Originally published as an Advisory Hour Substack post on 2026-08-28. This first-party site copy preserves the text, localizes public supporting media, and connects the record to its project trail.
Original post
TikTok provided inspiration for a new human user interaction idea to make sense of the abundance of music that’s spilled out of local AI systems. I find myself under the weight of abundance in hard to explain ways. I have galleries and filters, entire content libraries that are deep in the ways. When you start a system that generates new assets every two to four hours, it just changes what you see with all this content.
I’ve tried many forms of navigation. Websites, interactive arcades, galleries, filtering tools. The subject of this post is a TikTok inspired idea where I don’t have to find quality in the gallery, but rather the gallery tries to surface to me relevant content. The idea takes the “interest graph” algorithm of TikTok and just makes it local. Yes, if you’re curious, you can technically vibe-code a slop-copy of TikTok. It turns out this is reasonably useful in solving abundance.
The challenge with local AI systems is how fast you flip into this positive problem of “I have more assets than I know what to possible do with.” Pre-AI, this would be a top-shelf kind of problem. In a world of AI, there’s this sickening realization that there’s more content getting generated per second than you can possibly make sense of using.
Examples are good in trying to illustrate this problem.
I have setup an AI agent framework (Hermes) and with a little coaching the Hermes tool creates an entire album per day. It takes about that long to generate on my hardware. I don’t mind. There’s several hours of the day when I’m just busy with real world obligations. Are the tracks any good? Some of them surprisingly so.
The nuance here is that the local AI model isn’t just generating each track. It’s assembling lyrics, running the music production, cutting the audio, and generating per track artwork in addition to the album cover. It’s an infinite music box. How do you find music that you might like when you’re staring down hundreds, if not thousands, of songs? It’s the same problem with 3D assets, and videos, too. With the advent of advances in video generation where 5 second video clips take less than 2 seconds to generate, these problems will start to plague a lot more people.
The interest graph works if the content has enough “hooks” that an AI model can build a data science like annotation over your data, then subscribe an interest graph upon that annotation, and then ultimately serve up fresh picks based on those similarity scores. A deep pipeline would produce ever better results and the signals from the graph can feed right back into future asset production. Having the model build more things like the things you like. Taste automation, effectively.
The UI
The use of full color backgrounds per track was a nice solution to add visual contour to each track. Some of the tracks are plucked out of game projects, and so the featured images are more tuned to the particular game project. Other images are pulled directly from an album, such as this image below. The images were created using a pipeline setup for Z-image. It has the advantage of being free over paid offerings. Some of the images are quite good; others are weirdly AI.
MusicLab doesn’t play audio if your phone screen goes into lock mode. It will cycle tracks automatically. You can also dive into an album if you find one you enjoy. I am debating whether to broaden the number of albums on the website. If you’re just interested in the mechanic or a sampling of music, then click the link and jump into “Keep the Comet Loud” on MusicLab.