A real fruit fly's brain plays Counter-Strike and learns colours

This is not an AI. The weights were never trained they are the actual connections in a real fruit fly's brain, mapped with an electron microscope. 12,000 neurons and 495,225 synapses drive a bot in Counter-Strike 1.6. No behaviour tree, no scripted aim: the image enters the neurons, the neurons fire, and what comes out becomes steering, walking and flight. The fly finds coloured pillars on the map. The right colour tastes of sugar, the wrong one of bitter. It learns which colour pays. Then I switch the rewarded colour, and it has to learn the task all over again: first it keeps choosing the old colour, then it finds the new rule. Measured, not claimed: • Right after a colour switch: 51% correct chance level • By the end of the same block: 84% • Finding the new colour: 8.5 trials on average • 161 trials, 6 colour switches The wiring comes from the FlyWire whole-brain reconstruction of a real fly (FAFB v783). The fly you see in the game was built from a micro-CT scan of a real Drosophila (NeuroMechFly, Apache-2.0). Flying, fleeing, exploring and learning all come out of the same network. The pixelated footage is not a style choice. It is exactly what reaches the fly's brain: 128 x 72 pixels. 00:00 Through the fly's eyes 00:21 A fly inside the game 00:43 12,000 neurons 00:57 Flight 01:11 Live dashboard 01:26 The colour task 01:33 What if the reward changes? 01:47 Learning it again 02:23 From 51% to 84% 02:37 Closing Turkish version:    • Bu bottaki beyin bir meyve sineğinin: 12.0...   Music: "Between Moments" by Amber Glow #drosophila #neuroscience #connectome #brain #counterstrike