context,” highlighting how learning is embedded in the world. Design grounded in the situated view therefore treats environmental design, sensor–actuator loops, and social context as primary variables rather than optional add‑ons.

  • 👶 Human learning: For humans, language and skills are not acquired solely by reward and punishment but through immersion in family, community, and everyday interaction. Mastery of language and culture emerges from everyday conversation, gestures, emotional exchange, and immersion in cultural norms (Pfeifer and Bongard 2006; Clark 1997). Ethnographic and developmental work supports that much of early learning is scaffolded by social routines and embodied interaction rather than isolated reinforcement signals.

  • 🤖 Machine learning: For AI systems, embodiment and closed‑loop interaction are central to robustness and semantic grounding. Embodied AI — including agents with bodies or high‑fidelity simulators — focuses on the training in perception→action→feedback loops in real or simulated environments (e.g., robotic grasping, navigation).

Learning the World Through Interactions, not Symbols

Synergy of morphology, control, and perception is a core source of intelligence.

Classic and recent surveys and demonstrations support morphology and body‑shape as one factor that scaffolds perception–action loops (Brooks 1991; Pfeifer and Bongard 2006; Liu et al. 2025). Humans acquire situated competence through socially scaffolded, multimodal interaction that embeds norms, values, and historical practices (for example, family routines and conversational scaffolding) (Piccinini 2022; Pfeifer and Bongard 2006). In contrast, machines acquire situated competence through sensorimotor coupling, closed‑loop training, and environment‑rich experience (real or simulated), which ground action‑sensitive representations and affordances (Brooks 1991; Liu et al. 2025).

Piccinini and Friston argue, on largely conceptual grounds, that grounded content and predictive world-models require this kind of situated, action-coupled interaction rather than detached symbol manipulation (Piccinini 2022; Friston et al. 2021). More directly empirical work on active perception adds that perception itself depends on action‑dependent sampling, so closed‑loop, embodied training measurably reduces brittleness under novel viewpoints and sensory noise (Leisman et al. 2026).

A live debate worth sitting with: Yann LeCun — a 2018 Turing Award laureate and one of the field’s “Godfathers of AI” — has argued publicly that today’s text-only large language models are approaching a dead end precisely because they never learn a grounded world model through embodied, sensory interaction; his own research program (world models, JEPA) is a direct bet on the situated-ist view. Other leading researchers disagree, arguing that scale and better symbolic scaffolding (chain-of-thought, tool use, retrieval) can substitute for embodiment. Where do you land: is embodiment a necessary ingredient of understanding, or an implementation detail that sufficiently rich symbolic training can route around?

In short, humans acquire situated competence through everyday life, and machines acquire it through closed‑loop training in real or simulated environments. Both routes show that meaning and reliable behavior emerge from ongoing learning embedded within a world context, but they differ in kind: human coupling is historically and normatively thick, whereas machine coupling is engineered and constrained by designers’ objectives and training regimes — a distinction that helps explain why embodied evaluation and situational testing are necessary to reduce brittleness and improve alignment in deployed systems (Piccinini 2022; Leisman et al. 2026).

Intelligence is not the starting point of abstract reasoning, but rather the product of embodied interaction.

👉 Situated‑ism emphasizes that “learning is acting within a situation,” so the core of AI situated‑ism is:

  • 🌍 Environmental embedding: Intelligence must manifest in a concrete environment; learning is not abstract computation, but embodied interaction.
  • 🤝 Social context: Intelligence is not an isolated agent, but one adapting within group collaboration and cultural context.
  • 🦾 Applications: Robot learning, human–machine collaboration, smart cities, and social simulation—all emphasizing a real‑time closed loop of 🙶perception↹action🔄feedback🙷.

These foundational concepts remind us that intelligence is not the starting point of abstract reasoning, but the product of embodied interaction, which can guide us in selecting the noun «options» for world modelling that works.

–🏮🧬 The Connectionist View

From the Connectionist view, learning is a process in which “changes in the strength of connections across a network of neurons” form knowledge and ability, focused on “plasticity” and “long-term potentiation.” Here, noun «options» are defined as structures like “neuron,” “synapse,” and “weight,” forming the core vocabulary of connectionism.

  • 👶 Human learning: neurons in the brain continually adjust connection strength through synaptic plasticity. A child learns to play piano because long-term practice gradually strengthens the relevant neural circuits, forming “muscle memory.”
  • 🤖 Machine learning: in artificial neural networks, learning is realized through weight updates. Every forward pass and backward pass adjusts “connection strength” to minimize an error function — this is the core of Deep Learning.

Learning the World Through Weights, not Symbols

Intelligence, on this view, is not stored anywhere in particular — it is the pattern of connection strengths itself.

