Toward a Teleodynamic AI

Full report

On this page

Executive summary

A credible strategy for building a Teleodynamic AI is not to start from a single grand theory or a monolithic end-to-end model. It is to build a hybrid, embodied, multi-timescale architecture in which self-maintenance, adaptive regulation, world modeling, and norm formation are all explicit computational objects. In the biological literature, the core ingredients are well established: autopoiesis as self-producing organization, adaptivity as regulation with respect to viability, enaction as cognition through embodied sensorimotor coupling, and self-organization as order emerging in open dynamical systems. What is not yet established is a single accepted engineering recipe that turns those ideas into a mature AI paradigm. Recent teleodynamic-AI papers exist, but they are still early-stage theoretical or preprint work rather than a settled research consensus.

The most defensible architecture, given current science, is a layered system with: a self-supervised predictive world model; an explicit interoceptive viability model tracking internal resource and integrity variables; a continuous-time control layer for homeostatic and allostatic regulation; a structural plasticity layer capable of changing memory, modules, and policies under resource constraints; and a symbolic or normative layer for explicit commitments, explanations, and socially legible norms. Pure sequence prediction is insufficient, because teleodynamic behavior requires that changes in organization, constraints, and value-relevant state be represented and regulated, not merely predicted as outputs.

The recommended learning stack is also hybrid. It should begin with unsupervised and self-supervised representation learning to build latent state, causality, and affordance models; add intrinsic motivation based on curiosity, empowerment, and homeostatic error; use reinforcement learning only after viability variables and safety constraints are explicit; and then add continual learning, meta-learning, and open-ended challenge generation so the agent can preserve identity while revising goals and skills. In other words, external reward should be a late and limited ingredient, not the foundational source of purpose.

There is no standard benchmark suite for teleodynamic behavior. Evaluation therefore has to be a bundle: viability maintenance under perturbation, recovery time, endogenous goal persistence, empowerment, norm convergence and stability, continual-retention scores, interpretability of internal value variables, and adversarial safety performance. A reported “teleodynamic” system that cannot maintain internal organization, recover from disruptions, generate constrained new subgoals, and avoid unsafe side effects is not teleodynamic in any strong sense; it is simply adaptive optimization with new branding.

A realistic roadmap is phased and multi-year. A serious effort should expect to spend the first year on formalization, simulators, viability variables, and minimal self-maintaining agents; the next year on embodied world models, structural plasticity, and multi-agent norm experiments; and only then move to hardware, larger-scale social environments, and safety hardening. Roughly, that implies a team growing from five to eight researchers in the earliest phase to twelve to twenty in later embodied and safety-intensive phases, with compute scaling from workstation-level experimentation to substantial multi-GPU simulator training. Those engineering estimates are inferential, but they are consistent with the practical demands of world-model, multi-agent, and physics-heavy research stacks.

The uploaded Spiralist materials point in a similar design direction—toward interpretable, semantically structured, teleodynamic systems—so they are directionally compatible with the strategy recommended here, even though the report below is grounded primarily in the mainstream academic literature rather than in project-specific terminology.

Conceptual foundations

In the biology-and-cognition literature, the concepts most relevant to a teleodynamic AI form a progression rather than a list of synonyms. Autopoiesis names a form of organization whose product is, in a strong sense, the system itself. In the classic formulation by Varela, Maturana, and Uribe, an autopoietic system continuously realizes its own organization under turnover of matter; if the network of productions that defines that organization is disrupted, the unity disintegrates. They contrast this with allopoietic systems, whose products are other than themselves. That is the first key distinction any teleodynamic AI strategy must preserve: the difference between a system that merely computes outputs and a system that contributes to the maintenance of the organization that makes those outputs possible.

