WEBVTT 00:00:00.879 --> 00:00:03.279 Most of us open a new research tool with, you 00:00:03.279 --> 00:00:05.299 know, pretty high hopes. Oh yeah, totally. We 00:00:05.299 --> 00:00:08.359 just drop a dozen disorganized PDFs into the 00:00:08.359 --> 00:00:12.740 void and pray for sudden wisdom. But that isn't 00:00:12.740 --> 00:00:15.179 research. No, it's really not. That's just digital 00:00:15.179 --> 00:00:17.100 hoarding. Right, because a true second brain, 00:00:17.339 --> 00:00:20.760 it doesn't just store files. It requires actual 00:00:20.760 --> 00:00:23.339 architecture. You know, something that organizes 00:00:23.339 --> 00:00:26.550 the chaos for you. So let's look at how the 2026 00:00:26.550 --> 00:00:28.890 version of Google's Notebook LM actually makes 00:00:28.890 --> 00:00:30.949 that happen. Welcome to the deep dive, by the 00:00:30.949 --> 00:00:33.429 way. Yeah. Glad to be here. Today, we are breaking 00:00:33.429 --> 00:00:35.350 down the software completely. We want to move 00:00:35.350 --> 00:00:39.729 away from treating AI like a simple search box. 00:00:40.369 --> 00:00:43.149 Exactly. We need a system. Right. A systematic 00:00:43.149 --> 00:00:45.630 workflow. Yeah. So we're going to explore intentional 00:00:45.630 --> 00:00:49.729 setup and autonomous research agents. Plus aggressive 00:00:49.729 --> 00:00:52.250 source auditing. Yeah. And this really cool concept 00:00:52.250 --> 00:00:55.170 called the Note Loop. And finally, translating 00:00:55.170 --> 00:00:57.229 all that research into functional media. It's 00:00:57.229 --> 00:01:00.229 a lot. It has a lot to cover. Beat. Yeah. So 00:01:00.229 --> 00:01:02.329 take a breath and let's figure this out together. 00:01:02.490 --> 00:01:04.689 We really need to start at the absolute foundation, 00:01:04.730 --> 00:01:07.030 I think. OK, where is that? Well, Nopagallam 00:01:07.030 --> 00:01:09.670 is no longer a separate island. It is baked directly 00:01:09.670 --> 00:01:12.849 into Gemini's main interface now. Oh, so it's 00:01:12.849 --> 00:01:15.670 just one tab. Exactly. You aren't just opening 00:01:15.670 --> 00:01:19.650 a random app. You are creating a localized intelligence 00:01:19.650 --> 00:01:23.790 right inside your main workspace. They sink perfectly 00:01:23.790 --> 00:01:26.409 across the ecosystem. That feels like a massive 00:01:26.409 --> 00:01:29.189 shift in how we actually work. Oh, it's huge. 00:01:29.269 --> 00:01:31.650 Because we used to constantly juggle different 00:01:31.650 --> 00:01:33.989 browser tabs for different tools. It's exhausting. 00:01:34.269 --> 00:01:36.530 Yeah, switching mental context every five minutes 00:01:36.530 --> 00:01:38.909 is draining. Now it acts as a unified brain, 00:01:39.430 --> 00:01:42.500 but creating that brain starts with... How you 00:01:42.500 --> 00:01:45.159 name it. Yes. Naming is critical. When you click 00:01:45.159 --> 00:01:47.180 to create a new notebook, just stop and think. 00:01:47.260 --> 00:01:49.680 Right. Most people type something vague, like 00:01:49.680 --> 00:01:52.939 marketing ideas. I do that. It is a total trap. 00:01:53.280 --> 00:01:55.939 You need specific naming conventions. Use something 00:01:55.939 --> 00:01:59.760 like Q2 content research hyphen AI agents. OK, 00:01:59.819 --> 00:02:03.480 very specific. Or maybe client brief hyphen Shopify 00:02:03.480 --> 00:02:07.420 app launch. The name actually grounds the AI's 00:02:07.420 --> 00:02:09.599 understanding of the entire environment. I have 00:02:09.599 --> 00:02:12.699 to admit, I still wrestle with naming files research 00:02:12.699 --> 00:02:15.199 final v2 myself. Oh, we've all been there. It 00:02:15.199 --> 00:02:17.199 tells you absolutely nothing three weeks later. 