pouët.net

Effectively maybe half an hour, but it was drawn out because of waiting for feedback from and (and i ran out of tokens in between). It debugs by compiling a version of the demo that runs on my windows machine and takes screenshots from that to inspect and debug.

In principle it should be possible to come up with a skill to instruct models to do all this, but i am lazy.


There were zero bugs that led to non-compiling or nonfunctional code, mostly design stuff. The reaction-diffusion effect was parametrized incorrectly and looked bad.



yes you can.

But can anyone do it, or will 95% of the times fail?

If it was real, everybody would just go and say "make me this complete game/demo/application the way I describe it" to the point you get complete software, same as with image generation where indeed you see make me a picture of something and you get something complete, even if it's not what you wanted, you can make it a poster and it feels complete in some ways.

But with software, I expect things to break, sometimes programmers who tried LLMs can't even make a function that is only 1/100% to work correctly without babysiting. So how much a software where hundreds of these functions and logic has to tie together?

Yes, maybe one in a rare case can one-shoot theoritically something (and a lot of these are exagerrated a bit, or it was "one-shoot" with some extra help in various ways), but is it the case that anyone can one-shoot easilly things like that? I doubt it.. and even if it was there would be something lost in the software quality among other things.

Optimus, you're not up to date. That was already possible with older models.
Also interesting to note that the most relevant demos (in the original sense: demonstrate what's possible) are done outside of this demoscene now. :-)

They do something, I tried recently,. like ok they will build some basic SDL framework with the basic loop and populate it with 4 random effects I ask it, but not one-shoot the complete demo except if you are very lucky. And then baby sit it to fix things. Also I see the code and I want to delete it and write it myself. Because it is impressive they can even do some stuff, people think it's the second coming. In the future, some programmers would still write the normal way and others using LLMs, like many people are lost today in web and frameworks and other write pure native C/C++ or assembly and that's ok.

For the general case I'm sure it's going to be hybrid for a while. You just keep on doing modules and test-driven development, and move away from being a software engineer (and code monkey) to being a software architect and supervisor.

Not unwelcome. Although AI assisted development is rather annoying and stressful - because it's so fast, you don't have an excuse for the relaxing phases of "typing in" and "contemplating" code. Code is getting eliminated from the process. In refactoring AI really excels.

For demos I currently don't see the point beyond proof of concept, but it might be interesting to watch if and when classic demos are left in the dust, or if AI is helpful to the point that they relieve the artists to do what they really want to.

Code monkey, software engineering, architect, no-code code, I see we created these categories because some people entering the tech didn't want to look like life-less nerds, they wanted to be part of the tech industry when it became successful but wanted to diferrentiate from themselves. Code monkey is very demeaning. Code at the lowest level still demands a deep thought process that resembles what people call software architect. Even if I coded low level C or assembly I would still think of what to type before typing it, making big plans in my mind or on paper of how to structure things, so there is not much difference and its a false dichotomy imho to say "I am not a code monkey, I am this higher more sophisticated and away from the way these nerds used to do it". Difference is that on coding culture a lot of people just want to be part of the club without getting their hands dirty. You never hear an artist say "Can I create art without touching a pencil, it's disgusting!" or "Can I compose music without knowing anything about music and not touch instruments, they are all disgusting for nerds and I am only a high fidelity music specialist or some title". Only coding culture have this sickness and I wish all of you get stuck to AI coding, forget your skills, so the real coders become mystic wizards like the 70-80s again.

Not meant in a demeaning way, but with proper modularization and testing you can have unskilled code producers code against specifications until they succeed, and finally succeed in finishing your projects. That was already state of the art long before AI. Now with AI the logical next step is to replace the unskilled programmers (which still cost a lot of money) with automatization.

In demomaking, it's just one more option in the toolbox. Democoding is so easy only because it's not dealing with error cases, delineation, specifications, and I/O. It's freeform, delineated only by deadlines, if at all. But in the end it also blurs the line between what's code that can be taken seriously, and animation. So if you want to excel as a human wizard, make games on old platforms, not demos.. or interactive demos.

It's true that demos are just effect after effect which makes it easier, there is not much state reliance between each part. I agree with that. And yes, demoscene is less focused on pushing the limits nowadays.

I am sorry I am coming a bit angry in here. Whatever happens happens with AI coding, afterall it's out of control. I still predict software done in various ways in the future but the "traditional" way will be still there I believe not dissapear. Maybe some things will be hybrid between two approaches.

Optimus, of course a person who is more skilled in a domain will get better results from an LLM.

Also note that there can be a lot of difference depending on the model you use. Gemini 3.5 Flash seems do better in oldschool demos than other models for some reason - maybe they had some kind of demoscenish RL environment, or maybe there is spillover from their omni model. no idea.

One thing is for certain: We are not in the decelerating part of the s-curve yet.

Its quite pointless trying to point out single imperfections right now as they are most likely gone a few model updates down the road. The better question is how to utilize these models in a way where they will augment your skills in a useful way.

So if AI didn't exist I wouldn't be able to augment my skills? It's always presented like the gift from god or something, the one that will raise the sky and if you didn't have it how the hell you could do anything? Also different models, different brands for the same exact stochastic hack.

Lol fuck AI.

That's a weak rhetoric device, a false dilemma.

Well, maybe maybe my reply looked like it was glazing AI a bit too much. In any case, using AI or not using it is a personal decision. In regard to the demoscene, the conflicts are very obvious.

I'd consider skills more a function of time and training, so no divine intervention needed if you have plenty of neurons, dedication, discipline, and examples. The same applies with AI. You can sense the teachers' school of thought in how AI work in programming languages and fields of problems, and I presume this comes from the most basic principles taught when the models were babies. I found some of them uncreative, unexciting, and stubbornly running counter to the way I solve problems. My suggestion is to "refactor with a human touch" frequently, and present the results to the AI as fait accompli to work with. It should adapt to your style.

So Opus4.8 was released last week and here is its debug to the rp2350 demoscene.

Two demos, I explicitely asked for them to be different from the others and also to win the compo.

https://github.com/cpldcpu/PicoDemos/tree/master/13_Singularity
https://github.com/cpldcpu/PicoDemos/tree/master/14_Origami

Also tried some other models. GPT5.5 is ok, but creates only very boring demos. Chinese models not so good so far.

>debug
debut

nice typo.. :D

"to win the compo"? Does the LLM understand that "Oooh,. the user wants me not to just make another demo, but make the winning demo, I must try harder and be more original"

Maybe I shouldn't reply, seems like you are joking here :)

Only half joking, because it literally did what you are saying.

Take a look at the implemenation plans (implementation.md in the folder).

>Why it wins:

Btw, if this teaches my anything, it is the astonishing speed of a 300MHz CM33. Some of the code looks scary in ways I would have never attempted - e.g. sampling with bilinear interpolation and plenty of float/int casts, or transcendal operations in the inner loop. But it still runs full framerate, so it's probably ok.

The RP2350 is seriously underexplored as a platform.

So are you going to use this to make a GOOD demo?

Thats a tricky question. First of all, are you referring to me or the creator of the demos?

However, I would also argue that each of the demos is TOP10 worthy on the RP2350 platform, considering that there are only 9 demos in total for it:

https://www.pouet.net/lists.php?which=244

(Maybe even TOP3? But "GOOD" is subjective and people will be inclined to vote machine generated demos down.)

I will try a demo that makes better use of the hardware.

A demo with SOUL, that's all that needs. It can be mediocre but has SOUL, like created by human.