Georeactor Blog

RSS Feed

NotebookLM vs. human podcast



Tags: podcastsml

This past week was a little exciting as I went peak-unemployed and made a podcast. It's a little funny that NotebookLM has launched with a podcast demo feature around the same time, so I'll discuss them in one post.

My podcast topic is a developing news story which I posted before on the blog: fugitive mayor Alice Guo. After becoming a minor celebrity and villain in the Philippines, their Senate and House held hours and hours of hearings. That's content!
When I made the first episode, it was an experiment for me to get all the audio software and an external mic (which I've had since 2022). Reading from a script with intention was an improvement on presentations which I've recorded in the past. Conceptually I thought of the audience as someone like my cousin, who has no other exposure to this story. So it was a bit of a surprise when YouTube kept recommending it to Filipinos.
The second episode was smoother in planning and recording. I think it's one of the best pieces of media that I've created!

Creating this content is a bit too obscure and hyper-focus-y. There's a 4chan post (hopefully fake) where someone's grandfather obsesses over Rwanda conspiracies, even though his family has no connection to the country. If I can keep the podcast episodes to 10 minutes with mostly Filipino primary sources, I think that we're in the green zone.
The third episode process was not as fluid. I was happy assembling updates and video links into a script over a few days, but wasn't as confident when I read it aloud the next day. Mania / hyperfocus wearing off? Another issue was that I summarized the updates too much (Al-Jazeera had the Chinese agent record a warning in Mandarin for Alice, but I didn't want to include it, then me reading the subtitles, then me talking about it).
And there's something funny about how Cassandra is in a totally different movie, but I'm not sure that I nailed it in this episode.

Now for NotebookLM - it's a UI to collect documents and media to feed a long-context LLM, and includes a feature to generate a 'podcast' dialogue. voices and dynamic remind me of the 'Armchair Expert' podcast. No one else Tweeted this yet, so I'm probably wrong.
First I heard about this demo as a tool for praising your own GitHub or homepage (which it does well) but I wanted to go meta and explore my own podcast.
There's a suggestion to let the app generate a timeline, and that feature worked really well (considering it's multiple websites and audio clips).
When I gave NotebookLM the links to my first two YouTube episodes, it did a decent job of summarizing the content (except for pronouncing POGOs, "Pee.Oh.Jos"). I have subtitles for each episode, but the third episode is too recent and the app wouldn't accept it (?).
Then I added Wikipedia, news stories, the third episode transcript, and a couple of other videos for context. BBC and CNN are forbidden. With this additional context, the topics got too disparate and repetitive, especially around platitudes (she was like a wholesome farmer, big money is sketchy, in this globalized world, will she actually face trial, transparent government is important).
NotebookLM likes to up the ante and call Alice a 'double agent'. And as much as I like when the AI hosts sound human, I don't like when they claim human experiences like "I saw her name pop up somewhere" or an aside that a specific interviewer is "really tough".
Finally I removed most sources and uploaded all three episodes as MP3s. I could see that they did a good job of transcribing and summarizing Filipino text (including a TikTok video that I'd downloaded).
The new version of the 'podcast' revolved around my Episode 2 (because it has English at the start, or from the Friends & Foes title?). The narration oddly starts in Alice's shoes ('you're the mayor of this totally normal town and now you're in the middle of an international incident') - when the obvious narrative is that you're a normal person wondering if your mayor is an international cybercriminal spy.

Real talk on the analytics - the podcast is getting almost all of its audience on YouTube.

Part of this goes back to a ramble that I've had before. It supposedly doesn't matter to an artist who sees or likes their art. But when I make an online course, or upload this podcast, it's going to be significant if there is no student or listener. Or if I've unintentionally made low quality clickbait.

I could see repurposing these skills for an AI-voice radio script for a future topic or fiction. Or maybe I need some help from AI research and voices to make a series of SE Asian true crime / political podcasts? More seriously, I think the next steps are going to be: