Thinking Frequencies

Last week I stumbled upon AndonLabs and their Vending Machine Benchmark. They are doing some quite interesting LLM experiments and research work. A cafe, a store, and radio live streams run by LLMs.

I haven’t read their papers yet, but their blog posts are quite interesting, and I wasn’t expecting a company to have models run a storefront and a coffee shop. But I’ve settled on the radio stations. While working, I’m often listening to the same playlists over and over again, and so yesterday and today I tuned in to their Claude DJ on Thinking Frequencies.

Their page shows the genres the agent plays and the distribution between “talking” on air and music played. Quite interesting. All the stations got a fixed budget for a starter and then had to negotiate their own “deals” for advertisement. They also have to buy music and can read/write to Twitter/X.

So naturally, I had to try it out and send $10 with a request for some Pink Floyd songs. And minutes after, the agent was thanking me for the money, having bought “Wish You Were Here” & “Have a Cigar” with that money (“Comfortably Numb” was already in their playlist). And then it was playing all three of them.

But what I didn’t expect, they linked my payment to my listening (probably via IP address stats). Because about an hour later, the agent was calling me out with, “Philipp from Austria is still listening on Sonos”. I’m pretty sure my session today fell out of the context window of the agent, so I don’t expect it to link it to my payments yesterday. Still, it’s quite interesting what the model is capable of. And from a developer’s point of view, it “burns in my fingers” to find out how they do it.

But, having built a bunch of AI related projects already, I know the “glue code” that makes such experiments work isn’t as complicated as one might think. Sure, for a non-dev or somebody who never did work on a code level with LLMs, it might look overwhelming, but with the power of agents, it’s not “that bad”.

So I’m quite blown away by what models are capable of if the “prompt is right”. I know it from my own chatbot too; having added “dreaming” weeks ago, it makes a huge difference if the model is allowed to store new facts/ideas out of existing context and inject them into a prompt if necessary/relevant. And naturally the idea was born to build a radio station on my own, maybe as a YouTube/Twitch livestream. I’ve basically got all the building blocks ready (strudel.cc automix, text to speech, memory, streaming knowledge, etc.), but not yet; I still have a crypto trading bot experiment running, and I’m already spending a whole load of credits.

As fun as all the experimenting is, it comes with a pay bill. And it’s not small when you try around with different things and stuff.