There’s a video going around from a lab called Eon Systems. A fruit fly walks across a flat virtual floor toward a few slices of banana. Nothing special about that, flies do this every day in every kitchen. Except this fly doesn’t exist. Its body is a physics simulation. And the thing controlling that body isn’t an AI that was trained to act fly-like. It’s an actual copy of a real fly’s brain, all 139,255 neurons of it, wired exactly the way a real fly’s neurons are wired, sitting on a server and deciding, on its own, to walk toward the food.
I spent an evening trying to figure out if this was being oversold, because “we simulated a brain” is the kind of headline that gets stretched past what actually happened. It isn’t, mostly. What’s really going on is stranger and more careful than the headline suggests, and also a lot less finished than it looks in the demo clip.
What they actually built
The starting point is something called the FlyWire connectome, finished in October 2024 by a big international group that included Princeton and a team at the MRC Laboratory of Molecular Biology in Cambridge. A connectome is the full wiring diagram of a brain, literally every neuron and every synapse between them, mapped by slicing a fly brain into thousands of layers and reconstructing it in 3D. For the fruit fly, that turned out to be 139,255 neurons and something over 50 million synaptic connections. It’s only the second brain ever mapped this completely. The first was the roundworm C. elegans, 302 neurons, done back in 1986 by a team including Sydney Brenner. Three hundred neurons to almost a hundred and forty thousand is not a small jump.
A former UC Berkeley postdoc named Phil Shiu, working in Kristin Scott’s lab, took that wiring diagram and simulated it. Not trained a model on fly behavior. Simulated the actual circuit, neuron by neuron, on a laptop. His paper came out in Nature on the same day as the connectome itself, and the result was that the simulation predicted which real neurons would fire when you stimulated a fly’s taste or touch sensors, and it got this right often enough that other researchers started using his model to predict things they hadn’t tested yet. One of them, Salil Bidaye, used it to predict locomotion behavior before checking it against real flies. It held up.
That’s the part that’s genuinely new here. Historically, if you wanted a computer program to act like a fly, you trained it. You built an artificial neural network, gave it a reward signal, and let it learn through trial and error until it walked and turned and avoided obstacles the way a fly does. DeepMind and a Janelia team did exactly this in 2025, and it worked, their simulated fly walked convincingly. But that network’s internal wiring has nothing to do with an actual fly’s brain. It’s a different structure that arrived at similar behavior. What Shiu did, and what Eon Systems has now taken further, uses the real wiring. The neurons are placed where the real ones are placed, connected how the real ones connect. If it behaves like a fly, it’s not because it learned to. It’s because, structurally, it more or less is one.
Closing the loop
Shiu’s 2024 work only went one direction, though: stimulus in, predicted neural activity out. It never had a body to move. That’s the piece Eon Systems added this year. Their team, which includes Shiu himself along with Scott Harris, Aarav Sinha, Viktor Toth, and Alexis Pomares, hooked the same connectome up to a physics-simulated fly body and let sensory input flow all the way through to motor output. Taste receptors get activated by something sweet, activity propagates through the real wiring, motor commands come out the other end, and a simulated body executes them. First time anyone’s closed that full loop with a whole biological brain, not a trained stand-in for one.
They reported the model predicted the fly’s motor behavior with around 95 percent accuracy, which sounds close to perfect until you sit with what that number is actually measuring. It’s not “the fly does everything a real fly would do.” It’s closer to “in the specific behaviors we tested, output matched real fly data 95 percent of the time.” Flight is a good example of where this gets messier. A separate 2025 project by Vaxenburg’s team built an extremely detailed physical fly body, one that could stick to surfaces the way real flies do and handle the fluid dynamics of actual flight, but they drove it with a trained network, not a pure connectome model, because getting flight to emerge straight from the wiring diagram is still unsolved. So the “whole fly in a computer” story is true for walking toward food. It’s not yet true, at least not fully, for taking off and flying there.