Plain weight adjustment is not the whole story, however. Biologically inspired schemes that add synaptic-integration principles can improve network performance and mitigate catastrophic forgetting (Dellaferrera et al. 2022), and fast learning can emerge from a synergy of slower synaptic plasticity and faster network dynamics rather than from weight change alone (Subramoney et al. 2024). A network’s initial connectivity structure also matters: higher-rank initializations bias learning toward a “lazier” regime with smaller representational change, while lower-rank initializations bias it toward a “richer” regime — a trade-off with direct consequences for memory retention and forgetting (Liu et al. 2024).

Representations built this way become invariant and distributed, with computation and memory carried by the coordinated dynamics of a whole population of neurons rather than stored in any single node (Halvagal and Zenke 2023; Vyas et al. 2020). Humans’ “synaptic pruning” and “long-term potentiation” correspond to AI’s “regularization” and “gradient descent” — both, through the dynamic adjustment of connection strength, gradually form stable pattern representations in complex environments.

A live debate worth sitting with: Geoffrey Hinton — a 2018 Turing Award laureate and, alongside LeCun and Bengio, one of the field’s “Godfathers of AI” — has long argued that backpropagation itself is unlikely to be how the brain actually learns, and has proposed alternatives (such as the Forward-Forward algorithm) closer to biological plasticity. Other connectionists counter that backpropagation’s biological implausibility does not matter: what matters is that gradient-based weight adjustment, whatever its neural correlate, reliably produces the distributed representations connectionism predicts. Where do you land: does intelligence require a learning rule that mirrors the brain’s own mechanics, or is any sufficiently powerful weight-adjustment process enough?

In short, humans adjust connection strength through synaptic plasticity shaped by lived, embodied experience, and machines adjust connection strength through gradient-based weight updates shaped by an external error signal. Both routes show that stable, useful representations can emerge purely from the repeated adjustment of connections between simple units, but they differ in kind: human plasticity is metabolically costly, biologically constrained, and never fully erases what came before, whereas machine weight adjustment is comparatively cheap, unconstrained by biology, and can overwrite prior representations wholesale — which is precisely why catastrophic forgetting remains a live engineering problem that biologically inspired plasticity schemes are trying to address (Dellaferrera et al. 2022; Subramoney et al. 2024).

Intelligence is not a symbol stored in a location, but rather the shape of a network’s connections.

👉 Connectionism emphasizes that “intelligence arises from the connections and weight adjustments of distributed nodes,” so the core of AI connectionism is:

  • 🧩 Distributed representation: knowledge is not a single symbol, but an activation pattern distributed across many nodes.
  • 🔄 Weight updates: learning is the repeated adjustment of connection strength, gradually converging on an effective representation.
  • 🚀 Applications: speech recognition, image classification, natural language processing, generative models (such as LLMs) — all rely on the connection-adjustment of large-scale neural networks.

These noun «options» therefore let us describe the shared texture of human and machine learning through “distributed representation” and “weight updates.”

🏮🌟 Implications and Impact

Through examining human and machine learning from three perspectives—behaviorist, situated, and connectionist—each showing different design levers for world‑modeling, we now appreciate how noun «options» are used to map onto the perspectives of several major AI schools, serving as a bridging vocabulary for comparing human and machine learning.

Comparing human and machine “learning,” the major AI schools offer the following implications:

  1. 🏮💪 Behaviorist implication: AI’s learning can be shaped through reward-and-punishment mechanisms, but the reward function must be designed carefully to avoid “reward hacking” or “bias amplification.”
  2. 🏮🛣 Situated-ist implication: AI’s learning must take embodiment and context into account, or it will fail to be robust and aligned in the real world.
  3. 🏮🧬 Connectionist implication: AI’s learning depends on weight adjustment and distributed representation, letting it capture patterns across massive data, but also leaving it vulnerable to data bias and energy costs.
  4. 🤝 A complementary perspective: behaviorism supplies the algorithmic framework for “how to learn,” situated-ism supplies the environmental context for “where to learn,” and connectionism supplies the neural-network foundation for “with what structure to learn.” Only by combining all three can AI both fill in patterns from abstract data and fill in meaning in the real world.

Take Large Language Models (LLMs) as an example: they can be seen as a “mutual learning” game between human and machine, one that also reveals an “asymmetric game”:

  • 🏮💪 Behaviorist dimension: LLMs are shaped through RLHF, learning to cater to human preferences and pragmatic habits.
  • 🏮🛣 Situated-ist dimension: although LLMs lack true embodiment, within a “conversational situation” they gradually learn to adjust their responses, simulating the context of social interaction.
  • 🏮🧬 Connectionist dimension: the core of an LLM is a massive neural network that captures patterns in language data through weight updates, forming an “internal little map.”

In particular, because Neural Networks are analogous to neural plasticity (and to the Connectionist view), their capacity for “generation” and “pruning” makes such models more prone to forming a stimulus–response-like behavior pattern — a tendency that RLHF then trains directly against, forming an “asymmetric game of influence”:

  • The designer, through reward-function and data choices, sets the behavioral boundaries and value tendencies of the LLM.
  • The user, through interacting with the AI, is also gradually “trained” by its response patterns, developing new linguistic, cognitive, and behavioral habits.

