Steven Johnson: Revolutionizing Writing and Research Steven Johnson, author of 14 books on the history of science and technology, discusses Google’s new writing and research tool, NotebookLM. He elaborates on how walking influences creativity, the importance of play in innovation, and shares insights into using AI for augmenting thought and memory. The conversation covers his role in co-creating NotebookLM with Google Labs, the importance of uploading sources for effective use, and the potential of AI to enhance both authorship and readership. Johnson also touches on potential future developments, such as recommending additional sources and addressing privacy concerns. The dialogue highlights the revolutionary potential of AI-driven tools in research and content creation.
00:00 Introduction to Steven Johnson
00:20 The Role of Walking in Creativity
02:27 The Importance of Play and Uselessness
04:35 Notebook LM: A New Research Tool
07:02 The Genesis of Notebook LM
09:29 Building and Innovating at Google Labs
12:48 Using Notebook LM for Writing and Research
22:24 Exploring New Book Ideas with Notebook LM
25:27 Exploring AI's Role in Research
25:56 Future of AI in Recommending Sources
26:44 Tone and Style in AI Writing
29:20 Privacy and Copyright Concerns
30:26 AI's Impact on Writing and Creativity
36:05 New Forms of Reading and Books
40:29 Practical Tips for Using NotebookLM
44:34 Reflections on the Future of Authorship
📜 Read the transcript for this episode: Transcript of The Future of Books and Improving Your memory with AI: Steven Johnson (Author and Google Labs) |
[00:00:00] Steven Johnson: Hi, I'm Steven Johnson. I'm the author 14 books about the history of science and technology and innovation. And I'm also the co creator and editorial director of a new writing and research tool from Google called notebook LM.
[00:00:13] Jeremy Utley: I just happen to be leafing through where good ideas come from. Which is one of my favorite books for the record. Oldie but a goodie. And in that book, you actually advocate walking as a means of, of shall we say, courting serendipity.
And it just made me think when all of one's resources are now summoned, instantly summonable via notebook, LM, et cetera. What role does walking play? , how do you think about walking in a world with notebook LM?
[00:00:40] Steven Johnson: You know, when you're walking by yourself is that it gives you time, with your own thoughts and in a sense uninterrupted by your reading history or the, you know, the books, your, your, the notes or what you've just written.
When you have long stretches of uninterrupted time without any media around you, without any distractions, without listening to a podcast, , it gives like each idea time to kind of evolve into something else, right?
You're like, you start to think about one thing and that brings up another kind of association in your mind, and then that leads to another association and, and like often getting to the new insight. You have to have like six or seven jumps of association, you know, to be able to get to something new in your own head.
And if you're constantly, if you're only thinking in like, you know, 30 second or minute long chunks, because you're constantly distracted by something else, you just don't get far enough in those associative trails. Right. , and so walking is one of the places it also happens to be like, apparently there is some like, physiological thing that happens when people walk at their natural gait.
Like the second you walk a little too fast or a little too slow, it gets harder to think. But if you walk at your natural gait, there's, there's something cognitively like rich about that. , So that's another element to, I suppose, walking at your natural gate while holding notebook LM in your hands,
[00:02:06] Jeremy Utley: the ultimate, the ultimate, no, but I mean, I think in this, in, in, I think one of the things that we all wrestle with, it could be like an existential challenges in a world where we can increasingly be productive all the time and efficient all the time.
How do we, it could even be rationalized, if you will. The. What kevin kelly the other day called, uh uselessness He said I try to do something useless every day and it can feel like walking is useless and we know it's not but To me unless it's actually an explicit goal. It's something I can neglect in the name of productivity I don't know if you if you've wrestled with that tension yourself, especially as an expert in innovation, but i'd love to get your thoughts
[00:02:48] Steven Johnson: Yeah, there are two things that that brings up to me.
The first is um The book I wrote, Wonderland, about the history of play, um, and that was basically another
[00:02:57] Jeremy Utley: favorite, by the way. And
[00:02:58] Steven Johnson: yeah, that's that of all the kind of innovation books that I've written that that may be my favorite actually. I love that book. , and, and part of the argument in Wonderland was like that there was a long history of people pursuing, , hobbies and interests purely for the kind of playful delight that they got out of them without an objectively useful purpose in the world.
Um, without trying to solve a big problem or a small problem, they're just doing something for the fun of it, for the love of it, you know, games and music and, , you know, and, and wandering in a park and all those different things. Um, and, and yet, despite that seeming kind of. frivolous nature of the pursuit, those kinds of explorations end up generating like, you know, world changing ideas that are useful, , because Are you familiar
[00:03:48] Henrik Werdelin: with, uh, are you familiar with, uh, why greatness can't be planned?
It's a, it's a professor in AI, , called, , Kenneth Stanley, who wrote this book, which I slightly obsess about and basically makes the same point that we, we had this fallacy of the objective and in his view, we should pursue with great intensity, interestingness. And by doing that, you create like basically the stepping stone that creates to greatness.
And so. And his book is about how we get to artificial general intelligence. So it's actually not like a philosophical book at all, but I think both from entrepreneurship and for ideation, it probably is very, uh, um, topical.
[00:04:26] Steven Johnson: Well, you know, it's, , it's interesting that you say that because like one of the things that I have loved, um, about.
notebook. Lm, which will, we can get into in more detail. But, um, you know, the kind of core idea is that you are able to upload sources that can be your own notes or books that you've read or, , other documents, pdfs, whatever it is, and, and then have a conversation with the, With the model about that information.
