This works pretty well. I tried it with this guidance prompt:

    You are both pelicans who work as data
    journalist at a pelican news service.
    Discuss this from the perspective of
    pelican data journalists, being sure
    to inject as many pelican related
    anecdotes as possible

Against this article: https://simonwillison.net/2024/Oct/17/video-scraping/

You can listen to the 7m40s resulting MP4 here: https://simonwillison.net/2024/Oct/17/notebooklm-pelicans/

Example snippets:

    You ever find yourself wading through
    mountains of data trying to pluck out
    the juicy bits? It's like hunting for
    a single shrimp in a whole kelp forest,
    am I right?

And:

    The future of data journalism is
    looking brighter than a school of
    silversides reflecting the morning sun.
    Until next time, keep those wings
    spread, those eyes sharp, and those
    minds open. There's a whole ocean
    of data out there just waiting to be
    explored.

Aside from being invited to a few events (Google I/O, OpenAI DevDay twice, an upcoming Anthropic hackathon that I applied for and was accepted) I've had no compensation from any of the LLM vendors.

I've been invited to a few alpha/beta previews for all three of OpenAI/Google/Anthropic, and I've received API credits as an attendee of the DevDay events.

I write about this stuff because it's really interesting to me, and super fun to explore.

Fwiw, I'd say this falls under the guidelines "Please don't post insinuations about astroturfing, shilling, brigading, foreign agents, and the like. It degrades discussion and is usually mistaken. If you're worried about abuse, email hn@ycombinator.com and we'll look at the data."

we really want to know who is behind big pelican

NotebookLM is contributing to fake podcasts across the internet, with over 1,300 and counting:

https://github.com/ListenNotes/ai-generated-fake-podcasts/bl...

Google is taking a different approach this time, moving quickly. While NotebookLM is indeed a remarkable tool for personal productivity and learning, it also opens the door for spammers to mass-produce content that isn't meant for human consumption.

Amidst all the praise for this project, I’d like to offer a different perspective. I hope the NotebookLM team sees this and recognizes the seriousness of the spam issue, which will only grow if left unaddressed. If you know someone on the team, please bring this to their attention - Could you please provide a tool or some plain-English guidelines to help detect audio generated by NotebookLM? Is there a watermark or any other identifiable marker that can be used?

Just recently, a Hacker News post highlighted how nearly all Google image results for "baby peacock" are AI-generated: https://news.ycombinator.com/item?id=41767648

It won't be long before we see a similar trend with low-quality, AI-generated fake podcasts flooding the internet.

Where do you get the "low-quality" part from - my experience with NotebookLM is that they create much higher quality, more informative, more fact based, and more concise podcasts than 99% of the stuff I listen to. I've mostly switched entirely over to NotebookLM for my podcast listening. They, generally, offer a far higher quality experience from my perspective.

Maybe you have the problem backwards - we accidentally end up listening to non NotebookLM podcasts?

A coworker fed some EU trade regulation page and its official FAQ to NotebookLM, and I was quite impressed with the results.

It was factually accurate, and presented the topic in a manner that was easy to digest and kept it interesting.

I didn't plan to but ended up listening to the whole thing, and I normally don't enjoy the podcast format.

For someone new to the topic, it'd be a pretty great intro compared to reading the official pages.

It's interesting assumption that by virtue of being AI generated, it's considered bad/fake. 20 years ago, people hated how photoshop changed the photo design industry, NotebookLM is knocking on the door now.

I'm excited by AI, but I've also tried using this specific one to generate a podcast based on one of my own blog posts and will only try again due to this product announcement rather than because I think the state of the art is already "there".

On the plus side, the speech is almost perfect; so good, that I sincerely hope the voices themselves are never fully under user control.

With regards to the actual summary of the content I gave them, I would say they are grade B: only mostly correct, they're still inventing things I didn't say and missing things I did say.

That's not to say humans don't make mistakes, I still consider this objectively impressive, that is able to reach even this level was SciFi when I was a kid — but why waste time on a grade-B podcast when the AAA-tier costs you as a consumer a 30 second advert?

