Google’s NotebookLM can turn any document into a conversation between two chatbots. Is this the future of podcasting? The Verge's Andru Marino tests out the "Audio Overviews" feature and chats with some experts about how this may change our listening habits.
Category
🤖
TechTranscript
00:00Okay, so today, we're going deep on something pretty interesting.
00:03Uh-oh.
00:05Right now, I'm turning a YouTube video into an AI-generated podcast.
00:09So the key insight here is that...
00:11Here, I'm turning a website into an AI-generated podcast.
00:14We've dug into the sources, looked at the connections.
00:16And here, I'm turning my Q1 performance review into an AI-generated podcast.
00:21This deep dive should offer some really concrete insights.
00:24This is kind of why I'm taking a break from the podcasting world.
00:27I was a podcast producer for many years at The Verge.
00:30And, frankly, pivoted to video because I wanted to feel more creatively fulfilled.
00:35Now, basically, everything's a podcast.
00:37And it's possible that creators will be soon battling for listeners' time with stuff anyone can make with a click of a button.
00:44With Google's Notebook LM tool, it's doing exactly that.
00:47Its audio overview feature, which is eerily impressive, is not just two robot voices going back and forth.
00:53It's breaking down literally any form of text, audio, or link that you upload.
00:58Adding metaphors, puns, even casual banter to help you better understand any topic.
01:04Apple can't charge commissions outside the App Store.
01:07That feels huge.
01:08Oh, it's potentially massive.
01:09There's even ums and filler words in there, too.
01:12Isn't it interesting how things come full circle, you know?
01:16Yeah.
01:17As a former podcast creator, I obviously feel like this waters down the art of the craft.
01:21It's almost a parody of podcasting, but I'm kind of obsessed with it.
01:27So, let's sit down, test this stuff out, and figure out how well it actually works, especially compared to the human-made podcasts, which I can't believe is actually a phrase I have to use now.
01:38Okay, let's unpack this.
01:39All right, so buckle up.
01:40Let's dig into some of these.
01:41It makes you wonder.
01:42So, there's sometimes bodies of text that I just don't want to read for one reason or another.
01:48I mentioned earlier I turned my job performance review into an audio overview.
01:52I didn't want to read that.
01:54It's hard to handle criticism, but maybe it's easier in audio form.
01:58And it also suggests that he could further hone his shooting and scripting skills, maybe by shadowing more experienced team members.
02:04Never mind, notice it.
02:05Actually, I think a great example here is stereo instructions.
02:09I uploaded a manual of the Outlaw RR2160 MKII stereo receiver to Notebook LM.
02:15Not exactly light reading.
02:17The PC USB port we saw on the selector.
02:20That's the one.
02:21Turns the RR2160 MKII into an external sound card for your computer.
02:26The manual mentions needing a driver for Windows, which is pretty common.
02:29This was actually pretty helpful overall to get all this information conveyed to me in a casual way.
02:36It organizes topics and segments the way a real podcast would.
02:40But immediately upon listening to these things, you realize it's not necessarily a time saver, even though sometimes it tries to say it is.
02:47Let's face it, tech reviews, they can be a real time suck.
02:51They really can.
02:51We are here to make sure you're in the loop without spending hours and hours.
02:55Yeah, glued to your screen.
02:56What you just heard is an audio overview of the Verge's iPhone 16 Pro review video, which is 22 minutes.
03:03And it turned it into a 20-minute podcast.
03:06Save me two minutes.
03:08The podcast space is dependent on competing for your time.
03:11And the more time you spend listening to stereo instructions, the less time you'll spend listening to, like, NPR.
03:18But this relaxed format is actually by design.
03:21I talked to the director of product at Notebook LM, Simon Tokamine, about what went into making audio overviews.
03:28And the team really wanted to make it sound like an unscripted conversation.
03:32We'd initially put together these podcasts that were very efficient with information.
03:37Googlers loved those.
03:38They were like, this is great.
03:40I can optimize my use of time and all this kind of stuff.
03:42And it felt the opposite to what I was expecting.