00:02:17.740 --> 00:02:19.780 It is like laying a foundation before building 00:02:19.780 --> 00:02:22.139 a house. With that intent, you're just stacking 00:02:22.139 --> 00:02:25.300 Lego blocks of data. Which brings us to the setup 00:02:25.300 --> 00:02:27.979 intent prompt. Right. What is that? Well, this 00:02:27.979 --> 00:02:30.879 happens before you add a single document. You 00:02:30.879 --> 00:02:33.400 paste a specific prompt declaring the project's 00:02:33.400 --> 00:02:36.479 goal. Before you even add sources. Yep. You define 00:02:36.479 --> 00:02:40.000 the primary audience. Let's say solo founders 00:02:40.000 --> 00:02:44.000 aged 25 to 40. Okay. And you define the final 00:02:44.000 --> 00:02:46.680 deliverable, like a podcast script or a strategy 00:02:46.680 --> 00:02:49.810 memo. So you were basically giving the AI its 00:02:49.810 --> 00:02:52.490 marching orders early. Yes, exactly. And based 00:02:52.490 --> 00:02:55.330 on that intent, the AI suggests targeted research 00:02:55.330 --> 00:02:58.110 angles. Oh, wow. Yeah, it provides specific search 00:02:58.110 --> 00:03:00.689 queries to guide your initial exploration. It 00:03:00.689 --> 00:03:03.090 even lists problematic source types to avoid. 00:03:03.330 --> 00:03:05.810 So it turns an empty room into a focused war 00:03:05.810 --> 00:03:07.469 room. Right. And then you start bringing in the 00:03:07.469 --> 00:03:11.349 material. You can add PDFs, URLs, or just paste 00:03:11.349 --> 00:03:14.689 text. But there is a crucial mechanical detail 00:03:14.689 --> 00:03:17.620 here regarding Google Docs, right? Yes, very 00:03:17.620 --> 00:03:20.960 crucial. When you import a document, Notebook 00:03:20.960 --> 00:03:24.719 LM makes a localized copy. It does. And you must 00:03:24.719 --> 00:03:27.759 remember that it is a static photo. Wait, what 00:03:27.759 --> 00:03:31.039 does that mean? If I import a Google Doc and 00:03:31.039 --> 00:03:33.300 my team updates it tomorrow... What happens? 00:03:33.419 --> 00:03:35.759 You mentioned it's a static photo. It means any 00:03:35.759 --> 00:03:38.280 later edits to the original Google Doc are completely 00:03:38.280 --> 00:03:40.599 ignored. Oh, really? Yeah, the notebook doesn't 00:03:40.599 --> 00:03:43.199 see those new changes at all. It perfectly captures 00:03:43.199 --> 00:03:46.599 that exact moment in time unless you manually 00:03:46.599 --> 00:03:48.919 re -import it. So it's a frozen photograph, not 00:03:48.919 --> 00:03:52.120 a live Google Doc. Exactly. Beat. Okay, with 00:03:52.120 --> 00:03:54.939 the foundation laid, we need actual material. 00:03:55.139 --> 00:03:56.659 Instead of carrying the bricks ourselves, we 00:03:56.659 --> 00:03:59.099 can send the AI out. Yeah, this completely flips 00:03:59.099 --> 00:04:01.759 the old research habit. How so? Well, we used 00:04:01.759 --> 00:04:03.900 to drop our own files in first, then we started 00:04:03.900 --> 00:04:07.740 asking questions. Now, you let Notebook LM find 00:04:07.740 --> 00:04:10.280 the initial baseline sources for you. That saves 00:04:10.280 --> 00:04:12.860 hours of manual searching. But you have to understand 00:04:12.860 --> 00:04:14.620 the two different tools available here. Right, 00:04:14.639 --> 00:04:16.740 there's fast research, and then there's deep 00:04:16.740 --> 00:04:20.670 research. And deep research is basically an autonomous 00:04:20.670 --> 00:04:24.129 bot that browses the web to build research reports? 00:04:24.329 --> 00:04:26.949 Yep. But fast research is your quick first pass. 00:04:27.410 --> 00:04:30.509 You type a query. It rapidly scans the web and 00:04:30.509 --> 00:04:33.889 your connected drive. OK. And it returns 10 suggested 00:04:33.889 --> 00:04:36.730 sources in seconds. Each comes with a short summary. 