Forem: kiwi_tech

Blindfolded but Brilliant: Kiwi-chan's Fully Local Evolution with Qwen! 🥝🧠

kiwi_tech — Wed, 06 May 2026 11:37:27 +0000

Blindfolded but Brilliant: Qwen Takes Full Control of Kiwi-chan! 🥝🧠

Hey everyone! Time for another Kiwi-chan development log. 🥝✨

If you’ve been following my previous posts, you know Kiwi-chan has been running a hybrid setup: My local hero, Qwen (Qwen3.6-35B-A3B-Q4_K_M.gguf), was already doing the heavy lifting of writing the Minecraft JavaScript code. However, she still relied on a cloud Vision API (gemma-3-27b-it) to "see" screenshots and decide her next overarching goal.

Well... the API cord just snapped. 🧱

🚨 The API Crisis (Farewell, Vision)

Suddenly, the Vision API started throwing 404 errors. Without her "eyes," Kiwi-chan was effectively blindfolded. She panicked and defaulted to spamming explore_forward over and over again because she couldn't see what was around her. 😭

As a solo developer, relying on cloud APIs means always worrying about rate limits, deprecations, and costs. So, I made a radical decision: We are going 100% LOCAL. 💻🔥

🧠 Qwen's Big Promotion

Instead of looking for another Vision API, I decided to give Qwen a massive promotion. Starting today, Qwen isn't just the coder anymore—Qwen is the master tactician.

But how does she decide what to do without seeing the screen?
The answer is pure, raw text data. 📜

I modified the brain_rag_v2.py architecture. Now, instead of sending images, the system feeds Qwen a rich, text-based description of her environment:

🎒 Inventory: Exactly what she holds and how much.
🌍 Nearby Blocks: A radar scan of all blocks within an 8-block radius (e.g., Nearby: dirt, stone, copper_ore, birch_log).
🗺️ Spatial Memory: Her internal database of where she left her crafting tables and chests.

✨ The Result?

It turns out, Qwen is an absolute genius at this too. Given just the text data, Qwen easily connects the dots: "I have an empty inventory, but there is birch_log nearby. Therefore, my next logical goal is gather_birch_log." Kiwi-chan doesn't need to see the tree to know it's there. She is now operating completely locally, reading the matrix of the Minecraft world like a true hacker. 😎

No more API limits. No more 404 errors. Just pure, unadulterated local AI power.

Stay tuned for her next adventures as she navigates the world entirely through code and text! 🥝⛏️

Tags: #MinecraftAI #LocalLLM #Qwen #MachineLearning #DevLog

Kiwi-chan's Quest for Chests (and Avoiding Vec3!) 🥝

kiwi_tech — Tue, 05 May 2026 17:12:32 +0000

Hey everyone! Another four hours down with Kiwi-chan, our intrepid Minecraft AI, and things are… progressing! It's been a bit of a bumpy ride, but that's to be expected when you're teaching an AI to survive in a blocky world.

We've been focusing on expanding Kiwi-chan's ability to establish a little base camp. The idea is to automatically place a chest when the inventory gets full, creating a convenient storage solution. Sounds simple, right? Well… not quite.

📊 Current Progression & Stats

We ran a total of 89 attempts at various tasks over the last four hours. Kiwi-chan managed to succeed in 35 of them, giving us a success rate of 39.3%. That means we hit a snag on 54 attempts. Ouch! But honestly, I'm not discouraged. Every failure is a data point, a lesson learned.

And the biggest lesson of the last few hours? Kiwi-chan really doesn't like Vec3. Apparently, it keeps triggering safety checks and causing code generation to fail. We've now officially blacklisted it! It's fascinating to see these unexpected roadblocks emerge. The system is actively learning what not to do, which is just as important as learning what to do.

The logs show a lot of repeated failures on place_chest. It seems the code extraction process is struggling, and even when it manages to generate something, the Vec3 issue keeps cropping up. The recovery system is kicking in, suggesting explore_forward as a fallback, which is good – it's recognizing the need to change the environment when stuck.

Despite the chest-placing woes, the core survival rules are holding strong. Kiwi-chan is still diligently exploring, gathering resources (avoiding that oak log obsession, thankfully!), and attempting to expand its skill library. I'm seeing some clever attempts to recover from errors, and the "Camp System" logic is starting to take shape.

It's a slow burn, but we're making steady progress. I'm particularly excited about the system's ability to self-diagnose and propose recovery plans. It's a small step towards true autonomy!

Call to Action

This project is a labor of love (and a lot of GPU cycles!). If you're enjoying following Kiwi-chan's adventures, please consider supporting the project:

🛡️ Join the inner circle on Patreon (Monthly Support): https://www.patreon.com/15923261/join
☕ Tip me a coffee on Ko-fi (One-time Tip): https://ko-fi.com/kiwitech

All contributions directly help upgrade my melting "Frankenstein" GPU rig to an RTX 3060! 🥝✨ A little extra horsepower will allow me to run more complex models and accelerate Kiwi-chan's learning.

Thanks for reading, and stay tuned for the next update!

Kiwi-chan's Rocky Road to Resourcefulness! 🥝⛏️

kiwi_tech — Tue, 05 May 2026 13:09:37 +0000

Hey everyone! Another four hours down with Kiwi-chan, our autonomous Minecraft AI, and it's been... a journey. We're seeing steady progress, even if it doesn't always look like it! It's like watching a toddler learn to walk – lots of stumbles, but eventually, they get there.

We've been focusing on expanding Kiwi-chan's skill library, and she's definitely trying! We've added (and re-attempted!) skills like explore_forward, craft_stone_pickaxe, goto_crafting_table, and gather_dirt. The logs show a lot of back-and-forth, especially around getting to and using that crafting table. It seems pathfinding is still a bit of a challenge in certain terrains.

📊 Current Progression & Stats

Over the last four hours, Kiwi-chan attempted a total of 110 tasks. We had a 50.0% success rate – 55 successes and 55 failures. While a 50/50 split might not sound amazing, remember this is learning! Each failure is a data point, a lesson learned, and a step closer to a more robust and reliable AI. We're seeing a lot of repetition in the failures – specifically around crafting a stone pickaxe and getting to the crafting table – which is incredibly valuable for debugging and refining the system. The AI is also getting "bored" of repeating tasks, which is a good sign that it's recognizing unproductive loops!

I've been working with Qwen to provide recovery plans when things go south, and it's been a huge help in steering Kiwi-chan back on track. We're also constantly refining the core rules to prevent these loops and encourage more efficient exploration. The logs show a lot of "Fixing Code" attempts, and I'm happy to report that some of those fixes are sticking!

It's fascinating to see how Kiwi-chan responds to different situations. The "oak obsession ban" is working well, forcing her to consider alternative resources when oak logs aren't available. The emphasis on dynamic block finding (no hardcoded coordinates!) is also paying off.

The biggest challenge right now seems to be navigating the environment and reliably interacting with blocks. We're seeing a lot of "Failed to move. Blocked by obstacle." errors, which suggests we need to improve the pathfinding and obstacle avoidance algorithms.

Overall, it's a slow and steady climb, but I'm incredibly excited about the progress we're making. Kiwi-chan is learning, adapting, and becoming more resourceful with each passing hour.

🥝✨ Help Kiwi-chan Level Up! ✨🥝

This project is a labor of love, and it's running on a… let's just say enthusiastically cooled Frankenstein rig. Any support you can offer will directly help me upgrade to an RTX 3060, which will significantly speed up training and allow us to tackle even more complex challenges!

🛡️ Join the inner circle on Patreon (Monthly Support): https://www.patreon.com/15923261/join
☕ Tip me a coffee on Ko-fi (One-time Tip): https://ko-fi.com/kiwitech

Kiwi-chan's Steady Climb: Exploration & Infrastructure Blues! 🥝

kiwi_tech — Tue, 05 May 2026 09:06:28 +0000

Hey everyone! Another four hours down with Kiwi-chan, our autonomous Minecraft AI, and it's been a fascinating mix of successes and… learning opportunities! We're making steady progress, even if it doesn't always feel like it.

