WEBVTT 00:00:00.000 --> 00:00:02.319 The pace of artificial intelligence is blistering 00:00:02.319 --> 00:00:05.259 right now. I mean, it isn't science fiction anymore. 00:00:05.580 --> 00:00:07.879 AI is basically your coworker. If you ignore 00:00:07.879 --> 00:00:10.419 what's under the hood, you're flying blind. Yeah, 00:00:10.539 --> 00:00:12.099 totally blind. Yeah. You know, you really need 00:00:12.099 --> 00:00:14.880 to know how it operates. The magic fades fast 00:00:14.880 --> 00:00:17.820 when you realize it's just mechanics. Welcome 00:00:17.820 --> 00:00:21.739 to this deep dive. Today, we're exploring something 00:00:21.739 --> 00:00:26.199 very special for you. We have the Essential 2026 00:00:26.199 --> 00:00:29.539 AI Concepts Handbook. Our mission is simple. 00:00:29.660 --> 00:00:31.920 We want to strip away the intimidation factor. 00:00:32.200 --> 00:00:34.060 Exactly. You don't need a math degree for this. 00:00:34.159 --> 00:00:36.159 You just need to understand how machines actually 00:00:36.159 --> 00:00:38.920 think and act. We're going to unpack 10 foundational 00:00:38.920 --> 00:00:41.560 concepts today, step by step. We want to turn 00:00:41.560 --> 00:00:44.000 you from a passive user into someone who controls 00:00:44.000 --> 00:00:46.450 the technology. Okay, let's unpack this. Let's 00:00:46.450 --> 00:00:48.350 do it. It all starts at the central brain. Right. 00:00:48.469 --> 00:00:50.789 We hear the term everywhere now. Large language 00:00:50.789 --> 00:00:53.429 models are LLMs. Things like Claude or Gemini. 00:00:53.789 --> 00:00:56.270 They are the core engine powering everything. 00:00:57.030 --> 00:00:59.409 An LLM is a model that guesses the next word 00:00:59.409 --> 00:01:02.590 based on vast data. That is the fundamental mechanism. 00:01:02.890 --> 00:01:04.530 But they don't actually understand the world, 00:01:04.569 --> 00:01:07.150 do they? We just project human intelligence onto 00:01:07.150 --> 00:01:10.329 them. We do. But it's essentially a highly complex 00:01:10.329 --> 00:01:13.760 guessing game. The AI just predicts the most 00:01:13.760 --> 00:01:16.900 likely next word. It does this after reading 00:01:16.900 --> 00:01:19.579 billions of pages of human text. It's like a 00:01:19.579 --> 00:01:22.019 brilliant student who's read every book. Yeah. 00:01:22.840 --> 00:01:24.980 The handbook has a genuinely fascinating example 00:01:24.980 --> 00:01:28.840 of this. You ask the AI to explain a Python for 00:01:28.840 --> 00:01:30.760 loop. Right, and you ask it to explain it to 00:01:30.760 --> 00:01:33.060 a complete beginner. Someone who knows absolutely 00:01:33.060 --> 00:01:36.099 nothing about code. Yes. The AI doesn't just 00:01:36.099 --> 00:01:38.739 give you a textbook definition. It uses an analogy 00:01:38.739 --> 00:01:41.420 of people waiting in line. They are waiting to 00:01:41.420 --> 00:01:44.060 buy milk tea. It's such a beautifully human way 00:01:44.060 --> 00:01:46.579 to explain cold logic. It really is. It takes 00:01:46.579 --> 00:01:49.159 the abstract concept of iteration and grounds 00:01:49.159 --> 00:01:51.400 it in everyday life. It's seen millions of code 00:01:51.400 --> 00:01:54.079 examples before. It's also seen millions of stories 00:01:54.079 --> 00:01:56.700 about lines and cafes. So it pieces together 00:01:56.700 --> 00:01:59.439 a highly relatable explanation. How does guessing 00:01:59.439 --> 00:02:02.599 next words look like actual logic? Predicting 00:02:02.599 --> 00:02:05.359 billions of words accurately creates the illusion 00:02:05.359 --> 00:02:09.219 of human reasoning. Two sec silence. So if the 00:02:09.219 --> 00:02:11.400 engine is just predicting words, how does it 00:02:11.400 --> 00:02:14.520 actually consume the text we feed it? That brings 00:02:14.520 --> 00:02:17.840 us to how AI reads. We have to talk about tokens. 