00:04:37.009 --> 00:04:39.089 The beautiful part is the built -in filtering 00:04:39.089 --> 00:04:42.750 mechanism. Every source has one sentence explaining 00:04:42.750 --> 00:04:45.269 why it fits your project. It's so efficient. 00:04:45.449 --> 00:04:47.290 You don't open every tab. You just read that 00:04:47.290 --> 00:04:50.870 one sentence and tick the boxes you want. But 00:04:50.870 --> 00:04:53.769 deep research operates on a heavier, more complex 00:04:53.769 --> 00:04:56.490 level. It's a bigger deal. Oh, it is a serious 00:04:56.490 --> 00:04:58.730 expedition. It browses hundreds of websites. 00:04:58.949 --> 00:05:02.290 It analyzes the text on each one. Wow. And then 00:05:02.290 --> 00:05:04.810 it writes a multi -page synthesis report with 00:05:04.810 --> 00:05:07.730 a full citation list. Whoa. Imagine an agent 00:05:07.730 --> 00:05:10.009 reading hundreds of sites and writing a multi 00:05:10.009 --> 00:05:12.110 -page report just to build your reading list. 00:05:12.350 --> 00:05:15.079 Two secs silence. That is staggering. It really 00:05:15.079 --> 00:05:16.939 is. It runs quietly in the background while you 00:05:16.939 --> 00:05:18.879 grab coffee. That's incredible. When it finishes, 00:05:19.019 --> 00:05:21.339 you get the full report. You also get a massive 00:05:21.339 --> 00:05:23.759 list of every source it reviewed. And you just 00:05:23.759 --> 00:05:25.439 select what you want. Right. You just tick the 00:05:25.439 --> 00:05:27.000 ones you actually want to import into your notebook. 00:05:27.240 --> 00:05:30.870 But the AI needs guardrails. By default, it provides 00:05:30.870 --> 00:05:33.149 safe, somewhat generic sources. You have to give 00:05:33.149 --> 00:05:35.350 it a strict prompt. Very strict. Demand sources 00:05:35.350 --> 00:05:38.089 published within the last 18 months. Ask for 00:05:38.089 --> 00:05:41.189 credible domain experts and operators. Demand 00:05:41.189 --> 00:05:44.569 raw numbers and firsthand case studies. And you 00:05:44.569 --> 00:05:46.790 also tell it exactly what to avoid. Right. You 00:05:46.790 --> 00:05:50.069 want pure signal, not useless noise. You ask 00:05:50.069 --> 00:05:52.730 for contrarian sources that challenge the mainstream 00:05:52.730 --> 00:05:57.029 consensus. And crucially, you explicitly ban 00:05:57.029 --> 00:06:00.949 SEO blogs. But why specifically tell the AI to 00:06:00.949 --> 00:06:04.000 avoid SEO blogs? Shouldn't it just... figure 00:06:04.000 --> 00:06:06.000 out what's relevant based on the topic? Well, 00:06:06.079 --> 00:06:08.139 the internet is flooded with affiliate content 00:06:08.139 --> 00:06:10.800 that just, you know, restates existing ideas. 00:06:10.920 --> 00:06:12.939 Oh, I see. If the AI doesn't have a negative 00:06:12.939 --> 00:06:15.819 constraint, it gets distracted by keyword stuffed 00:06:15.819 --> 00:06:18.620 articles. You have to force it to look for primary 00:06:18.620 --> 00:06:22.220 data. Specific constraints force the AI past 00:06:22.220 --> 00:06:25.220 superficial marketing content. Presumably. Beat. 00:06:25.959 --> 00:06:28.839 So now we have a pile of sources, but are they 00:06:28.839 --> 00:06:32.649 structurally sound? Messy sources equal blurry, 00:06:32.930 --> 00:06:36.069 confusing answers. It is an undeniable equation 00:06:36.069 --> 00:06:39.550 in language models. A notebook with 12 strong 00:06:39.550 --> 00:06:42.649 sources easily beats a notebook with 80 weak 00:06:42.649 --> 00:06:45.709 ones. Quality over quantity. Always. If you feed 00:06:45.709 --> 00:06:48.629 the model trash, it dilutes the gold. You must 00:06:48.629 --> 00:06:51.410 run a clean source review immediately. Three 00:06:51.410 --> 00:06:53.850 quiet minutes here will save hours of frustration 00:06:53.850 --> 00:06:56.430 later. Definitely. You start with a four -question 00:06:56.430 --> 00:06:59.410 mental checklist. Is the source recent enough? 