Agreed. I finally got around to listening to a NotebookLM generated podcast this weekend, and found it absolutely unlistenable.

For some reason the LLM seemed to latch on to the idea that one host should say something, and the other host should just repeat a key phrase in the sentence for emphasis -- and say nothing else, and they'd do it like multiple times in a sentence, over and over again throughout the entire thing.

Slightly less weird -- but it seems like the LLM caught on that a good narrative structure for a two-host podcast is that one host is the 'expert' on the topic, and the other host plays dumb and ask questions. Not an unreasonable narrative structure. Except that the hosts would seamlessly, and very weirdly, switch roles constantly throughout the podcast.

And ultimately the result was just a high level summary of the article I had provided. They told me in the intro and the outro about the interesting parts they were going to dive into, but they never actually got around to diving into those parts.

I tried it with something fairly abstract about a decision I was working on making. Fed it a bunch of information, details about me, my background, the factors I'm considering in the decision and the impact of getting it right/wrong.

It was interesting. I certainly wouldn't say it was useless, I think the contrived dialogue actually touched on some angles I hadn't considered and I think it was useful. Not in a 'oh shit it's clear to me now' kind of way but it definitely advanced my thinking.

I find it as listenable as the podcasts the people I'm around play in their cars during roadtrips

To me that's absolutely unlistenable, but to them its interesting and engaging. I find NotebookLM replicates that perfectly. Its not at all the issue that the OP encountered with the Hawaii news service, as those lacked tone and pronounciation, which NotebookLM would not.

Regarding the outcome being a long winded summary, yeah thats what I see about the aforementioned 2 hour podcasts as well. They take the intro paragraph of a wikipedia article, pretend that the topic is a novel mystery they just discovered through hours of scouring microfilm at a municipal library, and then interject each other every other word with nonsense, before getting back to the point for just one sentence.

> For some reason the LLM seemed to latch on to the idea that one host should say something, and the other host should just repeat a key phrase in the sentence for emphasis -- and say nothing else, and they'd do it like multiple times in a sentence, over and over again throughout the entire thing.

cough https://en.wikipedia.org/wiki/ELIZA

Hmm, does Google not use the Claude or ChatGPT API for it? I still don't hear of people chatting with Gemini nearly as often as with the other two.

Edit: looks like the subreddit is still called Bard. Well played, Internets. https://www.reddit.com/r/Bard/comments/1g0egad/gemini_vs_not...

There are a lot of "uh huh", "yeah", "agree" types of statements in odd places as well.