Been thinking about this: https://bsky.app/profile/jay.bsky.team/post/3micpg7z2h22g

[aparrish@friend.camp]

@cwebber i can only read that statement as an outrageous troll. or at least i hope it's an outrageous troll, because otherwise it's the most misguided thing i've ever read about AI and programming. "we also dislike AI slop. which is why we use AI to generate choreography, not dancing" "which is why we use AI to generate sheet music, not musical performance" "which is why we use AI to generate blueprints, not buildings" "which is why we use AI to generate plots, not novels" etc

[eniko@mastodon.gamedev.place]

@cwebber AI code is definitely slop

[dvshkn@social.treehouse.systems]

@cwebber For my own stuff I find code as art to be a spectrum, with different projects falling at different points on the line. It correlates heavily with the extent to which I feel like using LLMs for that project or not.

[ainmosni@social.ainmosni.eu]

@cwebber I consider AI code definitely as slop, the sloppiness just shows itself in different ways. It shows in being inefficient, in having subtle, weird, illogical bugs, in having obvious security holes.

Also, I hate that he's using the word "content" when he means any form of art.

[mttaggart@infosec.exchange]

@cwebber Paul's reply here has been living in my head.

Paul Frazee (@pfrazee.com)

As I said elsewhere I am basically radicalized about this. We fight now for personal computing and personal agency or we lose another decade to closed clouds We push now for an open internet and open models. Nobody is going to hand it to us because they’re nice

Bluesky Social (bsky.app)

I fundamentally can't understand this position. Pinning all your hopes for free and open computing on "open models," a thing that doesn't meaningfully exist, is so confusing to me.

But this does appear to be dogma for them.

[eramdam@social.erambert.me]

@eniko @cwebber the jankiness of Bluesky as a website in the past few months after the dev team has admitted to be fully vibecoding the thing definitely makes that argument lol

[rysiek@mstdn.social]

@eniko @cwebber yeah, +1 on that.

To me "slop" is about the (low) quality of it, the throw-awayness of it. Not the form.

[aparrish@friend.camp]

@cwebber the existence and grossness of AI slop is a *direct outcome* of choices made in the code that trains and does inference from AI models! the code *determines* the content! you'd think someone who has worked in social media and has thought for five seconds about the consequences of (e.g.) "algorithmic curation" would have a clue about this

[swetland@chaos.social]

@cwebber One can argue whether code can be art (I'm in the "yes" camp on this one), but doing the job well and building clean, understandable, maintainable software, even if it's not art, pays off. Building it on piles of slop... well, not so much.

I fell like if you're taking so little pride in your work and care for your craft that you outsource it to the garbage machine, maybe you should find a different calling.

I guess it's creating job opportunities for people to fix your mess someday.

[jaredwhite@indieweb.social]

@cwebber Hoo boy does that make my blood boil! 🤪

Code 100% is content. Claiming otherwise is a false distinction. It's like saying horrible awful ingredients in food is bad, that's why they only add horrible awful ingredients to skin creams. Hello! It still gets absorbed into the body ya dorks!

Slop is slop is slop is slop.

(Also I agree with you: code is a form of art. And our industry has been on a warpath to dissuade people from believing that, to society's great detriment.)

[octaviaconamore@cutie.city]

@cwebber it's definitely art, but also, I think anybody that respects their craft, whether that be coding, music, writing, or anything else that takes human skill, would be ok handing some large chunk of it to generative AI

meanwhile, people that just want the result (including money)? usually the ones in favour of genAI

[cwebber@social.coop]

@mttaggart But Attie is built on top of Claude so how is it a push for open models?

[mttaggart@infosec.exchange]

@cwebber Also that! Apparently though components of it use smaller, purposed models that are not Claude? I don't know what the thinking is here, other than an unshakeable belief that generative code must be the way to do things now, and all other reasoning walks backwards from that starting position.

[pawv@tech.lgbt]

@cwebber This is a bit of an obscure reference, but among gamedevs that use C++, you'll find that many went all-in with C++ features, inheritance, and encapsulation earlier in their careers. Then as they were forced to maintain code that spanned many many files, many layers of indirection, regret sank in. Many pulled a full about-face, especially those that had worked in C, to adopt a more C-like style in C++ (or C only), as it greatly improved maintainability. The happy medium for most seasoned developers lands somewhere in the middle.

I personally see a lot of these "ALL IN ON AI" leads as misguided as we were going "ALL IN ON C++". The consequences are something you can only truly appreciate once you're forced to maintain something so large that navigation doesn't make sense.

[wordshaper@weatherishappening.network]

@mttaggart @cwebber We don't even *have* any open models as far as I'm aware, and short of someone sitting down with Project Gutenberg and maybe a copy of Wikipedia I can't see any way we'll get one for english text, and I'm pretty sure there's *no* properly licensed corpus of code for any programming language to do even minimal training there.

Every model I'm aware of is based on theft. (I'd love to be wrong, but that doesn't seem likely alas)

[lumi@snug.moe]

@cwebber art and code are the same thing, they're the creative output of humans

and even if they were different, we still should be rejecting genai for all purposes, to stand in solidarity with affected people (and many other social, ethical, political and environmental reasons)

[cwebber@social.coop]

@mttaggart I asked the question here https://bsky.app/profile/did:plc:dyyvywontyeuaegemczcushz/post/3miei3zqook2a

[mttaggart@infosec.exchange]

@wordshaper @cwebber Exactly! That's my understanding as well. At bare minimum, the pirated indie books corpus is in almost all training datasets.

Now it's happened before that the utility of a thing is so great that courts will handwave copyright law (e.g. YouTube). But in this case, the precedent has not been established—and even once it is, legality and ethicality are two different things. I expect many (most here on Masto) will forever be uncomfortable with the original sin of large language models.

[cygnathreadbare@retro.pizza]

@mttaggart @cwebber I was really interested in genAI trained only on public domain content, but even the ones that apparently do are like:

-well we use this dataset that was made from flickr public domain and creative commons content
+okay but creative commons can (and almost always does) require attribution or limits derivative works to specific conditions, can you ensure you don't include images with these requirements?
-lol no we did the model, not the dataset

[tom@tomkahe.com]

@cwebber It sounds to me like they're trying to say the distinction is if you generate code then a human can validate the result (and/or make changes as necessary) because it's less abstract. But they also seem to be pushing the idea that it's intended to be used by people who don't know how to code, which means those people would not be able to validate and fix issues without going through the LLM again.

With art/content generation directly you'd likely get a different result each time, with feed code generation once the code is made that's what the feed uses and there is no longer the abstraction to the LLM.

I don't think I personally agree with this logic and that the LLM's "opinion" during that code generation stage is still pretty relevant to the overall output.