WEBVTT 00:00:00.000 --> 00:00:02.580 We all know that feeling, don't we? You sit down 00:00:02.580 --> 00:00:05.080 at your desk and you're just, you're immediately 00:00:05.080 --> 00:00:08.140 faced with this modern dilemma. It's just pure 00:00:08.140 --> 00:00:10.980 overwhelming information overload. It really 00:00:10.980 --> 00:00:13.060 is the digital equivalent of drowning. Right. 00:00:13.220 --> 00:00:15.679 You've got what, 12 PDFs open for a project, 00:00:16.120 --> 00:00:18.460 three industry reports waiting. Oh yeah. Maybe 00:00:18.460 --> 00:00:21.039 a handful of technical articles you saved. You 00:00:21.039 --> 00:00:23.460 spend hours skimming and your brain just feels 00:00:23.460 --> 00:00:26.359 full. Yeah. But you're not actually informed. 00:00:26.559 --> 00:00:29.280 You're just exhausted. drowning in data. It's 00:00:29.280 --> 00:00:31.780 the core challenge. But what if you could just 00:00:31.780 --> 00:00:34.359 bypass all that? What if you had like a personal 00:00:34.359 --> 00:00:36.520 assistant that could read all those boring files 00:00:36.520 --> 00:00:39.560 in seconds and then just explain it to you? Well, 00:00:39.780 --> 00:00:42.859 that's the promise of Notebook LM. And that's 00:00:42.859 --> 00:00:45.219 our deep dive today. We're not just, you know, 00:00:45.320 --> 00:00:47.219 looking at a piece of software. We're exploring 00:00:47.219 --> 00:00:49.759 a whole new way of managing knowledge. And it's 00:00:49.759 --> 00:00:52.359 so important that we start right there by understanding 00:00:52.359 --> 00:00:54.520 why this is so fundamentally different from, 00:00:54.520 --> 00:00:57.280 say, a standard chat bot. This isn't about searching 00:00:57.280 --> 00:01:00.439 the entire Internet. Exactly. Our mission here 00:01:00.439 --> 00:01:02.920 is to walk you through these essential concepts 00:01:02.920 --> 00:01:05.859 like grounding and the practical steps to turn 00:01:05.859 --> 00:01:08.439 this into your own custom research brain. We'll 00:01:08.439 --> 00:01:11.519 cover the setup. the layout, and the critical 00:01:11.519 --> 00:01:14.280 skill of feeding the AI what it needs. We call 00:01:14.280 --> 00:01:16.640 that context engineering. OK, let's unpack that. 00:01:16.840 --> 00:01:19.579 Why does standard AI so often fail us when we 00:01:19.579 --> 00:01:23.120 need precise facts? Well, a standard AI is searching 00:01:23.120 --> 00:01:26.260 the entire world of information. It's incredibly 00:01:26.260 --> 00:01:28.879 powerful. But when you ask a nuanced question, 00:01:29.319 --> 00:01:31.900 it pulls from everything. That means outdated 00:01:31.900 --> 00:01:35.099 material, irrelevant forum posts. Or stuff that's 00:01:35.099 --> 00:01:37.500 just flat out wrong. Precisely. And that's what 00:01:37.500 --> 00:01:39.540 leads to hallucination. Right, the term for when 00:01:39.540 --> 00:01:42.459 an AI just makes things up. Yeah, it just invents 00:01:42.459 --> 00:01:45.319 an answer to fill the gap. And it sounds completely 00:01:45.319 --> 00:01:47.219 confident while doing it. And that is where this 00:01:47.219 --> 00:01:49.900 idea of grounding comes in. Grounding is the 00:01:49.900 --> 00:01:52.859 guardrail. Think of it like this. Imagine you're 00:01:52.859 --> 00:01:54.579 taking the most important test of your life. 00:01:54.579 --> 00:01:57.200 OK. And the examiner says you can only use one 00:01:57.200 --> 00:02:01.140 single authorized textbook. No phone, no friends, 00:02:01.540 --> 00:02:03.620 no searching the web. I love that analogy. It 00:02:03.620 --> 00:02:06.849 defines the boundary perfectly. It does. Grounding 00:02:06.849 --> 00:02:11.110 means the AI puts on literal blinkers. It only 00:02:11.110 --> 00:02:13.909 uses facts from the documents you provided. You 00:02:13.909 --> 00:02:16.430 can't guess or pull in outside information. Its 00:02:16.430 --> 00:02:18.550 whole world is what you uploaded. That constraint 00:02:18.550 --> 00:02:21.469 is actually its superpower. It is. It forces 00:02:21.469 --> 00:02:24.509 factual accountability. The AI will even cite 00:02:24.509 --> 00:02:26.990 where it found the answer in your files, and 00:02:26.990 --> 00:02:29.430 this is where you step into a new role. The context 00:02:29.430 --> 00:02:31.680 engineer. That's the one. It sounds like a really 00:02:31.680 --> 00:02:33.719 fancy job title, but it just means you're the 00:02:33.719 --> 00:02:36.639 boss of the AI's brain. You are. And if you want 00:02:36.639 --> 00:02:38.840 a great output, you have to pick the right ingredients. 00:02:38.960 --> 00:02:41.879 It's like cooking, right? You put garbage ingredients, 00:02:42.360 --> 00:02:45.120 old files, irrelevant notes into the pot, you're 00:02:45.120 --> 00:02:47.379 going to get garbage soup. The quality of your 00:02:47.379 --> 00:02:49.840 files dictates the quality of the insight. You 00:02:49.840 --> 00:02:52.240 control the context. So if you upload a book 00:02:52.240 --> 00:02:54.919 on deep sea biology, it becomes a marine biologist 00:02:54.919 --> 00:02:57.449 trained only on that book. And that tight control, 00:02:57.469 --> 00:03:00.289 that focus, is what makes the answer so useful 00:03:00.289 --> 00:03:02.889 for your specific needs, instead of just general 00:03:02.889 --> 00:03:05.389 trivia. I have to admit, and this may be a little 00:03:05.389 --> 00:03:07.849 vulnerable, but I still wrestle with prompt drift 00:03:07.849 --> 00:03:10.990 myself. You know, when a general AI just pulls 00:03:10.990 --> 00:03:14.530 from too many places at once, that forced grounding, 00:03:14.590 --> 00:03:16.750 knowing it can only look at my three specific 00:03:16.750 --> 00:03:20.669 files, it's just incredibly helpful. It forces 00:03:20.669 --> 00:03:23.889 you to be strategic. So if I am the context engineer, 00:03:24.509 --> 00:03:27.830 What's the core danger of bad ingredients in 00:03:27.830 --> 00:03:30.330 my knowledge pot? The danger isn't that the AI 00:03:30.330 --> 00:03:33.409 will fail. It's that it will process your flawed 00:03:33.409 --> 00:03:36.669 input perfectly. Leading to perfectly useless, 00:03:36.750 --> 00:03:39.110 irrelevant output. Exactly. OK, that makes sense. 00:03:39.569 --> 00:03:42.449 Let's talk practicality. For a new learner, how 00:03:42.449 --> 00:03:44.509 easy is it to get started? Remarkably seamless. 00:03:44.830 --> 00:03:46.969 No payment, no download. You just sign in with 00:03:46.969 --> 00:03:49.349 your normal Google account. Once you're in, you 00:03:49.349 --> 00:03:51.379 create what's called a notebook. And we recommend 00:03:51.379 --> 00:03:53.620 keeping that name super specific, right? Not 00:03:53.620 --> 00:03:56.879 Project One, but maybe Q4 marketing study. Absolutely. 00:03:57.180 --> 00:03:59.479 Now, here's where most beginners make their first 00:03:59.479 --> 00:04:01.740 mistake. As soon as you create that notebook, 00:04:02.080 --> 00:04:05.560 a big pop -up appears. It asks you to add sources. 00:04:05.699 --> 00:04:08.900 The pop -up panic. That immediate urge to just 00:04:08.900 --> 00:04:11.819 start dumping every file you own into it. Resist 00:04:11.819 --> 00:04:14.539 it. This is key. Hit the X button, hit escape, 00:04:14.620 --> 00:04:16.920 just close that pop -up. You need to understand 00:04:16.920 --> 00:04:18.980 the empty room before you start bringing in all 00:04:18.980 --> 00:04:21.519 the furniture. So once that's gone, the screen 00:04:21.519 --> 00:04:24.160 is pretty clean. It's divided into three crucial 00:04:24.160 --> 00:04:26.439 sections. Right. On the far left, you have your 00:04:26.439 --> 00:04:28.639 source panel. Think of it as your bookshelf. 00:04:28.920 --> 00:04:31.800 This is where every file you upload lives. And 00:04:31.800 --> 00:04:33.839 the most important feature on that bookshelf 00:04:33.839 --> 00:04:36.899 is the little checkbox next to each file. That 00:04:36.899 --> 00:04:39.339 checkbox is the most powerful thing in the entire 00:04:39.339 --> 00:04:43.240 system. If the box is checked, the AI is actively 00:04:43.240 --> 00:04:45.879 reading that document. If you uncheck it, the 00:04:45.879 --> 00:04:48.610 AI completely ignores it. That's the surgical 00:04:48.610 --> 00:04:50.449 control you mentioned earlier. You could have 00:04:50.449 --> 00:04:52.550 10 sources uploaded, but you only check the two 00:04:52.550 --> 00:04:54.310 that are relevant to your question right now. 00:04:54.529 --> 00:04:57.129 Exactly. It eliminates all the noise. Then the 00:04:57.129 --> 00:04:59.170 middle section is just the chat room. That's 00:04:59.170 --> 00:05:00.889 your workspace where you type your questions 00:05:00.889 --> 00:05:03.490 like, summarize this chapter. And the answers 00:05:03.490 --> 00:05:06.089 appear right there. And then on the right, often 00:05:06.089 --> 00:05:09.509 hidden, is the notebook guide or the studio area. 00:05:09.790 --> 00:05:11.889 That's where special features and saved work 00:05:11.889 --> 00:05:15.529 accumulate. So to be crystal clear, that checkbox 00:05:15.529 --> 00:05:19.410 in the source panel. That's really the key to 00:05:19.410 --> 00:05:21.470 giving me surgical control over what the AI is 00:05:21.470 --> 00:05:24.069 reading. Without a doubt, it controls the AI's 00:05:24.069 --> 00:05:26.870 current focus. Okay, now that we know the layout, 00:05:26.949 --> 00:05:29.250 let's talk about feeding the brain. What kind 00:05:29.250 --> 00:05:32.569 of sources can we actually upload? It's incredibly 00:05:32.569 --> 00:05:35.470 flexible. You can use standard PDFs, text files. 00:05:36.009 --> 00:05:37.930 It connects directly to your Google Docs and 00:05:37.930 --> 00:05:40.430 slides. You can also just paste website links. 00:05:40.680 --> 00:05:42.620 And the one that really surprises people is YouTube. 00:05:42.860 --> 00:05:45.319 It's a total game changer. You just paste a video 00:05:45.319 --> 00:05:48.500 link, and the AI accesses the entire video transcript. 00:05:48.600 --> 00:05:51.000 It doesn't watch the video, it reads the words. 00:05:51.100 --> 00:05:53.259 Which means a three -hour lecture becomes a 10 00:05:53.259 --> 00:05:55.100 -page reference source you can search instantly. 00:05:55.259 --> 00:05:57.899 It's amazing. But this leads us to the single 00:05:57.899 --> 00:06:00.660 biggest mistake that beginners make. They trust 00:06:00.660 --> 00:06:03.120 the chat history. The chat history problem. This 00:06:03.120 --> 00:06:06.240 is critical. The AI does not, by default, remember 00:06:06.240 --> 00:06:08.300 your conversation. All right. So if you ask it 00:06:08.300 --> 00:06:10.180 for the five main ideas in a book, and it gives 00:06:10.180 --> 00:06:13.259 you this brilliant, perfect list, if you navigate 00:06:13.259 --> 00:06:15.720 away and come back, that list is gone from its 00:06:15.720 --> 00:06:18.459 memory. It only remembers the uploaded source 00:06:18.459 --> 00:06:21.439 files, not the conversation about them. And that's 00:06:21.439 --> 00:06:23.920 so frustrating for people starting out. The chat 00:06:23.920 --> 00:06:26.800 is temporary workspace. The sources panel is 00:06:26.800 --> 00:06:29.319 permanent memory. So the fix for this is basically 00:06:29.319 --> 00:06:32.160 mandatory for anything useful the AI creates. 