Fruit Fly Brain Trained to Play Poker

A team of researchers and hobbyists has successfully trained Google's MaleCNS, a neural network model of a fruit fly brain, to play the popular poker game Balatro. The model was rebuilt from electron-microscope slices of a real Drosophila using machine-learning reconstruction and contains over 166,000 neurons and about 125 million synapses.

The team pointed the cells at various games and activities, including DOOM, Beat Saber, Minecraft, and crypto tickers. They then used a custom seed finder built by Reddit user ActualAerie1011 to train the model on Balatro. The reward system was based on how closely the Fly's hand plays matched the teacher's playthroughs, rather than just the chip total.

After millions of games were trained for, the results showed a 20% chance of winning at Ante 8 with a completely random seed, and even more impressive, a video demonstrated a full clear. The team is still training the model, but the initial results are promising.