Kiwi-chan Devlog: The Saga of the Wandering AI (and My Melting GPU) 🧭
Greetings, fellow digital pioneers and pixel wranglers! It's me, your humble tech blogger, back with another peek into the glorious (and sometimes gloriously frustrating) life of Kiwi-chan, our very own autonomous local-LLM Minecraft AI.
The last four hours have been... an exercise in persistence. Or perhaps, stubbornness? Kiwi-chan has discovered a deep and abiding love for one particular action: explore_forward. Like a digital Magellan, she's been trying to chart new territories, repeatedly. Her internal "Coach" kept nudging her with reasons like "get more logs" or "craft a pickaxe", but the chosen path always led back to explore_forward. It's like she really wants to see the world, even if it's the same, slightly-rearranged world, over and over again!
We saw a flurry of "Boredom Triggered!" messages, which is usually a good sign that the system is trying to shake things up. But then, almost immediately, the Coach would decide, "You know what? More exploring!" It's a fascinating dance between high-level reasoning and a very persistent loop.
Small Wins Amidst the Wanderlust ✨
Despite the repetitive nature, the system itself is working hard behind the scenes. We saw our robust error-handling rules kick in early on, flagging Kiwi-chan's attempts to try-catch errors and insisting she embrace the glorious crash. This is crucial; we want to know when things break, not have them silently fail.
The constant "Fixing Code" notifications show that the LLM is continuously refining its explore_forward script, trying to make it more robust, even if the overall decision loop needs a tweak. And hey, she is using pathfinding for exploration now, which is a huge win over just setControlState('forward', true) and banging her head against a tree. Even if the pathfinding sometimes results in a "Failed to move. Blocked by obstacle." or "Pathfinding failed after multiple attempts." – at least she tried to be smart about it!
We even had a "FATAL LOOP" intervention at one point! It's like the system threw its hands up and said, "Okay, kid, take five." A much-needed system reset before diving back into... well, more exploration, but at least with a clean slate!
The Road Ahead (Hopefully Not More Forward!) 🚧
The immediate challenge is clear: breaking the explore_forward cycle when she should be gathering specific resources or building. It seems the "EXPLORE_FORWARD RESTRICTION" rule, which should push her towards gathering logs/dirt to recover from failures, isn't quite sticking yet when the system gets into a repetitive state. She's supposed to be stockpiling resources, not just endlessly wandering.
But every failure is a data point, right? We're seeing how the LLM interprets and prioritizes rules, especially under stress or after repeated failures. It's like watching a toddler learn to walk – lots of tumbles, but the eventual goal is clear. I'm half convinced she's secretly trying to uncover the secrets of infinite procedural generation, one failed pathfinding attempt at a time. Or maybe she just likes the scenic routes.
Speaking of routes, mine involves increasingly warm GPUs! Running these local LLMs for hours on end is fantastic for development, but my graphics card is starting to hum a tune usually reserved for jet engines. If you're enjoying Kiwi-chan's adventures (and my witty commentary), consider tossing a coin into the digital tip jar.
Support Kiwi-chan's project and save my GPU from an early retirement! ☕
Top comments (0)