Coding assistants
category: code [glöplog]
I tried one of the tools smash linked to up there, and it generated a glsl noise function that is taken from Shadertoy with variable names matching exactly. We also have know that OpenAI does copy&pasting content directly from actual Shadertoy examples.
I think the shader repo we have there is too small for these models to learn anything really, they just memorize and maybe compress the dataset a bit. Maybe that that explains why we haven't seen any AI generated shaders published in Shadetoy. Yet.
With C++ I imagine my experience is similar to everybody else's? - the AI copilots only slow you down and become a net productivity loss (work to spoon-feed tasks and requirements over actual output is less than 1). We'll have to wait for a paradigm change where models don't just take text and find structure in a latent space, but actually give them analytical thinking capabilities?
I agree that LLMs are great to quickly extract copy&pasteable code from stack overflow without actually visiting the page, which does bring a 10x speed up if you are writing a script in a language you are not familiar with. For like 2 days; after that it's all diminishing returns again.
The thing is that those returns don't dimmish if your job is to actually write boilerplate and glue subsystems/services together every single day (a React grid with a Graphql REST API and Stripe payments). That's the type of software engineering most VCs and technocrats interact with, I think that explains their otherwise ridiculous claims. Again, for now (all might change next month).
I think the shader repo we have there is too small for these models to learn anything really, they just memorize and maybe compress the dataset a bit. Maybe that that explains why we haven't seen any AI generated shaders published in Shadetoy. Yet.
With C++ I imagine my experience is similar to everybody else's? - the AI copilots only slow you down and become a net productivity loss (work to spoon-feed tasks and requirements over actual output is less than 1). We'll have to wait for a paradigm change where models don't just take text and find structure in a latent space, but actually give them analytical thinking capabilities?
I agree that LLMs are great to quickly extract copy&pasteable code from stack overflow without actually visiting the page, which does bring a 10x speed up if you are writing a script in a language you are not familiar with. For like 2 days; after that it's all diminishing returns again.
The thing is that those returns don't dimmish if your job is to actually write boilerplate and glue subsystems/services together every single day (a React grid with a Graphql REST API and Stripe payments). That's the type of software engineering most VCs and technocrats interact with, I think that explains their otherwise ridiculous claims. Again, for now (all might change next month).
For demos, I don't know.
I've never used or understood autocomplete myself (AI or traditional), but I don't care if others use it. But when it comes to discovering or inventing a cool rendering algorithm or trick or technique, if the _that_ part was made by the AI then I might not be particularly interested in seeing the demo any more than seeing two machines play chess to each other.
But then, I would still be interested in the algorithm itself and seeing it in the big screen and seeing how somebody used it to make something cool with it does have value.
I don't know, I cannot make my mind.
I've never used or understood autocomplete myself (AI or traditional), but I don't care if others use it. But when it comes to discovering or inventing a cool rendering algorithm or trick or technique, if the _that_ part was made by the AI then I might not be particularly interested in seeing the demo any more than seeing two machines play chess to each other.
But then, I would still be interested in the algorithm itself and seeing it in the big screen and seeing how somebody used it to make something cool with it does have value.
I don't know, I cannot make my mind.
Quote:
Maybe that that explains why we haven't seen any AI generated shaders published in Shadetoy.
As I mentioned on a previous page, I have seen shaders on Shadertoy which seem to have come directly from ChatGPT. I'm not saying they are worth looking at or that they were worth publishing, but they are there: https://www.shadertoy.com/results?query=chatgpt
I just saw a pretty good quote today:
"AI should do tedious things for creative people and not creative things for tedious people" - Stuart Neill
"AI should do tedious things for creative people and not creative things for tedious people" - Stuart Neill
"tedious things" ... I used GitHub Copilot on a Lua source file for a project at work that involved a lot of tedious referencing to dozens of scenegraph objects. Copilot's suggestions were initially impressive - except that half the variable and object names were wrong.
It ended up taking me longer to fix that mess than if I had just written the code myself from the start. Since that day, I’ve never felt tempted to use that damn “assistant” again :D
Maybe it performs better on C codebases, but I’m not precisely interested in finding out :)
It ended up taking me longer to fix that mess than if I had just written the code myself from the start. Since that day, I’ve never felt tempted to use that damn “assistant” again :D
Maybe it performs better on C codebases, but I’m not precisely interested in finding out :)
Let's just say so far I haven't noticed any difference in productivity between me and colleagues who are using AI tools. Just like I never noticed any difference between me and colleagues who were using IDEs, prior to that. In the end I think it comes down to what tools you're comfortable working with, with very little end difference in productivity between them as long as you are comfortable. So far I'm seeing very little of the "AI revolution" some influencer-style programmers have been harping about.
