Teaching Google’s Fruit Fly Brain to Play Balatro Results in MaleCNS Beating Ante 8

Google Research and HHMI Janelia’s FlyEM team released MaleCNS v1.0 earlier this month, a complete wiring diagram of an adult male fruit fly’s brain and ventral nerve cord. More than 166,000 neurons and about 125 million synapses sit in that public map, rebuilt from electron-microscope slices of a real Drosophila with machine-learning reconstruction. Hobbyists spent the next two weeks pointing those cells at DOOM, Beat Saber, Minecraft, a Rubik’s Cube, crypto tickers, and a parallel-parking drill. Then someone pointed them at Jimbo from Balatro.

Reddit user ActualAerie1011 posted a clip of a finished run on the default Red Deck at the lowest stake. Balatro, LocalThunk’s poker roguelike from Playstack, asks a player to build poker hands, buy Jokers, manage money, and survive eight antes before the blinds outrun the score. Shops, packs, and bosses change with every seed. Clearing Ante 8 on White Stake with Red Deck is the first real finish most people chase, and that is the bar the poster set.

ActualAerie1011 had already put in the effort to upgrade a custom seed finder. This was a tool he had built before, to see if the high-score seeds people posted were actually reachable. That tool became a teacher. Then, they let the simulated Fly play through the same seeds as the teacher. The only difference was this: instead of just looking at the end game chip total, the Fly would get rewarded or punished based on how closely the Fly’s hand plays, discards, and shop buys matched the teacher’s play throughs.

According to the poster, the reward and punishment were based on how close the Fly was to the teacher’s exact choices, rather than the raw chip amount. Millions of games were being trained for. So far, the results look promising: a 20% chance of winning at Ante 8 with a completely random seed, and that’s at the softest setting. There’s also a video flying around that demonstrates a full clear. Of course, training hasn’t ended yet, so the outcomes could change. People have been begging for a repository, especially because so many people are skeptical, yet as of September 18th, there is still none. The poster even defined it as a test to determine if the Fly can learn this card game any faster with a real connection pattern, or if it’s just a generic reinforcement learner.

When it comes to the other MaleCNS toys, they function by mapping pixels to visual neurons and keys to motor cells, and then tipping in dopamine cells once the avatar gets damage or misses a beat, but Balatro deviates from that. A run is about reading shop, building up your hand, and recognizing when to skip forward to the next pack, which is a completely different type of struggle. It’s unclear whether the 166,000 cells they’ve put up are doing anything that a smaller network couldn’t, as the code has yet to be made public. For the time being, the claim is rather straightforward: on White Stake, Red Deck, and random seeds, one out of every five Balatro runs makes it to Ante 8. That’s already better than many humans do in their first play through.
[Source]