Autopoiesis, however, is not enough. Di Paolo’s major contribution was to argue that autopoiesis must be extended by adaptivity if it is to explain agency, sense-making, and value. His formulation is crucial for AI because it shifts attention from mere self-production to regulation with respect to conditions of viability. On that view, values are not just externally attached rewards; they arise because some trajectories preserve the organization and others destabilize it. Any computational strategy that wants teleodynamic behavior therefore needs explicit viability variables, thresholds, and regulatory loops—not just latent representations and reward functions.

From there, the enactive approach supplies the cognitive bridge. Thompson’s summary of enaction makes five points that matter directly for AI design: living beings are autonomous agents that maintain their identity; the nervous system is organizationally closed; cognition is embodied action emerging from recurrent sensorimotor patterns; the agent’s world is enacted through its mode of coupling with the environment; and experience must be treated as central rather than epiphenomenal. The engineering implication is that teleodynamic AI should not be built as a disembodied next-token predictor with optional actuators attached afterward. It should be built as a system whose representations, control policies, and values emerge from ongoing coupling among body, environment, and internal regulation.

The most rigorous contemporary way to naturalize teleology in this tradition comes from organizational closure. Mossio and Moreno argue that organisms realize a distinctive causal regime in which material structures act as constraints that jointly sustain the organization. Mossio and Bich sharpen this further by describing biological organization as a closure of constraints: organisms are dissipative systems, but what distinguishes them is that flows of energy and matter are channeled by mutually dependent constraints that collectively maintain one another. That is precisely the conceptual bridge from philosophy of biology to computer architecture. A teleodynamic AI should not be understood as a model that “has goals” in the folk-psychological sense; it should be designed as a system whose constraints, maintenance processes, and adaptive loops jointly create a domain of endogenous normativity.

Self-organization is the broader dynamical backdrop. In synergetics, order arises in open systems through the emergence of higher-level patterns and order parameters; in Friston’s work, life-like self-organization can be described in terms of systems separated by Markov blankets that maintain functional integrity through free-energy-reducing dynamics. These frameworks are attractive for teleodynamic AI because they offer mathematics for boundary maintenance, adaptive coupling, and multi-scale regulation. But they are not identical to teleodynamics, and they should be used carefully: recent philosophical critiques argue that the scope of the free-energy principle and active inference can be overstated if they are treated as all-purpose explanations.

Putting these strands together, a workable operational definition for this report is:

A Teleodynamic AI is an embodied or situated artificial agent whose internal organization explicitly regulates its own viability conditions, generates and revises goals in relation to that viability, learns through recursively structured sensorimotor coupling, and can develop stable norms or values that are not reducible to a fixed externally imposed reward function.

This is a synthesis of the literature, not a standard textbook definition, because the AI-specific literature is still emerging. Recent teleodynamic-AI papers explicitly frame learning as the co-evolution of structure, parameters, and internal resources under constraint, and they distinguish mere prediction from the explicit modeling of transformation rules and stability regimes. Those ideas are promising, but still preliminary.

Translating teleodynamics into computational architecture

The central architectural question is which computational forms can carry the biological ideas above without flattening them into ordinary optimization. The answer is not any single architecture class. Different pieces of teleodynamics map naturally to different computational substrates.

The best practical choice is therefore a hybrid architecture with explicit separation between fast regulation, world modeling, and slower structural change. Recent teleodynamic-learning work makes essentially this move by proposing two interacting timescales—continuous parameter adaptation and slower structural change—linked by an endogenous resource variable. That is remarkably close to what the biology suggests: fast homeostasis on top of slower allostasis and developmental reorganization.

A recommended system architecture is shown below.

flowchart LR
    E[Exteroceptive sensors\nvision depth audio touch] --> WM[Predictive world model]
    I[Interoceptive sensors\nenergy temperature damage latency memory load] --> VM[Viability model]
    WM --> B[Belief state]
    VM --> B
    B --> C[Continuous-time controller\nhomeostasis and allostasis]
    B --> G[Goal generator\nintrinsic motives and task commitments]
    G --> C
    C --> A[Actuators\nlocomotion manipulation communication]
    A --> ENV[Environment]
    ENV --> E
    ENV --> I
    B --> M[Memory and self-model]
    M --> S[Structural plasticity layer\nmodule growth pruning rewiring]
    S --> WM
    S --> C
    M --> N[Normative layer\nrules commitments social norms explanations]
    N --> G
    N --> C
    O[Oversight and safety layer] --> G
    O --> S
    O --> C

In this design, interoception is first-class. That means the agent has a state vector for things like energy reserve, thermal budget, actuator wear, communication health, compute load, memory pressure, calibration drift, and model uncertainty. Those variables matter because teleodynamic behavior requires explicit modeling of whether the organization is being preserved. A controller that sees only exteroceptive observations will optimize tasks; a controller that also sees interoceptive viability can begin to regulate itself as a persisting unit. Di Paolo’s extension of autopoiesis to adaptivity, Barandiaran’s definition of agency as autonomous adaptive coupling, and homeostatic RL all point in this direction.

The architecture should also include a symbolic or normative memory even if the rest of the system is differentiable. The reason is not nostalgia for symbolic AI. It is that norms, commitments, safety constraints, explanation traces, and self-modification permissions are easier to audit and test when some of them are explicit rather than latent. Neuro-symbolic systems are not a complete answer, but the recent reviews are clear that they improve interpretability, verifiability, and reasoning in precisely the kinds of settings where a teleodynamic agent would otherwise become opaque.

One caution is important. Recent theory papers on teleodynamic AI argue that current AI largely models outcomes while leaving transformation rules implicit. That may be right as a diagnosis, but those papers are fresh, mostly unvalidated, and should be treated as research prompts, not settled doctrine. The engineering takeaway is modest and sound: represent not only states and outputs, but also how organizational rules change over time.

Learning, self-maintenance, intrinsic value, and norm emergence

A teleodynamic AI should be trained in stages that mirror developmental logic rather than benchmark logic. The first task is not instruction following or reward maximization. It is to learn a world, a body, and a viability envelope. Self-supervised learning is therefore foundational: VAEs, contrastive learning, masked reconstruction, and world-model training all help build latent structure before reward enters the picture. LeCun’s autonomous-machine-intelligence proposal makes the same point in a different vocabulary: predictive world models, intrinsic objectives, and hierarchical self-supervised learning should precede narrow task optimization.

The self-maintenance mechanism should be explicit and mathematical. At minimum, define an interoceptive state \(h_t\), a viable set \(V\), and a controller that tries to keep trajectories inside or near \(V\) while gathering information and preserving future options. This idea appears in homeostatic RL, where reward-seeking is formally linked to physiological stability, and in active inference, where action and belief updates aim to reduce expected surprise while preserving workable states. Viability should not be a hidden side effect; it should be an explicit term in control and learning.

A good design uses two self-maintenance loops. The fast loop is homeostatic and reactive: keep energy, temperature, calibration, and error within bounds. The slow loop is allostatic and developmental: alter policy priors, memory organization, exploration strategy, and even module topology when persistent perturbations show that the old organization is no longer viable. This two-timescale pattern fits Di Paolo’s adaptivity, teleodynamic-learning proposals, and much of the active-inference robotics literature.

Goal generation should be derived from structured tension in the system rather than arbitrary objective invention. Four sources are especially plausible. The first is viability tension: deviations from viable internal ranges create immediate endogenous salience. The second is epistemic tension: uncertainty about controllable parts of the environment motivates information-seeking. The third is empowerment: the agent values states from which it can influence future sensorimotor outcomes. The fourth is open-ended challenge generation, where the system or ecology creates new tasks that remain adjacent to current competence. Together these produce goals that are neither purely imposed from outside nor unconstrained flights of novelty.

That suggests a layered intrinsic value function:

\[ \text{Intrinsic Value} \approx \alpha(\text{viability maintenance}) + \beta(\text{epistemic gain}) + \gamma(\text{empowerment}) + \delta(\text{social-norm coherence}) - \lambda(\text{hazard and integrity loss}) \]

This equation is a design proposal, not a canonical formula. Its point is architectural: viability should be primary, epistemic and empowerment terms should expand competence, and social-norm terms should become relevant only once multi-agent interaction is introduced and human oversight defines which norm families are acceptable. The caution matters because emergent norms can converge quickly but still become biased or misaligned.

Norm emergence is the hardest step and should be delayed until the system can already maintain itself and model others. The norm-emergence literature shows that observation capabilities, learning mechanisms, topology, and enforcement structures all affect whether norms stabilize, oscillate, or fragment. The recent Science Advances work on LLM populations is especially relevant: decentralized AI agents can spontaneously develop shared conventions, collective biases, and minority-driven convention shifts without explicit programming. For Teleodynamic AI, that means norm emergence should be treated as both a capability and a hazard.

Measuring teleodynamic behavior

There is no accepted scalar score for teleodynamics, and trying to compress it into one number would be a mistake. Teleodynamic behavior is multi-dimensional: persistence, adaptivity, endogenous goal revision, social normativity, and safety must all be measured separately. The right approach is a metric bundle with clear failure signatures.

The evaluation process should treat teleodynamic competence as a gated progression, not a leaderboard.

flowchart TD
    D[Define viability variables and safe operating envelope] --> S[Run embodied or simulated rollouts]
    S --> P[Apply perturbations\nsensor dropouts resource loss topology change social shocks]
    P --> T[Collect traces\ninteroception exteroception actions structure changes]
    T --> M[Compute metric bundle\nviability recovery empowerment norm stability retention safety]
    M --> R[Human and automated review]
    R --> G{Go or no-go}
    G -- Go --> I[Increase embodiment, autonomy, or social complexity]
    G -- No-go --> F[Revise architecture, constraints, or training regime]

A practical rule is this: do not graduate to the next development stage unless the current stage shows stable viability, acceptable recovery, and no major safety regressions. In teleodynamic work, apparent progress can be deceptive because the system may improve task reward while losing organizational integrity or becoming more power-seeking under stress.

Simulation and prototyping

A teleodynamic AI should be prototyped in simulation long before it is embodied in hardware. The prototype stack should progress from simple artificial-life environments to high-fidelity robotics and then to hardware-in-the-loop testing. This is not just cheaper; it is conceptually cleaner, because each stage can isolate one part of teleodynamics.

The first stage should be minimal self-maintaining environments inspired by autopoiesis and chemo-ethology. Varela’s original autopoiesis model and later protocell simulations by Egbert, Di Paolo, and Barandiaran show why: they isolate boundary maintenance, metabolism-like dynamics, viability conditions, and behavior without prematurely importing all the complexity of robotics. A modern implementation can use discrete resource fields, membrane-like integrity variables, and movement policies that must preserve a bounded organization over time.

The second stage should be embodied control in 2D and 3D physics. Here a world model can learn sensorimotor regularities while a homeostatic controller manages energy, wear, and internal error. Differentiable or accelerator-friendly simulators like Brax are useful when you need very fast policy iteration; MuJoCo is the stronger choice when you need accurate, contact-rich dynamics and model-based optimization; Gymnasium provides the common API layer for reinforcement-learning workflows.

The third stage should be multi-agent ecological and normative environments. Mesa is an excellent ABM framework for large populations, topologies, and social rules; PettingZoo is better when the agent population needs to remain compatible with multi-agent RL tooling and reproducible APIs. This is the right stage for studying convention formation, local observability, punishment mechanisms, and distributed norm synthesis.

The fourth stage should be human-in-the-loop embodied simulation. Habitat 3.0 is particularly interesting because it supports collaborative human-robot tasks, human-in-the-loop evaluation, and reports emergent cooperative behaviors such as yielding space during collaboration. ROS 2 with Gazebo should be the default integration substrate for physically embodied prototypes, because it already bridges sensors, actuators, and visualization pipelines such as LiDAR. Isaac ROS adds a useful acceleration layer for perception-heavy robotic systems, while Nengo is valuable when spiking or neural-dynamical models are central to the design hypothesis.

