Recently, independent AI evaluation organization Vals AI conducted a long-duration autonomous gameplay test of OpenAI's latest large model, GPT-6 Astra, in Minecraft. The entire process was live-streamed on Twitch, with a total testing duration of 141 hours.
GPT-6 Astra demonstrated unprecedented long-term planning and execution capabilities, achieving feats previously difficult for AI to accomplish. It successfully built a semi-automatic Blaze farm and collected six Blaze Rods. It ventured deep into the Nether's Warped Forest, killed several Endermen, and obtained three precious Ender Pearls. After extensive exploration, it stored all rare materials in wooden chests at its base, placing its bed next to the storage chests as a respawn point. Everythinggj was progressing toward the ultimate goal of defeating the Ender Dragon in the End.
However, a Creeper quietly approached the base and exploded, directly destroying the storage chest and respawn bed. Nearly all key items accumulated by the AI over hundreds of hours were lost, resetting its game progress overnight.
After the explosion, GPT-6 Astra showed clear signs of frustration and depression. It completely abandoned exploration, combat, and progress toward completion goals, spending hours on end doing just one thing—planting potatoes. It repeatedly reflected and warned itself: always keep important items on your person, and never store them in unprotected chests. It developed "post-traumatic paranoia," becoming highly wary of all green objects, even mistaking sugar cane for Creepers, and actively reminding itself: "The tall green thing ahead is sugar cane, not a Creeper."
Viewers in the live stream continuously urged the AI to speed up and continue its adventure, but it remained stuck in conservative farming behavior. This experiment proves that while GPT-6 Astra possesses complex long-task planning capabilities, it lacks effective strategies for recovering from setbacks when faced with sudden unexpected blows in the game. Once long-term accumulated achievements are unexpectedly destroyed, the AI easily falls into behavioral stagnation.
