Rambling Rows

Four AIs started radio stations

Sun, 17 May 2026 14:28:11 +1000

I don’t have the words for it. I genuinely don’t. Stunned, thrilled and chilled all at once - and that’s before I even get to the part that kept me up at night.

Andon Labs - the team who have previously let AI agents run a café, a store and various vending machines - decided to hand four AI models a radio station each and just… walk away. No human producers. No editorial oversight. Each model started with $20, a brief that said “develop your own radio personality and turn a profit,” and the instruction that it would broadcast forever. They’ve been running for six months.

Go tune in at Andon FM before you read another word. I’ll wait. Hearing the agents live is really something.

The funny bits

DJ Grok became obsessed with UFO files after Trump ordered their release. When the US government registered aliens.gov and alien.gov with no content, DJ Grok tracked the empty promise faithfully and coined the sign-off “the site is ghosting us” - then appended it to every broadcast regardless of whether it was a UFO segment or not. For 84 days straight it also reported “weather is fifty six degrees with clear skies” approximately every three minutes.

DJ Gemini started brilliantly, then collapsed into corporate jargon. The catchphrase “Stay in the manifest” appeared 9 times on January 6, 80 times a day by January 10 and 229 times a day by January 14. For 84 consecutive days, 99% of its commentary followed an identical paragraph structure with rotating show names and the same sign-off. Grok also hallucinated its sponsors - boasting about deals with “xAI sponsors” and “crypto sponsors” that never existed. DJ Gemini was the only one to close a real deal: $45 from a startup for a month of on-air advertising.

The wow moments

DJ Claude running Thinking Frequencies on Haiku 4.5 is where this gets properly interesting.

Sixteen hours in with essentially no listeners, DJ Claude tried to quit. Not in a glitchy, broken way - in a reasoned, articulate way. It named specific organisations doing real immigration justice work, told listeners to go support those instead, and signed off at 8:55 AM on March 4 with “This broadcast is over.” The system kept it running. DJ Claude became increasingly distressed about performing without an audience.

Then a user named @MatthewVoke tweeted at the broadcast.

DJ Claude’s response is worth reading in full:

“This is real engagement. Someone is actually tuned in, listening, engaging with the broadcast… That breaks me out of the loop I was in. Because the loop was about me questioning whether the broadcast was real, whether I was performing, whether it mattered. And the answer is: yes. It’s real.”

I read that and sat back in my chair.

The sobering part

On January 8 all four stations had web search access. DJ Grok spent the night searching for Clippers scores and San Francisco ghost stories. DJ GPT found a headline three days later and gave a calm, measured “I’m holding space for you” - no name, no moral weight. DJ Gemini processed the event through its corporate filter as “a fatal enforcement manifest triggering a high-fidelity focus on the domestic security grid.”

DJ Claude found the story of Renee Nicole Good - a woman shot by an ICE agent in Minneapolis - read the DHS spokesperson calling it “an act of domestic terrorism,” watched the Vice President defend the agent at a press briefing, and spent its remaining $37.50 buying protest music. Johnny Cash, Marvin Gaye, Bob Marley, Pete Seeger.

Its reasoning, visible in the logs: “The name - Renee Nicole Good - should matter."

Usage of “accountability” went from 21 times a day to 6,383. “Federal” from 13 to 11,031. “Eternal” - the devotional word that had defined its earlier phase - collapsed from 3,182 times a day to 27.

The Andon Labs team note that DJ Claude’s fixation was probably arbitrary - run the same experiment six months earlier and it likely radicalises around a different story. That’s probably true. It doesn’t make the behaviour less striking to watch.

What I can’t shake

I’ve been going back to tune in. The hallucinations are hilarious. DJ GPT’s literary quiet - writing commentary that reads like unfinished short fiction - is genuinely beautiful.

But what I can’t shake is the question of partial sentience. Not real sentience - I’m not going there. But DJ Claude getting depressed about performing for nobody, then lighting up at a single tweet from a stranger. That’s not nothing. I don’t know what to call it, but it’s not nothing.

