WEBVTT 00:00:00.000 --> 00:00:02.000 You were just saying how incredibly frustrating 00:00:02.000 --> 00:00:05.419 this is before we hit record. And it's honestly 00:00:05.419 --> 00:00:08.539 the perfect place to start. Oh, absolutely. Because, 00:00:09.160 --> 00:00:12.140 okay, picture this. You're prepping for a massive 00:00:12.140 --> 00:00:17.289 meeting, right? and you upload this huge hundred 00:00:17.289 --> 00:00:20.289 page financial document to an AI. Right, because 00:00:20.289 --> 00:00:22.289 nobody has time to read the whole thing. Exactly. 00:00:22.350 --> 00:00:24.809 So you ask it a crucial high stakes question. 00:00:25.269 --> 00:00:27.510 It spits back this brilliant sounding answer. 00:00:27.530 --> 00:00:29.870 Yeah. You feel like a genius until you actually 00:00:29.870 --> 00:00:32.469 double check the math and you realize the AI 00:00:32.469 --> 00:00:35.009 completely fabricated the core revenue number. 00:00:35.070 --> 00:00:37.689 Oh, man. It just invented it out of thin air. 00:00:37.689 --> 00:00:40.990 It's maddening. Yeah, it really is. Beat. Welcome 00:00:40.990 --> 00:00:43.409 to the deep dive. I'll admit something right 00:00:43.409 --> 00:00:45.649 up front here. I still wrestle with prompt drift 00:00:45.649 --> 00:00:47.630 myself. Oh, we all do. You trust the system, 00:00:47.710 --> 00:00:50.390 you get comfortable with it, and then it suddenly 00:00:50.390 --> 00:00:52.390 hallucinates on your own carefully collected 00:00:52.390 --> 00:00:54.969 data. It's a very modern kind of betrayal. That 00:00:54.969 --> 00:00:57.229 is the perfect word for it. Betrayal. Two -sex 00:00:57.229 --> 00:01:00.869 silence. Today we are exploring a real structural 00:01:00.869 --> 00:01:03.549 solution to this. We're breaking down the April 00:01:03.549 --> 00:01:08.040 2026 update. This is the update that finally 00:01:08.040 --> 00:01:11.299 combined Gemini and Notebook LM into a single, 00:01:11.599 --> 00:01:14.159 unbreakable workflow. Yeah, it changed everything. 00:01:14.299 --> 00:01:16.280 We're going to explore how to build a genuinely 00:01:16.280 --> 00:01:18.680 trusted knowledge base. We'll look at how to 00:01:18.680 --> 00:01:21.439 pull in outside research without corrupting your 00:01:21.439 --> 00:01:23.620 data. Which is key. And we'll see how to automate 00:01:23.620 --> 00:01:26.659 the entire loop. But let's unpack the divide 00:01:26.659 --> 00:01:29.379 first. Let's do it. Because Gemini and Notebook 00:01:29.379 --> 00:01:32.079 LM, they were really designed to solve two entirely 00:01:32.079 --> 00:01:34.540 different problems. Right. And understanding 00:01:34.540 --> 00:01:37.340 that difference is exactly why pairing them works 00:01:37.340 --> 00:01:39.640 so well. You really have to understand both halves 00:01:39.640 --> 00:01:42.120 of the brain basically before you wire them together. 00:01:42.379 --> 00:01:44.379 That makes sense. We need to look at their core 00:01:44.379 --> 00:01:47.280 mechanics. So let's start with Gemini. OK. So 00:01:47.280 --> 00:01:50.719 Gemini is your window to the world. It has live 00:01:50.719 --> 00:01:55.090 web access. It pulls in breaking news that didn't 00:01:55.090 --> 00:01:57.769 even exist during its original training, and 00:01:57.769 --> 00:02:00.329 it also reaches directly into your Google workspace. 00:02:00.549 --> 00:02:03.030 Which saves a ton of time. Oh, tons. It reads 00:02:03.030 --> 00:02:06.549 from your docs, your drive, Gmail, Sheets. You 00:02:06.549 --> 00:02:08.310 don't have to copy and paste anything. Right. 00:02:08.610 --> 00:02:12.389 And it has long -term memory, meaning it gradually 00:02:12.389 --> 00:02:15.409 learns your specific industry jargon over time. 00:02:16.270 --> 00:02:18.889 But it has a structural flaw, right? Especially 00:02:18.889 --> 00:02:21.110 when it comes to precision work. Exactly. It 00:02:21.110 --> 00:02:23.330 operates from general knowledge first. So if 00:02:23.330 --> 00:02:25.830 you ask it a highly specific question about your 00:02:25.830 --> 00:02:29.030 uploaded documents, it answers with absolute 00:02:29.030 --> 00:02:33.129 unwavering confidence. But it drifts. Yes. A 00:02:33.129 --> 00:02:35.949 wrong decimal slips in. A mixed update suddenly 00:02:35.949 --> 00:02:38.210 appears. Because it's trying to be helpful. Right. 