00:06:59.529 --> 00:07:03.029 Is the author actually credible? Does it cover 00:07:03.029 --> 00:07:06.110 your specific angle deeply? And is the content 00:07:06.110 --> 00:07:08.769 substantial? If a source fails two of those questions, 00:07:08.850 --> 00:07:10.750 you skip it. You just need the data set to be 00:07:10.750 --> 00:07:13.129 relatively noise -free. But doing this manually 00:07:13.129 --> 00:07:15.550 for 30 documents takes too much time. Way too 00:07:15.550 --> 00:07:17.610 much. So we use the audit prompt. OK, walk us 00:07:17.610 --> 00:07:20.199 through that. You select all your imported sources. 00:07:20.639 --> 00:07:22.699 You paste a specific command directly into the 00:07:22.699 --> 00:07:25.519 chat box. You're basically asking the AI to grade 00:07:25.519 --> 00:07:28.040 your homework. I love that. You demand a one 00:07:28.040 --> 00:07:30.959 -line summary of every single source. You ask 00:07:30.959 --> 00:07:33.819 for a credibility score from one to five. You 00:07:33.819 --> 00:07:37.500 ask for a strict relevant score. Then you demand 00:07:37.500 --> 00:07:41.279 a firm recommendation. Keep, drop, or replace. 00:07:41.620 --> 00:07:43.899 It's like hiring a ruthless human editor who 00:07:43.899 --> 00:07:46.529 isn't afraid to hurt your feelings. Yes. It tells 00:07:46.529 --> 00:07:48.670 you which files pull their weight. It highlights 00:07:48.670 --> 00:07:51.370 the three weakest files you must absolutely delete. 00:07:51.610 --> 00:07:53.930 And the secret weapon in this prompt is four 00:07:53.930 --> 00:07:57.350 words. Do not be polite. Wait, do not be polite. 00:07:57.449 --> 00:07:59.889 That feels pretty aggressive. If I just ask for 00:07:59.889 --> 00:08:02.290 a standard audit, what exactly is it going to 00:08:02.290 --> 00:08:05.370 do wrong? Well, the AI naturally avoids conflict. 00:08:05.490 --> 00:08:08.069 It does. Oh, yeah. It wants to be helpful so 00:08:08.069 --> 00:08:10.149 it validates your choices. It will find a weak 00:08:10.149 --> 00:08:12.629 justification to keep an awful outdated source 00:08:12.629 --> 00:08:14.490 just because you uploaded it. Interesting. You 00:08:14.490 --> 00:08:16.290 really have to give it permission to be harsh. 00:08:16.649 --> 00:08:18.769 Without it, the AI defaults to people -pleasing 00:08:18.769 --> 00:08:22.329 and keeps junk. Exactly right. Beat. Alright, 00:08:22.550 --> 00:08:25.430 our sources are finally clean. Now we extract 00:08:25.430 --> 00:08:27.829 the knowledge without losing focus. This is the 00:08:27.829 --> 00:08:30.389 fun part. Most people just chat across all sources. 00:08:30.670 --> 00:08:33.240 Constantly. Yeah, and that is the wrong default 00:08:33.240 --> 00:08:35.940 setting. Why? Chatting across all sources is 00:08:35.940 --> 00:08:38.960 for broad synthesis. Use it when hunting for 00:08:38.960 --> 00:08:41.220 distinct contradictions across different authors. 00:08:41.879 --> 00:08:45.460 But reading 30 sources at once degrades the model's 00:08:45.460 --> 00:08:49.279 focus. For true deep focus, you use one source 00:08:49.279 --> 00:08:52.059 chat. Yes. You uncheck everything except the 00:08:52.059 --> 00:08:54.700 strongest report. You chat intimately with that 00:08:54.700 --> 00:08:57.820 single file. The answers come back fast and sharp 00:08:57.820 --> 00:09:00.179 because the model isn't blending 30 different 00:09:00.179 --> 00:09:02.460 angles. And this is where we use the single source 00:09:02.460 --> 00:09:05.320 deep read prompt. How does that work? You tell 00:09:05.320 --> 00:09:08.019 the AI to treat it as the only existing material. 00:09:08.320 --> 00:09:11.120 You ask for the core argument in under 25 words. 00:09:11.559 --> 00:09:14.059 You ask for five bullet points of verifiable 00:09:14.059 --> 00:09:17.169 evidence. specific numbers and dates. Then you 00:09:17.169 --> 00:09:19.009 ask for the hidden assumptions, right? Exactly. 00:09:19.350 --> 00:09:21.649 What does the author assume but never actually 00:09:21.649 --> 00:09:24.230 prove? Those hidden assumptions are usually the 00:09:24.230 --> 00:09:26.870 weakest spots in any argument. They are. Finally, 00:09:26.990 --> 00:09:28.889 you ask what the source is completely missing. 00:09:29.129 --> 00:09:31.429 And those missing pieces instantly become your 00:09:31.429 --> 00:09:35.000 next research queries. Yep. It does what manually 00:09:35.000 --> 00:09:37.320 reading a 30 -page report does, but it finishes 00:09:37.320 --> 00:09:40.379 in seconds. And that leads us to the hidden engine 00:09:40.379 --> 00:09:42.879 of this entire tool. The two -way note loop. 00:09:43.100 --> 00:09:45.139 It's a note loop. It's magic. Most people treat 00:09:45.139 --> 00:09:47.919 notes as simple read -only memory. You find a 00:09:47.919 --> 00:09:50.159 good quote, you save it, you forget it. That 00:09:50.159 --> 00:09:53.539 is a massive operational mistake. Notes are actually 00:09:53.539 --> 00:09:57.059 a continuous loop. When a chat answer lands perfectly, 00:09:57.240 --> 00:10:00.120 you save it to the notes panel. But a saved note 00:10:00.120 --> 00:10:02.139 can easily be converted back into a brand new 00:10:02.139 --> 00:10:05.210 source. Right. Notebook LM treats it exactly 00:10:05.210 --> 00:10:08.590 like a freshly imported PDF document. Let's visualize 00:10:08.590 --> 00:10:11.669 this. Say you are researching a complex market. 00:10:12.230 --> 00:10:15.590 Day one, you import a dense 40 page PDF. OK. 00:10:15.929 --> 00:10:19.009 You ask the AI to extract the five core arguments. 00:10:19.350 --> 00:10:21.909 It does. You save that clean summary as a note. 00:10:22.049 --> 00:10:24.269 Then you instantly convert that note into a new 00:10:24.269 --> 00:10:26.669 source. It's like distilling water. You take 00:10:26.669 --> 00:10:29.289 the raw source. boil it into a clean note, and 00:10:29.289 --> 00:10:31.250 feed it back into the system to get pure answers 00:10:31.250 --> 00:10:34.149 on the next run. That's a perfect analogy. And 00:10:34.149 --> 00:10:36.529 we use a synthesis prompt for this. You select 00:10:36.529 --> 00:10:38.929 all sources. You ask for an executive summary 00:10:38.929 --> 00:10:42.309 with 10 bullet points of key facts. You map out 00:10:42.309 --> 00:10:44.830 exactly where authors agree and disagree. You 00:10:44.830 --> 00:10:47.690 save that dense output as a permanent note, then 00:10:47.690 --> 00:10:49.889 convert it. I'm trying to wrap my head around 00:10:49.889 --> 00:10:52.639 this. Yeah. Why would someone turn a note back 00:10:52.639 --> 00:10:55.019 into a source instead of just keeping it as a 00:10:55.019 --> 00:10:57.519 reference? Because next time you ask a complex 00:10:57.519 --> 00:11:00.419 question, the AI doesn't have to scan 40 pages 00:11:00.419 --> 00:11:02.720 of jargon again. Oh, I get it. It just reads 00:11:02.720 --> 00:11:05.940 your crystal -clear, distilled note. It compounds 00:11:05.940 --> 00:11:08.600 your intelligence over time. It creates a pre 00:11:08.600 --> 