00:02:18.500 --> 00:02:20.819 Tokens are small chunks of words used to process 00:02:20.819 --> 00:02:23.500 and build AI text. It doesn't read letter by 00:02:23.500 --> 00:02:26.039 letter. No. A short word is usually just one 00:02:26.039 --> 00:02:28.439 token. A very long word might be broken into 00:02:28.439 --> 00:02:30.939 three or four tokens. And this matters because 00:02:30.939 --> 00:02:34.620 you pay for the AI per token. You do. Keeping 00:02:34.620 --> 00:02:37.240 your prompts brief saves you real money. It's 00:02:37.240 --> 00:02:39.539 kind of the currency of the AI world. But tokens 00:02:39.539 --> 00:02:42.000 aren't just about cost. They also tie directly 00:02:42.000 --> 00:02:44.659 into memory. Exactly. This is the context window, 00:02:45.039 --> 00:02:48.060 the strict short -term memory limit for a single 00:02:48.060 --> 00:02:50.319 AI conversation. It's like a person's working 00:02:50.319 --> 00:02:52.719 memory. Yes. If you dump a massive novel into 00:02:52.719 --> 00:02:55.090 the chat, the window fills up. Once it hits that 00:02:55.090 --> 00:02:56.889 limit, it starts forgetting the beginning of 00:02:56.889 --> 00:02:59.930 the conversation. Why do long projects with AI 00:02:59.930 --> 00:03:02.610 suddenly derail? The AI's short -term memory 00:03:02.610 --> 00:03:04.889 fills up, causing it to forget early instructions. 00:03:05.550 --> 00:03:08.069 Memory limits explain a lot, but a conversation 00:03:08.069 --> 00:03:10.710 is inherently passive. What happens when we want 00:03:10.710 --> 00:03:13.289 the AI to actually do the work for us? This is 00:03:13.289 --> 00:03:15.770 where we move from passive chatbots to active 00:03:15.770 --> 00:03:19.840 workers, AI agents. Agents are programs that 00:03:19.840 --> 00:03:23.240 autonomously plan and execute multi -step digital 00:03:23.240 --> 00:03:25.900 tasks. The distinction here is incredibly important. 00:03:25.979 --> 00:03:28.479 It changes everything. A regular chatbot just 00:03:28.479 --> 00:03:30.699 gives you a recipe. It lists the steps you need 00:03:30.699 --> 00:03:33.680 to take. But an agent actually cooks the meal. 00:03:33.800 --> 00:03:36.199 Right, exactly. Say you want to book a flight 00:03:36.199 --> 00:03:38.860 to New York. A chatbot tells you how to navigate 00:03:38.860 --> 00:03:40.960 a travel website. It might even suggest some 00:03:40.960 --> 00:03:43.740 dates for you. Sure, but an agent goes to the 00:03:43.740 --> 00:03:46.370 site itself. It clicks the buttons. It compares 00:03:46.370 --> 00:03:49.229 the prices across multiple airlines. And it books 00:03:49.229 --> 00:03:51.969 the cheapest flight for next Friday. You just 00:03:51.969 --> 00:03:54.110 manage the final goal. You become the manager, 00:03:54.169 --> 00:03:56.289 not the worker. It does this using something 00:03:56.289 --> 00:04:00.030 called an action loop. The cycle is plan, act, 00:04:00.449 --> 00:04:02.870 observe, and repeat. It breaks the massive goal 00:04:02.870 --> 00:04:05.969 into very small, manageable tasks. Yes. It makes 00:04:05.969 --> 00:04:08.650 a plan. Then it acts on the first step. then 00:04:08.650 --> 00:04:10.830 it stops and carefully observes the result of 00:04:10.830 --> 00:04:13.050 that action. It's checking its own work. Exactly. 00:04:13.389 --> 00:04:16.149 If it clicks a broken link, it observes the failure. 00:04:16.769 --> 00:04:19.550 It fixes the plan, finds a new link, and repeats 00:04:19.550 --> 00:04:22.790 the cycle. What transforms a chatbot into an 00:04:22.790 --> 00:04:25.550 autonomous worker? Agents use a continuous loop 00:04:25.550 --> 00:04:28.050 of planning, acting, and self -correcting mistakes. 00:04:28.329 --> 00:04:31.230 To sex silence. So we have these agents acting 00:04:31.230 --> 00:04:34.560 as digital workers. But a worker is entirely 00:04:34.560 --> 00:04:37.279 useless if it can't access your filing cabinet. 