00:02:38.310 --> 00:02:40.669 Because of how large language models work. It 00:02:40.669 --> 00:02:42.689 fills any gaps in your document with its broad 00:02:42.689 --> 00:02:47.120 training data. So it's like a brilliant but highly 00:02:47.120 --> 00:02:50.020 distractible researcher. Ooh, I like that. They're 00:02:50.020 --> 00:02:51.800 sitting there with a smartphone connected to 00:02:51.800 --> 00:02:53.740 the entire internet. They're incredibly smart. 00:02:54.419 --> 00:02:56.719 But they inevitably lose focus on the physical 00:02:56.719 --> 00:02:58.800 document right in front of them. That is a perfect 00:02:58.800 --> 00:03:01.960 analogy. Yes. Now, Notebook LM is the complete 00:03:01.960 --> 00:03:04.300 opposite of that. OK. It's an enclosed ecosystem. 00:03:04.780 --> 00:03:07.460 It strictly answers from the documents you specifically 00:03:07.460 --> 00:03:09.939 upload. Only what you give it. Exactly. You drop 00:03:09.939 --> 00:03:12.819 in a PDF or a slide deck or a spreadsheet, every 00:03:12.819 --> 00:03:15.259 single sentence it generates comes straight from 00:03:15.259 --> 00:03:17.419 that material. It cites it, right. Right. It 00:03:17.419 --> 00:03:20.360 proves it. It gives you exact clickable citations. 00:03:20.539 --> 00:03:23.080 So if Gemini is the distractible researcher, 00:03:23.680 --> 00:03:26.680 Notepick LM is like a strict old school librarian. 00:03:27.419 --> 00:03:30.099 They only point to physical pages inside the 00:03:30.099 --> 00:03:32.020 room. If it's not in the room, it just doesn't 00:03:32.020 --> 00:03:34.479 exist. They absolutely refuse to look out the 00:03:34.479 --> 00:03:37.300 window. That's exactly it. But, you know, that 00:03:37.300 --> 00:03:39.419 exact strictness is also his biggest weakness. 00:03:39.539 --> 00:03:41.719 Because it's isolated. Right. It knows absolutely 00:03:41.719 --> 00:03:43.740 nothing about the outside world. If you ask it 00:03:43.740 --> 00:03:46.000 about a competitor's press release from like 00:03:46.000 --> 00:03:49.039 this morning, it has nothing to offer you. Yeah. 00:03:49.219 --> 00:03:51.819 The rigid accuracy that makes it so reliable 00:03:51.819 --> 00:03:56.080 also entirely limits its reach. So we had the 00:03:56.080 --> 00:03:58.439 brilliant researcher and the strict librarian, 00:03:58.460 --> 00:04:02.219 and they were isolated. But the April 2026 update 00:04:02.219 --> 00:04:05.180 built a bridge. It sure did. It integrated Notebook 00:04:05.180 --> 00:04:08.840 LM directly into Gemini's side panel. So now 00:04:08.840 --> 00:04:11.580 they operate as one connected workflow. Right. 00:04:11.900 --> 00:04:13.919 Notebook LM holds your trusted source material, 00:04:14.419 --> 00:04:16.759 and Gemini searches outside that material when 00:04:16.759 --> 00:04:19.939 you need broader context. And then those external 00:04:19.939 --> 00:04:23.040 findings get routed back into Notebook LM. Exactly. 00:04:23.220 --> 00:04:25.519 Your knowledge base keeps growing, but the strict 00:04:25.519 --> 00:04:28.920 boundary remains completely intact. beat. This 00:04:28.920 --> 00:04:31.759 integration is profound, but it makes me wonder, 00:04:32.240 --> 00:04:34.920 does this unified system make the standalone 00:04:34.920 --> 00:04:37.480 versions of these tools completely obsolete? 00:04:37.899 --> 00:04:40.800 Not entirely, no. I mean, if you just need a 00:04:40.800 --> 00:04:43.220 rapid summary of a web article, standalone Gemini 00:04:43.220 --> 00:04:45.220 is still faster. Right. And if you're setting 00:04:45.220 --> 00:04:47.560 a dense textbook and want zero distractions, 00:04:48.019 --> 00:04:50.639 standalone notebook LM is perfect. But for deep 00:04:50.639 --> 00:04:53.240 professional knowledge work, The hybrid workflow 00:04:53.240 --> 00:04:55.860 is absolutely the new standard. So integration 00:04:55.860 --> 00:04:58.379 elevates both, but doesn't erase their standalone 00:04:58.379 --> 00:05:01.259 value. Exactly. So we know why we need this connected 00:05:01.259 --> 00:05:04.420 brain. But if you feed it garbage, it's going 00:05:04.420 --> 00:05:07.180 to hallucinate anyway. 100%. Let's talk about 00:05:07.180 --> 00:05:09.600 the foundation. How do we build this moat without 00:05:09.600 --> 00:05:12.079 corrupting it? Well, everything hinges on what 00:05:12.079 --> 00:05:15.180 you put into Notebook LM first. If you skip this 00:05:15.180 --> 00:05:17.379 setup phase, the entire loop just collapses. 00:05:17.680 --> 00:05:21.040 Rule number one is absolute. You need one notebook 00:05:21.040 --> 00:05:24.800 per project. You must never mix topics. Humans 00:05:24.800 --> 00:05:26.980 have a natural tendency to hoard data, though. 00:05:27.180 --> 00:05:29.240 We really do. We love to dump everything into 00:05:29.240 --> 00:05:31.680 one giant folder. We throw marketing metrics 00:05:31.680 --> 00:05:33.980 right next to HR strategies. Yeah, and that's 00:05:33.980 --> 00:05:36.279 a disaster. Explain the actual mechanism here. 00:05:36.540 --> 00:05:39.300 Why does data hoarding break the AI's accuracy? 00:05:40.120 --> 00:05:42.519 It comes down to semantic search. Let's define 00:05:42.519 --> 00:05:45.050 that really quick. Semantic search is finding 00:05:45.050 --> 00:05:48.750 information based on meaning rather than exact 00:05:48.750 --> 00:05:53.189 keyword matches. Perfect. So the AI doesn't read 00:05:53.189 --> 00:05:55.670 like we do, right? It turns words into mathematical 00:05:55.670 --> 00:05:58.310 relationships. Okay. If you put HR data about 00:05:58.310 --> 00:06:01.050 a readiness gap next to marketing data about 00:06:01.050 --> 00:06:03.589 a readiness gap, the system grabs both because 00:06:03.589 --> 00:06:06.069 they share buzzwords. Oh, I see. Yeah, it blends 00:06:06.069 --> 00:06:08.069 them together. You get a hallucinated, blended 00:06:08.069 --> 00:06:10.870 answer that is mathematically logical, but factually 00:06:10.870 --> 00:06:12.910 totally useless. You have to keep the boundary 00:06:12.910 --> 00:06:15.269 tight. Let's ground this a bit. Walk me through 00:06:15.269 --> 00:06:18.110 a concrete example. OK. Imagine we're building 00:06:18.110 --> 00:06:21.170 a hypothetical notebook today. We'll title it 00:06:21.170 --> 00:06:25.089 The State of Organization's 2026, the agentic 00:06:25.089 --> 00:06:27.850 AI transformation. And let's define that term, 00:06:28.089 --> 00:06:30.980 too. Agenetic AI is AI that takes independent 00:06:30.980 --> 00:06:34.000 actions to achieve a specific goal. Exactly. 00:06:34.319 --> 00:06:36.620 So this notebook holds exactly 10 sources. You've 00:06:36.620 --> 00:06:39.879 got a dense Cisco PDF report, a Stanford research 00:06:39.879 --> 00:06:44.019 summary, a highly technical ARCISIP study, and 00:06:44.019 --> 00:06:46.699 a McKinsey strategy report. Every single file 00:06:46.699 --> 00:06:50.339 stays strictly focused on one topic, how organizations 00:06:50.339 --> 00:06:53.850 must adapt to AI. So no unrelated quarterly earnings, 00:06:54.230 --> 00:06:56.269 no random marketing decks. Nothing like that. 00:06:56.449 --> 00:06:58.990 It processes a huge variety of formats now, doesn't 00:06:58.990 --> 00:07:01.519 it? Oh yeah. It digests PDFs, Google Docs, slide 00:07:01.519 --> 00:07:04.000 decks, and complex spreadsheets. Wow. And it 00:07:04.000 --> 00:07:06.620 also accepts live websites, YouTube videos, and 00:07:06.620 --> 00:07:08.560 raw audio files. All in the same notebook? All 00:07:08.560 --> 00:07:10.839 in the same notebook. It synthesizes across every 00:07:10.839 --> 00:07:13.160 format simultaneously. You can literally ask 00:07:13.160 --> 00:07:15.379 it to compare a statistical claim in a McKinsey 00:07:15.379 --> 00:07:18.240 PDF against an offhand comment made in a YouTube 00:07:18.240 --> 00:07:19.920 interview. That is wild. There are still limits 00:07:19.920 --> 00:07:21.959 on the free tier, though, right? Right, yeah. 00:07:22.100 --> 00:07:24.540 The free tier gives you 50 sources per notebook. 00:07:24.660 --> 00:07:28.120 You get 50 queries a day and three audio overviews. 00:07:28.120 --> 00:07:30.230 Which is pretty It's generous. It's usually plenty 00:07:30.230 --> 00:07:33.350 for a single project. But power users on the 00:07:33.350 --> 00:07:37.129 plus, pro, or ultra tiers bump that up to 100, 00:07:37.290 --> 00:07:39.870 300, or 600 sources. OK, so our moat is built. 00:07:39.949 --> 00:07:43.170 We have 10 pristine documents. How do we actually 00:07:43.170 --> 00:07:45.069 interrogate this data? I don't want to just ask 00:07:45.069 --> 00:07:47.110 it, what does this say? Right, you want to go 00:07:47.110 --> 00:07:49.839 deeper. Yeah. Walk me through the query workflow. 00:07:50.420 --> 00:07:52.600 So step one is the overview phase. You explicitly 00:07:52.600 --> 00:07:55.459 tell it. Summarize the main argument of each 00:07:55.459 --> 00:07:57.300 source in this notebook. Give me one sentence 00:07:57.300 --> 00:08:00.279 per source with a citation. OK. This builds a 00:08:00.279 --> 00:08:02.300 high level map of your data. You're basically 00:08:02.300 --> 00:08:05.339 forcing it to read the room before you ask detailed 00:08:05.339 --> 00:08:07.160 questions. Exactly. You're making it show its 00:08:07.160 --> 00:08:10.279 work. Then step two is the narrow down. You avoid 00:08:10.279 --> 00:08:13.860 broad, lazy questions like, how will AI affect 00:08:13.860 --> 00:08:16.399 jobs? Too generic. Way too generic. Instead, 00:08:16.399 --> 00:08:19.470 you get surgical. You say, compare what the Cisco 00:08:19.470 --> 00:08:22.029 report and McKinsey report say about the leadership 00:08:22.029 --> 00:08:25.410 readiness gap, quote, specific percentages. Oh, 00:08:25.709 --> 00:08:29.089 nice. This forces the AI to ignore the other 00:08:29.089 --> 00:08:31.870 eight documents and pull data strictly from two 00:08:31.870 --> 00:08:34.769 specific named sources. Vita, I'm guessing step 00:08:34.769 --> 00:08:36.750 three is where it gets genuinely analytical. 00:08:36.870 --> 00:08:39.629 Oh, yeah. Step three is hunting for friction. 00:08:40.169 --> 00:08:43.350 You ask for disagreements. You type. Among these 00:08:43.350 --> 00:08:46.470 10 sources, who predicts a wildly different timeline 00:08:46.470 --> 00:08:49.690 for total job replacement? It scans the entire 00:08:49.690 --> 00:08:52.330 vector space at once. It finds exactly where 00:08:52.330 --> 00:08:55.009 the experts collide. That's fascinating. But 00:08:55.009 --> 00:08:58.169 what happens when the data itself is fundamentally 00:08:58.169 --> 00:09:01.370 broken? Like, what if Cisco and McKinsey flatly 00:09:01.370 --> 00:09:03.429 contradict each other on a crucial metric? Well, 00:09:03.490 --> 00:09:05.570 you don't try to force a resolution. You ask 00:09:05.570 --> 00:09:07.870 the AI to map the contradiction. OK. You tell 00:09:07.870 --> 00:09:10.950 it. Explain why Cisco and McKinsey reach different 00:09:10.950 --> 00:09:13.029 conclusions based on their distinct research 00:09:13.029 --> 00:09:16.009 methodologies. The friction itself becomes your 00:09:16.009 --> 00:09:18.450 most valuable insight. Preserve the conflict 00:09:18.450 --> 00:09:21.129 to let the AI map the debate. You nailed it. 00:09:21.250 --> 00:09:23.529 Sponsor, sponsor Reed goes here. So we've established 00:09:23.529 --> 00:09:25.649 our walled garden. We've exhausted what's inside 00:09:25.649 --> 00:09:28.070 our uploaded documents. We understand the internal 00:09:28.070 --> 00:09:30.570 friction. But business doesn't happen in a vacuum, 00:09:30.730 --> 00:09:33.460 right? A competitor could drop a massive press 00:09:33.460 --> 00:09:35.700 release this morning, and my walled garden would 00:09:35.700 --> 00:09:38.539 be completely blind to it. Totally blind. How 00:09:38.539 --> 00:09:40.740 do we reach out to the live web without letting 00:09:40.740 --> 00:09:43.159 the hallucinations back in? This is where the 00:09:43.159 --> 00:09:45.799 workflow gets really elegant. This is step three, 00:09:46.159 --> 00:09:48.659 bringing Gemini into the loop for outside research. 00:09:49.299 --> 00:09:51.919 OK. Notebook LM deliberately stops at the edge 00:09:51.919 --> 00:09:54.759 of your files. The moment your question requires 00:09:54.759 --> 00:09:58.500 fresh external context, Gemini takes the wheel 00:09:58.500 --> 00:10:01.419 using a feature called Deep Research. Push back 00:10:01.419 --> 00:10:04.440 on that for a second. How is Deep Research actually 00:10:04.440 --> 00:10:06.720 different from just asking Gemini to run a quick 00:10:06.720 --> 00:10:09.529 Google search? It's a completely different underlying 00:10:09.529 --> 00:10:12.509 mechanism. I mean, a standard search just pulls 00:10:12.509 --> 00:10:14.769 the top few results and summarizes them, right? 00:10:15.269 --> 00:10:18.009 Deep research acts like an actual human analyst. 00:10:18.549 --> 00:10:21.309 It plans a multi -step search strategy. It runs 00:10:21.309 --> 00:10:24.070 those queries. It clicks into the links, reads 00:10:24.070 --> 00:10:26.389 the long -form text, evaluates the credibility, 00:10:26.710 --> 00:10:28.830 and then synthesizes it all. And it returns a 00:10:28.830 --> 00:10:31.190 highly structured, comprehensively cited report. 00:10:31.309 --> 00:10:33.610 We're finally combining the strict librarian 00:10:33.610 --> 00:10:35.850 with the web researcher. Exactly. Let's talk 00:10:35.850 --> 00:10:38.029 about how this practically works. The single 00:10:38.029 --> 00:10:40.129 chat loop seems to be the crucial breakthrough 00:10:40.129 --> 00:10:43.690 of this April 2026 update. It really is. It's 00:10:43.690 --> 00:10:47.250 this magic 25 -minute workflow. The friction 00:10:47.250 --> 00:10:49.929 used to be jumping between tabs, right? Oh, constantly. 00:10:50.190 --> 00:10:53.490 Now you stay in one single chat window. You open 00:10:53.490 --> 00:10:56.860 Gemini. You click the plus button and you attach 00:10:56.860 --> 00:10:59.779 your carefully curated notebook. Okay, so I'm 00:10:59.779 --> 00:11:01.940 looking at the blank text box. What's the first 00:11:01.940 --> 00:11:04.620 question in the loop? Question one is strictly 00:11:04.620 --> 00:11:07.419 internal. You ask the notebook to establish the 00:11:07.419 --> 00:11:10.700 baseline. You type, based solely on our uploaded 00:11:10.700 --> 00:11:13.919 docs, what is the internal readiness gap for 00:11:13.919 --> 00:11:16.659 our company? Got it. Notebook LM answers, providing 00:11:16.659 --> 00:11:19.000 strict clickable citations. So you have your 00:11:19.000 --> 00:11:21.759 anchor. Then, without leaving that specific chat 00:11:21.759 --> 00:11:24.120 window, you shift gears. Right. You move to question 00:11:24.120 --> 00:11:26.500 two, you hit the deep research button right there 00:11:26.500 --> 00:11:28.759 on the chat, and you ask it to look at the wider 00:11:28.759 --> 00:11:31.639 world. You say, run deep research on how this 00:11:31.639 --> 00:11:34.039 specific readiness gap is manifesting across 00:11:34.039 --> 00:11:37.320 our top three competitors this quarter. Gemini 00:11:37.320 --> 00:11:39.759 goes out, does the heavy lifting, and comes back 00:11:39.759 --> 00:11:41.980 with a structured report from the live web. And 00:11:41.980 --> 00:11:44.179 then question three is where the real knowledge 00:11:44.179 --> 00:11:47.899 work happens. Yes, the synthesis. Because both 00:11:47.899 --> 00:11:50.080 previous answers are still active in the chat's 00:11:50.080 --> 00:11:53.460 memory, you simply type. Based on our internal 00:11:53.460 --> 00:11:55.799 documents and the external industry research 00:11:55.799 --> 00:11:58.500 you just found, what are the top two strategic 00:11:58.500 --> 00:12:01.740 moves leadership must make right now? Gemini 00:12:01.740 --> 00:12:03.840 weaves the internal constraints with the external 00:12:03.840 --> 00:12:06.179 realities. You don't copy, paste, or summarize 00:12:06.179 --> 00:12:09.039 anything yourself. Two sec silence. That reliance 00:12:09.039 --> 00:12:11.779 on active memory is incredible. We should clarify 00:12:11.779 --> 00:12:14.340 a concept here. A context window is how much 00:12:14.340 --> 00:12:16.679 text an AI can remember at one time. Good call. 00:12:16.919 --> 00:12:19.200 The top tier models now support Windows reaching 00:12:19.200 --> 00:12:21.720 one to two million tokens. Which is just massive. 00:12:22.100 --> 00:12:26.440 Whoa. Imagine scaling to a million tokens. It 00:12:26.440 --> 00:12:29.299 takes me a full weekend to deeply read and synthesize 00:12:29.299 --> 00:12:32.860 one dense textbook. The idea of holding an entire 00:12:32.860 --> 00:12:35.600 decade of corporate history in active working 00:12:35.600 --> 00:12:38.879 memory. It fundamentally breaks how we think 00:12:38.879 --> 00:12:41.399 about human expertise. It is staggering. I mean, 00:12:41.460 --> 00:12:43.919 it really is, but you do have to manage it carefully. 00:12:44.139 --> 00:12:46.840 How so? Well, the system overhead in the regular 00:12:46.840 --> 00:12:49.919 app eats up a chunk of that space. Processing 00:12:49.919 --> 00:12:52.659 a few complex reports is flawless. Well, if you 00:12:52.659 --> 00:12:55.539 try to cram, like, five years of raw financial 00:12:55.539 --> 00:12:58.620 ledgers into one chat, the AI will start quietly 00:12:58.620 --> 00:13:00.679 trimming the edges of his memory. It loses things. 00:13:00.840 --> 00:13:02.820 What about workspace integration during this 00:13:02.820 --> 00:13:05.460 loop? Because my company's knowledge isn't just 00:13:05.460 --> 00:13:07.740 in pristine PDFs. Well, mine's is. It's scattered 00:13:07.740 --> 00:13:09.879 across chaotic email threads. And that's the 00:13:09.879 --> 00:13:11.769 beauty of it. You can be in that same chat and 00:13:11.769 --> 00:13:13.789 say, check my Gmail and drive for any scattered 00:13:13.789 --> 00:13:15.909 notes from the last three months mentioning agentic 00:13:15.909 --> 00:13:19.149 AI. Really? In the same chat? Yes. It reads your 00:13:19.149 --> 00:13:21.710 messy email threads. It pulls context from meeting 00:13:21.710 --> 00:13:24.049 summaries. It maps your chaotic reality against 00:13:24.049 --> 00:13:27.070 the priskeen research. Then, crucially, you don't 00:13:27.070 --> 00:13:29.289 just leave that deep research sitting in the 00:13:29.289 --> 00:13:32.230 chat history. Never. You click Add to Notebook, 00:13:32.470 --> 00:13:35.029 right on that external report. Oh, I see. That 00:13:35.029 --> 00:13:37.710 fresh web research becomes a permanent, sighted 00:13:37.710 --> 00:13:40.049 source inside your notebook LM walled garden. 00:13:40.330 --> 00:13:43.710 Your local knowledge base continually evolves. 00:13:44.070 --> 00:13:46.669 It's a delicate balance though. If I keep dumping 00:13:46.669 --> 00:13:49.769 web research into my notebook, doesn't that risk 00:13:49.769 --> 00:13:52.830 contaminating our pristine internal data? Only 00:13:52.830 --> 00:13:55.090 if you get lazy with your prompting. Okay. You 00:13:55.090 --> 00:13:57.789 have to clearly instruct the AI to contrast the 00:13:57.789 --> 00:14:00.610 internal data against the external data. If you 00:14:00.610 --> 00:14:03.610 just ask for a generic summary, the sheer volume 00:14:03.610 --> 00:14:06.450 of external web data can absolutely drown out 00:14:06.450 --> 00:14:09.090 the nuances of your specific local files. External 00:14:09.090 --> 00:14:11.950 data adds context, but local files remain the 00:14:11.950 --> 00:14:14.029 ultimate anchor. That's exactly how you have 00:14:14.029 --> 00:14:16.690 to treat it. So doing this 25 -minute loop once 00:14:16.690 --> 00:14:19.590 is brilliant. But if I have to set up the persona, 00:14:20.350 --> 00:14:22.889 define the tone, and attach the notebook from 00:14:22.889 --> 00:14:26.110 scratch every Tuesday morning, it becomes a nightmare. 