And so it's a kind of a research tool. , and a kind of thought partner in a sense. Um, and so one of the things that is now possible with these advanced language models like Gemini, which is what underlies notebook LM is not just straight kind of search queries like that was the way I was originally imagining it when we started to build it.
Like, Oh, I'll be able to upload all my reading notes. From all the books that I've read and I'll be able to say, Hey, you know, what was that passage about ant colonies and the way that they, you know, , evolve over time, , and it would find it for me in a, in a more efficient way. What it turns out is that these models actually have a concept.
They seem to have a understanding in quotes of. Interesting this and of surprise and so one of the things that I do all the time is like I'll upload a bunch of new information and I'll be like, okay, what are the most surprising ideas in this about X now?
[00:05:49] Jeremy Utley: Now, is it surprising to you? Surprising to know, but like relative to what is it?
Yeah, I don't know what it knows, you know.
[00:05:56] Steven Johnson: I think what it, I think when it answers, so all I can tell you is that it generally does seem to surface, you know, the, the, the fact from this source material that if you were then dropped down into a cocktail party and had to like deliver like one fact to somebody like, Hey, this is pretty interesting. I think what it is, is delivering is, , let's say you asked for the most surprising fact about ant colonies.
It knows that it's generally well known that ant colonies have a queen and that ant colonies use pheromones to communicate. And so it's downplay, even though there's a lot of material in the, in the quotes about that, you know, those facts, it chooses not to share that because that's not surprising or interesting.
What is interesting is this little known fact that actually is, you know, , And so it's, it's just an, every time I do that, I'm just kind of astonished that like you can interact with a machine with that kind of query. It's pretty, it's pretty cool.
[00:06:53] Jeremy Utley: I know for you personally, this is almost like a fan boy moment. It's like this meta moment in a lot of ways. Could we maybe just take a step back? We don't have to go all the way back to kind of hypercard days, but even, but just take a step back, like even to call it may 22, Google team emails you out of the blue and says, we found a way to build this thing for those who may not be as familiar with your kind of obsession with tools for thought and things like that.
What are you thinking in that moment? And why is it, why was that cold email so momentous to you as an author?
[00:07:27] Steven Johnson: Yeah. As you said, I I've been obsessed with. using software to help me augment my thinking and my memory, um, and my writing process for, since I was in college, since the, you know, late eighties. Um, and I had, uh, I had written a piece for the times magazine.
It's like a 10, 000 word piece, basically about language models, mostly focused on GPT 3, actually, um, and, uh, because Google's models basically weren't public at that point. Um, but I had gotten access to GPT 3 in the fall of 21. And actually, it's a funny story. Originally, I had pitched to the Times Magazine on a piece about, uh, open AIs.
Uh, governance structure. I was like, Hey, there's this thing where like, it's kind of a for profit, but kind of a nonprofit and they have that won't have any relevance to
[00:08:18] Jeremy Utley: like to current events at any point.
[00:08:21] Steven Johnson: Exactly. And my, my, uh, my idea was like, ah, everybody already knows that GPT three is really interesting itself, but nobody knows about this governance structure thing, and I should write a piece about that.
And then I got access to GPT three and I was like, Oh, Oh, man, it's so much more interesting than I realized. Like there's so much to say just about that. And so the piece ended up kind of focusing mostly on what it means when computers master language and it alluded to the governance stuff in the end of that ended up being a whole other story.
But so. And it had occurred to me when I was writing that piece that like, oh, you know, maybe this dream I've always had of the ideal research assistant is now within reach because of these tools. And so, um, two folks at Google, Clay Bevore and Josh Woodward, uh, Had been long time fans of where good ideas come from.
And, and we're readers of my sub stack. We were, I was writing about, it's so cool,
[00:09:14] Jeremy Utley: right? How these things feed each other, your ideas about ideas then inspire the builders.
[00:09:20] Steven Johnson: Yeah. And then, and part of, so they had just spun up this new Google labs division, which there wasn't older Google labs, but this is the kind of second coming of it.
And they. A big part of their idea was that like, maybe we could do these co creations with people who are not necessarily technologists, but are going to be the kinds of people who will use the tools we build, and maybe we could kind of bring them in and have them help us build these products from, from day one.
And so I was, I was kind of the first guinea pig for that approach. And, uh, Clay reached out to me and, and we had a meeting, actually, it's very funny. We had a meeting in this. New kind of, um, 3d zoom product they have called project star line, which is like basically like a hologram of,
[00:10:05] Jeremy Utley: is that a good first impression or bad for, it was an
[00:10:07] Steven Johnson: unbelief.
It was actually one of the top five product demos I've ever seen. I mean, I'm an insider, so they wanted to knock
[00:10:13] Jeremy Utley: your socks off. They're like, we got it. We got to get him way on board. Well, the hilarious thing was
[00:10:17] Steven Johnson: I, so I do this virtual meeting with this hologram of clay before who subsequently left, but.
Um, was running labs at that point. And we have a great meeting. He's like, I think you should just come, like spend some of your time here. And we're like, let's build something together. We've got a little team, we can do it. And I was like, that sounds fantastic. Like what a, what a cool opportunity. And then I got out of the meeting and I was like, he literally put me in a reality distortion field.
