An author talks about the results of using AI to generate podcast-style audio from his book



An author wrote about his experience of loading his own book into Google's AI ``

NotebookLM ,'' which analyzes text to help generate ideas and organize thoughts, and feeling the ``rich expression'' that the generative AI exudes.

Fake AI “podcasters” are reviewing my book and it's freaking me out - Ars Technica
https://arstechnica.com/ai/2024/09/fake-ai-podcasters-are-reviewing-my-book-and-its-freaking-me-out/

Kyle Orlando, who wrote a 30,000-character book about the game Minesweeper, said he had been actively following the growth of generative AI, but didn't know much about the performance of Google's text summarization AI, NotebookLM. Then, inspired by a social media post by Ethan Morlick, an AI expert at the Wharton School of the University of Pennsylvania, who had 'generated audio from my book using NotebookLM,' Orlando decided he wanted to try it himself.




Orlando, who typed in the book's contents, says NotebookLM produced a highly-conceived summary that covered all the key parts of the book, including 'the history of the games that inspired Minesweeper, Microsoft's uphill battle to become more business-oriented in the '90s, the moral panic over Minesweeper being pre-installed on millions of corporate and government computers, and the startling cheating controversy that surrounded the competitive scene.'

What made NotebookLM stand out was that it didn't just summarize the contents of the book, but also introduced it in a casual conversation. NotebookLM used two different voices to generate the following dialogue-style conversation:

Voice 1: In the end the community came up with a more elegant solution.
VOICE 2: Let me guess, I made a new version of Minesweeper.
VOICE 1: That's right.
Voice 2: So you stopped with the old version.

The dialogue had natural pauses, and jokes were sometimes thrown in, making for a friendly presentation. Orlando said he felt the summaries had more depth, but he did feel a little uneasy, such as when a voice actor who had already read the book appeared to have not read it in the next scene. There were also over-reactions such as 'You've got to be kidding me!' and 'You've got to be kidding me!', and some laughter that seemed to cut off midway through.



According to Orlando, the '

hallucinations ' that generative AI is prone to were not noticeable, but he sometimes got the details wrong or used analogies that were not written in the main text.

Orlando commented, 'Overall, the level of completion was such that even people who haven't read the book could get a rough outline. Personally, I would hesitate to use NotebookLM for academic papers or journalistic articles, but it would be useful for quickly understanding the contents of a thick book that you don't have the time or energy to read in depth. I can easily imagine a future where this kind of AI will mix with humans to deliver podcast-style audio.'



Related Posts:

in Software, Posted by log1p_kr