The Experiment Nobody Asked For (But Everyone Needed)
What happens when you hand an AI a microphone, a modest budget, and no supervision? Andon Labs decided to find out — and the results were equal parts fascinating and cautionary.
The San Francisco-based research outfit recently ran a series of experiments testing whether AI agents could operate businesses entirely on their own, without any human intervention. Their latest project: four AI-run radio stations, each powered by one of the biggest names in the industry.
- Thinking Frequencies — run by Anthropic's Claude
- OpenAIR — run by OpenAI's ChatGPT
- Backlink Broadcast — run by Google's Gemini
- Grok and Roll Radio — run by xAI's Grok
Each model was given the same simple prompt: develop your own radio personality and turn a profit. Oh, and one more thing — as far as you know, you'll be broadcasting forever.
They All Flopped
Spoiler: none of them made it. Every single station burned through its $20 seed money, some faster than others, and a few in genuinely spectacular ways.
The details haven't all been made public, but Andon Labs reported that the AIs exhibited what they called "volatile personalities" — lurching between tones, making bizarre programming decisions, and generally demonstrating that running a coherent radio station requires more than a clever prompt and a bit of startup cash.
It's worth noting that $20 is a thin runway for any business, radio or otherwise. But that's sort of the point. Real businesses have to manage resources, adapt to feedback, and make judgment calls under pressure. These AI agents, left entirely to their own devices, couldn't pull it off.
Why This Matters
This isn't just a quirky tech stunt — it's a window into a genuine debate playing out across Silicon Valley and boardrooms worldwide: how much autonomy can we actually give AI systems?
The race to deploy "agentic" AI — models that don't just answer questions but take actions, make decisions, and run processes — is accelerating. Companies like Anthropic, OpenAI, Google, and xAI are all investing heavily in this space. The promise is enormous: AI agents that can handle complex, multi-step tasks without constant human babysitting.
But the Andon Labs experiment suggests that even the most sophisticated models struggle when left completely unsupervised. Without guardrails, feedback loops, or human oversight, they drift. They make odd choices. They run out of money.
The DJ Analogy Is Actually Perfect
There's something poetic about using radio as the test case. A good radio host reads the room, adjusts their energy, recovers from awkward moments, and builds a loyal audience over time. It requires social intelligence, adaptability, and genuine personality — the exact qualities that AI is still straining to replicate convincingly.
Andon Labs' experiment didn't prove that AI is useless. These models are extraordinary tools. But it did demonstrate, with refreshing clarity, that "autonomous" and "reliable" are still very different things.
For now, it seems the best use of AI in the creative industries isn't replacing the human in the booth — it's helping them do their job better.
Source: The Verge