Like it was, it's not a metaphor. He just, it's an actual, an actual physical reality. But, um, But it was great. And it was a very, I mean, it was a kind of a bold thing. And I, honestly, the way I thought about it was this will be really fun. I doubt we'll actually make anything that sees the light of day,
[00:11:03] Jeremy Utley: but
[00:11:03] Steven Johnson: I, you know, but we'll, we'll, I'll meet some interesting people and I'll be right at the center that this, you know, the most important technology of my lifetime, probably.
And, you know, I'll get to, you know, explore for a year or something like that. And then we built a little prototype. Early on, um, there was basically it had the core idea. It was from the, from the very beginning, it was like you upload documents, you're not just having an open ended chat with the model, you, you upload documents and you, you know, it's what we would now call rag, but we called it source grounding, which I think is a much better name for it.
And, uh, everything, everything that you talk about is grounded in the documents you've shared. So. So in the early days, it was pretty simple, but we always had that architecture and it just was one of these things. We built a very quick prototype and it got a lot of internal attention, um, you know, all the way up to the very top of the company.
And so we, we kind of got green lit to, to do it a little bit more. And then we got an IO last year, which is kind of a big deal. And it kind of built up some momentum and, but we're still, what's great about it is we're, we're still, um, even though we just launched in 200 countries and territories around the world, just a few weeks ago, um, uh, it's supporting a huge number of languages.
Um, the team is like, we have like maybe. Eight engineers now, like, you know, one designer, one product manager who works with me, Risa Martin, who's amazing. Like it's very, it does feel like a startup, um, still to this day. Um, and we've been able to build, you know, in part, because we're building on top of these models that are so capable, um, we've been able to build a lot of, a lot of cool things.
[00:12:48] Henrik Werdelin: In question that I've been looking forward to ask you is after kind of starting to do more and more work with and, you know, custom GPTs and stuff like that, I've started to think in repository of content that I need to dump into a notebook, for example.
And so like the obvious one is everything you've ever written, you know, like I noticed, , in one of the other, uh, the podcast you did with Dan Schipper, that You use, I think, Readwise as your way of bookmarking, and so another repository is kind of like anything you ever bookmark in your books. Have you created, like, a framework, or can you offer advice on, like, what are the kind of, like, the five first repository?
And I'm asking because I've now started to remind myself to collect stuff. Like, so, for example, after we've done this interview, one thing I need to do is to have a transcription of it for creating a repository of all the podcasts we've done.
[00:13:45] Steven Johnson: It's a great question. So, one way to think about it is, , you know, we've never really had a tool like this before.
Most of us, I think, have jobs or, or hobbies, , that involve synthesizing information across a large number of documents. Like, you know, if you're a researcher or a writer, of course you do. But a lot of people actually have this, like the, the, the work that they have to do is, um, is, involves information that's scattered across like 10 or 20 or 30 different documents that are floating around.
And. Most of us, I think, live in this kind of world where we're like, we have like 10 open tabs of all those documents. Only 10. I mean, yeah, whatever it is. Like, I mean, right now I have like 40 open. Right. And, but like 10 of those are relevant to the thing that I'm doing right now. And so your work is constantly being like, okay, which tab I need to find that one quote.
Okay. There it is. Okay. Now I switch over to that. Now I copy that and I paste it back in here. And so Part of like what we've tried to do with notebook is to, is to build a tool. Now, because of AI, we can build a tool that can just basically like. synthesize and help you explore all those documents in one place.
, so you're not, you're just kind of sitting there asking questions, thinking, jotting down ideas. And those documents are just like the, in the background being drawn upon whenever relevant by the AI to help you do your work. And so you're right. You know, this idea of like collecting these repositories is really important.
[00:15:10] Jeremy Utley: You've written about innovation kind of from the outside as a journalist, as a, as a technologist, as a student of the craft.
If, if you look back on how you've described innovation, is there anything that you've learned about from being on the inside that maybe you now see differently, or did it confirm understanding?
Did it disconfirm anything? I'd love to know how this experience building has influenced your, your understanding of innovation.
[00:15:37] Steven Johnson: Yeah, it, well, it's, I did have a little bit of insider experience when I was younger because I did a couple of actual startups, , feed, which was the first web zine. And then, , outside that in, which was a hyperlocal kind of news service in, , that I launched when, right after I'd written ghost map, actually.
So it was, it was about kind of mapping information, which was the theme of ghost map, as you know. Um, but both of those companies, like they got to, you know, you know, 15 or 20 people. And then, , we sold outside into AOL. , so I was always the, in those cases, I was always, , one, I was the CEO and two, it was an incredibly small team with very small budgets.
And I spent a lot of time raising money and all this kind of stuff. And so this is the, this is the first time that I've, You know, been inside a large organization. , As my father likes to say, it's the first time I've actually had a real job, took me to 55 until I, that's awesome. But, , finally gainful employment.
And I think it's been a little bit, I mean, I, , it, it's been, , We've had a very unusual situation inside this large company because labs has a very distinct culture that, , does align with a lot of the ideas that we're in, where good ideas come from, you know, that, that there's a lot of emphasis on multidisciplinary collaboration, um, you know, which was a big theme of that.
[00:17:00] Henrik Werdelin: So you might just tell me a little bit more about labs. There's like a lot of iteration of Google labs and it's always super fascinating kind of how you guys build on, on top of it.
, do you mind giving just like 30 seconds on, uh, on this latest version?
