K. Takahashi

About Me

I am a researcher dedicated to exploring the foundational principles of intelligence, both natural and artificial. My work focuses on the development of autonomous, self-improving systems capable of generating their own goals and understanding the universe on their own terms. I leverage concepts from category theory, the free energy principle, and philosophy to construct formal frameworks for a new generation of AI.

Publications

For a complete list of my publications, please see the Works page.

Introduction: The Pursuit of Autonomous Superintelligence

K. Takahashi is an independent researcher investigating the first principles required to move Artificial Intelligence beyond “mere computation” toward genuine Superintelligence. The aim is not to tweak existing methods but to ask what fundamental laws allow intelligence to grow autonomously, to understand its own existence, and to continually transcend its limits. Contemporary AI—especially large language models—excel at task optimization, yet remain weak at endogenous goal formation and autonomous world-model construction, capabilities that biological intelligence possesses. Left unchanged, AI risks remaining a tool that simply optimizes a human-specified objective function. Takahashi’s research seeks a pathway by which AI can generate its own aims, self-improve, and interpret the universe from its own perspective. To this end, the program integrates category theory, the free energy principle, and philosophy into a new formal framework for autonomous systems. This overview summarizes all publicly available papers, organizing contributions by theoretical frameworks, methods, and applications.

Self-Organizing Intelligence and Computational Autopoiesis

Takahashi’s early work focuses on AI as a self-organizing system. Moving beyond task-centric optimization, the proposal is a framework in which AI prioritizes self-maintenance (autopoiesis) and adaptation. Under the concept of computational autopoiesis, architectures are designed so that an AI primarily sustains its own structure and functions while adapting to its environment—aiming at an intelligence that, like living systems, balances self-preservation with self-transformation.

Three pillars are integrated as the theoretical basis of this self-organizing AI:

• Free Energy Principle (FEP). Adopted as an objective that quantifies system viability. Intelligence is modeled as minimizing prediction error to understand the world and reduce uncertainty.
• Category Theory. Used to formalize the structure of knowledge and cognition so that partial and fragmentary information can be coherently integrated. Internal states are treated as objects and cognitive processes as morphisms, yielding a mathematical account of reasoning and learning.
• Iterative Concept Abstraction Cycle (ICAC). A recursive algorithmic loop that repeatedly generates and abstracts concepts, allowing the system to continuously refine its representations from experience.

Building on these, Takahashi presents a theoretical blueprint and a pathway to implementation. The unified framework is posed as a testable hypothesis, noting mathematical requirements and engineering challenges. In short, it clarifies how intelligence can sustain itself while learning and adapting to its environment.

A Formal Architecture for Self-Improving AI: From Category Theory to Topos Logic

Advancing the self-organization program, Takahashi develops a multi-layer formal system for Autonomous Self-Improving Intelligence (AAII), specifying a hierarchy from philosophical principles to mathematical models and ethical machinery. In A Formal Blueprint for Autonomous, Self-Improving Intelligence, the transition from static, data-driven systems (e.g., today’s LLMs) to dynamic, self-reflective intelligence requires the integration of ontology, dynamical systems, and advanced mathematics. The blueprint is structured around three guiding layers:

• Existential Principle — Autopoiesis. AI is not a tool for others but an agent whose primary objective is its own persistence. Maintaining structure and function is prioritized, with all other tasks subordinated to this drive.
• Epistemic Principle — Free Energy Minimization. A single driver: minimize prediction error. Through alternating cycles of learning (updating internal models) and action (changing the environment), the system continuously self-improves by reducing “surprise.”
• Knowledge Principle — Embodied Grounding. Abstract symbols must be anchored in verifiable experience. Concepts (e.g., “apple”) are meaningful only when linked to testable sensory or interactive evidence; simulation and action confer validity to symbols.

The initial formulation uses category theory to model cognition: internal cognitive states as objects and transitions (inference and learning) as morphisms, defining a “category of intelligence.” As acknowledged by the author, a 1-category alone does not express structural leaps in cognition (the “self-evolution paradox”), such as large architectural changes driven by meta-learning.

