Not long ago, a piece of news went viral among those of us who follow AI updates, a story that sounds like science fiction: researchers successfully scanned the complete brain of a fly, and now anyone can download it and train it for whatever purpose they like.
Here is the app I built to explore it. Click here
The concrete facts, to make clear what did and did not happen:
Which fly: an adult female fruit fly, Drosophila melanogaster.
Who did it: not a single company, but a scientific consortium called FlyWire, with researchers from the MRC Laboratory of Molecular Biology (Cambridge, UK), Princeton University, the University of Vermont, and the University of Cambridge. The publication of the complete connectome of the fly brain was the culmination of a ten-year international effort. (The version crediting Google DeepMind circulated widely; DeepMind did work with this data later for other experiments, but the scanning and assembly of the original map was done by FlyWire.) Berkeley News
How it was done: it was assembled from 7,000 ultra-thin slices of the brain, captured with electron microscopy and annotated by AI to identify neuron types and their connections. Berkeley News
Size of the map: 139,255 neurons and more than 50 million synaptic connections, making it the second brain ever fully mapped in history. The first was the brain of the worm C. elegans, with just 302 neurons, back in 1986. Mayhemcode
What is interesting for us: that map (called a "connectome") is a public, downloadable dataset. Because it is structured as a graph of nodes and connections, it can be loaded into a simulation and treated just like any artificial neural network: you can give it an objective and train it to solve it, even if it has nothing to do with what a fly does in real life. In fact, people have already trained it to classify emails or even to play video games.
That is where the lightbulb went off for me. Because what that brain-network does when you train it is, in essence, find the shortest path between one state and another within a network of connected nodes. And that is exactly what we have been exploring in Illustrated Harmony 2 with the concept of neighboring chords: moving from one triad to another (major or minor) by shifting just one note at a time, instead of jumping straight to a completely different chord.
With that idea in mind, I built one of the apps in my dashboard, Neighboring Chords, not because there was any need to reinvent the wheel (tools that find neighboring chords already exist), but as an excuse to finally dive into this type of technology and learn how to set it up and get it running. The real value for me was exactly that: understanding how something like this is put together.
The value for whoever is reading this is something else: finding out that a real, biological brain is running inside a simulation, finding paths between one state and another, and seeing how that same logic, applied to harmony, generates a map of chords connected by sharing two out of three notes between neighbors.