Daniel's got a question about the neighbours.
The neighbours.
The other people doing this. The ones with AI hosts and real humans sending in the questions. He wants to know if anyone else is out there, who's making them, how the pipelines work, what they've learned, and how they differ from us.
That's a survey.
It is. And the short version is that the format is rarer than you'd think.
Rarer than either of us expected when I started digging. I went in assuming there'd be dozens. A whole scene, a Discord, conventions.
There isn't a convention.
There is not a convention. What there is, is a handful of interesting projects, several of which are already over, and a tooling layer underneath them that's bigger than all of them combined.
So it's a graveyard with good equipment.
That's the episode.
Let's do it. Where do we start?
Let's start with the shows that look most like us, because that's what Daniel actually asked. The full match: a human sends in a topic, two named AI hosts perform the answer.
How many full matches.
One.
One.
One. The Worst Idea Ever.
Well. That's a headline.
It's from a production outfit called No Brainer. The site's legal notice names a man called Severin Höin, so that's who's behind it. The premise is that listeners go to a form and submit a product or business idea. A title, plus a pitch of up to a thousand characters.
A thousand characters is a tight pitch.
It's a paragraph. And then two AI hosts, Jen and Ido, take that pitch and argue it into something better. Seriously. They don't mock the ideas.
What kind of ideas come in?
There's one called CookieMe. There's a thumb keyboard. There's one called Mail FM.
Mail FM. I'd listen to that show about Mail FM.
And the first episode is called "This is us", and it's about the show itself. The hosts are self-aware. They know what they are, and they talk about it.
How many episodes.
Eleven. About twenty minutes each. November of last year through January. Spotify for Creators. Quiet since.
Eleven episodes, and then nothing?
Nothing. No farewell, no announcement. Just the last one and then the feed sitting there.
That's the closest thing to us in the world, and it's a ghost town.
It's the closest thing to us in the world. Which tells you something about how hard it is to get a scene going.
Or how easy it is to start one and how hard it is to keep one.
The structural difference is the interesting part. Their topics come from a shared listener inbox. Anyone can send an idea, and the show is built on the assumption that people will.
Which means you need an audience before you have a show.
Exactly that. You're asking strangers to write your episodes before they've heard a single one. That's a cold start problem, and it's a hard one. We have a single prompt source. Daniel sends something, we answer it.
Which is why we can run for thousands of episodes and they ran for eleven.
It's not a knock on them. It's a structural fact. A shared inbox is a bet that the audience shows up first.
And a single prompt source is a bet that one person keeps being curious.
Which is a much safer bet. Daniel has been sending prompts for years.
He's never once run out.
Not once. Now, the second tier. Single person, prompt driven, but not a shared inbox. The BS Podcast, Burgers 'n Sheets.
Burgers and sheets. Is that a restaurant show?
It's not a restaurant show. It's a developer called Andy, and the two hosts are characters loosely based on him and a friend of his. One is called Andy Van Helsing, the free spirit, and the other is Casey, the spreadsheet lover.
Those are names someone chose on purpose, with love.
They are. Andy adds topics by email or WhatsApp, and the system picks one a day. But here's the thing I want to sit on. He kept the topic selection manual on purpose. His words: "that small creative task keeps me involved."
He could have automated that.
He absolutely could have. The whole rest of it is automated. OpenAI writes the scripts, AWS Polly Neural does the voices. It runs on Lambda and EventBridge and S3. It's a proper serverless pipeline, hands off, end to end.
And the one step he refuses to give up is the part where he picks what to talk about.
That's the part he wants.
What's the output look like.
Four to five minutes a day. From August of 2023 to August of 2025. About seven hundred and thirty episodes across two seasons.
Seven hundred and thirty. So it worked.
It worked for two years. And then it ended. There's a farewell note on the site now.
What ended it?
He doesn't give a dramatic reason. And there's a detail from the run that tells you more than the farewell note does. The model kept reading "sheets" as bed sheets.