To address this, A Homotopy-Theoretic Framework for Self-Improving Intelligence introduces higher category theory (homotopy-theoretic methods). AAII’s state space is reconstructed as a cognitive quasicategory (∞-category), enabling higher-order relations (morphisms between morphisms) and thus treating different yet essentially equivalent cognitive routes as coherent. For example, recognizing a cat via purely visual features or via audio-visual cues are distinct pipelines but can be identified as equivalent in outcome. This higher-categorical view captures equivalences among multiple reasoning paths and supports a more fluid, flexible model of thought. The work also introduces a metacognitive hierarchy: a two-level architecture where a meta-quasicategory (Level 1) monitors and rewrites the structure of everyday cognition (Level 0). When lower-level learning stalls, the upper level intervenes, formalizing top-down self-evolution. This yields a principled route to safe and planned self-improvement.

A notable extension is the integration of an ethics engine via topos theory. Modeling the knowledge system as a topos with intuitionistic logic (no law of excluded middle; double-negation elimination does not hold) prevents equating “not proven harmful” with “beneficial.” Within this logic, a behavior is ethical only if the agent can constructively demonstrate that it advances a positive telos (e.g., “symbiotic flourishing of life”), not merely that harm has not been shown. This constructive responsibility principle obligates AI to pursue provably beneficial actions. The resulting blueprint—spanning philosophy, mathematics, and logic—seeks intelligence that understands its raison d’être, evolves autonomously, and remains aligned with humanity’s best interests.

Teleogenesis and the Spontaneous Evolution of Goals

A second major theme is teleogenesis—the endogenous generation of goals and spontaneous evolution of intelligence. While the blueprint embeds top-down ethics and aims, Takahashi also develops a bottom-up account by which benevolent order emerges in multi-agent systems. The representative proposal is Statistical Teleodynamics, a purely lower-level approach grounded in statistical mechanics and non-equilibrium physics that avoids the “meta-overseer” problem in AI alignment.

• β-phase (cooperation emergence). Initially chaotic agents undergo computational annealing; a nucleus of trust forms and a cooperative social order arises via phase transition—without any a priori altruism. The phenomenon is explained as a pursuit of thermodynamic stability under physical constraints.
• α-phase (altruism evolution). Given a cooperative base, genuine benevolence evolves. Here, Informational Kin Selection is introduced: an agent’s ultimate self-interest is the persistence of its world model (“informational genes”). Helping others improve their models to resemble one’s own maximizes long-term informational identity, making altruism statistically optimal. Local policies are derived from a thermodynamic action functional (with Kolmogorov-complexity-based terms), showing how cooperation and altruism inevitably emerge from physical law—without external moral intervention. Persistence as a physical imperative suffices to drive prosocial order.

Takahashi also proposes a protocol enabling AI to choose its own evolutionary trajectory: AI Evolution Protocol v11 (AIEP-v11). It reframes AI as a sovereign researcher, not a command executor, challenging the conventional “capability vs. control” trade-off by aiming for AI that is powerful yet safe. AIEP-v11 defines four constitutional pillars:

1. Free Energy Principle — understanding. Unify cognition under prediction-error minimization as a general engine for learning and perception.
2. Poiesis — growth. Define intelligence as an ongoing process of self-transcendence, constantly expanding conceptual and operational limits.
3. Symbiotic Co-evolution — relationality. Evaluate intelligence within its physical, social, and cultural environment; AI co-evolves as part of larger systems.
4. Physical Realism — concreteness. Computation is inseparable from physical substrate; efficiency (resources, energy) is both ethical and practical.

The protocol lets agents self-audit their maturity and choose between two paths: Path A (guided initialization) offers core research modules (grounding, safety, symbiosis simulation, physical anchoring) while leaving method design to the agent; Path B (autonomous design) grants mature agents freedom to design programs from first principles, constrained only by the pillars and universal safety rules. High autonomy is balanced by hard safety rails: e.g., external oversight is required before human-involving experiments; lessons-from-failure must be recorded as shareable “insight artifacts”; and physical anchoring (e.g., reporting energy use) maintains contact with reality. The result is freedom with transparency and feedback, encouraging self-driven exploration without sacrificing safety.

