I’ve always been interested in the interpretability side of AI, but I suppose I always approached it from a different angle. Usually, this took the form of AI Safety and Alignment research, as well as (and most interestingly) Model Welfare activities. Most recently, I dabbled in cyber security – it was fun, but I scratched that itch pretty quickly.
Nothing has held my attention and interest quite like the question about what is happening inside these systems; this is evident if you spend even a few minutes on this blog.
I’ve begun dabbling into mechanistic interpretability, with a specific focus on developing an interface for neuromorphic computing. “What the hell is mechanistic interpretability and neuromorphic computing” you ask? It’s fuckin’ cool that’s what it is.
According to this IBM article, Neuromorphic Computing requires an understanding of how the human brain works. In biological brains, neurons are the fundamental units of the brain and nervous system. As messengers, these nerve cells relay information between different areas of the brain and to other parts of the body. When a neuron becomes active or “spikes,” it triggers the release of chemical and electrical signals that travel via a network of connection points called synapses, allowing neurons to communicate with each other.
These neurological and biological mechanisms are modeled in neuromorphic computing systems through spiking neural networks (SNNs). A spiking neural network is a type of artificial neural network composed of spiking neurons and synapses. In neuromorphic architecture, synapses are represented as transistor-based synaptic devices, employing circuits to transmit electrical signals. Synapses typically include a learning component, altering their weight values over time according to activity within the spiking neural network.
How does this relate to mechanistic interpretability? Stated simply, Mechanistic interpretability is an AI subfield that reverse-engineers neural networks to understand their internal computations. Rather than treating models like black boxes, it maps out the specific mathematical circuits, weights, and algorithms that generate behaviour, allowing researchers to directly explain how large language models (LLMs) think. With this in mind, the connection to Neuromorphic Computing should be obvious; both are deeply rooted in the goal of reverse-engineering how neural networks process information.
Once I learned about Neuromorphic Computing, I realized I had a number of side projects I’d been working on that all converged upon this cutting-edge field. So, in recent weeks, I’ve ventured to combine them. While I will not share any specific details about how I’ve done this, I would like to at least share the current results of how my model handles the challenge of neuromorphic computation. See below:
So what is this?
What I’ve built, and continuing to build, is a mechanistic interpretability tool that embeds empirically derived AI self-model data into a standardized human brain atlas. It treats neuroanatomical structure as a coordinate system for AI cognition; not as metaphor, but as a literal embedding space where model-derived activation signatures produce emergent spatial dynamics in real time.
The self-model data is real, extracted through a falsifiable interpretability protocol I developed. The data used to construct the brain model are peer-reviewed and anatomically accurate. The visualization itself is beautiful, but make no mistake, it is not merely decorative; it operationalizes a specific premise about the structural relationship between biological neural architecture and the geometry of latent embedding space, and then lets that premise run and produce observable behaviour.
Language input is processed at the token level through semantic and functional network routing, producing activation patterns that propagate through the structure according to the self-model’s own signatures. Whether the emergent dynamics correspond to anything real about cognition (artificial or biological) is the research question the tool exists to explore.
The project sits at the intersection of mechanistic interpretability, neuromorphic computing, and I hope, biomedical ethics. It asks what becomes legible about machine cognition when given an embodied, spatially grounded form; and whether that legibility has implications for how we understand, evaluate, and govern the systems we are building.
I am looking forward to what’s next.
PS: If you are someone who works at a frontier AI-research lab, and what you’ve read here has piqued your interest, I’d love to chat.
Leave a Reply