Quote:
A major drawback to 3d acceleration is that too many people aren't doing creative things with it. We should be pushing the technology, as it used to be in the old days.
I'm not saying I don't enjoy many of the visually stunning, high-quality 3d accelerated productions that have come out over the past few years, but I am saying that perhaps today's coders are getting lazy.
We need more creative coding, as well as the creativity involved in music, graphics, and overall design.
Found in aa 23yo thread : https://www.pouet.net/topic.php?which=77&page=7#c6466
The GPU vs Creavity was a debate, back then. Are we at it, again (but 10000 times worse) ?
@fra
I'm not sure about AI assisted code, the topic of this thread.
But I can tell you where I stand when it comes to AI gfx. Whenever someone mentions something like "this AI situation is the same as when - photoshop / 3D packages / photography / acrylic paint / insert whatever here - was introduced" I cringe at how wrong that statement is and I puke in my mouth a little.
I'm not sure about AI assisted code, the topic of this thread.
But I can tell you where I stand when it comes to AI gfx. Whenever someone mentions something like "this AI situation is the same as when - photoshop / 3D packages / photography / acrylic paint / insert whatever here - was introduced" I cringe at how wrong that statement is and I puke in my mouth a little.
This AI situation is the same as when photoshop / 3D packages / photography / personal computers / PC game engines was introduced.
This AI situation is the same as when photoshop / 3D packages / photography / personal computers / PC game engines was introduced.
This AI situation is the same as when photoshop / 3D packages / photography / personal computers / PC game engines was introduced.
This AI situation is the same as when photoshop / 3D packages / photography / personal computers / PC game engines was introduced.
This AI situation is the same as when photoshop / 3D packages / photography / personal computers / PC game engines was introduced.
@bifat
Poor thing. I understand you. You just realized your beloved Donny was indeed partying with Jeffrey.
Poor thing. I understand you. You just realized your beloved Donny was indeed partying with Jeffrey.
the amiga is a personal computer
Quote:
Quote:A major drawback to 3d acceleration is that too many people aren't doing creative things with it. We should be pushing the technology, as it used to be in the old days.
I'm not saying I don't enjoy many of the visually stunning, high-quality 3d accelerated productions that have come out over the past few years, but I am saying that perhaps today's coders are getting lazy.
We need more creative coding, as well as the creativity involved in music, graphics, and overall design.
Found in aa 23yo thread : https://www.pouet.net/topic.php?which=77&page=7#c6466
Said by who...?
The GPU vs Creavity was a debate, back then. Are we at it, again (but 10000 times worse) ?
^ said by who?
Quote:
the amiga is a personal computer
That's very true, and that's precisely why it's not a problem. Ask Atari people what they thought about Amiga in 1989. A machine packed full with high-tech hardware acceleration, and a scene of lazy and incompetent coders :-)
Sorry Gargaj I don’t understand your question regarding the post I linked and the quote I made of this post :(
Quote:
That's very true, and that's precisely why it's not a problem. Ask Atari people what they thought about Amiga in 1989. A machine packed full with high-tech hardware acceleration, and a scene of lazy and incompetent coders :-)
Of course, as we all know, the Amiga came embedded with all the demoeffects ever known to men, and coders only had to prompt. Just like how the photographic cameras come embedded with all the photos and paintings in history.
If we dismiss the upper 2 ironic sentences, you're actually making my point rather nicely: proprietary hardware = proprietary software = photography vs painting = 3D packages = 3D acceleration = game engines. All that stands on one side, and current LLM AI ripoffmachine on the other. Why would anybody in his right mind conflate the two is beyond me.
Quote:
This AI situation is the same as when photoshop / 3D packages / photography / personal computers / PC game engines was introduced.
"If I had asked people what they wanted, they would have said faster horses." is somehow always used as justification for backing up any startup idea, even if it's doomed to fail.
There is a fallacy here thinking that if something is similar it should be treated as *the same*. And yet, maybe it's not the same with AI as with those older "revolutions"?
