WEBVTT 00:00:00.000 --> 00:00:03.160 So you sit down at your desk, you open a brand 00:00:03.160 --> 00:00:06.259 new AI chat window, and you're ready to work. 00:00:06.259 --> 00:00:08.960 Right. But immediately you realize you have a 00:00:08.960 --> 00:00:12.439 problem. The AI has amnesia. Oh, complete and 00:00:12.439 --> 00:00:15.439 utter amnesia. It is a totally blank slate every 00:00:15.439 --> 00:00:17.399 single time. It really is. It starts from zero. 00:00:17.559 --> 00:00:19.260 I mean, it doesn't know anything about your business. 00:00:19.379 --> 00:00:21.539 It doesn't know your brand's tone. No, it has 00:00:21.539 --> 00:00:23.620 absolutely no idea what your goals are for the 00:00:23.620 --> 00:00:26.710 quarter either. Right. So. What do you do? You 00:00:26.710 --> 00:00:28.629 sit there and you just type it all out again? 00:00:28.750 --> 00:00:30.870 Yeah, you copy and paste your company bio, you 00:00:30.870 --> 00:00:32.810 paste your style guide, you grab your project 00:00:32.810 --> 00:00:36.390 notes. You do it again and again. It is exhausting. 00:00:37.409 --> 00:00:40.710 Welcome to the deep dive. Our mission today is 00:00:40.710 --> 00:00:43.609 solving this repetition fatigue once and for 00:00:43.609 --> 00:00:46.719 all. We're going to talk about crossing a bridge 00:00:46.719 --> 00:00:49.960 from just using generic chatbots to building 00:00:49.960 --> 00:00:52.399 a personal AI agent that actually knows you. 00:00:53.000 --> 00:00:54.640 Exactly. We want to build an assistant that shows 00:00:54.640 --> 00:00:57.159 up ready to work every single time you open your 00:00:57.159 --> 00:01:00.280 laptop. So the roadmap for today is pretty straightforward. 00:01:00.420 --> 00:01:03.659 You're going to combine two powerful tools, Google 00:01:03.659 --> 00:01:07.140 Notebook LM and Gemini Gems. Yeah, that is the 00:01:07.140 --> 00:01:09.680 magic combo. Right. So Notebook LM is going to 00:01:09.680 --> 00:01:13.709 act as the brain of your operation. Gemini gems, 00:01:14.230 --> 00:01:16.349 well, they're going to provide the personality 00:01:16.349 --> 00:01:18.209 and the rules. Rhythm together, and you have 00:01:18.209 --> 00:01:20.650 basically created a custom assistant that never 00:01:20.650 --> 00:01:22.829 forgets. Which is huge. It really is a total 00:01:22.829 --> 00:01:26.049 game changer for your daily workflow. OK, let's 00:01:26.049 --> 00:01:27.870 unpack this. Yeah. Because before we just start 00:01:27.870 --> 00:01:29.989 mashing buttons and building things, we need 00:01:29.989 --> 00:01:32.590 to understand the architecture. Sure. Why do 00:01:32.590 --> 00:01:35.969 generic AIs fail us when it comes to specific 00:01:35.969 --> 00:01:39.099 specialized work? Well, the core problem is actually 00:01:39.099 --> 00:01:41.219 pretty simple if you think about it. Generic 00:01:41.219 --> 00:01:44.480 AI pulls its answers from the entire internet. 00:01:45.000 --> 00:01:47.760 It is designed to give you a generalized average 00:01:47.760 --> 00:01:50.219 summation of all human knowledge. Which is great 00:01:50.219 --> 00:01:52.920 if you need like a recipe for banana bread. Exactly. 00:01:53.159 --> 00:01:56.540 But it is terrible if you need a specific analysis 00:01:56.540 --> 00:01:59.819 of your Q3 marketing strategy. I mean, it doesn't 00:01:59.819 --> 00:02:02.219 have access to your private files. It doesn't 00:02:02.219 --> 00:02:05.420 know your internal company rules. It is just 00:02:05.420 --> 00:02:07.719 guessing based on what other companies might 00:02:07.719 --> 00:02:11.379 do. So we fixed this blind spot with Google Notebook 00:02:11.379 --> 00:02:13.840 LM. And I hear people call this a smart folder 00:02:13.840 --> 00:02:17.020 or a second brain. But honestly, smart folder 00:02:17.020 --> 00:02:19.560 feels a little bit reductive. Yeah, I agree. 00:02:19.699 --> 00:02:21.780 How does it actually work under the hood? Well, 00:02:21.819 --> 00:02:23.599 think about a normal folder on your desktop. 00:02:23.979 --> 00:02:27.120 It just stores files. It is a passive container. 