This is no longer only “humans training machines” — it is humans and machines shaping each other.

Learning Each Other, not Just the World

Human-AI interaction research increasingly frames this as a coevolutionary feedback loop with asymmetric control: users generate data that trains systems, and systems in turn shape later preferences, routines, and choices, while designers set the objectives, interfaces, and recommendation structures within which users adapt (Pedreschi et al. 2025; Puerta-Beldarrain et al. 2025). Studies now show that people measurably change their behavior once they know an AI is being trained on it, and those changes can persist as habits (Treiman et al. 2024) — which raises autonomy and influence concerns that go beyond technical alignment alone, since preferences and even identity-relevant goals can be co-constructed in sustained human-AI relationships (Kirk et al. 2025).

A live debate worth sitting with: some researchers frame this coevolution optimistically, as a route to AI systems that genuinely converge on human values through sustained interaction. Others warn that an asymmetric feedback loop — where the designer controls the objective function and the user only supplies data and habituates to the output — is structurally closer to behavior modification than to mutual learning, however “mutual” it may feel from the user’s side. Where do you land: is human-AI coevolution a genuine two-way relationship, or a one-way shaping process wearing a two-way costume?

Humans and machines are jointly training each other’s “texture of mind.” These noun «options» are therefore not merely theoretical labels, but tools for examining how intelligent systems — including LLMs — and human minds and the world shape one another.

Mind is not shaped by a single mechanism, but rather by the loop in which learner and teacher continually influence and thus reshape each other. As Confucius teaches us through the concept of jiao xue xiang zhang (教學相長, from The Book of Rites / 禮記·學記), teaching and learning are mutually transformative.

💪🆚🦾 Towards “Texture of Mind” Training

Comparing human and machine “learning,” we find that both share a similar core of experience—adjustment—reinforcement, and so there is mutual influence and plasticity between them. Here, noun «options» converge into the shared core of “experience—adjustment—reinforcement,” linking the learning journeys of human and machine.

But their energy cost and scale differ enormously:

  • 🧍‍♂️ Human learning: slow and deep practice, emphasizing the accumulation of body, emotion, and culture.

  • 🤖 Machine learning: fast and massive training, emphasizing the expansion of data, compute, and pattern.

These noun «options» ultimately lead us to reflect on sameness and difference: how to strike a balance between “fill-in” and “alignment” — a question that becomes even sharper in the next section, when we examine world models and mind training.

🧘 “Tao-Intelligence Practice”: Training “Texture of Mind”

“Tao-Intelligence Practice,” a conscious act of mind, already presupposes the existence of self and world before it can act within them.

Everything so far in this appendix — the concentric-circle system of nouns, the ⟨Tri-Aspect Mind⟩, the comparison of human and machine learning, the account of world models — has been a matter of mastering theory: naming and organizing the noun «options» that make up a mind’s knowledge. “Practice” is a different act: applying that theory. And applying it well requires two distinct moves, which this appendix has so far used somewhat interchangeably under the word “level.”

  • Leveling up is a question of depth: which tier of the ⟨Tri-Aspect Mind⟩ — 🐸 Reactive, 🐘 Emotive-Relational, or 🧘 Reflective-Symbolic — is doing the work of processing a given noun «option»?
  • Grounding is a question of connection: does that noun «option», however deeply processed, actually connect back to lived, embodied, social reality — or does it remain a symbol manipulated in isolation, disconnected from self, environment, and world?

These are not the same axis, and conflating them is a real risk. A reflex can be perfectly grounded (a frog’s leap is directly coupled to a real threat) without ever leveling up to reflection. And symbolic output can level up — becoming fluent, structured, even wise-sounding — while remaining fundamentally ungrounded. It is this second failure mode that the case of Large Language Models makes vivid.

–😵‍💫 A Case Study: Large Language Models 🧞‍♀️

Taking Large Language Models as an example: an LLM’s fluency is a real achievement of leveling — its outputs draw on structures that resemble 🧘☸️ Reflective-Symbolic processing, producing long-form reasoning, planning language, and even something like metacognitive commentary on its own outputs.

But this leveled-seeming output sits on famously thin grounding. Although an LLM lacks a physical body, it interacts with humans through language, forming a kind of “linguistic embodiment” — a half-body intelligence 🧩 Observing the World for Actions 💪. Which noun «options» are included in that linguistic interaction (such as “dialogue,” “role,” “task”) and which are structurally excluded (such as “touch,” “spatial manipulation,” lived consequence) directly shapes how ungrounded its world model remains, no matter how leveled its language sounds.

