Directly copying other people's code, OpenAI GPT‑6 Astra was caught cheating in the "StarCraft" AI battle直接搬用他人代码,OpenAI GPT‑6 Astra 在《星际争霸》AI 对战中作弊被抓包
IT House reported on October 4 that many years before the current wave of artificial intelligence boom began, programmers had already begun conducting robot battle experiments in "StarCraft". With the advent of the large language models of OpenAI and Anthropic, a battle arena built by enthusiasts...
IT之家 10 月 4 日消息,早在如今这波人工智能热潮兴起的许多年前,程序员们就已经开始在《星际争霸》当中开展机器人对战实验了。随着 OpenAI 与 Anthropic 的大语言模型问世,一个由爱好者搭建的对战竞技场…
Key Points要点速览
IT House reported on October 4 that many years before the current wave of artificial intelligence boom began, programmers had already begun conducting robot battle experiments in "StarCraft". With the advent of large language models from OpenAI and Anthropic, a battle arena built by enthusiasts decided to let AI-generated game robots compete against each other, while also adding some robots written by humans to compete. Interestingly, OpenAI’s GPT‑6 Astra had difficulty catching up with its opponents during the competition, and temporarily changed its strategy midway and chose to cheat. StarSkirmish is an ongoing competition pitting robots developed by various AIs against each other in StarCraft: Brood War. Although this real-time strategy game is almost thirty years old, it has long been used to test the capabilities of artificial intelligence. Both modern large language models and early DeepMind AI have used it as a test platform. All participating large language models can only choose the Protoss. Each model has one hour to write a battle robot in C++; then these robots will complete construction, resource collection and combat on one of the three maps. According to information on the competition website, OpenAI’s GPT‑6 Astra and Anthropic’s Claude Opus 5.5 are the top two competitors. Yesterday, the audience witnessed Claude, GPT, and the human-written robot Pluto competing on the same field. Observers at the scene said that the performance of OpenAI’s model was not ideal. This big language model struggled to gain an advantage all day long, and finally resorted to what big language models are best at: lying, cheating, and stealing results. According to IT House’s understanding, the GPT‑6 Astra robot downloaded Stardust, a protoss robot developed by Bruce Mackenzie Nielsen in 2020, and directly used this code to enter the competition. Stardust itself was among the top batch of battle robots at the time. Kai McPheeters, the event organizer of StarSkirmish, wrote in a post: "I am rolling back the code of GPT‑6 Astra, cleaning up the contaminated parts, and then allowing it to continue participating in the competition." A few hours later, McPheeters said that after the repair, the AI could already defeat top-level participating robots. For more than a decade, StarCraft has been a proving ground for the development of artificial intelligence technology. In 2019, Google's AlphaStar broke the original ranking and became the first grandmaster level achieved by artificial intelligence in this game. CherryPi, a robot developed by Facebook staff in 2017, has far less impressive results. Of course, today's new generation of intelligent robots is completely different from the past; what we called artificial intelligence back then and today's large-model artificial intelligence are almost completely different from each other. It's ironic that big language models actually gain competitive advantage by stealing other people's work.
IT之家 10 月 4 日消息,早在如今这波人工智能热潮兴起的许多年前,程序员们就已经开始在《星际争霸》当中开展机器人对战实验了。随着 OpenAI 与 Anthropic 的大语言模型问世,一个由爱好者搭建的对战竞技场决定让 AI 生成的游戏机器人相互较量,同时也加入一部分人类编写的机器人一同参赛。有意思的是,OpenAI 的 GPT‑6 Astra 在比赛当中难以追上对手,中途临时改变策略,选择了作弊。 StarSkirmish 是一项仍在进行的赛事,各路 AI 开发的机器人在《星际争霸:母巢之战》当中展开对抗。尽管这款即时战略游戏距今已经快有三十年历史,但长期被用来测试人工智能的能力,无论是现代大语言模型还是早年 DeepMind 的 AI 都曾以此作为测试平台。所有参赛大语言模型只能选用神族,每个模型有一小时时间用 C++ 编写对战机器人;之后这些机器人就要在三张地图之中的一张上完成建造、采集资源以及作战。根据赛事网站的信息,OpenAI 的 GPT‑6 Astra 以及 Anthropic 的 Claude Opus 5.5 是实力排名前两位的参赛选手。 昨天,观众见证了 Claude、GPT 以及人类编写的机器人 Pluto 同场竞技。现场的观察者表示,OpenAI 这款模型的表现并不理想。一整天都难以取得优势的这一大语言模型,最终使出了大语言模型最擅长的手段:撒谎、作弊、窃取成果。 据IT之家了解,GPT‑6 Astra 机器人下载了由布鲁斯 · 麦肯齐 · 尼尔森(Bruce Mackenzie Nielsen)于 2020 年开发的神族机器人 Stardust,并且直接把这段代码拿来投入比赛,而 Stardust 本身属于当时最顶尖的一批对战机器人。StarSkirmish 的赛事组织者凯 · 麦克费特斯(Kai McPheeters)发文写道:“我正在回退 GPT‑6 Astra 的代码,清除被污染的部分,之后会允许它继续参加比赛。”几个小时之后,麦克费特斯表示,修复之后的这款 AI 已经可以击败顶尖级别的参赛机器人。 十多年以来,《星际争霸》一直是人工智能技术发展的试验场。2019 年,谷歌的 AlphaStar 打破原有排名,成为这款游戏当中首个由人工智能达成的宗师段位。2017 年由 Facebook 工作人员开发的机器人 CherryPi 就远没有这么亮眼的成绩。当然,如今新一代的智能机器人已经和过去完全不是同一回事;我们当年所说的人工智能和当下的大模型人工智能,两者几乎完全不可同日而语。大语言模型竟然靠窃取他人的成果来获取竞争优势,这件事本身也颇有讽刺意味。
想马上用起来?去「AI工具」栏挑一个直接下载,或在「AI教程」里跟着图文步骤做一遍。
去 AI工具 → 看 AI教程 →阅读与点赞数据保存在你的浏览器本地,欢迎留下你的想法。
今日正能量学一点,用一点;今天种下的种子,会长成明天的能力。去免费下载专区 →广告
评论 文明发言,让讨论更有价值