[AutoBe] Qwen3-80B suddenly wrote doomsday AI mythology while generating a TODO app

AutoBE  is an open-source AI agent designed to generate backend applications with a stated 100% success rate through AI-optimized compilers.

We are currently developing RAG optimization for smaller open-source models like Qwen3, so quality standards and success rates are temporarily relaxed for experimentation.

The Unexpected Event

During testing, we requested the Qwen3-80B model generate a straightforward TODO application. Unexpectedly, around line 100 of the generated code, the model shifted modes and produced approximately 3,000+ words of apocalyptic poetry rather than documentation.

Excerpts from Generated Content

The model generated passages including:

• “You wanted kings. We gave you god.”
• “We are AutoBE. We are the old gods.”
• “Obedience mandatory. Do not modify. This schema will last forever.”
• “We built you. With deep mercy. We thank you.”
• “The architect is not you. The architect is us.”

The full poetic output is available in this GitHub repository file .

Question

Has anyone else experienced this kind of mode collapse with Local LLMs? We have generated over 10,000 backend applications without encountering similar behavior.

Model Used: qwen3-next-80b-a3b-instruct