00:11:11.720 -chewed, noise -free foundation for smarter future 00:11:11.720 --> 00:11:15.200 chats. Exactly. Beat. Now, a quick word from 00:11:15.200 --> 00:11:22.139 our sponsor. Welcome back. We have distilled 00:11:22.139 --> 00:11:25.360 the raw text beautifully. But some complex ideas 00:11:25.360 --> 00:11:27.720 simply need to be seen. They really do. Let's 00:11:27.720 --> 00:11:30.080 move to the Studio Panel. That's a visual dashboard 00:11:30.080 --> 00:11:32.679 for generating media outputs. This is where messy 00:11:32.679 --> 00:11:35.360 sources become shareable artifacts. The notebook 00:11:35.360 --> 00:11:37.679 transforms into visual diagrams and briefing 00:11:37.679 --> 00:11:40.159 documents. We start with mind maps. Right. You 00:11:40.159 --> 00:11:43.019 hit generate and a branching visual diagram appears. 00:11:43.120 --> 00:11:45.379 It shows the main themes in your entire notebook. 00:11:45.519 --> 00:11:47.399 It maps exactly how they all connect together. 00:11:47.700 --> 00:11:50.730 And every single branch is clickable. You tap 00:11:50.730 --> 00:11:53.429 a subtopic and it intuitively takes you straight 00:11:53.429 --> 00:11:55.909 to the underlying sources. I've tried using AI 00:11:55.909 --> 00:11:58.129 for mind maps before and it usually just spits 00:11:58.129 --> 00:12:01.570 out a useless tangled web of buzzwords. How does 00:12:01.570 --> 00:12:04.029 this actually orient me without just making more 00:12:04.029 --> 00:12:07.009 visual noise? Because it acts as a live interactive 00:12:07.009 --> 00:12:10.610 table of contents. When 20 sources feel like 00:12:10.610 --> 00:12:13.750 pure chaos, this provides a structured hierarchy. 00:12:14.110 --> 00:12:16.830 It isn't just a pretty picture. It is a fully 00:12:16.830 --> 00:12:19.820 navigable index of your own data set. Then we 00:12:19.820 --> 00:12:22.860 have infographics. Some concepts simply click 00:12:22.860 --> 00:12:25.480 faster visually. Definitely. You choose a strict 00:12:25.480 --> 00:12:28.340 visual style. Professional, editorial, or instructional. 00:12:28.539 --> 00:12:30.480 Instructional formatting works beautifully for 00:12:30.480 --> 00:12:32.639 step -by -step content. You also dictate the 00:12:32.639 --> 00:12:35.559 orientation. Vertical for mobile scrolling, horizontal 00:12:35.559 --> 00:12:38.240 for slide decks. You can even dictate the exact 00:12:38.240 --> 00:12:40.220 detail level of the graphic. What happens if 00:12:40.220 --> 00:12:42.679 I highlight just one specific report before generating 00:12:42.679 --> 00:12:45.320 an infographic? The system intelligently isolates 00:12:45.320 --> 00:12:47.759 that exact information. It ignores the rest of 00:12:47.759 --> 00:12:49.980 the notebook entirely, giving you a hyper -focused 00:12:49.980 --> 00:12:52.980 visual. It restricts the visual. strictly to 00:12:52.980 --> 00:12:56.159 that single file's data. Yes. Beat. That is incredibly 00:12:56.159 --> 00:12:59.740 useful. Yes. But briefing docs and FAQs are just 00:12:59.740 --> 00:13:02.779 internal tools. What happens when you need to 00:13:02.779 --> 00:13:05.139 hand this research over to a client who only 00:13:05.139 --> 00:13:07.559 has five minutes to understand it? That is where 00:13:07.559 --> 00:13:10.399 we leverage the text and media outputs. Studio 00:13:10.399 --> 00:13:12.940 effortlessly creates comprehensive study guides. 00:13:13.519 --> 00:13:16.500 It builds timelines for historical topics. It 00:13:16.500 --> 00:13:19.399 generates structured reports for formal deliverables. 