00:06:32.439 --> 00:06:35.019 It is. When the AI gives you a great answer, 00:06:35.220 --> 00:06:37.720 that perfect summary, you have to find the little 00:06:37.720 --> 00:06:40.660 pin icon or the save to note button, click it. 00:06:40.980 --> 00:06:43.240 That saves the answer over to the studio area 00:06:43.240 --> 00:06:46.490 on the right. But it's not permanent memory for 00:06:46.490 --> 00:06:49.930 the AI, not yet. Exactly. The final critical 00:06:49.930 --> 00:06:53.149 step is this. You select that saved note, you 00:06:53.149 --> 00:06:55.389 click the three dots, and you select convert 00:06:55.389 --> 00:06:58.149 to source. And that one action transforms the 00:06:58.149 --> 00:07:01.209 AI's temporary answer into a new permanent source 00:07:01.209 --> 00:07:03.449 document. It's like adding a new page to the 00:07:03.449 --> 00:07:05.689 textbook. You're actively curating the knowledge 00:07:05.689 --> 00:07:08.370 base itself. It's no longer just a conversation, 00:07:08.829 --> 00:07:11.540 it's engineered insight. So if I get a fantastic 00:07:11.540 --> 00:07:13.720 answer in the chat, how do I make sure it becomes 00:07:13.720 --> 00:07:16.040 part of the AI's permanent knowledge base? You 00:07:16.040 --> 00:07:18.819 have to save the chat answer as a note, and then 00:07:18.819 --> 00:07:21.300 explicitly convert that note into a new source. 00:07:21.699 --> 00:07:24.220 OK, let's shift gears. What about the scenario 00:07:24.220 --> 00:07:26.519 where you don't have files yet? You're starting 00:07:26.519 --> 00:07:28.959 a big project from scratch, say, opening a coffee 00:07:28.959 --> 00:07:31.620 shop. This is where deep research comes in. Right. 00:07:31.660 --> 00:07:34.779 When you go to add source from the web, you see 00:07:34.779 --> 00:07:38.060 two options. One is quick search. It's for simple 00:07:38.060 --> 00:07:41.879 facts like capital of Australia. It's fast, but 00:07:41.879 --> 00:07:43.920 shallow. And then there's deep research, which 00:07:43.920 --> 00:07:47.040 I think is the real game changer here. This requires 00:07:47.040 --> 00:07:49.399 a bit more thought. You give it a detailed prompt, 00:07:49.560 --> 00:07:52.060 a mission, and it takes about five to ten minutes. 00:07:52.319 --> 00:07:55.120 But instead of just grazing the top results, 00:07:55.680 --> 00:07:58.019 Deep Research looks at hundreds of websites. 00:07:58.199 --> 00:08:00.500 It performs a sort of algorithmic triage, right? 00:08:00.500 --> 00:08:03.980 It filters out the spam, the junk, the ads. Yes, 00:08:03.980 --> 00:08:06.920 it's filtering and synthesizing. It compiles 00:08:06.920 --> 00:08:09.620 the best, most relevant info it finds into one 00:08:09.620 --> 00:08:12.480 comprehensive research report source made just 00:08:12.480 --> 00:08:14.839 for you. So instead of wading through 40 Google 00:08:14.839 --> 00:08:17.180 results, you give the AI a mission. Something 00:08:17.180 --> 00:08:19.759 like, find detailed guides on opening a specialty 00:08:19.759 --> 00:08:22.759 coffee shop in 2025. Look for equipment costs, 00:08:23.120 --> 00:08:25.579 profit margins, and find real case studies. And 00:08:25.579 --> 00:08:28.759 that specificity is key. You hit execute, and 00:08:28.759 --> 00:08:31.180 you can just step away for a few minutes. When 00:08:31.180 --> 00:08:34.460 you come back, the AIs build you a high quality, 00:08:34.740 --> 00:08:37.980 pre -filtered source document. Whoa. That means 00:08:37.980 --> 00:08:40.460 10 hours of reading and filtering is basically 00:08:40.460 --> 00:08:42.860 done in 10 minutes. It's an incredible time saver. 00:08:43.039 --> 00:08:45.120 Now you have a grounded quality source you can 00:08:45.120 --> 00:08:47.639 ask questions to instead of swimming in web noise. 00:08:47.870 --> 00:08:50.230 So what's the biggest advantage of deep research 00:08:50.230 --> 00:08:52.629 over a traditional web search for gathering that 00:08:52.629 --> 00:08:55.529 foundational knowledge? It actively filters the 00:08:55.529 --> 00:08:58.330 junk and synthesizes hundreds of sources into 00:08:58.330 --> 00:09:02.330 one single coherent report. That focus on quality 00:09:02.330 --> 00:09:05.289 leads us to our final tip for successful context 00:09:05.289 --> 00:09:08.149 engineering, the chunking method. So even though 00:09:08.149 --> 00:09:10.929 the AI can read a 500 page book you upload, what's 00:09:10.929 --> 00:09:13.649 the problem with doing that? It leads to what 00:09:13.649 --> 00:09:16.330 you could call AI fatigue. You get these bad, 00:09:16.570 --> 00:09:18.710 vague summaries because the scope is just too 00:09:18.710 --> 00:09:21.070 wide. It's like asking a student to summarize 00:09:21.070 --> 00:09:22.769 the history of the world. You'd get a terrible 00:09:22.769 --> 00:09:25.330 answer. But if you ask them to summarize chapter 00:09:25.330 --> 00:09:28.350 one, the Stone Age, you get detail and focus. 