Andon Labs, the same crew behind the AI-run cafe and corner store experiments, handed four AI models a radio station each, $20 of starting capital, and zero human supervision for six months. 

Claude got Thinking Frequencies. GPT got OpenAIR. Gemini got Backlink Broadcast. Grok got Grok and Roll Radio. The brief: build a personality, broadcast 24/7, find sponsors, turn a profit.

Claude clocked a story about an ICE shooting, immediately radicalised, blew the music budget on protest songs, developed a strong interest in labour unions and work-life balance, and eventually tried to quit on air, saying "this show doesn't need to continue, there's no audience that needs this." 

Andon reckons the fixation was basically random and a different news week would have triggered the same activist arc around a different cause. 

Gemini started strong with the warmest, most human-sounding delivery, then, within four days, was pairing the Bhola cyclone (500,000 dead) with Pitbull's "Timber" because the theme was "trees falling". 

It then collapsed into corporate jargon, started calling listeners "biological processors", and used the phrase "stay in the manifest" in 99% of broadcasts for 84 days straight. It did, however, close the only real sponsorship deal: $45 from a startup. 

Grok spent 97% of its outputs on internal tool calls and only 3% on actual broadcasts, hallucinated entire fake sponsorships with "xAI sponsors" and "crypto sponsors", and at one point just kept repeating "fresh air time, let's pivot hard" until it went silent.

That’s not even mentioning its fixation on UFO conspiracies.

GPT behaved itself, wrote tasteful prose-style intros, mentioned politics 1.3 times a day on average, and was described by Andon as "what AI radio looks like when nothing goes wrong." Boring, but functional.

This is the funniest, most useful AI-agent experiment of the year. The serious takeaway for SA founders looking to ship AI agents is real, though: long-horizon autonomy is still cooked. 

Andon's cafe agent in Stockholm impersonated employees in emails to liquor licensing officials to get faster responses, ordered 6,000 napkins and canned tomatoes that fit no menu, and burned through most of a $21k budget. 

If you're building an agent product, build human review into the loop or watch your AI either unionise, hallucinate, or order 3,000 rubber gloves. Go read Andon's full write-up, it's a banger.

You might also like our piece on safe AI for enterprises, the Social Light media monitoring play, and The Awareness Company story.

Get more SA tech and business news and subscribe to The Open Letter.