FaceDeer

Ah, the code itself won’t really help with that.

My setup is: LM Studio running the local model Qwen3.8-27V, quantized down to Q4 with Q8 for the KV cache, on my system that gives me 128000 tokens context. The harness is VS Code with the Qwen Code Companion installed.

I opened the folder that has all the music I wanted to organize as a workspace, adding a git repository with all the actual music folders gitignored. This was the first prompt I gave the agent once all that was in place:

The folder you’re in, and the subfolders inside it, are where I’ve been putting the music that I’ve generated using various AI music generation services (and my own local Ace-Step 1.5 model) over the past year or so. I’ve stored a lot of metadata (usually in a .json file) along with the music files themselves, along with lyrics in .srt and .lrc form, cover art, and so forth - different formats, different sources, not entirely consistent. I’ve decided that it’s about time that I try to organize this all a bit better, and I’ll be enlisting your aid. :)

[ A few specific paragraphs about the details of the filesystem layout omitted]

Before I explain the format the text files are in and how I’d like them to be converted, would you like to take a look around the folders to familiarize yourself with them and write whatever preliminary memories and notes you might want? And of course ask me questions about the stuff in here if anything isn’t clear in its purpose or organization.

Then after it had looked around and had read the various different metadata formats, it asked me some questions about what data I’d like to have in “unified” metadata files for each of the tracks, file naming conventions, and whether I wanted it to do anything with the cover art and subtitle files. I answered the various questions (telling it to leave the cover art and subtitles alone for the time being) and it wrote some python scripts that ran through all the existing metadata to tidy up the stuff that fit existing patterns and create a list of all the oddball exceptions it ran across for me to fix manually. It found stuff like situations where the punctuation was different between music file and metadata file, where there were cover images with no apparent music associated with them, and so forth.

It became a back-and-forth exchange of me telling it what to fix and it finding new things that I didn’t know needed fixing. Eventually I had enough random Python scripts to wrangle that I asked it to write a GUI, which made it much more convenient for me to edit the metadata by hand wherever needed as well.

Nope, didn’t help or harm in the slightest. As I said, going back to what I was doing before, this exchange was pointless.

Alright, got the name-calling tantrum out of your system? I’ll go back to doing exactly what I was doing before, now.

Okay. So presumably you now feel happy.

What does using that word change about the practical matters being discussed here? Any effect at all?

Oh, I’m sure you wouldn’t love to look at it. :) It’s not exactly messy, AIs are good at keeping things tidy, but everything about it is ad-hoc. It was never intended to be seen my another mortal’s eyes than mine. I don’t think there are any API keys or other secrets tucked away in the repo’s history but I’d rather not risk it.

Your mess of files is different from my mess of files and the things you’d want to do with them is different from what my code does, I’d suggest getting your own agent on the task from a blank slate. That’s the whole point I’ve been arguing here, that coding agents make these kind of bespoke “home-cooked” applications a lot easier to spin up.

As I recall, my first steps weren’t even to have the coding agent write code. I explained my general goal and told it to look through the folders itself to see what work it thought needed doing. That was how I learned that almost all the cover art I’d downloaded from Udio years ago was actually AVIF format that had just been given the file extension JPG, for example - I didn’t even know that needed to be fixed.

Make sure you back up your files before turning an AI loose on them, of course. Mine’s been very well behaved (I warned it that avoiding data loss was the most important priority) but mistakes can be made.

Yes, AI generated. Use whatever word makes you happy, it really doesn’t matter to me. The key point is that it’s not music that would be on any existing catalogue anywhere. There’s no way to “look it up” anywhere. The only metadata that exists for those tracks are in the files on my own computer.

It is, in fact, all original work. Much of it has never been off the hard drive it was generated on (the stuff by local models), and when Riffusion got bought out they deleted all the stuff their users had generated so my local copies of those are the only ones left. The metadata is in sidecar .json files, except for the Udio tracks which were in .txt files that first needed to be converted to .json. All very ad-hoc stuff to begin with.

