How I Made NotebookFM with NotebookLM

Simon Willison wrote about Google’s NotebookLM podcast generator yesterday and included some example outputs and some fun examples of having the program talk about itself – resulting in existential meltdown lol:

NotebookLM is effectively an end-user customizable RAG product. It lets you gather together multiple “sources”—documents, pasted text, links to web pages and YouTube videos—into a single interface where you can then use chat to ask questions of them. Under the hood it’s powered by their long-context Gemini 1.5 Pro LLM.

Once you’ve loaded in some sources, the Notebook Guide menu provides an option to create an Audio Overview:

These tools are so cool and moving really fast. A few people have asked me how I made the NotebookFM Episode of 301 over the weekend.:

Here’s how I made it:

Making NotebookFM

I added three initial bits of media to the Notebook:

• Wordrunning.guide collection (GDoc)
• Worlds a Walk-Though talk (Youtube)
• Myth-making Mechanisms Talk (Youtube)

Then I had it generate a summary of my own work.

I actually think if you’re interested in trying out NotebookLM as a tool, it’s worth testing it on your own corpus/body of work. You can calibrate on what it’s good at, where its limits are, and what it gets wrong.

NotebookLM gave me this 13m42s summery of my work:

Which I then transcribed it in my podcast editor Descript:

I then uploaded the transcript to ChatGPT (you can read the full exchange at the link) and gave it this prompt:

This transcript between chad and stacy is about my work. I'm jay springett. What I would like you to do is insert me as third participant in this conversation..

Other than cropping their side of conversation do not change what they have said as it's already recorded - Though I can crop splice and cut the existing audio file. Do not put words into chad or stacy's mouths I can't change anything. 

Please keep it feeling natural I would like for this whole thing to be about 1000 words

The output was about 90% of the way there. I then pasted in my side of the conversation, as generated by ChatGPT (with a few tweaks as I went), at the appropriate points. I deleted all the extra material from Chad and Stacy, then generated my side of the conversation using the voice model I made with Descript’s tools.

It’s meant for fixing single flubbed words here and there, not generating paragraphs from scratch—and for some reason, it always makes me sound Australian???? WTF?

That was basically it. I edited the show down to hit 301 seconds, set up the live captions over the video, and exported. The whole thing probably took about 90 minutes.

NotebookLM was only released the other day, and I haven’t yet seen anyone else insert themselves into a conversation. Dunno if I’m the first?

The podcast summaries that NotebookLM produces are very useful tool/technology, and I’m looking forward to getting to grips with the rest of the notebook tools.

Anyways it was a fun exercise.

As I’ve said before: Right now AI is the worst it’s ever going to be. There’s going to be much more of this and it’s only going to get better fast.