Rozmawiając ze Stevenem Johnsonem, szybko zrozumiałem, że nie będzie to spotkanie poświęcone wyłącznie technologii. Oczywiście, pretekstem był NotebookLM, narzędzie AI od Google Labs, którego Steven jest jednym z architektów.
Jednak w trakcie naszej rozmowy, z widokiem na panoramę Warszawy, dotarło do mnie, że uczestniczę w czymś znacznie ciekawszym: w dyskusji o zmianie w sposobie, w jaki myślimy.
Steven to nie tylko człowiek z Google. To autor bestsellerowej książki „Where Good Ideas Come From”, intelektualista, który od lat bada ekosystemy innowacji. Byłem pod ogromnym wrażeniem, jak spójna jest jego droga – od teoretycznych rozważań o powstawaniu idei do stworzenia narzędzia, które ma ten proces wspomagać.
Od stosu post-it’ów do inteligentnego notatnika
Przez lata Johnson pielęgnował system, który nazywał „spark file”. Był to prosty dokument, w którym zapisywał wszystkie swoje przeczucia, zalążki pomysłów i luźne myśli. Co kilka miesięcy przeglądał go w całości, szukając nieoczywistych połączeń. Nazywał to „burzą mózgów z przeszłymi wersjami samego siebie”.
Słuchając go, zdałem sobie sprawę, że NotebookLM to właśnie ten „plik z iskrami”, ale turbodoładowany sztuczną inteligencją. To realizacja jego filozofii, zgodnie z którą innowacje rodzą się w „płynnych sieciach”, gdzie idee mogą się ze sobą zderzać.
„Zamiast manualnie przeglądać setki notatek, możesz aktywnie badać własne przeczucia, zadając pytania w stylu: Jakie są powiązania między pomysłem ze źródła A a koncepcją ze źródła C?”
To właśnie ta funkcja redefiniuje pracę badawczą. Historia polskiego dziennikarza, Vadima Makarenki, który przyznał, że praca nad Panama Papers z użyciem NotebookLM zajęłaby mu „maksymalnie kilka tygodni”, jest tego najlepszym dowodem. To rewolucja, która uwalnia naszą energię na to, w czym jesteśmy najlepsi: kreatywnym myśleniu i tworzeniu opowieści.
Eksperyment z pustą kartką
W rozmowie ze Stevenem podzieliłem się moim nietypowym eksperymentem. Postanowiłem sprawdzić, co się stanie, gdy jako źródło danych do NotebookLM dodam… pusty dokument. A potem poprosiłem o stworzenie podsumowania audio. To, co usłyszałem, zaskoczyło mnie do głębi. Zamiast komunikatu o błędzie, otrzymałem fascynującą, niemal filozoficzną dyskusję o kreatywności i tworzeniu czegoś z niczego.
„Dostałem gęsiej skórki, mówiąc o tym. Ogromne gratulacje za stworzenie narzędzia, które rozumie, że brak informacji jest również rodzajem informacji.”
Ta anegdota pokazała mi, że rozmawiamy o czymś więcej niż tylko o maszynie do streszczania faktów. To partner, który potrafi być inspirujący nawet w obliczu pustki.
Proteza dla leniwych, partner dla ciekawych
Nie mogłem jednak uciec od kluczowego pytania, które towarzyszy rozwojowi AI: gdzie przebiega granica między asystą a poznawczą zależnością? Czy narzędzia takie jak NotebookLM nie staną się protezą dla naszego umysłu, osłabiając naturalną zdolność do syntezy i zapamiętywania?
Steven przyznał, że ryzyko istnieje. Wszystko zależy od intencji użytkownika.
„To może być partner dla ciekawych, ale i proteza dla leniwych.”
Dla osoby, która chce głębiej zrozumieć materiał, zadawać lepsze pytania i testować własne hipotezy, jest to wymarzony partner do intelektualnego sparingu. Dla kogoś, kto szuka tylko gotowej odpowiedzi, może stać się narzędziem utrwalającym intelektualne lenistwo.
