WEBVTT 00:00:00.000 --> 00:00:02.759 Imagine for a second that you could take every 00:00:02.759 --> 00:00:05.580 single Harry Potter book, all seven of them, 00:00:06.059 --> 00:00:09.000 stack them up, hand them to someone. Now imagine 00:00:09.000 --> 00:00:11.619 that person could read every single word just 00:00:11.619 --> 00:00:14.490 instantly. But more importantly, imagine that 00:00:14.490 --> 00:00:16.510 while they're reading the very last sentence 00:00:16.510 --> 00:00:19.370 of the last book, they still perfectly remember 00:00:19.370 --> 00:00:21.670 the first sentence of the first book. That is 00:00:21.670 --> 00:00:24.429 the scale we're talking about here. It's a staggering 00:00:24.429 --> 00:00:26.390 amount of information to hold in your head at 00:00:26.390 --> 00:00:28.750 once, without getting fuzzy on any of the details. 00:00:28.890 --> 00:00:31.510 It really is. And that concept, that ability 00:00:31.510 --> 00:00:33.810 to hold all that information without losing the 00:00:33.810 --> 00:00:36.130 thread, is technically called a context window. 00:00:36.490 --> 00:00:38.509 It's pretty central to what we're exploring today. 00:00:39.450 --> 00:00:42.469 Welcome back to the Deep Dive. Today we are unpacking 00:00:42.469 --> 00:00:44.670 Google Gemini Advanced. It's good to be here. 00:00:44.750 --> 00:00:47.450 This is a big one. It is, yeah. And looking at 00:00:47.450 --> 00:00:49.170 the numbers, the interest is definitely there. 00:00:49.350 --> 00:00:51.850 The user base for Gemini has, what, quadrupled 00:00:51.850 --> 00:00:54.429 in just a year? That's massive growth. But digging 00:00:54.429 --> 00:00:57.240 into the reports, there's this... this interesting 00:00:57.240 --> 00:00:59.340 tension. We have millions of people using these 00:00:59.340 --> 00:01:02.659 tools, but there's a massive gap between knowing 00:01:02.659 --> 00:01:05.319 a feature exists, like, oh, I heard it can read 00:01:05.319 --> 00:01:07.439 a PDF, and actually weaving that feature into 00:01:07.439 --> 00:01:09.939 a daily workflow. That's the classic adoption 00:01:09.939 --> 00:01:12.780 curve, right? We treat these tools like novelties 00:01:12.780 --> 00:01:16.159 or just really big search engines. We ask a question. 00:01:16.159 --> 00:01:19.219 We get an answer. It's transactional. But the 00:01:19.219 --> 00:01:21.219 shift we want to talk about today and what the 00:01:21.219 --> 00:01:23.859 source material really emphasizes is moving from 00:01:23.859 --> 00:01:27.569 just chatting with a bot to building a persistent 00:01:27.569 --> 00:01:29.870 personal assistant. Which is a heavy promise. 00:01:30.090 --> 00:01:32.030 So to keep us on track, we have a bit of a roadmap. 00:01:32.329 --> 00:01:34.109 We're going to look at three specific pillars 00:01:34.109 --> 00:01:37.030 from the documentation. First, multimodality, 00:01:37.170 --> 00:01:38.909 which is a fancy word we will definitely break 00:01:38.909 --> 00:01:42.090 down. Second, the new thinking model versus the 00:01:42.090 --> 00:01:44.489 fast model. And third, creating these custom 00:01:44.489 --> 00:01:46.870 gems. And we'll wrap it all up with a master 00:01:46.870 --> 00:01:50.189 class example, a gardening workflow that actually 00:01:50.189 --> 00:01:53.870 ties all these disparate pieces together into 00:01:53.870 --> 00:01:55.939 something really practical. So let's start with 00:01:55.939 --> 00:01:59.560 that first pillar, multimodality. To me, this 00:01:59.560 --> 00:02:01.959 often sounds a bit like marketing jargon. When 00:02:01.959 --> 00:02:04.159 you strip it back, what are we actually talking 00:02:04.159 --> 00:02:07.439 about here in terms of utility? It sounds complex, 00:02:07.439 --> 00:02:10.000 but it's actually the most human part of the 00:02:10.000 --> 00:02:12.479 AI. I mean, think about how you and I interact 00:02:12.479 --> 00:02:14.219 with the world right now. We don't just read 00:02:14.219 --> 00:02:16.960 text floating in a void. We see things. We hear 00:02:16.960 --> 00:02:20.039 sounds. We watch movement. Multimodality just 00:02:20.039 --> 00:02:22.930 means the AI can do that too. It isn't limited 00:02:22.930 --> 00:02:25.229 to text anymore. It can process photos. It can 00:02:25.229 --> 00:02:28.050 watch videos. It can listen to audio. So it's 00:02:28.050 --> 00:02:30.849 effectively breaking that text -only barrier. 00:02:31.150 --> 00:02:33.650 Exactly. And the efficiency gain here is massive. 00:02:33.849 --> 00:02:36.169 Think about the friction of trying to describe 00:02:36.169 --> 00:02:38.949 a photo to a computer. You're typing, OK, there's 00:02:38.949 --> 00:02:40.610 a tree in the left corner, and the light is hitting 00:02:40.610 --> 00:02:43.069 it this way. It's just tedious. With multimodality, 00:02:43.250 --> 00:02:45.689 you don't describe the context. You just show 00:02:45.689 --> 00:02:48.229 it. You just show it. And this ties back to that 00:02:48.229 --> 00:02:51.259 opening hook we had, the context window. The 00:02:51.259 --> 00:02:53.620 source material mentions Gemini Advanced has 00:02:53.620 --> 00:02:57.199 a capacity of 1 million tokens. We use the Harry 00:02:57.199 --> 00:02:59.719 Potter analogy, which is vivid, but technically, 00:02:59.919 --> 00:03:02.979 what is a token? Good catch. A token is basically 00:03:02.979 --> 00:03:06.319 a unit of text, roughly four characters. So a 00:03:06.319 --> 00:03:11.120 million tokens is about 700 ,000 words. Which 00:03:11.120 --> 00:03:13.520 is hard to wrap your head around. It is. But 00:03:13.520 --> 00:03:15.680 practically, it solves the goldfish memory problem. 00:03:16.250 --> 00:03:18.949 Most older models, if you gave them a long document 00:03:18.949 --> 00:03:20.629 by page 50, they've kind of forgotten what was 00:03:20.629 --> 00:03:23.229 on page one. With a million tokens, you can dump 00:03:23.229 --> 00:03:25.870 a massive amount of data in, and it retains high 00:03:25.870 --> 00:03:27.810 -fidelity recall from start to finish. Let's 00:03:27.810 --> 00:03:29.650 look at a concrete example from the research 00:03:29.650 --> 00:03:31.370 to see if this holds up. Say I have to learn 00:03:31.370 --> 00:03:34.009 about quantum computing, and I have a dense, 00:03:34.330 --> 00:03:38.289 79 -page PDF. It's dry, it's academic, it's painful. 00:03:38.430 --> 00:03:40.469 The nightmare scenario. Right. So I drag that 00:03:40.469 --> 00:03:42.569 PDF into Gemini, and because of that massive 00:03:42.569 --> 00:03:44.889 context window, I can ask it, explain this to 00:03:44.889 --> 00:03:46.590 me like I'm a five -year -old. And it works. 00:03:46.750 --> 00:03:48.909 It synthesizes the whole thing. But here's my 00:03:48.909 --> 00:03:51.949 pushback on that. It reads the text. It summarizes 00:03:51.949 --> 00:03:54.590 it. That's great for saving time. But how does 00:03:54.590 --> 00:03:57.289 that actually help us learn a complex topic? 