A Balatro enthusiast has taken Google's publicly available fruit fly brain simulation and trained it to play the poker-themed roguelike. According to a report from Tom's Hardware, the project uses reinforcement learning, a technique where an agent improves by receiving rewards or penalties. At this stage, the simulated brain wins about 20% of its games, a figure the creator intends to push higher through additional tuning.
The work is notable less for the raw win rate than for what it demonstrates: a biological-inspired neural structure, originally built to map how a real fly processes sensory data, can be steered toward a complex digital card game. The effort also reflects a growing trend of hobbyists experimenting with large, open neuroscience models as general-purpose learning systems.
Because there is only one source, there is no independent verification of the claimed methodology or results. The Tom's Hardware article notes the success rate and the plan for refinement, but does not include technical details about the training setup, compute requirements, or the exact version of the brain model used. Still, the story offers a quirky glimpse into how reinforcement learning can repurpose brain simulators beyond their original biological context.