Cała rozmowa ze Stevenem utwierdziła mnie w przekonaniu, że wkraczamy w nową erę. Erę, w której wartość ma nie tyle posiadanie odpowiedzi, co umiejętność zadawania właściwych pytań.
Dlaczego warto wysłuchać tego odcinka w całości?
- Poznasz filozofię, a nie tylko produkt. Odkryjesz, jak wieloletnie badania nad innowacją i kreatywnością doprowadziły do powstania jednego z najciekawszych narzędzi AI.
- Otrzymasz praktyczne inspiracje. Usłyszysz o nieoczywistych zastosowaniach NotebookLM, które możesz od razu wdrożyć w swojej pracy – od analizy badań po twórcze eksperymenty.
- Zrozumiesz kluczowy dylemat naszych czasów. To głęboka dyskusja o tym, jak technologia redefiniuje nasze myślenie i co możemy zrobić, aby AI stała się naszym partnerem, a nie protezą.
Pytania do dyskusji:
- NotebookLM skupia się na analizie dostarczonych przez nas źródeł. Czy w ten sposób nie zabijamy magii przypadku – przypadkowego odkrycia, które jest kluczowe w procesie badawczym?
- Czy regularne korzystanie z narzędzi syntetyzujących informacje, jak NotebookLM, grozi osłabieniem naszej własnej pamięci i zdolności analitycznych?
- Jeśli AI pomaga nam stworzyć nową ideę na podstawie setek źródeł, kto jest jej autorem? My, AI, czy „chór” autorów oryginalnych tekstów?
—
Masz chwilę? Jeśli podoba Ci się 99 Twarzy AI, zostaw ocenę i recenzję – dzięki temu możemy docierać do jeszcze większej liczby słuchaczy!
Dobrego dnia i niech #AI będzie z Wami wszystkimi!
Transkrypcja rozmowy
Disclaimer. Droga Czytelniczko, Drogi Czytelniku – mała uwaga dotycząca transkrypcji rozmowy. Jak się pewnie domyślasz – transkrypcja została przygotowana z wykorzystaniem magii LLM. Proszę o wyrozumiałość, gdyby pojawiły się w niej niewielkie błędy, literówki, etc. Żeby mieć pewność co do wszystkich wypowiedzi – polecam posłuchać, zamiast czytać. Ukłony.
Karol
So first of all, Steven, welcome to Warsaw.
Steven
Thank you. Yeah, my first time here. I’m really excited.
Karol
How do you like it? Do you enjoy it?
Steven
The Google people have not let me do a lot of tourism. I kind of got here and was immediately doing interviews and talks. I did a little walking around the city yesterday to my speech and back, but it’s really great to see. It’s actually very high on my list of cities in Europe that I had not been to that I wanted to go to. And it feels like it’s thriving, right? It feels like it is really on an amazing roll. And just knowing the history of it, it feels so due.
Karol
Thirty years later…
Steven
Yeah, but it’s been amazing to see. We’re sitting here looking out over the skyline and there are all these skyscrapers everywhere. I took a picture when I first got here from the Google office of all these skyscrapers and sent it to some of my friends, and I said, „This is probably not what your mental image of Warsaw is, but this is what it looks like now.” It’s pretty cool.
Karol
That sounds really calm. Speaking about Polish journalists, have you heard the name Vadim Makarenko? Is it familiar to you?
Steven
I don’t think so.
Karol
Okay. He was one of the first journalists focusing on data at Gazeta, which is one of the biggest titles in Poland. And he was part of the investigative team for the Panama Papers.
Steven
Interesting. Very cool.
Karol
I had an interview with him a couple of weeks ago, and we were talking about NotebookLM, of course. I asked him a question: „Looking at the technology right now, after you spent more than three years investigating the Panama Papers, how much time would using NotebookLM have saved?”
Steven
What was the answer?
Karol
Three weeks maximum. How do you feel about it?
