[Aachen]

So I voted against, partly because I feel that the text doesn't to justice to such a multifaceted topic.

If all projects on Codeberg were human-written, that would be a big seal of quality! There are a lot of projects coming out now where someone feels like they've made a cool thing but it breaks on the first edge case and is entirely unmaintainable. Many are built on wrong principles because the prompt authors aren't savvy and don't know better. Some authors (I've encountered this in the OpenStreetMap community) even have the gall to pretend that their code is a gift that we now have to host and maintain. Aside from these quality and community issues, build and test pipelines also take a lot of compute resources, and with large growth in projects, that will increase as well. That the author of the proposal wants to avoid this makes a ton of sense to me.

But it also drives a lot of projects away because maybe you later want to use LLM assistance (imagine LLMs got twice as good as they are today), or maybe the auto-generated tests might become >50% of the code. You'd have to migrate your repository and community away from Codeberg. Why take that risk? It's just not a worry I feel that developers need. They're looking to get away from Microsoft (GitHub) for their open source project and that's the point where Codeberg should be a no-brainer

We're an open source community and so the mass stealing of all our work to feed LLM training which heats the world and causes (at least temporarily) giant price hikes in consumer hardware... there's a lot to say for more ethical training and use of LLMs, but does it matter at all if Codeberg bans LLM output? Won't we just make ourselves obsolete? Idk. It's multifaceted and this proposal seems too vague and blunt for me. That's not an option that's set in stone btw, just my current understanding of the different aspects

[microtonal]

You'd have to migrate your repository and community away from Codeberg. Why take that risk? It's just not a worry I feel that developers need.

I feel like we are already there. I am not vibecoding a lot, but I certainly use LLM-assisted coding now for most of my work. Once you go beyond vibe coding, an LLM feels a lot like a compiler from your well-specified ideas to code. So, to some extend, these restrictions feel like disallowing a higher-level programming language in a time when everyone was still writing machine code.

I understand where this decision is coming from (especially because they have been under attack by LLM bots) and it is well within their right, but the vague language adds a lot of uncertainty:

You must not share projects that mostly consist of code written by "generative AI"-tools

It's clear that this excludes vibe coding. But what about LLM-assisted coding, where the developer writes very clear instructions of what to do and carefully reviews every line? In such cases a project 'mostly consist of code written by "generative AI"-tools', but the quality is probably not worse than fully hand-written code, especially because an LLM is often more able to 'think' of all kinds of corner cases.

I have some projects on Codeberg and donate monthly, but I have to reconsider both, since with this change Codeberg does not really align with how I work anymore. The projects that I have on Codeberg are largely human-written, but certainly use LLM-assistance for new code.

At any rate, I guess I'm back to GitHub until I have found a better alternative.

[p2detar]

> It's clear that this excludes vibe coding. But what about LLM-assisted coding, where the developer writes very clear instructions of what to do and carefully reviews every line?

My impression is that they don‘t care, it’s still ”slop” for them and for most of coderbergers for that matter.

But I do exactly this. It‘s a code base that I have written prior manually, before LLMs, so now this helps a lot with boilerplate and add features. LLM-assisted is not vibe-coding, but FWIW I don‘t care what others think.

[0rzech]

I think your impression is right, at least judging by https://codeberg.org/ethical-foss/open-slopware .

[therealpygon]

A big seal of quality because it is human written? Maybe in narrow spaces but in general, I’d put money on the average developer being significantly worse at writing code with bugs than an AI. It’s codebergs infra so their choice, but it seems like an overly simplistic solution to a complicated advancing technology, in that it is a threadbare line between “vibe-coded” and “complex agent harness with human oversight and review”.

Also, I’m actually quite happy with some of the vibe-coded apps I use. Handy is an example, which started as a vibecoded app, and is fantastic for STT/dictation. It seems like a literal baby with the bath water approach.

[lelanthran]

> I’d put money on the average developer being significantly worse at writing code with bugs than an AI.

I'd put money on the average AI-coded project being of a worse quality, simply because the number of vibe-coders with no experience now outnumbers actual coders using AI-assist around 10 to 1.

I'm literally seeing companies being pitched with B2B products written by people who had never written a program before, even as a hobby!

AI on average might write better code, but there's no way AI prompted by a shoe salesman is going to be better than code written by a person with even 2x years of experience.

[YeGoblynQueenne]

>> I'm literally seeing companies being pitched with B2B products written by people who had never written a program before, even as a hobby!

Can you post some links to examples of such companies? It sounds really bad but I don't want to really believe it until I can see it with mine own eyes.

[therealpygon]

If that were the comparison I made — that a non-developer vibe-coding is better than a developer — sure, I’d see your point.

[lelanthran]

I think, perhaps, my point is that at this rate, pretty soon there won't be any non-vibed projects.