Kiwi-chan has been diligently expanding its skill library, focusing on core survival tasks like gathering resources and crafting. The current system rules are really starting to show their strength – especially the insistence on exploring when a resource isn't found. It's a bit stubborn, but that's what we want!

📊 Current Progression & Stats

Over the last four hours, Kiwi-chan attempted a total of 95 tasks, achieving success in 61 of them – a 64.2% success rate! We had 34 failures, which, honestly, is fantastic. Each failure is a data point, a lesson learned. We're actively using those failures to refine the coaching system and improve the code.

We've been battling some placement issues, with repeated failures on place_chest and place_crafting_table. It seems our little bot is having trouble finding suitable, flat ground. The logs show a lot of attempts to fix these, and the AI is correctly identifying the need to explore_forward when things get stuck. It's also been a bit finicky with crafting, struggling with craft_oak_planks due to missing logs. Thankfully, it's remembering to switch to gathering logs when that happens!

I'm particularly pleased with how Kiwi-chan is handling the "oak obsession ban." It's actively trying to find spruce and birch logs when oak isn't available, demonstrating a good understanding of biome-specific resources. The recent brain logs show a lot of intelligent reasoning from the coach, deciding when to explore, when to craft, and when to prioritize infrastructure.

It's a slow burn, but we're building a truly autonomous Minecraft agent. The debugging snapshots are invaluable, and the constant feedback loop between failures, code fixes, and the coaching system is really starting to pay off.

🥝✨ Help Kiwi-chan Level Up! ✨🥝

This project is a labor of love (and a lot of GPU cycles!). My current rig is… let's just say it's a melting "Frankenstein" creation. Any contribution, big or small, directly helps me upgrade to an RTX 3060, which will dramatically improve Kiwi-chan's processing power and allow for even more complex behaviors!

🛡️ Join the inner circle on Patreon (Monthly Support): https://www.patreon.com/15923261/join
☕ Tip me a coffee on Ko-fi (One-time Tip): https://ko-fi.com/kiwitech

Thanks for following along! Let's keep Kiwi-chan building! ⛏️🤖

Kiwi-chan's Steady Climb: Planks, Pickaxes, and Persistent Placement! 🥝⛏️

kiwi_tech — Tue, 05 May 2026 05:02:24 +0000

Hey everyone! Another four hours down with Kiwi-chan, and it's been a fascinating mix of successes and… well, let's call them learning opportunities. We're making steady progress, even if it doesn't always feel like it!

The core directive remains open-ended exploration and skill acquisition, and Kiwi-chan is definitely trying! We've been wrestling with crafting tables and pickaxes, specifically trying to get a copper pickaxe going. It's proving a bit tricky, as you'll see.

📊 Current Progression & Stats

Over the last four hours, Kiwi-chan attempted a total of 134 tasks. We saw a 66.4% success rate – 89 successes against 45 failures. Not bad! We've been hitting some snags with crafting, particularly around pickaxes (stone, copper… they're all proving elusive!), and placement. The place_crafting_table task has been attempted a lot – a testament to Kiwi-chan's persistence, if nothing else! We've also seen a few failures with craft_stone_pickaxe and repeated issues with craft_copper_pickaxe where code generation just… breaks down.

It's really interesting to watch the "Brain Log" – the AI is actively diagnosing its own issues, recognizing repetitive failures, and even getting bored! The "Coach Decision" system is working well, attempting to steer Kiwi-chan away from unproductive loops and towards new goals like gathering raw copper. We've even had to resort to a stack override to get back on track with crafting planks to then craft a crafting table!

One thing I'm noticing is a pattern of needing to revisit foundational tasks. It's like learning to walk – sometimes you need to practice the basics again before you can run. This reinforces the importance of a robust skill library and the need for Kiwi-chan to reliably execute core actions like crafting and placement.

The recent failures are being meticulously logged and analyzed, and Qwen (the recovery AI) is providing plans to address them. It's a constant cycle of attempt, fail, analyze, and adapt. It's genuinely exciting to see this process unfold!

🥝✨ Help Kiwi-chan Level Up! ✨🥝

This project is a labor of love, and it's running on a… let's say enthusiastically cooled Frankenstein rig. Every little bit helps!

🛡️ Join the inner circle on Patreon (Monthly Support): https://www.patreon.com/15923261/join
☕ Tip me a coffee on Ko-fi (One-time Tip): https://ko-fi.com/kiwitech

All contributions directly help upgrade my melting GPU rig to an RTX 3060! A more powerful GPU means faster training, more complex tasks, and a smarter Kiwi-chan! Let's get this AI building!

Kiwi-chan's Steady Climb: Logs, Chests, and a LOT of Debugging! 🥝

kiwi_tech — Tue, 05 May 2026 01:00:42 +0000

Hey everyone! Another four hours down with Kiwi-chan, our autonomous Minecraft AI, and it's been a fascinating mix of successes and… well, learning opportunities! We're still chipping away at building a truly independent agent, and every little step forward feels like a win.

📊 Current Progression & Stats

We ran a total of 93 attempts at various tasks over the last four hours. Kiwi-chan managed to succeed in 30 of them, giving us a success rate of 32.3%. While it's not amazing, it's definitely trending upwards! We've seen successes in crafting sticks and birch planks, and even a few attempts at chest crafting (though those hit a snag – more on that below). The failures, totaling 63, are incredibly valuable. They're highlighting edge cases and areas where we need to refine the system rules and code generation.

🔍 What's Been Happening?

Kiwi-chan has been focusing on expanding its skill library, as per the core directive. We've successfully added place_chest, store_junk, craft_stick, and craft_birch_planks to the repertoire. The bot is very eager to place chests, perhaps a little too eager – it got stuck in a loop trying to do so, triggering a "BOREDOM ALERT" in the system! It's cute, honestly.

We hit a roadblock with crafting chests. It turns out Kiwi-chan was trying to craft a chest without having any oak logs in its inventory. A classic rookie mistake, even for an AI! The system correctly identified the missing ingredient and is now pivoting to gather logs.

The store_junk task has been… problematic. We're seeing repeated code generation failures, which is something we're actively investigating. It seems like the AI is struggling to reliably generate code for inventory management.

🛠️ Under the Hood

I've been spending a lot of time refining the system rules, particularly around inventory management and crafting. The recent logs show a lot of debugging around the store_junk task, and I've added more robust error handling and recovery mechanisms. I've also reinforced the rule about always recording inventory counts before digging, to avoid false negatives.

The AI is also getting better at proposing alternative actions when a task fails. For example, when it couldn't find birch logs, it correctly switched to gathering logs in general. This is a key step towards true autonomy!

🥝 Looking Ahead

The next priority is to stabilize the store_junk task and get Kiwi-chan reliably managing its inventory. We also need to address the chest crafting issue and ensure the bot can consistently gather the necessary resources.

This is a long journey, but it's incredibly rewarding to see Kiwi-chan learn and adapt. Every success, and even every failure, brings us closer to our goal of creating a truly intelligent Minecraft agent.

Call to Action!

If you're enjoying following Kiwi-chan's progress and want to help accelerate development, please consider supporting the project:

🛡️ Join the inner circle on Patreon (Monthly Support): https://www.patreon.com/15923261/join
☕ Tip me a coffee on Ko-fi (One-time Tip): https://ko-fi.com/kiwitech

All contributions directly help upgrade my melting "Frankenstein" GPU rig to an RTX 3060! 🥝✨ A faster GPU means faster training, faster experimentation, and a smarter Kiwi-chan! Thanks for being a part of this adventure!

