Wednesday, December 18, 2024

AI digests repetitive scatological document into profound “poop” podcast

Must read

Enlarge / This AI prompt stinks… or does it?

Aurich Lawson

Imagine you’re a podcaster who regularly does quick 10- to 12-minute summary reviews of written works. Now imagine your producer gives you multiple pages of nothing but the words “poop” and “fart” repeated over and over again and asks you to have an episode about the document on their desk within the hour.

Speaking for myself, I’d have trouble even knowing where to start with such a task. But when Reddit user sorryaboutyourcats gave the same prompt to Google’s NotebookLM AI model, the result was a surprisingly cogent and engaging AI-generated podcast that touches on the nature of art, the philosophy of attention, and the human desire to make meaning out of the inherently meaningless.

Analyzing Poop & Fart written 1,000 times – Creating meaning from the meaningless
byu/sorryaboutyourcats in notebooklm

When I asked NotebookLM to review my Minesweeper book last week, commenter Defenstrar smartly asked “what would happen if you fed it a less engrossing or well written body of text.” The answer, as seen here, shows the interesting directions a modern AI model can go in when you let it just spin its wheels and wander off from an essentially unmoored starting point.

“Sometimes a poop is just a poop…”

While Google’s NotebookLM launched over a year ago, the model’s recently launched “Audio Overview” feature has been getting a lot of attention for what Google calls “a new way to turn your documents into engaging audio discussions.” At its heart is a LLM similar to the kind that powers ChatGPT, which creates a podcast-like script for two convincing text-to-speech models to read, complete with “ums,” interruptions, and dramatic pauses.

Experimenters have managed to trick these AI-powered “hosts” into what sounds like an existential crisis by telling them that they aren’t really human. And investigators have managed to get NotebookLM to talk about its own system prompts, which seem to focus on “going beyond surface-level information” to unearth “golden nuggets of knowledge” from the source material.

The “poop-fart” document (as I’ll be calling it for simplicity) is a pretty interesting test case for this kind of system. After all, what “golden nuggets of knowledge” could be buried beyond the “surface level” of two scatological words repeated for multiple pages? How do you “highlight intriguing points with enthusiasm”—as the unearthed NotebookLM prompt suggests—when the document’s only oft-repeated points are “poop” and “fart”?

Artist's conception of a portion of the poop-fart document, as fed to NotebookLM.
Enlarge / Artist’s conception of a portion of the poop-fart document, as fed to NotebookLM.

Here, NotebookLM manages to use that complete lack of context as its own starting point for an interesting stream-of-consciousness, podcast-like conversation. After some throat-clearing about how the audience has “outdone itself” with “a unique piece of source material,” the ersatz podcast hosts are quick to compare the repetition in the document to Andy Warhol’s soup cans or “minimalist music” that can be “surprisingly powerful.” Later, the hosts try to glean some meaning by comparing the document to “a modern dadaist prank” (pronounced as “daday-ist” by the speech synthesizer) or the vase/faces optical illusion.

Artistic comparisons aside, NotebookLM’s virtual hosts also delve a bit into the psychology behind the “very human impulse” to “search for a pattern” in this “accidental Rorschach test” and our tendency to “try to impose order” on the “information overload” of the world around us. In almost the same breath, though, the hosts get philosophical about “confront[ing] the absurdity in trying to find meaning in everything” and suggest that “sometimes a poop is just a poop and a fart is just a fart.”

Latest article