00:03:57.770 --> 00:03:59.949 Because just reading a summary isn't the same 00:03:59.949 --> 00:04:01.990 as understanding the mechanics of it. That's 00:04:01.990 --> 00:04:04.509 a crucial distinction. And if you just ask for 00:04:04.509 --> 00:04:07.110 a summary, you are only getting surface -level 00:04:07.110 --> 00:04:11.069 info. The power is in interrogation. The source 00:04:11.069 --> 00:04:13.889 material suggests asking the AI to transform 00:04:13.889 --> 00:04:17.069 that static text into interactive tools. Interactive 00:04:17.069 --> 00:04:19.870 tools? Yeah. You aren't just getting a book report 00:04:19.870 --> 00:04:22.310 back. You can ask it to generate an infographic 00:04:22.310 --> 00:04:25.089 based on the text. Or, and this is where it gets 00:04:25.089 --> 00:04:27.709 wild, you can ask it to create a simple interactive 00:04:27.709 --> 00:04:30.750 simulation. Code, essentially. It lets you play 00:04:30.750 --> 00:04:33.170 with the variables mentioned in the PDF. It turns 00:04:33.170 --> 00:04:36.050 passive reading into active simulation. So I 00:04:36.050 --> 00:04:38.310 could ask... Based on this paper, write me a 00:04:38.310 --> 00:04:40.569 Python script that simulates particle spin. And 00:04:40.569 --> 00:04:43.050 it'll do it. You run the code, and suddenly you're 00:04:43.050 --> 00:04:45.189 seeing the concept in action rather than just 00:04:45.189 --> 00:04:47.730 reading about it. It's bridging that gap between 00:04:47.730 --> 00:04:50.529 information and real understanding. That's fascinating. 00:04:50.709 --> 00:04:52.350 OK, so we've got text and documents covered, 00:04:52.610 --> 00:04:54.910 but I want to pivot to the second pillar, which 00:04:54.910 --> 00:04:57.389 I think is even more mind -bending. The ability 00:04:57.389 --> 00:05:00.889 to watch. Video analysis. The source makes a 00:05:00.889 --> 00:05:03.529 really crucial distinction here. When we say 00:05:03.529 --> 00:05:05.829 Gemini watches a video, we aren't just saying 00:05:05.829 --> 00:05:07.930 it reads the automated captions, are we? Because 00:05:07.930 --> 00:05:09.709 I can do that. I can just read a transcript. 00:05:09.970 --> 00:05:12.490 Correct. And that is the game changer. Most tools 00:05:12.490 --> 00:05:16.910 just scrape the transcript text. Gemini is processing 00:05:16.910 --> 00:05:20.089 the pixels. It's using its native vision capabilities, 00:05:21.110 --> 00:05:23.610 analyzing the visual information frame by frame. 00:05:23.850 --> 00:05:26.629 I have to pause on that because... Just imagine 00:05:26.629 --> 00:05:28.730 the implication. It's not just hearing what is 00:05:28.730 --> 00:05:31.310 said. It's seeing the editing speed. It's seeing 00:05:31.310 --> 00:05:34.110 the color grading, the facial expressions. It 00:05:34.110 --> 00:05:35.870 captures the whole texture of the video, not 00:05:35.870 --> 00:05:37.910 just the content. So let's look at a use case. 00:05:38.209 --> 00:05:40.750 The guide mentions learning from YouTube. Say 00:05:40.750 --> 00:05:42.670 I want to make a video and I find a tutorial 00:05:42.670 --> 00:05:46.110 on how to make viral AI videos, but it's super 00:05:46.110 --> 00:05:48.529 fast paced. It's chaotic. Read the kind where 00:05:48.529 --> 00:05:50.129 you're pausing every two seconds just trying 00:05:50.129 --> 00:05:52.750 to catch up. Exactly. So you can paste that link 00:05:52.750 --> 00:05:55.430 into Gemini. What happens next? You just say, 00:05:55.750 --> 00:05:58.670 watch this and write a prompt for me to recreate 00:05:58.670 --> 00:06:01.589 this exact style. Because it's watching the pixels, 00:06:01.990 --> 00:06:04.550 it understands the visual language, the pacing, 00:06:04.829 --> 00:06:06.910 the aesthetic, and it gives you a command you 00:06:06.910 --> 00:06:09.410 can use to replicate it. It's like having a director 00:06:09.410 --> 00:06:12.029 analyzing the film for you. There was another 00:06:12.029 --> 00:06:13.689 example for creators that I thought was really 00:06:13.689 --> 00:06:17.149 sharp. If you're a YouTuber or even just doing 00:06:17.149 --> 00:06:19.629 marketing videos for work, you can upload your 00:06:19.629 --> 00:06:22.439 top five performing videos. This is where it 00:06:22.439 --> 00:06:24.699 becomes an analyst. You ask it, look at these 00:06:24.699 --> 00:06:27.420 five videos. Why did they succeed? And because 00:06:27.420 --> 00:06:30.100 it sees the video, it doesn't just say, oh, the 00:06:30.100 --> 00:06:32.779 topic was good. It looks at the structure. It 00:06:32.779 --> 00:06:35.680 might say, in all five videos, you used a rapid 00:06:35.680 --> 00:06:38.480 visual hook in the first three seconds. Or your 00:06:38.480 --> 00:06:40.959 editing pace accelerated at the one minute mark, 00:06:41.240 --> 00:06:43.779 which kept retention high. It's spotting patterns 00:06:43.779 --> 00:06:45.600 that you might not even realize you're doing. 00:06:45.800 --> 00:06:48.500 Precisely. It's objective feedback based on visual 00:06:48.500 --> 00:06:51.800 data. OK, now. This works great for immediate 00:06:51.800 --> 00:06:56.040 analysis, checking a few videos, a PDF. But what 00:06:56.040 --> 00:06:58.920 if I have a year's worth of scripts? What if 00:06:58.920 --> 00:07:01.459 I have a huge archive of data that I don't want 00:07:01.459 --> 00:07:03.639 to re -upload every single time? The context 00:07:03.639 --> 00:07:05.980 window is big, but it's not infinite. That's 00:07:05.980 --> 00:07:07.759 what you need to change tools. You connect it 00:07:07.759 --> 00:07:10.439 to Notebook LM to give the AI that long -term 00:07:10.439 --> 00:07:13.519 memory of your entire creative history. Notebook 00:07:13.519 --> 00:07:15.860 LM is often mentioned alongside Gemini. Can you 00:07:15.860 --> 00:07:18.199 just clarify the relationship there? Think of 00:07:18.199 --> 00:07:20.560 Gemini as the processor. the brain that's thinking 00:07:20.560 --> 00:07:23.139 right now. Think of Notebook LM as the library. 00:07:23.740 --> 00:07:26.600 You store your deep archives, hundreds of PDFs, 00:07:26.699 --> 00:07:29.819 old scripts, research notes in Notebook LM. Then 00:07:29.819 --> 00:07:31.740 when you're chatting in Gemini, you can reference 00:07:31.740 --> 00:07:33.639 that library without having to re -upload it. 00:07:33.939 --> 00:07:36.839 It grounds the AI in your specific history. Okay, 00:07:36.939 --> 00:07:38.600 moving on. I want to talk about something that 00:07:38.600 --> 00:07:41.500 I think everyone can relate to. The pain of repetition. 00:07:41.819 --> 00:07:44.339 Oh, absolutely. I still wrestle with this myself. 00:07:44.860 --> 00:07:46.800 I find a prompt that works, but then I have to 00:07:46.800 --> 00:07:49.560 type it out or copy paste it from a note. And 00:07:49.560 --> 00:07:52.259 I have to remind the AI of the formatting rules 00:07:52.259 --> 00:07:55.560 every single time. Don't use bullet points. Use 00:07:55.560 --> 00:07:59.279 a table. Don't do this. Do that. It's just exhausting. 00:07:59.540 --> 00:08:01.939 It's the friction that kills adoption. If it 00:08:01.939 --> 00:08:04.819 feels like work to ask the AI to do the work, 00:08:05.139 --> 00:08:07.750 you just stop doing it. So enter gems. Gems. 00:08:07.949 --> 00:08:10.209 These are described as custom versions of the 00:08:10.209 --> 00:08:13.269 AI, but how is that different from just, you 00:08:13.269 --> 00:08:16.370 know, a saved chat? Think of a gem as a preset 00:08:16.370 --> 00:08:19.170 persona. It's a version of Gemini where you have 00:08:19.170 --> 00:08:22.149 preloaded a specific set of system instructions 00:08:22.149 --> 00:08:24.730 that it never ever forgets. You're effectively 00:08:24.730 --> 00:08:26.930 programming the AI, but you're using natural 00:08:26.930 --> 00:08:28.970 language instead of code. The example in the 00:08:28.970 --> 00:08:31.589 deep dive material is the accountant gem. I think 00:08:31.589 --> 00:08:33.490 this illustrates it perfectly. It's so simple, 00:08:33.549 --> 00:08:35.909 but so effective. The problem is universal. You 00:08:35.909 --> 00:08:37.210 come back from a business trip, your pockets 00:08:37.210 --> 00:08:39.710 are full of crumpled receipts. It's a mess. The 00:08:39.710 --> 00:08:42.190 worst part of any trip. Easily. So the old way 00:08:42.190 --> 00:08:45.149 is, you take a photo, upload it, and you type 00:08:45.149 --> 00:08:47.909 a paragraph. Please extract the dates, amounts, 00:08:47.990 --> 00:08:51.110 and vendors. Put them in a table, categorize 00:08:51.110 --> 00:08:53.149 them. You have to type that every time. Which 00:08:53.149 --> 00:08:55.629 is so annoying. Right. With a gem, you build 00:08:55.629 --> 00:08:58.529 it once. You call it Expense Helper. And inside 00:08:58.529 --> 00:09:00.889 that gem, you write those instructions one time. 00:09:01.120 --> 00:09:04.440 Take numbers, format into an Excel table, categorize 00:09:04.440 --> 00:09:08.019 it to food, travel, hotel. Do not wait for me 00:09:08.019 --> 00:09:10.659 to ask. So the next time I open that gem, you 00:09:10.659 --> 00:09:12.980 drop the photo, you say absolutely nothing. Silence. 00:09:13.259 --> 00:09:15.820 Silence. It just does the job. It knows its role. 00:09:16.039 --> 00:09:17.600 It's like walking into your office and handing 00:09:17.600 --> 00:09:19.220 a file to an assistant who's worked with you 00:09:19.220 --> 00:09:21.620 for 10 years. You don't need to explain what 00:09:21.620 --> 00:09:23.139 to do with it. They already know the protocol. 00:09:23.259 --> 00:09:26.179 That is the shift from prompting every time to 00:09:26.179 --> 00:09:28.759 delegating to a system that you've built. Exactly. 00:09:29.019 --> 00:09:31.320 And you could do this for anything. a coding 00:09:31.320 --> 00:09:34.240 buddy that always comments your code, a writing 00:09:34.240 --> 00:09:36.059 editor that always checks for passive voice. 00:09:36.460 --> 00:09:38.480 You build the tool once, you use it forever. 00:09:38.919 --> 00:09:41.240 But, okay, let's say we have the data organized. 00:09:41.639 --> 00:09:43.659 The receipts are in a table, but that's just 00:09:43.659 --> 00:09:46.960 data entry. Where do we actually do the work 00:09:46.960 --> 00:09:49.919 of refining it? Where do we turn that data into, 00:09:49.919 --> 00:09:53.059 say, a report or a script? We move to Canvas, 00:09:53.919 --> 00:09:55.940 which the documentation describes as a hybrid 00:09:55.940 --> 00:09:58.399 between a Google Doc and a code editor. Canvas 00:09:58.399 --> 00:10:01.139 seems to be the answer to the chat interface 00:10:01.139 --> 00:10:03.720 problem because chatting is actually terrible 00:10:03.720 --> 00:10:06.299 for editing documents. It is. Chat is linear. 00:10:06.580 --> 00:10:09.940 You ask, it answers. If you want to change one 00:10:09.940 --> 00:10:11.659 sentence in the middle of a generated email, 00:10:12.000 --> 00:10:13.820 you usually have to ask it to rewrite the whole 00:10:13.820 --> 00:10:15.679 thing and then hope it doesn't mess up the parts 00:10:15.679 --> 00:10:18.289 you liked. Or you just copy paste it into Word 00:10:18.289 --> 00:10:20.549 and edit it yourself. Right. Canvas solves that. 00:10:20.669 --> 00:10:22.429 It opens a separate window right next to the 00:10:22.429 --> 00:10:24.009 chat. So you have your conversation on the left 00:10:24.009 --> 00:10:26.049 and the actual document on the right. So it looks 00:10:26.049 --> 00:10:28.750 more like a document editor. Yes. So let's say 00:10:28.750 --> 00:10:30.649 you're writing a video script based on those 00:10:30.649 --> 00:10:33.049 receipts. You generate the draft. It appears 00:10:33.049 --> 00:10:35.769 in the Canvas window. Now, you can highlight 00:10:35.769 --> 00:10:38.129 just the introduction and type, make this punchier, 00:10:38.590 --> 00:10:40.730 or highlight a technical paragraph and say, remove 00:10:40.730 --> 00:10:43.769 the jargon. It's surgical. It is. And crucially, 00:10:44.110 --> 00:10:46.210 you aren't losing the context of the rest of 00:10:46.210 --> 00:10:48.610 the document. It's acting like a collaborative 00:10:48.610 --> 00:10:51.129 editor sitting next to you rather than a bot 00:10:51.129 --> 00:10:54.590 just throwing text at you. Now, within this ecosystem, 00:10:54.870 --> 00:10:57.210 there's also a distinction made between the models 00:10:57.210 --> 00:10:59.889 themselves. We have the fast model and the thinking 00:10:59.889 --> 00:11:02.309 model. I feel like most people just leave it 00:11:02.309 --> 00:11:05.149 on default and never touch this. Which is a mistake. 00:11:06.090 --> 00:11:08.210 Because they are fundamentally different tools. 00:11:08.269 --> 00:11:12.000 How so? The fast model is... Well, it's fast. 00:11:12.460 --> 00:11:15.019 It uses what psychologists call system one thinking. 00:11:15.440 --> 00:11:18.580 It's instinctive, rapid, pattern matching. Great 00:11:18.580 --> 00:11:21.220 for brainstorming or quick chats. But the thinking 00:11:21.220 --> 00:11:23.879 model uses system two. System two being that 00:11:23.879 --> 00:11:26.259 slow, deliberative, logical part of the brain. 00:11:26.419 --> 00:11:28.559 Exactly. When you switch to the thinking model, 00:11:28.879 --> 00:11:31.700 you'll actually notice the AI pauses. It might 00:11:31.700 --> 00:11:33.919 take 10 or 15 seconds before it even starts typing. 00:11:34.179 --> 00:11:36.580 What's it doing during that pause? It's reasoning. 00:11:37.279 --> 00:11:39.990 It's mapping out a chain of thought. If you ask 00:11:39.990 --> 00:11:42.830 it a complex logic puzzle or a math problem, 00:11:43.370 --> 00:11:46.429 or to plan a travel itinerary with five different 00:11:46.429 --> 00:11:49.529 constraints, the thinking model maps out the 00:11:49.529 --> 00:11:52.210 steps before it generates the final answer. It 00:11:52.210 --> 00:11:54.629 actually checks its own work. The guide compares 00:11:54.629 --> 00:11:57.070 it to choosing the right tool to fix a car. You 00:11:57.070 --> 00:11:59.509 don't use a hammer for everything. Exactly. My 00:11:59.509 --> 00:12:02.409 recommendation. Keep thinking as the default. 00:12:02.559 --> 00:12:05.220 for anything complex. If you're doing real work 00:12:05.220 --> 00:12:08.240 coding, planning, analysis, the extra 10 seconds 00:12:08.240 --> 00:12:10.519 of wait time is worth the accuracy if you're 00:12:10.519 --> 00:12:13.240 just chatting about a movie. Use fast. So if 00:12:13.240 --> 00:12:15.620 we zoom out, we have multi -mobility, which is 00:12:15.620 --> 00:12:17.600 the eyes and ears. We have the thinking model, 00:12:17.639 --> 00:12:20.320 the brain. We have gems for the specialized training. 00:12:20.639 --> 00:12:22.700 And Canvas is the workspace. The full stack. 00:12:22.860 --> 00:12:24.460 It feels like a lot of separate tools when we 00:12:24.460 --> 00:12:26.179 list them like that. I want to see how they fit 00:12:26.179 --> 00:12:28.519 together. The source material lays out a smart 00:12:28.519 --> 00:12:31.039 garden workflow that I think perfectly illustrates 00:12:31.039 --> 00:12:33.639 the stacking concept. This is my favorite part 00:12:33.639 --> 00:12:35.700 because it takes someone with zero knowledge 00:12:35.700 --> 00:12:38.519 and gives them a result that looks like an expert 00:12:38.519 --> 00:12:41.129 produced it. OK, so let's walk through it. Step 00:12:41.129 --> 00:12:43.909 one, you want to grow a garden. You're in California. 00:12:44.029 --> 00:12:46.470 You have a balcony. You know nothing about plants. 00:12:46.690 --> 00:12:50.529 What do you do? Step one is multimodality. Don't 00:12:50.529 --> 00:12:53.029 type, I have a balcony that faces south. Just 00:12:53.029 --> 00:12:55.789 go outside, take a photo, upload it to Gemini, 00:12:56.250 --> 00:12:58.710 ask what grows here in California weather. So 00:12:58.710 --> 00:13:01.970 the AI analyzes the light, the space, the context 00:13:01.970 --> 00:13:04.450 right from the pixels. Right. It suggests, say, 00:13:04.610 --> 00:13:06.230 strawberries and lettuce because it sees you 00:13:06.230 --> 00:13:09.070 have Limited floor space, but good railing space. 00:13:09.289 --> 00:13:12.549 Okay. Step two? Visualization. This is the motivation 00:13:12.549 --> 00:13:15.190 step. You ask, show me a picture of what this 00:13:15.190 --> 00:13:17.370 will look like when it's fully grown. It uses 00:13:17.370 --> 00:13:20.049 image generation to show you a lush green balcony. 00:13:20.289 --> 00:13:21.809 Now you're excited. You can actually see the 00:13:21.809 --> 00:13:23.970 goal. I'm motivated, but I still don't know how 00:13:23.970 --> 00:13:26.929 to keep anything alive. Step three is deep research 00:13:26.929 --> 00:13:29.389 with a thinking model. You ask for a detailed 00:13:29.389 --> 00:13:31.830 plan. Create a weekly schedule telling me when 00:13:31.830 --> 00:13:34.070 to plant, when to fertilize, when to harvest. 00:13:34.549 --> 00:13:37.370 It digests all that agricultural data into a 00:13:37.370 --> 00:13:39.429 structured timeline for you. Now this is where 00:13:39.429 --> 00:13:41.870 most people would stop. They have the plan, they 00:13:41.870 --> 00:13:44.169 print it out, they lose it, and the plants die. 00:13:44.649 --> 00:13:47.149 Exactly. But we're going to build a system. Step 00:13:47.149 --> 00:13:50.870 four. Create a gem. You call it garden expert. 00:13:51.570 --> 00:13:53.870 You upload that plan and that original photo 00:13:53.870 --> 00:13:56.549 of your balcony into the gem's knowledge base. 00:13:56.850 --> 00:14:00.169 So the gem knows your specific garden. It knows 00:14:00.169 --> 00:14:01.990 your garden and the instruction you give it is, 00:14:02.370 --> 00:14:05.029 you are the caretaker of this balcony. When I 00:14:05.029 --> 00:14:07.370 send a photo of a sick plant, tell me how to 00:14:07.370 --> 00:14:09.929 fix it naturally based on the plants I own. That 00:14:09.929 --> 00:14:12.470 is brilliant. So two months later, the strawberry 00:14:12.470 --> 00:14:14.529 leaves are turning yellow. You don't Google yellow 00:14:14.529 --> 00:14:17.549 leaves. No. You snap a photo, you drop it in 00:14:17.549 --> 00:14:19.769 your gem, and it says, hey, remember those strawberries 00:14:19.769 --> 00:14:22.070 in the South Rail? They're nitrogen deficient. 00:14:22.330 --> 00:14:25.529 Add coffee grounds. It's context -aware troubleshooting. 00:14:25.730 --> 00:14:28.509 And finally, step five. Management via Canvas. 00:14:28.809 --> 00:14:31.289 You ask it to create a visual dashboard, a tracking 00:14:31.289 --> 00:14:33.190 table for watering and harvesting that you can 00:14:33.190 --> 00:14:35.889 edit. You just keep that open to track your progress. 00:14:36.269 --> 00:14:38.690 So in minutes, you went from zero knowledge to 00:14:38.690 --> 00:14:41.990 a personalized plan, a visual goal, a custom 00:14:41.990 --> 00:14:44.730 troubleshooter, and a management system. That's 00:14:44.730 --> 00:14:46.909 the power of the stack. You aren't just chatting. 00:14:47.250 --> 00:14:49.929 You are orchestrating. So the core shift here, 00:14:50.549 --> 00:14:53.539 it really is moving from just asking the AI a 00:14:53.539 --> 00:14:57.700 question to stacking the AI skills into a workflow. 00:14:57.919 --> 00:15:00.039 Exactly. It's not a robot you chat with. It's 00:15:00.039 --> 00:15:02.019 a multi -skilled assistant you build workflows 00:15:02.019 --> 00:15:05.120 for. You're the architect. The AI is the builder. 00:15:05.320 --> 00:15:07.480 I love that framing. It makes it feel much more 00:15:07.480 --> 00:15:09.600 active. But we've thrown a lot of features at 00:15:09.600 --> 00:15:12.039 you today. If you take away just one thing, what 00:15:12.039 --> 00:15:15.480 is the big idea here? The big idea is a mindset 00:15:15.480 --> 00:15:18.700 shift. It's resisting the urge to be overwhelmed. 00:15:19.419 --> 00:15:21.940 Don't try to change your entire work life tomorrow. 00:15:22.179 --> 00:15:24.100 Don't try to build the accountant gem and the 00:15:24.100 --> 00:15:26.879 garden gem and a coding gem all at once. Right. 00:15:27.100 --> 00:15:29.539 Start small. Pick one thing. One repetitive, 00:15:29.919 --> 00:15:32.320 boring, daily task. Maybe it's meal planning. 00:15:32.480 --> 00:15:34.940 Maybe it's summarizing that weekly meeting no 00:15:34.940 --> 00:15:36.840 one pays attention to. Maybe it's a good best 00:15:36.840 --> 00:15:38.820 report. It's one thing that annoys you. And build 00:15:38.820 --> 00:15:41.139 one simple gem for it. That's it. And once you 00:15:41.139 --> 00:15:43.299 save time on that one thing, the utility of the 00:15:43.299 --> 00:15:46.000 other features, multimodality, canvas, the thinking 00:15:46.000 --> 00:15:48.259 model, it all becomes obvious. It just clicks. 00:15:48.519 --> 00:15:50.279 You realize, oh, I can use this for that other 00:15:50.279 --> 00:15:52.480 thing, too. And suddenly, you aren't just a user 00:15:52.480 --> 00:15:55.159 anymore. You're a master of the tool. That is 00:15:55.159 --> 00:15:58.879 a great place to leave it. So here is our challenge 00:15:58.879 --> 00:16:02.000 to you listening right now. Identify one boring 00:16:02.000 --> 00:16:05.620 task today, just one, and attempt to create a 00:16:05.620 --> 00:16:08.440 gem for it. See if you can delegate that friction 00:16:08.440 --> 00:16:10.620 to the machine. And let us know how it goes. 00:16:10.779 --> 00:16:12.720 Thanks for diving deep with us today. We'll see 00:16:12.720 --> 00:16:14.159 you next time. Safe travels.