Tag
作者提出一个衡量AGI的新标准：一个通用大模型能够仅凭纯视觉/音频输入和人类相同的控制方式，实时、可靠地在多种专业电竞比赛中与职业选手同场竞技，且具备真实反应、机制理解、策略竞争性和极低幻觉错误率的能力。
GPT-6 Astra has beaten the game Factorio: Space Age in 165 in-game hours over 2 days, which is significantly longer than the standard Factorio.
FactorioBench, a new benchmark tool, has just been released as announced in a Twitter thread.
A social media post shares results of GPT-6 playing Pokemon FireRed, demonstrating AI capabilities in gaming.
The article questions why AI agents haven't been reliably integrated as NPC characters in video games, despite demonstrations like NVIDIA's AI-NPC from three years ago and the existence of capable open models like Gemma 4. It argues that the gaming industry's lack of effort in this area undermines claims about AI agent capabilities.