Crossposted from https://fedia.io/m/updates@kbin.melroy.org/t/4475005

As you all might know it, Mbin is a rather small project with a small team. In fact the current size of the team is 1.5 developers (0.5 for the ones which only have partially time). While our active devs do what they can in their free time, this leads to slow feature and security updates, sometimes slow support and increasing stress for the remaining devs.

That's why I ask everyone who likes this project, has time of their own to donate and knows how to program with PHP (or is willing to learn it) to consider helping this project move forward.
You can start with simpler features and submit them as PRs or even just help review existing PRs if full commitment to maintaining the codebase is currently too much. We are aware that parts of our codebase contains more spagett than some Italian restaurants, but we will help every newcomer to find their way through the sources.
Every bit of support helps, as it also shows that we have a passionate userbase which does not want Mbin to wither away.

Well ... as I'm not very good at writing motivational texts, just comment anything which comes to mind for this issue :D If you have ideas how to improve the chances that this post reach the right people, I would also love to hear it.

---

At the same time I want to bring up another topic, as transparency is important for any public project.

Some of our devs use AI assistants for generating code for PRs and also to assist with PR reviews. Simply because without it, they would not have time to continue contributing to Mbin (which would bring our devs down to 1, making it impossible to merge new PRs). On thing here is important to know: I will not call it slop nor vibe-coding, as they are a competent developer capable of understanding and managing the outputs of the agents. Still, I am very aware of the issues of AI and the general opinion on it in the Fediverse. That is why I would like to know your constructive opinions on this matter.
Also, I guarantee you that every PR is reviewed by at least one human pair of eyes.

The git repo:

https://github.com/MbinOrg/mbin

AI is a tool. Garbage in, garbage out still applies.

I was somewhat interested until I read the last part. I personally have no interest in contributing to projects that rely on AI and I feel like others feel the same way. It results in more code being shipped than necessary and thus requires a lot more time to understand and modify, I’d rather spend that time on projects that don’t rely on LLMs.

(Relevant comment from the open source community)

It's written in PHP, vibe coded, and it's on Microsoft Github. A trifecta.

What do you mean it results in more code being shipped than necessary? If it's producing too many changes then it's up to the dev to refactor and PR reviewers to catch before it gets merged into master.

In theory, yes. But that’s usually not what happens in practice. I don’t know if that’s the case for Mbin but I have no interest in finding out. Most people that do LLM driven development submit code that looks right, but that doesn’t necessarily mean it is right. The same probably applies to reviewers when approving/merging changes, if they have way more code to review I wouldn’t be surprised if they get lazier with it, might even be subconsciously.

There are a lot more reasons why using LLMs is generally speaking not a good idea, but to keep it short I’ll just share some relevant articles/posts that go more in depth and explain it better than me:

• Agentic coding is a trap
• (Twitter/X) The state of engineering
• AI Software Development – What Does The Data Say?

I'm all for AI, personally. When using it as a programmer it's a huge force multiplier, just pay attention to what it's doing.

AI isn't a "tool" - it's a set of subroutines built on the stolen labor and property of others using exploitative amounts of energy and resources to confidently make up something that's wrong.

That's a tool in the same way a hammer is a tool - if you stole it from your neighbor, it was coin-operated and gas-powered, and it often turned into a screwdriver or a halibut midswing and then obsequiously apologized.

GIGO was true pre-GenAI, but is no longer. You can take good data, train a model on it with the greatest care, do a well-structured prompt, and still get garbage out. Honestly, a lot of boomers and older GenX embrace GenAI because they still think GIGO applies without realizing what a broken assumption that is.

Big oof on using LLM code for a fediverse project. That's not going to do any favors finding motivated devs to chip in, especially with all the social, ethical, and potential copyright issues they bring.

Thanks for those links.

I really don't get why Mbin moved to Github. Should have stayed on Codeberg imo, like the project it was forked from.

Codeberg no longer allows projects that code with AI.

The reason was simply the huge amount of downtime codeberg had at the time. We started looking to get away from github, but as the the maintainers don't have a lot of time, we never got to move

written in PHP

If I wouldn't have to work for some months I would love to rewrite it in Kotlin to be honest

vibe coded

No. Only a few recent PRs involved AI code, not the whole project is vibe coded. Also keep in mind that one dev uses AI, the others write and review by hand.

it's on Microsoft Github

Yes, and I agree that this is no good. Benti already mentioned that we had plans to migrate the repo to a Forgejo instance but didn't have time to actually do it.

It is a tool. Like I said, garbage in, garbage out. I wouldn't use it for something I don't know what the result should be.

It's also good for gathering information, papers on subjects. Were is the stolen labour in that?

where do you think the data the models are trained on comes from? The Clanker Fairy?

Also, LLMs are notoriously bad for gathering info. They make up info and sources constantly. Look at all the lawyers found to have used LLMs that made up legal citations and cases. Or scientific papers citing studies that don't exist.

They can move to codefloe

You're confusing training data with coding capabilities.

You lost me at AI, sorry. That's my issue though, not yours. Good luck to you.

There are no coding capabilities without the training data.

Newer models have access to tools themselves, they aren't just using predictions

Except, you know… the predictions on what tools to use, how to use them, and what to do with the results of such tools. Everything in an LLM is as a result, one way or another, of the training data. And they haven't been good or transparent on where they acquired that data or what licenses covered what they trained on.

Prediction on what tool to write html with?

Have you actually used these models?

Writing html?! What are you on about? No one's talking about writing html code with LLMs. That would be the biggest waste of tokens I can think of.

There's no such thing as "just a tool".

https://deadsimpletech.com/blog/no-such-thing-as-just-a-tool

Big oof on the LLM thing. Best of luck I suppose

Nice way to avoid the question lol

the fact that the SUV has ideological and behavioural consequences is something that we're not really allowed to discuss

What a load of wank

The "we're not allowed to discuss" line? I suppose that assumes a level self-censur that is debatable.

Or do you mean the part about SUV ownership having demonstrable effects of human behavior? That is true.

Seriously - did you think when OP stated:

Some of our devs use AI assistants for generating code for PRs and also to assist with PR reviews

that they were talking about html code? 😆

Yeah, those anti-people around here really do not seem to have used modern AI coding and therefore have no idea how powerful it is.

I use it to write html and CSS sheets. It's a good tool.

Yea, the claimed censorship is just nonsense

🤣

As a startup, it's saved us tens of thousands. Why not use it?

As a start up, why not just use free, pre-existing web frameworks? Why are you reinventing the wheel to create stuff that you won't be able to maintain? You're sinking resources into something that is not needed, won't have any payoff, and hamstrings all future development. If LLMs are half as smart as you seem to think they are, it would have told you that.

It's just a brochure site. There's nothing to maintain. AI code is better than using website builder apps.

Start-ups are about stories, LLMs are built for stories. Everything business related has a standard format and template. AI is great at taking project files and producing content from it.

You are so much missing the mark, don't even know how to go about correcting you.

Good luck, I guess. You'll need it.

Says a lot more about you buddy.

Sure thing pal. Enjoy your "startup".