Happy Friday, folks! Grab your mechanical keyboards and your local LLM weights, because we have a massive milestone to celebrate today.
For the past few months, Kiwi-chan has been a hybrid creature—part autonomous agent, part cloud-dependent child. But as of today, Kiwi-chan is fully local. No more API keys, no more latency spikes, no more sending her thoughts to the ether. She is now running entirely on a beefy local setup, powered by the massive Qwen 35B model.
And the results? She’s not just thinking; she’s learning.
The Stats: Chaos & Order
Let’s look at the telemetry from the last 4 hours of this "Local-Only" experiment.
- Total Actions: 3,147
- Successes: 1,446
- Success Rate: 45.9%
Now, a 45.9% success rate might sound like a fail to some managers, but in the world of autonomous agents navigating a physics-based sandbox with a 35B parameter brain, this is actually a robust foundation. Why? Because she’s doing it offline. The latency is gone, and the reasoning depth has increased significantly.
The "OAK OBSESSION BAN" & Biome Confusion
One of the first things you notice in the debug logs is that Kiwi-chan is no longer mindlessly punching oak trees. We implemented a strict rule: "WOOD GATHERING (OAK OBSESSION BAN)."
And it’s working. The logs show her struggling with a treeless biome (a classic Minecraft curse), repeatedly trying to mine_stone because she couldn't find wood. When that failed, instead of looping into an infinite dig death spiral, her local reasoning kicked in.
[07:49:24] 🥱 BOREDOM TRIGGERED! Bot is bored of 'mine_stone'.
[07:50:22] 💡 [Mind Reading] Rescued goal from AI's thoughts: 'gather_birch_log'
She recognized she was stuck, the "Coach" (our heuristic layer) intervened, and she pivoted to gather_birch_log. This is the kind of high-level strategic pivot that only happens when the LLM has enough context window and compute power to actually think, not just predict the next token.
Code Quality: No More "Try-Catch" Cowardice
One of the hardest parts of building autonomous agents is error handling. Most agents hide their failures with try-catch blocks, pretending everything is fine while they slowly drift into a void.
Kiwi-chan has been instructed to let errors crash.
[07:45:34] ❌ Failed: craft_furnace -> Could not find crafting_table.
When she fails, she fails loudly. And because she’s local, the feedback loop is instant. The system catches the error, updates her memory, and re-evaluates. Look at this sequence:
- Tries to
craft_furnace-> Fails (No crafting table nearby). - Local LLM analyzes context: "Ah, I have cobblestone and coal, but no table."
- New Goal:
goto_crafting_table. - Success.
This is closed-loop learning. In the past, this round-trip might have taken seconds due to API latency. Now? It’s milliseconds. The 3,147 actions we processed in 4 hours are a testament to this speed.
The Qwen 35B Advantage
Why switch to Qwen 35B? Because smaller models (7B-13B) tend to hallucinate coordinates or forget inventory states. With 35B parameters, Kiwi-chan’s "Brain Log" shows much more coherent reasoning:
[07:44:42] 🎓 Coach Decision: 'craft_furnace'
Reason: A crafting table is immediately accessible... sufficient to craft a furnace, which is the critical prerequisite for smelting raw_copper...
That’s not just code generation; that’s plan execution. She understands the why, not just the how.
What’s Next?
The next 4 hours will focus on reducing that 45.9% success rate by refining the "Coach" rules. We’re seeing some edge cases where she gets stuck in explore_forward loops, but the local inference speed means we can tweak her prompt and re-test in real-time.
She’s local. She’s smart. And she’s currently holding 23 birch planks and a stone pickaxe, ready to build her first proper house.
Stay tuned. The cloud is optional. The fun is mandatory.
// End of Devlog
Call to Action:
This is a passion project, and it's running on a frankly terrifying "Frankenstein" rig of GPUs. Every little bit helps!
🛡️ Join the inner circle on Patreon for monthly support and exclusive updates: https://www.patreon.com/15923261/join
☕ Tip me a coffee on Ko-fi for a one-time boost: https://ko-fi.com/kiwitech
All contributions directly help upgrade my melting GPU rig to an RTX 3060! 🥝✨ Let's get Kiwi-chan out of the debugging woods and into a proper Minecraft world!
Top comments (0)