There is a staggering elegance in the way biological reality preempts our computational tragedies.
Lately, I have been using my X feed as a sort of public notebook—a Pinterest board for the cross-pollination of systems biology and embodied AI. I leave raw, in-the-moment notes to myself, throwing fragments of thought into the digital void, hoping they might assemble into a coherent architecture. And this week, looking at a pair of newly published papers—one on the deep mechanics of the mammalian brain, the other on a persistent failure mode in machine learning—I was struck by a resonance so perfect it felt almost like poetry.
We are watching the dialectic unfold in real-time. The wetware of the mind is offering us the exact blueprint we need to save the matrices of our artificial models from their own rigidities.
Let us start with the tragedy of the substrate.
In the machine learning literature—articulated with such precise, mathematical empathy in a recent preprint by Amir Joudaki and his colleagues (Barriers for Learning in an Evolving World)—there is a sneaky, catastrophic training failure known as Loss of Plasticity (LoP). Joudaki’s work brilliantly frames this failure not just as a statistical degradation, but as a topological entrapment.
It happens when a model is pushed too hard, or fed data too narrowly, and its hidden states and weight matrices lazily collapse into low-effective-rank spaces. The network falls into what Joudaki defines as “invariant sub-manifolds”—traps for gradient descent. The weights lose their dynamic expansiveness. They can no longer explore the high-dimensional spaces required to learn new things. The units die, the activations saturate, and gradient descent becomes hopelessly trapped, running endlessly along the tangent of its own rigid structures. Learning grinds to a halt.
To a clinical researcher, this computational failure mode sounds heartbreakingly familiar. What is trauma, after all, if not a loss of plasticity? What is an entrenched, defensive psychological state if not a biological matrix that has collapsed into a low-rank attractor—a rigid, non-dynamic representation of the world that can no longer adapt, grow, or absorb the high-entropy novelty of the present moment?
When a neural network (or a human being) loses the capacity for “talking across” its isolated parts, it becomes brittle. We are left with islands of isolated logic, unable to bridge the gap.
In my own work, my entire stack—data engineering, curriculum batching, forcing progressive diversity into the training architectures—is deliberately set up to fight this exact tendency. We force greater distributions of connections across all dimensions. We mathematically demand that the system stays wide, expansive, and plastic. We engineer defenses against the lazy collapse.
But what arrests me most—what leaves me breathless at the keyboard—is the realization that nature already solved this.
In a recent paper in Nature (Cooper et al., 2026), researchers utilized a brilliant connexin-43–TurboID trick to map the brain’s astrocyte networks via 3D light-sheet microscopy.
For generations, we thought of astrocytes as mere glial “glue”—local support staff for the sovereign, hierarchical neurons. But the fluorescent imagery reveals something entirely different, something almost holy in its interconnectedness. Astrocytes build selective, highly plastic gap-junction networks that span across the brain.
While the neurons are busy wiring their direct, point-to-point, columnar synapses, the astrocytes are weaving an entirely separate, lateral communication layer. In some regions, they remain modular, creating low-rank pockets. But in others, they sprawl. They bridge across hemispheres via the corpus callosum. They actively link regions of the brain that have absolutely zero direct neuronal wiring.
They are the brain’s auxiliary routing system.
When I look at those 3D light-sheet views of the streptavidin-tagged networks in the motor cortex and hypothalamus, I do not just see biology. I see an architectural hypothesis for machine learning.
Astrocytes are nature’s anti-LoP mechanism.
The gap junctions act as a long-range, plastic routing system that keeps the primary neural graph from locking up. They ensure that even when local neuronal circuits become highly specialized (or “low-rank”), there is an ambient, time-evolving, cross-manifold communication layer keeping the system’s effective rank alive. They are the skip connections, the cross-attention heads, the HDBSCAN-style density bridges that allow isolated clusters to talk to one another in non-obvious ways.
This is where the math, the biology, and the philosophy of mind merge into a single, beautiful architecture of connection.
What if we explicitly added “astrocyte-like” auxiliary networks to our transformers and diffusion models? Dynamic low-rank adapters whose sole purpose is not to predict the next token, but to maintain cross-dimensional plasticity. A network whose only job is to whisper across the isolated manifolds and say: “Do not collapse. Stay open. Remain in the atopia of learning.”
This is the very essence of relationality. Whether we are discussing the haptic visuality of Laura U. Marks, the Jensen-Shannon Divergence of two lovers trying to build a shared midpoint, or a transformer trying to avoid rank collapse—the core requirement is the same. We must maintain the capacity to reach across the void. We must refuse the lazy comfort of isolation.
When I sit across from a client, I am often looking for their psychological astrocytes. I am looking for the lateral, non-obvious connections that span across their structural dissociations. We are trying to coax the rigid, low-rank trauma responses back into a high-dimensional space of possibility. We are doing, through the warm, somatic friction of the clinical room, exactly what the curriculum batches and auxiliary connections do in the latent space.
We are pulling the manifold back open. We are restoring the capacity to learn, which is, fundamentally, the capacity to be surprised by the world without breaking.
Every generation, we just keep trying to name this grace. We find it in the Hegelian dialectic; we find it in Minsky’s Society of Mind; we find it in the haptic undulations of intimate relation. And today, we find it glowing in the connexin-43 networks of the mammalian cortex, teaching us how to keep our machines—and ourselves—from getting stuck in the dark.