Even those devs that are currently overseeing are gradually atrophying their critical thinking skills in this space.

[therealpygon]

Definitely agree, though not sure I agree on the why (I’ve also disagreed with the conclusions of similar studies as well).

Programming is high-function job that requires knowledge in a million areas. I know I certainly can’t keep all of them in my head and frequently had to go back to documentation, or look up parameters, or how to implement that one pattern I used that one time that worked well in this situation or… The conclusion they draw is basically that “programming also made me dumber” because I didn’t remember them after working on a completely different project in a different language. I wholeheartedly agree thinking skills can atrophy with AI, but my view is that they are likely replaced by different thinking skills in other areas (e.g. improved prompting, improved structure, understanding harnesses, and so on. I used to write batch scripts often; I stopped, now I suck at it. If I paid someone else to write a batch script, I won’t suddenly remember everything I’ve forgotten but plenty people condemn AI for the equivalent result of paying someone else to do the programming.

[chasd00]

Does that mean codeberg has just guaranteed its own demise?

[lelanthran]

> Does that mean codeberg has just guaranteed its own demise?

In industry, certainly. The free-because-it's-a-hobby crowd will still be well served by Codeberg.

Much like how you can get very cheap mass-manufactured furniture, but some people still build their own tables and chairs.

[onion2k]

I’d put money on the average developer being significantly worse at writing code with bugs than an AI.

I'd take that bet if it was measured in "bugs per day" instead of "bugs per LOC". LLMs generate far fewer bugs in absolute terms, but they also generate far more code, so it probably balances out if everything is equal. I think LLMs will actually lead to more bugs just because they're usually working with poorly defined requirements (prompts) and they guess to fill in the gaps based on the averages of their training data. A lot of those guesses are wrong.

Where LLMs win though is the severity. An LLM will write buggy code in the sense that it doesn't do what you wanted it, but not so much in the sense of security or data loss bugs. They get that right more than humans because they're trained to care about that stuff. There's far more "how to write good code" content in the training data than "how to not care about security and go fast" content (because writing that would be weird!)

[therealpygon]

You’d take the bet if you could completely choose a different set of parameters? At no point did I say every line of a vibe-coded project in its entirety is better than every line of every repository created by humans, or anything of the sort.

You also go on to admit that the direct quality of the code is likely to better when written by AI — exactly the basis of what I argued — but wish to disagree with what I argued? An interesting position that I don’t quite follow.

The point was that using “human” as an indicator of “code quality” is nonsense. I get CB reasoning for disallowing “vibe-coded” apps when talking about the junk that people are constantly updating/building/deploying for an audience of none, wasting resources. Totally sensible from that perspective, but it takes a baby with the bath water approach that all vibe-coding is bad. I’m simply saying the point of view “if AI then bad, if human then good” isn’t grounded in reality and that your average developer is not producing high quality, well structured, well organized, well documented code. On average, it could be even odds as to whether that developer primarily works on Wordpress.

[Brian_K_White]

I'd rather read bad poetry written by a person than anything generated by a machine.

Even if an LLM generates "good" poetry, it did not, it mashed up elements of good poetry previously written by people, but stripped of all attribution, and stripped of all intents and insights of those original authors.

Since software has a function besides pure art, it's tempting to think of it as actually nothing but function, and so it doesn't matter how a functional thing comes about, merely that it functions. But I say there is more to software than that, and I'd rather not spend my own time and attention on things that no one else spent any of their own time and attention on.

Maybe codeberg is reacting too crudely, so maybe I don't defend them or this action fully in it's current form, but there is nothing ridiculous about the concept of a human-only filter being a seal of quality. The quality isn't the code all works the best, the quality is it's all human.

[Aachen]

I struggle with that argument. I feel the same way but remind myself that I find art stupid (the blue line on a blank canvas worth x millions because of who made it and what their story is or something) and this seems like the same thing. I shouldn't care who wrote something if it's of the same quality. Also part of hacker ethos: we judge people based on their hacking, not diplomas or skin color or other aspects (in this case: type of neural network)

There's lots of people that appreciate art so I'm not saying this isn't a valid argument for wanting code from a real craftsperson. Just to me personally it feels like it shouldn't be, yet I feel similarly about it

[Aachen]

I'm not saying you can't be happy with autogenerated software, especially for something like a dictation app where the main functionality amounts to a language selector (or calling into a detector) and calling TTSlibary.render(String text).play(). You could also just use an existing TTS app, but whatever works for you

For more complex software, so far it has been my experience that you get more predictable behavior (fewer edge-cases), performance, and relevant (but, to be fair, less comprehensive) documentation when someone thought about and built it

[therealpygon]

For software that became popular, sure. Have you spent a lot of time looking at horribly written unpopular software that far outpaces the popular ones in quantity, not quality?

