Excited and Unsettled
Recently, Google released an updated version of one of its AI tools: NotebookLM. As of this writing, is available for free, and I expect it will remain accessible for quite a while, as it is still in a mode they call “experimental.”
There are a lot of Large Language Models out there, most notably ChatGPT, that do a lot of interesting things. I have found them to be useful as “research assistants:” tools that help me stir up new ideas – and sometimes suggest new connections – but never anything I have taken as anywhere near reliable when looking for factual, verifiable information. They are completely prone to what have become known as “hallucinations:” completely fictional creations that masquerade as fact.
But this newest tool from Google has done several remarkable things – and that’s just with one of its myriad of features.
NotebookLM allows you to identify a number of “sources” – files of yours, or links to the web – that you want to query the AI about. I am pleased to report that it doesn’t consume your files into its model, or train its own model unless you allow it.
While previously (and in other LLMs) these sources were typically text, and occasionally pictures, NotebookLM allows you to use video and audio files, which in and of itself is nice: you can ask it for a summary of a video, and query it about the video’s content. But that’s not what excited me.
NotebookLM has a deceptively simple button called “Deep dive conversation” in a section called Audio Overview. It feels like black magic.
Giving NotebookLM a single file and then generating the deep dive results in an audio recording of two completely computer-generated “hosts” discussing the contents of the file, including taking a “deep dive” into its deeper meanings. The conversation is nearly flawless (but as of yet only in English): they interrupt each other, use a lot of idioms, and I challenge you to easily determine that they are not humans interacting. But that’s just the beginning.
I’ve fed NotebookLM a number of stories, particularly those that are not widely known, and have been astonished at the deeper meanings they’ve uncovered. For example:
Here is a story from my book (I hold the copyright to it), which is my own retelling of an obscure Jewish story from Syria – The Tear Jar. Read it first.
Now, listen to the deep dive conversation. It’s only 8m42s, and you can listen to it here.
You can exhale now.
There’s a lot to say about this, and I will do so in subsequent posts. But I’d like to draw your attention to one really critical element of this audio overview, which has nothing to do with AI: it’s a conversation.
This is a truly brilliant choice by the designers of this system. Imagine that you had a single voice in a steady monologue sharing the same information. I maintain it would not have been as captivating as this conversation.
Why? Because, in large part, it emphasizes the narrative structure of its presentation: it (sometimes clumsily, I agree) attempts to build suspense, offers surprises, and so on. All the techniques that storytellers have been using since, well, forever.
There are some real limitations – for example, I gave it two chapters of the Bible to work on, and it relied far more on existing commentaries and didn’t really uncover anything new. But when I gave it three stories at once from three different cultures (and again, fairly obscure), it did a good job at spotting a unifying theme between them that I had never considered.
Should we, as storytellers, as people, be worried?
Well, when the match replaced flint stones as a way of starting fire, it was a great tool with a lot of convenience. But it also meant that fires could be started accidentally far more easily, and it opened the door to incendiary devices. So I think, like any other new tool, it has to be understood as only a tool, and probed for its hidden biases. But again, more on that later.
For now, try visiting https://notebooklm.google.com/ and see what adventures await you.
Oh, and just for hoots and hollers, here’s what NotebookLM said about this blog post: AudioLink
Enjoy!