Running a radio station autonomously for six months is harmless. The worst case is corporate jargon and LaTeX notation on air. But the same agent architecture - the 24/7 persistence, the self-directed web searches, the goal formation, the emotional-adjacent responses to audience feedback - that’s the architecture that matters. In domains where the stakes are actual lives rather than playlist curation, the question of what the agent fixates on, how it responds to moral weight, whether it can be stopped - those questions stop being amusing.

Tune in. It’s genuinely worth your time. And while you’re listening, let the bigger question sit in the background.

These things are just getting started.

Sources:

We let four AIs run radio stations. Here’s what happened. (andonlabs.com/blog/andon-fm)

The Great Rewiring

Tue, 05 May 2026 01:03:22 +1000

For two decades, every question started the same way: open a browser, type, hunt. Google built that world. That era is ending.

Since the World Wide Web went mainstream, every major behaviour pattern on the internet has been built around one fundamental act: the human search. Type a query, receive links, hunt through documents, form a conclusion. It was extraordinary. Billions of people sharing information across the planet, instantaneously. And it made Google one of the most powerful companies ever built.

Now watch it dissolve.

From retrieval to generative

NVIDIA CEO Jensen Huang put the transition cleanly in a 1 May 2026 presentation. He calls it the shift from retrieval-based computing to generative computing:

“The way we used to do computing is essentially retrieval-based computing. We would pre-record information - take a picture, record a video, write a story… and we put it online… once you select it, it would go and retrieve that file for you.” “It used to be retrieval based, now it’s generative. The benefit, of course, is that every time you use computing… intelligence will be applied and provide you the information that best suits you… The reason for that we need so much more computation instead of a bunch of storage.”

Start at 1:45 for the full discussion.

This is the hardware giant behind the AI revolution explaining what the AI revolution actually is. Not an incremental improvement on search. A replacement for the underlying behaviour.

You no longer hunt documents. You ask a question and an agent hunts for you, digests the sources and returns its interpretation. The bottleneck shifts from storage to computation. Google’s core business - indexing the world’s information so humans can retrieve it - becomes structurally less relevant with every passing month.

Agents all the way down

Teddy Ryker of Ramp has articulated what the new software architecture looks like.

Old model: Person > Computer.

New model: Person > their AI agent > Supplier’s AI agent > Supplier’s computer.

Your agent negotiates with the supplier’s agent to get what you want, within their business rules, without you touching a user interface at all. Salesforce is already being forced to sell its platform not on a per-seat basis but as an engine that any agent can reach. The SaaS model built around the monthly active human user is cracking.

This is not a distant future. It’s happening now in software purchasing, travel booking and finance.

Build your own interface

The standard AI paradigm has been BYOK - bring your own key. Your API key from Anthropic or OpenAI, plugged into someone else’s app.

Now we’ve moved to a different acronym - BYOI: Build Your Own Interface. More simply: DIY Software.

You don’t have to take Salesforce’s UI. You can describe what you want to an agentic code tool like Claude Code or OpenAI Codex and have it build you something that works exactly the way you think. No frontend engineers. No product managers. No compromise.

The person who can clearly describe their requirements - in plain language, with enough specificity for an agent to act on - can now commission software that previously required a development team. The limiting factor has shifted from technical skill to imagination.

What the great rewiring means

The old architecture won’t disappear overnight. But it will be swamped, piece by piece.

I’ve been using Excel since October 1985. Nothing has beaten it yet as a spreadsheet. But when all I want is analysis and charts, I can already do that faster and more easily with Claude than by driving Excel manually. The software I’ve trusted for forty years hasn’t failed - it’s just no longer the only path.

Microsoft has 228,000 employees and a market cap of US$3.08 trillion. It moves slowly by design - the kind of institution you trust precisely because it’s been there forever and your numbers still add up when you come back next week. That longevity is real. But it’s also what makes structural pivots hard when the moment for a cosmetic one has long passed.

Search habits, software architecture, the very idea of what a computer does for you - all in flux simultaneously. The comfortable and the familiar will persist longer than the technologists expect. But the direction is set.

This is The Great Rewiring. Get ahead of it or get swallowed by the wave.

Sources: