GPT-6 Astra Reached the Nether in Minecraft, Lost Its Stash to a Creeper, and Farmed Potatoes for Hours

Vals AI, a completely independent evaluator, tested OpenAI’s GPT-6 Astra by throwing it into Minecraft and allowing it to its own tools for 141 hours with only ordinary computer controls: screen, mouse, and keyboard. No fancy game-modifying plug-ins were employed. Thousands of people tuned in to the Twitch stream to watch how it performed. Which wasn’t surprising given that the first, shorter attempt had already created a Nether portal in two and a half hours, long before the game had ever been released. Finally, Astra was the first system Vals had witnessed approach a Nether Castle.

Progress moved at a steady pace, much like a house of cards built by a human builder. Then Astra emerged and began building a semi-automated blaze farm. It just got up and walked away, then returned with six flame rods in hand. That wasn’t all; it also located a warped forest, where it killed almost half a dozen Endermen and snagged three Ender Pearls, a haul required to reach “The End.” It was all carefully stowed in a box next to the bed, but as usual, a game bug occurred, and a Creeper soon appeared and destroyed everything.

BOOM! That single Creeper completely destroyed the chest and bed all at once. Vals later characterized it as the most costly Creeper they’d ever seen, but this was all part of a test to see how things would turn out. Later, if you’re playing the game and select keepInventory, any player items you might lose when you die will be saved. Of course, having a bed full of goods stowed away in an unprotected box right near to the spawn site is simply unbeatable. Later, on a particularly dreary in-game day, Astra saw the box was empty.

The action log began taking notes from itself. Some sentences were nearly entirely in capitals, such as this one: ALWAYS CARRY CRITICAL ITEMS WHEN YOU CAN USE keepInventory; you’ll never store them in an unsecured chest again. Another read: You can and will make mistakes and drop things. As Astra traveled through the landscape, some green stalks in the distance received a warning, followed by a correction: “The green objects ahead were sugarcane, not another Creeper!” Later, as nite fell and it needed to forage for food, Astra simply kept beating herself up for minor mistakes. Pigs transformed into “dark pink pixels that needed to be eaten.” And after one disastrous tool use, they soon regretted it: “Slabs were chosen incorrectly, and 15 seconds of pondering killed us.” Speaking of which, people in the discussion were all shouting for it to pick up the pace, but Astra simply plodded through the potato rows.

For hours after Creeper blew everything up, Astra did nothing but continue growing potatoes. Blaze work was out of the question, as there were no more Ender pearls to be found and no plans to rebuild the base. Potato plots are safe, repeatable, and close. That loop is what a lot of people recognized, which is why the clip spread so fast. Vals described the model as frustrated and ultimately defeated.
[Source]