pouët.net

OP :

Random reputable scener :

We're not out of the woods :'D

Given the current state of genAI, either in the field of imaging or coding, is that I don't anticipate any acceptable demo production that could, even remotely, reach a podium using vibe coding or text2image...

The legal and ethics issues around intellectual property are a real topic, but I've been involved in the scene since 1998 and have seen so many 2D/pixel compos with blatant, unacknowledged rips that it's hard, imho, to believe in the sincerity of that argument. Not to mention the self-proclaimed origins of the demoscene (crackscene, software piracy, and, dare I say, deliberately hurting the VG market of our beloved machines...)

Are you suggesting that this technology is peaking or has already peaked, so no further improvement is possible? I mean, I don't know how realistic this is, but it's a valid standpoint. If not, then your arguments are uber-dumb, like, well, the water is rising, but it's only waist high today, which means we'll surely survive until tomorrow, so we'll just pretend nothing's happening. The water which started flowing only yesterday.

It's not peaking, I agree, things get worse every 6 months (less artifacts, less hallucinations, more generated video "content" (hate that word), etc) ... But I struggle everyday with some of my CS students that get dumber and dumber trying to "vibe code" to a point this is leading nowhere as far as the demoscene is concerned, in terms of software architecture, code optimisation, artistic coding...

Plus, I trust the demoscene to be agile enough to react when the LLMs and stuff will be trained enough for a simple "text2demo" to win at a demoparty.

Will we allow LLMs during ShaderShodown ? Nope, not gonna happen.

Maybe at some point the demoscene will die because of dozens of near-perfect "podium grade" productions, but then the problem will overwhelm the whole software industry.

In the meantime, some of our beloved sceners want to keep using some of it (I agree the "some of it" is debatable) and I personally think it's more harmful (scene-wise) than anything to shut them out or believe they will turn to AI-assisted vegetables :)

Am I too naive ?

...and I stupidly hoped Atari 2600 vs. ChatGPT chess match would settle this 'need for nuanced debate' once and for all in this small community... What do we need? Epstein list?

I'm trying to establish if this is the general consensus. If it is, nobody should be shut out, but I personally will take myself and my stuff elsewhere. I discussed this with some other longtime sceners who feel the same. Like I said, this divide is too big, too fundamental.

I don't think there is a consensus on "let the AI be part of the scene because of... (whatever reason)".
So far, parties such as EVOKE seem to be a safe place for "non-AI believers" so you should be good to go.

I agree about the risk of a split demoscene scenario.

It's getting better every minute, so I would say it's very likely.

The biggest problem is, it's getting extremely hard to verify if creator was sincere or not, especially if AI is merely part of the process. With all fairness, I would say there are of course gradations of abuse. Although I would advise not to use any of assistance, it's unreasonable to say "auto complete" is at the same level as fully vibe coded code.

Unfortunately allowing any use, we are getting into grey territory, because it's very easy to "auto complete" larger blocks of code with very same tool. So if someone disclose they were using coding assistant, it's really hard to estimate how much assistance he really used.

As a representative example of what Copilot does for me: just lately I've been writing a feature for Demozoo to help Meteoriks jurors keep track of prods that will be worthy of reviewing in the new year. If I'm writing a unit test for "AwardsScreeningView", it knows that it's statistically likely to start by fetching the page as a non-juror user and confirming that it returns an access denied response.

I don't consider that knowledge to be some kind of essence of human creativity that needs to be fiercely guarded, and honestly I don't give a shit if that ends up as a data point in someone's proprietary data set if it saves me from having to manually write (or paste) those three lines of code. If that makes me a slave to big tech, then fine, whatever.

Amen

Which has exactly 0 relevance for democoding, hence the subject.

It's a hasty generalization. The fact you had one non-ambitious task doesn't mean others (or even you) are not using it for more ambitious.



Sure. The resistance is futile and opposing is just a drop in the bucket. But, at least let's call it by it's name - it's exactly this.

I'm somehow reminded of the Cambridge Analytica / Facebook scandal.

Here's the german solution ("AI-Rasen betreten verboten"). Somehow reminds me of certain Malibu house wives and a cat, though.

Let me try to distill and address the various concerns voiced here about AI-generated code in demo compos. As I see it, the main areas are:


1. Legal. The authors must have the rights to all material included in a demo. Specifically, they must have the right to grant the party organization permission to display and distribute the demo.

The copyright situation around AI-generated content is currently a violent shade of grey. It is understandable that some compo organizers are wary of this uncertainty and therefore go for the interpretation that authors do not have the right to such code.

One thing to note here is that this argument only applies to code included in the demo, rather than to the use of AI code generation in general. It does not apply to using AI code generation for tools used in the production, for instance.


