A network brain choosing the
best route across the planet
The Earth below is the physical network — real cities, real great-circle distances, links that follow the actual submarine-cable map, and a LEO satellite constellation overhead. The overlay is the decision layer, drawn as the fruit fly's actual olfactory→mushroom-body circuit: it scores every candidate path by latency, throughput, jitter and loss, then commits traffic to the cheapest one. Pick a city pair, drag the weights, congest the winning route — and watch the brain re-decide, sometimes straight onto the satellites.
Latency = distance ÷ speed of light in the medium (fiber ≈ c/1.468; LEO laser links ≈ vacuum c) + per-hop processing. Best route = argmin of the weighted per-hop cost over your sliders — computed live with Dijkstra + Yen's k-shortest paths. Cities and distances are real; capacities, load, jitter and loss are plausible illustrative figures, not live telemetry.
Physics decides the floor
Light in fiber travels at roughly 204,000 km/s— about two-thirds of vacuum c — and cables never run straight. That's why Toronto→Mumbai can't beat ~80 ms one-way on glass no matter how good your gear is. Every latency figure in the demo is computed from that physics, not typed in.
Why satellites can win
LEO inter-satellite laser links run in vacuum — at full c. On very long routes that ~47% speed advantage can outweigh the up/down hops, so the NTN path genuinely wins when you prioritize latency. The trade is honest too: less capacity, more jitter, and rain fade. Slide the weights and watch the flip happen.
Best-path is a policy, not a fact
There is no single "best" route — only the best for a given intent. A trading firm weighs latency; a backup job weighs throughput; a voice call weighs jitter and loss. That's exactly what SD-WAN and traffic-engineering systems encode as policy. The sliders are that policy; the reroute you see is the math obeying it.
How this relates to the fruit fly
A fruit fly runs its whole life on about 140,000 neurons — and we now have the complete wiring diagram. In its olfactory pathway, ~50 types of receptor neurons (ORNs) pick up a smell, projection neurons relay it inward, and the signal fans out into ~2,000 Kenyon cells per hemisphere. Each Kenyon cell samples ~6 random inputs and only fires when several arrive at once — a sparse, fault-tolerant code. That code then converges onto just 34 mushroom-body output neurons (MBONs) whose balance commits the fly to one action: approach or avoid. Many noisy inputs → sparse robust representation → one committed output. No central controller, and it degrades gracefully when parts fail.
Routes sensory input (smells) through a sparse coincidence layer and commits to a single action at the MBON readout — hundreds of times a second, on milliwatts, with no neuron in charge of the whole decision.
Routes candidate paths (with noisy latency/throughput/loss metrics) through a scoring function and commits traffic to a single winner — recomputing the instant conditions change. Same shape of problem: many inputs, one committed choice, survive failures.
That's exactly how the decision-cortex overlay above is built: its neurons are the ORN → PN → Kenyon-cell → MBON architecture from my fly-brain interactive. Candidate routes enter as input channels, pass through the sparse Kenyon layer, and the MBON corresponding to the route the algorithm actually chose is the one that fires. To be plain about it: the connectome data, the Neuroglancer viewer below and the routing math are real; mapping routes onto that circuit is my analogy. No fly routes packets — but the architecture it uses to commit to one action out of many is the same shape engineers keep reinventing.
Real data · not a drawingThe real connectome behind the metaphor
This panel streams the actual FAFB electron-microscopy volume of a fruit-fly brain — the raw imagery the FlyWire connectome (Dorkenwald et al., Nature, 2024) was traced from — into Neuroglancer, Google's open-source connectomics viewer. Somewhere in this tissue are the very ORNs, Kenyon cells and MBONs the overlay above is modeled on.
This panel is not a drawing — it is the real connectome substrate rendered in Neuroglancer, Google's open-source WebGL viewer for petabyte-scale connectomics. Scroll to move through the tissue; drag to pan.
Open full-screen ↗What's real here — and what's illustrative
- RealThe geography and physics: city coordinates, great-circle distances (haversine), speed of light in fiber (~c/1.468) vs. vacuum-c laser links, and per-hop processing delays. The 11-PoP topology follows real submarine-cable and backbone corridors.
- RealThe routing math: candidates come from Yen's k-shortest-paths algorithm over Dijkstra, and the winner is the argmin of a weighted per-hop cost built from your sliders. Nothing is scripted — congest a link and the decision genuinely recomputes.
- IllustrativeThe numbers on the links: capacities, utilization, jitter and loss are plausible hand-set figures sized for a legible demo — not telemetry from any production network or vendor.
- RealThe connectome: the Neuroglancer panel streams the actual FAFB fly-brain EM volume, and the decision-cortex overlay uses the real architecture of the fly's olfactory→mushroom-body pathway — ORN → PN → sparse Kenyon cells → MBON, with the divergence, random ~6-input sampling and convergence the connectome papers describe (drawn with reduced node counts so a browser can run it at 60fps).
- AnalogyThe mapping: assigning each candidate route to an input channel and reading the chosen route out at an MBON is my illustration, borrowed from my fly-brain demo. No fly is running the internet, and no neural circuit computes this decision — Dijkstra + Yen's does, and the fly wiring is how I show it committing. The winning MBON always agrees with the algorithm because it is driven by the algorithm.
Connectome panel & open tools: the FAFB fly-brain EM volume in Neuroglancer; FlyWire whole-brain connectome (Dorkenwald et al., Nature, 2024) · nature.com.
This page routes between cities; that one routes inside the network — RU/DU/CU, backhaul, UPF, CDN — with faults you can localize and fix.
The interactive the decision cortex is built from — the same ORN→PN→Kenyon→MBON circuit at full size. Watch a smell become a memory, neuron by neuron.
My consultancy applies exactly this decision layer to production telco, satellite/NTN and enterprise networks.
Twenty years in telecom taught me that routing is where physics meets policy. I wrote about the decision layer in what a fly brain teaches telecom networks.