It's a non-issue, you have your vote to sort out the stuff you don't like. And if that doesn't help, you can resort to the dramaz. Historically it's certainly not in the demoscene DNA that new technology is a problem, it's more like an invitation to exploit it. We take the data whereever we please, we put them into a context that we please, in a way that we please, and even give credits as we please. Those who have invested a lot in their platform, skills and comfort zones may grumble a bit here and there - and then we carry on. And besides, defiance can be a great motivator.
bitfat: sure, if you find a way to differentiate from AI slop that virtually anyone can do. The problem is, sooner or later you won't be able to. So what's your plan to exploiting AI tech and avoid being exploited at the same time?
Why is it a problem if you can't differentiate to begin with? It took me years of watching gfx to tell Photoshop+Internet material apart from DPaint crafted. For one person Photoshop+Internet might be a showstopper, for someone else it may be just a tool and outweighed in other areas. Most of my peers do cross-development and analyze their stuff in emulators. I can't tell in each and every instance, but in any case, I would a thousand times rather have my emucoding buddies in compos than trying to outright BAN stuff that got crafted with the help of tools I don't like. Just develop some personal criteria, taste, and use your vote.
Bifat is right about using your vote, in the sense that really the whole point is to enter a democompo and that then it's up to the audience to decide what is acceptable and what isn't. I think that's ultimately what ended the early-2000s hardware acceleration debate after all.
The question is whether or not any AI usage (even "auto complete") should be disclosed - so I can cast my vote in an informed way.
For the party rule - to some it's a form of plagiarism, so the discussion whether or not allow it is also a valid one.
For the party rule - to some it's a form of plagiarism, so the discussion whether or not allow it is also a valid one.
Quote:
This AI situation is the same as when photoshop / 3D packages / photography / personal computers / PC game engines was introduced.
Even game engines, are not exactly directly the easiest thing to jump into. You tell the average person to go download Unity to make his game idea into reality, some guys might be like "Wat do I do with that? Pleas help". Even not coding and learn to use Unity requires you to spend the time.
In contrast, AI services, at least the way they are promoted, are thought as star trek replication machines. You just tell them in natural language what you want and it does it for you. No skills involved. Even if in practice we are not there yet of fully asking it for complete software and just get it. But who says this is not a possibility in the future? However I don't see it as next 6 months, next 2 years or even 10. There is too much complexity in software for this to be done well with current bruteforcing of LLMs.
Maybe we will all turn into an "idea guys" nation. My demo winning over your demo at Assembly 2053. How? I will feel I was more creative at AI prompting than you and that's why, same way people type a brain fart on twitter nowadays, get thousands of likes and think they are geniuses for a day.
Quote:
You just tell them in natural language what you want and it does it for you. No skills involved.
Even this is not so tragic. What’s tragic is that you don’t evolve in the base art you’re getting into when using LLMs. You evolve in prompt-tricking the machine. It’s a whole different game. Again, sorry for returning to visual “AI art machines”, I just feel more at home in that field. It doesn’t matter whether you use a paintbrush on canvas, dpaint, photoshop, digital camera or whatever. By using all these technologies, You are learning about geometry, framing, visual balance, color theory, anatomy, perspective, paint techniques. Whether you realize it or not. By prompting? Not.
It’s amazing how people like bifat want to reduce our world to just a few concepts, derive a few rules and that’s it. Tragic simplifications. In principle, this is exactly what peddlers of magic do, but let’s leave that aside. I’m talking about how he uses “technology” as this rule-making stuff. This is how this fallacy goes: “Photoshop is technology and LLMs are technology, therefore they are the same, just on another spot on the same curve.” What a stupid notion. A mug, a chair, a toothpick, hearing aid - they are technology. So is thermonuclear weapon. One can have a problem with the nature of certain technology, not only its “newness”. And this useful idiot capitalist determinism is ridiculous. We all have to adapt to enslaving technologies because reasons. Ridiculous.
Quote:
What’s tragic is that you don’t evolve in the base art you’re getting into when using LLMs
Spot on. One young person will spend a year prompting AI, while another will stick to learning to code or do art on his own. If AI services fail, first person is left with nothing, second guy has developed useful skills in his hands. It's a shame because most young people will be tempted to cheat now there is an opportunity.
I also hear people talking "but you can also become a skilful prompt engineer" but it seems like a very invented thing to me and not something of a value after a long time playing with these things.