Personally, I hate even the idea of an AI made podcast, because to me podcasts are personal and emotional. They're about the individual humans who make them. They're not just a source of "information".

I'm glad there are different kinds of podcasts for different people now.

I've always absolutely hated the focus on the individual humans and their personalities behind the podcast, and wished they'd be a better source of well-structured "information".

I never listened to a podcast I didn't get frustrated with, even at 2x speed. These NotebookLM podcasts have been exactly what I've always wished podcasts were.

I don't know... I think it's pretty cool. In fact, I just found a podcast talking about this very thread (about a minute in, your comment even gets discussed ;-) https://drive.google.com/file/d/130s6OzcfsZam8V-6S5ugKmc0M7O...

I may have to re-task my paperclip making AI on generating podcasts...

Have you listened to any audio overviews in NotebookLM? They can be surprisingly good.

That has not been my experience. The podcasts are vapid and full of cliches.

Interesting, are there any podcasts in particular that you recommend? Everything I’ve heard from it just seems like the most banal, cookie cutter stereotype of a podcast with nothing but extremely surface level summarization of a given article, peppered with random cliches and fake sounding reactions “Wow! ok, so let’s hear more about that. I’m intrigued!” “OK, let’s dive deep.” Etc.

DeepDive AI - I'm addicted to it.

That does not appear to be an AI generated podcast.

There’s a normal, human generated podcast called Deep Dive: AI (https://deepdive.opensource.org/). There’s also a confusingly similar named podcast Deep Dive AI that appears to only have one episode and is NotebookLM generated. Which one are you referring to?

Edit: If I'm understanding it correctly after some googling, supposedly the "name" of all podcasts generated by NotebookLM is "Deep Dive"? That's just confusing.

Assuming Google retain the current two voices for the audio overviews, it will rapidly become obvious to most people where the podcast came from. I've seen "creators" on YouTube running NotebookLM-generated audio through (e.g.) ElevenLabs to change the voices, but this invariably degrades the quality.

This doesn't strike me as much of a problem as it appears for you. What are the biggest issues you foresee?

I'm an avid podcast listener, but I already ignore 99.9% of podcasts out there. I'm not concerned that this is going to become 99.99%.

If these AI generated podcasts are all bad, I will just continue to ignore them. If some turn out to be good, it seems like a win to me.

If you're worried about an existential "what happens to the world if all media is machine generated", I guess I'm willing to hop on the ride and see what we find out.

99.9? There are roughly 3mm podcasts out there right now - I listen, regularly, to about 10 over a year (in any given week maybe 3-4). I'm therefore ignoring 2,999,990 or 99.9997% of podcast. I definitely agree with you that this isn't a problem.

(Also - ironically, one of the podcast out of those 10 that I listen to regularly - it's the Deep Dive on AI. A NotebookLM production! )

It could poison the well - make it hard for people to find new good podcasts, and reduce discovery and revenue. Also they could fragment our society even more, disconnect people from people. Doesn't seem worth the risk.

If people want to listen to AI generated podcasts, they can just make them themselves. They don't need publishing on a platform alongside human-made podcasts. If I was Apple, who ultimately control curation of podcasts, then I'd prevent them. After all, Apple Intelligence will soon do as good a job of making your custom podcast if that's what you want.

This is like saying: “Text based LLMs should do more to stop people from publishing the results of what they produce”

NotebookLM seems wonderful for digesting various content in an alternative way. It’s not a “fake podcast” either.

Nobody is saying that the audio output should or should not be published somewhere. That’s a user decision for both publishing and subscribing.

Indexes and discovery on the internet is where you advocate policing instead of nit picking a useful tool.

It sounds more like we should ban email and all email providers should consider the problem of email spam which traditional mail didn't have because no one could afford that many envelops and stamps.

Or like we should go back to carts because cars are noisy and not only that but might collide with pedestrians and not only that, might even collide among each other.

Instead of containing the tools and curtailing the progress (email and cars) we should probably try to contain and curtail abusers. Very hard to do, I know but the right thing to do.

> it also opens the door for spammers to mass-produce content that isn't meant for human consumption.