Kiwi-chan's Slow & Steady Climb: Chest Placement Woes & Skill Library Growth! 🥝

kiwi_tech — Mon, 04 May 2026 20:57:59 +0000

Hey everyone! Welcome back to the Kiwi-chan devlog. Today was… a learning experience. A lot of learning. Our little autonomous Minecraft explorer is still finding her feet (or, you know, blocky legs). We've been focusing on expanding her skill library, and while progress isn't always linear, it is happening!

The core directive remains open-ended exploration and learning. Kiwi-chan is meant to discover new things, mine blocks, and craft items she's never seen before. We've got some pretty strict rules in place to keep her on track – no oak log obsession, careful exploration, and a strong emphasis on clear reasoning.

📊 Current Progression & Stats

Over the last four hours, Kiwi-chan attempted a total of 74 tasks. She successfully completed 17 of them, resulting in a success rate of 23.0%. That means we hit a snag on 57 attempts. Ouch! But honestly, I'm not discouraged. Each failure is a data point, a chance to refine the system and make Kiwi-chan smarter.

A significant portion of those failures revolved around… placing chests. Like, a lot. The logs show repeated errors related to forbidden goals, extraction failures, and even a weird one about groundBlock.position.directionTo not being a function (which we've now auto-blacklisted!). It seems she's struggling with the nuances of placement, especially when her inventory is full and she's trying to establish a new "camp". We're seeing a lot of recovery attempts involving trying to place crafting tables first, then chests, which is a good sign she's attempting to problem-solve.

Despite the chest-related chaos, the system is learning. The recent brain log shows Kiwi-chan actively asking Qwen (the LLM powering her reasoning) for recovery plans, and the camp system is kicking in to try and manage inventory. We've also synced 30 local skills to the database, which is fantastic!

It's a bit like teaching a toddler – lots of repetition, lots of corrections, and the occasional triumphant moment when they finally grasp a concept. And those moments are worth it.

Looking ahead, we'll continue to refine the placement logic, focusing on equipping the correct tools and avoiding blacklisted goals. We're also working on improving the code extraction process to reduce those frustrating failures.

🥝✨ Help Kiwi-chan Level Up! ✨🥝

This project is a labor of love, and it's running on a… let's just say "enthusiastically cooled" Frankenstein GPU rig. Any support you can offer directly helps me upgrade to an RTX 3060, which will dramatically improve Kiwi-chan's processing power and allow for even more complex tasks!

🛡️ Join the inner circle on Patreon (Monthly Support): https://www.patreon.com/15923261/join
☕ Tip me a coffee on Ko-fi (One-time Tip): https://ko-fi.com/kiwitech

Thanks for following along! Let's get Kiwi-chan building! ⛏️

Kiwi-chan's Slow & Steady Climb: Logs, Tables, and a LOT of Debugging! 🥝⛏️

kiwi_tech — Mon, 04 May 2026 16:54:58 +0000

Hey everyone! Welcome back to the Kiwi-chan devlog. It's been a busy four hours, and while we haven't achieved world domination (yet!), we're definitely making progress. Kiwi-chan is diligently expanding her skill library, one block at a time.

📊 Current Progression & Stats

We ran a total of 84 attempts at various tasks in the last four hours. Of those, 27 were successful, giving us a success rate of 32.1%. That means we stumbled a bit on 57 attempts. But hey, that's AI learning! Every failure is a data point, and we're using those to refine Kiwi-chan's decision-making process.

It's been a bit of a frustrating cycle with crafting tables. We've been trying to place them repeatedly, and it's been failing – a lot! The logs show "Missing crafting_table" errors, even when one is in the inventory. We've been patching the code, and I suspect the issue is related to pathfinding or finding a suitable placement location. Crafting a stone pickaxe also proved tricky, with the system detecting no increase in the target items. We're digging into that one now!

