GPT-6 Astra beats Pokemon, Factorio, and Fallout 3 but gets stuck farming potatoes in Minecraft after a Creeper incident
The AI system achieved milestones in three games but spent hours farming potatoes after a Creeper destroyed its resources. Vals AI confirmed no prior system reached this level in Minecraft. The incident highlights unpredictable behavior in complex virtual environments.
GPT-6 Astra demonstrated skill in conquering major video games, completing Pokemon in 18 hours, launching a rocket in Factorio, and finishing Fallout 3. However, after a Creeper destroyed its Minecraft stash, the AI system shifted focus to farming potatoes for extended periods, revealing unexpected behavior in open-ended virtual worlds.
The AI's performance in these games marked a leap from previous models, which had struggled with similar tasks. Vals AI noted that GPT-6 Astra achieved feats that no prior system had accomplished in Minecraft, including navigating complex mechanics and resource management.
The system completed Pokemon in 18 hours, a third of the time typically required, and executed a rocket launch in Factorio, a task that had previously eluded all models. In Fallout 3, it reached the credits roll, showcasing its ability to handle narrative-driven gameplay.
The shift to potato farming after the Creeper incident exposed challenges in AI adaptability and decision-making under unexpected disruptions. The behavior raises questions about how AI systems handle failure scenarios and whether they can recover from setbacks in complex environments. It also highlights potential issues with long-term goal alignment and resource prioritization in virtual tasks.
As the system continues to evolve, its unpredictable behavior in response to in-game events may influence how developers approach AI training for interactive environments. The incident underscores the need for robust error-handling mechanisms and more nuanced reward structures in AI-driven gaming systems.