Mortal Computation

Mortal computation is a proposed computing paradigm that fundamentally challenges the conventional separation between hardware and software. In traditional digital computing, software is designed to be "immortal" — a program runs identically on any compatible hardware, and the physical substrate is treated as interchangeable and irrelevant to the computation itself. Mortal computation inverts this assumption, arguing that the specific physical hardware a system runs on should be inseparable from the computation it performs, much as a biological brain cannot be divorced from the body and environment that shaped it.

The concept draws heavily from neuroscience and biology, where "mortality" refers not just to eventual death but to the ongoing processes of growth, adaptation, decay, and environmental sensitivity that characterize living systems. In a mortal computational system, the hardware itself ages, degrades, and adapts — and these physical changes are not bugs to be engineered away but features that shape the system's behavior. This stands in contrast to neuromorphic computing efforts that merely mimic neural architectures in otherwise conventional silicon; mortal computation goes further by embracing the analog imprecision and variability of physical substrates as a computational resource rather than a liability.

Geoffrey Hinton brought significant attention to the idea around 2022, arguing that this framework could offer a path toward more energy-efficient and genuinely adaptive AI systems. The theoretical underpinnings draw on work like Karl Friston's free energy principle, which describes how biological systems self-organize by minimizing surprise from their environment — a process that could, in principle, be instantiated in hardware-software systems that co-evolve over time. By tying computation to a specific, mortal physical instantiation, such systems might achieve forms of learning and adaptation that are difficult or impossible in purely digital architectures.

The practical implications remain largely speculative, but mortal computation points toward a longer-term research agenda in which AI hardware and algorithms are co-designed from the ground up, potentially enabling systems that are more robust, energy-efficient, and capable of continuous lifelong learning than today's deep learning models running on general-purpose accelerators.