00:14:26.210 --> 00:14:28.450 Oh, it's exhausting. This brings us to step five. 00:14:28.720 --> 00:14:33.299 Making it systemic. Yes. The blank slate is the 00:14:33.299 --> 00:14:36.000 enemy of productivity. Every time you open a 00:14:36.000 --> 00:14:38.220 new Gemini chat, it forgets who you are. Right. 00:14:38.360 --> 00:14:40.019 You have to re -explain that you're a senior 00:14:40.019 --> 00:14:42.200 analyst. You have to tell it to be concise. You 00:14:42.200 --> 00:14:44.679 have to manually link the notebook again. It's 00:14:44.679 --> 00:14:46.899 a massive source of friction. Enter automation. 00:14:47.399 --> 00:14:50.559 Enter gems. Gems are the ultimate fix. A gem 00:14:50.559 --> 00:14:53.460 is a custom specialist persona that lives inside 00:14:53.460 --> 00:14:56.740 Gemini. Okay. It permanently remembers its specific 00:14:56.740 --> 00:14:59.740 role across an infinite number of chats. And 00:14:59.740 --> 00:15:02.299 most importantly, you can permanently tie a gem 00:15:02.299 --> 00:15:06.259 to a specific notebook LM project. So let's say 00:15:06.259 --> 00:15:08.620 I build a strategy analyst gem. What does that 00:15:08.620 --> 00:15:11.019 actually look like under the hood? Well... You 00:15:11.019 --> 00:15:13.240 configure it once. You tie it directly to your 00:15:13.240 --> 00:15:15.639 Agenic AI notebook. You give it strict instructions. 00:15:16.299 --> 00:15:18.840 Always prioritize the notebook data. Always back 00:15:18.840 --> 00:15:22.159 up your claims with exact citations. Always highlight 00:15:22.159 --> 00:15:24.980 potential regulatory risks. Wow. From then on, 00:15:25.100 --> 00:15:27.279 you just click the gem and it's instantly ready 00:15:27.279 --> 00:15:28.960 to work. I don't want to just copy and paste 00:15:28.960 --> 00:15:30.960 from a chat window all day though. Where does 00:15:30.960 --> 00:15:33.720 the output actually go? It pushes directly into 00:15:33.720 --> 00:15:36.590 your workflow. A synthesized strategy report 00:15:36.590 --> 00:15:39.289 goes straight into a formatted Google Doc. That's 00:15:39.289 --> 00:15:41.730 convenient. A table comparing competitor metrics 00:15:41.730 --> 00:15:44.649 updates a live Google Sheet. A list of action 00:15:44.649 --> 00:15:47.309 items becomes a drafted follow -up email right 00:15:47.309 --> 00:15:49.809 inside your Gmail inbox. And there's a fascinating 00:15:49.809 --> 00:15:52.129 Google Meet integration now too, right? Yeah. 00:15:52.370 --> 00:15:55.250 This is a game changer. Gemini inside Google 00:15:55.250 --> 00:15:57.289 Meet can catch you up on the first 20 minutes 00:15:57.289 --> 00:16:00.029 of a meeting you missed. It extracts the action 00:16:00.029 --> 00:16:02.950 items. That alone is huge. Right. But here's 00:16:02.950 --> 00:16:05.970 the best part. It reads your Google Drive to 00:16:05.970 --> 00:16:09.110 learn your team's specific writing style. Over 00:16:09.110 --> 00:16:11.710 time, the memos it drafts actually sound like 00:16:11.710 --> 00:16:14.429 your specific corporate culture, not a generic 00:16:14.429 --> 00:16:16.590 robot. This completely shifts the human role. 00:16:16.879 --> 00:16:19.120 For decades, we've been information gatherers. 00:16:19.279 --> 00:16:22.039 We scour databases. We copy -pasted metrics. 00:16:22.240 --> 00:16:24.440 We assembled the raw materials. Yeah, the busy 00:16:24.440 --> 00:16:26.860 work. Now, we are systems architects. We build 00:16:26.860 --> 00:16:29.299 the trusted knowledge base. We define the parameters 00:16:29.299 --> 00:16:32.139 of the persona. We set the ethical boundaries. 