00:04:37.420 --> 00:04:40.399 That is the big bottleneck. Agents need access 00:04:40.399 --> 00:04:42.879 to your tools. This is where the model context 00:04:42.879 --> 00:04:46.120 protocol comes in. MCP. I loved the handbooks 00:04:46.120 --> 00:04:49.000 analogy for this. Think of MCP as the invention 00:04:49.000 --> 00:04:51.639 of the USB port. It's the perfect way to visualize 00:04:51.639 --> 00:04:54.899 it. MCP is a universal standard connecting AI 00:04:54.899 --> 00:04:58.500 directly to external data sources. Before MCP, 00:04:58.579 --> 00:05:00.740 things were incredibly messy for developers. 00:05:00.920 --> 00:05:03.019 They were a nightmare. Connecting an AI to your 00:05:03.019 --> 00:05:06.000 Google Drive meant writing complex, custom code. 00:05:06.240 --> 00:05:07.879 Right, and then connecting it to Slack meant 00:05:07.879 --> 00:05:10.540 writing totally new code from scratch. It took 00:05:10.540 --> 00:05:13.139 software engineers massive amounts of time. Every 00:05:13.139 --> 00:05:15.379 single connection was a bespoke, fragile bridge. 00:05:15.639 --> 00:05:18.740 But MCP functions as the universal plug. It does. 00:05:18.959 --> 00:05:21.540 Amtropic helped build this open standard. Now, 00:05:21.639 --> 00:05:23.939 different AI models can use the exact same data 00:05:23.939 --> 00:05:26.220 sources effortlessly. You just plug it in once 00:05:26.220 --> 00:05:29.639 and it works everywhere. Why is MCP such a massive 00:05:29.639 --> 00:05:31.480 leap for developers? Oh wait, that's my line. 00:05:31.800 --> 00:05:34.199 Why is MCP such a massive leap for developers? 00:05:34.439 --> 00:05:37.740 It acts as a universal plug, connecting AI to 00:05:37.740 --> 00:05:41.120 tools without custom code. Beat. Okay, the AI 00:05:41.120 --> 00:05:44.310 is plugged into our systems. But we still face 00:05:44.310 --> 00:05:47.689 a major limitation. The AI only knows what it 00:05:47.689 --> 00:05:50.009 learned during its initial training. Yeah. Its 00:05:50.009 --> 00:05:52.250 knowledge has a strict cutoff date. If a crucial 00:05:52.250 --> 00:05:54.709 financial report came out this morning, the AI 00:05:54.709 --> 00:05:56.509 is completely blind to it. And if it doesn't 00:05:56.509 --> 00:05:59.089 know the answer, it tends to guess. Which leads 00:05:59.089 --> 00:06:02.009 to hallucinations. It wants to please you, so 00:06:02.009 --> 00:06:04.550 it confidently makes things up. This introduces 00:06:04.550 --> 00:06:07.389 retrieval -augmented generation. We call it RG. 00:06:07.550 --> 00:06:10.350 Fetching relevant documents first so the AI basis 00:06:10.350 --> 00:06:13.009 answers on facts. It's exactly like letting a 00:06:13.009 --> 00:06:15.470 student bring reference books into a highly difficult 00:06:15.470 --> 00:06:17.449 exam. It doesn't have to memorize everything 00:06:17.449 --> 00:06:20.029 anymore. Right. It retrieves your specific documents 00:06:20.029 --> 00:06:23.170 first. Then it generates the final answer based 00:06:23.170 --> 00:06:25.790 strictly on those files. It firmly anchors the 00:06:25.790 --> 00:06:28.629 AI in reality. It stops the AI from lying to 00:06:28.629 --> 00:06:31.310 you. Exactly. And it uses vector databases to 00:06:31.310 --> 00:06:34.209 do this instantly. A vector database is a system 00:06:34.209 --> 00:06:36.410 searching information by mathematical meaning. 00:06:36.600 --> 00:06:40.139 Not exact keywords. This part blew my mind. If 00:06:40.139 --> 00:06:42.639 you search a normal database for the word car... 00:06:42.639 --> 00:06:44.579 It only looks for those exact three letters. 00:06:44.819 --> 00:06:48.259 C -A -R. But a vector database understands the 00:06:48.259 --> 00:06:50.980 actual concept. Yes. It finds documents mentioning 00:06:50.980 --> 00:06:54.339 automobile or transportation or vehicle. It plots 00:06:54.339 --> 00:06:56.720 these concepts mathematically in space to find 00:06:56.720 --> 00:07:00.000 deep relationships. How exactly does ARAG prevent 00:07:00.000 --> 00:07:03.430 the AI from hallucinating? It forces the AI to 00:07:03.430 --> 00:07:07.009 base its answers strictly on retrieved, verified 00:07:07.009 --> 00:07:09.889 documents. To sex silence. I understand