Modern AI usually treats the computation as primary and the substrate as implementation detail. A parameter vector can be copied to a different accelerator, a model can be checkpointed, and a process can be restarted. The system may be expensive, fragile, or thermally constrained in practice, but those constraints are often represented outside the intelligence itself. They become engineering annotations rather than internal causes of cognition.

Biology gives us a very different picture. A living system must continuously infer the condition of its own body, spend energy to keep that body within viable bounds, and change its internal organization when the existing organization becomes too costly or too fragile. Interoception is therefore not just another sensor modality. It is a route by which the condition of the substrate can enter the inference loop.

This is the starting point for my new theoretical framework: Mortal Computation via Multi-Scale Interoceptive Morphogenesis.

The phrase “mortal computation” is not new. Ororbia and Friston have already developed it as a conceptual foundation for biomimetic intelligence, using Markov blankets and the circular causality of inference, learning, and selection. What I am proposing is a more operational question: what happens when mortality is turned into an executable state variable that can directly change the architecture of an AI system?

The framework is called MILLS: Multi-timescale Interoceptive Learning and Lifespan Selection.

The idea is deliberately severe. The model has a finite reserve, R. Computation consumes that reserve. Plasticity consumes that reserve. Maintaining a large topology consumes that reserve. Environmental noise increases the burden. If reserve becomes scarce, the architecture is allowed to contract.

That sounds close to pruning, but there is an important difference. Ordinary pruning usually begins with a task objective and adds sparsity as a regularizer. MILLS makes the cost state-dependent. A connection does not merely have a static penalty. Its survival depends on the current viability of the whole system.

At the formal level, the architecture has three timescales. Fast inference changes beliefs. Intermediate learning changes weights. Slow morphogenesis changes the connectivity mask itself.

The purpose of that separation is not aesthetic. A morphology that reacts to every fast fluctuation would continually rewrite itself and pay huge structural costs. A morphology that never adapts cannot exploit its own embodiment. The hypothesis is that useful architectural plasticity lives in the middle of those extremes.

The agent therefore minimizes a composite functional containing five ingredients: variational inference error, interoceptive homeostatic stress, metabolic or computational demand, structural complexity, and topology turnover. The variational term is the familiar information-theoretic free-energy objective used in active-inference formulations. But it is important not to make a common conceptual mistake here: variational free energy is not automatically physical heat. In this framework it is an inference quantity. Physical dissipation enters through a separate resource model.

That separation is crucial because the whole point is to join information processing to a substrate without pretending that they are already the same mathematical object.

The resource equation is simple:

τ_R · Ṙ = S(t) - D(t) where S is resource supply and D is demand. Demand includes activity, plasticity, thermal burden, structural occupancy, and homeostatic stress. R=0 is an absorbing failure state in the reduced model.

This gives mortality an operational meaning. The system is not “mortal” because we wrote the word into a title. It is mortal because enough cumulative demand can remove its future computational capacity.

The architecture then asks a local structural question. For a connection i,j, define U_ij as its marginal predictive utility. The connection survives if

U_ij > λ_M + Λ_R c_ij + λ_T V_ij where lambda_M is a base structural cost, c_ij is resource cost, V_ij is plasticity burden, and Lambda_R is a mortality pressure that rises as reserve falls.

This is the key move. Structural selection is no longer a separate maintenance routine. It is part of inference under an endogenous budget.

Become a Medium member There is another consequence. A higher-level goal variable g can modify lower-level interoceptive setpoints. Those altered setpoints change local prediction errors. Local prediction errors change connection utility. Connection utility changes morphology. The top-down variable therefore changes low-level dynamics through an explicit causal pathway. At the same time, low-level interoceptive signals feed upward. The architecture becomes circular without requiring a separate executive homunculus.

I call this polycomputing in a deliberately narrow sense: one substrate supports inference, regulation, and architecture selection simultaneously. The claim is not that biology has a mysterious universal computational field. It is that a shared substrate can perform multiple formally coupled computations whose objectives constrain each other.