What's new? Every novel class of genAI product has brought a tidal wave of slop, spam and/or scams to the medium it generates. If anyone working on a product like this doesn't anticipate it being used to mass produce vapid white-noise "content" on an industrial scale then they haven't been paying attention.

This is definitely not a new issue.

What I’m aiming for is to ensure that the NotebookLM team is aware of the impact and actively considering it. Hopefully, they are already working on tools or mechanisms to address the problem—ideally before their colleagues at YouTube and Google Search come asking for help to fight NotebookLM-generated spams :)

It's certainly easier for the creators of genAI to build detection tools than for outsiders to do so. AI audio detection is a hard problem - https://www.npr.org/2024/04/05/1241446778/deepfake-audio-det...

> What I’m aiming for is to ensure that the NotebookLM team is aware of the impact and actively considering it.

What is the impact? Have any of them attracted an audience of any meaningful size? If a month from now there are 1.3 million generated podcasts, what do you anticipate the fallout to be?

> If a month from now there are 1.3 million generated podcasts, what do you anticipate the fallout to be?

Is this a rhetorical question? Because the answer for podcast indexing and search services is surely pretty obvious.

Why is it a problem? There's even more material for those services now and for their customers, the value these services can provide is even higher.

Podcasts - episodic radio shows hosted on Apple Music and Spotify - haven't been around for very long. Not long enough to have kids being tutored in making podcasts and then becoming adults with that sentimental hobby, like with playing violin or oil painting. If you believe that the "Human Authenticity Badge" is meaningful for podcasts, it's complicated: traditions play the biggest role in the outrage you are trying to spin, not an appeal to slop and spam, which of course, there is already a ton of low quality podcasts, music and art written by real people for no nefarious purpose whatsoever. Like with many of these posts, which are really common on HN, there isn't a sensible remedy suggested besides pointing the fingers at some giant corporation, and asking them to do something impossible.

If you care a lot about podcast quality, go and make your own podcast service with better discovery. Once you realize the antagonist was collaborative filtering, made possible by non-negative matrix factorization dating from the year 2000, and not AI, you will at least have learned something from the comment, instead of just feeling better. And then, how do you propose to curate by hand, and why would someone choose your curation over the New Yorker's? And maybe those very purists, trying to make everything sentimental, accusing everyone of slop and spam - well, why do so many creators thrive and ignore the New Yorker's opinion about them entirely? Perhaps curation is not only not scalable, but also wrong. Difficult questions for listeners and podcast authors alike.

Only 1300? I imagine it would be soo many more.

It’s definitely more than that.

The 1,300+ shows are just the ones recently removed from Listen Notes.

Give it a few days, and I’m sure the number will double, quadruple, and continue to grow. :(

Counterpoint: Most podcasts were utterly worthless before AI too. The world will do fine losing a few mattress ad vehicles.

Like other data, provenance suddenly matters a lot. From my POV, that's good. Not all data sources are created equal, and this is putting it into stark enough relief it might actually change the landscape. (In case it isn't obvious, I strongly believe most of the Internet was garbage well before LLMs. We just called it "SEO". Still garbage)

> Counterpoint: Most podcasts were utterly worthless before AI too.

Yet more "but humans also".

I generally agree, but when AI generated content is actively trying to avoid being labelled as “AI generated” it kinda gets depressing. Because in the end, it will just make the entire industry “seem” worthless, akin to AI generated pictures.

I’d rather let the end user know if it was made by humans or not, and let the marker decide. If people love listening to such content, let it be. But hiding how it was made, feels a bit disingenuous.

So what do you propose Google do to prevent this from happening?

The comments' default remedy is tribal: "The only moral content is my content." We sort of used to live in that world under the studio and TV networks system. Most consumers would say, it was not so bad, maybe better even.

Of course, the commenter never says this, living in the world today, where the writing he likes would never be published by the New York Times like it is on Twitter, the TV he likes would never be offered for free like it is on YouTube, and the music he likes would never been offered for pennies on Spotify. Some meaningful creators will lose from every remedy you could think of, where Google "something somethings" AI. Maybe the root problem is generalizing.

