Người hâm mộ dùng mô hình não ruồi của Google huấn luyện chơi Balatro đạt tỷ lệ thắng 20%
Balatro fan claims they trained Google fruit fly brain simulation to beat the game — reinforcement learning currently has the model at 20% success rate
Tóm tắt nhanh
- Less than two weeks after Google released a complete adult male fruit fly connectome, a Reddit user trained that brain structure to play Balatro and shared a sped-up video of it playing.
- The player, ActualAerie1011, says they used a trainer algorithm they developed to discover useful Balatro seeds and trained the connectome-plus-model by rewarding or punishing choices.
- The model currently achieves a 20% success rate on a random seed at White Stake difficulty with the default Red Deck, according to the user.
- The user states the model only perceives what's immediately visible on-screen and that training is ongoing; they hinted the brain will improve.
- Commenters have cast doubt on the project, and the player did not share detailed training methods, a repo, or references to other open-source tools.
- Tom's Hardware notes similar Balatro projects (BalatroBot, BalatroLLM) exist, and training is hard due to complex joker interactions, perk ordering, boss abilities and scoring rules.
Tóm tắt tự động bằng AI từ bài gốc, có thể có sai sót. Hãy kiểm chứng ở bài gốc.
Trích đoạn bài gốc
One Balatro player says they've taken Google's mapped fruit fly brain and trained it to play Balatro, currently at a 20% success rate with plans for further refinement.Tapview chỉ hiển thị tiêu đề, trích đoạn ngắn, tóm tắt và đường dẫn tới bài gốc. Nội dung gốc thuộc bản quyền của Tom's Hardware.
