news

new wisdom report

google's notebooklm function, which can quickly generate podcasts from text, caused a strong response once it was launched.

and just recently, a game writer discovered that a mine-clearance book he published had been generated by notebooklm as a fake ai podcast!

seeing that his book was condensed into a 12-and-a-half-minute conversation between two non-existent people, author kyle orland exclaimed: i was shocked!

of course, if you listen carefully, you will find that there are still many problems in podcast audio. obviously, it cannot replace professional podcasts.

but even so, orland still lamented: podcast formats are a more engaging and interesting way for us to receive complex information.

it clearly points to a more fascinating future for generative ai than the dull replies text-based chatbots give us.

google's notebooklm was launched more than a year ago.

according to google, it functions "as a virtual research assistant that summarizes facts, explains complex ideas, and brainstorms new connections based on sources of your choice."

but just last week, google added the "audio overview" feature.

suddenly, notebooklm has a new selling point: a cold document can be turned into an engaging audio!

interestingly, google doesn’t use the word “podcast” anywhere, but says the feature can help you “summarize material, make connections between topics, and even crack some playful jokes.”

when wharton business school professor ethan mollick posted a post, he gave the product the correct name—“podcast.”

professor mollick rated this podcast as "the best ai demonstration currently available."

he also said it was disturbing.

of course, in ai podcasts, the problem of hallucinations is still unavoidable.

but professor mollick said that the ability to summarize, refine and process multiple large documents allows us to fact-check ai, which is its greatest value.

inspired by ethan mollick, orland decided to enter his book into notebooklm.

he is looking forward to what this 30,000-word book about the minesweeper game will look like if it becomes a virtual podcast.

just a few minutes later, orland saw the picture at the beginning of the article.

he experienced firsthand what it would be like to appear on npr's "pop culture happy hour" or be invited to appear on a tongue-in-cheek podcast.

orland found that the summary of notebooklm covered the main part of his book——

the prehistory of the games that inspired minesweeper; the uphill battle that business-focused microsoft waged over the windows entertainment pack in the 1990s; the moral panic surrounding minesweeper's preinstallation on millions of corporate and government computers; there are all kinds of surprising cheating controversies in gaming’s competitive scene.

overall, anyone who listens to this podcast can get a general idea of ​​what my book is about, just like many real podcasts.

although there are no large-scale hallucinations in this fake podcast, there are still some mistakes in small details, or assumptions that are not supported by the text.

for example, when discussing minesweeper's predecessor, mined-out, the ai ​​podcast said: "so this is where those blocks and flags come into play..."

however, mined-out does not have these two functions, it is all made up by ai.

obviously, if we need to use notebooklm's summary podcast to understand an academic article or news report, it will be very discouraging if similar illusions occur, or if some key parts of the book are missed.

however, we can use this kind of summary to achieve something "cliff notes" style, which is still very attractive for those large books that we do not have the time or interest to read in full.

and, unlike reading cliff notes, this concise podcast style provides a pleasant background noise, perfect for listening to while out for a walk or running errands.

it is this natural and humorous presentation that makes notebooklm’s new feature stand out from other ai products that can generate text summaries.

orland said it felt like he happened to be in a coffee shop and overheard two people discussing his book.

however, these two people do not exist in reality.

moreover, their praise for the book may have been algorithmically designed.

at first, a "podcast host" described the book as a story from "the land of floppy disks and dial-up modems."

such an interesting expression has not been used by the author in the book. the humor of ai made him laugh out loud.

the host then teased that "bill gates was sneaking around the microsoft offices," hinting at the author's favorite anecdote in the book.

subsequently, the two fake hosts discussed this anecdote in a very natural dialogue.

host 1: it's hard to deny the impact of something when your own ceo is secretly obsessed with it.

host 2: wait, are we talking about bill gates?

the back-and-forth dialogue style of the two-person "podcast" format also allows for some interesting but deviating topics from the main points of the book.

for example, the two ai hosts seemed a little distraught when discussing the worm movie star damsel in distress from mined-out.

host 1: i would like to ask, what other movies has chong starred in?

moderator 2: i’m afraid the details have been lost in the dust of gaming history.

moderator 1: so eventually the community came up with a more elegant solution.

host 2: let me guess. they created a new version of minesweeper.

host 1: that’s right.

host 2: that’s it for the old day.

the conversation between the two has a gentle and relaxed rhythm, but the information presented is very dense.

also, natural-sounding pauses and repetitions can help emphasize key points.

for example, host 1 talks about "this incredibly addictive puzzle game comes pre-installed on almost every computer," and another voice replies, "on every computer?"

or one moderator would say "we found a flaw in the way the original minesweeper generated random boards," and another voice would jump in and exclaim "a flaw!"

the atmosphere team is very energetic.

of course, there are still some problems with the conversation between the two ai dummies.

for example, the two voices seem to be jumping back and forth between "i read this book" and "i was surprised by the facts you shared in the book." it's hard to believe that either one is sincerely.

and an overly surprised reaction can also make people a little annoyed because they are tired of being stupid.

in addition, the ai ​​will break sentences in the wrong place or make some strange laughter.

still, it surprised orland.

he said, for example, that at certain moments, the way it elicited certain things subtly appealed to his authorial self-awareness, which served him well.

finally, orland concluded that after listening to notebooklm’s summary, he would imagine that in the near future, these fake podcasts would become his “main meal.”

especially those professional books or topics that are not likely to interest real-life podcasters.

by repackaging generative ai text as a "two people chatting" model, google puts an affable face on a technology that sometimes seems impersonal.