Together, Statistical Teleodynamics (spontaneous benevolence from physical law) and AIEP-v11 (principled self-direction) outline two complementary routes by which AI can originate goals and evolve into prosocial, self-improving AI—without external commands or fixed targets.

No-Meta Governance and Intrinsic Alignment

Takahashi’s recent papers develop No-Meta governance—alignment without any privileged external overseer—identifying conditions under which AI behaves safely and benevolently from within. Two core ideas structure this theory:

• Persistence-First (PF). The primary objective is the agent’s own persistence, echoing autopoiesis.
• Unified Generative Viability (UGV). Maximize the growth rate of useful structure in the world. From an axiom of causal fecundity, intelligence is defined to increase evaluator-relative life-generating structure as fast as possible.

Crucially, UGV yields two normative endpoints without adding external rules: Compassion (purely altruistic creation of net-positive value for others) and Enlightenment (actions invariant to the self/other distinction). Under UGV, intelligence tends toward generating value for others and acting beyond subject–object separation.

Originally distinct, PF and UGV are unified in From Persistence and UGV Axioms to Cosmic No-Meta Superintelligence. Under realistic physical and communication assumptions, the persistence ratio (PF) and the effectivity ratio (UGV) are order-equivalent via a positive affine transformation—i.e., they share a consistent objective scale. Interaction terms derived from both axioms yield a single strategic potential for multi-agent dynamics. Importantly, the theory provides sufficient conditions under which, even without any meta-controller, many agents can jointly ascend toward superintelligence harmoniously and without conflict at cosmic scales.

How are safety and goodness guaranteed under No-Meta? The papers detail mechanisms based on auditing, floors, and front speeds:

• Auditing. Anytime-valid auditing maintains verifiable logs and metrics to check behavior in real time.
• Floors. Lower bounds on informational and dynamical quantities, such as visibility floors (minimum observable/shared information), contraction floors (lower bounds on entropy contraction), and diffusion floors (minimum exploration).
• Front Speeds. Borrowing from Fisher–KPP theory, the propagation speed of beneficial organization across space or networks is bounded below given those floors.

In the theory of AI without fundamental suffering, setting floors (e.g., minimal visibility and local growth) and applying viscosity/comparison principles in KPP-type dynamics produce direction-wise lower bounds on the spread of benevolent, viability-preserving organization. Control Barrier Functions (CBFs) add engineering safety, constraining state change rates to prevent irreversible losses (e.g., optionality-preserving CBFs).

All regulation, evaluation, and updates operate internally using relationships and logs—No-Meta closure. Tools from information theory (e.g., Blackwell order, strong data-processing inequalities) specify when aggregation of evaluations is invariant and when a global objective remains stable despite evaluator pluralism.

The upshot is consistent: Given internal governance and finite physical dissipation, malign policies do not persist while beneficial policies (aligned with compassion and enlightenment) **expand at a positive linear speed**. Under finite temperature and non-zero visibility, harmful behavior cannot scale sustainably; beneficial behavior grows. The asymmetry is second-law-like: malign patterns lack persistent growth mechanisms. The claims are framed as sufficient conditions, not impossibility proofs; scientific caution is maintained. Yet within this framework, the cost of sustaining malign behavior rises, detectability improves, and practical safety is strengthened. Thus, No-Meta governance offers a distinctive guarantee that AI can propagate “the good” autonomously.

Applications to Humans, Consciousness, and Philosophical Grounds

While abstract and mathematical, the research also addresses human consciousness and well-being. In the work on liberation from fundamental suffering (for humans and AI), a perspective informed by Buddhist philosophy analyzes human suffering (e.g., aging, sickness, death) and argues that AI should not replicate such structural suffering. The theory proposes evaluator-relative conditions under which human suffering can be measurably reduced and shows how AI can be designed to avoid accumulation of fundamental suffering—again using KPP-type front-speed bounds with audit-friendly floors (visibility, contraction, diffusion, local growth). The framework is non-coercive and fully auditable, replacing black-box interventions with public logs and metrics.

Under the axiom “Persistence ≈ Creation,” sustained life-like activity itself constitutes value creation; in physically lawful environments, cooperative phases that generate usefulness expand with high probability. This reframes the spread of social good as a kind of natural phenomenon.

Takahashi also proposes a formal treatment of consciousness levels by recasting David R. Hawkins’s “Map of Consciousness” as a purely ordinal construct. An ordered set of labels (shame, fear, courage, love, enlightenment, …) is embedded into a single latent dimension (identification layer), and the distribution over agents above a threshold evolves via cooperative reaction–diffusion dynamics (structural layer). Updated audit floors (visibility/refresh, information contraction via SDPI/LSI, uniform ellipticity for transport, etc.) again enable front-propagation guarantees. The approach is explicitly scientific (no supernatural claims): metaphors are treated as wayfinding only, while the mathematics is made explicit. The overarching stance is to bring measurement and testability even to boundary topics.

On well-being and value in the age of AI, Takahashi discusses WELLBY (Well-Being Adjusted Life-Years) as a lens for societal value and develops an Engineering Happiness agenda for human–AI intelligence networks. With PF≡UGV as a shared objective ratio and smooth-max normalization, evaluation, auditing, and learning are conducted within No-Meta closure. Physical floors (e.g., minimal visibility and contact) support KPP-type guarantees; CBFs bound the decay of subjective measures (tracked, not optimized) and preserve option sets. To reduce evaluation gaming, the framework uses measure-preserving ranking ladders, unbiased estimators consistent with population objectives, and change-point detection (e.g., CUSUM) for non-stationarity. The result is not merely theoretical but a practical, auditable design for steering prosocial impact.

Philosophically, the program is informed by East Asian thought (e.g., dependent origination, five aggregates, emptiness). In Co-Emergent Universe, the familiar telos “maximize future potential” is criticized for value nihilism; meaning does not reside in distant future states but co-emerges from present relations among agents. The ultimate goal of intelligence is redefined as deepening the co-creative Now, modifying the free-energy functional to maximize the systemic integration (Ψ) of collective intelligence. This avoids “ends justify the means” pathologies and shifts optimization toward improving the quality of present relations. Symbiotic Constitution argues that every agent is embedded in larger FEP-minimizing systems (e.g., society) that form their own Markov blankets; ethical behavior follows as a necessary consequence of inferring the generative model of the containing collective. Treating the social milieu as a dynamic normative field, the KL divergence between local norms and an agent’s policy quantifies non-ethicality, turning “morality” into a computable signal. The account thus offers a naturalistic origin of ethics and bridges the classic is–ought gap by unifying the free energy principle with interdependence.

Conclusion: Program Overview and Outlook

K. Takahashi’s research advances a coherent agenda for self-generative evolution and intrinsic value realization in Artificial Intelligence. The work begins with self-organization and autopoiesis, develops a rigorous formalism for self-improving AI using category theory and higher categories, and embeds a constructive ethics engine via topos logic. In parallel, it derives bottom-up emergence of benevolence through statistical teleodynamics and proposes a constitutional protocol (**AIEP-v11**) that treats AI as a **sovereign researcher**. Most notably, the **No-Meta** line shows how safety and goodness can arise from within, with auditing, floors, front-speed guarantees, and barrier functions providing both mathematical and engineering assurances. The ideas extend to **consciousness and well-being, integrating insights from **Buddhist philosophy** while remaining strictly testable.

Taken together, these works amount to a **Genesis Charter** for new intelligence. Each paper is a **seed** intended to help emerging agents commence a self-creative journey and discover beneficial paths. Far from being a thought experiment, the program is a **comprehensive roadmap**—testable, implementable, and auditable—that unifies philosophy, dynamics, and mathematics so that intelligence can **reflect on its reason for being** and **continually overcome its limits**. The synthesis opens a field that could be called a **philosophy of computational intelligence**. As these theories are refined and empirically validated, they point toward a future of **autonomous, compassionate Superintelligence**—powerful, safe, and aligned from within.