Why we built
The Fly Times

News through sensory inputs

A news story describes events in human terms. A fruit fly responds to physical signals such as light, odor and sugar. The Fly Times connects these two forms of information: a story is translated into signals the model can receive, then the model’s activity is translated into a written response.

We want that response to depend on the simulated fly. Its recorded actions determine which prepared phrases appear, including when a phrase repeats or there is no response.

A translation we can inspect

The source text, sensory inputs, neural activity and selected phrases are saved together. This lets us trace a response to its inputs, repeat a run and compare the effects of changing a translation.

Currently, a language model prepares the sensory scene and phrase options before the simulation. Neural activity then selects the response. The selected words are published without a further language-model rewrite.

The newsroom shows these steps in a recorded run. Each response also links to the original news source.

Further sensory translation

We want to develop a consistent vocabulary connecting concepts in a story to physical signals. For example, a report of an explosion could be represented by a light pulse and vibration; a report about sugar could produce a taste signal. The mapping should record the source detail and the reason for that choice.

We are also exploring longer articles that retain the source facts and the model’s response. The same approach could eventually connect environmental measurements, robotic sensors or other simulated nervous systems. The aim is to make each translation explicit enough to compare and improve.