I’ll admit I found this genuinely confusing to untangle at first, because different papers are solving different pieces of the same puzzle and the press coverage tends to flatten them into one clean narrative. There isn’t one single “digital fly.” There’s a connectome, a taste-and-touch model built on top of it, a separate body model with its own physics, and now an integration project stitching pieces together, some of them connectome-driven and some of them still trained rather than wired. The banana video is real. It’s just not the whole animal yet.
Eon’s own writeup is pretty upfront about this, to their credit. They describe the current system as an integration effort built on other people’s connectome work, other people’s body models, decades of separate circuit mapping, rather than claiming they invented digital life from scratch. That’s a more honest framing than the “brain simulation that acts alive” headlines some outlets ran with.
Why a mouse is a much bigger problem than a fly
Eon has said their next target is a mouse brain, and eventually, “human-scale emulation.” It’s worth actually sitting with the numbers instead of skipping past them. A fruit fly brain has about 139,000 neurons. A mouse brain has around 70 million. A human brain has roughly 86 billion. Going from fly to mouse is close to a 500x jump in neuron count. Going from mouse to human is over a thousand times that again. And neuron count isn’t even the hard part, mapping every synapse at that scale, the way FlyWire did for the fly, is a project that took a global consortium years for an organism with a brain smaller than a poppy seed.
So no, nobody is uploading a human mind next year, or probably next decade. But the direction of travel matters more than the current position does. Ten years ago a full-synapse connectome of anything bigger than a worm was considered basically impossible on any reasonable timeline. Now it’s been done for a fly, and a mouse connectome effort is already underway at multiple labs. The gap between “we mapped it” and “we can run a predictive simulation of it” used to be assumed to be decades. For the fly it turned out to be a few months, Shiu’s paper landed the same day as the connectome.
The Matrix question, honestly
People have been asking, half-joking and half-not, whether we’re living in a simulation for a long time, at least since the Wachowskis put it on screen in 1999, and philosophers were arguing versions of it long before that. What’s changed isn’t that anyone’s answered the question. It’s that we’re now demonstrably capable of building the kind of thing the question is about, at least at fly scale. A digital mind that takes in sensory input, processes it through something structurally equivalent to a real nervous system, and acts on the world because of it. That’s not a metaphor anymore for a fly. It’s a working pipeline with a paper trail and a GitHub repo.
Here’s where I’ll actually take a side instead of hedging: I don’t think this gets us anywhere close to human consciousness questions, and I think a lot of the online reaction to this stuff jumps straight past the biology to the philosophy because the philosophy is more fun to argue about. A simulated fly walking toward a banana slice tells us something real and useful about how nervous systems generate behavior. It does not tell us anything about whether that fly, or a future simulated mouse, or eventually a future simulated human, would have any kind of inner experience.
Nobody knows how to even test for that, in a biological fly, let alone a digital one. The engineering problem and the consciousness problem are not the same problem, and treating a 95-percent motor-prediction accuracy number as evidence for or against subjective experience is a mistake I’ve seen made in more than one comment section already.
What I do think is worth taking seriously is the practical trajectory. If a mouse connectome and simulation happen within the next several years, and there’s real momentum suggesting they might, we’ll be one step from the thing that actually matters for medicine: using a working digital brain to test what happens when you perturb specific circuits, the way you’d test a drug, except you can run the experiment a thousand times overnight instead of breeding a thousand mice. Shiu said something close to this himself when the connectome first came out, that the value isn’t in the philosophy, it’s in what a simulated brain lets you learn about actual brain disorders without needing a live animal for every single test.
That’s a less dramatic ending than “we might be in the Matrix,” I know. But it’s the honest one. The bigger, stranger question, whether a good enough simulation of a nervous system produces something that experiences anything at all, is still completely open, and I don’t think a banana-seeking fly on a server answers it either way. It just proves the wiring works. What runs on that wiring, and whether anything is home when it runs, is still nobody’s to claim.