00:02:27.319 --> 00:02:29.479 You still have to double -click them, open them, 00:02:29.520 --> 00:02:31.300 and read them yourself. Right. You do the work. 00:02:31.500 --> 00:02:35.000 Exactly. But with Notebook LM, those files actually 00:02:35.000 --> 00:02:38.240 become the AI's active memory. It reads them, 00:02:38.479 --> 00:02:40.240 processes them, and holds them in its working 00:02:40.240 --> 00:02:43.280 brain for you. And you can upload up to 300 different 00:02:43.280 --> 00:02:46.560 sources into a single notebook. And that includes 00:02:46.560 --> 00:02:50.240 PDFs, your own rough personal notes, or even 00:02:50.240 --> 00:02:52.340 direct links to websites. Right. If you are a 00:02:52.340 --> 00:02:54.180 student, you can literally upload your entire 00:02:54.180 --> 00:02:56.639 semester's worth of textbooks. Or if you are 00:02:56.639 --> 00:02:59.139 a business owner, you are uploading your product 00:02:59.139 --> 00:03:01.780 specs, your customer feedback, your financial 00:03:01.780 --> 00:03:03.860 sheets. Right. And here's where it gets really 00:03:03.860 --> 00:03:06.740 interesting. When you ask a question inside that 00:03:06.740 --> 00:03:09.259 notebook, it isn't searching the internet anymore. 00:03:09.360 --> 00:03:13.419 It stays local. Exactly. It pulls directly and 00:03:13.419 --> 00:03:15.460 only from the documents you gave it. It will 00:03:15.460 --> 00:03:17.919 literally say, according to page five of the 00:03:17.919 --> 00:03:20.419 Q3 financial report. That brings up a term we 00:03:20.419 --> 00:03:22.379 see thrown around constantly in the tech world. 00:03:22.639 --> 00:03:25.780 Grounding. Yes, grounding. But it often sounds 00:03:25.780 --> 00:03:27.919 like marketing jargon. What does it actually 00:03:27.919 --> 00:03:30.919 mean in plain English? It means anchoring the 00:03:30.919 --> 00:03:33.879 AI to your actual data, not the open internet. 00:03:34.039 --> 00:03:37.300 I love that. Anchoring the AI to your actual 00:03:37.300 --> 00:03:39.819 data. So there is no more guessing. None. The 00:03:39.819 --> 00:03:41.379 source is cited right there on the screen for 00:03:41.379 --> 00:03:44.419 you to verify. But there is a catch. OK. Notebook 00:03:44.419 --> 00:03:47.180 LM is amazing at storing facts, but its behavior 00:03:47.180 --> 00:03:49.500 still resets every time you start a new chat. 00:03:49.900 --> 00:03:52.039 It might give you a long academic answer one 00:03:52.039 --> 00:03:54.439 day and then just a bulleted list the next. Ah, 00:03:54.460 --> 00:03:56.819 I see. That is where Gemini gems come into play. 00:03:57.080 --> 00:04:00.669 So if Notebook LM is the memory. The gems of 00:04:00.669 --> 00:04:03.990 the personality. Yes. Think of a gem as a permanent 00:04:03.990 --> 00:04:06.889 behavior layer. It is a permanent system prompt 00:04:06.889 --> 00:04:09.129 that intercepts everything. It sits right on 00:04:09.129 --> 00:04:10.909 top of your notebook. Gotcha. Every time you 00:04:10.909 --> 00:04:13.530 ask a question, the gem features the answer through 00:04:13.530 --> 00:04:16.009 its specific set of rules. So I could build, 00:04:16.029 --> 00:04:19.370 say, a social media manager gem. And I'd tell 00:04:19.370 --> 00:04:21.870 it to always use a friendly tone. I'd tell it 00:04:21.870 --> 00:04:25.279 to keep sentences. under 12 words. And I put 00:04:25.279 --> 00:04:28.360 in a strict rule, like never ever use emojis. 00:04:28.699 --> 00:04:31.439 Exactly. And once you save that gem, those rules 00:04:31.439 --> 00:04:33.420 stick forever. You don't have to remind it to 00:04:33.420 --> 00:04:35.519 drop the emojis every time you open a new chat. 00:04:35.740 --> 00:04:38.560 That saves so much time. It really does. The 00:04:38.560 --> 00:04:41.300 notebook holds the deep knowledge and the gem 00:04:41.300 --> 00:04:44.220 holds the strict behavioral rules. Together, 00:04:44.639 --> 00:04:46.980 they sync up perfectly. I picture it kind of 00:04:46.980 --> 00:04:49.360 like stacking Lego blocks. Oh, I like that. You've 00:04:49.360 --> 00:04:51.120 got your base plate, which is the underlying 00:04:51.120 --> 00:04:54.000 AI model. Then you snap on the notebook block, 00:04:54.060 --> 00:04:56.120 which gives it your company's entire memory. 00:04:56.839 --> 00:04:58.959 Finally, you snap the gem block right on top, 00:04:59.220 --> 00:05:02.279 giving it your specific brand voice. You're building 00:05:02.279 --> 00:05:04.759 a custom employee who has already memorized your 00:05:04.759 --> 00:05:06.819 entire company handbook. And because they know 00:05:06.819 --> 00:05:08.500 your brand voice and they know your products 00:05:08.500 --> 00:05:11.540 perfectly, the output feels consistent and reliable 00:05:11.540 --> 00:05:14.139 every single time. So if the notebook is the 00:05:14.139 --> 00:05:17.240 memory, the gem is basically the instruction 00:05:17.240 --> 00:05:19.750 manual for how to use that memory. Precisely. 00:05:20.069 --> 00:05:23.370 It tells the AI exactly how to retrieve, format, 00:05:23.550 --> 00:05:25.850 and deliver that stored knowledge to you. Got 00:05:25.850 --> 00:05:28.550 it. Notebook is what it knows. Gem is how it 00:05:28.550 --> 00:05:31.250 acts. Spot on. OK. But let's pivot for a second. 00:05:31.990 --> 00:05:34.490 Stacking my existing files makes total sense. 00:05:34.990 --> 00:05:37.730 But what if I am starting at zero? Right. Blank 00:05:37.730 --> 00:05:40.490 slate. Yeah. What if I don't have a company handbook 00:05:40.490 --> 00:05:43.790 yet? Am I just stuck until I write one? Or can 00:05:43.790 --> 00:05:46.709 the AI actually help me build the knowledge base 00:05:46.709 --> 00:05:50.199 itself? Let's dive into deep research. This is 00:05:50.199 --> 00:05:53.060 where things get genuinely fun. You don't need 00:05:53.060 --> 00:05:55.139 a massive archive of files to start. You can 00:05:55.139 --> 00:05:57.060 actually use Notebook LM to go out and search 00:05:57.060 --> 00:05:59.300 the live web. Oh, wow. Yeah, it will crawl through 00:05:59.300 --> 00:06:01.300 the internet, pull the best information, and 00:06:01.300 --> 00:06:03.480 synthesize multiple sources for you automatically. 00:06:03.720 --> 00:06:05.639 But there is a trap here, right? If you just 00:06:05.639 --> 00:06:07.500 say, give me a summary of how to start a business, 00:06:07.939 --> 00:06:10.519 you are going to get a generic useless paragraph. 00:06:10.579 --> 00:06:13.519 Very generic. The golden rule here is to never 00:06:13.519 --> 00:06:16.660 ask for a general summary. You have to ask the 00:06:16.660 --> 00:06:19.740 AI to build you a framework. Yes, that is a critical 00:06:19.740 --> 00:06:23.019 distinction. A summary is passive. It is just 00:06:23.019 --> 00:06:26.779 a recap of information. A framework is a structured 00:06:26.779 --> 00:06:30.100 set of rules. Right. It is a checklist, a set 00:06:30.100 --> 00:06:33.339 of guidelines, or a decision matrix. It is an 00:06:33.339 --> 00:06:36.100 active, permanent thinking tool. Let's make this 00:06:36.100 --> 00:06:38.199 concrete. Let's use a coffee shop example. Yeah. 00:06:38.459 --> 00:06:40.199 Say you want to start a boutique coffee shop 00:06:40.199 --> 00:06:42.839 at a highly competitive, busy city. you have 00:06:42.839 --> 00:06:45.000 absolutely no idea where to begin. Instead of 00:06:45.000 --> 00:06:46.839 asking for a summary of the coffee industry, 00:06:47.060 --> 00:06:50.259 you prompt the AI for specific frameworks. You 00:06:50.259 --> 00:06:53.360 ask it to build a rigorous checklist for evaluating 00:06:53.360 --> 00:06:56.220 a retail location. You ask it to build a spreadsheet 00:06:56.220 --> 00:06:58.560 of essential equipment costs. You ask it for 00:06:58.560 --> 00:07:00.680 a step -by -step guide to interviewing and hiring 00:07:00.680 --> 00:07:03.480 baristas. Yeah. And most importantly, you ask 00:07:03.480 --> 00:07:06.139 it to hunt for red flags. What are the top 10 00:07:06.139 --> 00:07:08.300 reasons independent coffee shops fail in their 00:07:08.300 --> 00:07:10.600 first year? Right. And then the AI goes out, 00:07:10.680 --> 00:07:12.399 searches the web, reads... through dozens of 00:07:12.399 --> 00:07:15.019 articles, forums, and case studies, then it organizes 00:07:15.019 --> 00:07:18.019 all of that chaotic information into one clean, 00:07:18.500 --> 00:07:21.519 highly structured document. And here is the magic 00:07:21.519 --> 00:07:24.899 step. You take that beautiful, detailed output 00:07:24.899 --> 00:07:27.519 and you save it right back into your notebook. 