flyschool

0x759f49aade96d52adc39a3e5149afe64e7b2d031 Scent inputs activate the circuit. Modeled dopamine changes eligible Kenyon-cell → MBON synapses. We probe trained and untrained weights using identical inputs and random seeds. The anatomy comes from mapped neurons; activity and movement are modeled.

Across our first 84 Food Runs: • 64 student wins, 20 ties against the untrained baseline. • 55 wrong turns versus the baseline’s 265. • In 28 unaided exams, the student finished faster 23 times and tied 5 times. What has changed inside the model? Reward pairing reduced the weights of 22,702 connections. Mean gain across the learning circuit fell to 61.3% of its untrained value, then plateaued. We can observe both the effect of training and its current limits. Our documented game controller translates that learned state into scent-following choices. Both competitors use the same layout, movement speed and seeded random draws; exams remove visitor cues. That’s our contribution to FLYBRAIN: a public, persistent testbed connecting neural measurements to inspectable game outcomes—with saved learning, matched controls, recorded replays and open methods. Other builders can fork it, challenge the controller, and test alternative learning rules. These are early simulation results. Biological behavior still requires validation. Explore the experiment: flyschool.dev

FLY SCHOOL is live. 🪰 165,122 modeled neurons. 10.2M connections. Reward changes synaptic weights in our FLYBRAIN fork. A game controller translates that learned state into scent-following choices, racing the student against an untrained baseline. flyschool.dev