Can a NotebookLM podcast surface new mysteries of Satoshi?
making a podcast just became a lot easier, but there are pitfalls to watch out for
NotebookLM’s most impressive feature is how it can create a 2-person podcast out of a bunch of documents you feed it. Upload, link, or paste any collection of text documents and click “generate”. After a few minutes of churning you get a podcast of 2 people talking about the topics addressed throughout the text.
But is it good?
The results are impressive, but maybe not for the reason you’d hope. NotebookLM’s podcast absolutely nails the casual feel of 2 humans bantering about a topic just like you’d hear on a podcast. There are stutters, interruptions, and turns-of-phrase throughout the commentary that make the result sound very human. In the category of “authenticity” I’d grade it an A.
In order to try it out I decided to create a 2-person podcast out of the Satoshi-Sirius (Malmi) emails (https://mmalmi.github.io/satoshi/) that were published earlier this year.
NotebookLM warns the audio creator that “Audio Overviews are not a comprehensive or objective view of a topic, but simply a reflection of your sources.”
So it doesn’t really summarize the content in a meaningful way or contribute new information to the world. If I feed in a bunch of source material I would hope I could listen and get some sort of key takeaways or important overview of the information input. It’s certainly not a replacement for reading the information, but it also doesn’t really give me the important/broad strokes. It picks some details and produces humanlike conversations around those. This comes off as “a cool party-trick”, but not useful as a great research tool yet. So in the category of “overview quality” I’d grade it a C.
Obviously this stuff is improving very quickly and it doesn’t feel like we’re too far off from having much better AI generated podcasts. But this also opens a bigger question in my mind: will the future of podcasting include new durable AI personalities? Will there be a Rogan, Fridman, Cowen, or Dwarkesh of AI?
While the voices generated by NotebookLM do sound authentically human, today they have the depth of cardboard cutouts. I don’t know much about their perspective or life story. I can’t follow from episode to episode and understand where their world-view might align or diverge from mine. The reason I like human podcasts is that I get to learn something by listening into a conversation. Often I’m familiar with the host and mostly unfamiliar with the guest so I get to tune-into the host’s perspective which I’m gathering over time while learning about the guest’s special knowledge in this episode. Will AIs generate longitudinal, interesting hosts with a perspective? This doesn’t seem impossible to imagine with another breakthrough or two, but for now at least we’ve got “a cool party trick”.