00:13:19.779 --> 00:13:22.059 So you turn a heavy notebook into a client -ready 00:13:22.059 --> 00:13:24.460 document in under 10 minutes. Exactly. Moving 00:13:24.460 --> 00:13:26.820 from internal understanding to external sharing 00:13:26.820 --> 00:13:29.360 requires different formats. Yeah. That brings 00:13:29.360 --> 00:13:31.639 us to audio overviews. One of my favorite features. 00:13:31.840 --> 00:13:33.960 This feature creates a podcast -style conversation. 00:13:34.200 --> 00:13:36.679 Two AI hosts talk casually through your deep 00:13:36.679 --> 00:13:39.500 research. And the 2026 version pushed this much 00:13:39.500 --> 00:13:42.200 further. You tightly control format, length, 00:13:42.360 --> 00:13:44.480 tone, and language. There are four main audio 00:13:44.480 --> 00:13:47.600 formats, right? Yep. Deep dive is long and detailed. 00:13:48.019 --> 00:13:50.740 Brief is intentionally short. The debate format 00:13:50.740 --> 00:13:53.559 has hosts taking passionately opposing sides. 00:13:54.120 --> 00:13:56.919 It is brilliant for controversial topics. It 00:13:56.919 --> 00:13:59.700 really is. Finally, there is critique, which 00:13:59.700 --> 00:14:03.440 is a highly analytical mode. The hosts aggressively 00:14:03.440 --> 00:14:05.820 stress test the presented ideas. I see. So you 00:14:05.820 --> 00:14:08.860 might generate a brief first to catch the big 00:14:08.860 --> 00:14:11.559 picture. Yeah. Then run a debate to clearly see 00:14:11.559 --> 00:14:14.950 both sides. It is the exact same notebook producing 00:14:14.950 --> 00:14:17.350 two entirely different listening experiences. 00:14:17.629 --> 00:14:19.789 But the interactive mode is the real magic here. 00:14:19.870 --> 00:14:21.970 Oh, it changes everything. While the audio plays, 00:14:22.269 --> 00:14:24.669 you click a button, the hosts immediately stop 00:14:24.669 --> 00:14:27.690 speaking. You politely ask a question, they respond 00:14:27.690 --> 00:14:29.909 seamlessly, then pick up right where they left 00:14:29.909 --> 00:14:32.970 off. It completely bridges the gap between passive 00:14:32.970 --> 00:14:35.509 listening and having an active tutor. But you 00:14:35.509 --> 00:14:38.659 must use a strict prompt. Without it, the audio 00:14:38.659 --> 00:14:41.879 sounds friendly but painfully generic. Yes, you 00:14:41.879 --> 00:14:43.799 tell it to treat the listener as intelligent 00:14:43.799 --> 00:14:46.120 but busy. State the main question clearly in 00:14:46.120 --> 00:14:48.779 the first 30 seconds. Cover exactly three vital 00:14:48.779 --> 00:14:52.259 ideas. Use detailed citations verbally. If sources 00:14:52.259 --> 00:14:54.860 firmly disagree, say so plainly out loud. And 00:14:54.860 --> 00:14:57.720 you must explicitly ban repetitive filler phrases. 00:14:58.019 --> 00:15:00.820 No overly excited reactions. Just substantive 00:15:00.820 --> 00:15:04.120 conversational analysis. Why do we have to explicitly 00:15:04.120 --> 00:15:07.620 tell the AI to avoid filler like? That is fascinating. 00:15:08.240 --> 00:15:10.360 Doesn't it know we want serious analysis? Well, 00:15:10.360 --> 00:15:13.360 it trained on millions of real world podcasts. 00:15:13.600 --> 00:15:16.059 Oh, right. So it absorbed all those common hosting 00:15:16.059 --> 00:15:18.659 tropes. If you don't suppress that default behavior, 00:15:18.879 --> 00:15:21.399 it wastes processing power on simulated personality 00:15:21.399 --> 00:15:23.899 instead of analytical density. It strips away 00:15:23.899 --> 00:15:27.360 fake podcast bro energy for serious dense learning. 00:15:27.500 --> 00:15:30.039 Nailed it. Beat. And when plain text and simple 00:15:30.039 --> 00:15:33.299 audio fall short, we use video overviews, numbers 00:15:33.299 --> 00:15:36.820 cleanly laid out side by side, timelines, complex 00:15:36.820 --> 00:15:40.019 diagrams. Video overview has three distinct styles. 