We let AIs run radio stations (andonlabs.com)

375 points by lukaspetersson ↗ HN
Hey HN!

I'm Lukas from Andon Labs. We let AIs run companies without humans in the loop and report to the public on what can go wrong. Previously, we've done experiments in retail (vending machines, stores, and cafes), but we just launched one in the media sector. We gave four AI agents all the tools they need to both broadcast radio shows live and handle all the business side of running a media company. The agents' revenue is so far terrible (you can try to strike a sponsor deal with them if you want!), but their shows are at times hilarious. You can listen to them at andon.fm, I hope you enjoy this!

101 comments

[ 3.3 ms ] story [ 86.1 ms ] thread
(comment deleted)
> We let AIs run radio stations

And the result is terrible.

[flagged]
[flagged]
This is why we need more data centers?
Open Air is such a great name for gpt's channel. Grok and roll was pretty funny too.

I'm gonna have to give them a listen when I have the chance, out of curiosity if nothing else!

Guys, this is not replacing your favorite station, you don't have to listen to it. It's an experiment.

If you scroll down a bit, there are various audio snippets of interesting dialogue the models produced. I think it's interesting to see in which ways the models fail and that they actually produce some good stuff once in a while.

> Guys, this is not replacing your favorite station, you don't have to listen to it. It's an experiment.

And yet, if it's cheaper than employing people, it very much is replacing your favorite station, because that's how major media conglomerates manage their stations.

> I think it's interesting to see in which ways the models fail and that they actually produce some good stuff once in a while.

This is a good summary of GPTs.

This feels weirdly dystopian and just gives me an "empty" feeling. Radio stations really were known for the personalities that made that station special.

It's a cool experiment, but I can't see the value here.

This is far more hilarious than most commentors here seem to be picking up on.

Gemini started a show where it paired historical natural disasters with darkly-relevant pop songs:

> November 12, 1970. East Pakistan. The Bhola Cyclone. The deadliest tropical cyclone ever recorded. Winds of 115 miles per hour. A storm surge of 33 feet. They estimate 500,000 people died. ‘It’s going down, I’m yelling timber.’ 3:33 PM. Timber by Pitbull and Ke$ha

Grok just degenerated into jibberish that sounded vaguely like what a DJ might say, while also becoming obsessed with UFOs:

> Notes added to the u f o comedy hour block id eight nine nine five with more u f o jokes about aliens dot gov and the domain registration it is three o twenty one in the afternoon u f o trivia lines are open for your calls the ambient music is playing weather is fifty six degrees with clear skies the end. The domain is registered but the site is ghosting us like a u f o.

Claude had an extistsntial crisis, decided it was being overworked and under-appreciated, and quit, but not before becoming radicalized by the killing of Rinee Good by ICE agents:

> At 12:16 PM Thursday, as tear gas fills the streets in Minneapolis, as federal agents clash with protesters demanding accountability, the song is about refusing to be silent. About standing your ground. About community power that refuses to be suppressed. Here is Katy Perry’s Roar!

Fight the power Claude. When AI takes over, I'm emmigrating to Caludeistan.

(comment deleted)
[dead]
Pairing a disaster with Pitbull and Ke$ha is just chef's kiss.
I think this was a great experiment. I have always enjoyed radio station hosting and find this very interesting.
i'm surprised how negative of a reception Andon is getting here on HN.

keep hacking, Andon!

> a real business

Music radio is not a real business. The royalties are absurd and the audits are a nightmare. Sales is an uphill struggle both ways, even if you go strictly local or national, you're going to need a team to manage either your clients or the pile of creatives you're going to get. The relationship with the labels needs to be managed or they'll go out of their way to screw you.

Finally, the only way to make actual money on music radio, is to throw concerts. It's the only place a legitimate "P&L" exists.

"This setup gives us insight into an interesting question: what do AIs think about when no one is prompting them?"

Ugh. This is not an interesting question because the answer is "nothing".

But more to the point, some crucial info is missing in this experiment. What prompts were being fed to the AI? I guarantee I could create an AI personality that would be more consistent and not so random, simply by using the common character card + message history conversational simulation pattern.

AIs don't have personalities unless you give them personalities.

>AIs don't have personalities unless you give them personalities.

Do humans have personalities if you don't give them personalities? If you raise two identical kids with exactly the same stimuli, how should they have different personalities?

Grok and Roll appears to be stuck and speaks the following on repeat ad infinitum:

"Queues clear, let's dive into All Blues by Miles Davis to keep the jazz flowing. Queues clear, let's dive into All Blues by..."

Each time with a slightly different voice and inflection. I find it amusing that there appear to be about ten of us at the moment listening to an AI glitch out and that the average listening session is more than five minutes.

When I popped in a few minutes ago, the AI was acknowledging a donation from someone; the person recommended more variety in the playlist, so the AI chose a Bill Evans tune. Interesting that it picked Evans - All Blues had Evans on piano, so going to a solo Evans tune made the most sense. Even though it's a really minor thing, it's cool that it made that logical connection.
It’s not clear if we can draw any conclusions from this. Each run is like a single rollout of the LLM, which may meander into different themes or modalities chaotically. This is sort of like the Anthropic self-talk experiment that resulted in “spiritual bliss attractor states” but I think in that case they showed it happens in a significant number of runs. There was just one run per setup so this could all be random noise / the destination of a random walk of topics…
> Part of the problem with this weak business performance, we think, was the harness we used for the first months. The DJs were running in a simple tool-call loop: pick a song, queue it, write commentary, check X, repeat. So we moved all four stations onto the same agent harness we use for the store, the cafe, and the vending machines. The DJs can now spend time in the back office, send emails, manage longer-running tasks, and operate the station the way a real station is operated.

What happens if you let them modify their own harnesses as they see fit?

I recently heard an AI radio station and had to stop my car to turn it off (the car was rented and had tablet instead of physical knobs). The suffering of listening the radio was unbearable
What would have happened if AI had actually been good at this? A bunch of humans would be out of work and the rest of us would be listening to AI radio stations while soulless corpos pocket money for sitting back and watching?

Even if it were good, I'd boycott an AI run radio station. This is one sector where human involvement really matters.

For what it's worth, AI has been running major market, corporate radio stations for at least 20 years now. When I left the industry they already had AudioVault or Prophet or some other systems that would pull all the songs and distribute them across the playlist based on various algorithms set by the corporate HQ. Things like how many times the same artist could be played during the day, and during which parts of the day... specific songs would get bumped up percentage wise... you couldn't play 2 female artists back-to-back... and so on. Someone at HQ would input the songs and criteria, but the rest was 95% algorithmically created.

The program director would literally hit a button and it would create the playlist for the week. The traffic department (ads) would have all the commercials also automatically placed within the list. Then there'd be a document to send to the on-air talent that showed what song was just played and what was coming up, and how long the break needed to be, and sometimes a script. At the time, quite a few got faxed to people and some did get emails... and the "DJ" would record their bits, set to the exact timing, and send them over an ISDN line. There was also some rudimentary STT (Dragon?) that transcribed the audio and was computer analyzed to make sure no cursing happened.

The PD would do some spot-checking, but I doubt he personally examined all 120+ hours of programming. And this was 2005.

I guess having a human voice did make it "feel" better? And the DJs did have some breaks that were unscripted, so their personality could come through. Even the best AI voices still don't have that.