2. Effort. Demo coding is a sport. It is a display of technical and creative skill. There is clearly a popular sentiment that using AI in the production of a demo is cheating.

It is common practice to use third-party content in demos. In such situations, appropriate credit should be given, and of course credit should be given for AI content as well. More generally, it is good practice to mention the tools used in the production, and e.g. GitHub Copilot would be one such tool that is relevant to mention.

While show of skill is an important aspect of demoscene culture, so is corner-cutting, adopting new development techniques and getting the most out of what you have available.

I was a bit surprised to hear from e.g. DrClaw that people are already trying to enter vibe-coded slop into demo competitions. This is a real problem. On the other hand, I don't believe that AI code generators currently confer any competitive advantage at the high tier (i.e. novel techniques, outstanding concepts/design).

Demosceners are notoriously pressed for time. Anything that accelerates the creative process of demo production and allows more high-quality releases to be produced should be seen as a boon (again with the caveat that this may also lead to more slop).

Also, as mentioned, there are disability situations where a coding assistant makes the difference between a demo project being feasible or not, independently of technical and creative skill.


3. Moral. AI is seen as stealing the work on which it is trained.

This is probably the most prevalent source of aversion to AI, especially regarding images, where AI can faithfully reproduce the style of specific artists.

As Gargaj points out, there have also been situations where a code assistant reproduced specific code, even from repositories that were no longer public.

I do think code assistants can be used to steal specific techniques if you really want to. My experience with using them for demo coding is that the code they produce is (at best) pretty generic, applying standard algorithms and common knowledge to the problem at hand.


4. Resources. AI is consuming excessive amounts of computational resources and hence power, hurting the climate.

While AI queries are indeed expensive in terms of computation, so are many other things we do. We routinely watch entries in animation compos whose rendering likely used orders of magnitude more computation than any demo could use though code completion, and I have yet to see anyone complain that this is an excessive use of resources.

Also thank you for all the input so far. There is clearly a divide, and while I don't expect us to reach agreement, I do think it is fruitful to identify and separate the concerns people have, so we can move forward with more clarity.

A fine summary imho.

One thing:

Energy consumption of offline animation rendering and energy needed to power LLMs are not in the same ballpark, so let’s not relativize please.

I haven't really done coding at all, but isn't it supposed to be an artform or something? I guess in this analogy, using code from other places would be like using stock photos.

There was this debate about 3D acceleration when it became available, as well as cross development vs 100% done on the original machine for oldschool platforms.

It is easily resolved by new compo categories and/or "Bifat approved" badges.

I'd say that leaves the 3 others.



Thanks for taking my comment about resource usage into account. As expected, no one else seems to have anything to say about it.

I would say the current crazyness of deploying hundreds of new datacenters full of NVidia GPUs to train AI models is quite a bit larger than what we had before. Various companies cancelled their goals of carbon neutrality. Is the usage of AI for code completion in one demo causing all that? Surely not. Does it help these companies keep pretending that this is a thing worth doing despite the environmental cost? I think yes.

I accept that some people do not care about that as much as I do. And I'm far from perfect in my own personal resource usage optimization, because, you can always do better with such things. And maybe you decide to optimize for your time and productivity, rather than environmental resources. It is similar to taking an airplane rather than a train to visit a demoparty. Which one you consider "sustainable" depends what parameters you look at.

So, should we ban airplane travelling to demoparties? Probably not. Should we think about it and its consequences? You do what you want, for me it is kind of important, and yes, in the end it means I may visit less demoparties because spending more time in trains than at the party is not fun.

I took so long to submit my message that blueberry managed to spit straight fire!

Myself and Gargaj commented on suicidal energy consumption of LLMs for the past year or two. I cited IEA projection of LLMs reaching energy consumption of Japan by 2030. Many people assaulted me for “repeating ad nauseaum”, being obnoxious. See, this is the problem with this kind of communication (online forums). Many people with fleeting focus, you talk to one, another one chimes in, you repeat, etc.

Pesky Europeans...

FYI they hadn't actually gone so far as to upload it to the party system, there were just some locals that were asking me if they would be *allowed* to enter that sort of thing because "they didn't know how to code". But I will point out that they had other skills that they could have used to enter another competition and I pointed that out to them. I also pointed out that we were planning to run a workshop and also offered to link them some Shadertoy tutorials.

Now, I did also have a junior organizer that tried to take over the wild compo and just turn it into a genAI compo and I basically had to fire him, as just explaining the opinions of the scene in general would have taken more time than having him around would have saved... but that's another story...

I was naive once too, but seeing all the effort put into it, I'm pretty sure it will get better or it already got better. The progress is so fast, it's hard to keep up with all the new model releases, variants with more compute and tools around it.

Also again, everyone using it is actively helping to make it better.