Coding assistants
category: code [glöplog]
It's an interesting observation, that during the course of the debate on this topic, the capabilities of "assistants" (now "agents") improved exponentially.
One of the agentic benchmarks I always had was to single shot a pixel-perfect conversion of the tube 256b intro into a web app. This failed only two weeks ago. Well GPT5.3-Codex can do it now, and I am certain Opus 4.6 can do it too, but I spent my token "allowance" on something more useful.
Not sure what the message is here; but these capabilities will not disappear and it will be very hard to lock them out.
A sane way to look at this is as another level of abstraction.
In the past, the demoscene never cared about additional levels of abstraction. It was always about making the bare metal doing something it was not made for. The AI models will not implement a novel trick for you where you can display 11 sprites in a row on a C64. They may help in one way or another, but its still up to you to conceptualize this.
One of the agentic benchmarks I always had was to single shot a pixel-perfect conversion of the tube 256b intro into a web app. This failed only two weeks ago. Well GPT5.3-Codex can do it now, and I am certain Opus 4.6 can do it too, but I spent my token "allowance" on something more useful.
Not sure what the message is here; but these capabilities will not disappear and it will be very hard to lock them out.
A sane way to look at this is as another level of abstraction.
In the past, the demoscene never cared about additional levels of abstraction. It was always about making the bare metal doing something it was not made for. The AI models will not implement a novel trick for you where you can display 11 sprites in a row on a C64. They may help in one way or another, but its still up to you to conceptualize this.
Quote:
Quote:Why do we even care about AI/no AI prods? (that's a real question, not a rhetorical one)
Just like we have cared that the person credited in a demo has actually written the code himself (and not for instance his mother).
1. Is this a real problem we run into at every demoparty? Do we disqualify people often who didn't invent their own algorithms, but instead read some paper, or (gasp) got help from someone else (or, god forbid, copied those bit state tables for marching cubes from somewhere)?
2. Modern AI doesn't have agency. It's a tool. It still doesn't do a demo for you.
Quote:
Quote:Our entire subculture substrate is not ethically sourced
And this should imply that therefore nothing in this subculture actually matters?
Not necessarily, but I wanted to point out the disproportionality of it.
From my PoV I see at least the following concerns with AI:
1. AI slop. Either the amount of it ("democratization" -- now more people can generate something that would barely pass as an entry), or deterioration of quality (now it's easier to do "something" quickly than to focus on a more thoughtful result).
2. "Unfairness". Now something that required years of expertise can be vibe-coded in a few minutes. People who hand-craft things would have harder time competing with automation. Basically, a running track turning into race track, and we all shift into OpenGL and D3D and forget the art of assembling software rasterizers.
3. Marginalization of fun. It's fun to make things with your bare hands. It's fun to spend a few days of inane amount of effort to draw a single Vulkan triangle in a C programming language using bare OS APIs without any external libraries whatsoever. But now there's no value in showing that "look, i made it myself"-triangle to your friend, who can just prompt-generate the very same thing under 3 minutes.
I don't really see too many ethical or copyright concerns as many people do here. Part of it is personal bias -- I was raised in a culture where copyright wasn't a recognized concept at all, not even legally. I understand the concept intellectually, but it's not internalized and not intuitive at all. Part of it -- you can kinda side-step a few issues there by running local models for free with 100% renewable energy.
Quote:
A sane way to look at this is as another level of abstraction.
Abstraction does not have a stochastic component. It's not only abstracting specific things 70% of the time while looking pretty abstract to the untrained eye the rest of the time
Quote:
Modern AI doesn't have agency. It's a tool. It still doesn't do a demo for you.
Did you try next-gen video production agents by any chance? I don't want to market any specific here, but those do exactly that. You type in your vision and it does the rest as an agent. It creates music, storyboard, then finally video shots. It can do all of this automatically or interactively - asking you for feedback at every step, where you can make as many adjustments you want (of course if you have enough tokens/credits).
This is of course for videos, but it's not hard to imagine to add coding component to it for real-time demos.
Quote:
Abstraction does not have a stochastic component.
Not sure what the argument is here. I could abstract coding by hiring a person to do it, and I am pretty sure there is a significant stochastic component to that.
Quote:
in fact there are just few people on the demoscene, unfortunately influential ones, that try to enforce those views on everyone else
I don’t know whether tom’s assessment is detached from reality, or in fact realistic. That’s why I mentioned that we need some kind of scene-wide survey. I gave my best to try and explain to these relentless people where I think they err. But to no avail. In fact I have occasionally been so persistent and boring, repeating myself endlessly, that I got criticised and ridiculed by some folks from “my camp”. I got overly confrontative, frustrated every now and then, not very likeable. Because I care(d). I gave it my best shot. Most of you keep quiet.
All I can do is speak for myself. The recklessness, the determination, blindness and deafness of those “AI” people, the will to reset to factory settings every morning, disregarding what was said a day before is such that I can hardly justify being in the same space with them, physical or online. So, if tom is right, and this is predominately “AI lover” scene, I’d like to know that for a fact so I can spend my time elsewhere. I’m tired boss. And that’s no big deal, it’s not “ragequitting”, it’s not drama, after all I’m a nobody, nobody will cry, it’s just that we all have limited time in this world, and I don’t want to waste mine. I simply have no interest in making, watching, or even thinking about “AI creativity” in this here space, because “AI” onslaught is all over the news and people’s thoughts too much as it is. I thought this subculture could perhaps be our “Nebuchadnezzar”, a shelter for human handicraft (and not in the stick-our-heads-in-the-sand way), but perhaps I was mistaken.
FWIW at @party I have had bad experiences with techbros trying to enter or encourage bottom of the barrel slop in our compos. That's why we have a fairly strict policy because as lead orga I do not want to see or encourage that behavior.
I *am* hoping that in the future we'll have agreed upon social norms we can all just follow WRT what counts as one's own work. But we aren't there yet, else this thread wouldn't be on lucky page #13 already ;)
I *am* hoping that in the future we'll have agreed upon social norms we can all just follow WRT what counts as one's own work. But we aren't there yet, else this thread wouldn't be on lucky page #13 already ;)
And by not encourage, I also mean "I just don't have time for this shit. You want a slop party, go run it yourself"
Quote:
Did you try next-gen video production agents by any chance? I don't want to market any specific here, but those do exactly that. You type in your vision and it does the rest as an agent. It creates music, storyboard, then finally video shots. It can do all of this automatically or interactively - asking you for feedback at every step, where you can make as many adjustments you want (of course if you have enough tokens/credits).
This is of course for videos, but it's not hard to imagine to add coding component to it for real-time demos.
I haven't, and for now I have no intention to try. Their proposition just doesn't resonate with me at all, fundamentally. I don't think in a language, and formulating art design doc into a language is pain in itself, it often is way easier to just do what you want to do by hand without translating any of that into language. And then it also removes the hard-real-time feedback-guided tinkering part of it, which is the crucial emotional part of doing art.
Also it naturally fills many gaps for you, and from what I heard from other artists, it is rather inconvenient to adjust these. But I have no personal experience here.
There are some very cool things made using AI-generated video (IGORRR ADHD video comes to mind), though.
I can envision future toolchains which utilize AI that help me with my workflow, e.g.:
- in a GLSL editor saying aloud "insert a `smin` snippet here" immediately adds its code there in a context-aware way; another example is saying "now subtract this sphere from that cube" would be just faster than editing all of that, rearranging expressions, vars, etc.
- "give me a slider that controls position of this object, and also give me a material editor and color picker for it" without navigating anywhere
- "give me a heat map of how many iterations each pixel consumes"
- makes weird beatboxing sounds "now generate a sequence pattern out of that and give me filters chains to make those sounds"
Quote:
in fact there are just few people on the demoscene, unfortunately influential ones, that try to enforce those views on everyone else
And how do people become influential on the demoscene? By organising parties, releasing cool things, running community resources, doing the work.
I'm okay with letting those people call the shots on the direction our community is going to take.
Quote:
That’s why I mentioned that we need some kind of scene-wide survey.
And this is why we don't.
Quote:
I thought this subculture could perhaps be our “Nebuchadnezzar”, a shelter for human handicraft (and not in the stick-our-heads-in-the-sand way), but perhaps I was mistaken.
From my observation, there are pockets of mutual respect from appreciation for certain aspects and skills. But these are informal and wildly variable across platforms, departments and interests. Even the realtime aspect, which I considered untouchable, I find increasingly contested. E.g. parties play recordings from demos which are basically anim players. Crazy? A reasonable number of people doesn't think so.
In extension to that, my guess is that this scene is only a loosely knit collection of subscenes, and only a minimum of tolerance can hold it together. This however may not be entirely undesirable for reasons of infrastructure efficiency and viability (e.g. parties, websites).
Quote:
I haven't, and for now I have no intention to try.
I tried, it made a better video in terms of flow (cuts) that I could do myself in minutes :'-(
Quote:
I can envision future toolchains which utilize AI that help me with my workflow
I think you are asking for the same in principle, just a bit more control.
But I get it - I would prefer more creative control as well. This would require AI tools to operate on more granular representations: single brush stroke/layer vs image, notes+synth/filter settings vs waveform, etc... and editing operations on them. The current generation is not good at it. But there are ways to train models like this, even if you don't have a database of editing patterns from professionals (probably collected as we speak BTW) - it's quite similar problem to "thinking tokens" really (hint: can be synthesized with a bit of RL). And I wouldn't be surprised if new genAI tools would go in this direction and just offer more and more control.
Quote:
From my PoV I see at least the following concerns with AI
You missed the main point that it's the programming itself which matters. AI coding in the demoscene is like giving cars for marathon runners (even though it may help less in the former case).
Quote:
I don't think in a language, and formulating art design doc into a language is pain in itself
That's why you have LVMs/VLMs now (multi-modal language + vision models).
Quote:
I'm okay with letting those people call the shots on the direction our community is going to take.
And if anyone has a problem with anything those people raise, they can simply just....leave. Or do their creative coding stuff that cannot/or chosen to not be branded as "demoscene" by the powers that be, elsewhere.
To those who are worried about compos drowning in AI slop, the same could've also been said about 3rd party game engines, and yet compo's aren't flooded with them, nor do they displace prods that are more in the spirit of the scene, in fact they very rarely make top 3 because somehow the (increasingly heterogeneous) audience still manages to differentiate.
Your friend could've copy pasted that same triangle in under 1 minute from the internet 20 years ago, or beat kasparov in chess with the help of a computer.
Pretty sure we've been on that track for decades now, and yet people still eventually become curious about how the sausage is made.
Quote:
But now there's no value in showing that "look, i made it myself"-triangle to your friend, who can just prompt-generate the very same thing under 3 minutes.
Your friend could've copy pasted that same triangle in under 1 minute from the internet 20 years ago, or beat kasparov in chess with the help of a computer.
Quote:
and we all shift into OpenGL and D3D and forget the art of assembling software rasterizers
Pretty sure we've been on that track for decades now, and yet people still eventually become curious about how the sausage is made.
Quote:
Quote:I can envision future toolchains which utilize AI that help me with my workflow
I think you are asking for the same in principle, just a bit more control.
But I get it - I would prefer more creative control as well. This would require AI tools to operate on more granular representations: single brush stroke/layer vs image, notes+synth/filter settings vs waveform, etc... and editing operations on them. The current generation is not good at it. But there are ways to train models like this, even if you don't have a database of editing patterns from professionals (probably collected as we speak BTW) - it's quite similar problem to "thinking tokens" really (hint: can be synthesized with a bit of RL). And I wouldn't be surprised if new genAI tools would go in this direction and just offer more and more control.
It isn't just control, it's also scale. I don't want it to do the entire thing for me. I want myself to do both the big picture, and also details I care about, but deliberately summon the AI to aid with some parts I either care less about, or find not a good fit for manual labor.
Also, no idea about music/video genAI, but for coding LLMs are pretty good about details (if fed properly), but are very bad about the bigger picture. So it is in principle already there, just has to be wired up properly.
Quote:
You missed the main point that it's the programming itself which matters. AI coding in the demoscene is like giving cars for marathon runners (even though it may help less in the former case).
I kinda made exactly this very point with exactly this comparison.
Quote:
Quote:I don't think in a language, and formulating art design doc into a language is pain in itself
That's why you have LVMs/VLMs now (multi-modal language + vision models).
It's orthogonal -- there's no way to feed the multidimensional time-warped object being implemented on my wet neurons directly into anything. It has to be serialized one way or another, either to spelled description in any human language, or a sequence of physical steps and inputs to directly perform on a machine with feedback. The latter is way easier to do in practice, as human language is just such a poor limited medium. (why is the entire ai field, and western culture at large, is so obsessed with it?)
Quote:
FWIW at @party I have had bad experiences with techbros trying to enter or encourage bottom of the barrel slop in our compos.
This is a new datapoint to me. If this is indeed already an issue, then I'm all in for trying to contain it for now.
azure: by abstraction I mean https://en.wikipedia.org/wiki/Abstraction
hiring someone to code something for you is not an abstraction, it's a delegation. neither is typing a prompt into a LLM. an abstraction must correct 100% of the time, by definition
hiring someone to code something for you is not an abstraction, it's a delegation. neither is typing a prompt into a LLM. an abstraction must correct 100% of the time, by definition
Quote:
Quote:This is a new datapoint to me. If this is indeed already an issue, then I'm all in for trying to contain it for now.FWIW at @party I have had bad experiences with techbros trying to enter or encourage bottom of the barrel slop in our compos.
If not already done, you might also want to add the location of @party to your dataset. =)
Interestring enough! Techbro invasion of the demospace could be an awesome topic for making a demo about it, any amount of details would be apprectiated.
Quote:
I kinda made exactly this very point with exactly this comparison.
You conceived in in the terms of fairness which is a different matter.
Quote:
100% agree here. This whole "ethical consumerism" is really a 2020s thing. And in fact there are just few people on the demoscene, unfortunately influential ones, that try to enforce those views on everyone else, while I'm sure majority doesn't give a shit.
This is gold
Quote:
hiring someone to code something for you is not an abstraction, it's a delegation. neither is typing a prompt into a LLM. an abstraction must correct 100% of the time, by definition
From the wikipedia link: Abstraction uses a strategy of simplification, wherein formerly concrete details are left ambiguous, vague, or undefined; thus effective communication about things in the abstract requires an intuitive or common experience between the communicator and the communication recipient. This is true for all verbal/abstract communication.
My interpretation here is that it is lossy by definition, and therefore it would be quite difficult to assume correctness in whichever way it would be defined.
When you use a C compiler, it can represent your code in many different ways. The degree of "correctness" is rather by convention (how about cycle accurate code)?
Compilers are usually reproducible, which is maybe more to the point. LLMs can also be sampled in a reproducible way with greedy sampling.
I can see a demo coder use "AI" in a way it helps them iterate on ideas faster without compromising on their "demo eliteness" at all. And the worst is that they'll go faster and faster and release bunch of things which will make people want to jump ship, use "AI" too, because they'll feel less productive than others.
Some may hold the fort, some may not.
Some may have troubles finding musicians for their projects, so they'll probably end up using suno, some others will have troubles figuring the design, the illustration or anything.
What makes the demoscene is not the single efforts of one person. It's groups, the collboaration between team mates… and where we're heading there's already a lot of OKish "one man efforts" and I'm afraid actually that AI will accelerate this.
Some may hold the fort, some may not.
Some may have troubles finding musicians for their projects, so they'll probably end up using suno, some others will have troubles figuring the design, the illustration or anything.
What makes the demoscene is not the single efforts of one person. It's groups, the collboaration between team mates… and where we're heading there's already a lot of OKish "one man efforts" and I'm afraid actually that AI will accelerate this.