As humanity transitions into an era of deep-space exploration and permanent off-world settlements, the computation requirements for autonomous systems have scaled exponentially. However, the current trajectory of Artificial Intelligence—dominated by Large Language Models (LLMs) and autoregressive Transformers running on massive von Neumann GPU clusters—is fundamentally incompatible with the physics of space.
GPUs require megawatts of power and massive cooling infrastructure. They are physically fragile; as transistor nodes shrink below 3nm, they become critically vulnerable to Galactic Cosmic Ray (GCR) bit-flips. Shielding them requires prohibitive mass (lead shielding or localized magnetic fields), and redundancy requires carrying triple the hardware footprint.
This discussion paper proposes a paradigm shift: abandoning top-down, centralized computation in favor of Bottom-Up Biological Computation, leveraging Michael Levin’s Technological Approach to Mind Everywhere (TAME) framework implemented on Spiking Neural Networks (SNNs) such as BrainChip’s Akida.
In a traditional von Neumann architecture, hardware failure is handled via a centralized, top-down diagnostic check (a CPU polling its subsystems for error codes).
If a cosmic ray strikes a critical GPU cluster on a Martian life-support system, the diagnostic center itself may be corrupted. The system crashes because it is completely dependent on a static, pre-programmed hierarchy of health checks. It is brittle. It does not bend; it breaks.
We propose replacing centralized diagnostic polling with Peter van der Made’s Event-Based Processing.
In a Spiking Neural Network (SNN), nodes (neurons) communicate exclusively via discrete electrical spikes, and only when there is active information to share. If a cell or a node is physically destroyed by a cosmic ray, it does not send an error code. It simply stops spiking.
The failure detection mechanism is entirely localized and bottom-up. The surrounding nodes detect a void—a sudden silence in the bioelectric rhythm of the network.
Drawing on Stuart Kauffman’s theories of biological purpose, the network is not optimizing for "next token prediction," but for a physical Function—such as maintaining 100% atmospheric oxygen pressure. Michael Levin refers to this homeostatic set-point as the Target Morphology or the Attractor (M*).
When a cluster of neuromorphic nodes goes silent due to radiation damage, the global output of the life-support system drops below the Attractor (M*). The network recognizes it is falling away from homeostasis.
Because the SNN is optimized to seek the Attractor, the surviving nodes do not wait for a central CPU to issue a repair order. The localized silence and the subsequent drop in system output trigger immediate Synaptic Plasticity.
The surviving neighbors autonomously strengthen their own synaptic weights and increase their firing rates. They essentially signal to one another: "Our neighbors have gone silent and we are missing our target output. We will increase our load and carry the weight until the hardware is replaced."
This creates an instantaneous, biological workaround. The SNN "heals" its operational output despite permanent physical damage to the hardware.
While Mars provides the ultimate radiation stress-test, this architecture has immediate, critical applications on Earth. Because BrainChip's Akida is ultra-low-power and can be deployed as an IP block on simple microcontrollers, this biological failsafe does not require a data center.
A neuromorphic failsafe switch can be embedded directly into remote edge environments running critical systems: a water pump in the Australian Outback, a pressure valve on a deep-sea oil rig, a drone monitoring a wildfire, or a satellite in low Earth orbit.
If the primary system is damaged by heat, pressure, or hardware failure, the embedded SNN does not crash. It detects the localized drop in function (e.g., water pressure falling) and automatically re-routes the signal, compensating for the failure to maintain a safe hold state until human engineers arrive.
The future of Martian operating systems will not be found in server farms running Transformers. It will be found in synthetic biological tissues running on Neuromorphic silicon.
By merging the event-driven hardware sparsity of BrainChip with the morphogenetic algorithms of the Levin TAME framework, we can deploy autonomous, low-power, radiation-resilient AI to the edge of the solar system.
Open Questions for the Community:
(This is an open-source engineering manifesto. Collaboration and critique are required to push this to the edge of the adjacent possible.)