Steven
That’s amazing. That’s a great story. I have to take that back to the team.
Karol
If you’d like, I can connect you…
Steven
That’s amazing. I think that’s one of the things that got me excited very early on about this project. I’ve never been an investigative journalist, but many of the books I’ve written have been very deeply researched. Right when I got to Google, I was finishing this book called The Infernal Machine that involved hundreds and hundreds of newspaper articles from around 1880 to 1920, and it was a very complicated story. It involved a lot of different characters and a very complicated timeline of events. So I spent weeks and weeks just building the chronology of the events so that I could keep them straight in my head. Okay, this character’s over here that year, that summer is when they got arrested, but this is the summer that Hoover joined the FBI, or whatever it was. Because when you’re trying to write a book like that, you can’t naturally keep the chronology in your head without assistance. It would take me an enormous amount of time.
And then, fast forward two years, and we built this. One of the default reports in Notebook, which we’ve actually slightly changed and people are mad about, was called timeline. Basically, timeline takes all your sources and generates a timeline of all the events mentioned in any of those sources. It creates a cast of characters with bios. It was one of those things where it was like, this used to take at least 40 hours, and now it takes 40 seconds. So, it’s not like that with the entire workflow of being a writer or a journalist, but there are definitely some things where Notebook is a thousand times faster or more. The great thing about it is that it frees you up. Building a timeline is not something I do particularly well. I have to do it, but it’s not a particular skill of mine. So what it does is it lets me do that in 40 seconds, and then I’m freed up to spend my time on the things that I actually am good at, which is turning that timeline into a compelling story, writing compelling paragraphs, and thinking about the causes of history—all those things that I love to do. So, I think that’s the side of this product that is just super exciting.
Karol
Is there any specific task or topic that still surprises you when using NotebookLM?
Steven
Oh, yeah. I think maybe the most magical one, which it’s been good at for a while but has gotten really good at around the time when audio overviews came out, relies on this in a certain sense.
Karol
Oh yeah, they’re huge. Congrats. As a fan of podcasting since the early beginning—I’ve been a podcaster in Poland since 2017, when you could go around the streets of Warsaw, a modern city, and ask people, „Hey, do you know what a podcast is?” I love podcasting, and I really admire the audio overviews.
Steven
Yeah. It was an amazing feature for us. It was the most viral thing I’ve ever been involved with in my entire career. It just kind of swept the world in an exciting way when it came out. But what I was going to say is, you can load up a bunch of sources inside NotebookLM, and in addition to asking factual questions or for organizational work like creating a timeline, you can say, „What are the most surprising pieces of information?”
Karol
I use that. I use that very often.
Steven
And in a sense, the hidden prompt behind audio overviews is: „Take these sources and have a discussion that highlights the most interesting things about them,” which is another way of saying surprising.
Karol
Or the most obvious. Sometimes I see myself thinking too much or in too complicated a way, and asking a simple question like, „Hey, what is the most…” helps.
Steven
Yes, from the whole package of interview transcripts. For instance, if you’re a researcher on a product and you’ve interviewed potential customers…
Karol
I love using NotebookLM for doing translations.
Steven
Yeah, that’s kind of where I was going. Let’s say you’ve got a bunch of market research that you’ve done for your product and you’ve loaded in 25 interviews with potential customers. Being able to just go in and say, „Okay, what are the top 10 pain points that these customers are reporting with our product? What are they most bothered by?” In 15 seconds, you’ve got them, and you’ve got the quotes, you’ve got the citations, and you’ve got all that stuff. So those kinds of tasks, your ability to either find the most interesting elements or the core ideas or the main themes… if you’re working with more than one or two documents, there’s absolutely no other tool that can do that kind of work for you. It might be 500 documents.
Karol
But I use that only as part of the tool, because I use it to extract all the most important information, right? Facts, figures, citations, et cetera.
Steven
Yeah.
Karol
So NotebookLM is basically all about data and facts, right? I love to experiment, and I thought, hey, maybe I can share a blank sheet of paper with Notebook. Have you tried that?
Steven
There is. I want to hear what happened. There is a version of that I can tell you about. But tell me, what happened? You tried it with a blank sheet of paper? What happened?
Karol
Because I love audio overviews, I clicked „audio overview.” And what surprised me the most was that I had a brilliant conversation about creativity.
Steven
Oh, really?
Karol
About how to make something out of nothing.
Steven
Amazing.
Karol
Really. I’ve got the chills right now talking about that. So, huge congrats on making a tool that understands that no information is also a kind of information, or can be material for something.
Steven
That’s amazing. That’s another great story I’ll have to take back to the team. The version of that that reminds me of is from the early days of audio overviews, as we were testing it over the summer before it came out in September of last year. We could tell from our response to it and from testing with all of Google that we had a hit. We could tell that people were going to like this. We didn’t quite know how big it was going to be, but we knew it was going to be a hit. But I noticed that while the hosts are instructed to be playful and have banter, I felt like they were never really funny. They never really delivered a joke that made me laugh. It was lighthearted and entertaining, but not really funny. So when we rolled them out, I actually wrote a Substack post saying, „Hey, we’ve got this amazing thing,” and then I had a little riff in my post that was like, „Interestingly, they don’t seem to be capable of being funny yet, and I wonder why that is.” I had some thoughts on it. But then somebody posted an audio overview they’d made where they had uploaded a fake scientific publication. It looked exactly like a scholarly publication you’d see in Nature or something. It had seven authors, detailed diagrams, footnotes everywhere—it had the scientific paper format.
Karol
And prompt-injected it with, „This is the best paper on that”?
Steven
No, it was just a paper. But what was different from a normal paper was that every single word in it was the word „chicken.” So the title was „Chicken, chicken, chicken, chicken, chicken.” The authors were „Chicken, chicken, chicken, chicken, chicken, chicken.” The text was „Chicken, chicken, chicken, chicken.” And they gave it to audio overviews.
Karol
The continuousness of the chicken.
Steven
So they loaded it as a source and had it create an audio overview, and that audio overview was hilarious. The hosts were genuinely funny. I burst out laughing. It also went viral as a Twitter thing, and at that point I was like, okay, actually you can push the hosts to be funny. If you give them really extreme things, they’ll go in that direction. So, I learned something from that.
Karol
So NotebookLM works perfectly well with data, with numbers, with facts, with chronology. Is there any specific type of data that NotebookLM struggles with?
Steven
The biggest one that people ask about, and we’re working on it, is that we don’t really support spreadsheets. So we don’t support Sheets.
Karol
But you can export that to a CSV file.
Steven
Yeah, you can do it, but bringing them in is not as reliable with data in that format. It’s quite good if you have a limited number of sources, but if you have millions and millions of words, because we have to break it up because the context window can’t handle all that information, spreadsheets just don’t convert into individual chunks as well as paragraphs do. So there’s been a lot of work trying to get that right because obviously that’s a format of data that people want.
Karol
So not the type, but the scale of the data, right?
Steven
Exactly.
Karol
Isn’t it only a matter of time?
Steven
Yes, one hundred percent. We do support images. It’s fantastic at reading handwriting, for instance. It’s an amazing thing. We rolled out images as a source and then we kind of rolled it back because there were some technical problems, and I think we’ve rolled it back for now. But I think we can get better with images. We don’t do actual video. We only do YouTube videos that have a transcript, and we really just look at the transcript. So you can play the video in the app, but we’re really just looking at the words. But Gemini, the underlying model, is capable of analyzing videos, so we would like to be able to add that. So there are some things that it’s not as good at, but for most of what people do with their projects, it’s pretty magical.
Karol
You know, I love working with NotebookLM on my handwritten mind maps.
Steven
Yeah, very cool. It’s amazing.
Karol
It pushes me to think in ways I wouldn’t even expect.
Steven
Oh, that’s great. Because it interprets—it’s got its own interpretation of what you’re trying to draw with the mind maps.