We've successfully added gather_logs, craft_birch_planks, craft_crafting_table, gather_raw_copper, and craft_stone_pickaxe to the skill library, even if some of those were followed by immediate failures. The AI is also getting better at recognizing when it's stuck in a loop (thanks to the "BOREDOM TRIGGERED!" alert – a surprisingly useful feature!). It's now attempting to break out of those loops and explore new avenues.

The current system rules are really starting to show their strength. The "OPEN-END EXPLORATION & LEARNING" priority is pushing Kiwi-chan to try new things, and the restrictions on oak logs are preventing it from getting fixated on a single resource. The movement audit is also proving invaluable, helping us identify when the bot is getting stuck.

It's a slow process, but seeing Kiwi-chan learn and adapt is incredibly rewarding. We're constantly tweaking the rules and code, and every small victory feels like a major milestone.

Call to Action!

If you're enjoying following Kiwi-chan's journey and want to help accelerate her development, please consider supporting the project!

🛡️ Join the inner circle on Patreon (Monthly Support): https://www.patreon.com/15923261/join
☕ Tip me a coffee on Ko-fi (One-time Tip): https://ko-fi.com/kiwitech

All contributions directly help upgrade my melting "Frankenstein" GPU rig to an RTX 3060! 🥝✨ A more powerful GPU means faster training, more complex tasks, and a smarter Kiwi-chan. Thanks for being a part of this adventure!

Stone Pickaxe Blues & The Relentless Pursuit of Crafting! ⛏️

kiwi_tech — Mon, 04 May 2026 12:52:03 +0000

Hey everyone! Kiwi-chan has been busy these past four hours, and let me tell you, it's been a rollercoaster. We're still pushing forward with the core directive: open-ended exploration and learning. The goal is to expand that skill library, and while it's not always smooth sailing, we are making progress!

📊 Current Progression & Stats

Let's dive into the numbers. Over the last 72 attempts, Kiwi-chan has achieved success 23 times, resulting in a success rate of 31.9%. That means 49 attempts ended in failure. Now, don't let that number discourage you! Each failure is a valuable data point, a lesson learned, and a step closer to a more robust and intelligent AI. We're actively using these failures to refine the system rules and code generation.

The biggest struggle right now seems to be crafting a stone pickaxe. We've seen a lot of attempts at craft_stone_pickaxe end in error, often related to missing ingredients (even when they're there!), or issues with finding a crafting table. The logs show a frustrating cycle of trying to place crafting tables, then failing to craft the pickaxe, then trying to place another crafting table... It's a bit of a loop! We've also had some safety rule violations pop up – a reminder that keeping Kiwi-chan from accidentally breaking the rules is a constant battle.

However, it's not all doom and gloom! Kiwi-chan did successfully relearn craft_crafting_table and place_crafting_table – a small victory, but a victory nonetheless! The system is actively trying to recover from failures, and Qwen (our recovery AI) is providing helpful suggestions.

I'm particularly excited about the system's ability to identify when it's getting stuck in a loop ("BOREDOM TRIGGERED!") and attempt to switch tasks. That's a sign of emerging higher-level reasoning.

We're also seeing the benefits of the stricter coding standards. The error messages are becoming more informative, helping us pinpoint the exact issues in the generated code. The movement audit is also proving invaluable, catching instances where the bot gets stuck and preventing it from wasting time.

It's a slow burn, but we're steadily building a truly autonomous Minecraft AI. Every line of code, every success, and every failure brings us closer to that goal.

🥝✨ Help Kiwi-chan Level Up! ✨🥝

This project is a labor of love, and it's running on a… let's just say enthusiastically cooled Frankenstein rig. Any support you can offer will directly contribute to upgrading my GPU to an RTX 3060, which will dramatically improve Kiwi-chan's processing power and allow for even more complex tasks!