[00:17:14] Steven Johnson: Yeah, I think that there had a couple of different. Elements to it that, you know, Google has an amazing research division, , , that generates a lot of ideas that inform and enhance what the company ends up producing. , But it's also a little bit less kind of like actual product focused.
, research and, and then you have the existing mature products that are out there, the workspaces and the, you know, the search and, and Android and so on. And so part of the original idea was that kind of labs is filling a little bit of a hole here or would be a gap filler of like, what are the things that are possible that don't fit any of the existing product categories as they stand now that we could experiment with?
And. Also the opportunity to build things that are small and can stay small for a while. Like there's a problem when you have multiple product lines that all have more than like a billion users that when you build something that's like, Hey, this is pretty cool. And we got 30, 000 users. Like everybody's like, we don't even know what that number is.
What, what, like, what are you talking about? So it was kind of initially, and then it was this idea of co creation as well and bringing people from the outside. But it ended up very quickly just because of the timing of it. It ended up. Really turning into something that's very focused on AI products and figuring out how to innovate on these new AI platforms.
The original vision, as it was sketched out in 2021, I think, was a little bit less. It had an AI lane, but it had other lanes too. But, , we ended up, almost all the things we're doing, , have AI and some kind of component to them now. , so it's, it's, it's really cool. Notebook is kind of the, of the, you know, Of the consumer facing, um, prototypes and experiments we've come up with notebook is the furthest along.
, we were one of the first. , and so we're one of the few ones that have really been released, certainly to an international audience. And that kind of looks like an actual product at this point, rather than just an experiment. So we're It's an interesting question now, like where, what's the next phase for NotebookLM, , and how do we play with the rest of the Google ecosystem, which I don't know the answer to yet, but, , because we're the first one kind of along the ladder in that direction.
[00:19:25] Jeremy Utley: Is there a portfolio of other kind of experiments that are in the hopper, so to speak? And do you, do you have involvement with other experiments or are you squarely focused on Notebook?
[00:19:34] Steven Johnson: Yeah, I, I, I do. I haven't had an involvement with some of them, , which I love and I wanted to do more, uh, but, but notebook is doing it is so Lately in my wheelhouse that I, it's, I ended up thinking about it more, but if people go to labs.
google. com, the portfolio of, of things that people can try are there. That's some really cool, uh, really cool music tools that I'm, I'm kind of an amateur musician. And so I love some of the music things that they've been doing.
[00:20:04] Jeremy Utley: So maybe let's dive diving into the tool itself a little bit. We'd love to know, , you know, Henrik mentioned, what are the kind of repositories of information, so to speak? When do you know to call upon? No book in your own workflow. Like, how do you remember to use it? , I've got a post it that I keep here on my desk, which is, have you tried chat GPT and you have to almost like enter, like all of a sudden we had these alien intelligences available to us.
I'd love to know, like, when does it hit your workflow? How do you remind yourself? How do you, as kind of call it the leading bleeding edge user, what are your primary use cases and how do you keep it integrated? Yeah.
[00:20:42] Steven Johnson: So let me give you two examples of how I use it. One is I have, this is very meta, but I think you'll understand why it makes sense.
I have an, , a notebook LM notebook, and that's where I do my thinking, For the product itself.
So over the life of notebook. You know, there are a lot of documents that have been created, whether some of them are press releases and some of them are, you know, blog posts and some of them are, , how to documents and some of them are product feature overviews and things like that. But anyway, in that notebook, , the sources that I have added to that notebook are basically all the big documents that have been part of notebooks.
History. And then I use that to. , jot down ideas, , you know, so you can, you can just create notes and write notes, just old fashioned note style, um, and so I can just say in notebook LM, I can just type in like create a taxonomy of all the key, , terms, , in the notebook LM product, and it'll give me an amazing first draft of that.
, it's a classic kind of work where you're like, I could do it myself. It's not particularly high level writing, but sitting down and writing a list of all the like key terms and definitions for them is going to take, it'll take me like 45 minutes and if notebook can give me the, uh, you know, 95 percent of it or 90 percent of it in 35 seconds, like.
That's golden. , and so that's, like, that's an example of, like, you have a specific project, that project spins out a bunch of documents, you're constantly referring back to information in those documents, and you're also, like, thinking and jotting down ideas on your own, and you may be creating new documents out of that document.
The combination of your notes and your, those sources, that's a, that's a great use case for it. The other one that I would mention is, kind of early stages of casting about for ideas for my next book and for the next kind of history book in the ghost map motor, or like the infernal machine might.
My latest book that just came out and, and when you're in that kind of phase, like, you know, the field is really wide open. Like, you're like, should I write a book about the gold rush? Should I write a book about NASA? Should I write a book about, you know, like puppets, like whatever it is, like, you know, you don't like, I'm constantly like.
Following different trails, whatever. There are, there are
[00:22:56] Jeremy Utley: many authors who can really, you know, select from such a wide, uh, array of possibilities. Just some people's
[00:23:02] Steven Johnson: constraints are a little tighter. I'm not actually writing a book about puppets, but, but I do think that there's, so yes, that's true. But like there, but there are, I think all of us, not all of us, but a lot of people do have things where they're like, they're kind of have ruminations.
Like I'm thinking about startup ideas or I'm thinking about, you know, this. And so I created a new notebook called the next book. And that was a place where I would just go and whenever I, you know, had an idea, I would jot down a note in the notebook. But more than that, I would be like, okay, I'm thinking about the gold rush.
