> I've seen "vibe coded" projects before and couldn't see any clear indicators of AI-generated code of any mod. "vibe coded" projects are mud pile - https://en.wikipedia.org/wiki/Spaghetti_code#Big_ball_of_mud.


This isn't remotely true. You're living in the past. I use Claude Code daily. I'm not claiming it makes no errors, but it absolutely does not create the kind of code you are suggesting.



> I liked making mods. I'm passionate about open source and good code.

Blatant contradiction. Opus, Fable, Sol, etc. are writing most of the code being contributed to open source projects these days, and most of it is of very high quality.



https://github.com/Fiztba/RustMUD

The project is roughly 90,000 lines of code at last check. Divided into 85 different files, that are thematically structured, and also inside of subdirectories that break them down based in to what kind of purpose they have.

Yet, this is what you claim is: "A big ball of mud that lacks a perceivable architecture. "

Maybe you've only seen code written by old or cheap models. I pay $200 a month for a Claude subscription, and it makes me far more than its cost back in return. Over 100x that each month in fact.

Edited 2026-09-01 01:51:25

https://github.com/Fiztba/Wayfarer

I wrote this with AI as well. Let me know if it lacks structure too.


Both of those projects were initially started by Fable, and later completed by Opus 5 using Fable as its advisor.

Stop looking at slop written by Luna or Flash 2.5 and writing off all AI-generated code as being of similar quality. The frontier models produce code better than most software engineers, and at a pace faster than they could ever dream of matching.

Edited 2026-08-30 15:54:34

Unless you actually know how to program, you have absolutely no say in engineering decisions. You are using an AI to entirely write the program. This demonstrates total lack of understanding.

I'm not going to go through the entire source code to find structural issues, but have found these in https://github.com/Fiztba/Wayfarer :

https://github.com/Fiztba/Wayfarer/blob/master/src/main/SessionManager.ts

https://github.com/Fiztba/Wayfarer/blob/master/src/main/telnet/TelnetSocket.ts

The first uses a massive array in a loop. Maybe they are better as object properties, rather than completely disconnected even though they appear to have a direct relationship with an object.

The second uses comments to group constants, rather than using a TS enum or regular JS object property list.

I liked making mods. I'm passionate about open source and good code.

Blatant contradiction. Opus, Fable, Sol, etc. are writing most of the code being contributed to open source projects these days, and most of it is of very high quality.

The models are actively harming open source by not following open source license conditions, if you read the Arxive study that I posted on the page before. It's also worth noting that curl specifically ended its bug bounty program because they were getting a flood of worthless AI security reports that were already fixed or don't actually exist and the AI operator didn't validate the claims. The Linux kernel has kicked out support for older hardware in hypothetical issues that AIs reported, even though the drivers are stable because they don't have the time to validate the claims AI makes. Linux is meant to make hardware get a new life, not continue with what's essentially planned obsolescence.

The code that the models produce likely isn't copyrightable anyway, so neither you or your company owns the code: https://fsfe.org/news/2026/news-20260825-01.en.html . There's a precedent of it being public domain, which allows for licensing washing like with the chardet Python project that I also linked on the other page.

If models were only trained on source code that use the Unlicense, 0BSD or any other public domain license that don't carry attribution conditions, that would be a lot more reasonable. Theoretically models aren't even allowed to use MIT licensed code because they fail to follow through with the attribution requirements.

@Dan: neither example demonstrates a structural defect.

1. In SessionManager.ts, the “massive array” contains 13 subscription names. The loop sends the same request for each name. That is exactly the kind of operation an array and a loop express clearly.

These are names of variables to request, not the values of a character’s health, mana, or room. Calling them “completely disconnected” overlooks their explicit shared purpose. Replacing them with object properties needs a justification: which object, what property values, and what useful relationship would that representation add? Without that explanation, this is a proposed change with no demonstrated benefit.

2. In TelnetSocket.ts, grouping constants into enums or objects is a reasonable option. It is not a requirement that makes module-level constants grouped by comments structurally wrong.

The constants already have meaningful names and, where appropriate, prefixes such as CHARSET_, TTYPE_, and ST_. The file also already uses an OPT object with as const for Telnet options. An enum could provide an additional benefit if used to constrain the parser state’s type; an object could provide a common namespace. Those would be specific arguments for a change. Merely pointing out that comments group constants does not establish a defect.


You’ve suggested how you would organise the code. You haven’t demonstrated what is wrong with its current organisation. If the alternative improves type safety, readability, or maintenance, explain specifically how. “I would have written it differently” is not, by itself, a finding of bad design.

------------
That was GPT-6 Astra BTW.

Edited 2026-09-07 08:58:44

> Unless you actually know how to program, you have absolutely no say in engineering decisions. You are using an AI to entirely write the program. This demonstrates total lack of understanding.

No, Dan. I got my Master's Degree in Computer Science in 2013. I work as a senior software engineer for a Fortune 500 company, and have since long before AI was capable of outcoding you.