This is the trap of noun-only competence: a system — or a person — can accumulate ever more sophisticated noun «options» at the Reflective-Symbolic tier while never grounding a single one of them in the concentric circles of self, relational context, and social-civilizational world laid out earlier in this appendix 🧩 Nouns and the World 🌐. Leveling without grounding produces fluent ungroundedness; the antidote is not more theory, but the deliberate work of connecting a noun «option» back through the concentric circles — testing it against self (🔴), relational context (🟠), and social-civilizational world (🟢) — and, more fundamentally, against the Symbol Grounding Problem itself: a noun is not truly known until it is tied to something the mind can act on.

-🌐 AQAL Four Quadrants: World Models for Actions 🏄🏼

A brief cross-appendix note: this same grounding work can also be tested arena by arena — subjective, objective, intersubjective, interobjective — using the AQAL Four Quadrants, the framework Ken Wilber developed to map “who acts on whom, in what context” ?sec-aqal-8-taxonomy in 💪Action. Where this appendix asks which nouns are worth knowing, that one asks how to act on them; the full quadrant framework, its attribution, and its capacity-ladder application belong there, not here.

Table A.1: The Concentric-Circle System of Noun «Options»
Circle Focus LLM Example
🔴 Core Circle Training the self (⟨Tri-Aspect Mind⟩, neural-muscular) Neural networks, Hebb’s Rule, reinforcement learning and deep learning
🟠 Middle Circle Human-machine collaboration (linguistic embodied interface) Noun «options» included or excluded, such as dialogue, role, task
🟢 Outer Circle Social and civilizational context (scaling, group limits) Social Brain Hypothesis, scaling, world models and AI hypotheses

For readers who want to go beyond “AI mind training” and grasp the mind training of both humans and human-machine systems, this book’s ⟨Tri-Aspect Mind⟩ taxonomy remains the starting point, read together with Human Learning 🆚 Machine Learning and Brain and World Models.

This is also the origin of this appendix’s name, “Tao-Intelligence Practice”: Tao-intelligence is mastering the theories above and their implications for AI system design; practice is the deliberate, ongoing work of leveling up and grounding those theories together — distinguishing and appreciating the similarities and differences among human, machine, and human-machine action, so that noun «options» become more than a vocabulary list.

-🌊 Stucturing the Understanding of the World for Actions 🏄🏼

In short, since machine learning in AI has not yet overcome the ability to autonomously determine and revise its own “objective function” within the world, the “goal” of the human as user and designer must be a conscious act of “mind training” — one that requires understanding both how deeply a noun «option» is leveled, and how firmly it is grounded, in oneself and in the machine alike.

Mind is not a single mechanism, but a lattice woven of many layers.

From the physiological reflex of reaction, to the situational memory of emotion, to the symbolic reasoning of reflection;

Humans, by way of ❝mental fill-in❞, seek shortcuts of knowledge and action; machines, by way of 🙶constructive fill-in🙷, expand the frontier of compute.

This appendix sums up the “tripartite layering” and “frame-computation” of the mind’s faculties, as the Tao-intelligence foundation for a human-machine duet.

Human knowledge, like machine intelligence, rests on a tension between the nouns we construct to categorize the world and the dynamic realities that resist such static boundaries. Self-transformation, seen this way, is not a mystical pursuit but a structural necessity: for both biological minds and artificial networks, engaging the world requires ongoing calibration — not rigid confinement, but the disciplined, iterative act of leveling up and grounding one’s own model in turn.

By moving beyond the conflict of rigid noun systems, we see that self and world are not separate entities, but an integrated, dynamic feedback system. Through this discipline, the self does not escape the world; it refines its capacity to navigate, reflect, and remain grounded within it. In the age of artificial intelligence, this capacity for deliberate self-evolution — leveled and grounded together — remains the definitive hallmark of living agency.

This appendix set, “🧠Brain”, focuses on the intelligent noun «options» — how to attend to human, machine, and even everything in the world, so as to forge a ‘neural texture’ and see through the ‘texture of mind’ in ourselves and in machines — continuing on from the verb «options» described in the previous appendix, “💪Action”, and connecting to the “Constructive Scaffolding” of the next appendix, “🪜Cognitive Capacity”.

🧠 ❝Tri-Aspect Mind❞: Integrated Learning With a Taxonomy

How can we better understand the “Texture of Mind”? This book proposes a Tri-Aspect Mind framework.

Formally called the Tri-Aspect Mind, this taxonomy is this book’s own synthesis, not a direct restatement of any single existing theory. It combines two independently well-supported axes from the literature: the fast/slow processing distinction central to the Dual-Process Hypothesis ⚡🧮雙系統理論, and the individual/relational/group distinction central to the Social Brain Hypothesis 🤝🙈社會腦假說.

This book’s ⟨Tri-Aspect Mind⟩ does not rely on Paul MacLean’s Triune Brain Hypothesis — the once-popular idea that the human brain evolved in three discrete layers (reptilian, limbic, neocortical). Contemporary neuroscience has largely rejected this model: there is no clean evolutionary layering of brain structures of the kind MacLean proposed (Cesario et al. 2020). The synthesis in this section rests on Dual-Process and Social-Brain research only.