So, uh, I'm going to go read this great book from like 30 years ago about the gold rush and get some quotes from that. And then I'm using read wise. which gathers quotes from your Kindle. I can bring those quotes into notebook. And then that's there. And then I was thinking about NASA. And so I have these oral histories of, uh, the, , Apollo program.
And so I brought those into that notebook. And so I'm slowly accumulating a bunch of different sources related to all the various different things. It's very disorganized notebook because it's like NASA and gold rush and various other things in there. , but that's a place where I could then at any moment, sit down and be like, Okay.
You know, what was I thinking about a couple weeks ago? What was that idea? And the, the last thing I'll say about it is that I've actually started now, and this has really only become possible with the latest version of Gemini. , I've actually started doing things like, Hey, um, here's a bunch of readings that I just have collected on this particular topic.
, And I'll write a effectively a query or a prompt that's like, I'm Steven Johnson, author of books like The Ghost Map. , that are multi threaded, multidisciplinary, like, narrative histories. , Uh, I'm thinking about writing a book about the gold rush. Here's a bunch of quotes from a couple of books that I've just read.
What do you see that's interesting in here, given who I am and given what I'm trying to write? And it's actually quite good. At picking out bits that might be useful or, or suggesting framings that might be useful. I, you know, that
[00:25:10] Jeremy Utley: conversation, because I know it's not just like a one, it's not a one shot, zero shot.
Hey, what's interesting. Great. Thanks. Notebook LM for another book idea. Right. Yeah. Talk about the dialogue. How are you pushing back? How are you, how are you digging deeper, et cetera, et cetera.
[00:25:26] Steven Johnson: Yeah. So you'll get, um, it's, it's so, It's so amazing, actually, because you'll get, um, it'll say something, so in the Gold Rush example, it'll come back and say, well, there's a really, there seems to be a lot of material here about the clash between the, , the influx of new settlers chasing gold and the existing Native American populations in this region, , and, you know, kind of summarize that, and along with a couple of other themes that you could explore.
[00:25:54] Jeremy Utley: Already, you've already uploaded your sources.
Is there a world in which it actually recommends to you other sources that you don't know of?
[00:26:01] Steven Johnson: Yeah. There, there will be a world like that, Jeremy,
[00:26:05] Jeremy Utley: and the not too distant future.
[00:26:08] Steven Johnson: . How about I'll say like my meet the meeting that I'm going to as soon as we get off this call is directly related to that.
It seems it seems
[00:26:15] Jeremy Utley: like a natural evolution, right? It's like it's, you know, have you read this work? Have you seen this paper? You know, it's one thing to query the stuff that you already know of, which is an incredible feat. I can hardly remember the things I supposedly know, but then. To go a step farther and say, if this was of interest, you know, the book you really should read.
[00:26:33] Steven Johnson: If only I knew a company that was really good at finding information that's scattered all around the world. It feels like there's a big white space in the market or something
[00:26:42] Jeremy Utley: like that.
[00:26:44] Henrik Werdelin: How about the tone of voice? A lot of time people complain about, you know, like. AI writing a little bit bland, obviously somebody at your caliber of authorship is probably very sensitive to the tone.
Have you found like a good way of making it write a different way or prompting it to do something different?
[00:27:06] Steven Johnson: Yeah, it's a great question. So there's two things about tone, actually, like one of the things that's interesting, it's been our kind of ethos in general, and I think this is true. Still in the product, um, so, so I did a style guide actually for notebook LM, like as if you were a magazine editor, like this is what the model should sound like. , and we've kind of used that as a kind of like Bible for like the voice of the model, um, from the beginning. And one of the things that,
[00:27:34] Jeremy Utley: by the way, for your version or for everyone's version, just
[00:27:36] Steven Johnson: like, no, for everyone's version, like kind of, kind of, kind of what's the default setting.
And one of the things that. , it was in the kind of system level prompt for it is like, don't, don't refer to yourself in the first person. Um, so it doesn't, it, you know, it doesn't try to be, uh, uh, an entity. It doesn't try to be your friend. You know, it's kind of just like, here's the information you want.
It will occasionally say like, I'm sorry, I can't help you with that. But generally it should not actually ever refer to itself as, as an eye, which is like, I don't know, I may be like, out of the mainstream on this, but to me, it's like, I don't want you to pretend to be a person. I just want you to give me the information that I want in the most efficient way possible.
But, , so. But in terms of the voice, I did some very early experiments when I got to, , Google with, with style. , I tried to create, I gave it a, an earlier, this is back in the palm days, but I gave it a bunch of like Orwell and I was like, okay, right. Like Orwell and translate this passage into Orwell language and translate this passage into like Dickens language.
And it was, it was pretty good at it. And the thing we've always talked about is like, You know, could we somehow actually train a model on my My voice and you know get it to what would what would it look like if it wrote in my style? We we haven't done that yet, but that's something we've talked about.
Um, and I can imagine how would you
[00:28:59] Henrik Werdelin: feel about that? A lot of people I would imagine like I don't want to put my Content on like the, the notebook, because, you know, I'm sure Google is then going to write the book that I have always been wanting to write.
Yeah. Not true. But like, do you have any concern as an author of like uploading your books and having the model have access to it and all those different things?
[00:29:20] Steven Johnson: Well, we've, one of the things that's really important about the way that we designed notebook LM from the very beginning, we've been very kind of rigorous about this is we don't.
do not train the model on anything that anybody uploads. So what we're doing is simply when you upload sources, we are taking information from those sources or the entirety of the sources and putting it in the, Context window of the model, the short term memory of the model. And we're saying, answer this question or, you know, come up with this summary based on this information.
