San Francisco startup Eon Systems claims that it has created the first digital simulation of a fruit fly brain that can control a virtual body and produce recognizable behaviors.
We'd call this "huge if true" but the research papers behind it look solid. This is science-fictional stuff – but not the SF you might think.
The "first multi-behavior brain upload" is based on several substantial pieces of previous work, including some hefty scientific research papers in reputable journals. Otherwise, we would have dismissed it. On first look, the story sets several of our alarm bells ringing (brain uploading, AI, virtual reality, the Singularity).
However, on the face of it, it seems to hang together. The researchers at Eon Systems have taken several pre-existing components: a fruit fly brain scan, a tool for modelling neurons, a model of some of the fly's muscles and body, and a very simple virtual environment, connected them together and ran it. The team claims that the result displays some of the behavior of the real insect.
One part is a whole-brain connectome of a fruit fly, Drosophila melanogaster – one of the classic laboratory organisms, and a species this vulture spent hours peering through binocular microscopes at university in the 1980s (back then, he could separate male from female anesthetized flies in a dish at dozens to hundreds per hour). Effectively, this is a 3D brain scan of a fly's brain, including 125,000 neurons and 50 million synapses.
The scan is from an adult female fly. This is important because one of the reasons Drosophila is a model organism is that some of the parts of their larvae are unusually large and so easier to work with. Relatively speaking, maggots don't do much, so a maggot brain scan would be easier – but no, this was an adult fly.
The connectome comes from the splendidly named Flywire effort. That was published in 2024, and the same year, a team demonstrated that it seemed to respond to some signals, described in Nature in the paper "A Drosophila computational brain model reveals sensorimotor processing."
Eon Systems ran the brain scan in a spiking neural network simulator called Brian2, and no, that is not a typo. Brian2 is FOSS and the source code is on GitHub. Eon also drew upon the NeuroMechFly v2 model, as described in the paper "Simulating embodied sensorimotor control in adult Drosophila."
So, it took a synapse-level fly brain scan. It linked the scan's 125,000 neurons back up to one another using an AI tool to reconnect its 50 million synapses. Then it ran the resulting dataset inside an existing tool for modeling networks of neurons. It connected this digital model fly brain up to another model representing some of a fruit fly's body: its legs, antennae, mouth parts, and so on. The resulting model was fed some inputs provided by the MuJoCo advanced physics simulation.
A video accompanying Dr Alex Wissner-Gross's Substack article about the work shows the model fly moving. In summary, the model fly brain, controlling a model fly body, managed to walk around, stop and clean its antennae, and when the model was presented with a signal resembling the smell of sugar, it stopped and extended its proboscis. In other words, it stuck out its tongue for a lick.
We took the story to someone with considerable knowledge of running models of neurons in silico, Dr Steve Furber, co-creator of the ARM processor whose subsequent research included creating the SpiNNaker 2 hardware. He told us:
I've come across this work before, and I think it's pretty impressive! I was hoping to get a mouse brain model running on SpiNNaker to connect to a virtual mouse embodiment as part of the work in the Human Brain project, but finding a suitable mouse brain model proved too difficult, so we never got there. The million-core SpiNNaker machine is at least in principle capable of running a whole mouse brain model. Getting this running convincingly for a fruit fly is a really good first step, and I applaud it!
Now, to be fair, at least one of the comments on the Substack post is severely critical:
The simulated brain is not operating as a fly's brain does. It's machine learning, clamped into the shape of a fly connectome, that has a resting state of 0 Hz, being zapped with simple inputs, not virtual sensory data.
We specifically asked Dr Furber about that, and he said:
I think it does weaken some of the claims made for the work.
Is Eon Systems' work an uploaded fly, waking up in the fruit-fly Matrix? No. But it does demonstrate interconnecting a modeled fly brain with a modeled fly body and getting fly-like behaviour out of it, and that is pretty impressive.
Dr Wissner-Gross has been writing and presenting about AI and related matters for some 15 years, so he is clearly highly motivated to spin this for all it's worth, but even so, it looks like an impressive result to us.
There are a number of things left unstated about the results so far. One of them is how long it took. This is going to be an absolutely massive model, which means running it will be quite slow. We're absolutely certain that the model and the video are not running in real time. It could well be that to get a few seconds of fly-like behavior took weeks or months of computer time. We are confident that there won't be brain scans of flying insects acting as onboard drone pilots in the next couple of years.
We're sure that researchers around the world will be scurrying to replicate this, and indeed, to find any problems and highlight them. Among other things, though, it might result in easier access to funding. This is the sort of demo that you can wave in front of an uncomprehending administrator to good effect.
The resonances from science fiction are inescapable, but they might not be as positive as you might immediately think.
For instance, you might be excited that this could be a measurable, if tiny, step towards brain uploading. Be careful what you wish for, and read our favorite thought experiment about human brain uploads in recent years, Lena. It's a superb short story presented in the form of a fake Wikipedia article, and it is a chilling read. (The title is taken from one of the reference images used in digital image processing. The eponymous model herself now wishes the image was retired.)
For a much longer version, there's Neal Stephenson's recent novel Fall; or, Dodge in Hell. This has the approximate shape of a happy ending if you squint, but the real subtext is right there in the title.
What it directly and closely reminded us of, though, is no coincidence because the author is the person who shared the research with us. It was a short story called Lobsters by Charles Stross. This later became part of his 2005 novel Accelerando, which tells the story of humanity undergoing a technological singularity. A lot of folks in the tech world enjoyed this book hugely (it has 22,431 ratings and 1,448 reviews on Goodreads), without noticing the gigantic genocidal atrocity happening in the background – which should tell you important things about tech bros.
For us, the takeaway was that, as the Russian invasion of Ukraine drags into its 12th year, the Gaza war is nearly 30 months in, and the new 2026 Iran war continues to send oil prices skyward along with the smoke, we are rolling into William Gibson's "Jackpot" scenario:
There are no comets crashing, nothing you could really call a nuclear war. Just everything else, tangled in the changing climate: droughts, water shortages, crop failures, honeybees gone like they almost were now, collapse of other keystone species, every last alpha predator gone, antibiotics doing even less than they already did, diseases that were never quite the one big pandemic but big enough to be historic events in themselves. And all of it around people: how people were, how many of them there were, how they'd changed things just by being there.
Through the 21st century, it kills 80 percent of humanity, and most other life on Earth. As it happens, it is accompanied by a period of vast scientific and technological advance. ®
Source: The register