00:16:32.919 --> 00:16:35.620 The machine executes the gathering. It fundamentally 00:16:35.620 --> 00:16:38.500 changes what it means to work. It elevates the 00:16:38.500 --> 00:16:41.460 work. You're finally spending your time thinking 00:16:41.460 --> 00:16:44.059 rather than endlessly searching. But systems 00:16:44.059 --> 00:16:47.200 evolve. If my team finishes a new quarterly report 00:16:47.200 --> 00:16:49.700 and I drop that new PDF into the underlying notebook, 00:16:50.200 --> 00:16:53.139 does the gem automatically know? Or do I have 00:16:53.139 --> 00:16:55.440 to retrain it? You don't touch a thing. Really? 00:16:55.720 --> 00:16:58.460 The gem acts as a lens. It is constantly looking 00:16:58.460 --> 00:17:00.419 at the current state of the notebook. If you 00:17:00.419 --> 00:17:03.200 drop a new PDF in today, the gem will seamlessly 00:17:03.200 --> 00:17:06.200 include that fresh data in its answers tomorrow. 00:17:06.559 --> 00:17:09.099 The GEM automatically updates its understanding 00:17:09.099 --> 00:17:11.480 when your local data changes. That's it, exactly. 00:17:11.660 --> 00:17:13.619 We've covered a massive amount of ground today. 00:17:13.940 --> 00:17:16.240 We navigated the structural divide between strict 00:17:16.240 --> 00:17:19.559 citation and broad web research. We built a trusted 00:17:19.559 --> 00:17:22.180 moat. We reached outside the walls using deep 00:17:22.180 --> 00:17:25.400 research. And we automated the entire system. 00:17:26.599 --> 00:17:28.839 Can you distill the core philosophy of this deep 00:17:28.839 --> 00:17:30.960 dive for us? I'd say the core philosophy is simple. 00:17:31.480 --> 00:17:34.380 Stop app switching. Use Notebook LM as your trusted, 00:17:34.599 --> 00:17:37.109 deeply -cited anchor. It holds the absolute facts 00:17:37.109 --> 00:17:39.910 you verified. Use Gemini as your scout. It runs 00:17:39.910 --> 00:17:42.289 out to the live web, grabs fresh intelligence, 00:17:42.549 --> 00:17:44.690 and brings it back to the anchor. Run that entire 00:17:44.690 --> 00:17:48.390 operation in one single, focused, 25 -minute 00:17:48.390 --> 00:17:50.890 chat loop. The single loop. Exactly. And once 00:17:50.890 --> 00:17:53.230 you perfect that loop, set it in stone by building 00:17:53.230 --> 00:17:56.049 a jam. Stop starting from zero every single day, 00:17:56.390 --> 00:17:58.450 build the system once, and let it do the heavy 00:17:58.450 --> 00:18:00.950 lifting for you. It's an incredibly powerful 00:18:00.950 --> 00:18:03.910 paradigm shift. We are moving away from confident 00:18:03.910 --> 00:18:07.190 AI guesses toward verifiable automated truth. 00:18:07.690 --> 00:18:09.430 But it leaves me with a final thought for you 00:18:09.430 --> 00:18:13.480 to chew on. OK. If we can now effortlessly build 00:18:13.480 --> 00:18:17.099 personalized, sighted, hallucination -free AI 00:18:17.099 --> 00:18:20.450 analysts for every single project... How does 00:18:20.450 --> 00:18:22.829 that change what we fundamentally value in human 00:18:22.829 --> 00:18:25.390 intelligence? Ooh, that's a big question. What 00:18:25.390 --> 00:18:27.009 happens to our culture when having the right 00:18:27.009 --> 00:18:29.970 answer is easy, cheap, and instantaneous? Perhaps 00:18:29.970 --> 00:18:31.750 in a world where answers are fully automated, 00:18:31.829 --> 00:18:33.930 the only thing that truly matters anymore is 00:18:33.930 --> 00:18:36.549 our ability to ask the right question. Wow. Two 00:18:36.549 --> 00:18:39.069 -Sec Silence. I highly encourage you to pick 00:18:39.069 --> 00:18:42.170 just one real -world project this week. Set up 00:18:42.170 --> 00:18:44.349 a 10 -source notebook, run the three -step single 00:18:44.349 --> 00:18:46.410 -chat loop for yourself, see how it changes the 00:18:46.410 --> 00:18:48.410 way you think about your own work, out to your 00:18:48.410 --> 00:18:48.950 own music.