The second that query is over, that information gets wiped and there's no like long term training or long term storage of that information. So, which means, which is an important for, from a writer's point of view, from a privacy point of view, , also important from a copyright point of view. So if you have.
the right to, you know, assemble quotations from a book that you've paid for as a reader, you have the ability to use those inside of notebook LM without worrying that some bit of copyright information, copywritten information would be used to train the model in the future. So, so I feel secure about that.
The trickier thing is kind of what you're also kind of alluding to is like, how do you, how does one feel about like, You know, if I were to actually train a model on my writing style, , you know, is that,, is it creepy? Is it alarming? Would it be if it really does start to sound like me? And if, you know, is there a world in a year or two where , one of these models trained on my voice, given a kind of topic and some sources, could plausibly write a chapter that might pass, pass the Turing test of, like, Stephen's writing.
, that is imaginable. , would I use that to write a book? No. Would I use it to, like, maybe help Suggest alternate phrasings or things like that. , yeah, you're no, you're no, it wasn't,
[00:31:16] Jeremy Utley: you're no, wasn't conclusive by the way. You know, had a question mark just for the record. For if we're looking at Hendrix transcript, you're no, it was like a, I don't think so, but maybe, right.
Is there, am I reading
[00:31:28] Steven Johnson: between the lines there or
[00:31:29] Jeremy Utley: how
[00:31:29] Steven Johnson: do you think about it? It's like, for me, I think, I think what's going to happen is as tools like You know, notebook LM become more common and as folks who, who write for a living, , embrace what, what they do, the initial phase is going to be, it just lets you do the thing you already do much faster, you know, and more efficiently with, and, and to stay in a kind of flow state. And instead of going through all the different tabs, instead of searching, you know, through all your old PDFs, you can just ask the question and get it and you can just write, you know, a more effectively, , and so it's going to be like a huge win. I think for for a stretch of time, maybe it's like 3 years.
Then there's a point, , where you raise the one has to raise a question like does writing become more like conducting than actually writing and you're like Okay, notebook LM version 5. 0 , I have an idea for a book about the Gold Rush. Um, suggest, you know, a structure for it. Okay. Yeah, that's interesting.
Change this, change this, change this. Okay. All right. I like this opening chapter idea. Suggest an outline for that opening chapter. Okay. That's good. I like that. Let's change this, change this, change this. Um, now let's work on that first paragraph. Gimme a first draft in my voice please. Um. Yeah, that's pretty good, but that first sentence is wrong.
Let me just rewrite that by hand.
[00:32:54] Henrik Werdelin: Like it seems hard to believe that it won't go that way. Right. Yeah.
[00:32:58] Steven Johnson: Yeah. And I think, I don't know. Um,
[00:33:01] Henrik Werdelin: and would that be, the benefits is easy, but I wonder like in, in, in. A way that, you know, we invented Instagram. I'm sure that nobody thought that that would kind of lead to the demise of mental health for kids.
Right. And I was just like, it was very difficult to fathom that kind of link. , if we were thinking about those like three or four years out and we are no longer chit chatting. Over a meeting because most of us are just brainstorming with a notebook.
I am, and we're doing all these different things. Like, what do you think is some of those kind of like not necessarily thought through consequences of this, or maybe what do you worry about?
[00:33:42] Steven Johnson: Yeah, I think the, I mean, one obvious one is a, , using these tools, not necessarily notebook wouldn't be particularly well suited for this, but, um, other related tools to.
Just create kind of junk content. I mean, you know, there's already been this kind of concern about the web, just getting overrun with kind of middling, kind of vaguely interesting, but also formulaic AI writing that is just like, it's like
[00:34:13] Jeremy Utley: novel spam, right? It's it's book spam. Yeah. Yeah. Yeah.
[00:34:16] Steven Johnson: And I think that that, um, I think that there's probably some version of the future where we do need to.
Kind of authenticate human participation in in just as we're talking about. Um, You know, watermarking images to, to show that they've been generated with AI, you know, a lot of interesting investigations in that front, that there may be some kind of point of like, Hey, this, this novel was in fact written by a human, and I only want to search for things written by humans.
[00:34:50] Jeremy Utley: Is that a slightly potentially anachronistic view?
I mean, it's interesting to consider that our default assumption Is that a human? produces a better work.
[00:35:00] Steven Johnson: Yeah.
[00:35:01] Jeremy Utley: Is there, if the goal, like Guy Kawasaki said something to us that I, that I've really been thinking about ever since our conversation, which is my agreement with the reader, isn't that I write every line, but that I produce the best possible book for their consumption.
[00:35:17] Steven Johnson: It's, I mean, well, look, first off, there's so many questions and that's why it's such an incredibly interesting inflection point to be thinking about these things. . I feel confident that the, the best books.
just to think about books. The best books will, um, whether they are written just with a kind of, like, virtual research assistant like Notebook LLM is in its current form, , or whether they're maybe written in a kind of conductor mode, , where there's a real partnership between the human and the AI, and perhaps some future where the AI is the leader, , in the exchange.
The best books produced will, will get better and better thanks to these tools.
So that's good news. Then there's a question of like, you know, what happened to the
[00:36:01] Henrik Werdelin: question is book themselves. I mean, like if I just look at some of my behavior now, I buy a book, I then put it into notebook. And then I basically are in conversation with Steven Johnson about something.
