Google's fruit fly connectome sparks headline bots, stock trades
WIRED's Will Knight built PitchFly, a project that generates story-pitch ideas by running the outlet's own past headlines through an open-sourced connectome, a detailed map of a male fruit fly's brain. Knight scraped hundreds of WIRED's best-performing headlines from the past year, turned them into text and phrases, then converted that into a representation a neural network could process; he says he had Codex do the technical work of wiring it together, calling the result "vibe coded." Fed on that training data, the fly brain now spits out pitches such as "The Hidden Weather Problem Inside Surveillance" and "Everyone Wants Cooking. Nobody Has Solved Donald Trump." Knight is upfront that the connectome does not understand any of the words; it is remixing patterns from the headlines it was shown, not reasoning about them.
The connectome itself was developed by Google together with a number of unnamed academic institutions and released as open source in early September. It maps 166,000 neurons and 125 million connecting synapses, modeling how a real fly's brain would fire in response to stimuli; the specific build behind PitchFly is credited with 165,112 neurons. Because it is open, Knight says anyone can import it into a project with a bit of AI-assisted prompting.
Since the release, a string of other hobbyist projects has used the same connectome for very different ends. An X user called Lyra Bubbles trained a version to play the VR rhythm game Beat Saber. Alex Wormuth, a software engineer at Coinbase, built StonkFly, which lets the fly brain decide how to trade stocks; Knight reports it is losing money but "doing surprisingly well, all things considered." Wormuth told Knight in a direct message that "the fly brings lighthearted, humorous relief" at a time when people worry about AI's existential risks, and that experimenting with the connectome raised questions for him about the ethics and possible consciousness of replicating a brain in software. Knight also mentions fly brains reportedly trained to play Doom, Minecraft and Pong, to solve Rubik's cubes, and to drive virtual cars, where they turned out to be bad at parallel parking. He cautions he has not verified that every one of these projects genuinely runs on the connectome; some, he suspects, may just be animations or spoofs. Alongside the novelty projects, he notes there are also serious scientific efforts built on the same map, including visualizations for exploring the connectome and tools for manipulating neural circuitry.
Knight frames the whole wave as evidence of how cheap and easy it now is to train and deploy a small, specialized AI model, in contrast with a full human brain's roughly 86 billion neurons, which remains impossible to map with current techniques. His conclusion is that the best use of the giant LLMs run by big AI companies may end up being to help build smaller, task-specific models like this one, rather than to do every job themselves.
Key facts
- WIRED's Will Knight built PitchFly, feeding hundreds of the outlet's past headlines into an open-sourced fruit fly connectome with 165,112 neurons to generate new story pitches.
- The connectome, developed by Google and unnamed academic institutions and released in early September, maps 166,000 neurons and 125 million synapses in a male fruit fly's brain.
- Coinbase engineer Alex Wormuth built StonkFly, which uses the same fly brain to trade stocks; it is losing money but, per Knight, "doing surprisingly well."
- Other hobbyists trained fly-brain projects to play Beat Saber, Doom, Minecraft and Pong, solve Rubik's cubes, and drive virtual cars, though Knight has not verified all of them actually use the connectome.
- Knight argues the trend shows how cheap it now is to train small specialized AI models, unlike a human brain's roughly 86 billion neurons, which cannot yet be mapped.
Why it matters
A research connectome released as open data turned, within weeks, into a playground for hobbyists with no neuroscience background: a journalist, a Coinbase engineer and assorted X users all built working projects on top of it. That speed is the story: it takes very little now to go from a raw biological dataset to a running, if silly, AI application, which is the same accessibility argument usually made about large language models applied instead to a completely different kind of model.
Who it affects
AI hobbyists and engineers looking for cheap ways to experiment with alternative, non-LLM model architectures; neuroscientists who study brain mapping and can reuse the same connectome for visualization and neural-circuitry research; and casual readers, since the throughline of the piece is entertainment and curiosity rather than a product launch.
How to use it
The connectome is open source, so Knight says it can be imported into a project with some AI-assisted coding rather than deep machine-learning expertise; he had Codex handle the technical steps of turning text headlines into a representation the network could learn from. No price, license terms or access restrictions are mentioned in the piece.
How solid is it
This is a firsthand account by the person who built PitchFly, complete with specific figures for the connectome's neuron and synapse counts, so the core claims about his own project are well grounded. The wider roundup of other fly-brain projects is looser: Knight explicitly says he has not verified that all of them genuinely run on the connectome rather than being fun animations, and some, he suspects, may be spoofs.
Risks and caveats
Several details are simply unstated: the exact release date and year for the connectome, which academic institutions besides Google were involved, and who built the separately-mentioned fly-based language model. The piece is also self-aware about its own limits: the connectome has no understanding of language or meaning, it is only remixing patterns, and Wormuth's musing about consciousness and ethics is offered as personal reflection, not an established finding.
“The fly brings lighthearted, humorous relief”
— Alex Wormuth, software engineer at Coinbase, creator of StonkFly