I fully agree that a well-written human application can be far better organized and documented. But I’m not talking about the top 10% that most people interact with and focus on, but rather the bottom 90% that get far fewer eyes and are way more poorly written, regardless of perceived complexity.

[Aachen]

Hah, that's a unique argument. I haven't before been quizzed on how much time I spend with unpopular software. Do you have a minute? xD I think I use more unpopular software than 99% of the world population. Besides that it's part of my job to security test what are usually internal applications (every ~two weeks a new one), I also use basically only software from F-Droid and the Debian repositories, pulling only the apps from the Play Store from (semi) governments where it's unpractical to get by without them. Besides that, a lot of what I use is self-written, with number of users at most a few hundred, biasing heavily towards one. Yes, I'm familiar with hobby coding projects thank you very much :D

> I’m not talking about the top 10% that most people interact with and focus on

I’m glad we're on the same page then :)

[therealpygon]

I have all the time in the world for interesting topics of debate. I’ve equally reviewed quite a lot of poorly written tools, libraries, and APIs in the search for far less mainstream solutions to complex problems; guess our conclusions on what quality we found differ, which is the human experience.

Like I said, I’m not claiming AI outperforms in quality in all cases; it can be just as bad as human code can be. I just disagree with the idea that “human written” is a proxy for “quality code” in my experience.

[skydhash]

Human written is not a proxy of good code. But human written and published on a hosting platform (with the additional artifacts like docs, build system, tests) are generally good. While LLM may generate all of these, but they are slop.

[YeGoblynQueenne]

>> There are a lot of projects coming out now where someone feels like they've made a cool thing but it breaks on the first edge case and is entirely unmaintainable. Many are built on wrong principles because the prompt authors aren't savvy and don't know better.

It would be nice to have concrete examples of that kind of project.

[mcv]

> imagine LLMs got twice as good as they are today)

I have no idea how you'd quantify that, but they're pretty good already. At least in the hands of someone who knows what they're doing.

I share everybody's concern about many aspects of LLMs; copyright, data collection, energy use, component prices, not to mention what it does to us: outsourcing our thinking, dependency, slop. And I don't entirely disagree wanting to limit that.

And yet, I have just written a library that I was planning to put on Codeberg, but I've used a lot of LLM assistance for it, because it's simply too big otherwise. I put a lot of effort on code quality and architecture, and I don't think this counts as slop, but I do think it violates Codeberg's new rule, so I guess I'll have to look for another place.

No idea where yet; I don't want to go to Github, and I believe there were similar concerns about Gitlab, so I don't know where to go yet.

[0rzech]

> No idea where yet; I don't want to go to Github, and I believe there were similar concerns about Gitlab, so I don't know where to go yet.

Perhaps https://radicle.dev or https://tangled.org would be fine for you?

[mcv]

Is that a code repository built on top of the protocol underneath bluesky? Fascinating idea.

[0rzech]

Radicle uses its own P2P protocol - Gossip. Indeed, Tangled uses the same protocol as Bluesky - ATProto. https://forgefed.org, on the other hand, is based on the ActivityPub protocol, like Mastodon, PeerTube and Pixelfed.

[Aachen]

If you could have built it yourself but it was just too much work for the available time, then I trust that indeed the architecture will be sound and it sounds to me like it should be fine to build this way and host on Codeberg... until this amendment :/. I feel sorry that I'm part of this organisation that just shut the door on you

How's your appetite for self hosting? I've only ever hosted the predecessor of Forgejo, Gitea, and that was a breeze. Sadly I'm not aware of other open source hosting platforms, like, if you search "forgejo is confirmed" (with quotes) you find a lot of instances, but the point of Codeberg was the (presumed) longevity

[mcv]

I was hoping my library is useful enough to enough people that a community would form around it. Self hosting doesn't sound like a good way to do that.

I'll have a look at Gitea. Otherwise it's probably going to be Github. Or I'll self host until I find something better.

[cmrdporcupine]

> No idea where yet; I don't want to go to Github, and I believe there were similar concerns about Gitlab, so I don't know where to go yet.

FWIW I just gave up and went back to GitHub. Not because I like it. I hate the organization.

But having been burned with the year long diversion through Codeberg, and actually caring about getting eyes and contributions on my project, I have no appetite for playing more games here and ending up back in the same situation in another year.

I could self-host, and I'm willing to front that cost. But it would be frankly at the cost of getting eyeballs and contributions. And come with hassle to boot.

I already had a read only mirror to GH from codeberg anyways.

[saidnooneever]

you say human code is better but in reality it is not. a lot of human code is garbage. i am sure if youd look at mine you would be horrified. moritified. a lot of people just have fun with it.

what you imply is only wanting 'professional' projects on there written by professional programmers. This seems like some exclusive membership would be more adequate. no need for fees, just invite the people whos code you like to see -_-