Fruit fly brain model learns to play Balatro, wins 1 in 5 games
A Reddit user has applied reinforcement learning to Google's recently released fruit fly brain connectome, training it to play the card game Balatro. The model currently achieves a 20% win rate on the easiest difficulty, using only on-screen information. The trainer algorithm also discovers favorable game seeds, and the user plans further training to improve performance.
The connectome, released by Google less than two weeks ago, maps the complete brain and central nervous system of an adult male fruit fly. Since its debut, hobbyists have repurposed it for tasks like day trading and parallel parking. The Balatro project applies reinforcement learning, where a trainer algorithm discovers favorable game seeds and compares its own play against the brain model's choices, rewarding or punishing accordingly.
The model currently wins 20% of random seeds on the easiest difficulty with the default deck, relying only on visible screen information. The user plans continued training, stating the fruit fly will return "strong and smarter." Skeptics note the absence of shared code or methodology, though similar projects like BalatroBot and BalatroLLM have existed for roughly a year.
This project could demonstrate that biological connectomes can be adapted for complex game-playing tasks, potentially informing neuromorphic computing research. For gamers, it may normalize AI opponents trained on unconventional architectures. However, the lack of shared methodology could fuel skepticism about hobbyist AI claims, affecting how such projects are received by the broader community. Society may see growing interest in repurposing biological structures for practical applications, though independent verification will remain a key challenge.