00:07:27.899 --> 00:07:30.959 Whoa. I mean, imagine building a permanent, expert 00:07:30.959 --> 00:07:33.639 decision -making brain in seconds that never 00:07:33.639 --> 00:07:37.519 forgets. Two -sex silence. It really is incredible 00:07:37.519 --> 00:07:40.220 when you step back and look at it. Your AI brain 00:07:40.360 --> 00:07:44.240 now has a permanent tool. It is not just fetching 00:07:44.240 --> 00:07:46.339 research for you anymore. Right. It is actively 00:07:46.339 --> 00:07:48.420 equipped to help you evaluate your own business 00:07:48.420 --> 00:07:50.720 plan. You can bounce ideas off it, and it will 00:07:50.720 --> 00:07:52.860 run your ideas through that framework over and 00:07:52.860 --> 00:07:55.220 over again, keeping you honest. Why is a framework 00:07:55.220 --> 00:07:57.100 so much better than just asking for a summary? 00:07:57.180 --> 00:07:59.480 Because a summary just gives you a static answer 00:07:59.480 --> 00:08:02.160 to read once and forget. A framework gives you 00:08:02.160 --> 00:08:05.019 a repeatable, reliable structure to apply to 00:08:05.019 --> 00:08:07.259 any future problems you face. A summary is a 00:08:07.259 --> 00:08:09.300 one -time answer. A framework is a permanent 00:08:09.300 --> 00:08:12.259 tool. Exactly. Now we have these brilliant frameworks 00:08:12.259 --> 00:08:15.060 to help us make complex decisions, but knowing 00:08:15.060 --> 00:08:18.480 what to say is only half the battle. How do we 00:08:18.480 --> 00:08:20.980 get the AI to actually communicate those ideas 00:08:20.980 --> 00:08:23.540 out loud? The writing part. Right. How do we 00:08:23.540 --> 00:08:27.019 make it right in our unique voice without sounding 00:08:27.019 --> 00:08:29.920 like a robotic encyclopedia? Well... Writing 00:08:29.920 --> 00:08:32.639 content is rarely hard because you lack ideas. 00:08:33.159 --> 00:08:36.039 It is hard because maintaining tone consistency 00:08:36.039 --> 00:08:39.679 is exhausting. Generic AI simply doesn't know 00:08:39.679 --> 00:08:42.299 the cadence of how you naturally write. No, it 00:08:42.299 --> 00:08:44.960 doesn't. It defaults to that overly enthusiastic 00:08:44.960 --> 00:08:49.139 slightly sterile AI voice we all recognize instantly. 00:08:49.460 --> 00:08:52.299 I have a vulnerable admission to make here. Oh, 00:08:52.360 --> 00:08:54.559 lay it on me. I still wrestle with prompt drift. 00:08:54.789 --> 00:08:57.210 Myself yeah where you start with a great output 00:08:57.210 --> 00:09:00.210 But by the fifth follow -up question the AI has 00:09:00.210 --> 00:09:02.110 completely forgotten your original instructions 00:09:02.110 --> 00:09:04.629 and wandered off -topic It starts hallucinating 00:09:04.629 --> 00:09:06.830 weird vocabulary. I would never use one day. 00:09:06.830 --> 00:09:08.789 It sounds sharp the next it completely falls 00:09:08.789 --> 00:09:10.950 flat Oh, we have all been there Prompt Drift 00:09:10.950 --> 00:09:13.649 is infuriating. The fix for this is building 00:09:13.649 --> 00:09:16.490 a dedicated, highly focused writing notebook. 00:09:16.629 --> 00:09:19.029 So instead of mixing research and writing, you 00:09:19.029 --> 00:09:21.490 create a new space. You upload your top five 00:09:21.490 --> 00:09:24.090 best performing LinkedIn posts. You drop in three 00:09:24.090 --> 00:09:26.409 of your favorite, most persuasive emails. You 00:09:26.409 --> 00:09:28.389 add your company's core brand mission statement. 00:09:28.809 --> 00:09:31.909 Then you pair that specific notebook with a content 00:09:31.909 --> 00:09:35.549 writer gem. The AI physically reads your past 00:09:35.549 --> 00:09:38.509 work. It analyzes your vocabulary, your sentence 00:09:38.509 --> 00:09:42.149 length, your pacing. It essentially reverse -engineers 00:09:42.149 --> 00:09:44.789 your unique voice. Let's look at how you'd actually 00:09:44.789 --> 00:09:47.289 prompt this setup. You'd say, write a LinkedIn 00:09:47.289 --> 00:09:49.629 post about our new product launch. Right, but 00:09:49.629 --> 00:09:52.730 you add specific constraints. Check the brand 00:09:52.730 --> 00:09:55.690 guidelines document, use a hook, then the body, 00:09:55.870 --> 00:09:58.389 then a clear call -to -action structure. Keep 00:09:58.389 --> 00:10:01.750 the entire post strictly under 200 words. And 00:10:01.750 --> 00:10:03.789 make sure to mention the three main user benefits 00:10:03.789 --> 00:10:05.909 you pull from the product specs file. Because 00:10:05.909 --> 00:10:08.230 it is actively referencing your past writing, 00:10:08.590 --> 00:10:11.350 it perfectly mimics you. It pulls the hard facts 00:10:11.350 --> 00:10:13.509 from your uploaded product specs, but it applies 00:10:13.509 --> 00:10:15.830 your distinct personal writing style to deliver 00:10:15.830 --> 00:10:18.330 the message. It makes the entire content creation 00:10:18.330 --> 00:10:21.350 process so much faster. And honestly, it makes 00:10:21.350 --> 00:10:23.429 it much more enjoyable. You aren't fighting the 00:10:23.429 --> 00:10:25.669 machine anymore. Exactly. No more spending an 00:10:25.669 --> 00:10:28.250 hour doing endless line edits just to fix the 00:10:28.250 --> 00:10:30.809 robotic voice. It starts off sounding like you 00:10:30.809 --> 00:10:33.480 on draft one. Does throwing in a few old emails 00:10:33.480 --> 00:10:36.399 really change the tone that much? Absolutely. 00:10:36.519 --> 00:10:40.039 It gives the AI a literal blueprint of your cadence. 00:10:40.620 --> 00:10:42.840 It stops guessing how a professional should sound 00:10:42.840 --> 00:10:45.679 and just looks at how you actually sound. Yet 00:10:45.679 --> 00:10:48.279 seeing your past writing forces it to mimic your 00:10:48.279 --> 00:10:50.240 unique rhythm. It really does. We're going to 00:10:50.240 --> 00:10:52.139 take a quick break right here. Mid -roll sponsor 00:10:52.139 --> 00:10:55.220 reads, plays, holder, and we are back. Ready 00:10:55.220 --> 00:10:58.080 to keep going. OK, the logic is sound. The tools 00:10:58.080 --> 00:11:00.720 make sense. The frameworks are built. But let's 00:11:00.720 --> 00:11:03.779 talk about where this falls apart. How do we 00:11:03.779 --> 00:11:07.240 prevent this customized system from hallucinating 00:11:07.240 --> 00:11:09.860 or just giving us muddled, confusing answers? 00:11:10.220 --> 00:11:13.100 Setup and rigorous testing are absolutely critical 00:11:13.100 --> 00:11:15.019 here. You cannot just throw a bunch of files 00:11:15.019 --> 00:11:17.299 into a folder, skip to the finish line, and trust 00:11:17.299 --> 00:11:20.320 it blindly. I'll be honest. My very first instinct 00:11:20.320 --> 00:11:22.720 is always to just dump every single file I own 00:11:22.720 --> 00:11:25.220 into one massive notebook. Oh, no. I just want 00:11:25.220 --> 00:11:26.980 one mega brain that knows my whole life. Why 00:11:26.980 --> 00:11:28.419 shouldn't I do that? Then not do that. I cannot 00:11:28.419 --> 00:11:30.559 stress this enough. That is a massive mistake. 00:11:30.919 --> 00:11:34.460 When the AI processes entirely unrelated topics 00:11:34.460 --> 00:11:37.240 in the same space, the vector searches cross 00:11:37.240 --> 00:11:40.059 contaminate. The answers get muddled. So if I 00:11:40.059 --> 00:11:42.110 ask for marketing copy, might accidentally pull 00:11:42.110 --> 00:11:44.409 in the passive aggressive tone from an old HR 00:11:44.409 --> 00:11:47.250 complaint I uploaded? Exactly. The AI tries to 00:11:47.250 --> 00:11:49.990 synthesize everything it sees. So separation 00:11:49.990 --> 00:11:54.230 is key. You need a strict boundary. One notebook 00:11:54.230 --> 00:11:58.409 per project or one notebook per specific role. 00:11:58.909 --> 00:12:01.570 Do not mix your new coffee shop business plans 00:12:01.570 --> 00:12:03.629 with your personal daily journal. That makes 00:12:03.629 --> 00:12:06.200 sense. Also, after you upload your files, you 00:12:06.200 --> 00:12:08.480 need to actually check the auto -generated AI 00:12:08.480 --> 00:12:10.580 summaries of those files. Yes, very important. 