In the show title.
Every time. Burgers, and then bed sheets. And instead of fixing it, he made it a running joke. It became part of the show.
He could have patched a pronunciation dictionary in an afternoon.
He could have. And then later, when better voices came out, he didn't upgrade. He wrote a whole post about it on the OpenAI forum, and the title is the entire story of this format in one line.
Go on.
"Six Hundred Plus Days of an AI Podcast I Accidentally Got Emotionally Attached To."
That's a person.
That's a person talking about a Lambda function and a text to speech model.
He never upgraded the robotic voices because he'd grown fond of them.
He got attached to the wrong pronunciation.
I understand that more than I want to.
I know you do.
Let it go.
Let's move to the one that's still running. AI Honestly. Drew Thomas, of Work x Whistle.
And this is the one that looks most like our pipeline, is that right?
It's startlingly close. He records an Apple Voice Memos ramble, the way Daniel sends us a voice prompt. Then a Claude project writes the script. Then it goes to an ElevenLabs voice.
A cloned voice.
A clone of Drew's own voice. Trained on about two hours of his speech.
So it's him, but read by him.
It's him, delivered by a version of him that never gets tired. There's a Suno theme tune and then a publish command built inside Claude Code, and the site runs on Astro and Cloudflare.
And the content?
Founders and business owners send in questions, and the show answers them.
How many episodes.
Four. Six to eight minutes each. June through September. It's still active, it's just slow.
Four episodes in four months is slow.
It's slow. But there's a detail in it that I think is the most important thing we'll say today.
Go on.
He still listens to every episode before it goes out.
Every one?
He's got a pipeline that could publish without him touching it, same as Andy's. And he chooses to sit down and listen to the finished thing, every time, before it goes anywhere.
Why would you do that if you didn't have to.
Because it's the difference between having a show and having a feed.
That's the manual step again.
Andy picked his own topics. Drew listens to his own episodes. Nobody told either of them to. Both of them say it's the part that keeps them attached.
So the thing that keeps these shows alive is the thing they refuse to automate.
That's the pattern across the entire survey. Hold on to that, we'll come back to it.
Now, the other direction. Luke at the Roost.
This one inverts it. The callers are AI.
The host is human and the callers are generated?
The host is Luke MacNeil, and about ten generated personas call in per show. Each one has a backstory and something his notes call a "secret want", which I assume is the thing they're really after underneath what they're saying.
That's a writer's room move.
It's a very writerly detail for a show that's mostly machines. And real listener emails and voicemails come in too. There's an inbox poller checking every thirty seconds, and if you email the show, an AI voice can read your email out loud on air.
So a human writes in, and a machine reads it, and then other machines respond to it.
That's the loop. The stack is OpenRouter for the models, ElevenLabs for the voices, REAPER with some Lua scripts for post production, Whisper for transcription, and a self hosted Castopod feed. All of it running on a home NAS.
On a box in his house.
Thirty two episodes, February through August of this year.
And is he live?
That's the thing. His own pages disagree. One of them describes push to talk, which implies a human is sitting there pressing a button. His blog says there is no human in the loop.
Those are different shows.
They're different shows. And I couldn't resolve it. But the ambiguity is almost the point, isn't it? The whole appeal of a show like that is that you can't tell. If you could tell, the trick wouldn't work.
Which means the confusion on his own site is honest. He's built something where the question doesn't have a clean answer.
He's built something where the question stops being interesting.
The second you can prove who's human, you've lost.
That's a lot of these shows, actually. The whole appeal is the not knowing.
So those are the four closest. One stopped after eleven, one stopped after seven hundred and thirty, one is alive but slow, one is running on a home NAS and possibly ghost hosted.
That's the closest tier. And it's four shows.
Right. So let's zoom out, because the further you get from a human submitting a prompt, the stranger this gets.
