In other uncanny-valley AI voice news...
Google has this new thing called "NotebookLM," which allows you to upload any document, click a button, and then a few minutes later receive an entire AI-generated podcast episode (!) about the document. The generation seems to occur somewhat faster than real-time.
(This is currently offered for free as a demo, all you need is a Google account.)
These podcast episodes are... they're not, uh, good. In fact, they're terrible – so cringe-y and inane that I find them painful to listen to.
But – unlike with the "AI-generated content" of even the very recent past – the problem with this stuff isn't that it's unrealistic. It's perfectly realistic. The podcasters sound like real people! Everything they say is perfectly coherent! It's just coherently ... bad.
It's a perfect imitation of superficial, formulaic, cringe-y media commentary podcasts. The content isn't good, but it's a type of bad content that exists, and the AI mimics it expertly.
The badness is authentic. The dumb shit they say is exactly the sort of dumb shit that humans would say on this sort of podcast, and they say it with the exact sorts of inflections that people would use when saying that dumb shit on that sort of podcast, and... and everything.
(Advanced Voice Mode feels a lot like this too. And – much as with Advanced Voice Mode – if Google can do this, then they can presumably do lots of things that are more interesting and artistically impressive.
But even if no really especially likes this kind of slop, it's highly inoffensive – palatable to everyone, not likely to confuse anyone or piss anyone off – and so it's what we get, for now, while these companies are still cautiously testing the waters.)
The first thing I tried was my novel Almost Nowhere, as a PDF file.
This seemed to throw the whole "NotebookLM" system for a loop, to some extent because it's a confusing book (even to humans), but also to some extent because it's very long.
I saw several different "NotebookLM" features spit out different attempts to summarize/describe it that seemed to be working off of different subsets of the text.
In the case of the generated podcast, the podcasters appear to have only "seen" the first 8 (?) chapters.
And their discussion of those early chapters is... like I said, pretty bad. They get some basic things wrong, and the commentary is painfully basic even when it's not actually inaccurate. But it's still uncanny that something like this is possible.
(Spoilers for the first ~8 chapters of Almost Nowhere)
The second thing I tried was my previous novel, The Northern Caves.
The Northern Caves is a much shorter book, and there were no length-related issues this time.
It's also a book that uses a found-media format and includes a fictitious podcast transcript.
And, possibly because of this, NotebookLM "decided" to generate a podcast that treated the story and characters as though they existed in the real world – effectively, creating fanfiction as opposed to commentary!
(Spoilers for The Northern Caves.)
