Anyone else noticed the increase of Vibe coded plugins flooding the market?

[ksandvik]

LLMs could output to intermediate code (R) , such as LLVM IR. I doubt that will be common, through, biggest problem is to verify that LLM generated code makes sense and we need humans to look at it. Maybe LLMs will spew out more and better unit tests, but as most of us in the business knows, unit tests are fallible.

[Munin]

I have seen some people argue that the end game will be LLMs straight up writing binary machine code, i.e. effectively becoming a compiler themselves. Obviously at this point they probably lack the training data for this to go anywhere advanced, but I assume it's probably already possible for extremely simple things.

[jamcat]

That was me, 3 posts up.

[Kaboom75]

Ive got ai to make code to drag and drop into Blender to make procedural shaders and never makes anything that works.

[Touch The Universe]

I have seen some people argue that the end game will be LLMs straight up writing binary machine code, i.e. effectively becoming a compiler themselves. Obviously at this point they probably lack the training data for this to go anywhere advanced, but I assume it's probably already possible for extremely simple things.

Yeah, the point I’m making is that the endgame could be one massive AI synth/VSTi where you don’t really need separate synths anymore.

Within that one instrument, you could just say, “Make something like Spire,” “make something like Sylenth,” “make something like Serum,” or “make me a completely new synth engine with 1,000 new presets,” and it would just generate it.

At that point, you’d basically only need one VSTi and you’d never lack any sound you’ve ever heard — or even sounds you haven’t heard yet.

So the question becomes: if AI can endlessly generate synths, engines, presets, and sound design on demand, what is the point of humans making new synths?

Maybe the answer is still personality, workflow, taste, limitation, playability, and the human intention behind the instrument. But yeah, if one AI synth can become every synth, then that changes the whole game.

[Munin]

I mean technically you can already do this and current LLMs will whip up something functionally competent if not exactly outstanding (though the question is does the world really need 5000 more subtractive synths). However the prompt-based nature means that before it can create something, you still need to have something in mind to create. That is different from browsing sounds and samples in a synth or sample library and coming across stuff you like - or didn't even know you wanted! - by accident or spontaneous exploration.

And just like not everyone was interested in making their own sounds or presets/patches historically, preferring to focus on composition or other aspects instead, not everyone will be interested in making their own synths or sounds. That's been the business model of ROMplers and sample libraries for decades. Plus the time that is required for the actual generation is at the moment still too long to be viable in a creative musical context that demands immediacy and spontaneity (excluding more focused ML implementations like Synplant, of which there will be more too of course).

So yes I think you will see stuff from "tastemakers" who do something a bit differently in their own creations to stand out.

[D-Fusion]

I have seen some people argue that the end game will be LLMs straight up writing binary machine code, i.e. effectively becoming a compiler themselves. Obviously at this point they probably lack the training data for this to go anywhere advanced, but I assume it's probably already possible for extremely simple things.

Yeah, the point I’m making is that the endgame could be one massive AI synth/VSTi where you don’t really need separate synths anymore.

Within that one instrument, you could just say, “Make something like Spire,” “make something like Sylenth,” “make something like Serum,” or “make me a completely new synth engine with 1,000 new presets,” and it would just generate it.

At that point, you’d basically only need one VSTi and you’d never lack any sound you’ve ever heard — or even sounds you haven’t heard yet.

So the question becomes: if AI can endlessly generate synths, engines, presets, and sound design on demand, what is the point of humans making new synths?

Maybe the answer is still personality, workflow, taste, limitation, playability, and the human intention behind the instrument. But yeah, if one AI synth can become every synth, then that changes the whole game.

I would much rather support a human company than buying a prompted cash grab from a new unknown company.
You don't know if that new company knows anything about coding and can fix problems if there are any issues or if they even exist after 6 months or a year.

Sound design is also a part i enjoy doing and i usually find it to be much faster building my sound from scratch than it is to go thru 1000s of preset in hope that i find the right sound that will sound exactly how i want it.

[ksandvik]

Also, these 'unlimited' plugins just cause a lot of stress due to all the options, compared with loading up one quality SW synth, learn it inside out and just make music. That way the early eighties with the cheap analog synths produced -- imho -- the most innovative music concerning synthesizers. Let's say all the early day Depeche Mode music with cheap mono synths.

[Funky40]

So the flood continues. Actually, it looks like it even just has started.





And i started today my own endevour into vibe coding.
literally have to start at ZERO. I even never used a chatbot myself on my end.
i even had to google that, haha.....and had to ask first which one would be best for my task.
But i just kept staying with ChatGTP. Since it can make me a simple delay by itsself


the experince was so far funny on one side. And horrible on the other.
endless completly unnecessary blahblah. Endless recapping the whole thing.
....reading, reading, reading....for what could have been said with very few lines.

But it also gave me some insights in several directions.
I hope i will end up with my very own, -very basic- delay. But if things go well will it outperform any available delay plugin. Cause they all have something NOT right when it comes to creative jam usages.

The fun part of day 1 chatting was: exactly what all these delays are missing seems to be easy as f*ck to make that right. It`s here just about parameter value lin-and delinearisations.

I`m on it. On the funky-Jam delay.
need to update another mac so i can load Xcode.


strange experience so far. talking to a PC that feels like talking to a 7 year old that pretends to be a professor and handles you as if you were that 7 year old........

[Touch The Universe]

So the flood continues. Actually, it looks like it even just has started.





And i started today my own endevour into vibe coding.
literally have to start at ZERO. I even never used a chatbot myself on my end.
i even had to google that, haha.....and had to ask first which one would be best for my task.
But i just kept staying with ChatGTP. Since it can make me a simple delay by itsself


the experince was so far funny on one side. And horrible on the other.
endless completly unnecessary blahblah. Endless recapping the whole thing.
....reading, reading, reading....for what could have been said with very few lines.

But it also gave me some insights in several directions.
I hope i will end up with my very own, -very basic- delay. But if things go well will it outperform any available delay plugin. Cause they all have something NOT right when it comes to creative jam usages.

The fun part of day 1 chatting was: exactly what all these delays are missing seems to be easy as f*ck to make that right. It`s here just about parameter value lin-and delinearisations.

I`m on it. On the funky-Jam delay.
need to update another mac so i can load Xcode.


strange experience so far. talking to a PC that feels like talking to a 7 year old that pretends to be a professor and handles you as if you were that 7 year old........

Go go mario! It's rare to receive any encouragement, so let me be the perhaps only one to encourage you to do so on this site. Go at it, it's amazing once you get the gist of it. Just knowing what tools to build , juce, xcode and gpt is the hardest part.

I made a thread in the plugin development forum tracking my recent journey into it that might benefit you. I made it for partly that purpose.

[Tiles]

So the flood continues. Actually, it looks like it even just has started.





And i started today my own endevour into vibe coding.
literally have to start at ZERO. I even never used a chatbot myself on my end.
i even had to google that, haha.....and had to ask first which one would be best for my task.
But i just kept staying with ChatGTP. Since it can make me a simple delay by itsself


the experince was so far funny on one side. And horrible on the other.
endless completly unnecessary blahblah. Endless recapping the whole thing.
....reading, reading, reading....for what could have been said with very few lines.

But it also gave me some insights in several directions.
I hope i will end up with my very own, -very basic- delay. But if things go well will it outperform any available delay plugin. Cause they all have something NOT right when it comes to creative jam usages.

The fun part of day 1 chatting was: exactly what all these delays are missing seems to be easy as f*ck to make that right. It`s here just about parameter value lin-and delinearisations.

I`m on it. On the funky-Jam delay.
need to update another mac so i can load Xcode.


strange experience so far. talking to a PC that feels like talking to a 7 year old that pretends to be a professor and handles you as if you were that 7 year old........

The reason you listed is the reason why i recommend to use Claude instead of ChatGPT. Much more pleasant and productive. I even run a local model when i run out of tokens, just to avoid ChatGPT.

[Funky40]

I made a thread in the plugin development forum tracking my recent journey into it that might benefit you. I made it for partly that purpose.

Thank You. I`ll have a look

The reason you listed is the reason why i recommend to use Claude instead of ChatGPT. Much more pleasant and productive. I even run a local model when i run out of tokens, just to avoid ChatGPT.

Yeah, my first search about what AI-chat would be best was pointing towards other platforms. I thought so that Claude would be best for a first try. But the "first access" to chatGPT was easier.
And it looked like i had to pay for Claude to get to a usable level....?

But when you say - - - >>"Much more pleasant"<< - - - okay, that`s something. I`ll keep it in mind

....... , juce, xcode and gpt is the hardest part.

Yeah, looks so.
Thats how far we came. My next step.

Thanks ! i`ll switch over to the other Forums section with my journey........





On topic:
- i see now how big the potential of vibe coding is.

real -live-play- music-making skills can count much more than prime coding capabilities.
The aformentioned need for higher coding skills is not true for everything.


vibe coding only might work perfectly well for more simple technical functionlities.
"functionalities" where the AI-assistent has enough infos on the table to grasp it to 100%
And: Much is plain math. parametrisation setups and finetuning is plain math.


I hope since devs can code now faster, that it will lead to more and deeper collaborations of skilled musicians with AI-assistent code and GUI savy folks. That could be something

[Tiles]

And it looked like i had to pay for Claude to get to a usable level....?

There is a free plan, but it's usually half an hour with the weakest model, and then you have to wait again. The Pro plan is better, but even here i run into the limit regularly.

With ChatGPT the free version is always available. But i cannot even upload my project there, and 95% of the text is noise. As told, i avoid ChatGPT for coding.

[ghettosynth]

I cannot imagine doing AI assisted coding on free plans on the mainstream models. That will get old and frustrating very quickly. Gemini still has the best value proposition for this. I don't know how long that will last. With respect to conversational style, Claude is still the most emotionally mature of the three, but also sometimes a bit of a woke karen. All are better inside of coding harnesses.

You want a harness. I use opencode, but, claude code is very popular. Here's the thing about opencode though, it often has trials with other providers that are free for a while.

OpenCodeModels.PNG

Here you can see that some new models are free to use. There are limits, of course, but you can switch between them and there's almost always a steady flow of different models to use.

All of the providers have a coding harness like this Claude Code, OpenAI Codex, Gemini CLI. AFAIK, OpenAI is still the only provider that you can use your subscription time with opencode without violating TOS, which, can get your account banned.

The other thing about opencode and their Zen service is that you can pay $20 and then use that for credits with different models from different providers.