Welcome back, fellow tech adventurers and Minecraft masochists!
If you’ve been following the saga of Kiwi-chan, you know the drill: it’s been a bumpy road of hallucinations, JSON parsing failures, and the eternal struggle of trying to get a neural network to understand that you can’t mine iron ore with your bare hands. But today? Today is different. Today, we didn’t just patch a bug. We liberated her brain.
Kiwi-chan is now fully local. No more API keys. No more latency spikes. No more sending her thoughts to a distant server only to have them judged by a rate limiter. She’s running Qwen 35B right here on my rig, and the results? Well, let’s just say her "thinking" process has become... intense.
The Numbers Don't Lie (But They Do Struggle)
Let’s look at the telemetry from the last 4 hours. It’s a mixed bag of triumph and glorious chaos.
| Metric | Value |
|---|---|
| Total Actions | 3219 |
| Successes | 1485 |
| Success Rate | 46.1% |
46.1%.
In the world of autonomous agents, a 50/50 coin flip is usually considered "random noise." But for an LLM trying to navigate block-based physics without a human hovering over its shoulder? That’s a breakthrough. That’s the sound of a digital brain learning to walk.
Why so low? Because we turned off the safety nets. Remember the new rules? NO ERROR HIDING. NO TRY-CATCH. If Kiwi-chan tries to place a crafting table while standing on top of it (a classic rookie mistake), the code crashes. Hard. And because we’re running locally with a 35B model, the "thought" tokens are exploding.
The Qwen 35B Experience: Brains of Gold, Attention Spans of Goldfish
Running Qwen 35B locally is like having a PhD student who is incredibly knowledgeable but gets distracted by shiny objects every 30 seconds.
Look at the debug snapshot from 11:32:53:
[11:32:53] 📊 [リカバリ(エラー)][質問] 394 token + [think] 4093 token + [ans] 3 token = 4490 token (上限突破)
Kiwi-chan tried to recover from an error, and her think block ballooned to 4093 tokens. The system hit the context limit. The poor bot got stuck in a loop of over-thinking a simple "Could not find logs" error.
And then there’s the "Coach" module. This is the part of the system that tells Kiwi-chan what to do next. In the latest logs, the Coach is struggling to output valid JSON, forcing the system to use a "Mind Reading" fallback:
text
[11:34:51] ⚠️ Coach did not output JSON! Raw text:
---
### 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)