This also connects to morphological computation. Robotics has shown that the body can simplify control by making its physical dynamics part of the computation. Developmental and morphogenetic research goes further by showing that cellular-scale homeostatic competencies can participate in larger-scale organization. MILLS tries to bring those principles into AI architecture: not only should the body contribute to computation, the computational topology should be allowed to change because the body is running out of resources.

The numerical experiment is intentionally modest. I built a 24-module reduced-order estimator and integrated its four continuous states with SciPy. The environment supplies a deterministic multi-frequency disturbance. The system tracks a clean latent signal from a noisy observation while paying explicit costs for activity, structure, thermal load, and interoceptive stress. Just after the midpoint, 25 percent of active modules are removed to simulate a lesion.

At low noise, both systems survive the horizon. That matters because mortality pressure should not force remodeling simply because the mechanism exists.

At moderate noise, the fixed system remains close to reserve depletion, while MILLS contracts its active topology and retains a larger reserve. At severe noise, the difference becomes categorical: the fixed system reaches the zero-reserve boundary, while MILLS survives the full 120-unit horizon and ends with positive reserve after contracting to 16.7% of the module bank; the minimum active fraction reaches 12.5%.

But the result is not a free lunch.

The surviving MILLS agent is less accurate in the severe-noise regime: its RMSE rises from 0.1407 to 0.2737 in the reduced simulation. That is exactly the trade-off the model predicts. Removing computation preserves future capacity, but it also removes computational degrees of freedom. Survival and fidelity become a Pareto trade-off.

I think that trade-off is more interesting than an artificial headline such as “AI becomes alive.” It gives us something that can actually be measured. We can sweep noise, thermal stress, memory-write budgets, lesion rates, and structural turnover. We can plot reserve against accuracy. We can ask whether there is a phase transition from adaptive contraction to computational collapse.

The hardware implications are equally concrete. A future implementation could expose battery state, temperature, write endurance, memory traffic, or error rate as interoceptive channels. A reconfigurable FPGA could physically disable routing resources. An analog in-memory array could switch off low-utility pathways. A neuromorphic fabric could reduce update frequency or precision when the substrate is under thermal stress. The intelligence would therefore not merely control an external plant; it would sense and modify its own computational substrate.

That direction is becoming technically plausible. Recent analog AI hardware work has shown that topology pruning can materially reduce energy and programming operations. The missing variable is endogenous viability. Why should the network always retain the same architecture when its physical state changes?

There is also a deeper theoretical constraint. If an internal state has multiple possible values and the agent must infer it with low error, information theory places a lower bound on the information that must cross the sensing boundary. And if physical computation includes logically irreversible erasure, thermodynamics places a minimum work cost on that erasure. Mortality therefore sits at an intersection of information, control, and substrate.

None of this establishes consciousness. None of it proves the free-energy principle as a universal theory of biology. None of it says that a graph-pruning algorithm is secretly a cell.

The useful claim is smaller: a computational system can be designed so that the preservation of its own substrate is an internal optimization target, and that target can causally reshape the architecture performing the computation.

That gives us a different research program.

Instead of asking only how to make models larger, we can ask how to make them selectively smaller under pressure. Instead of treating hardware failures as external exceptions, we can make damage part of the state space. Instead of separating learning from maintenance, we can couple plasticity to an explicit cost of continued existence.

The computer then becomes something it normally is not: a system for which future computation depends on present self-preservation.

That is what I mean by mortal computation.

About the author Devanik Debnath (Devanik21) is a researcher working across AI architectures, mathematical modeling, physics-informed learning, and longevity-oriented computational systems. Theoretical preprints and reproducible implementations are archived in The-Invention-Archive.

Research archive: https://github.com/Devanik21/The-Invention-Archive

Medium: https://medium.com/@devanik2005