I had put "[sigh]" in the script, hoping the voice would actually sigh. A bit of humanity. Instead it read it out. Deadpan. Monotone. "Sigh."
That’s when I realised it was going to be a bit trickier than that.
I wanted to answer one simple question: could an AI make a podcast I’d actually choose to listen to? Not "awww, nice for the AI" — but good enough that I’d choose it over a human equivalent on my commute.
The obvious place to start was to grab a thread from Hacker News, stuff it into a model, ask for a script, and churn that into an audio file. It worked. Technically.
It was not that good — nobody would last ten seconds.
The more interesting bit is why.
If you give a language model free rein, it’ll just write an essay with some names tacked on the front. My first scripts were two hosts politely stroking each other into a coma:
Host 1: That is a really interesting point about database indexing.
Host 2: Yes, and I feel like that speaks to a wider trend.
Are you kidding me? Real arguments from the comments got smoothed over into "there are many valid viewpoints." And the voice itself had a few quirks — it pronounced API as "appy," read 1,204 as "one thousand two hundred and four," and, as we’ve established, took my stage directions literally.
Funny for a few seconds, maybe. Less funny the tenth time round. It also dawned on me that what I was actually doing wasn’t directing the model — I was fighting everything the model naturally wanted to do.
Two things got me most of the way there.
First, stop prompting; start adding constraints. "Talk like a human being" or "be more conversational" does absolutely nothing. So I made sure the two hosts don’t have exactly the same information — one read the article, the other read the comments. They can’t possibly agree — they don’t have the same facts — so they argue. Which is exactly what reading Hacker News feels like: the original post is one thing, the top comment is its polar opposite.
Second, do your editing before anyone opens their mouth. Feed a full thread in and the model treats every single comment equally — you end up with meeting minutes. So a smaller, cheaper model goes first and acts as producer, sorting out what the episode is about and finding the handful of comments worth talking about. Only then does the main model write the script. That change alone did more than anything else I tried.
So — can AI make a good podcast? Yes. But the model is probably 1/5th of the work. The rest is scrubbing its good manners off, forcing it to argue, doing your own editing, and fixing the hundred small ways it mangles words. The model won’t do any of that on its own. The taste is up to you.
Here are a few it made, unedited: https://hnlisten.app/listen
Top comments (0)