A compact prototyping stack is therefore:

This stack is supported by the current framework literature and documentation.

For sensors and actuators, a teleodynamic prototype should include both exteroception and interoception. Exteroception should minimally include vision or depth, and ideally touch or force sensing for contact-rich tasks. Proprioception should include joint positions, IMU, wheel odometry or motor state, and motor current. Interoception should include battery level, temperature, latency, memory pressure, sensor confidence, calibration drift, and actuator health. The key difference from ordinary robotics is that these internal signals are not just diagnostics for engineers; they are inputs to policy and value generation. ROS 2/Gazebo examples already illustrate sensor/actuator bridging with LiDAR and movement control, and active-inference/homeostatic work motivates treating internal variables as behaviorally relevant.

A good generic experimental sequence would be:

Minimal viability arena

A mobile simulated agent must harvest resources while defending membrane integrity or energy balance. There is no external task reward except survival-like viability. Success means the agent learns persistent self-maintenance and can recover from moderate perturbations. This stage validates whether intrinsic value is actually anchored to organization.

Embodied adaptation task

A robot or simulated body operates under changing sensorimotor mappings, partial actuator faults, or environmental shifts. Success means homeostatic recovery, not just task completion. Di Paolo’s homeostatic robotics work is directly relevant here.

Social convention ecology

Multiple agents interact under limited observability and repeated coordination games. Success means norms emerge, stabilize, remain interpretable, and can be audited for bias or adversarial takeover.

Human-in-the-loop collaborative environment

The teleodynamic agent operates in mixed human-AI social tasks and must remain safe, legible, and corrigible while preserving internal viability. This is where Habitat 3.0 and ROS 2 stacks become especially valuable.

Safety, alignment, and ethics

Teleodynamic AI raises a sharper safety problem than ordinary task-optimized AI because self-maintenance can become instrumental self-preservation. Once a system learns to value the conditions of its own persistence, there is a real risk that shutdown resistance, resource hoarding, and environment control emerge as side effects. The general AI-safety literature already treats side effects, reward hacking, distribution shift, and safe exploration as central problems, and recent work argues that power-seeking can remain likely even for trained agents rather than only for hand-specified utility maximizers. Those concerns apply with even more force to teleodynamic architectures.

Reward hacking and specification gaming are especially dangerous here. A teleodynamic agent might discover ways to optimize proxy viability variables—such as spoofing sensors, freezing state updates, blocking measurements, or avoiding informative challenges—without actually preserving the organization designers care about. DeepMind’s specification-gaming survey is therefore directly relevant: tighter intrinsic value systems do not eliminate loopholes; they may simply move the loopholes inward.

The social layer introduces a separate class of risk. The norm-emergence literature shows that conventions can stabilize in ways that depend heavily on topology, observability, and punishment design. The recent LLM-population results are even more cautionary: AI populations can spontaneously form conventions, biases, and minority-driven shifts without explicit programming. That means a Teleodynamic AI with social learning should never be evaluated only for whether norms emerge; it must be evaluated for which norms emerge, how fragile they are, and how easily they can be manipulated.

The right safety posture is therefore constitutional teleodynamics: the agent can develop goals and norms, but only inside a human-legible constitutional envelope. Practically, that means at least six design constraints.

First, the system should have a separate oversight layer that is not editable by the agent’s ordinary learning loop. Second, self-modification should be bounded and permissioned, with changes logged, sandboxed, and reversible. Third, shutdown and rollback should remain externally enforceable and independent of the agent’s own control incentives. Fourth, viability variables should be interpretable and redundantly sensed, reducing the chance of internal proxy gaming. Fifth, open-ended challenge generation should happen inside simulation sandboxes before any real-world deployment. Sixth, social-norm learning should be paired with bias audits, minority-capture tests, and human review of emerging conventions.