Using AI for things it is incredibly capable of, is not demonstrating a lack of understanding. You believing AI writes slop still does.

It is not a massive array, it is 13 protocol variable names.

The second file groups Telnet options into an exported OPT object with as const.

> There's a precedent of it being public domain

There is not. That precedent is for AI-generated art. There is speculation about code, not precedent.

Edited 2026-09-08 02:30:00

With my own skill, I have hit a wall a few years ago where I simply lacked the intelligence to work simultaneously on Kubernetes, Ansible, Spring Backend, complex business processes, Angular frontend and much more all on my plate to jump around. If I was an enthusiast pure frontend dev, I could jump on each new UI framework but I work on the whole stack and having learned many UI frameworks after work which are now outdated, I am tired. AI is the magic I need.

AI is better at me at coding. It is not close, it is a slaughter. I am not crazy elite but top tier in my profession, one of the few senior devs who develops software outside of work. I have also peaked in my skills. IDK if AI is improving exponentially but clearly the AI in a few months will again be vastly superior to the current AI.

Edited 2026-09-08 03:44:06

SE is business, not art where people pay me to paint with my feet. All that counts is
- features
- bugs
- delivery speed
- maintainability
(- communication)

If I have a tool which helps me on all those axes (especially first three) to an absurd degree, it would be some kind of malpractice to not use it. People like me are serious cost factors. If my pride is hurt by people without my skill set being able to develop awesome apps, nobody cares. I am paid to deliver, not to gatekeep the normies.

Edited 2026-09-08 04:44:48

This video describes the split in AI being used for code generation in open source projects: https://youtube.com/watch?v=UudhdvY946c .

From someone's blog: https://antman-does-software.com/i-will-never-use-ai-to-code-or-write .

Academic papers:

• https://arxiv.org/html/2503.06327v2
  
• https://arxiv.org/pdf/2605.24521
  
• https://arxiv.org/pdf/2605.02741
  

It's also worth reading their references.

QEMU also rejects AI code: https://www.qemu.org/docs/master/devel/code-provenance.html#code-provenance .

I'll just leave this link here: https://techcrunch.com/2026/10/02/pope-leo-xiv-is-not-a-fan-of-ai-generated-art/

Leo is probably buthurt because of all the AI memes someone made about him. I deny all the rumors that it was me.

Anyhow, who cares about krinid when there are bigger fish to fry?
https://war.app/Forum/897130-mulis-userscript-update

> I'll just leave this link here: https://techcrunch.com/2026/10/02/pope-leo-xiv-is-not-a-fan-of-ai-generated-art/

Why the fuck should anyone care about that? I'm an atheist, is a cult leader opposing something supposed to scare me away from it? And yes, that's what I view religions to be. Cults.

Your academic papers are also obsolete.

The first is from March 2025, second and third May, 2026. That March 2025 article relates to today's AI about as much as an article from 2000 about Pentium processers would relate to modern hardware. Yes, that is really how fast this field is advancing.

The May articles are less outdated - they are still quite outdated.



Opus 4.7 on the right there was the best model in the world in May. It gets completely smoked today by Argon, Astra 6, and Opus 5.5, the current best models from Google, OpenAI, and Anthropic.

> IDK if AI is improving exponentially but clearly the AI in a few months will again be vastly superior to the current AI

Possibly, arguably probably, but not necessarily true. No one knows for certain when the current crop of ideas will stop producing improvements. When new ideas are needed progress will likely hit a wall. Will AI itself provide those new ideas? Again, no one knows for certain. There's an assumption/fear that an AI that can improve itself will rapidly evolve into some god-like entity, but another possibility is that an AI that can improve itself gets diminishing gains upon each iteration of improvement until it can no longer improve itself.

@Kurtosis: Given that a non released OpenAI model codenamed "Bell" has single handedly demolished the field of mathematics from the pinacle of human intellect towards something like chess where you watch chess players for fun, knowing that Stockfish will beat them easily, I would give a high likelihood to better AI arriving. And at the end of the day, does it really matter how good Stockfish is, apart from you having fun with removing its starting pieces until you can win against it?

AI is already better than humans, it is just not on the headlines because humans believe that some royal baby or random Trump tweet is more important.

Edited 2026-10-15 11:32:47

Well I did say "arguably probably"

"AI is already better than humans, it is just not on the headlines because humans believe that some royal baby or random Trump tweet is more important."

Believe me, I've been paying attention. There's been a gulf for years between the behind doors top of the line frontier research models and what is available to the public for purchase and further between that and what is available to the public for free. The free AI is what most people make a comparison too, and as anyone with experience with makes plenty of mistakes. To make an imperfect human analogy, judging AI's by the free versions is like judging human intelligence by a high school dropout.

That said there's a distinction between AI being better than humans, to AI being better than all of humanity, and further to AI being better than the union set of AI and humanity. I don't think human mathematicians are done yet. For starters they have to verify the results released by OpenAI are not "proof by intimidation."