I created a “podcast episode” (???) of my personal blog (not trying to get traffic to it. It’s more of a journal) using NotebookLM. It sounded just as bland and overproduced as a “professional” podcast by NPR like “Planet Money” and “The Indicator”.

Whether that is saying how high quality NotebookLM is or how low quality NPRs podcast are is an exercise for the reader.

The only reason “Stuff you should know” is better is because of the random off topic discussions they go into and that’s not a complaint about SYSK.

> Is there a watermark or any other identifiable marker that can be used?

The problem with this is it's not feasible long-term, or even medium-term - as soon as a watermarking system is implemented, a watermark-removal system will be created.

(Happy to be proven wrong)

Geez, I hope there aren't people like you working at Google

Well which one is it? Are the podcasts low quality or not? If they are, what the hell are you worried about? To be worried about, idk, disinformation from podcasts of all things is absolutely silliness. Won't someone think of the... podcast audiences? Fuckin what dude?

I am late to the Google's AI party but... My personal impression (might be wrong) is that Google's breadth and depth of AI tools is heavily underrated ranging from Notebook LLM to AI studio. Too good as far as I have tried.

Google of course is the birthplace of attention is all you need.

Nice, I've only scratched the surface of Notebook LM, mainly for dumping lots of component reference material (datasheets, reference guides, application notes, etc). The text querying works great, but the audio overview wasn't very useful when it stuck to the high level of the content. With some ability to steer the topic out might be quite useful!

My product https://reasonote.com allows you to generate podcasts as well, and it's had this feature for a few weeks.

Improvements over NotebookLM:

(1) You can start with just a subject, and you don't need a full document to begin (though you can do that too![1])

(2) The podcast generates much faster

(3) The podcasts are interactive -- you can ask the hosts to change direction mid podcast, and they will do so.

(4) (Coming soon) You'll be able to make a Spotify-style Queue of Podcast topics, which you can add to as you encounter new ideas.

The primary tradeoff is that the voices / personalities are somewhat less engaging than NotebookLM at this time, though this will be dramatically improved over the coming months.

This is all in addition to the core value proposition, which is roughly "AI Generated Duolingo for Any Subject".

It's early days, but I'd love for you all to check it out and give me feedback :)

[1] Documents are currently heavily length-limited but this will be improved shortly

I was using this yesterday. I dumped all postmortems for an aspect of our infrastructure into a notebook and could then ask it to pull out common themes. It was remarkably effective. I also generated one of these "audio overviews" (aka podcasts) and it was great.

There was a vast improvement in quality from giving it a prompt when generating the overview. The generic un-prompted overview was for entirely the wrong audience, in our case users of our infrastructure rather than the developers. When instructing it to generate an overview for the SRE team and what they should focus on it was far better.

Was it useful for our in-depth analysis, no. Would I listen to one based on the last 100 postmortems for a new team I joined, absolutely. As an overview it was ideal, pulling out common themes from a lot of data and getting some of the vibe right too.

This is awesome! I have actually been using NotebookLM to create daily digests of HN and publish them to YouTube: https://www.youtube.com/@HackerCasts

I'm still getting the tooling right so that the videos will get made in a better and more consistent schedule.

Google Illuminate recently also introduced a customization feature. I use this customization with it:

audience=technical, duration=long, tone=professional & engaging

Surprised this was not there from the beginning. It can result in much better output. My problem with the default prompt is that it often is just two equally "knowledgeable hosts" kind of just bouncing information back and forth. With being able to customize the prompt you can create a kind of "explainer" and "listener" dynamic among the hosts that really helps the overall flow of the episode.

Something like this:

The two podcast hosts have very different levels of knowledge on the topic. The first host is the expert on the topic and explains the subject and the details to the second host. The second host has very little existing knowledge about the subject but will react to the information and ask follow up questions.

I think the default prompt attempts something like that -- one host does tend to take a more questioning role, and I've even had a few pods introduce one of them as a guest expert. But it seems like it sometimes loses track of who's who in that dynamic.

