AI通过观看YouTube视频学会了玩《我的世界》

OpenAI的最新人工智能(AI)项目显然已经学会了玩视频游戏《我的世界》。这是根据该公司官方博客上的一篇文章获得的信息，其详细介绍了如何训练一个神经网络来完成这一壮举。

OpenAI是一个致力于AI研究的组织，由埃隆·马斯克在2015年创立。据悉，研究人员使用了一种被称为“视频预训练(VPT)”的技术，涉及从实际人类游戏的视频中收集样本数据如按下按钮和鼠标移动，并创建一个注释这些动作的算法。

研究人员随后训练了一个“逆动力学模型(IDM)”以预测游戏视频中所采取的行动顺序。最后，向受训的IDM展示了7万小时的在线视频。

该模型能从视频中复制动作。它能够收集原木，将原木制作成木板，然后将木板制作成一个工艺台。该模型还能游泳、猎杀动物和吃食物。事实上，它甚至可以完成相当复杂的动作–反复跳跃并将木块放在玩家身下。

该AI甚至还能制作钻石镐，这是一项复杂的技能，人类玩家需要花费约20分钟才能完成。

OpenAI is an organization dedicated to AI research, founded by Elon Musk in 2015. The researchers are reported to have used a technique called video pre-training (VPT), which involves collecting sample data from videos of actual human games, such as pressing buttons and mouse movements, and creating an algorithm to annotate these actions.

The researchers then trained an “inverse dynamic model (IDM)” to predict the sequence of actions taken in the game video. Finally, 70, 000 hours of online video was shown to the trainee IDM.

The model can copy actions from the video. It can collect logs, make logs into boards, and then make boards into a craft table. The model can also swim, hunt animals and eat food. In fact, it can even perform quite complex movements-jumping over and over again and placing pieces of wood under the player.

The AI can even make diamond picks, a complex skill that takes about 20 minutes for human players to complete.