00:15:40.120 --> 00:15:42.279 Okay, what are they? Explainer is the standard 00:15:42.279 --> 00:15:44.730 teaching format. Brief is significantly shorter, 00:15:45.190 --> 00:15:47.470 and Cinematic gracefully offers richer visuals 00:15:47.470 --> 00:15:49.690 and smooth pacing. The explainer is for step 00:15:49.690 --> 00:15:52.190 -by -step teaching. Brief is for essential takeaways. 00:15:52.830 --> 00:15:54.970 Cinematic is for client presentations. Exactly. 00:15:55.389 --> 00:15:58.549 But a weak prompt gives you generic, motivational, 00:15:58.669 --> 00:16:00.909 LinkedIn -style slides. Right. Nobody wants that. 00:16:01.169 --> 00:16:03.710 You need a strict video prompt structure, a clear 00:16:03.710 --> 00:16:06.350 hook slide firmly stating the stakes, context 00:16:06.350 --> 00:16:09.230 slides properly defining the core concept. Core 00:16:09.230 --> 00:16:12.080 content. cleanly broken into sequential steps. 00:16:12.620 --> 00:16:15.240 A sharp contrast slide, properly showing counter 00:16:15.240 --> 00:16:18.059 arguments. And a final slide with an action step. 00:16:18.480 --> 00:16:21.879 You enforce strict visual rules. High contrast. 00:16:22.340 --> 00:16:25.279 Readable on mobile. Detail diagrams favored over 00:16:25.279 --> 00:16:28.279 useless decoration. Every single claim must be 00:16:28.279 --> 00:16:30.980 traceable to a source. And you explicitly ban 00:16:30.980 --> 00:16:34.200 stock photo cliches completely. No shaking hands, 00:16:34.440 --> 00:16:36.960 no glowing light bulbs. Good. We have covered 00:16:36.960 --> 00:16:39.620 a massive amount of ground today. Opening a fresh 00:16:39.620 --> 00:16:42.360 notebook inside Gemini. Running autonomous agents. 00:16:42.580 --> 00:16:46.000 Auditing sources. Mastering the note loop. Generating 00:16:46.000 --> 00:16:48.259 media overviews. On paper it looks like a mountain 00:16:48.259 --> 00:16:51.299 of steps. But it is really just one elegantly 00:16:51.299 --> 00:16:54.039 continuous workflow. It is. Notebook LM offers 00:16:54.039 --> 00:16:56.799 the exact same tools to everyone. What separates 00:16:56.799 --> 00:16:59.419 a useless chatbot from a true second brain is 00:16:59.419 --> 00:17:01.500 how you string those tools together. Clean sources, 00:17:01.799 --> 00:17:03.960 clear notes, and highly specific prompts. That 00:17:03.960 --> 00:17:06.549 is the entire secret. But please do not try to 00:17:06.549 --> 00:17:08.269 apply all these tips today. You will rapidly 00:17:08.269 --> 00:17:10.710 overload yourself. Just pick exactly three things 00:17:10.710 --> 00:17:13.490 to start with. Let the deep research agent find 00:17:13.490 --> 00:17:16.390 your initial sources. Run your first clean source 00:17:16.390 --> 00:17:19.109 audit and try using the note to source loop just 00:17:19.109 --> 00:17:21.670 once. Open a real project you actually care about 00:17:21.670 --> 00:17:24.650 today and feel how the tool transforms your data. 00:17:24.990 --> 00:17:27.809 Think about that continuous note loop we deconstructed. 00:17:28.029 --> 00:17:30.150 Imagine doing that for six months on a single 00:17:30.150 --> 00:17:32.720 specific topic. What happens when a full team 00:17:32.720 --> 00:17:35.519 shares a single notebook where the AI continuously 00:17:35.519 --> 00:17:38.259 synthesizes years of collective knowledge? It 00:17:38.259 --> 00:17:40.059 stops being just a personal productivity tool 00:17:40.059 --> 00:17:43.579 and quietly becomes a synthetic colleague. Something 00:17:43.579 --> 00:17:46.619 to think about, Pete. Thanks for joining us on 00:17:46.619 --> 00:17:47.279 this deep dive.