Source: A Drosophila Computational Brain Model — Shiu et al., Nature 2024
I enjoy these papers because they pull AI out of the browser window. They remind me that intelligence has always been physical, adaptive, and annoyingly hard to reduce to one neat benchmark.
In October 2024, Nature published what I consider one of the landmark papers of the decade in computational neuroscience: a complete leaky integrate-and-fire model of the entire Drosophila melanogaster central brain. Built from the FlyWire connectome — the most complete map of any brain ever assembled. The paper does something that most computational neuroscience papers stop short of: it makes specific testable predictions and then validates them experimentally. The model predicts which neurons should drive feeding behaviour when sugar-sensing neurons are activated. The researchers then use optogenetics — light-activated tools that can stimulate specific neurons in living flies — to test those predictions. What makes this significant for AI is the implicit proof of concept it provides. This model wasn’t trained on data in the conventional machine learning sense. It was built from structural connectivity information and basic neuroscience principles. The fact that it accurately predicts behaviour demonstrates that intelligence-like function can emerge from structural prior alone — without gradient descent over billions of examples. The obvious question: if a 125,000-neuron model can predict complex behaviour from structure, what would a model of a more complex nervous system, built from its connectome, be capable of?
In plain English, that is why the result matters beyond the chart. It changes where people should look, what they should question, and which comfortable assumption probably needs to be retired.
For me, these studies are a reminder that intelligence is not just answering. It is sensing, moving, adapting, and surviving contact with reality. Reality remains undefeated.