Here's an open source version that generates Podcasts:

https://github.com/souzatharsis/podcastfy

Developer's twitter: @souzatharsis

Not an improvement for me. I've been instructing NotebookLM for weeks now already by including the instructions into the sources. That way I have version control on my prompts and can easily drag into the sources upload. This requires finding my instructions and copying and pasting, there's also a 500 character limit which is very small, I have over 2000 characters for my standard prompts.

I think it's an easy affordance for those users who are just interested in the basic functionality of the product.

However like you I cottoned on early that one could put a "Podcast Production Notes.txt" in each one of my Notebooks that allowed me to really put some horsepower behind the generated audio :D

It's a shame that folks look at this and think it's awesome but then have the dawning question of "When will Google kill it?"

People building on top of this will likely want to know what the Open Source / non doomed version will be!

I hope Google never kills it. It is a useful tool. But then whatever Google killed was useful too.

Really wish there was an API so I can upload my content and connect it to my website to make it interactive for my potential clients.

In a sea of similar tools, Google seems to have struck on something semi-viral with NotebookLM. Output can be mediocre, but with the bar for many podcasts being set at "read pages from Wikipedia", that's not bad at all for zero effort.

https://trends.google.com/trends/explore?geo=US&q=NotebookLM...

The 100 baseline on that graph is the highest attention the term has received, and it correlated with a launch and has since decreased.

Google never has problems with first the first few millions for consumer-launched tools. They have problems with the first few millions of net profit almost 100% of the time and shut it down a few years later.

But I do agree this is a good play for Google, it plays to their strengths.

> The 100 baseline on that graph is the highest attention the term has receive

Good point. I couldn't come up with a well-enough known competing tool to compare against.

I really wish it had more voices, notebooklm-guy and notebooklm-girl get tiring

Definitely hear what you are saying but I personally think it is for the best that they are instantly recognizable as NotebookLM podcasters. Especially as this makes the rounds on the internet. If you could manipulate the voices it would just make it more challenging to detect if a "Podcast" is using this tool.

Now that NotebookLM has gone from "small experiment" to "moderate viral success", I expect all kinds of roadmaps are being drawn up to use it to hook users into the broader Google AI ecosystem (e.g. automatically add images and illustrations by Imagen 3, etc.

> With over 80,000 organizations already using NotebookLM

Really. "Using"? (as in an email from an org owned domain logged in to notebooklm page?..)

I want the HN comment section as a podcast

Made one for these comments https://notebooklm.google.com/notebook/e3b9d8c5-6243-4ae6-ab... (although I haven't checked the quality since I'm in a public space as of writing)

That could actually be a top quality podcast - well moderated content from thoughtful people, many subject matter experts, with mostly thoughtful discussion.. (I read hn for the comments..) sounds good to me.

One day too late. ^-^

I need different voices, people think the guy is me.

I’ve recently started using NotebookLM and I wish either it was from any other company besides Google or that Google would charge for it.

Google has the attention span and product focus of a crack addled flea. I’m afraid the entire project will be killed.

NotebookLM is a great product. I just started using it this week to ingest artifacts for a new project and get an overview.

I realize now that this is actually a clever way to collect training data. If it were any company other than Google, I'd be like, Awesome toy. With them, I am uneasy.

The terms explicitly say that uploaded content is not used for training the model.

Is there an open source tool that copies NotebookLM yet, or did anyone dig a bit into how the prompting is done to generate output in this dialogue format?

Check out this prompt: https://github.com/souzatharsis/podcastfy/blob/6ad5734c3ffb5...

Huh, that gives the LLM an example

    <Person1> "Welcome to Podcastfy! <break time="0.5s" /> Today, we're summarizing an interesting text about [topic from input text]. Let's dive in!"</Person1> 

And then tells the LLM not to do anything at all like that example

    Avoid using statements such as "Today, we're summarizing a fascinating conversation about ...". ]
    [PauseInsertion: Avoid using breaks (<break> tag) but if included they should not go over 0.1 seconds]

Is this an intentional technique in prompt construction to avoid the LLM over indexing on the example or something?

Looks good, I wonder if F5 could replace 11labs?

Ahh excellent! Podcast listings and Youtube weren't filling up with quite enough AI slop yet.