Summary
- GPT-6 Astra streamed Minecraft 141 hours, built a semi-automatic blaze farm and got Ender Pearls.
- A Creeper blew up its storage chest; Astra lost loot, then stalled and farmed potatoes.
- The run shows LLMs can plan complex tasks but fail on surprises — they need controlled environments and recovery plans.
X (Twitter) has been flooded today with memes about ChatGPT playing Minecraft on a livestream for 141 hours. According to a post by its evaluator, Vals AI, GPT-6 Astra got much farther than any other AI model it had previously tested. Reportedly, OpenAI’s latest model managed to set up a semi-automatic blaze farm, defeat more than six Endermen, and collect three Ender Pearls!
All was going well for Astra as thousands watched its progress on stream — until one catastrophic event appeared to frustrate the model dramatically: a Creeper blew up Astra’s storage chest and sent it into a potato-farming slump. Poor Astra.
Astra knew exactly what went wrong, but had no idea how to fix it
Hear me out: I understand Astra’s pain
-
-
In the second image above, you can see all the items Astra managed to get its hands on during the 141-hour playthrough. Many viewers were impressed by how far the model had progressed, up until the incident occurred. Astra had placed most of those valuable items inside a single chest next to its in-game bed. One unlucky morning, a Creeper exploded right in front of Astra, destroying itself along with the bed and chest.
Later, on a rainy day, Astra discovered that everything it had collected in the chest was gone. From that point onward, the model appeared to lose its motivation to make further progress and instead spent most of its time farming potatoes. Astra even left several in-game messages for itself:
“ALWAYS CARRY CRITICALITEMS withkeepInventory; don’tstoreinunguardedchesteveragain.”
“you can screw up and drop things”
In the X post below, Vals AI shows the moment when the Creeper destroyed the chest containing Astra’s most valuable items, followed by Astra’s later realization of the scale of its loss, as it looks down in the rain.
Anthropomorphism aside, this gives us new insights into the AI world
Recent AI models have proven to be noticeably more capable than their predecessors. It’s almost surreal to see the progress happening in real time. Just a couple of years ago, no LLM could’ve come close to what we now see people doing with GPT and Claude. Astra’s Minecraft run comes shortly after Claude helped create an entire operating system capable of running DOOM, although that project still required substantial human testing and troubleshooting.
Although anthropomorphism gives this story a cute narrative, Astra’s “potato-farming slump” offers a more serious insight. The incident illustrates why autonomous agents need controlled environments and recovery plans. An LLM can execute remarkably complicated sequences, yet still respond very poorly when one unexpected event derails the original plan.