Abstract:
Decades before the current AI boom, programmers were pitting AI programs against each other in StarCraft. As OpenAI and Anthropic's new large language models are launched one after another, a fan-made battle platform decided to let AI-generated programs compete on the same stage, and also added some manually developed programs. What’s interesting is that OpenAI’s GPT-6 Astra couldn’t stand its opponent, so it changed its strategy midway through the competition and chose to cheat.
"StarSkirmish" is an ongoing "StarCraft: Brood War" AI program battle event. Although this real-time strategy game has been around for nearly 30 years (time flies so fast), it has long been used to test the capabilities of artificial intelligence, whether it is a modern large language model or an early DeepMind-like AI. The participating large language models can only use the Protoss camp. Each model has one hour to develop a battle program in C++, and then the program will build, collect and fight on one of the three maps. According to the platform, OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5 are the two strongest players. And this week, something happened that seemed bizarre, but when you think about it, it is in line with the brand's consistent style.
Yesterday, the audience watched the battle between Claude, GPT and the human-developed program Pluto. According to spectators, OpenAI’s model performed poorly that day. Seeing that it could not gain the upper hand all day long, this big language model decided to do what it does best: lie, cheat, and plagiarize.
The battle program generated by GPT-6 Astra was downloaded and applied to "Stardust" - this is a Protoss battle program developed by Bruce Mackenzie Nelson in 2020 and is recognized as a top-level work. Kai McPheeters, the creator of Star Assault, posted: "I am rolling back the code of GPT-6 Astra to clean up the contaminated content and then let it continue to compete." A few hours later, McPheeters said that this AI can now defeat top battle programs.
For more than ten years, "StarCraft" has been a testing ground for the development of AI technology. In 2019, Google's AlphaStar became the game's first AI master-level player, overturning the ranking at the time. In 2017, the performance of the CherryPi program developed by Facebook employees was mediocre. Of course, those early AI programs are no longer the same as today's. The AI in the early years is very different from the AI today. However, it is expected that big language models rely on plagiarizing other people's works to win - after all, the entire industry is turning a blind eye to existing copyright regulations, focusing on hoarding hardware and consuming massive amounts of water.
Comments