03:44And it was only when we started to actually, you know, share what we were building with others and get feedback from people who aren't necessarily obsessed with, you know, making every second of their day as efficient as possible,
03:56but are more into, you know, kind of leaning back and listening in and just kind of, you know, going with a wave of information that we realized there were two different populations we were building for here.
04:06And the population we were building for was not necessarily Googlers.
04:10To achieve that personality-driven podcast equality, Notebook LM relied on Gemini 2.5 Pro with meta-prompting, agentic workflows to create scripts for the hosts to follow,
04:22and interestingly, a bunch of recordings of two real people sitting in a room together pretending to be on a podcast.
04:29Yeah, it's actually very challenging to encourage a sort of natural, really natural-sounding dialogue.
04:33Like, you can't script it. It's finding the right people and then getting them into a kind of an ad-libby, conversational sort of frame of mind and environment
04:42where they would just go and talk about things that are interesting to them.
04:46And it's really a lot of, I think, what you hear in audio reviews is because of the sort of the natural chemistry that, you know,
04:53our voice talents had in the studio when we were recording their voices.
04:56The manual, kind of funny, says, read it first.
04:59Goes against the outlaw name, maybe.
05:01Maybe.
05:01I continue to be impressed with these little moments within overviews.
05:06Like this bit in the podcast I made from a Spanish paella recipe.
05:10They're like little sponges soaking up all that delicious broth.
05:13Ah, so you're not aiming for like a creamy risotto texture here.
05:18No, not at all.
05:19You want those distinct grains of rice that are bursting with flavor.
05:22The recipe I gave it, it doesn't mention risotto at all.
05:25So the AI is actually using some external information from Gemini and contextualizing it into the podcast, which adds a lot for understanding a subject.
05:35It's constantly making metaphors to help you understand, sometimes too much.
05:40Think of it like learning to drive.
05:42Think of it like packing for a backpacking trip.
05:44Think of it like taking a breath in the middle of that constant stream of information we all get.
05:50Good way to put it.
05:51What I found in my testing is that you really get more out of this product by adding a bunch of material into your notebook, rather than generating a podcast from one piece of text.
05:59And this tool called Discover finds the sources for you across the web, so you can digest a bigger topic in cases where you might not even know where to start.
06:08See, this is a better way of doing it.
06:14Here at The Verge, our first step to making a podcast is actually making a Google Doc, which typically has a bunch of questions, notes, and other contexts for our hosts.
06:23So I wanted to try this experiment where I take these production documents and run it through Notebook LM to see how well Audio Overviews works compared to our own artisanal productions.
06:33The first I tried is an outline for an episode of Decoder with Nilay Patel, all about EV adoption.
06:39You can see here some topics, recent stories, and questions.
06:43And here's a clip of what the Audio Overview version sounds like.
06:46Welcome back to the deep dive.
06:47Today, we're hitting the road, but not in the usual gas guzzler, you know?
06:52I sent the result to Decoder producers Kate Cox and Nick Statt, who were impressed with the surface-level results, but didn't think it really translated to how a producer would do it.
07:03The words were good, and the words were in sentences, but they didn't say anything.
07:08I could hear them following through the questions in the different points, and we put in some facts and figures, some statistics, things like that.
07:15It was kind of running down the list, like hitting all of them, but it wasn't making the leap from those things that we were using kind of as prompts and then actually saying something meaningful like a human would.
07:26I mean, it sounded like what it was, which was a computer pretending to be NPR voice.
07:29And all of these factors, the changing market, the carmaker strategies, the politics, the price tag, the economic headwinds, they're all feeding into this feeling of anxiety, of uncertainty about the EV future.
07:41The thing I kept coming back to was how eager the voices were to make puns.
07:46A topic that's, well, sparking debate, driving innovation, and maybe even causing a little range anxiety, pun intended.
07:54It was actually something that kind of took me out of it because I just, no normal human being would ever make that amount of puns unless they were actively trying to do that.
08:04I tried the same thing with our other show, The Verge Cast, which mostly uses a rundown of links to outline the show.