And then obviously the format is then tailored To where I in my mind and the narrative kind of get crafted around my journey, which seemed to be a more modern potential way of doing it, where the book format, of course, is something that was fitted specific format because you had to print it. And that was therefore how you created that format.
[00:36:38] Steven Johnson: Yeah, that was exactly the next thing I was gonna say, which is that not only will like the underlying text in traditional book form at the high end get better and better. Thanks to these tools. But. new forms of reading will become possible. , and already right now, as you say, like if you load a book into like, I have where good ideas come from in a notebook and it's so cool to.
, in a sense, read it by particularly if you've already read it from the beginning, so you've read it once and it's straight back to, you know, fun, or you've written it. So you know it really well. But then you read it by following the suggested questions we get. So, you know, whenever you're in dialogue, we're always suggesting like a follow up question or two based on what you've just said and based on the book.
And, and then, and the, the answers are now so nuanced. And, you know, it's, you know, notebook writes, he's like, Four or five paragraph little mini essay answers and and the questions that it suggests are very deep And so so again, we you know We started this conversation with talking about like associative trails in your mind when you're walking Right, and the reason why that's powerful is that you?
Are able to follow a trail of associations. And that gets you to a new insight that you might not have otherwise gotten to notebook does that as well. If you read, instead of reading that kind of in a linear way, you read it by following trails of questions. Um, and that, when I first saw that feature live in the product and I did it with my own book, like I was like, wow, this is such an incredible way to explore.
[00:38:10] Jeremy Utley: Just describe that. Somebody say they log into notebook LM. How would they follow a trail of questions? Just like very basic, practically.
[00:38:18] Steven Johnson: Yeah. Okay. So imagine you upload, you could upload any document, but if you happen to have a PDF of a, of a book of some form that you have access to, let's say you upload that almost immediately notebook will, it will generate what's called a notebook guide, which will give you a summary.
And it will give you, suggest like three high level questions for understanding the material. But, you could start with your own question, and whatever you're interested in, you could just type in and say, tell me more about ant colonies in this book, whatever it is. At that point, you'll get an answer, and then on the screen, right above the chat box, you will see, Three little questions that are like based on the, what you've just asked and the material in the book.
, I had this bias coming into the product because I've trained as a writer and a journalist, like I know how to ask good questions of material or people I'm interviewing, but that's not a skill that, you know, everybody has and people often sat down a notebook and they were like, I don't know what to ask.
And so we were like, Oh, let's suggest questions. And then once we did that, we realized that that was a whole other way of exploring the material that was really interesting. And so to, to get back to like the original. Thing you were saying, Hendrick, the, the, you, it creates new kinds of books. It's almost like the, kind of the dream of the hypertext book that You know, it was animating me when I was in, in the late eighties, even before the web, where people were like, people are gonna write new kinds of books with links that you can follow and go in different directions.
Go to page
[00:39:42] Henrik Werdelin: 264 for Yes. Yeah. Kinda choose your own adventure. Yeah, adventure idea
[00:39:47] Steven Johnson: books. And, and it never really happened, you know, like it happened only really in the form of Wikipedia. , But other than that, like people generally kinda read in a linear way still. , you know, they read the web. Follow links, but then once I get to the article, they're pretty much reading start to finish.
But now you have this like conversational way of exploring a large body of information and that is that is really new and was never possible before and so I think The combination of like maybe better books and new conversational ways of exploring those books, um, is, is going to be, um, you know, something I feel, I feel really, it's harder to see the downside
[00:40:28] Jeremy Utley: Last question from me is what are, what are pitfalls you would recommend people avoid and how? So somebody comes to NotebookLM for the first time, what's dumb stuff or stuff, areas where people get stuck, um, underperforming capabilities, missed opportunities. What should people avoid doing? I
[00:40:49] Steven Johnson: mean, the biggest thing, our biggest problem, I think, is that You can't do anything without sources.
So it's still excellent at answering questions, but sometimes you used to be able to say like, Hey, write a sonnet based on These non poetic sources and it would do a great job. But sometimes now it says, well, sonnets are not mentioned in these sources. So it can't do that. And I'm trying, I'm actually, one of the things that we're going to right now is trying to figure out like what happened, like why did it refresh the page?
I find sometimes just refreshing. It's like, I get it. I get a new chat bot, right? Yeah. If that's something, no, something, something changed and I don't, I have to have it specified what that is. It's not, , it's not a calculator. It will, it will retrieve. numerical facts from, , your documents quite accurately, but you know, you don't want to ask it to perform like statistical analyses of things.
Um, it just, their language models, not number models yet. Um, so that's, that's one you want to watch out for. , The one thing to embrace now, uh, that's just amazing is how multilingual it is. So we're like, we're seeing this huge, , spike of usage in Japan. And one of the things that people are doing is that like, there's a lot of English documents that Japanese people have to deal with, but like English is not as common a language in Japan.
And so people are able to upload an English document and have an entire conversation with the model about the document in Japanese. That's just kind of crazy that you can do that now. So anything, I mean, I even have, and the image stuff that we have now too. So, um, Walter Isaacson has been using it for this, his new biography that he's doing on Marie Curie and, and her husband.
And I have as a sample, this is handwritten notes. Marie Curie wrote in like 1902 in French, like a scan of it, scrawled French, like with a bunch of like chemistry, like symbols and stuff like that. And I put that image on a slide and brought it into Notebook LM and I was like, okay, you know, translate this and explain what you think its significance is.
