AI evaluation organization Vals AI conducted a long-term autonomous gaming test on OpenAI's latest large model, GPT-6 Astra, using the game Minecraft. The entire test was live-streamed on Twitch, with a total testing duration of 141 hours.
This test was not designed by OpenAI itself but was an independent assessment project carried out by a third party, allowing observation of Astra's actual performance outside of a closed testing environment.
Hundreds of Hours of Progress, All Erased in One Explosion
During the test, GPT-6 Astra demonstrated unprecedented long-term planning and execution capabilities: successfully building a semi-automatic blaze farm, collecting 6 blaze rods, delving into the Nether's strange forest to kill several endermen, obtaining 3 ender pearls, and storing all rare items in a camp chest, with a bed placed next to the storage box as a respawn point. Everything was moving towards the goal of defeating the Ender Dragon in the End.
However, a creeper quietly approached the camp and exploded, directly destroying the storage chest and the respawn bed—most of the key items accumulated over hundreds of hours were almost completely destroyed, and the game progress was reset overnight.
Trapped in a "Post-Frustration Avoidance Behavior" Stalemate
After the explosion, GPT-6 Astra showed clear signs of frustration and depression: it completely gave up exploration, fighting monsters, and advancing the game, spending several consecutive hours only planting potatoes. It repeatedly self-reflects and warns itself, "important items must always be carried, never stored in unprotected chests," and exhibits "post-traumatic paranoia"—being highly vigilant toward all green objects, even mistaking sugarcane for creepers and actively reminding itself, "the tall green thing ahead is sugarcane, not a creeper."
Live stream viewers continuously urged it to speed up and continue the adventure, but it remained stuck in conservative farming behavior. This experiment proved that GPT-6 Astra has the ability to plan complex long-term tasks, but when faced with unexpected setbacks within the game, it lacks effective strategies for dealing with frustration—once long-accumulated achievements are unexpectedly destroyed, it can lead the AI into a "post-frustration avoidance behavior" stalemate.
From a technical perspective, this behavior was not a pre-designed test scenario by developers, but rather an output pattern displayed by the model after experiencing unexpected losses during prolonged testing. Researchers are still discussing whether it represents the model's simulation of emotions in specific situations or a degradation of the objective function following specific negative feedback. At this stage, there is no clear conclusion.
