Google has an AI feature called NotebookLM. You can upload files to it, web links, YouTube videos, and so on, and the service supplies you a Retrieval Augmented Generation environment. I use it often for research and for creative work. Recently, with very little fanfare, they rolled out a new feature called "Audio Overview." This feature creates an AI podcast where a male and female host discuss the contents of the notebook.
The audio above is from an experiment I did, uploading my first eleven essays and short stories from this Substack. The results are uneven, but two weeks ago if you had played this for me, I would have thought "Who are these hosts who read my essays, made a podcast about them, and didn't bother to tell me they were doing so?" Even now, I think the only reason I can still tell it is AI generated is because I clicked the button that created it. If I didn’t I would excuse the few audio glitches, mispronunciations, and the occasionally rambling conversation as just being another mid-tier podcast.
Keep in mind I had no control over the generation: I didn’t tell it the tone to adopt, the order of the works to review, or even what to say about them.
I don’t agree with the system’s “interpretation” of my short story Career Day, but I think it is trying to interpolate between the story and the other writing I gave it. It’s a B+ student attempt. But that is still incredible, that it is a B+ effort.
I’m still a bit dumbstruck at just how well it works. I’ve given it the most random content - Clark Ashton Smith short stories, single essays of mine, academic articles, a 117-page monograph about Soviet peaceful nuclear explosive experiments, a supplement for the tabletop roleplaying game GURPS - and it comes up with something sensible and even more incredibly listenable each and every time.
All I can say is: try it for yourself.
First: create a New Notebook
Second: upload a single document or link (this will give you a better sense of how it works before you experiment with generating based on corpuses of documents)
Third: From the Notebook Guide click to generate an Audio Overview
Then wait 5-15 minutes, and it’ll be done. Be sure to experiment with the Chat feature as well.
To end with some throwaway thoughts to prompt myself to keep thinking about this:
This feels fantastically useful for studying or comprehension. You can upload something you’re about to start reading or studying, and the system will give you a reasonably competent overview of the content. I haven’t been able to stump it yet, though I will post when I find the conditions that cause it to lose its metaphorical mind. You can learn more by checking what the system said about the document as you read, as it will provide a contrastive experience that engages your attention more.
Like with my first experience with ChatGPT 3.5 or Midjourney v5, it’s good enough I can imagine a lot of ways it could be better. I’d like the ability to get the mode to focus particular details of the texts, or explore a certain type of interpretation, or engage in more free associative discussion and interpolation. In a future experiment I’ll try putting in two or three very conceptually distinct items and see whether the model tries to interpolate between them or just discusses them serially.
I guess there are pareidolia and uncanny valley issues here, but they’re not something I experience personally. The realism of the content doesn’t make me feel like I can’t trust things I see and read online: checking sources remains a valuable practice, as does critical thinking about why certain stories and ideas are being promoted, whether or not they are true.
For reference, the essays I gave the model were:
AI Podcast Discusses My Essays