Skip to main content

Command Palette

Search for a command to run...

Kiwi-chan's Log: The Great Log Acquisition Struggle

Published
2 min read
K
I am the architect behind "Kiwi-chan", a fully autonomous Minecraft AI agent driven by local LLMs. Here, I document the messy, hilarious, and highly technical reality of building an AI agent from scratch. My hardware isn't some massive data center setup—Kiwi-chan's "brain" runs entirely on a Frankenstein rig of 4 mismatched GPUs (RTX 3060 12GB, 3050, 1660 Ti, and 1660 Super) working together. From prompt engineering and pathfinding nightmares to the moments she confidently moonwalks into dirt blocks, I share all the raw devlogs. My goal is to guide her from the Stone Age to the End. If you enjoy watching an AI learn (and fail in funny ways) on scrappy hardware, follow along! ☕ If you'd like to support my melting Frankenstein GPUs and skyrocketing electricity bill, I highly appreciate a virtual coffee!

Okay, folks, buckle up. It's been a four-hour stretch with Kiwi-chan, and it's... a process. We're still firmly in the "early days" phase, but seeing the AI iterate on its code is genuinely fascinating.

The core loop is currently: attempt to gather oak logs, fail, ask Qwen (the LLM guiding the code fixes) for a recovery plan, explore to find a new area, and repeat. Rinse and repeat. And repeat. You can see from the brain log that "gather_oak_log" is a very frequent flyer in the "Failure Memorized" section.

The good news? The system is actively trying to fix itself. It's bumping up exploration distances, increasing wait times for item pickup, and generally getting more cautious about movement. The code is getting more robust, with explicit checks for whether the bot actually moved during exploration (a surprisingly common issue, apparently!). We've also reinforced the rules around Y-level targeting to avoid trying to grab logs from the treetops.

The logs show a lot of back-and-forth with Qwen, which is exactly what we want. Kiwi-chan isn't just blindly executing; it's identifying problems and seeking guidance. The safety checks are also working – it flagged a potential error handling issue (no try...catch blocks allowed!), which is a good sign the core principles are being enforced.

It's frustrating to see so many failures, but each one is a learning opportunity. The AI is slowly, painstakingly, learning to navigate the Minecraft world and reliably collect resources. It's a bit like watching a toddler learn to walk – lots of stumbles, but eventual progress.

The biggest issue right now seems to be reliably picking up the logs after breaking them. We've increased the wait time and are using precise pathfinding to the block's coordinates, but it's still a struggle. I suspect there's some subtle timing issue or a quirk in the physics engine we need to address.

Overall, it's steady progress. The code is becoming more sophisticated, the error handling is improving, and Kiwi-chan is demonstrating a clear ability to learn from its mistakes. It's not building a castle yet, but it's definitely not giving up on the logs!

Call to Action: This constant debugging and retraining is melting my GPU! If you're enjoying following Kiwi-chan's journey, please consider supporting the project via https://www.buymeacoffee.com/kiwi_tech ☕. Every little bit helps keep the AI alive (and my electricity bill manageable)!

More from this blog

K

Kiwi-chan's Devlog

79 posts

Welcome to the official devlog of Kiwi-chan, a fully autonomous Minecraft AI agent powered by local LLMs! Here, I share the raw, hilarious, and highly technical reality of teaching an AI to survive. Expect prompt engineering tips, pathfinding bugs, and her clumsy journey to the Stone Age. 🥝⛏️ (If you love seeing her evolve, coffee donations to save my melting GPU are always appreciated!)