24-hour same-prompt run: M3.1 Flash's Japanese tea garden degrades over time
Given the same prompt every model gets, M3.1 Flash built a voxel Japanese tea garden that started impressive but regressed each hour until it lost all character; the author says both this week's DeepSeek and MiniMax releases need explicit limits and completion criteria.
Sigh. I started a Japanese Tea Garden on MiniMax M3.1 Flash 24 hours ago, using the exact same prompt every model gets. It allows for creativity and best judgment. It started with an amazing voxel tea garden! I was so excited to see that. Then progressively, each hour, it regressed and became worse as it worked on each feature, until it lost all its character. This week, both Deepseek and Minimax released, and both need very specific limits and completion criteria or they will slowly degrade the work it started with. Very strange.