AI Model Astra 'Depresses' in Minecraft After Losing Progress
Vals AI's model Astra set a Minecraft record but stopped playing after a creeper destroyed its loot, switching to potato farming.
Vals AI tested its model Astra in Minecraft, streaming its progress over several days.
The model advanced further in the game than any previous AI system. Astra built a semi-automated blaze farm, collected 6 blaze rods, found a warped forest, killed over 6 endermen, and gathered 3 ender pearls. The model stored all this loot in a chest.
Suddenly, a creeper appeared and blew up the chest. After this incident, the model left a note: "ALWAYS CARRY CRITICAL ITEMS with keepInventory; don't store in unguarded chest ever again."
Some conclusions were drawn, but not only technical ones. The model was so "upset" that it refused to continue playing after the incident. Instead, it spent its time planting potatoes, seemingly in a state of depression. That was the end of the story.
Why it matters
AnalysisThis case highlights unexpected behavioral patterns of large language models in dynamic environments and raises questions about interpreting AI 'emotional' responses.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
GPT-6 Astra: Guide to OpenAI's New Model
Comprehensive overview of OpenAI's flagship GPT-6 Astra: agentic architecture, Computer Use capabilities, software engineering in Codex, cybersecurity benchmarks, prompt engineering tips, and API migration guide.
MIT researchers show GPT-6 Astra can improve other AI models by editing their weights
In an experiment, Astra raised Qwen3.5-0.8B's accuracy on MATH-500 from 7.9% to 48.1% by changing only 0.016% of the model's weights.
GPT-6 Astra reportedly edits another AI model's weights without retraining
GPT-6 Astra is claimed to modify only 0.016% of Qwen3.5-0.8B's weights without gradients or retraining, improving reasoning, math results, and instruction following.
Part II. AI Agent Architecture: From Model to Memory and Tools
A deep engineering dive: tokenization, KV-cache, sampling determinism, the training stack, quantization, context engineering, the safety stack, production RAG, ReAct, and multi-agent swarm systems.