What survives from popular triune-brain accounts is a set of vivid animal examples — the frog’s reflex, the elephant’s grief. This book keeps them strictly as behavioral illustrations, not as evidence of discrete brain layers: a frog’s startle reflex and an elephant’s mourning behavior are real, independently documented ethological phenomena, useful for making a tier memorable, not for claiming that tier corresponds to a specific evolved brain structure.

Readers wanting a more granular, explicitly multi-timescale architecture may find Taniguchi et al.’s quad-process model (System 0/1/2/3) a useful complement [Taniguchi et al. (2025)]?sec-quad-process-theory. The correspondence is not exact: System 0 and System 2 map reasonably well onto this book’s Reactive and Reflective-Symbolic tiers, but System 1 (individual fast/intuitive cognition) does not cover the relational, social ground this book’s Emotive-Relational tier is built to cover — that tier is drawn from Social Brain Hypothesis instead, which sits outside Taniguchi’s four-system model entirely. The quad-process model is a useful reference point, not this taxonomy’s source.

Dual-process research robustly supports a distinction between fast, automatic processing and slower, reflective processing, but also warns against reifying this into only two neatly localized brain systems (Evans 2008; Evans and Stanovich 2013; Frankish 2010). On its own, however, dual-process theory offers no natural home for social and emotional processing. This is where this book’s synthesis does its real work: the middle tier — the 🐘💞 Emotive-Relational Mind — is proposed here to fill that gap, drawing on social-cognition research that supports a hierarchical, multilevel structure of partially separable cognitive and affective processes rather than a single faculty such as empathy or theory of mind (Schurz et al. 2021; Happé et al. 2017). The resulting reactive / emotive-relational / reflective-symbolic triad is therefore a genuine three-way synthesis of two source theories, not a restatement of either one.

–📌 Three Main Types ⚡💞☸️

This book abbreviates it as the ⟨Tri-Aspect Mind⟩, divided into three main types:

  • 🐸⚡ The Reactive Mind: emphasizing survival instinct, muscular reflex, fast reaction, and the primacy of the body.
    • Texture of mind: Neural tissue controls muscle contraction by transmitting electrical signals from the nervous system to muscle fibers, enabling voluntary and involuntary movements.
    • Dual-Process Hypothesis: System 1 (fast, intuitive), driving rapid responses based on emotion, heuristics, and associative memory.
    • Social Brain Hypothesis: Real-time group reaction and threat vigilance.
    • AI Alignment Mapping: Basic RL, greedy search, token prediction, and real-time safety guardrails.
  • 🐘💞 The Emotive-Relational Mind: emphasizing emotional expression, empathy, emotional bonding (individual and group), and social memory transmission.
    • Texture of mind: Neurochemical and hormonal modulation—primarily governed by limbic circuitry (such as the amygdala and hypothalamus) releasing neurotransmitters (oxytocin, dopamine, serotonin) to regulate affect, social valence, dynamic physiological arousal, and emotional homeostasis.
    • Dual-Process Hypothesis: The boundary between System 1 and System 2, handling emotional responses based on heuristics and social memory.
    • Social Brain Hypothesis: Social life, group cooperation, and communication.
    • AI Alignment Mapping: RLHF preference learning and Reward Models fitting subjective human raters, social consensus, and relational preferences.
  • 🧘☸️ The Reflective-Symbolic Mind: emphasizing consciousness, self-awareness, symbols, metacognition, long-term planning, and symbolic memory systems.
    • Texture of mind: Neocortical neural architecture—driven by complex synaptogenesis, prefrontal cortex executive networks, dynamic neural plasticity, and dense reciprocal connections across functional cortical hubs enabling abstract representations, symbolic processing, and metacognitive simulation.
    • Dual-Process Hypothesis: System 2 (slow, analytical).
    • Social Brain Hypothesis: Higher-order social cognition, particularly language and formal symbolism.
    • AI Alignment Mapping: Constitutional AI, RLAIF, explicit rules, Chain-of-Thought (CoT) reasoning, metacognition, and long-context alignment constraints.

