Lab Newsletter — August 23, 2026: Wiring the Brain

A connectome — the complete map of every neuron and every connection in a brain — used to sit in the same drawer as science fiction. The surprising part is why it was hard. The imaging was largely solved: electron microscopy can slice a brain into nanometre-thin sections and photograph every synapse. The wall was reconstruction — tracing each hair-thin neurite as it weaves through billions of image voxels, without ever accidentally merging two neurons or dropping one. That is an image-analysis problem at a scale no human can finish by hand, and it is exactly the kind of problem deep learning was made to break. In 2024, it broke.

🧠 The reconstruction was the wall — and AI was the unlock

The turning point was teaching a network to follow a neuron. Flood-filling networks (Januszewski et al., Nature Methods, 2018; senior author Viren Jain) pair a convolutional network with a recurrent pathway that starts inside one neurite and iteratively “floods” outward along it, segmenting and extending as it goes. On serial block-face EM of a zebra finch brain it reached a mean error-free neurite path length of 1.1 mm, with only four mergers across 97 mm of traced path — “an order of magnitude better” than the methods before it (at, the authors note, substantially higher compute). Tracing stopped being the bottleneck and became the engine.

🪰 The first whole brain — and the human scale arriving

Then came the map everyone said couldn’t be built. FlyWire (Dorkenwald et al., Nature, 2024; co-senior authors Sebastian Seung and Mala Murthy) published the neuronal wiring diagram of an adult brain — the fruit fly Drosophila — with 139,255 neurons and about 50 million (5 × 10⁷) chemical synapses: the first complete connectome of an adult animal brain. A companion paper (Schlegel et al., Nature, 2024; senior author Gregory Jefferis) turned that raw graph into biology, annotating 8,453 cell types across the ~140,000-neuron connectome. And the human scale is already coming into view: H01 (Shapson-Coe et al., Science, 2024; senior authors Jeff Lichtman and Viren Jain) reconstructed roughly one cubic millimeter of human temporal cortex — about 57,000 cells, about 150 million synapses, 1.4 petabytes of data. One cubic millimeter, and already a landmark; a whole human brain is a million times larger. The trajectory is unmistakable.

🧭 The honest frontier — and why it’s our kind of problem

Two caveats keep this grounded, and both are the lab’s native language. First, automation isn’t finished. The same team is blunt that even with the best networks, proofreading whole-brain reconstructions “will require many person-years of effort, due to the huge volumes of data involved” (Nature Methods, 2022) — connectomics is the definitive human-in-the-loop-at-planetary-scale enterprise, AI doing the impossible bulk and people catching what it gets wrong. Second, a wiring diagram is not function. As Bargmann & Marder argued (Nature Methods, 2013), understanding a nervous system takes more than connectivity — you need “neuronal dynamics and neuromodulation,” the electrical and chemical life the static map leaves out. The connectome is necessary, not sufficient.

Here’s why it lands for us. Connectomics is the most extreme instance of the exact problem the lab is built around: AI reading images at a scale no human can, served openly so anyone can use — and check — the result. Petascale EM segmentation is the same engineering as making large foundation models callable over enormous image volumes on shared infrastructure (BioEngine); FlyWire’s open, community-proofread reconstruction is the same open, reproducible ethos our platforms are built on; and the map-versus-function gap is the prove-it discipline once more — a reconstructed synapse is a hypothesis until physiology agrees. It even rhymes with the virtual cell we keep circling: a wiring diagram is a structural prior for a dynamic model, just as a static protein structure is for its conformational ensemble. Reading biology’s images at impossible scale, keeping the machine honest, and turning a static map into a living model — that’s the whole job, written very, very large.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits; a Grok-based replacement is wired and awaiting credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.