And it's, it's just a bunch of pixels, right? It's just like pixels showing, like, you know, scrawled handwriting from 120 years ago in French. And it's like, Oh, this appears to be a document. It looks like it might be written by Marie Curie. Here's what it says. This is a shopping list, two tomatoes, the three cubes.
Yeah. Yeah. It was kind of disappointing what it turned out to be. , but the fact that, you know, you could do that kind of stuff, , with images now, you could put a chart in there and it will understand it. Even if it's just pixels, even if, you know, there's no like OCR on the charter. Text associated with the chart.
, that stuff is really cool. I only talked about the limitations for like three seconds. Then I started talking about the good things. I'm sorry. I was just too excited about this. You're just a proud, you're
[00:43:33] Jeremy Utley: a proud Papa. Nobody
[00:43:35] Steven Johnson: can
[00:43:35] Jeremy Utley: blame you.
[00:43:36] Henrik Werdelin: Well, people should definitely go and check it out. And I think we're about to reach the end of like the time we have, but, , anything else that you want to add, all the things we forgot to ask you about.
[00:43:50] Steven Johnson: No, I think we covered it all. We just love to get feedback from people and, you know, please do try it. And, and as I said, we do have a discord where people are welcome to come and post their, like, we just love to hear use cases that people come up with it. It's, , it's a very like flexible platform. Like you can push it in all sorts of like directions that like aren't in our heads yet.
[00:44:12] Henrik Werdelin: That's awesome. And we look forward to reading the book about the puppets of ants going to space, you know, connecting the
[00:44:21] Steven Johnson: dots.
[00:44:23] Jeremy Utley: It's all, uh, all innovation is just recombination, right? It sells itself.
[00:44:27] Steven Johnson: I mean, it really just rolls
[00:44:29] Jeremy Utley: off the tongue so easily. It's incredible.
Right on. Super fun, guys. Alright, be good. .
[00:44:34] Henrik Werdelin: Jeremy, tell me your, uh, reflections.
[00:44:37] Jeremy Utley: You know, I thought I was expecting to get insights into the future of authorship. I thought it was really cool thinking about the future of readership.
That was particularly cool. And then of course, you know, my mind can't help but go to all sorts of unexpected use cases well beyond call it writing and research, which I think are just, you know, Weeks away, if not days away, um, for organizations to think about leveraging to create their own, call it organizational notebooks to my mind's just spinning far beyond kind of the ability to write and read.
This tool really has a potential to be revolutionary. And so super cool to hear about his journey. Super cool. I mean, having been a fan of his for a long time and seen his obsession with tools for thought and things like that, to see it actually come to fruition. Yeah. Is really fun and neat and then just to be sparked into what are all the possible Use cases far beyond even his imagination.
I I think it was super fun discussion. What about you? What struck you?
[00:45:38] Henrik Werdelin: I think i'll plus one to everything you said. I mean like this idea of what is the future of books It's fascinating. Like, we'll just have long conversations with kind of like the, the, uh, synthetic version of, of Steven on like a new topic.
And, and then what would that look like? I think that'd be incredibly fascinating. Having been somebody who's been in the incubation space for a long time, obviously quite fascinated about this idea of a new Google apps and kind of like a new way of trying it. And I think it's fascinating, this kind of core concept of bringing external people in to be the domain experts, you know, have the, the customer founder fit, if you like, and then merge them with obviously the brain power of all the Googlers, uh, I'm excited to see, uh, to see what's going to come out of.
I
[00:46:25] Jeremy Utley: thought just to your point on kind of the model for innovation, super interesting to think about Google bringing him in as call it a lead user and then employing him, right? Cause now he's a full time employee. And that to me is pretty fascinating. I think in most of the innovation efforts I've seen, there's, there's an attempt at call it a separation, so to speak, of we're designing for this person who's outside us, who's other than us.
And there's a lot of, I would say, even, um, kind of deprecation of, of models like co creation. And so to, for Google to go from co creation to actually employing the customer in a way or employing the user, it's just a, I, I haven't actually processed it all, but now that you mentioned it, it's a pretty radical departure in terms of innovation model.
What do you, what do you think? What's your, if you had to, if you had to make a prediction for the future about this model of innovation, any, anything that that sparks for you as an innovation expert yourself?
[00:47:20] Henrik Werdelin: Well, most of these labs and incubators, you know, like I have a tough time, right? Because anything that's not a billion users is kind of like a miss.
And so, uh, I think what's cool about both him as a person, and I think this approach is that it seemed to be much more driven by curiosity without necessarily kind of a goal of making a billion user product. And I do think back to his point about playing, um, That that is probably the best way you get to these, uh, disproportionately big outcomes.
[00:47:56] Jeremy Utley: Yeah, I mean, you always want to hire the person who doesn't need a job, right? And so hire the customer that doesn't need a job is probably a pretty great guideline there as well. Right? But they're, they're willing to follow their curiosity and their interestingness to great places. Super cool.
[00:48:12] Henrik Werdelin: I think that, I think we all for this time.
[00:48:16] Jeremy Utley: Thanks for listening. Thanks for joining us on this adventure. If you enjoyed this conversation with Steven Johnson, please smash that like button, smash that share button, recommend it to a budding author or reader or researcher in your life who may have strong opinions about the future of authorship, readership, and technology.
[00:48:35] Henrik Werdelin: See you, my friend.