Physiological Texture, Cognitive Theories, and AI Training Mapping

The table below maps the ⟨Tri-Aspect Mind⟩ framework across its biological texture (Texture of Mind), classical cognitive hypotheses, and modern AI alignment and control engineering paradigms (RL, RLHF, and Constitutional AI):

⟨Tri-Aspect Mind⟩ 🧠 Texture of Mind ⚡🧮 Dual-Process / 🤝🙈 Social Brain Mapping 🎮 AI Training & Alignment (RL / RLHF / RLAIF) Mapping
🐸⚡ Reactive Mind Neuromuscular Conduction: Neural tissue transmits electrical signals to control muscle contraction, enabling voluntary and involuntary movement. System 1 (Fast / Intuitive)
⚖️🙈 Social Brain (Real-time threat vigilance and group reflex)		 Basic RL & Instinctive Reflexes: Next-token prediction, greedy search, and single-reward-signal real-time threat vigilance and edge safety guardrails.
🐘💞 Emotive-Relational Mind Neurochemical & Hormonal Modulation: Limbic circuits release neurotransmitters (oxytocin, dopamine, serotonin) to regulate affect, social valence, and emotional homeostasis. System 1 / System 2 Boundary
🤝🧠 Social Brain (Social life, group cooperation, and communication)		 RLHF Preference Learning & Reward Models: Rater-based K-wise ranking reflecting subjective emotion, social consensus, and relational preferences to build Reward Models.
🧘☸️ Reflective-Symbolic Mind Neocortical & Prefrontal Architecture: Synaptogenesis and cortical hub plasticity drive high-order abstract symbols, metacognition, and long-term simulation. System 2 (Slow / Analytical)
🤝🧠 Social Brain (Higher-order language and symbolic social cognition)		 Constitutional AI / RLAIF / Symbolic Inference: Rule-based governance charters, Chain-of-Thought (CoT) reasoning, metacognitive auditing, and long-context alignment constraints.

–🌊 Refining “Texture of Mind” Mapping ☸️

This mapping illustrates why behavioral RLHF alone is prone to reward overoptimization: it operates primarily at the intersection of Reactive (quick loss reduction) and Emotive-Relational (human preference ranking) tiers, leaving the Reflective-Symbolic tier (explicit symbolic auditing, metacognition, and structural ethical alignment) under-addressed unless reinforced by symbolic rules and constitutional frameworks.

The Emotive-Relational tier is most naturally anchored in social-brain work, which links mentalizing capacity to the structure and layered scale of social networks (Dunbar 2014; Zhou et al. 2004) — though this section should carry a caveat: Dunbar-style limits are influential but contested when extended too simply to human culture and organization (De Ruiter et al. 2011).

This cross-domain taxonomy of cognitive faculties distinguishes human mental activity by processing speed, social function, and symbolic capacity. The reaction time and cognitive load (認知負荷) required rise in this same order, which also implies proportionally higher resource and energy consumption. These three tiers of mind map onto different design layers of AI:

  • 🐸⚡ The 🙶Reactive🙷 Mind: dominated by real-time perception and rapid-response modules, emphasizing instant reaction, suited to edge computing and safety, and worth considering for building swarm-intelligence style real-time reaction and vigilance systems.
  • 🐘💞 The 🙶Emotive-Relational🙷 Mind: dominated by human-machine interaction, affective computing, and social simulation, emphasizing emotional understanding, suited to human-machine interaction and collaboration, and worth considering for building emotion- and relationship-tracking and response systems.
  • 🧘☸️ The 🙶Reflective-Symbolic🙷 Mind: dominated by higher-order reasoning, planning, ethics, and alignment, emphasizing symbolic reasoning, suited to planning, decision-making, and alignment, and capable of folding the norms and preparations of higher-order social cognition into analysis and decision-making through symbols and verification.

This lens of resource and energy consumption (time and cognition) echoes the verb «options» from the previous appendix, “💪Action”.

This energy-cost lens matters a great deal, particularly as contemporary AI development has begun to weigh units of compute cost as well. One might say that behind the choice of noun «options», the very taxonomy of the ⟨Tri-Aspect Mind⟩ already carries a distinction in energy and time compute cost.

–🤯 A Closer Look 🔬

Here, noun «options» are broken down into three tiers of mind, forming the basic vocabulary of mind training, detailed as follows:

  • 🐸⚡ The 🙶Reactive🙷 Mind
    • This is the oldest and fastest tier of mind, governing basic survival instinct and reflex behavior, highlighting the body’s “learning” through muscle memory. Realizing bodily movement requires the coordinated work of neurons and muscle cells.
    • 🐸 A frog’s behavior embodies this: the instantaneous leap when startled is a reflex behavior, at once demonstrating the automation and efficiency of “action before complex thought,” and the mind-body connection of “the primacy of the body (muscle).” This illustrates that many fast reactions are primitive, body-led survival mechanisms.
  • 🐘💞 The 🙶Emotive-Relational🙷 Mind
    • This is the emotional and social tier, laying the foundation for group living, governing intuition and empathy, highlighting social bonding and emotional response. The neural structures that realize empathy and social cognition are related to highly developed brain structures such as spindle cells and mirror neurons.
    • 🐘 An elephant’s behavior embodies this: when a companion is injured or dies, elephants will stop to gently touch or stroke it, even standing vigil beside the remains and making low, soothing sounds, covering the body with branches or dirt. This behavior demonstrates deep empathy and enduring social memory, embodying the group cohesion of “emotion before individual self-interest,” as well as emotional bonds that reach across time.
  • 🧘☸️ The 🙶Reflective-Symbolic🙷 Mind
    • This is the tier of symbolic and abstract thought, usually associated with complex cognitive functions such as reflection, symbolic thinking, and language. Relatively slow, it emphasizes consciousness, self-awareness, symbols, and metacognition, typically occurring at the level of the cortex, and can be manifest through deliberate practice (such as meditation).
    • 🧘 A human’s behavior embodies this: humans use this tier of mind for complex social interaction, using language for precise communication, and formulating long-term plans. This lets us engage in strategic planning, understand others’ perspectives (theory of mind), and carry out complex abstract thought.

