jj4211 comments on Major investor is 'shocked and sad' that the games industry is 'demonizing' generative AI

[–] jj4211@lemmy.world 3 points 7 months ago (5 children)

While people may be opposed even in theory to more tame things like a little code completion, there's plenty of room to very obviously notice GenAI slop.

If people use LLM to generate text, they tend to make too much text, and it shows in how offputting it is. LLM may be able to generate a modest text without notice, but people will put in a two liner and get pages of garbage back and use that.

And of course famously the GenAI textures are generally offputting. Maybe you can have 'generic metal texture' and no one will notice, but try for specific details and it generally gets caught.

It is possible that human output that is similarly crappy gets mistaken for GenAI output, but oh well, slop is slop either way. It's just that GenAI extends the slop to unbelievable magnitude.

permalink

fedilink

source

parent

hideshow 5 child comments

[–] Dojan@pawb.social 2 points 7 months ago (3 children)

While people may be opposed even in theory to more tame things like a little code completion, there's plenty of room to very obviously notice GenAI slop.

I mean there's the regular "can you really sell code you don't own" kind of thing going for it. The companies have stolen all sorts of data; voices, music, raster, vector, video, books, film. It'd be shocking if they also haven't scraped all the code that's out there on the web.

Some of that is perfectly fine to alter, and sell. A lot of it isn't. There are plenty of FOSS licenses that are restrictive in the sense that you're free to use it and change it, but you can't alter the license of it, and in many cases not sell it.

So when an LLM produces code based on that, what applies?

Then there's obviously broader problems with ex-developers turned vibe coders coming out of the woodworks talking about how they can't code anymore. I've people at my company joking about this, and the notion scares me. The idea that they've outsourced their thinking and problem solving skills to the point that they're incapable of doing now it is terrifying.

I don't know why anyone would willingly do that.

permalink

fedilink

source

parent

hideshow 3 child comments

[–] jj4211@lemmy.world 3 points 7 months ago (2 children)

Well, unless you declare AI consumption fair use, only public domain is fair game, since every single license requires at least attribution. The courts regrettably seem to be buying the line that they are merely "learning" like a human and therefore exempt from the rules. All this ignoring that if a human reproduces something they "learned" close enough, they are on the hook for infringement, and in the AI scenario the codegen user has no sane way to know if the output is substantive and close enough to training material to count, since the origins are so muddled.

I just don't understand the "real" developer to vibe coding scenario. Like, it really sucks, even Opus 4.6, at being completely off the leash. I don't understand how anyone can take what it yields as-is if they ever knew how to specifically get what they want. I know people that might be considered "coding adjacent" who are enthusiastic at seeing a utility brought to life, though usually they haven't that is not quite what they wanted and get frustrated when it doesn't work right and no amount of "prompt" seems to get the things to fix it. They long were intimidated by "coding", but LLM is approachable. Many of these folks "scripted" far more convoluted stuff than many "coders", yet they are intimidated by coding.

permalink

fedilink

source

parent

hideshow 2 child comments

[–] pixxelkick@lemmy.world 1 point 7 months ago (1 child)

I just don’t understand the “real” developer to vibe coding scenario.

Software developer of 17 years here.

For any given project, even tremendously optimized and easy to maintain, is about 90%~95% easy boilerplate code anyone can understand.

With an existing project with already hundreds of examples of how to write that boilerplate code, I can point even just sonnet 4.5 at that, give it the business rules required, and tell it "go do that, use the code base as an example" and it'll pretty much always get it correct first try, with the occasional small thing wrong that is an easy fix.

Once the LLM has an entire codebase to build off of as an example, its efficacy skyrockets.

Add in stuff like LSP feedback, linting rules, a .editorconfig like, an AGENTS.md, and it will be very effective.

Then I can handle that last little bit of 5% of actually important code, allowing me to put way more of my time and energy into the parts that really matter (security hardening, business rules, auth, etc)

I still spend 8 hours on a task but before itd be:

4 hrs boilerplate
4 hrs important part

Now its:

30min boilerplate
5 hrs important part
2.5 hrs adding in rigorous integration tests for corner cases too

Its about removing all the mental overhead of all the annoying boilerplate easy stuff, like having a lil junior dev I can hand off all the simple tasks to, so I can focus on the "real" work.

That is where real productivity shines with these tools.

permalink

fedilink

source

parent

hideshow 1 child comment

[–] jj4211@lemmy.world 1 point 7 months ago

I've seen and managed to avoid so many boilerplate heavy projects, I suppose it's skewed.

But yes, I find it good at boilerplate, but I consider that short of "vibe" coding, as even if prompting I'm doing it in specific context to avoid having to dig back into its sea of codegen to get at the important parts. I might have it spin up to a whole specific file at a time, but I'm not going to let it roll a whole project at once.

permalink

fedilink

source

parent

[–] pixxelkick@lemmy.world 0 points 7 months ago

I agree, people tend to use it very poorly.

That largely stems from it still being a fairly new tool and, to be honest, quite unintuitive how to use it well.

Theres a lot of fundamentally bad ways to use AI that feel natural due to the way an LLM creates the illusion of thinking.

For example, one of the first things you learn in prompt engineering is dont correct the mistakes of an llm, this is unintuitive but it inherently reinforces the llm to make more mistakes.

Instead you have to go backwards in the history and edit your prior statement to "pre" correct it before it made the mistake, and regenerate.

Its a subtle thing but makes a huge difference in it producing stupid useless garbage vs actually not half bad output.

Pretty much every "trick" to it is unintuitive like this, so thats why so much of what you see AI producing from people in the industry is garbage, Id estimate like 95%+ of people just straight up are using it very wrong, becoming frustrated, and producing sloppy tier output.

Which is a big waste of resources atm. More work has to go into education on how to use this stuff efficiently so its not wasting resources and slop levels go down.

permalink

fedilink

source

parent