Ludwig von Bertalanffy

In the [YOU] on AI Field Guide

The cycle that began with [YOU] on AI describes AI systems as extraordinary capabilities that require new habits of mind to use well. Bertalanffy supplies the most precise theoretical tools for understanding what makes these systems so difficult to characterize—why they keep surprising their builders, why certifying their safety is harder than certifying a bridge, why the question “what can it do?” turns out to require a perpetual answer rather than a settled one. His open-system analysis says: a deployed model’s behavior is not a fixed property of its weights but a steady state maintained through continuous exchange with a changing environment. When the environment drifts—when users shift, the world’s text changes, or the deployment context evolves—the behavior drifts, even though not a single parameter has been updated. The closed-system intuition that a trained model is a finished artifact is precisely the intuition Bertalanffy spent his career demolishing.

His concept of equifinality—that many paths lead to the same competence—reframes both the triumphalism and the alarm surrounding frontier AI. It is deflationary toward any claim that a specific architecture or proprietary method is uniquely responsible for a given capability, since equifinal systems are path-independent and the competence belongs to the regime rather than the road. It is equally sobering toward those who hope to prevent dangerous capabilities by blocking one approach: if the capability is an attractor reachable from many starting points, closing one road does not close the destination.

Emergent capabilities—abilities that appear abruptly at scale, that no one designed in and that were not present in any smaller version of the system—are Bertalanffy’s constitutive/summative distinction made empirical. He argued eighty years ago, in exactly the language the field has rediscovered: that a complete inventory of the parts can leave you ignorant of the whole’s most important properties, because those properties are constituted by the relations among parts, not by the parts themselves. The challenge of mechanistic interpretability—the project of understanding AI systems by decomposing them into circuits and features—is fighting Bertalanffy’s wall directly, and its successes so far have been precisely in the summative cases, while the constitutive, whole-level behavior continues to resist decomposition.

He stands alongside Norbert Wiener as a systems thinker who arrived at the edge of the AI age with the right conceptual equipment and the wrong expectation about how far mechanism would reach. Wiener built cybernetics; Bertalanffy built General System Theory; each was right about the power of systems thinking and cautious about the reduction of life and mind to mechanism. Each has been partially vindicated and partially surprised by the scale of what mechanism has achieved.

Origin

Born in Vienna in 1901, Bertalanffy studied biology and philosophy at the University of Vienna and developed his first systematic ideas about open systems in the 1920s and 1930s. His 1932 Theoretische Biologie laid out the argument that classical physics, built for closed systems approaching equilibrium, was the wrong framework for living things, which maintain themselves far from equilibrium through continuous exchange with their environment. His concept of the open system predates by two decades the cybernetics movement with which it is often confused, and he was careful to distinguish his approach from Norbert Wiener’s: cybernetics was, in his view, a theory of regulation through feedback, while his General System Theory was a theory of the spontaneous, self-organizing, growth-oriented wholeness of living systems—a distinction, he insisted, that feedback loops alone could not capture.

General System Theory as a formal program was articulated in a 1945 lecture and expanded through the 1950s and 1960s, culminating in his 1968 book General System Theory: Foundations, Development, Applications. The Society for General Systems Research, which he co-founded in 1954 with Kenneth Boulding, Anatol Rapoport, and Ralph Gerard, became the institutional home for cross-disciplinary research on the mathematical forms common to systems of all kinds. He spent extended periods at the University of Alberta and the State University of New York at Buffalo, dying in Vienna in 1972, four years after the book that would become his most widely read legacy.

His anti-reductionism was not vitalism; he explicitly rejected the idea that living things require a special ingredient absent from matter. His claim was structural: that organization is causally real, that arrangement matters, that you cannot fully predict the behavior of a whole from the behavior of its isolated parts. This is the claim the age of AI has had to rediscover empirically, by watching emergent capabilities appear in scaled models and finding that the training objective—predict the next token—is a sufficient account of how the capability was trained but not of what the capability is.

Key Ideas

The open system. A living system is not a closed mechanism winding down to equilibrium but an entity constituted by continuous exchange with its environment, maintaining itself in a steady state far from equilibrium through inflow and outflow. A trained AI model, understood through this lens, is not a finished artifact but a behavioral steady state that depends on the exchange continuing. Alignment is not a property installed at training and then guaranteed; it is an ongoing accomplishment that can drift as the environment evolves.

Equifinality. The same final state can be reached from different initial conditions and by different paths. In biological systems, this means an organism can develop normally despite early damage or rearrangement. In AI, it means that many architectures, trained from different random seeds on shuffled data, converge on near-identical capabilities—because the competence is an attractor of the system class, not a property of the specific path. This is both deflationary (no architecture has unique claim to its capabilities) and sobering (capability attractors cannot be blocked by controlling any single road to them).

Constitutive vs. summative properties and emergence. Summative properties of a system are addable from its parts; constitutive properties depend on the specific relations among parts and cannot be derived from isolated-parts inventories. Emergent capabilities in AI are constitutive in exactly this sense: absent in every part and in the training objective, present in the organized whole. The wall that interpretability research keeps hitting—finding the summative structure readily and the constitutive behavior stubbornly—is Bertalanffy’s wall, encountered empirically eighty years after he described it theoretically.

System isomorphisms. The same mathematical forms recur across biology, physics, sociology, and economics not as coincidence but as evidence of genuinely common organizational structure. The equations governing population growth and chemical kinetics and feedback control are structurally identical because they describe the same abstract properties of systems-in-general. The structural resemblance between statistical mechanics and machine learning—the fact that energy landscapes in condensed matter physics and in neural networks are governed by related mathematics—is a Bertalanffian isomorphism, and it is why physicists from Boltzmann to Hopfield have been able to import their methods into AI without category error.