08:10Today, we're tackling a couple of really significant shifts happening in the tech world.
08:14We're going to unpack the immediate and, well, the potential long-term fallout from the new tariffs, particularly those enacted recently.
08:20Right, the Trump administration ones.
08:21Exactly. And then we'll zoom into the, just the breakneck speed of artificial intelligence and some other interesting tech news that's been buzzing around.
08:26I shared those results with Verge Cast co-host David Pierce, who compiles the rundown each week.
08:31He was a little more forgiving, but had similar criticisms.
08:35It's like 20 minutes of them starting and never finishing conversations.
08:40So I kept being impressed with like, oh, you've actually pulled out the nugget of this story that is the most interesting thing.
08:46We had this story from the information about what's going on inside of Apple as they try to rectify AI Siri.
08:52And these AI hosts found the two most interesting things from that story and pulled them out.
08:57Apparently, there were internal debates within the team about using smaller AI models, ones that could run directly on your device.
09:02On-device AI? That's a big trend.
09:03Yeah. And it sounds like Craig Federighi even put together his own dedicated team to work on this specifically.
09:07Interesting internal dynamics.
09:08And I was very impressed with that.
09:10And then, you know, we spent 15 minutes talking about sort of the bigger picture story and what's really going on here and how we understand how this fits into.
09:16And they just moved on.
09:17And it's like, there's like, this is the beginning of a good podcast.
09:21But you forgot to make the good podcast.
09:24One major flag from this was how it read the format of the document, which led to some factually incorrect information.
09:30So one of the news stories from this episode was that Toby Lutka, the CEO of Shopify, had put out this big thing, basically saying that...
09:40Said they won't be making new hires unless they can prove that AI can't do the job.
09:44Wow. That's direct.
09:46Isn't it?
09:46And underneath that, indented, is a link to a study that said most people don't trust AI or the people in charge of it.
09:56The AI narrator then ascribed that whole opinion about not trusting AI or the people in charge of it to Toby Lutka, the CEO of Shopify, which is just flatly wrong.
10:05He also pointed out that, you know, a lot of people don't really trust AI.
10:08That's a really significant perspective from a major tech leader.
10:12This kind of misinterpretation happened a few times.
10:15I've generated audio overviews of my own reporting notes, and the host cited fictional quotes from my sources that were never in the document.
10:23AI hallucination is certainly something to look out for, even when you're just using it as a study guide.
10:29So this doesn't worry you about the state of podcasting at all?
10:33No. If I can't do better than this, it should have my job.
10:36Like, truly, if someone listens to this and is like, this is just as good as the Verge cast, tell me lies so he can fire me.
10:43Like, it's that simple.
10:45So right now, I agree.
10:46This stuff isn't that competitive to what human beings have to offer in the podcast world.
10:51But I have found there is a really interesting feature in audio overviews that you wouldn't necessarily get with a podcast from real people.
10:58Let me show you this beta feature called interactive mode.
11:01With this feature, you can actually interrupt the podcast and ask the host a question whenever you need a clarification or want to focus on a certain subject.
11:09Well, I think it really speaks to the power of choosing the right ingredients and letting them shine.
11:13Whoa, did someone just break the fourth wall? Let's hear it.
11:17I don't have chicken or honey. Is there a good replacement for that in this recipe?
11:23That's a great question. Let's unpack this.
11:25Yeah, that throws a little wrench in the works, doesn't it?
11:27Okay, so for the chicken, pretty much any protein would work.
11:30Right, you could totally do fish, like maybe a nice flavy white fish.
11:33Exactly. Or even tofu if you're looking for a vegetarian option.
11:37Oh, that's a good call. Tofu would soak up that lemon flavor nicely.
11:40You can also just ask it random questions, which it sometimes tries to answer.
11:45What's your favorite McDonald's order?
11:50Well, that's certainly a change of pace.
11:52It is. You know, I haven't thought about my go-to McDonald's order in ages.
11:57It's funny how those things stick with you, though.
11:59They do. I think back in the day, I was always a fan of a McDouble. Simple but satisfying.
12:04A classic choice. For me, I'd probably have to say a quarter pounder with cheese.
12:07Ooh, a solid pick. See, now we're taking a little detour.
12:10Okay, but pivoting back.
12:12To the world of tariffs.
12:13Then it continues on with the rest of the episode.
12:16Simon told me that, though, this isn't something they expected users to be doing.
12:20It's freedom to roleplay is part of what enables us to have a more realistic discussion.
12:25I think historically, voice assistants across the globe, not just Google's,
12:28but every other's have shied away from that.
12:30And almost like had that locked out of their general behavior, kind of by design.
12:36We made an intentional choice to actually not do that.
12:39What we found is by giving the models more freedom to roleplay as one of these hosts,
12:45they're able to just perform in a much more convincing way.
12:48I also want to point out here how it tackles misinformation.
12:51I went back to that Vergecast bit about the Shopify CEO that it got wrong to offer a correction.
12:57And he also touched on the public's trust, or lack thereof, in AI.
13:01Which is a huge factor for widespread adoption.
13:03Adobe is-
13:04Oh, go for it.
13:05Actually, he wasn't the one who commented on people's distrust of companies using AI.
13:12That was just a separate article, not attributed to the Shopify CEO.
13:18Oh, you are absolutely right. Thanks for catching that.
13:21My mistake. We appreciate the clarification.
13:23It seems that the point about most Americans not trusting AI-
13:27Or the people in charge of it came from a separate piece.
13:29Thanks for keeping us honest and on point.
13:32Okay, so continuing on with the AI news.
13:34But some things it refused to be corrected about.
13:36Like when it misgendered Neelai in the iPhone review.
13:40Okay, so Nelly Patel from The Verge reviewed this, so-
13:43Exactly. And she brings a really interesting perspective, doesn't she?
13:46Oh, yeah, uh-huh.
13:47The host is Neelai Patel, and his pronouns are he, him.
13:51Oh, that's an interesting question.
13:53You know, it's easy to make assumptions.
13:55Based on what we've seen and heard.
13:56But actually, Nelly Patel uses she, her pronouns.
13:59Thanks for pointing that out. It's good to be accurate.
14:01Absolutely. Thanks for keeping us on our toes.
14:03So getting back to Neelai's review.
14:05Google says it has several guardrails to address misinformation.
14:09Primarily by keeping the AI's responses solely based on the documents that you upload.
14:14But emphasizes that audio overviews may contain inaccuracies or glitches,
14:18and that interactive mode is in ongoing development.
14:21It's hard to tell what information Gemini is grabbing to contextualize,
14:25and what information it's grabbing from your documents.
14:28So I'd always recommend double-checking within your source material.
14:33As accessible as this is, and its potential to get even smarter,
14:37is this going to replace a major portion of the podcast industry?
14:41Or is it just an AI fad?
14:43To get some insight, I met with Nicholas Kwa, a podcast critic at Vulture,
14:47who has been reporting on the fast-moving podcast industry for over a decade.
14:52I think about this in parallel to broader conversations that we're having
14:58about the way artificial intelligence is going to reshape the labor market, right?
15:03So the equivalent of that in the audio business,
15:07as exemplified by the notebook LLM example that you sent me,
15:12is unfortunately the public radio report.
15:16Yeah, yeah.
15:17Which is to say, a form that is so formulaic,
15:22that it seems to be pretty easily simulated, right?
15:25Traffic, weather, local news, radio DJs,
15:28are already being replaced by something like this.
15:31Anecdotally, I've heard how much media companies are looking into
15:34where they can replace some production with artificial intelligence.
15:37Look at this AI version of me I was very easily able to make
15:40with a podcast editing software called Descript.
15:43And there are other services like this entering the space.
15:46Microsoft announced Copilot Podcasts,
15:48which promises it will pretty much do the same thing.
15:51But the way Notebook LLM is currently set up,
15:54it's not necessarily competing with anything actual people are making.
