The complete FlyWire connectome of an adult fruit fly (Drosophila melanogaster), mapping all 139,255 neurons and over 50 million synaptic connections across 7,000 electron microscopy slices. Credit: Tyler Sloan for FlyWire, Princeton University.
Tech giants are burning billions of dollars on data centers in pursuit of AGI. LLMs are training on all the text garbage on the internet just to get slightly smarter. Meanwhile, a regular fruit fly, digitized by enthusiasts, is landing headshots in Counter-Strike 2 and trading crypto.
Let’s break down how a bug’s brain became the biggest AI trend, what it’s already been taught, and why it actually works.
What actually happened and how they digitized the brain
Scientists (the FlyWire consortium, and later Google Research) mapped the complete connectome of a fruit fly’s brain: a wiring diagram of all its neurons and synapses. An actual fly’s brain was sliced into thousands of 40-nanometer-thick layers, scanned under an electron microscope, and assembled into a 3D model. The result is a network of ~139,000 to 166,000 neurons and tens of millions of connections.
But it is not just a 3D picture. This schematic was ported into a software simulator. Turns out, if you attach a virtual body to this digital brain, the fly starts moving. It walks, looks for food, and grooms its legs. No datasets, no machine learning, and no thousands of hours of training. The biological wiring architecture works right “out of the box.”
The 78 functional regions (neuropils) in the fruit fly brain responsible for flight, vision, and olfactory processing.
The 50 largest projection neurons spanning across both optical lobes and sensory navigation centers.
What the digital fly has already learned
As soon as the connectome database went open-source, indie devs got their hands on it. It turns out this perfect natural neural network can be adapted for abstract human tasks.
Here is an incomplete list of what the digital fly is doing right now:
Esports: Through open initiatives such as DOOMFLY, the fly plays Counter-Strike 2 (getting sniper rifle frags with minimal latency), beats DOOM, slashes blocks in Beat Saber, and survives in Minecraft.
Trading: The Stonkfly bot converts stock quotes into neural impulses, feeds them into the fly’s brain, and makes crypto buy/sell decisions based on the virtual neurons’ reactions.
Reading documents (OCR): The FlyOCR project taught connectome fragments to recognize text in PDF documents and sort emails into spam and work folders.
Hardware control: The simulation is hooked up to drones, forced to park virtual cars, and in the Desktop-Fly environment, functions as an embodied desktop companion navigating window borders and live cursor movements.
Web surfing: Through projects like Fruitfly-Doomscroller, fly neurons are mapped to control browser tabs and the TikTok feed (which is where the viral videos about these experiments actually came from).
Architecture of the DOOMFLY project: 3,335 brightness inputs and 811 color receptors stream directly into 166,700 neurons of the MaleCNS connectome to govern combat steering and weapon triggers.
Why it works and what the catch is
The fly has no idea it’s playing CS2 and doesn’t possess a gamer’s “intellect.” The secret to success lies in the concept of Reservoir Computing.
A developer takes the connectome and uses it as a complex non-linear filter (the reservoir).
Screen pixels or market prices are converted into impulses and fed to the fly’s visual sensors.
The signal runs through hundreds of thousands of natural connections, delays, and loops, transforming at each stage.
At the output (motor neurons), the developer hooks up a simple classifier script that decides: “If pattern A lights up in this zone, pull the trigger; if pattern B, sell Bitcoin.”
Closed-loop sensorimotor framework in NeuroMechFly: environmental sensory stimuli feed into spiking neural networks, which drive biomechanical joint actuations and feedback loops.
Enthusiasts adapt the fly’s reactions via reinforcement learning, programmatically stimulating its dopamine nodes (reward centers) for the right actions. We aren’t teaching the network to think: we’re exploiting its phenomenal signal-processing capabilities, honed by millions of years of evolution.
Why the fruit fly is winning the AI news race
It is both absurd and fascinating. While corporations try to recreate intelligence through brute compute power, the open fruit fly connectome proves that massive potential lies in the right wiring architecture. The fly is doing what usually requires heavy, resource-intensive neural networks. This is a shift from abstract math to applied digital biology.
How to launch and train your own fly
You don’t need a secret lab for this. The tools are open, and anyone can replicate the experiments:
Where to get a brain: You can download the connectome graph weights and topology via the FlyWire Codex platform or through public datasets hosted by HHMI Janelia FlyEM.
Simulation environment: For biomechanical tests, NeuroMechFly (FlyGym) is perfect: a physical fly simulation available on GitHub from the EPFL lab.
Tech stack: You’ll need Python and a physics engine (MuJoCo or PyBullet) to integrate the spiking neural network.
Your own project: You’ll need to write two bridges: an encoder (turning your data into biological impulses at the input) and a decoder (reading the activity at the output).
NeuroMechFly v2 embodied sensorimotor simulation in MuJoCo: walking across complex terrain, visual navigation, and limb coordination. Developed by NeLy-EPFL.
Embodied fruit fly digital twin modeled with physical exoskeleton dynamics in MuJoCo.
Degrees of freedom across leg joints, tarsal adhesion pads, and proprioceptive sensors.
Instead of waiting for yet another startup to roll out a new neural network on a paid subscription, today any dev can fork a bug’s brain and make it solve their own tasks.
