Halla is an infinite AI music experiment -- an evolving radio station where every track is generated by artificial intelligence and has never existed before. No playlists. No repeats. Just an endless stream of sound that learns and improves from how people listen.
Think of it as a radio station from the future -- one that writes its own music, notices what you skip, and gets better every day.
The system runs on a loop. Prompt cards -- short text descriptions of mood, genre, tempo, and texture -- are fed into an AI music generation model running on GPU. Each card produces a unique track. Those tracks are streamed live to listeners.
As people listen, we observe anonymous patterns: what gets played through, what gets skipped, how long people stay. That signal feeds back into the system, shaping which prompt cards survive and which evolve into the next generation.
Halla uses two ideas from machine learning, explained simply:
Thompson Sampling is how we pick what to play next. Imagine each track has a score based on how well past listeners responded to it. But instead of always picking the highest-scoring track, we add a bit of randomness -- giving newer, less-tested tracks a fair chance to prove themselves. This balances exploration (trying new things) with exploitation (playing what works).
Genetic Evolution is how we create new tracks. The best-performing prompt cards "breed" -- we mix and mutate their descriptions to create offspring cards, which are then generated into new music. Bad performers die off. Good ones pass their traits forward. Over generations, the music improves.
Synthwave, cyberpunk, and retro-futurism. Neon-lit highways and chrome reflections. The default channel -- high energy, forward momentum.
Dark ambient, deep drones, and space. For when you want to dissolve into nothing. Slow, immersive, hypnotic.
Warm downtempo, lo-fi textures, and golden-hour vibes. Softer edges, gentle rhythms, a late-night companion.
Aggressive, phonk-influenced, and intense. Heavy bass, distorted edges, raw energy. Not for the faint-hearted.
What we track: anonymous listening patterns. Which tracks get played, how long, and what gets skipped. We count active listeners and session lengths. All of this is aggregated -- no individual tracking.
What we don't track: your identity, your location, your browsing history, or anything personal. No cookies. No login. No fingerprinting. We genuinely don't know who you are, and we don't want to.
The data exists for one reason: to make the music better. You can see our live analytics on the stats page.
For the curious:
We think you should know what this actually takes to run.
The server that streams music to you draws about 3 watts — less than a phone charger. It runs 24/7 on an AWS t3.micro for $8/month.
Music generation happens in batches on a rented GPU — a few hundred tracks for the cost of a coffee. Each track is generated once and played thousands of times. The algorithm gets smarter, so over time we generate fewer duds and waste less compute.
Total operating cost: roughly $20/month between batches. About what you'd spend on a streaming subscription, except this one writes its own music.
Halla isn't trying to replace musicians. It's exploring sounds that don't have a human equivalent — a bulldozer on a frozen lake quantized to a beat, throat singing layered over drum and bass, the electromagnetic field of a dying star turned into melody.
This isn't AI music vs human music. It's a new category. Music that exists because a machine can try ten thousand things a human never would, and listeners collectively decide which ones are worth keeping.
The best human music comes from lived experience, emotion, and story. No algorithm will ever write the song that makes you cry. This is the other thing — the weird, the ambient, the experimental. The stuff that doesn't need a story. It just needs to sound like nothing you've heard before.
Halla started as a weekend experiment and became an obsession. We believe AI-generated music isn't a gimmick -- it's the beginning of something genuinely new. This is our attempt to find out what happens when a radio station can learn.