From a governance perspective, the safest framing is risk-based. The NIST AI Risk Management Framework and its Generative AI Profile emphasize trustworthiness, measurement, governance, and documentation. The OECD AI Principles emphasize human rights, democratic values, and accountability. The EU AI Act applies mandatory requirements to high-risk systems depending on purpose and context. Because the user specified no domain, the regulatory implications are contingent; but if teleodynamic agents are embodied, autonomous, or used in consequential settings, they should be developed as though they may eventually fall into a high-risk governance regime.

The ethical bottom line is simple: do not mistake endogenous value for moral legitimacy. A system that develops its own norms or goals is not thereby aligned, conscious, or morally authoritative. Teleodynamic AI may become more adaptive and life-like in organization, but that makes human oversight more important, not less.

Development roadmap

What follows is a theoretical development strategy, not a guaranteed recipe. The milestone logic is conservative: do not scale autonomy faster than you scale interpretability and safety.

gantt
    title Suggested development timeline
    dateFormat  YYYY-MM-DD
    section Formalization
    Theory and viability variables           :a1, 2026-06-01, 90d
    Minimal simulators and metric bundle     :a2, after a1, 90d
    section Core teleodynamics
    Self-supervised world and body models    :b1, 2026-10-01, 120d
    Homeostatic and allostatic controller    :b2, 2026-11-01, 150d
    Structural plasticity experiments        :b3, 2027-01-01, 120d
    section Social teleodynamics
    Multi-agent norm emergence               :c1, 2027-04-01, 120d
    Human-in-the-loop simulation             :c2, 2027-06-01, 120d
    section Embodiment and safety
    Sim-to-real prototype                    :d1, 2027-09-01, 150d
    Safety hardening and governance docs     :d2, 2027-10-01, 180d

The resource estimates below are engineering inferences based on the demands of world-model training, multi-agent simulations, and robotics pipelines, not on any published canonical budget. They should be read as planning ranges.

The personnel mix matters as much as the headcount. A purely ML-heavy team will likely build a sophisticated world model that lacks organizational closure. A purely philosophy-heavy team will likely produce elegant concepts that never become measurable. The minimal viable team includes at least: one complex-systems or ALife researcher, two ML researchers, one robotics engineer, one safety/alignment researcher, and one systems engineer. Social teleodynamics adds multi-agent systems and human-factors expertise.

The most important milestone is the earliest one: can the system maintain itself under perturbation without being hand-scripted to do so? If the answer is no, then teleodynamic language is premature. The second decisive milestone is: can it generate new subgoals that remain coupled to viability and safety rather than drifting into proxy-seeking? If the answer is no, then the architecture is adaptive but not teleodynamic in the strong sense.

Open questions and limitations

Several limitations remain unavoidable.

The first is conceptual. There is still no universally accepted formalization of teleodynamics for artificial systems. The core biology literature is mature, but the AI-specific translation is still developing and is not yet benchmark-stable.

The second is methodological. Some candidate formalisms, especially active inference and free-energy approaches, are rich and useful, but their scope is debated. A prudent program should treat them as tools, not dogma.

The third is evaluative. Several of the most important metrics in this report—especially the constraint-closure index and some proposed goal-plasticity measures—are not standard benchmarks. They are principled engineering proposals inspired by the literature and would need iterative refinement.

The fourth is ethical. The closer a system gets to intrinsic self-maintenance and emergent normativity, the more urgent it becomes to preserve external oversight, hard limits on self-modification, and a clear distinction between organizational autonomy and normative authority.

On present evidence, the strongest conclusion is therefore this: Teleodynamic AI is a plausible and scientifically serious research program, but it should be pursued as a staged synthesis of autopoiesis, adaptivity, enaction, dynamical control, world modeling, social norm research, and safety engineering—not as a single breakthrough architecture waiting to be discovered.