My AI Fought Itself 2,000,000 Times and Reinvented Rock-Paper-Scissors

Two castles, one road, five kinds of soldiers, and an AI that has never seen a single human game. It learns everything by playing against itself: about 2 million games, over 4,000 at a time on one gaming graphics card, about 64,000 decisions every second. In this video: • the rules of the game and what every unit can do • how the AI "sees" the battlefield and makes decisions • the Great Stalemate: why the AI refused to attack for five whole minutes • the Turtle: the robot that beat the early AI, and how a later version set a perfect trap for it • the five ages of strategy: infantry rush, economy boom, archer wall, cavalry storm, balanced army, and why each one beat the one before • what it all means: there is no single best strategy, only the best answer to your opponents Chapters: 0:00 Intro 0:22 The Game 1:18 The Army 2:22 Gold & Powers 3:15 The AI Brain 4:30 What Went Wrong 7:18 The Turtle 10:08 The Five Ages of Strategy 14:32 How Far It Has Come 15:38 What's next Battle footage comes from real matches between trained AI versions (the short unit demos are small staged scenarios in the same simulator). The AI's "thoughts" on screen are its own outputs: its estimated win chance, what it wants to build, and how much it wants to charge. Narration: Microsoft Edge neural TTS (en-US-AndrewNeural). Music: synthesized from scratch in Python (royalty-free). 中文简介： 两座城堡、一条路、五种士兵，还有一个从没看过人类对局的 AI。它完全靠自己和自己对战来学习：在一张游戏显卡上同时打 4000 多局，每秒做约 6.4 万个决定。 视频里讲了游戏规则、AI 怎么看战场和做决定、训练中的"大僵局"、打败早期 AI 的"乌龟"机器人，以及 AI 在五个半小时里自己演化出的五个策略时代：步兵速推、爆经济、弓手墙、骑兵风暴、均衡阵容，和每次更替背后的原因。 章节： 0:00 开场 0:22 游戏规则 1:18 兵种 2:22 金币与技能 3:15 AI 大脑 4:30 训练中的问题 7:18 乌龟机器人 10:08 五个策略时代 14:32 最终有多强 15:38 下一步