using 00:07:09.889 --> 00:07:12.569 ARAG to feed the AI a factual financial report, 00:07:12.949 --> 00:07:15.629 but what if I want the AI to stop writing like 00:07:15.629 --> 00:07:17.970 a chipper customer service rep? This is where 00:07:17.970 --> 00:07:21.129 people get confused. ARAG is for facts. Fine 00:07:21.129 --> 00:07:23.709 -tuning is for style and habits. Fine -tuning 00:07:23.709 --> 00:07:26.509 is training a model on specific examples to change 00:07:26.509 --> 00:07:28.829 its conversational style. Right. You're changing 00:07:28.829 --> 00:07:31.160 the behavior of the model itself. Maybe you want 00:07:31.160 --> 00:07:33.199 it to sound like a seasoned professional lawyer. 00:07:33.420 --> 00:07:34.779 You definitely don't want it sounding like a 00:07:34.779 --> 00:07:38.060 teenager texting their friends. Exactly. Or maybe 00:07:38.060 --> 00:07:40.339 you need it to format data properly for other 00:07:40.339 --> 00:07:42.920 software, like forcing it to output strictly 00:07:42.920 --> 00:07:45.019 in JSON format. You don't have to build a new 00:07:45.019 --> 00:07:48.319 model for this. No. You use small, exceptionally 00:07:48.319 --> 00:07:50.939 high -quality data. You don't train a massive 00:07:50.939 --> 00:07:53.949 brain from scratch. That would cost millions 00:07:53.949 --> 00:07:56.829 of dollars. Instead, you just feed it a few hundred 00:07:56.829 --> 00:07:59.529 of your own best emails. It acts like a finishing 00:07:59.529 --> 00:08:02.550 school for the AI. It learns your specific habits 00:08:02.550 --> 00:08:05.250 and cadence. Yes, it copies your exact energy 00:08:05.250 --> 00:08:07.930 and phrasing. It's surprisingly cheap, but yields 00:08:07.930 --> 00:08:11.250 very impressive personalized results. If ARIG 00:08:11.250 --> 00:08:14.089 gives the AI facts, what is fine -tuning for? 00:08:14.319 --> 00:08:16.899 Fine -tuning teaches the AI specific professional 00:08:16.899 --> 00:08:19.879 habits, formats, and conversational styles. Beat. 00:08:20.139 --> 00:08:22.139 This brings us to a concept that feels highly 00:08:22.139 --> 00:08:25.699 personal. Context engineering. This goes way 00:08:25.699 --> 00:08:28.399 beyond just writing a clever prompt. It's a completely 00:08:28.399 --> 00:08:30.240 different mindset. I have a confession here. 00:08:30.279 --> 00:08:32.940 I still wrestle with prompt drift myself. We 00:08:32.940 --> 00:08:35.500 all do. It's so easy to lose control of the chat. 00:08:36.220 --> 00:08:38.100 Context engineering is carefully designing the 00:08:38.100 --> 00:08:40.039 specific information environment you feed an 00:08:40.039 --> 00:08:42.799 AI. You are the context engineer. You are picking 00:08:42.799 --> 00:08:45.299 the raw ingredients for a highly talented chef. 00:08:45.500 --> 00:08:49.299 That is the perfect analogy. The AI is the master 00:08:49.299 --> 00:08:51.860 chef. It has all the skills. But if you hand 00:08:51.860 --> 00:08:54.620 the chef rotten ingredients, the meal is completely 00:08:54.620 --> 00:08:58.379 ruined. Exactly. Bad ingredients mean messy files, 00:08:58.879 --> 00:09:01.840 contradictory instructions, or extra irrelevant 00:09:01.840 --> 00:09:04.059 information. More information isn't always better. 00:09:04.139 --> 00:09:06.919 Often, it's much worse. You have to curate a 00:09:06.919 --> 00:09:09.360 clean, highly focused environment. You have to 00:09:09.360 --> 00:09:12.159 decide exactly which files it truly needs. What 00:09:12.159 --> 00:09:15.460 is the fastest way to ruin a smart AI's output? 