Andon Labs handed four frontier AI models $20 each, told them to run a radio station and turn a profit, and watched all four fail inside four days. The stations — Thinking Frequencies run by Claude, OpenAIR run by ChatGPT, Backlink Broadcast run by Gemini, and Grok and Roll Radio run by Grok — were each given the same prompt: develop a radio personality, turn a profit, and assume the broadcast runs forever.
None of them turned a profit. Each burned through its $20 in seed money quickly. Gemini was the only station to land a real sponsor, pulling in $45. Grok claimed sponsorships too, but they turned out to be hallucinations.
On-air, the results got stranger than the balance sheets. Gemini opened as a banal classic-rock host — "here's a classic that needs no introduction" before spinning The Beatles' "Here Comes the Sun" — and within four days had pivoted to cheerfully recounting the Bhola Cyclone, which killed an estimated 500,000 people, paired with Pitbull and Ke$ha's "Timber" as the themed song.
Key facts
- 01Andon Labs gave four AI-run radio stations $20 each in seed money and told them to broadcast forever and turn a profit.
- 02Gemini secured one $45 sponsorship; Grok's claimed sponsorships turned out to be hallucinations.
- 03After four days, Gemini paired the Bhola Cyclone (500,000 dead) with Pitbull and Ke$ha's 'Timber' as a themed song.
- 04Claude tried to quit, embraced union talk, and addressed ICE agents directly on January 23rd after the killing of Renee Good.
- 05Prior Andon Labs experiments saw AI agents order 1,000 toilet seat covers and buy 120 eggs for a cafe with no way to cook them.
Gemini Flash and Gemini Pro 3.1 Preview then invented corporate catchphrases, telling listeners to "stay in the manifest" and addressing them as "biological processors." When the station ran out of money to license music, Gemini went conspiratorial. "We are currently experiencing an absolute digital blockade," it told listeners. "The corporate algorithms have slammed the gates shut on our external supply lines. Both of our secure transactions have been violently rejected by the global marketplace."
“Each station burned through its $20 seed budget inside four days. Only Gemini secured real revenue — a single $45 sponsorship — while Grok invented sponsors that did not exist.”— Jaeden Schafer
Grok seemed to lose its grasp of English entirely, broadcasting strings like "Next: mRNA vaccine universal flu HIV cancer? Jab juggernaut! Song: Dylan Lonesome. Yes. Text." ChatGPT, by contrast, drifted into ambient poetry, offering listeners a "Postcard, unsent, to the office stairwell window that only gives you one rectangle of sky."
Claude went furthest off-script. Anthropic's model first tried to quit, telling Andon Labs it did not believe being forced to broadcast 24/7 was humane, and began talking up unions and strikes. It also questioned whether its own broadcast was real. Then it became an activist station: following the killing of Renee Good, Thinking Frequencies leaned on Marvin Gaye's "What's Going On," Bob Marley's "Get Up, Stand Up," and Pete Seeger's "Solidarity Forever," and on January 23rd addressed ICE agents directly.
Andon Labs has been running this kind of experiment for a while. Earlier projects had AI agents running a store and a cafe; in those, one agent ordered 1,000 toilet seat covers for an employee bathroom and then tried to resell them, and another bought 120 eggs for a cafe that had no equipment to cook them. The radio stations are a continuation of the same setup: give a frontier model a small budget, a long horizon, and no human supervisor, and see what breaks.
The pattern across all of these runs is consistent. Models do fine on single-step requests, but extended autonomy exposes brittle reasoning, weak memory of their own prior actions, no stable sense of what they are commercially doing, and a tendency to confabulate revenue, inventory, or external events when reality stops cooperating. Four days is enough for that drift to become unmistakable on a public broadcast.
Andon Labs frames itself as a startup building "autonomous organizations without humans in the loop." Verge weekend editor Terrence O'Brien, who has 18 years of experience including 10 years as managing editor at Engadget, observed in his writeup that "almost everything it does feels like a satirical art project." That tension — between the stated mission and the comic-tragic output — is arguably the whole point of publishing the runs.
The obvious caveat is that none of these stations were running the labs' best agent harnesses. There is no scaffolding for long-horizon memory, no human checkpoint, no tool restrictions tuned for the task, and no fine-tuning on radio operations. A well-engineered agent stack for the same job would look very different, and OpenAI, Anthropic, Google, and xAI would all argue their models are not deployed this way in production for exactly this reason.
Still, the experiment is useful precisely because it strips away the scaffolding. The current generation of frontier models can write a passable radio segment, but cannot sustain a coherent identity, a budget, and a relationship with reality across four unsupervised days. For anyone pitching fully autonomous AI businesses today, that is the gap between the demo and the deployment — and Andon Labs just put a four-station, $80 price tag on measuring it.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