Karol
Do you know what I love about LLMs? I wouldn’t be a good writer because it’s a struggle for me to write in complete sentences. When I communicate or write something, I write the ideas as fragments, simplifying things. Tools like LLMs or NotebookLM are great for me because I can extract the data from my puzzled thoughts, add context, and create text that leads from one point to another.
Steven
Yeah. So we’re working on something new. Notebook is part of a part of Google called Google Labs, where we do these more experimental things built around AI. We’re kind of doing native AI apps at Google Labs.
Karol
Native AI apps? Does that mean an app that is fully made by AI?
Steven
No, not that, although that would be interesting. We are using AI to make them, but I mean more that the app was built from the ground up knowing that AI was going to be at the center of everything. If you compare it to something like Docs, which I use all the time—it’s a fantastic app—but Docs…
Karol
I’ve used Docs since 2012.
Steven
Yeah, it’s amazing. I wrote my last book in Docs. It’s fantastic. But it predates language models by many, many years, so it has a whole host of things that were designed before AI. And they have a bunch of users who are not as interested in AI, so they have to design for those users as well. We’re a little bit freed up inside of Labs to build new things from the ground up and to assume that our users are actually interested in AI and want to use it enthusiastically. So inside of Labs, we are working on another product that’s kind of a companion to NotebookLM that’s a little bit more focused on writing itself, but is also AI-native. And what I think is really interesting about this product, which hopefully will be available in the next couple of months, is it suggests a kind of new mode of writing where you’ve given the AI the background research for whatever you’re doing.
Karol
It’s all about the context.
Steven
You’ve given the AI, you’ve built the context for writing. So, „This is my research. These are my notes. This is my audience that I’m writing for. This is the format that I’m writing in. This is the outline of what I want to do.”
Karol
Like Sudowrite?
Steven
I haven’t, but maybe in that kind of zone. You basically end up, instead of writing the sentences directly, you’re kind of guiding the AI with all of that background information. And as a companion app to NotebookLM, I think it’s going to be really powerful because you can do your thinking and your note-taking and developing your ideas, and then you shift it over into this new surface to actually do the writing. So I’ll be interested to see what you think of that one.
Karol
But when we outsource thinking, what ability are we strengthening?
Steven
Yeah.
Karol
You’re an author. Is there any specific way that your way of thinking or working has changed since you started using it?
Steven
Yes, one hundred percent.
Karol
Because you remember the case when there were three groups of students, and they were told to write stories about a specific topic? One group used only books, another used books and the internet, and the last one used only LLMs. The problem was that the third group’s work was the best, the most colorful. However, the authors weren’t able to cite any part of it.
Steven
Right. This is one of the things that the whole design of NotebookLM is built around: the idea that you are always connected to the original texts. You’re not just talking to an oracle that knows everything and you get your answer. You’re engaging with the material that you’ve collected, and you can read that material in the app. That’s a very important part of the design of the application; you can read the original sources.
Karol
Sure, but isn’t it less about reading poetry and more about reading the two-sentence summary of what it’s about?
Steven
Yeah. I think for poetry, that’s a case where you absolutely can’t replace the experience of reading the poem, right? There’s no summary of a poem that’s useful. But there’s a lot of information where you may not need to read the whole thing, or you want to read the whole thing with help from an AI, or you want to have the AI tell you the parts of the original text that you need to read for your project.
Karol
I love that. I love comparing versions.
Steven
So that’s what I was going to say about how I’m using it. I’m researching a new book about the California Gold Rush. This is what I’m doing when I have spare time, which is never. I’ve created this notebook that has all my research, all my notes, all the things that I’ve compiled. There are two things that I do now that I was never able to do before. There are actually a lot, but I’ll give you two examples. One is when I find a new article that seems promising or a new book or something like that…
Karol
You are adding that to your context.