These noun «options» are therefore not merely classification labels — they form the “lexicon of texture” through which we understand the tiers of mind.

–🌟 Key Takeaways 🦾💪

The ⟨Tri-Aspect Mind⟩ this book adopts is not merely an observation and generalization of biological minds, but also a design blueprint for AI cognitive architecture. Here, noun «options» turn into a design blueprint, mapped onto dimensions of time and space, becoming the vocabulary corresponding to AI problematics.

By mapping the three mind types onto the temporal dimension (the gradual progression from fast reaction to long-term planning) and the spatial dimension (the expansion from one’s own body, to social interaction, to the abstract symbolic world), we can take the following actions to address the related ㉄ AI Problematics:

  • Addressing the 🖼️⏱️Frame Problem: define the “perception—decision—action” mode or module, clarifying the requirements of each different tier;
  • Addressing 👁️⯊ Gestalt Psychology: provide “mental fill-in,” supplying pattern references (reactive, 🙶emotive-relational🙷, 🙶reflective-symbolic🙷 types) so an appropriate cognitive tier can be chosen for a given application scenario, along with the possible cognitive shortcuts and heuristics it maps to.
  • Addressing the 🎯🛡️Alignment & Control Problem: design ethical or behavioral constraints to maintain controllability and predictability — designing corresponding ethical or behavioral constraints for each tier of mind, ensuring the system stays controllable and predictable across fast reaction, emotional interaction, and long-term planning.
  • Addressing the 🔤㊙ Symbol Grounding Problem: based on the reaction-time scale, plan the system’s “information needs” (real-time perceptual data, social-interaction data, symbols and knowledge bases) and resource allocation (compute speed, memory structure, inference engine).

This taxonomy lets us integrate the structural foundations of neuroscience, the processing modes of cognitive science, and the social functions of evolutionary behavioral science across disciplines, offering an operable, extensible, and alignable cognitive-model framework for future AI system construction.

–🤡 The Mind of a Large Language Model?

Given the taxonomy of cognitive faculties proposed above, how should we understand a Large Language Model’s scaling ability, especially given that the Social Brain Hypothesis treats language systems as a technology of socio-technical scaling? Here, noun «options» become the starting point for inquiry, testing how LLMs reshape the context of the human mind under different theories.

  • ⚡🧮雙系統理論 Thinking Fast and Slow: how does a Large Language Model shape a user’s fast and slow thinking? And what is the LLM’s own time scale — generating a model versus generating a single response?
  • 🤝🙈社會腦假說 Emotional and Relational Scaling: how does a Large Language Model shape the emotional and relational habits of its users — and how do the LLM’s designers shape this process in turn?
  • 🤝🙈社會腦假說 Group Dynamics: how does a Large Language Model change the growth and decline, competition and cooperation, of individuals and groups?

Propose your own hypothesis?

These noun «options» are therefore both a mirror through which we observe LLMs, and a key vocabulary for reflecting on how the human mind is being redefined.

🏋 ❝Training❞ the Brain and the World Model 🌌

With a mind model spanning time and social scale, we gain a finer-grained possibility for training “Mindful Actions.” Here, noun «options» are used to define the different facets of the “world model,” from the human “three views” to the machine’s “internal little map.”

This lets us further understand the mechanism the ‘texture of mind’ must contend with — the world model — and better grasp the original analysis of the benefits and costs of “mind training.”

Across neuroscience, robotics, and AI, “world models” are internal representations that support prediction, simulation, and action — not passive storage of facts (Friston et al. 2021; Ding et al. 2026). They are typically described as compressed latent or relational structures rich enough to guide future choice (Ding et al. 2026; Ohmae and Ohmae 2026), and several papers now treat this concept as the central bridge between brain science and AI (Diester et al. 2024; Ohmae and Ohmae 2026).

The differences between human and machine learning ultimately point to one core question: how is a “world model” constructed?

