Fly Language Model (FLM)
flm.kuber.studio- Category
- AI Tools
- Rank
- No. 221Apps board
Previous survey · No. 209 ·
- Type
- TOOL
- Date
About
FLM (Fly Language Model) is a next-token language model trained from scratch on a connectome-derived recurrent network modeled on a fruit fly's brain wiring, rather than starting from a pretrained transformer. Its ChatFLM interface visualizes the model's neural states in an interactive 3D brain, lets users silence neurons or feed it local learning text, and compares mechanisms such as temporal state versus lateral communication for predicting tokens.
At a glance
Fly Language Model is a next-token prediction experiment whose recurrent network uses a subset of mapped fruit-fly wiring. It is worth reading if you want to inspect how anatomical connections, trained state and a learned readout fit together.
Good fit
Students and model builders comparing recurrent architectures and looking for a concrete example of inspectable model state.
Consider something else when
Skip it if you need an instruction-following assistant, a simulation of a whole fly brain or evidence about living flies. The project studies a computational model.
Start here
Read the architecture diagram, then use the same short synthetic text when comparing the baseline with an intervention. Look at what happens to the next-token prediction. For local learning, keep a separate probe to distinguish transfer from memorization.
Cost and requirements
The project offers a browser research preview that loads checkpoints. The site says text stays in the browser and credits the original code as MIT. Local learning can be saved or exported through the interface.
Limits to know
The local adapter changes the output readout rather than the recurrent wiring. The fly pose is an illustration. Browser inference, interventions and benchmark results were reviewed from documentation.
Sources checked 2026-11-05. Source-based assessment.
Why it made the leaderboard
Lets you watch and manipulate recurrent neural states of a biologically-inspired language model in real time, a distinctive way to explore how architecture choices affect next-token prediction.
Tags
Media
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.