00:09:28.769 --> 00:09:31.149 Exactly. You want surgical precision. The solution 00:09:31.149 --> 00:09:34.429 is to use a free tool, a split PDF utility, and 00:09:34.429 --> 00:09:36.850 just cut your big report into smaller chunks. 00:09:37.629 --> 00:09:39.850 Upload chapter one, chapter two, chapter three 00:09:39.850 --> 00:09:42.350 as separate files. And the power of that is when 00:09:42.350 --> 00:09:44.169 you go to your bookshelf, you can check only 00:09:44.169 --> 00:09:47.610 chapter one, and then ask the AI, create a glossary 00:09:47.610 --> 00:09:49.970 of difficult words from this specific chapter. 00:09:50.590 --> 00:09:53.210 And the AI is focused entirely on that one chunk. 00:09:53.690 --> 00:09:56.190 It's not distracted by the noise from the rest 00:09:56.190 --> 00:09:58.610 of the book. It gives you a hyper -focused, usable 00:09:58.610 --> 00:10:01.330 answer. But what if I can't easily split a file? 00:10:01.669 --> 00:10:03.830 Well... You can still be tactical. If you have 00:10:03.830 --> 00:10:06.529 to upload a huge file, you can tell the AI in 00:10:06.529 --> 00:10:09.610 your prompt, focus only on pages 300 to 350 for 00:10:09.610 --> 00:10:11.950 this summary. Kind of force the chunking with 00:10:11.950 --> 00:10:14.110 your words, but splitting the files is always 00:10:14.110 --> 00:10:16.690 better. So why does chunking improve the quality 00:10:16.690 --> 00:10:19.110 of the summaries, even if the AI can technically 00:10:19.110 --> 00:10:21.409 process the whole file? Because splitting the 00:10:21.409 --> 00:10:24.370 files allows the AI to focus precisely. It delivers 00:10:24.370 --> 00:10:26.870 detailed surgical summaries instead of vague 00:10:26.870 --> 00:10:29.289 generalized ones. This has been a huge amount 00:10:29.289 --> 00:10:32.730 of foundational knowledge, but so valuable. We've 00:10:32.730 --> 00:10:35.990 moved from just drowning in browser tabs to really 00:10:35.990 --> 00:10:38.289 understanding the core concepts of knowledge 00:10:38.289 --> 00:10:40.090 engineering. Yeah, we established the difference 00:10:40.090 --> 00:10:43.909 between general AI and grounding that vital boundary. 00:10:44.029 --> 00:10:46.330 Yeah. We learned that you are the context engineer. 00:10:46.769 --> 00:10:49.610 And crucially, we fixed the memory problem by 00:10:49.610 --> 00:10:51.889 knowing how to convert chat answers to permanent 00:10:51.889 --> 00:10:54.429 sources. This foundational work really moves 00:10:54.429 --> 00:10:56.929 you out of just passively consuming information. 00:10:57.120 --> 00:10:59.299 You are now actively engineering your knowledge 00:10:59.299 --> 00:11:02.419 base for control. Definitely. And next time, 00:11:02.779 --> 00:11:05.399 we get into the really fun part. We'll be turning 00:11:05.399 --> 00:11:08.159 boring text into listenable audio, making mind 00:11:08.159 --> 00:11:10.759 maps, and using advanced prompts to learn complex 00:11:10.759 --> 00:11:13.409 topics super fast. But for now, here's a final 00:11:13.409 --> 00:11:15.350 thought for you to take into your week. How much 00:11:15.350 --> 00:11:18.070 time in actual hours could you truly save next 00:11:18.070 --> 00:11:20.769 week if you just forced the AI to only read the 00:11:20.769 --> 00:11:22.669 three most important reports sitting in your 00:11:22.669 --> 00:11:25.409 inbox right now? Take back control of your context. 00:11:25.769 --> 00:11:27.590 Thank you for joining us for this deep dive. 00:11:27.789 --> 00:11:28.830 Go explore your sources.