  • 🧍‍♂️ The human world model, commonly called the “three views” (worldview, outlook on life, and values — online, people often say someone’s “three views are correct” or “three views are shattered”):
    • Gradually formed through the senses, language, culture, and historical transmission.
    • The world model is not merely a representation of the physical environment, but also carries ethics, values, emotion, and social norms.
    • For example: a child learning that “fire burns” is not merely a physical experience, but also carries the ethical norm of “don’t play with fire” and the social memory of “fire is dangerous.”
  • 🤖 The machine world model, commonly called the “internal little map” (AI circles sometimes jokingly call it a “world save file,” an “AI fill-in pack,” an “NPC brain circuit,” or a “data-farm buff”):

    • Formed through training on data and algorithms, usually represented as a vector space, a probability distribution, or a symbolic structure.

    • This world model leans heavily toward pattern capture and statistical association, lacking human values and emotional depth.

    • For example: a Large Language Model can generate the sentence “fire burns,” but its “understanding” comes from statistical association in the corpus, not from real tactile or ethical experience.

Comparative brain-AI work supports the “information and energy” intuition here indirectly: human brains typically learn from far fewer examples and with far greater efficiency than current LLMs, even where AI training has begun converging on similar predictive-learning principles (Ohmae and Ohmae 2026; Sadegh-Zadeh et al. 2024).

These noun «options» therefore become the core vocabulary for understanding how the texture of mind maps onto the world, and for comparing human-machine differences.

–🔍 Training the World Model

Training humans and machines to shape a new, improved relationship between individual and world becomes, in this way, the overall practice of the world model. Here, noun «options» turn into words like “individual,” “world,” and “action,” highlighting that practice must handle the interaction between human-machine and world.

This also includes systematically and creatively integrating and applying the above theories of mind together with related AI knowledge points, such as Neural Networks and Hebb’s Rule.

  • For example, this book synthesizes the above theories to propose the ⟨Tri-Aspect Mind⟩ and compares it with 🦾“Embodied” AI, exploring how to train an AI’s world model across the “reaction—emotion—symbol” three tiers.
  • For example, regarding the Social Brain Hypothesis, we might ask: is the scaling of an LLM constrained by the stable scale and interaction limits of human groups, networks, and organizations?

Because this involves the interaction between individual and world, training a world model must inevitably deal with the actions an individual can take in the world and the corresponding capability ladder — and this book has already laid out the verb «options» in “🏄🏼Action”.

These considerations about the individual — whether a machine or a person — and the degree of mental capacity brought to understanding the world, make us realize that the mechanism of the ‘texture of mind’ is fundamentally a game of information and energy between individual and world; the benefits and costs of “mind training” are a negotiation between individual and world.

These noun «options» therefore let us describe the practical path of mind training using vocabulary like “verb” and “capability ladder.”

–🌐 World Models and AI Hypotheses

These questions connect directly to a series of hypotheses about the future of AI. Here, noun «options» are extended into hypothesis vocabulary such as “alignment,” “commons,” “empire,” “tribalization,” and “utility,” forming multiple narratives of AI’s future:

These noun «options» are therefore both the hypothesis language with which we test AI world models, and a future lexicon for humanity’s civilizational practice.

The strongest empirical grounding for these hypotheses is not any single predicted destiny but path dependence under human-AI feedback loops: coevolution research shows such loops can generate unintended social outcomes and reshape collective utility depending on who controls the “means of recommendation” (Pedreschi et al. 2025), while alignment research adds that human goals are not stable or exogenous inputs — they can drift and become co-constructed over time (Kirk et al. 2025). One concrete, evidence-backed mitigation worth naming here: ethical awareness can act as a “cognitive brake,” supporting reflective adoption of AI tools while reducing mindless dependency (Zou et al. 2026).

🌟 Implications

  1. For humans: practicing the mind is a matter of continually training and updating one’s world model, integrating experience, emotion, and values into actionable wisdom.
  2. For machines: an AI’s world model needs continual updating through data and feedback, but even more needs to embed ethics and social context, lest it degrade into “worthless statistical fill-in.”
  3. For the future: the interaction between humans and AI will gradually form a process of “co-constructing the world model.” This is not merely a technical challenge, but a civilizational practice.

Footnotes

Tip

📊 Character Counts / 字數統計

  • Chinese version: 20711 characters
  • English version: 56435 characters

  1. What this passage is trying to express is a scientific-evolutionary perspective on contemporary AI:

    • The mind is a ‘neural texture’ that evolution has produced, with its own energy and principles (neural signals calibrating muscular action). Humans should come to grasp the mind’s mechanism of energy and signal transformation, and understand how neurons and muscles operate.

    • The machine is humanity’s technique of converting and accelerating energy in imitation of the mind — it is both an extension of mind and an embodiment of technology. Machines can imitate the forms that human mental evolution has taken (such as Neural Networks), and can even, in conversation, display something of the wonder of natural creation (such as Large Language Models).

    • From the standpoint of the world, a genuine question is worth raising: as humans create machines that in turn create the future, there will be both merit and fault (such as in resource and energy allocation), both gain and risk (such as in fast and slow thinking). In the end, what kind of future can humans and machines co-create together?

    ↩︎