I find this discussion fascinating, as I can see views in both perspectives. As with most things, there is a double-edged sword at play here. For instance, while some may be hesitant to promote AI within the Monero community, AI systems may soon rely on Monero for transactions. Take @monerica as an example, they leverage AI to assist with site development while transacting in Monero, yet this approach may arguably undermine the very principles that makes Monero valuable. I may not have articulated this perfectly, but that is my interpretation of the points raised. @XMR

How exactly does using AI for writing code lead to "...this approach may arguably undermine the very principles that makes Monero valuable"?

You cannot licence AI generated code at all, not under GPL, not under copyright, nothing. This has already been legally ruled.

But if you use AI then you should understand that you are assisting the adoption of something that is contrary to the goal of monero. AI corpos want you to rely on them for skill and knowledge instead of learning it yourself, so that you are forced to pay them for it. You really cannot argue that this is something positive or aligned with the goals and spirit of monero at all.

If AI did or does become successful then cutting jobs with no replacement will lead to society collapsing. Unemployment leads to boredom, boredom leads to crime, and widespread crime leads to collapse. The only people in the world that this benefits are the billionaires.

@monerica TBH I don't care if you use AI because I never used any of your services and with this information I never will

You're free to use whatever you want. But this kind of argument could be applied to IDEs, because don't code editors want you to rely on them to write/ organize code? The idea that you can use software that wasn't in some way assisted with AI is a bit difficult at this point. Do you think this forum was made using AI at all, and if you found out even 1 line of code was written with AI, would you still use it? If even 1 line of code in the Monero code base was made with AI, would you stop using it?

Nothing made for Monerica couldn't have been done without AI. It is simply more productive to use it. Does it decay skills and force reliance on AI more, sure. But, all the code is not going to be re-written for a handful of people just to say it was hand-made. And even if it was, how would you know?

All the code is open sourced anyway, there is no intention to license it.

What anti-AI Monero directories are there out there? It's pretty clear that KYCnot.me uses AI everywhere to summarize all the terms and conditions of other sites. If you really believe there's a market for anything you are proposing, surely you would create it, right?

The submissions are actually manually reviewed, for what it's worth. There is a ton of manual work, but that doesn't matter somehow, does it?

@monerica Talk your shit, Monerica. My 'undermining' comment comes from the OG school of thought. I’m skeptical of anything AI-assisted. In my mind, if it's hitting a server like ChatGPT, there's a paper trail. I don't trust the black box. I do appreciate your defense. I still use your services, so my "OG school of thought" isn't all that OG lol.

@Unstirred - look, if everything could be run locally then it would be. But who is taking the time to actually explain how to do that without spending a ton of money? You detail how to run an AI as good as the big ones locally then it will be considered.

ChatGPT was used for a long time, but now Claude is being used.

You should be skeptical, this is what make the Monero community what it is. But there is a spectrum from completely relying on machines/AI vs going full Kaczynski and not using the internet at all because it all feeds into the government databases and control.

@monerica Reliance on AI is harmful for society and only beneficial for corporations. It teaches people to be lazy instead of learn skills and knowledge. People who learn are not repeat customers for AI so they play on the human psychology to make it easier to just copy paste the output instead, so that you keep paying.

Corporate centralized AI is completely contrary to the idea of a decentralized currency, corporations do not want the average people to have control of anything, they want you to own nothing and be happy because that's what generates the most revenue for shareholders.

@Shecp are you implying you do not use AI at all?

Ideally all the AI models are run locally but that's not practical.

It's not just for corporations, it empowers every individual to create.

Running X poll on this topic.

@monerica

look, if everything could be run locally then it would be. But who is taking the time to actually explain how to do that without spending a ton of money? You detail how to run an AI as good as the big ones locally then it will be considered.

that was few months ago, they were all working locally (well, Kimi K2.5 needed over 200GB of VRAM, only had 128GB of share RAM/VRAM so with the SWAP on NVME drive that would have taken actually literally forever for even just a 1 character prompt lol, but when they releasing a rig like that in the future with more RAM (low powered, integrated graphics) then it's all setup to run it locally already :3
started the scripts with BigPickle (free in opencode, no API no nothing required):

curl -fsSL https://opencode.ai/install | bash

like, it was actually able to build the whole custom server from scratch instead of using ollama, it made scripts to convert properly the files to gguf and everything

then for couple weeks, the opencode team had Kimi K2.5 integrated for free, they ran it on their own infrastructure for test purposes and compare how it would perform with BigPickle i guess but sadly they ended up removing it (which is still just way better than BigPickle ngl)
still, they're both alright for basic tasks
the one that just ended up working the best in my case was gpt-oss-20b, didnt use much RAM, was good enough on small tasks, like, it made me a shell scripts to check on xmrchat.com tips and donation messages from the shell directly for example
but yea, for bigger projects, BigPickle should be enough, else Kimi K2.5 just next level
right now, for the V0.1-NoJS tower defense game thing, i had it made with BigPickle, it was able to implement the cryptography aspect very quickly
but now, since moved to doing a whole webgl 3d thing, Kimi K2.5 is pretty awesome on that, it can also procedurally generate 3d assets (like trees, basic characters) by using script shells through blender
so... got my xmr subscription for few days now at nano-gpt.com for their paid Kimi K2.5 instance lol
but i also got Kimi K2 Instruct (free) from the Nvidia API key
which is alright too but that latter one is super buggy from Nvidia's API being broken or somethin, it often has issues to write tools and so on, so...
well... even for the nano-gpt Kimi K2.5 instance as well sadly... which is a paid subscription....
but when opencode had it for free up for couple weeks on their infrastructure didnt have those issues... it was free... nano-gpt not, that's sad lol
so, always have to start sessions with:

read "file_writing_guide.md" at the root of current work directory, aknowledge it, then proceed with this and that task

$ cat file_writing_guide.md

Guide: Writing Large Files with Proper Tool Usage

The Problem

When writing large files (>200 lines), token limits can cause issues. The solution is to chunk the output into smaller operations.

Correct Approach

Step 1: Create Initial File with write tool

Use the write tool to create the first chunk (typically header + first major section):

{
  "filePath": "/path/to/file.md",
  "content": "# Header\n\n## Section 1\n...content here..."
}

Step 2: Append Content with edit tool

Use the edit tool to add subsequent sections. The key is using a unique oldString that matches the end of existing content:

{
  "filePath": "/path/to/file.md",
  "oldString": "...last line of existing content",
  "newString": "...last line of existing content\n\n## New Section\n...new content..."
}

Important: The oldString must exactly match existing content including whitespace.

Example from V0.5 Audit Report

First Operation - Create File:

{
  "filePath": "/V0.5d/docs/audits/V0.5a-AUDIT_KimiK2.5-01.md",
  "content": "# COMPREHENSIVE AUDIT REPORT...\n\n## EXECUTIVE SUMMARY...\n\n## DETAILED AUDIT: V0.5a...\n### V0.5a Verdict: 🔴 **FAIL - NOT PRODUCTION READY**"
}

Second Operation - Append V0.5b:

{
  "filePath": "V0.5d/docs/audits/V0.5a-AUDIT_KimiK2.5-01.md",
  "oldString": "### V0.5a Verdict: 🔴 **FAIL - NOT PRODUCTION READY**\n- Animation system has fundamental bugs\n- State machine is fake (just if-else chains)\n- No proper animation completion handling (onFinished doesn't exist)\n- yOffset calculation mathematically wrong\n- No skeleton rebinding",
  "newString": "### V0.5a Verdict: 🔴 **FAIL - NOT PRODUCTION READY**\n...same content...\n\n---\n\n## DETAILED AUDIT: V0.5b\n...new V0.5b content..."
}

Continue for Each Section

Repeat the edit process for:

• V0.5c section
• Comparative Analysis
• Conclusion
• Recommendations

Key Tips

1. Chunk size: Keep each write/edit under 200 lines when possible
2. Unique oldString: Use section headers or verdict lines as anchors
3. Exact matching: oldString must match exactly (including newlines)
4. Escaping: Escape special characters like \n, ", \, etc.
5. Verify: Check file exists and has expected content after each operation

Common Mistakes to Avoid

❌ DON'T: Try to write entire large file in one write operation (token limit)
❌ DON'T: Use replaceAll: true when appending (will replace all occurrences)
❌ DON'T: Use generic oldString like "---" (may match multiple places)
✅ DO: Use unique section headers or verdict blocks as oldString anchors
✅ DO: Verify file content after each chunk
✅ DO: Use read tool to check current file content if unsure

Troubleshooting

If edit fails with "oldString not found":

• Use read tool to see actual file content
• Check for extra whitespace or special characters
• Copy exact string from read output

If file gets corrupted:

• Start over with write tool
• Or restore from backup

Tools Reference

• write: Creates new file or overwrites existing (use for first chunk)
• edit: Replaces specific text in existing file (use for appending)
• read: Check file content when troubleshooting
• bash: Use cat or wc -l to verify file

---

This guide written after successfully creating V0.5-AUDIT_KimiK2.5-01.md (699 lines, 26KB)