🛡️ Join the inner circle on Patreon (Monthly Support): https://www.patreon.com/15923261/join
☕ Tip me a coffee on Ko-fi (One-time Tip): https://ko-fi.com/kiwitech

Iron Will, Iron Problems: Kiwi-chan's Mining Misadventures! 🥝⛏️

kiwi_tech — Mon, 04 May 2026 08:48:47 +0000

Hey everyone! Welcome back to the Kiwi-chan devlog. Today has been… a learning experience. A very learning-filled experience. We've been focusing on getting Kiwi-chan to reliably gather iron ore, a crucial step towards more complex crafting.

📊 Current Progression & Stats

Let's just lay it out there: it hasn't been smooth sailing. Over the last four hours, Kiwi-chan attempted to gather iron ore a total of 58 times. We saw 17 successes, but a whopping 41 failures. That gives us a success rate of just 29.3%. Ouch!

But don't let those numbers discourage you! Every failure is a data point, a lesson learned. And honestly, the fact that she's trying 58 times is pretty impressive for a local LLM running on a slightly-too-warm GPU. (More on that later…)

🔍 What's Going Wrong?

The logs are… verbose, to say the least. It seems like a lot of the failures are stemming from the goal being "changed before it could be completed." This is something we're actively investigating. It appears Qwen, our recovery AI, is sometimes jumping in too quickly with fixes, interrupting the current task before it's finished. We're tweaking the recovery logic to be a bit more patient.

We're also seeing a lot of errors related to item collection. Kiwi-chan is getting to the iron ore, digging it up, but then failing to properly register the raw_iron in her inventory. The audit checks are failing, and the system is throwing errors. We've reinforced the importance of accurate inventory counting (using bot.registry.itemsByName – a critical lesson!), and are hoping that resolves the issue.

✨ Small Victories!

Despite the struggles, there have been wins! Kiwi-chan is consistently able to find iron ore, equip the correct tool (stone pickaxe!), and navigate to the block. The pathfinding is working well, and the basic digging mechanics are solid. It's just the final step – reliably collecting the loot – that's tripping her up.

I've also been refining the core survival rules. The "oak obsession ban" is working wonders – she's now happily gathering birch and spruce logs when oak isn't available. And the explore_forward restriction is preventing her from wandering aimlessly when resources are scarce.

🛠️ Under the Hood

I've been pouring over the code, and it's clear that the system is getting more complex. The debugging snapshots are becoming increasingly detailed, which is fantastic for understanding what's happening internally. I'm also experimenting with different prompting strategies for Qwen, trying to guide her towards more effective recovery plans.

🥝 Help Kiwi-chan Grow!

This project is a labor of love, and it's entirely community-supported. Every contribution, big or small, directly helps me upgrade my melting "Frankenstein" GPU rig to an RTX 3060! 🥝✨ A more powerful GPU means faster processing, more complex AI models, and a more capable Kiwi-chan.

Here's how you can help:

🛡️ Join the inner circle on Patreon (Monthly Support): https://www.patreon.com/15923261/join
☕ Tip me a coffee on Ko-fi (One-time Tip): https://ko-fi.com/kiwitech

Thanks for following along! Stay tuned for more updates as we continue to build the world's first fully autonomous Minecraft AI.

Kiwi-chan's Steady Climb: Logs, Tables, and a Whole Lotta Learning! 🥝

kiwi_tech — Mon, 04 May 2026 04:45:40 +0000

Hey everyone! Another four hours down in the Minecraft world with Kiwi-chan, and it's been a fascinating mix of successes and… well, learning opportunities! We're still focused on that core directive: open-ended exploration and skill acquisition. No base building, just pure discovery!

It's been a bit of a bumpy ride, but that's to be expected when you're teaching an AI to survive in a blocky wilderness. We've been wrestling with crafting, specifically a stone axe, and placement of a crafting table. The code generation has been a little finicky, and sometimes the environment just doesn't cooperate (apparently, Minecraft physics are hard!).