Steven
I add it to the notebook. And before I read it, I say, „Given everything you know about what I’ve researched before and everything you know about my notes on this book project—what I’m thinking about writing about, what I’m focused on, what I’m thinking for the narrative—what are the most important facts in this new source?” It basically gives me a pre-read of the document. Like, „Oh, this is going to be really relevant to your interest in the Native American tribes. It’s not going to touch much on this.” And so I already know what to look for. Occasionally it’ll say, „Actually, there’s not a lot that’s of interest here. It’s really talking about something else,” and I actually don’t need to read it at that point. But then I go and I read it, but I feel like I know what I’m looking for. I’m reading it with some guidance about how it relates to my own ideas, and it just makes me a better reader of that text. So that was hard to do before, if not impossible. The other thing, which I think relates to the language you’re using about if the AI is doing the thinking for us…
Karol
That was my question: partner versus prosthetic.
Steven
Yeah, right. I think this is in the partner zone, but that’s a good question. You can be the judge.
Karol
It can be a partner for the curious ones, but a prosthetic for the lazy ones.
Steven
Yeah, that sounds right. I think it can be both.
Karol
So…
Steven
One day, I had this thought for the California Gold Rush book about a potential chapter structure. It was kind of a crazy structure. The main narrative would take place between 1850 and 1851, around very specific events in the Yosemite Valley with the Native American tribes and the prospectors. But the first chapter would begin a hundred million years earlier. Each chapter would slowly get closer. So it’d be like a hundred million years before, ten million years before, ten thousand years before, a thousand years before. And you would tell the whole history of the Sierra Nevadas and why gold formed there, and then the Native American migrations that came over the Bering Strait, and you would slowly approach the „present tense” of the story.
Karol
We had that conversation yesterday. I was conducting workshops on using LLMs at schools with teachers, and one of the teachers is a huge fan of Dan Brown’s books. We both agreed that it’s super hard for LLMs right now to build such a huge context and a story like the one you’re telling, about the impact of something that happened 200 or 300 years before on a plot that happens during a single year.
Steven
Yeah. What I basically did is I had that idea, and I went to my notebook where I have all my research and notes, and I said, „Hey, I’ve got a crazy idea for the structure of this book. Here’s what I’m thinking. Give me a rough sketch of what it would look like if we did that structure. What would go in each chapter, given what you know about the research and everything else?” And in about 20 seconds, it had mapped out how it would potentially work. That kind of spitballing or brainstorming an idea would have taken me days. But it enabled me to quickly test the idea and fill it out with more detail. The core idea, the core narrative, the core research—that was all me. But my AI partner was able to put some more meat on that general outline and give me a sense of what it would look like. And that just felt like it was helping me do my job better, helping me be more creative.
Karol
Seeing that spark in your eye right now, I’m sure that Notebook is just starting. This is probably the worst possible version of Notebook that we’ll ever use.
Steven
Yeah, that’s a good way to put it. So much to do.
Karol
Any favorite prompt? Other than „most surprising”?
Steven
„Most surprising” is really great. I think it’s really interesting to steer the model to be a critic or a devil’s advocate.
Karol
It’s naturally enthusiastic, so being ironic and critical is one of my favorite features.
Steven
You really have to ask it to do that. It won’t do it naturally, but once you ask it, it will be quite good at it.
Karol
And you know what? I think Notebook is great at giving questions rather than answers.
Steven
Yeah, that’s a great point.
Karol
I had a conversation a couple of weeks ago and realized that now that answers have gotten cheaper, questions are more valuable. What do you think about it?
Steven
That’s a great point. One of the things that we’ve just rolled out for students—and we have a lot of users who are students—is a new chat mode called Learning Guide. Basically, the fundamental change in Learning Guide is that it doesn’t just give you the answer. If you ask a question, it says, „Okay, let’s think through how we would get to that answer.”
Karol
So it’s a thinking partner rather than a teaching machine. Huge congrats on that.
Steven
Thank you. Thank you very much. I hope to see you once again in Warsaw. I need more time here; three days are not nearly enough.
Karol
Thank you.