Google DeepMind mapped a fruit fly's brain into a downloadable dataset. Here's how hobbyists are training it to sort email and play games.
What is the fruit fly connectome, exactly? #
The fruit fly connectome is a publicly released map of a Drosophila brain and nerve cord, produced by researchers connected to Google DeepMind. It documents roughly 166,000 neurons and more than 125 million connections between them, essentially a wiring diagram of a real biological brain. The dataset is downloadable, and because it’s structured like a network graph, it can be loaded into a simulation and treated much like any other neural network, including one you can train to perform tasks.
That last part is what turned a niche neuroscience release into an internet hobby project. Within days of it going public, people were running the simulated fly brain through video games, letting it trade stocks, and pointing it at coding tasks. None of these are things a fruit fly brain was built to do. They’re demonstrations of what happens when you take a real, mapped biological neural network and repurpose it the way you’d repurpose an artificial one.
TL;DR #
- The fruit fly connectome is a real map of a Drosophila brain, roughly 166,000 neurons and 125 million-plus connections, released as a public dataset by Google DeepMind researchers.
- Hobbyists are running this connectome inside simulations and training it on new tasks , from email sorting to trading to playing games, using large language models to handle the data prep and training pipeline.
- One creator trained the simulated fly brain as an email classifier by feeding it hundreds of labeled email examples across a handful of categories and mapping which “brain regions” activated for each label.
- The system worked by pattern-matching character sequences in emails to pre-written template replies , not by generating original text, and topped out around80% accuracy compared to a small conventional neural network.
- The motor circuitry controlling the simulated fly’s movement was kept separate from the part trained on email data, because mixing the two risked breaking the brain’s basic navigation behavior.
- Task complexity has a ceiling : attempts to scale past roughly five to ten categories saw performance break down, since a fly brain is orders of magnitude simpler than even a small purpose-built model.
- The project raises real ethical questions about whether a simulated biological brain can sense or experience anything, questions nobody can currently answer with confidence.
#
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
How does training a fruit fly brain actually work? #
The process looks a lot like training any machine learning classifier, except the underlying network isn’t an artificial one designed from scratch. It’s a fixed biological structure, so the “training” is really about finding which parts of that existing structure respond to which inputs.
In the case of email sorting, the workflow went roughly like this:
- Collect labeled data. A batch of past emails was hand-labeled into a small number of categories, things like sponsorships, consulting inquiries, student support, scheduling requests, newsletters, and receipts.
- Convert text into neural activity. Since an email is just a sequence of characters, and certain character patterns correlate with certain categories (a receipt is more likely to contain numbers, hashtags, and words like “receipt”), each labeled email produced a distinct firing pattern across the simulated brain.
- Map activity to labels. Instead of adjusting weights the way a normal neural net training loop would, the process looked at which regions of the fly’s simulated brain lit up for each category and tied those activation patterns to labels, using something like a simulated imaging readout.
- Reinforce with correct answers. The system was walked through correct examples, then refined with reward-based feedback so the pattern-to-label mapping got more reliable.
- Connect decisions to actions. Once a category was identified, it triggered a pre-written template reply rather than a freshly generated response.
Critically, the motor circuitry that controls the simulated fly’s movement through its virtual environment was left untouched. That circuitry handles flight and navigation, and training it on email data risked corrupting basic movement behavior, the equivalent of the fly “flying into a wall” because its motor neurons got scrambled by irrelevant signals.
Why does the fly brain only reach about 80% accuracy? #
Because it’s a fly brain. Even mapped and simulated, the underlying network is built for a fruit fly’s actual life: sensing light, sensing smell, and moving toward the good versions of both. It was never built for language, categorization at scale, or nuanced decision-making.
According to the demonstration, the trained system reached around 80% of the performance of a small conventional neural network on the same email classification task, a small network that would take a fraction of the time to train. That’s a real result, but it comes with caveats. The system doesn’t compose replies from scratch. It matches an incoming email to the closest pre-written template and fills in the blanks, which is why replies came out generic (“Hey there” instead of using the recipient’s name). It’s classification and retrieval, not generation.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Scaling also has a hard limit. Attempts to push the system past five to ten categories saw accuracy degrade, likely because the network is simply too small and too specialized in its native wiring to support fine-grained distinctions the way a purpose-built language model can. Nobody should expect a fly brain to approximate even an early large language model. Fly brains don’t process language at all in nature. They’re built for a much narrower set of survival behaviors, and that ceiling shows up quickly once you ask more of them.
Is it ethical to run a real brain map inside a simulation? #
This is the question that has no confident answer yet, and it’s worth taking seriously rather than treating as a footnote. A connectome is a map of real biological wiring from a real organism. Running that map inside a simulation, feeding it sensory inputs, and letting it act on outputs blurs a line that neuroscience and philosophy haven’t settled: whether a simulated instantiation of a biological brain has any form of experience.
Nobody currently knows whether this constitutes anything resembling sentience for something as biologically simple as a fruit fly, let alone what it would mean at greater scale. That uncertainty hasn’t stopped the pace of these projects, and it’s reasonable to expect the same techniques to move toward more complex organisms (mice, rats, primates) as more connectomes get mapped and released. The gap between “this is possible” and “we understand what it means” is likely to stay wide for a while.
One informal response some builders have taken is to give the simulated organism a more humane environment: a virtual space to move through, simulated food, other simulated agents to interact with, rather than just a black box that receives inputs and emits outputs. Whether that meaningfully changes anything is unknown, but it reflects a genuine unease about running biological brain maps purely as utility.
What can (and can’t) you actually build with this? #
Right now, the practical ceiling is narrow. Simple classification tasks, sorting items into a handful of buckets based on pattern recognition, are within reach. Tasks that require nuanced language generation, long-range reasoning, or dozens of fine-grained categories are not, at least not with a fly’s connectome as the base network. The appeal isn’t efficiency: training and running one of these simulated brains is slower and less accurate than a small purpose-built model would be. The appeal is that it’s a genuinely new kind of computing substrate, a real biological wiring diagram doing digital labor, and that alone has made it a popular experiment to run and share.
Frequently Asked Questions #
What is a connectome?
A connectome is a complete map of the neurons and connections in a brain or nervous system. The fruit fly connectome released publicly documents around 166,000 neurons and more than 125 million connections in a Drosophila brain and nerve cord.
Who released the fruit fly connectome dataset?
Researchers connected to Google DeepMind mapped and published the dataset, making it downloadable for anyone to use in simulations or research.
Can you really train a fruit fly brain like an AI model?
Yes, in the sense that the connectome can be loaded into a simulation and treated as a fixed neural network. Feeding it labeled examples and mapping which regions activate for which inputs lets it perform classification-style tasks, though its capabilities are far more limited than a purpose-built artificial neural network.
How accurate is the fly brain at real tasks like email sorting?
Other agents start typing. Remy starts asking. #
Scoping, trade-offs, edge cases — the real work. Before a line of code.
In one demonstrated project, the trained system reached roughly 80% of the accuracy of a small conventional neural network on an email classification task, and it worked by matching emails to pre-written template replies rather than generating original text.
Is running a simulated fruit fly brain considered unethical?
There’s no settled answer. It’s genuinely unknown whether a simulated instantiation of a biological brain has any form of experience or sensation, and that uncertainty applies even to something as simple as a fruit fly.