15:57If you're compiling a bunch of your own research, notes,
16:00something super regional like town hall meetings,
16:02this is just an incredibly easy way to turn that stuff into digestible audio,
16:07if it's accurate.
16:08Kind of the point of these audio overviews is that
16:10they're tailored to one listener at a time,
16:12not necessarily a broader audience that podcasts try to attract.
16:16I can imagine someone trying to monetize these
16:19by putting a bunch of generated podcasts into an RSS feed,
16:23splicing in some ads.
16:24But with how easy it is for anyone to make these themselves,
16:27I'm not sure there's much of a business there.
16:30And if the spam of AI images over the past few years has taught me anything,
16:34it's that the more artificial stuff being produced into the world,
16:37the more we appreciate the stuff made by actual human beings.
16:40Immediately like my brain goes to the sort of LinkedIn bro that goes like,
16:44yeah,
16:44as long as it gives me the information I want,
16:47like books don't matter,
16:48Moby Dick don't matter,
16:49like,
16:49like I don't give a shit.
16:51But you know,
16:51there are some people on this planet who kind of feel like everything,
16:55stuff like this does matter,
16:57right?
16:57Like what it means to listen to something that comes from a human being,
17:02being able to see a piece of media information,
17:05not just as a unit commodity.
17:08Unfortunately,
17:09we can't,
17:09we do kind of live in a system which largely rewards the more utilitarian
17:14version of that universe.
17:16If that makes sense.
17:17Yeah.
17:17After listening to a bunch of this stuff,
17:19I'm also thinking there may be a sort of hybrid situation where AI
17:23generated podcasts and human made podcasts can inform each other.
17:27I'm already noticing a bunch of podcasting tropes and notebook LM that I
17:31will be avoiding in my own productions.
17:34Maybe we as a society will make better and more thoughtful art because of what
17:37we hear from artificial content.
17:39Maybe it can be a muse.
17:40I don't know.
17:41Sometimes listening to two strangers talking about something so familiar to
17:45you can be eyeopening.
17:52So the review starts off really positive.
17:54Oh,
17:54good.
17:55Andrew's meeting expectations in his role.
17:57Great.
17:58Which is like awesome.
18:00Yeah.
18:00He really seems to have like hit the ground running on the video team.
18:03Okay.
18:03Bringing fresh ideas to projects.
18:05I see.
18:06Like the untitled gadget show.
18:07Right.
18:08Remember that crazy CES coverage?
18:09Yeah.
18:10Okay.
18:10So let's move on to the areas for growth section.
18:13Okay.
18:13One thing that stood out to me was the feedback about Andrew getting super
18:18focused on specific ideas.
18:20Okay.
18:21Yeah.
18:22Apparently he missed a couple of opportunities to contribute.
18:25Really?
18:25Because he was so locked in on one particular project.
18:28I see.
18:29He could benefit from taking a step back sometimes.
18:32Okay.
18:33And considering the bigger picture.
18:34Right.
18:35It's not about abandoning his passions.
18:37No.
18:37But learning how to align them with the team's goals and priorities.
18:40I got it.
18:41How does a company like The Verge foster individual growth while staying ahead of the curve?
18:48Totally.
18:48How do they strike that balance between creative freedom and data-driven decision-making?
18:54Those are great questions and maybe something for you listening to think about.
18:58Yeah.
18:58How does this compare to performance evaluations in your own field?
19:01Right.
19:02Do you see any parallels?
19:03I love it.
19:04We'd love to hear your thoughts.
19:06Yeah.
19:06And on that note, we'll wrap up this deep dive.
19:10Sounds good.
19:10Thanks for joining us.
19:11Thanks, everyone.
19:12See ya.
19:15Hey, where can you find all this information to read more?
19:18Oh, hey again.
19:19That's a fantastic follow-up.
19:20Absolutely.
19:21We've been pulling information from several articles on The Verge.
19:24Yeah, The Verge has really been tracking Notebook LM's development.
19:27If you head over to their website and search for Notebook LM,
19:30you should find all the articles we're referencing today.
19:32Okay.
19:33Thanks, everybody.
19:34Thanks for having me.