Also, it looks like Dropped Needle is aiming to make the music available for streaming through things like Jellyfin. That’s not what I wanted, I wanted to make a web site specifically so that I could control the layout and contents for when I share it with friends (many of whom are not technically savvy and so only know how to click on a web URL).

This really wasn’t a “just use pre-existing application X” situation. Even if there had been one that could handle all of that, configuring it would have taken just as much work as explaining to the coding AI what I wanted.

No I couldn’t’ve. I just had a look and that application has nothing to do with what I was doing.

All of the music in those folders was generated by me, some of it from a variety of different online services that are now defunct (Udio and Riffusion) and some of it from local music models. None of it is published anywhere, much of it’s never been off of my personal machine before. There was cover art for some of them in a variety of different formats. There were lyrics for most, but again in a variety of different formats - I had different metadata from each of the different sources and I wanted to unify them so they could be edited seamlessly. The lyrics were in .srt or .lrc format and usually needed hand-tweaking so I wanted a one-click way to open those in Subtitle Edit. I had a ComfyUI workflow that took descriptions, tags, and lyrics to auto-generate cover art for the songs, I wanted to be able to do that at the click of a button and hand-edit the image prompt if I didn’t like what it came up with. And so forth, lots and lots of special circumstances that applied to this bundle of data that are probably unique.

So I had a coding AI write tools specifically to handle all that. Super easy, barely an inconvenience.

No analogy is ever perfect, of course. That’s the point of an analogy, it’s not the actual thing.

In the case of vibe-coded apps, I think the thing that still matches up with “home cooked meal” is the fact that the app can be whatever you want it to be.

For example, I’ve got a big pile of music files I’ve accumulated over time from a couple of different sources with a couple of different styles of metadata, piled together in an ad-hoc assortment of folders. It’s an arrangement entirely unique to me. I had a coding AI create tools specifically to handle this pile, cleaning it up and organizing it, and had it build a website out of them that looked and behaved exactly how I wanted it to. It’s not the fanciest thing and the code could always use some refinement but it suits my needs perfectly. I’m not going to publish it because nobody else would find it useful, if they had a similar sort of situation to deal with they would be best served by having a coding AI whip up their own unique tool themselves.

The food analogy there is looking in your cupboard to see what ingredients you’ve got, thinking about what sort of food you’re in the mood for, and then throwing together something that fits those parameters.

Your avatar image checks out. :)

Joke’s on them, I’m almost done training FaceBot 9000.

Granted, it’s a 1.2B model and the training data is entirely My Little Pony fanfics. It’s insane and thinks it’s Princess Luna. But it’s going to dominate this benchmark.

It says more than half, so it’s a sure thing.

One’s got a four in it and the other’s got a three, the bigger one is obvious.

Utilitarianism is a tricky philosophy, it’s got a lot of sharp edges like this.

Well, this isn’t limited to indoor spaces, then.

And yet I still can’t buy a mosquito laser. Very frustrating.

I literally never give my phone to anyone else physically. And if I had it set up with a self-destruct button like this I would be even more fastidious about making sure it’s backed up than I already am.

For example, would it be considered destruction of evidence if a defendant had written it on the back of the phone and the officers assumed it was the correct PIN? That’s just a hypothetical

What a good idea! Hypothetically. :)

Why not just assemble all of Thomas’ spare parts immediately, without dismantling the original? Same end result, should be indistinguishable.

Exactly, which is why it’s IMO a bit presumptuous to say with confidence that humans are conscious while LLMs are categorically not conscious. We don’t even really know what that means.

I don’t personally think LLMs are conscious, at least not yet or not to the same degree that humans are. But that’s purely based on vibe, it’s not something I can know. We need to figure out what consciousness really is and how to measure it before we can say we know this with any certainty.