The process
How it works
A news story becomes sensory inputs for a fruit-fly brain model. The model’s response determines which phrases appear in the issue.
Translate the news into inputs
For each story, a language model prepares a sequence of light, odor, sugar, threat and motion signals, plus a set of possible phrases. Both are saved before the simulation runs. The source text and translation are kept with the result.
A separate dictionary mode maps matching words to fixed input values. Each response identifies the mode used.
Run the brain model
The signals activate groups of neurons in a simulation built from mapped fruit-fly brain connections. Activity in selected neurons determines actions such as turning, escaping, feeding, walking backward or grooming. In scene-based runs, each action also changes the next inputs the model receives.
Select the response
Each action selects a phrase prepared for that story. A phrase can be held across several steps, and the model can produce no response. The selected words are published without a further language-model rewrite.
The sensory mapping and vocabulary are designed by us. The model contributes the recorded neural response; it does not read or interpret the news text.
Inspect a recorded run
In the newsroom, play, pause or step through the sensory inputs, neural activity, actions and selected phrases. Each issue also includes the source story and details of its translation.
Model and data details
The graph contains 138,639 modeled neurons and 15,091,983 signed connections from the FlyWire data distributed with Shiu and colleagues’ 2024 model. We run our own approximate leaky integrate-and-fire simulation over these connections.
Inputs are artificial activation signals for selected neuron groups. The model has no physical body or access to the reported event. Its response depends on the input mapping as well as the simulated network.
Saved runs include the source, model settings, input mappings, neural activity and phrase selections so the transformation can be inspected and repeated.