I am not too bothered about the vibecoding itself. there are plenty examples of Synth Edit and Synth Maker/FlowStone plugins that elevate them beyond what you would normally expect of either environment. my main issue with vibecoding is now any tom dick and harry can make plugins, and the issue with that is 98% of the people who will be interested in doing that, will be putting out extremely low effort shit.

The libraries and algorithms that LLM's can pull from is extremely limited and most stuff that is public domain outside of MIT and GPL license stuff on Github is decades old. usually you would have to go to specific communities like the Reaktor User Library to find the low effort stuff, but due to accessibility and ease of use of LLM's the market is already flooded with these plugins.

Just look how many Youtubers and Social Media twats have brought out plugins in just the past 6 months alone, and very few of them offer anything novel outside of their branding

 

If you can share more insight about your experience with Ai assisting you, I'd like to know more. Especially in comparison of the frontier models vs the realistic expectations of local models like Qwen3 Coder 30B 5-bit. I'm constantly probing Gemini and Copilot for perspectives and stil in process of setting things up, but it helps to have a human perspective.

For context: I just built an Ubuntu box with a 5060ti 16gb (still configuring it). My intent is to use it as a peer-programmer, or even do some of the heavy lifting/scaffolding remotely for projects on my laptop. I have a M4 mac with 64gb of ram, so I may utilize that to run agents and mcp (autogen, continue.dev, openclaw, etc) - essentially using the laptop's LLM to orchestrate and supervise the ubuntu server's LLM (actually 26.04 desktop version being used as headless). My primary dev stack is Microsoft .Net, but I'd give some time for DSP and VST too.