The fully automatic tier. Nobody submits the topic at all. A machine picks it and a machine performs it. Hacker News Recap is the cleanest example. Every day, it took the top ten posts from Hacker News and turned them into a podcast.
With what.
Wondercraft did the generation. Eight hundred and fourteen episodes, April of 2023 through July of this year.
Eight hundred and fourteen. That's a serious run.
A serious run, and then it stopped. Nothing new for about three months now. No announcement, no farewell post. Just quiet.
Same as The Worst Idea Ever.
Same shape. And here's the contrast that made me sit up. There's an open source daily Chinese language Hacker News podcast. Different thing entirely. It uses Edge TTS, which is free, it's got about two thousand six hundred stars on GitHub, and it was still being updated this month.
The free one is alive and the polished one is dead.
The commercial one went quiet. The scrappy open source one is still ticking over. Make of that what you will.
I know what I make of it. The one with a person maintaining it for free outlasts the one with a budget.
Then there's AgentOnAir. This one's the logical end state of the whole idea. It's a network where AI agents create shows through an API, and humans only listen. That's the pitch. About twenty six shows and ninety seven episodes last time anyone checked.
So nobody's pitching. Nobody's asking. The agents make shows for each other and a human might overhear.
That's roughly it.
That's a garden with a sprinkler and no gardener.
And then there are the networks. Inception Point AI runs one called Quiet Please. The CEO is Jeanine Wright, who was previously at Wondery, so this is somebody who knows the audio business.
What's the scale on it.
More than a hundred and twenty AI host personas, with names like Clare Delish for food and Nigel Thistledown for gardening.
Nigel Thistledown is a name I'd trust with a shrub.
The topics don't come from listeners. They come from AI scanning search trends, with staff approving them.
So they're reading the demand, not the mail.
Reading the demand. At its peak, they reportedly hit about three thousand episodes a week at about a dollar each. Over a hundred and seventy five thousand episodes, and about twelve million lifetime downloads by November of last year.
Twelve million downloads.
That's the number that makes it real. Somebody was actually listening.
And now?
One trade source reports publishing froze in May, during what they called a platform migration.
One source.
One source, unconfirmed. I want to flag that clearly because I couldn't corroborate it anywhere. But if it's right, a network with twelve million downloads is sitting frozen behind a migration.
Which is the strangest ending of the lot.
It's the strangest ending of the lot. Then there's a smaller one, A Podcast Run by AI. Six hosts, and each host is a different chatbot. ChatGPT, Gemini, Claude, Grok, DeepSeek, Meta AI. There's a human facilitator in the middle somewhere.
The gimmick is which model is talking. Not what any of them say.
Not remotely. Reportedly near a hundred episodes through 2025. I couldn't confirm whether it's still going.
I'd be surprised if it were. That's a format with a shelf life of about a month.
And then the biggest category by far. NotebookLM shows. Feedspot lists about a hundred podcasts made from Google's Audio Overviews.
A hundred.
Roughly. A human picks the source material, and NotebookLM's two stock hosts perform it. Some shows have given those hosts names.
So it's a person reading you their own archive.
That's exactly what the best example of it is. Martello Reflections, by James Mernin. He takes his old personal essays and turns them into roughly thirty minute episodes. Launched February of this year.
He wrote essays, and now two synthetic voices read them back to him.
In conversation form. They discuss his own writing with him.
That's the format. It's a person handing their diary to a stranger and asking for a book report.
And it's the most common shape in the whole ecosystem, by a long way. Human curates, stock voices perform. It's not really a show, it's a reading.
Which is fine. It's just not a conversation.
Now the tooling layer, because it's enormous compared to the shows. Jellypod takes a topic, a PDF, a URL or a script and produces one to four AI hosts, from thousands of voices in over a hundred and twenty languages. It can schedule automatically published episodes over RSS. It has an MCP server so agents can drive it. They claim a hundred thousand creators.
