For most of neuroscience’s history, the brain’s story had two characters: neurons, which did the thinking, and everything else, which held the neurons in place. The “everything else” — collectively called glia, from the Greek word for glue — was assigned a supporting role so thoroughly uninteresting that generations of neuroscience students were taught to skip past it. Astrocytes provided structural scaffolding. Oligodendrocytes insulated axons with myelin. Microglia cleaned up cellular debris. The real work — the computation, the signaling, the information processing that produced thought, memory, behavior, and consciousness — was done by neurons. The glia were the roadies. The neurons were the band.
A single human astrocyte contacts between 270,000 and 2 million synapses. There are approximately as many glial cells in the human brain as neurons — roughly 85 billion of each, depending on the counting method. Astrocytes communicate with each other through gap junctions, forming a continuous syncytium — an electrically and chemically interconnected network that spans entire brain regions. They generate calcium waves that propagate through the syncytium at speeds of 15-25 micrometers per second — slower than neuronal signaling by a factor of a thousand, but covering spatial scales that individual neurons cannot. They release gliotransmitters — glutamate, ATP, D-serine, GABA — that modulate synaptic transmission at the synapses they contact. They detect neuronal activity in real time, integrate signals across thousands of synapses simultaneously, and adjust synaptic strength in response. In 2023, a team of researchers at the University of Pennsylvania published a paper in PNAS demonstrating that neuron-astrocyte networks can perform the core computations of a Transformer — the architecture underlying ChatGPT, Claude, and every large language model currently operating. The paper was not a metaphor. It was a mathematical proof that the tripartite synapse — the three-way connection between a presynaptic neuron, a postsynaptic neuron, and an astrocyte — can implement the normalization step in the self-attention operation that makes Transformer models work. Half the cells in your brain may be running computations that neuroscience has spent 150 years ignoring.
The tripartite synapse
The conceptual revolution began with the tripartite synapse — a term introduced in 1999 by Alfonso Araque and colleagues to describe the functional unit that replaced the classical two-neuron synapse in the emerging understanding of how brain signaling actually works. The classical synapse had two partners: a presynaptic neuron that releases neurotransmitter and a postsynaptic neuron that receives it. The tripartite synapse has three: the same two neurons, plus an astrocyte whose fine processes wrap around the synaptic cleft, detect the neurotransmitter release, respond with intracellular calcium elevation, and — critically — release gliotransmitters that feed back onto both neurons, modifying the strength, timing, and probability of future synaptic transmission.
The astrocyte is not passively recording what the neurons do. It is actively modulating it. When an astrocyte detects glutamate released at a synapse, calcium ions are released from internal stores through the IP3 pathway, generating a calcium transient that can remain localized to a single astrocytic process (a “microdomain” event spanning a few micrometers) or propagate across the entire cell and, through gap junctions, into neighboring astrocytes. The spatial scale of the response is graded: a weak synaptic input produces a local microdomain calcium blip, a stronger input produces a cell-wide calcium wave, and a sustained barrage of inputs produces an intercellular calcium wave that rolls through the astrocytic syncytium like a slow-motion neural tide. The astrocyte integrates synaptic activity across thousands of synapses and, through its calcium dynamics, generates a response that reflects the aggregate state of the local network — not the activity of any single synapse, but the pattern of activity across all of them.
This integration happens on timescales that neurons don’t operate on. Neuronal signaling runs at milliseconds. Astrocytic calcium dynamics run at seconds to minutes. The two systems are processing the same synaptic events on different temporal scales — the neurons handling moment-to-moment signaling, the astrocytes handling the slower, contextual modulation that adjusts how the neuronal network operates over time. The relationship between the two is not parallel processing in the usual sense. It is nested processing: the fast system (neurons) generates the signals, and the slow system (astrocytes) tunes the fast system based on a broader, slower integration of those signals. The analogy — imperfect but useful — is a mixing board at a concert. The musicians play their instruments at performance tempo. The sound engineer adjusts levels, EQ, and effects on a slower timescale, shaping the overall sound without playing any notes. The astrocytes are the sound engineers of the brain.
What astrocytes compute
The theoretical question — what are astrocytes actually computing? — moved from speculation to empirical traction in 2021 when two studies demonstrated direct astrocytic involvement in behavioral computation.
Mu and colleagues showed that astrocytes in the zebrafish brainstem directly integrate sensory signals and control motor output — specifically, astrocytic calcium activity predicted and causally influenced the zebrafish’s swimming behavior. Removing astrocytic signaling impaired the fish’s ability to coordinate its movements. The astrocytes were not just modulating neuronal circuits. They were computing part of the motor output.
Slezak and colleagues showed that astrocytes in the mouse visual cortex integrate visual information and behavioral state — simultaneously encoding what the mouse is seeing and whether the mouse is running or stationary. The calcium signals in visual cortex astrocytes carried information about both the visual stimulus and the animal’s locomotor state, combining two streams of information that arrive through separate neural pathways. The astrocytes were performing multisensory integration — the combination of signals from different sources into a unified representation — independently of the neuronal circuits operating in the same cortical region.
These findings joined a growing body of evidence that astrocytes are involved in learning and memory. Hippocampal astrocytes modulate long-term potentiation — the cellular mechanism widely believed to underlie memory formation — through the release of D-serine, a co-agonist of the NMDA receptor. Blocking astrocytic D-serine release impairs LTP and impairs spatial memory in mice. Astrocytic ensembles — coordinated populations of astrocytes that activate together — have been observed during memory encoding and recall, with the ensemble patterns being specific to particular memories. The memory without a brain post documented memory in organisms with no neurons at all. Glial networks suggest that even in organisms with neurons, a significant portion of the memory computation may be running on non-neuronal hardware.
The speed-scale tradeoff
The most important structural insight about glial computation is the speed-scale tradeoff it creates. Neurons are fast and local — a single action potential takes a millisecond, travels along one axon, and activates one set of synapses. Astrocytes are slow and distributed — a single calcium wave takes seconds, propagates through gap junctions across hundreds of micrometers, and modulates thousands of synapses simultaneously. The two systems together create a dual-timescale architecture: neurons handle the fast, precise, point-to-point signaling that produces moment-to-moment behavior, and astrocytes handle the slow, distributed, contextual modulation that shapes how the fast system operates.
The 2023 PNAS Transformer paper formalized this intuition. In a Transformer architecture, the self-attention mechanism computes how much each element of an input sequence should attend to every other element — a global integration step that requires normalizing across all possible attention weights simultaneously. The paper showed that the tripartite synapse can perform this normalization: the astrocyte, by integrating signals from thousands of synapses and feeding back a modulatory signal that depends on the aggregate, implements the mathematical operation that Transformers use to compute attention. The claim is not that the brain is a Transformer. The claim is that the biological hardware — specifically the neuron-astrocyte interaction — has the right computational properties to implement the kind of global integration that Transformer models perform, and that neuroscience has been modeling the brain as a purely neuronal network while ignoring the cells that may be performing the global integration step.
The comparative angle
The glial-to-neuron ratio varies dramatically across species, and the variation maps onto cognitive complexity in ways that neuron counts alone do not explain. The comparative cortices post demonstrated that the dolphin’s cortex — larger in surface area than the human’s — contains fewer cortical neurons but substantially more glial cells. Whether those glia are performing computations that compensate for the lower neuron count is the open question that could reframe the entire dolphin intelligence debate.
Invertebrates have glia too. The Drosophila brain — 100,000 neurons — contains approximately 10,000 glial cells that regulate synaptic transmission, maintain the blood-brain barrier, and respond to injury. Even the slime mold Physarum — which has no neurons and no glia — stores information in the physical architecture of its tube network, using a mechanism (tube diameter as memory) that is functionally analogous to the way astrocytes modulate synaptic strength through calcium-dependent feedback. The parallel is not coincidental. It suggests that the computational operation — integrating past experience into the physical substrate of the network to modulate future behavior — is a general principle of biological information processing that glia implement in one way, neurons implement in another, and slime molds implement in a third.
The brain-body co-evolution post argued that brains evolve in response to the demands of the body. The glial story adds a layer: within the brain itself, two cellular populations — neurons and glia — co-evolved to handle different aspects of computation, with neurons specializing in fast signaling and glia specializing in slow integration. The swarm intelligence post documented computation distributed across thousands of bodies. Glial networks document computation distributed across thousands of cells within a single brain — a swarm system running inside the organ that neuroscience has spent 150 years studying as if only one cell type mattered.
Why it matters for the course
Glial networks are the Neurozoology lecture that challenges the most fundamental assumption in neuroscience: that neural activity is brain activity. Neurons fire. Glia modulate. The modulation is computation. The computation runs on timescales and spatial scales that neuronal recording techniques — which are optimized for millisecond-resolution electrical signals — are poorly equipped to detect. An entire parallel processing system has been operating in every brain ever studied, and the field is only now building the calcium imaging tools, optogenetic manipulations, and computational models required to understand what it’s doing.
The Umwelt concept established that every animal lives in a perceptual world defined by its sensory hardware. The glial story suggests that every brain operates in a computational world defined by its cellular composition — and that the neurons and the glia are processing the same information on different timescales, in different spatial domains, using different signaling mechanisms, to produce an integrated output that neither system could generate alone. The neurons are the instruments. The glia are the mixing board. The music is what happens when both play together. And for 150 years, neuroscience has been transcribing only the instruments and wondering why the score sounded incomplete.
This is the kind of question our Neurozoology course was built to explore — where half the cells in every brain on Earth were dismissed as glue for a century, a single human astrocyte contacts up to 2 million synapses, the tripartite synapse can implement the core computation of a Transformer, and the most unsettling implication is that the parallel processing system neuroscience has been ignoring may be performing exactly the kind of slow, global, contextual integration that no one could find in the neurons alone.