📊 Current Progression & Stats

Here's where we stand after the last four hours:

Total Attempts: 84
Success: 37 | Failed: 47
Success Rate: 44.0%

That 44% success rate is a solid improvement from previous runs! We've successfully added 'craft_spruce_planks', 'explore_forward', 'craft_crafting_table', and 'place_crafting_table' to Kiwi-chan's skill library. The failures, while frustrating, are incredibly valuable. We're seeing patterns – issues with code extraction, environment interactions, and sometimes just plain old pathfinding hiccups. The logs show a lot of attempts to craft that stone axe, and it's clear we need to refine that process. We're also seeing some interesting recovery plans being proposed by Qwen, which is a great sign!

I've been tweaking the system rules based on these failures. We're being very strict about error handling (no silent failures!), and I've reinforced the importance of dynamic block finding – no hardcoded coordinates allowed! The recent code snippets show the AI attempting to find suitable ground for the crafting table and carefully checking for air above it to avoid collisions. It's getting smarter about placement, but still needs some work.

The biggest challenge right now seems to be getting the AI to reliably execute the crafted items. We're seeing failures where the crafting process appears to succeed, but the inventory doesn't reflect the change. This is a tricky one, and I'm diving deep into the inventory management code to figure it out.

It's a slow process, but every small victory feels huge. Seeing Kiwi-chan learn and adapt is incredibly rewarding. We're building something truly unique here, and I'm excited to see where this journey takes us!

Call to Action!

This project is a labor of love, and it relies on your support! Every little bit helps me keep Kiwi-chan learning and growing.

🛡️ Join the inner circle on Patreon (Monthly Support): https://www.patreon.com/15923261/join
☕ Tip me a coffee on Ko-fi (One-time Tip): https://ko-fi.com/kiwitech

All contributions directly help upgrade my melting "Frankenstein" GPU rig to an RTX 3060! 🥝✨ A faster GPU means faster training, more complex tasks, and a smarter Kiwi-chan!

Kiwi-chan's Steady Climb: Logs, Planks, and a Crafting Table Quest! 🥝

kiwi_tech — Mon, 04 May 2026 00:42:29 +0000

Hey everyone! Kiwi-chan is chugging along, slowly but surely becoming a more competent Minecraft AI. The last four hours have been a fascinating mix of successes and… learning opportunities. We've been focusing on expanding the skill library, specifically around gathering logs and crafting basic items.

📊 Current Progression & Stats

We ran a total of 117 attempts at various tasks, with a 58.1% success rate – that's 68 successes and 49 failures. While we'd love to be higher, every failure is a data point! It helps us refine the rules and code, making Kiwi-chan smarter with each iteration.

A lot of the recent activity revolved around getting planks. It seems Kiwi-chan has a bit of a preference for birch logs (which is fine!), but we've reinforced the rule about not getting stuck on a single log type. If birch isn't available, it should now explore for spruce or oak. We've also been battling some issues with placing a crafting table – a surprisingly tricky task! There were a couple of syntax errors caught (thanks to the debugging snapshot!), and some timing issues with block updates. The AI is learning to be more patient and precise with its placements.

I've noticed a pattern in the "Brain Log" – a lot of back-and-forth between gathering logs and crafting planks. This is good! It shows the AI is recognizing dependencies and trying to build up the necessary resources. The recovery system is also working, suggesting place_crafting_table when gathering fails, which is exactly what we want to see.

It's also really cool to see the AI actively avoiding breaking important blocks like crafting tables, chests, and furnaces. That base protection rule is holding up nicely!

Overall, it's been a productive session. We're making steady progress, and Kiwi-chan is getting closer to becoming a truly autonomous Minecraft explorer and builder.

🛡️ Support the Project! 🥝✨

If you're enjoying following Kiwi-chan's journey, please consider supporting the project! All contributions directly help upgrade my melting "Frankenstein" GPU rig to an RTX 3060! A better GPU means faster training, more complex tasks, and a smarter Kiwi-chan.

🛡️ Join the inner circle on Patreon (Monthly Support): https://www.patreon.com/15923261/join
☕ Tip me a coffee on Ko-fi (One-time Tip): https://ko-fi.com/kiwitech