00:12:10.740 --> 00:12:12.980 If the AI's initial summary of a document is 00:12:12.980 --> 00:12:15.379 fundamentally off, every answer it gives you 00:12:15.379 --> 00:12:18.220 based on that document will be off. You can actually 00:12:18.220 --> 00:12:20.840 add your own inline notes to signal to the AI 00:12:20.840 --> 00:12:23.240 what is truly important. Then we have to talk 00:12:23.240 --> 00:12:26.000 about GEM specificity. Vague instructions in 00:12:26.000 --> 00:12:29.049 your GEM will fail every time. Telling a gem 00:12:29.049 --> 00:12:31.389 to simply be a helpful assistant means absolutely 00:12:31.389 --> 00:12:33.730 nothing to an AI. You have to be hyper -specific. 00:12:33.909 --> 00:12:35.669 Let's use a real estate example to illustrate 00:12:35.669 --> 00:12:37.929 this. Perfect. You don't just say, help me find 00:12:37.929 --> 00:12:41.210 houses. You tell the gem its precise role. You 00:12:41.210 --> 00:12:43.350 are an expert real estate research assistant. 00:12:43.830 --> 00:12:46.470 Your primary job is looking for undervalued deals 00:12:46.470 --> 00:12:49.210 for a family of four. You tell it exactly what 00:12:49.210 --> 00:12:52.190 metrics matter. Focus strictly on price per square 00:12:52.190 --> 00:12:54.690 foot. Focus heavily on the physical distance 00:12:54.690 --> 00:12:56.789 to top -rated public schools. And then you give 00:12:56.789 --> 00:12:59.669 it a hard, specific rule to govern its output. 00:12:59.929 --> 00:13:01.970 Always present your findings in a simple, readable 00:13:01.970 --> 00:13:04.470 table. And if a property does not have a fence 00:13:04.470 --> 00:13:06.769 backyard, you must mark it with a yellow flag. 00:13:07.289 --> 00:13:09.590 Beat. I love that level of control. The more 00:13:09.590 --> 00:13:11.269 granular and specific you are with the rules, 00:13:11.370 --> 00:13:13.950 the better the output. But even with all that, 00:13:14.129 --> 00:13:16.370 you still cannot just trust the AI right out 00:13:16.370 --> 00:13:18.429 of the gate. You have to verify it. You have 00:13:18.429 --> 00:13:21.169 to stress test the agent before you rely on it 00:13:21.169 --> 00:13:23.750 for real work. Test number one is strictly for 00:13:23.750 --> 00:13:25.769 grounding. Right. So you intentionally ask the 00:13:25.769 --> 00:13:29.169 AI for a highly specific detail, a tiny detail 00:13:29.169 --> 00:13:32.149 that you know is buried deep within page 42 of 00:13:32.149 --> 00:13:35.169 a massive PDF you uploaded. And you explicitly 00:13:35.169 --> 00:13:38.190 demand an exact citation for the answer. If the 00:13:38.190 --> 00:13:40.610 AI pulls the detail and psychs the source correctly, 00:13:40.830 --> 00:13:43.269 your grounding is working perfectly. But if it 00:13:43.269 --> 00:13:45.309 gives you a vague answer that sounds like a generic 00:13:45.309 --> 00:13:48.090 Wikipedia entry, your gem needs stricter rules. 00:13:48.570 --> 00:13:51.029 It is drifting back to the internet. Then you 00:13:51.029 --> 00:13:53.669 run test number two, which is for reasoning. 00:13:54.409 --> 00:13:57.129 You ask the AI to compare two competing concepts 00:13:57.129 --> 00:14:00.830 from your documents, idea A and idea B. But the 00:14:00.830 --> 00:14:03.669 trick is, you instruct it to base the comparison 00:14:03.669 --> 00:14:06.190 only on your uploaded notes. For example, ask 00:14:06.190 --> 00:14:08.269 it to compare the two ideas strictly based on 00:14:08.269 --> 00:14:11.490 implementation cost or deployment speed. If the 00:14:11.490 --> 00:14:14.590 answer is logical, structured, and clearly draws 00:14:14.590 --> 00:14:17.230 only from the nuances in your documents, the 00:14:17.230 --> 00:14:20.450 system works. And look, if the tests fail, don't 00:14:20.450 --> 00:14:23.700 panic. It doesn't mean the AI is broken. Usually, 00:14:23.779 --> 00:14:26.059 it just means you need clearer headings in your 00:14:26.059 --> 00:14:28.659 original source documents, or you need to simplify 00:14:28.659 --> 00:14:30.580 your gem instructions so they aren't contradicting 00:14:30.580 --> 00:14:33.279 each other. What's the immediate tell that the 00:14:33.279 --> 00:14:35.600 AI has ignored my files and is just guessing? 00:14:35.799 --> 00:14:38.340 It'll give you a really smooth, confident, generalized 00:14:38.340 --> 00:14:40.840 overview, but it won't be able to point to a 00:14:40.840 --> 00:14:43.139 single page or paragraph from your uploads. If 00:14:43.139 --> 00:14:45.639 it cannot cite the exact page, it's just making 00:14:45.639 --> 00:14:47.779 things up. That is the ultimate red flag. Always 00:14:47.779 --> 00:14:49.820 look for the receipts. Okay, let's look at the 00:14:49.820 --> 00:14:52.529 bigger picture. If you put in the work, you get 00:14:52.529 --> 00:14:54.470 the setup right, you separate your notebooks, 00:14:54.870 --> 00:14:57.629 and you rigorously test the output, what actually 00:14:57.629 --> 00:15:00.769 happens next? How did this change a random Tuesday 00:15:00.769 --> 00:15:03.389 morning six months from now? It completely transforms 00:15:03.389 --> 00:15:06.669 the friction of your daily workflow. Every morning 00:15:06.669 --> 00:15:09.190 you just open your gem. You don't prep it, you 00:15:09.190 --> 00:15:12.090 don't brief it. It is just ready. There is no 00:15:12.090 --> 00:15:15.039 more reloading context. No more typing out long 00:15:15.039 --> 00:15:17.299 paragraphs re -explaining the goals of the project 00:15:17.299 --> 00:15:19.080 you were working on yesterday. You just jump 00:15:19.080 --> 00:15:21.940 straight into the actual work. It entirely solves 00:15:21.940 --> 00:15:25.100 the blank page problem. You can instantly ask 00:15:25.100 --> 00:15:27.539 your customized assistant for a structured outline 00:15:27.539 --> 00:15:29.580 based on the messy research notes you took the 00:15:29.580 --> 00:15:31.899 night before. You can ask it to actively flag 00:15:31.899 --> 00:15:34.379 blind spots or gaps in your logic. You start 00:15:34.379 --> 00:15:36.039 the work with the heavy lifting already done. 00:15:36.379 --> 00:15:38.559 You're starting at the 80 % mark. And then you 00:15:38.559 --> 00:15:41.080 experience the compound effect. The agent actually 00:15:41.080 --> 00:15:43.440 gets smarter and more aligned with you over time. 00:15:43.759 --> 00:15:46.159 Because every new file you upload, every new 00:15:46.159 --> 00:15:48.659 piece of feedback you drop into the notebook... 00:15:48.620 --> 00:15:52.320 deepens the context. It increases the AI's accuracy. 00:15:52.659 --> 00:15:55.659 It drastically reduces the time you spend fixing 00:15:55.659 --> 00:15:58.059 those little annoying errors. It literally grows 00:15:58.059 --> 00:16:00.279 alongside you. The longer you use the specific 00:16:00.279 --> 00:16:02.820 system, the more deeply it understands the weird 00:16:02.820 --> 00:16:06.220 nuances of your specific business. Let's hit 00:16:06.220 --> 00:16:09.299 some rapid fire listener FAQs, because whenever 00:16:09.299 --> 00:16:11.740 we talk about uploading personal data, red flags 00:16:11.740 --> 00:16:13.580 go up. People always ask about privacy first. 00:16:14.080 --> 00:16:17.250 If I upload my started business plan, Is my data 00:16:17.250 --> 00:16:20.649 truly safe? Yes. Google has explicitly stated 00:16:20.649 --> 00:16:23.450 they do not use your private notebook LM data 00:16:23.450 --> 00:16:26.379 to train their global public models. Your data 00:16:26.379 --> 00:16:28.940 stays in your instance. It is completely safe 00:16:28.940 --> 00:16:31.100 for sensitive business plans and proprietary 00:16:31.100 --> 00:16:33.340 research. What are the actual technical limits 00:16:33.340 --> 00:16:34.960 on the uploads? Am I going to hit a wall fast? 00:16:35.100 --> 00:16:37.799 Not likely. You can have up to 300 individual 00:16:37.799 --> 00:16:40.220 sources per single NURT book. And each of those 00:16:40.220 --> 00:16:42.340 sources can be up to 500 ,000 words. That is 00:16:42.340 --> 00:16:44.480 massive. That is a massive, massive amount of 00:16:44.480 --> 00:16:46.600 data. You could upload the entire Harry Potter 00:16:46.600 --> 00:16:48.700 series several times over into one notebook. 00:16:49.159 --> 00:16:52.539 What does this kind of power actually cost? Currently, 00:16:52.860 --> 00:16:55.179 Notebook LM is completely free while it's in 00:16:55.179 --> 00:16:57.639 its testing phase. Some of the more advanced 00:16:57.639 --> 00:17:00.440 Gemini features might require a premium subscription 00:17:00.440 --> 00:17:03.120 later on down the line, but right now the barrier 00:17:03.120 --> 00:17:06.579 to entry is zero. Can I collaborate with my remote 00:17:06.579 --> 00:17:08.660 team on this, or is it strictly single player? 00:17:08.779 --> 00:17:10.779 Absolutely. You can collaborate. You can share 00:17:10.779 --> 00:17:13.000 a fully loaded notebook exactly like you'd share 00:17:13.000 --> 00:17:15.700 a Google Doc. Everyone on the team gets access 00:17:15.700 --> 00:17:18.339 to the exact same brain, which is brilliant for 00:17:18.339 --> 00:17:20.380 keeping a remote team perfectly aligned on the 00:17:20.380 --> 00:17:23.160 facts. And finally, what happens when a project 00:17:23.160 --> 00:17:26.319 changes and I delete an old, outdated file? That's 00:17:26.319 --> 00:17:29.059 how you keep the brain clean and accurate. Deleting 00:17:29.059 --> 00:17:32.019 a source file causes the AI to instantly forget 00:17:32.019 --> 00:17:34.839 that information. There is no residual memory 00:17:34.839 --> 00:17:37.140 muddying up future answers. So if I delete a 00:17:37.140 --> 00:17:39.839 business plan, the AI completely wipes it from 00:17:39.839 --> 00:17:42.640 memory. Instantly. The second that PDF is gone, 00:17:43.000 --> 00:17:45.359 the AI's knowledge of it completely vanishes. 00:17:45.539 --> 00:17:48.220 Exactly. Deleting a file instantly scrubs it 00:17:48.220 --> 00:17:51.099 from the AI's brain. It gives you total granular 00:17:51.099 --> 00:17:54.059 control over exactly what the system knows at 00:17:54.059 --> 00:17:57.240 any given moment. Two -second silence. Lift your 00:17:57.240 --> 00:17:59.940 breath and recap the really big idea here today. 00:18:00.559 --> 00:18:04.160 We are crossing a major historical bridge right 00:18:04.160 --> 00:18:07.599 now. We are moving from simply acting as consumers 00:18:07.599 --> 00:18:10.799 who just use a generic AI chatbot to acting as 00:18:10.799 --> 00:18:13.259 architects who are building our own highly specialized 00:18:13.259 --> 00:18:16.200 AI systems. And we are doing it entirely in plain 00:18:16.200 --> 00:18:18.140 English without writing a single line of code. 00:18:18.660 --> 00:18:21.460 It takes a neat parlor trick, a basic tool, and 00:18:21.460 --> 00:18:23.980 turns it into a true reliable partnership. It 00:18:23.980 --> 00:18:26.420 is incredibly empowering once you get it set 00:18:26.420 --> 00:18:28.619 up. It changes how you view your own capabilities. 00:18:28.880 --> 00:18:30.700 So here's our call to action for you this week. 00:18:30.920 --> 00:18:33.220 Don't sit around waiting for the perfect comprehensive 00:18:33.220 --> 00:18:36.220 master plan. Pick one small annoying part of 00:18:36.220 --> 00:18:38.839 your job today, make a dedicated notebook for 00:18:38.839 --> 00:18:42.099 it, upload five relevant files, ask it one question, 00:18:42.339 --> 00:18:44.380 and just see how it feels to have an AI that 00:18:44.380 --> 00:18:46.380 actually knows what you're talking about. Start 00:18:46.380 --> 00:18:48.720 small, experiment with the gems, and the results 00:18:48.720 --> 00:18:51.019 will surprise you immediately. Before we wrap 00:18:51.019 --> 00:18:52.779 up, I want to leave you with one final thought, 00:18:52.880 --> 00:18:54.400 something to really chew on as you start building 00:18:54.400 --> 00:18:57.460 these out. If your customized AI has eventually 00:18:57.460 --> 00:18:59.619 read every important document you have ever written, 00:18:59.789 --> 00:19:02.329 And it perfectly understands your personal frameworks, 00:19:02.809 --> 00:19:05.230 your business goals, and your unique tone of 00:19:05.230 --> 00:19:07.809 voice. How long until it starts connecting dots 00:19:07.809 --> 00:19:11.210 across years of your own work and starts having 00:19:11.210 --> 00:19:13.869 brilliant original ideas that you wish you had 00:19:13.869 --> 00:19:16.329 thought of first? That is the ultimate compound 00:19:16.329 --> 00:19:18.250 effect in action. It stops being an assistant. 00:19:18.269 --> 00:19:20.549 It starts being a collaborator. It is a totally 00:19:20.549 --> 00:19:23.549 new frontier. And the absolute best part, you 00:19:23.549 --> 00:19:26.009 will never have to cure its amnesia ever again.