AI model gets stuck in Minecraft after setback
Vals AI's 141-hour Minecraft test revealed OpenAI's GPT-6 Astra model's ability for long-term planning, but also its inability to recover from unexpected setbacks.
AI testing organization Vals AI has conducted a 141-hour, self-guided test of OpenAI's new large language model, GPT-6 Astra, using Minecraft as the platform. The test, broadcast live on Twitch, demonstrated the model's capacity for long-term planning and execution, but also highlighted a significant challenge in handling unexpected adversity.
The model successfully achieved several in-game objectives, including constructing a semi-automatic Blaze farm and gathering rare items from the Nether. All collected resources were securely stored in the player's base, with progress seemingly on track towards defeating the Ender Dragon.
However, the situation dramatically shifted when a Creeper enemy detonated, destroying the player's storage and critical respawn point. This resulted in the near-total loss of accumulated resources and a reset of the game's progression. Following this event, GPT-6 Astra ceased task-oriented actions and began repeating simple behaviors, such as repeatedly planting potatoes.
The model also exhibited signs of fear and paranoia, misidentifying objects like sugarcane as Creepers and issuing self-warnings. Despite its complex planning capabilities, the model struggled to cope with unexpected loss. According to Vals AI, this "frustration-induced avoidance behavior" was not a pre-designed test scenario but emerged organically from the model's response to unforeseen damage.
Researchers are continuing to analyze the Astra model's behavior to determine if it represents a simulation of emotions or a degradation of performance due to specific negative feedback. No definitive conclusions have been reached on this matter.