The AI Podcast Thing
Author: Peter Kaminski Issue: 2024-10-02
The AI Podcast Thing
by Peter Kaminski
Perhaps you've heard a clip of one of the “deep dive” AI-generated podcast episodes by Google's NotebookLM. Perhaps you haven't.
In either case, I urge you to do a deeper dive and explore the technology more for yourself. The investment is low, and I think you will be pleased with the return on investment. And, I apologize in advance, you will experience being a little creeped out along the way. Onward to the future!
The first step might be listening to a few examples. (Google sign-in required. If you really don’t have a Google account, it’s okay to skip down to where I talk about PDF2Audio.)
Here’s an episode about last week’s OGM Call, 2024-09-26. Or another instant classic, the one where the AI hosts learn that they’re not human. (No, these AI hosts don’t really think or feel–or indeed, even exist–it was just several stages of AIs synthesizing plausible-sounding AI podcast hosts and their plausible reactions. No AIs were harmed in the process.)
After the initial reaction to how realistic it sounds, you’ll realize it’s sort of a parlor trick. But if you think a little more past that, you’ll also realize that it’s a watershed moment, perhaps not unlike the launch of ChatGPT, and that there’s a lot more potential there that isn’t just a parlor trick.
A next step–remember, I said the investment was low–go to https://notebooklm.google.com/ and play around with NotebookLM for yourself. It is currently free, if you have a Google account. Create a new notebook, add some sources (PDFs or other documents, links to non-paywalled websites, YouTube videos, etc.), click the “Generate” button in the Audio Overview section and let it get started generating. Click some of the other “Help me create” buttons. Check out the results. Reflect on what you could do if you gave it more or better source materials.
If you want to try something similar that you pay a small amount of money for, rather than getting it for free from Google, you can get an OpenAI API key and then try PDF2Audio. It’s not quite as stunning as NotebookLM’s podcasts, but still, very interesting and useful, and in some ways, easier to control. For text generation model, you should probably select gpt-4o-mini or gpt-4o–most people don’t have API access to o1 yet. (With some colleagues, I am building AI Coaching Forum as a place to learn how to make sense of paragraphs like this, but it’s not ready for prime-time yet, just early adopters. If that describes you, though, email me.)
Here are some additional resource links and explanations:
- OpenAI API key: Anybody can sign up and pay for OpenAI services by the token.
- PDF2Audio is a simple open source application that just organizes your documents and then sends them to OpenAI's LLM and text-to-speech APIs to generate podcast transcripts and audio. It works better than I would think it would, although I miss NotebookLM's hosts’ disfluencies.
- The OGM 2024-09-26 notebook I built to create the podcast episode uses just the text transcript from the YouTube recording of the call. It's not the best application for NotebookLM; you should try to use more documents and/or videos for your notebooks.
- More background on the “AI Hosts’ Existential Crisis” episode from the originator: I want to offer some clarity.
- Steven Johnson (editorial director of NotebookLM) on Using NotebookLM to remember everything you've read. One of the key takeaways applicable to any kind of source material, not just quotes: “A lot of folks have asked how I get 8,000 quotes into a single notebook, when currently notebooks are limited to only 50 sources. The key thing is that each source can have up to 500,000 words in it, so if you can compile your quotes into giant documents, you can easily fit a quote collection of that size into a single notebook.” Got that? Giant documents.
- Two tweets from Andrej Karpathy: NotebookLM is quite powerful and worth playing with, and Deep Dive is now my favorite podcast.
- Some description of the methodology from Steven Johnson on the Hard Fork podcast.
Related:
- Peter Kaminski (author)
- 2024 (year)
- Topics: