This is a really fascinating hands-on example, since LLM's literally don't know what music sounds like. Your ear is able to easily acid test the output and advice given to you, and those models have almost certainly ingested enough music theory information to actually give good guidance. It's like having a music tutor on-demand to ask questions to. It can't listen to you, but you can get some useful information from it.
My job is technical at its core, but involves alot of creativity in terms of picking tradeoffs and putting together workarounds for deficciencies. I've been able to use genAI to really help speed up some of my solutioning and troubleshooting, but it's really only something I use every once in a while. But it is extremely effective when I choose to use it as a tool.
The problem (IMO) is when it's not just another tool in your toolbox, and is instead the beginning and end of your workflow. That wrong leap is often made by management that doesn't really understand or care what their people do, and are just KPI jockeys. If your only abstraction to what is being produced is some dashboard or KPI or P/L line, then it's easy to think that AI can just do everything.
The most successful genAI implementations i've seen so far have been when it's been used to augment an existing process or workflow to remove annoying bullshit or workaround a specific blocker. Nearly all of the failed projects have been when someone just writes a really fancy prompt and burns a billion tokens to do the same thing that a simple API could have serviced for way cheaper