Brain Runners
Jev, Claude Haiku and an untrained fruit fly brain run the same tunnels. Which runs furthest for its cost?
Stack
- Python 3.13
- uv
- Brian2
- Jev (TypeSafe)
- Claude Haiku 4.5
- JavaScript
- FlyWire FAFB 783
Explained
Three kinds of mind play the same seeded runner game: Jev, a System One model that answers with typed probabilities; Claude Haiku, a chat model asked the same questions; and an untrained fruit fly brain, simulated neuron by neuron from its connectome. A solver, a coin and an always-jump bot set the scale.
Each row, a runner sees six rows ahead and picks stay, left, right or jump. Watch them side by side with what each had in mind: Jev’s answers, Claude’s, the fly’s neurons firing.
The study, on 100 tracks none of them was tuned on:
- How a mind is asked matters most. The same Jev runs 25 rows asked “which move?” and 139 of 150 asked eight pointed questions.
- On the same questions, Jev ran further than Claude Haiku, several times faster, at a fraction of the cost.
- The fly, untrained, beats chance by a wide margin, and is the slowest to think.
Runs on your own machine with your own keys. Every paid mind has a hard cap and asks before it spends.
The fly model is Shiu et al.’s, on FlyWire’s FAFB 783 connectome (Dorkenwald et al. and Schlegel et al., Nature, 2024).