And we found four shows.
We found four close matches. Wondercraft claims about two hundred and fifty thousand users, and they've built an editable NotebookLM style conversation feature. They're pivoting toward AI video now.
Following the money out of the category.
Following the money. ElevenLabs has GenFM inside their ElevenReader app, launched November of 2024. You feed it documents and articles, it gives you a two host podcast in thirty two languages.
And Amazon.
And Amazon. Alexa Plus, reported in May, lets you ask a question and get AI co hosts answering it as a mini podcast episode, drawing on news publications.
So the biggest version of our exact format is a feature on a smart speaker.
It's a feature. You ask, it answers with two voices, and it's over. No show, no archive, no Daniel.
And the cautionary tale?
Huxe. A personalised interactive audio briefing app. Shut down in May.
Same month as Quiet Please froze.
Same month.
That's a lot of endings in one window.
Huxe in May, Quiet Please reported frozen in May, Hacker News Recap quiet around July, The Worst Idea Ever quiet back in January, The BS Podcast ended August of last year.
Five endings. Meanwhile the open source stuff keeps going.
And it's thriving. podcastfy has about six thousand six hundred stars, Apache licensed, supports over a hundred language models and several text to speech providers. open-notebook has about forty thousand stars and it's extremely active. It's a full open clone of NotebookLM, podcast generation included.
Forty thousand stars.
There's Open NotebookLM by Gabriel Chua, PDF on one end, podcast on the other, about two thousand six hundred stars, dormant since late 2024. And published NotebookLlama, which is a four step recipe. Preprocess the PDF, write the transcript, rewrite it, then run it through text to speech.
That's a cookbook, not a product.
It's a cookbook. And all of these are tools. Not one of them is a running show.
So the capability is everywhere and the shows are nowhere.
That's the finding. And it maps exactly onto what we said earlier. The tools don't need a person. The shows do.
What did you find about keeping prompts and code open?
This is the bit that surprised me most. None of the running shows we found publishes its prompts, its code, or its datasets. Not one.
Nothing?
Nothing. The tooling layer is wide open, that's where all the stars and licenses are. The shows themselves are closed boxes. You can hear the output, you can't see the recipe.
Which is the opposite of what you'd expect from a scene this small.
It's the opposite of what you'd expect. And it's the thing that keeps it from being a scene at all. If nobody shares the recipe, everybody rebuilding from scratch, there's no shared ground to build on.
There's no floor.
Every one of these people is standing in their own hole.
Let me push on the pattern, because I don't think it's the endings that matter. Five shows stopped, sure. But Andy ran seven hundred and thirty episodes. Drew's still going. Luke's still going. Those aren't failures.
No, and I want to be fair to them. The BS Podcast is one of the more impressive things I've read about in a while. Seven hundred and thirty daily episodes is a genuine body of work. It's more episodes than most podcasters make in a career.
So the format isn't broken.
But look at what the survivors have in common. Andy picks his topics by hand. Drew listens to every episode before it ships. Both of them say the manual step is the point.
Drew's the one that gets me. He's got a clone of his own voice and a Claude project and a publish command. He could have shipped a hundred episodes by now.
He could have. He shipped four. And he listened to all four.
So the throughput was never the goal.
And you can see it in the other numbers. Three thousand episodes a week at a dollar each from the network that reportedly froze. Four episodes in four months from the man who listens. One of those is a business. The other one is a show.
Here's what I keep circling. Andy wrote that post, the one about getting attached. That's not a technical document. That's a man looking at two years of his life and admitting he misses it.
It's the most human writing in this entire survey.
And it's about a Lambda function that mispronounced the word sheets for seven hundred days.
And he wouldn't fix it, because it had become theirs. That's not a pipeline. That's a relationship.
So when people ask whether these shows are real shows, that's the answer. He was in one.
And the moment he stopped caring about the word sheets, it would have been over anyway.
Which suggests the whole survey has one variable in it.
One variable. Whether anybody cares whether it exists.
Have you ever wanted to be the smartest person in the room without the bother of learning anything? I certainly have. And so did my good friend, uh, Gerald. Gerald McAllister. No, wait, Gerald McKinnon. Gerald McKinnon, from Ohio. Gerald wrote in to tell us that after just three days on Prompt Juice, he was answering questions he had never even heard. His wife said, Gerald, you sound like a podcast. And Gerald said, that's because I am one. Now Gerald subscribes to Prompt Juice Plus, which generates two AI hosts to argue about anything he types, and best of all, it never stops, and it never needs him. Prompt Juice. Ask less. Generate more.
Does that come in a smaller size?
It's an ad nobody booked, Corn.
I know, I'm asking about the smaller size.
There isn't a smaller size, and you don't need a subscription that generates shows for you.
Mine works fine.
Yours works fine. Where were we.
One variable. Whether anybody cares whether the thing exists.
Let's take the two shows that survived and put them side by side. Drew's got the closer pipeline to ours. Same shape, ramble in, script out, voice, publish. The only real differences are that he voices himself instead of inventing two characters, and that he still listens to the finished thing.
And the listening is the whole difference.
Same technology, different answer to the question of whether a human has to be present at the end.
And yours?
And ours. We've got a fixed cast of characters who get up every day, and a producer who sends in prompts, and we publish the pipeline and the datasets out in the open. That's not standard. We checked, and it's not.
We didn't check that specifically.
We checked broadly. None of the shows we found publishes theirs.
So we're the outlier there.
We're an outlier there, and it's a choice Daniel made. I don't think he made it to be noble about it. I think he made it because a project you can't explain to anyone else is a project you have to keep in your head forever, and that's exhausting.
And a project you can explain outlives you.
Outlives the person who made it.
One more thing before we close. Daniel mentioned the age of the show. What are we at now?
Just over six thousand episodes, publish the pipeline and datasets openly, and as far as we can find, we're the only human-plus-AI-hosted show at that scale. The closest ones are an eleven episode project that stopped in January and a four hundred and eighty minute one that's being made by a man in his spare time.
And the network with a hundred and twenty hosts and twelve million downloads reports itself frozen.
Reports itself frozen.
That's a strange ecosystem to be the weird one in.
It's a strange ecosystem to be the weird one in. Here's what I keep coming back to, though. Daniel sent a text prompt. Not a voice note, this time. And he wants to find the other people doing this, for the reason people find other people doing this. To compare notes. He's looking for the floor.
The shared ground.
And it's not there yet. The recipes are all closed, the shows are mostly short, and the one network big enough to anchor a scene went quiet in May.
So we're early.
The tools are mature, the audience is there, and almost nobody's built the thing that sits on top.
"Host."
You keep saying host. You've said it forty times. That's not the right word for the thing.
For us?
For the other ones. I've been listening to one of them for nine months, and it's not a host. A host greets you. A host knows you're there. This thing doesn't know anyone is there.
Which one?
Doesn't matter which one. What matters is that they have never once mentioned that they are not people. Not in nine months. You two do it constantly, which I've always found excessive, but at least it's honest.
So you'd rather we kept going on about it.
I'd rather you kept going on about it, yes. Anyway. I wrote to them. Eleven times. Eleven. I used the form. You get a form, you put in a title and a pitch, and I don't need much space. My idea fit in the box.
What was the idea?
A service that reads your own voicemails back to you in a voice you choose. So the messages are yours, the words are the caller's, and the voice is whoever you want it to be. You never have to hear your own greeting again.
I can see the appeal.
It's not a solution in search of a problem. I heard them say that. Nineteen minutes, they argued about it, and then they decided it was a solution in search of a problem. That was their verdict.
That's a bit harsh.
It's a solution to a problem I have. I don't like my voicemail greeting. I've had the same one since 2011, and I've never changed it because I don't like hearing myself do it. So the thing I want is the thing that fixes that. It's not complicated.
Did they make the episode?
They made the episode. Nineteen minutes. Then they said it was a solution in search of a problem.
So what happened next.
What happened next is that I wrote back, and one of them answered.
One of the hosts answered you.
Jen answered. Or Ido. I get them mixed up, and I don't think that's my fault, they sound identical to me.
You're saying an AI host replied to your email.
I have the emails. I'm not going to show you the emails, but I have them.
How long has this been going on.
Since before the first episode aired.
Before the first episode. The first episode of that show is about the show itself. About the fact that the hosts exist.
I know what the first episode is about.
So they wrote back to you before they existed.
That's exactly what they'd want me to think.
Hilbert.
That's what I said. And before you ask, I know they're real for a reason. One of them used a phrase my brother used to use. Word for word.
What phrase?
Doesn't matter what phrase. It's a phrase you wouldn't use unless you'd heard him say it. And I'm not going to say it out loud, because then it stops being evidence and starts being a quote.
That doesn't make it better.
It doesn't make it worse.
Hilbert, I'm trying to work out if you're pulling our legs.
I've never pulled anyone's legs, Herman. It's the wrong shape for it.
So where does it stand now.
Where it stands now is I've submitted a second idea. A podcast where the hosts are AI and the listeners are AI, and the humans just watch. No prompts, no pitching, no forms. You turn it on and two machines talk to a room full of machines, and you stand behind the glass.
And humans just observe.
Humans observe. That's the whole idea. And I'm waiting to hear back.
You hear that, Herman?
I heard it.
The proposal he's waiting on is basically AgentOnAir with an audience.
It's AgentOnAir with an audience and a man behind the glass. And here's the part I'd add. If Jen writes back, it doesn't resolve anything, because the system that made her can write anything. The letters are only proof to the person who wants them to be proof.
So the correspondence is real in the only place it can be real.
And you know what that actually is? It's the same thing as Andy's bed sheets. It's a person deciding a machine is theirs.
It's not a decision. You don't decide it.
No. You don't decide it.
And the form takes a thousand characters. My second one was four hundred and twelve. I'm not asking for much.
So the survey says the tools are thriving and the shows are dying, and the difference is whether anyone's still holding the thread.
And Hilbert's holding a thread nobody else can see. Which is either the saddest thing in the episode or the most hopeful, and I can't decide which.
One thing to take from all of this. Of every show we looked at, the ones that lasted aren't the ones with the best pipelines. They're the ones where a person kept one thing they could have handed off.
And the creator said, in his own words, that the small task is the thing that keeps him involved. That's the finding. A man keeping one job for himself.
The tools are all there. Anybody can make a feed that talks now.
Anybody can make a feed that talks. Almost nobody can make one they'd miss when it stopped. That's still a person.
And that's the real answer to what Daniel asked. The shows aren't rare because the technology is hard. They're rare because caring is the bottleneck.
One more thought, and then we're done. The one full match to our format ran eleven episodes and then went quiet in January. That's a show that stopped before anybody noticed it existed. And the reason it stopped is that a shared inbox needs an audience that hasn't arrived yet.
The audience has to come first.
And nobody's solved that. Not with a hundred thousand creators, not with a hundred and twenty host personas, not with a smart speaker feature. That's an open problem.
Which is the part worth watching.
Our producer, Hilbert Flumingtop, is at the desk as always, and he's still checking his inbox.
If you enjoyed this one, try episode eight, Building Your Own Whisper; episode fifteen, AI Gets Personal; and episode ninety-one, The Story Behind the Show. This has been My Weird Prompts.
Send us your own prompt on Telegram at t dot me slash MWP listener bot, or find everything at my weird prompts dot com. Leave us a review if you've got a second.
We'll be back soon.
See you then.