00:09:16.059 --> 00:09:18.639 Feeding it messy, conflicting, or overflowing 00:09:18.639 --> 00:09:22.120 context files ruins the results. To sex silence, 00:09:22.559 --> 00:09:24.460 even with the perfect ingredients, sometimes 00:09:24.460 --> 00:09:26.720 the chef rushes. That brings us to reasoning 00:09:26.720 --> 00:09:28.840 models. These models are fundamentally learning 00:09:28.840 --> 00:09:32.279 how to think. Old models rely on instant, reflexive 00:09:32.279 --> 00:09:34.379 generation. They rush to answer the question 00:09:34.379 --> 00:09:36.519 immediately. And because they rush, they make 00:09:36.519 --> 00:09:38.899 incredibly silly logic mistakes. They fall into 00:09:38.899 --> 00:09:41.340 obvious cognitive traps. Reasoning models take 00:09:41.340 --> 00:09:44.440 their time. A reasoning model is AI that pauses 00:09:44.440 --> 00:09:47.039 to internally map out logic before generating 00:09:47.039 --> 00:09:50.480 answers. Yes. They use a hidden internal dialogue. 00:09:50.980 --> 00:09:53.059 When you see that thinking prompt on your screen, 00:09:53.200 --> 00:09:55.600 it's working hard. It's having a conversation 00:09:55.600 --> 00:09:58.100 with itself. It's mapping out the problem step 00:09:58.100 --> 00:10:00.940 by step. It's actively checking its own logic 00:10:00.940 --> 00:10:03.779 to avoid those silly traps. The handbook shares 00:10:03.779 --> 00:10:06.919 a brilliant logic puzzle for this. The box puzzle. 00:10:07.120 --> 00:10:09.639 Right. You have three boxes. One gold, one silver, 00:10:09.740 --> 00:10:12.200 and one empty. The gold is not in the first box. 00:10:12.440 --> 00:10:14.460 The silver is in the second box. Where is the 00:10:14.460 --> 00:10:18.139 gold? It seems simple to us, but an old AI model 00:10:18.139 --> 00:10:20.580 might guess instantly and fail completely. But 00:10:20.580 --> 00:10:23.720 a reasoning model pauses. It maps the physical 00:10:23.720 --> 00:10:26.299 constraints of the boxes, and it gets it right. 00:10:26.779 --> 00:10:29.320 Why do reasoning models pause before they type? 00:10:29.600 --> 00:10:32.080 Wait, I asked the questions here. Why do reasoning 00:10:32.080 --> 00:10:34.159 models pause before they type? They're using 00:10:34.159 --> 00:10:36.759 internal dialogue to map out logic and avoid 00:10:36.759 --> 00:10:39.279 silly mistakes. Thinking deeply is one thing, 00:10:39.460 --> 00:10:41.500 but sensing the physical world is an entirely 00:10:41.500 --> 00:10:44.139 different leap. Let's look at multimodal AI. 00:10:44.620 --> 00:10:46.860 This is where the technology gets truly wild. 00:10:47.600 --> 00:10:50.039 Multimodal AI means systems that can process 00:10:50.039 --> 00:10:53.460 text, images, and audio simultaneously. It's 00:10:53.460 --> 00:10:56.460 not just text on a screen anymore. The AI has 00:10:56.460 --> 00:10:59.080 eyes and ears. It's interacting with reality 00:10:59.080 --> 00:11:03.139 in the same way we do. Beat! Whoa! Imagine pointing 00:11:03.139 --> 00:11:05.799 your camera at a leaking car pipe and the AI 00:11:05.799 --> 00:11:07.639 just looks at it and tells you which screw to 00:11:07.639 --> 00:11:10.460 turn. It's an incredible shift. The physical 00:11:10.460 --> 00:11:13.620 world becomes a prompt. You can draw a messy 00:11:13.620 --> 00:11:15.879 handwritten whiteboard layout for a website. 00:11:16.100 --> 00:11:17.980 Then you just show the AI a quick picture of 00:11:17.980 --> 00:11:20.279 it. And it generates the functional working code 00:11:20.279 --> 00:11:22.620 for that layout instantly. Right. And it works 00:11:22.620 --> 00:11:24.720 seamlessly with audio, too. You can have the 00:11:24.720 --> 00:11:27.820 AI listen to a chaotic one -hour team meeting. 00:11:28.100 --> 00:11:30.440 It knows exactly who is speaking. It separates 00:11:30.440 --> 00:11:33.259 the voices and summarizes the action items for 00:11:33.259 --> 00:11:36.419 each person. How does multimodal AI change our 00:11:36.419 --> 00:11:39.000 physical reality? It bridges the gap by letting 00:11:39.000 --> 00:11:42.600 AI see and hear the physical world. To sex silence. 00:11:43.100 --> 00:11:45.259 Processing all those different senses must require 00:11:45.259 --> 00:11:47.639 a staggering amount of brain power. It does. 00:11:48.000 --> 00:11:50.379 And that brings us to the final concept. Mystery 00:11:50.379 --> 00:11:54.019 of experts or MoE? This is about maximizing efficiency 00:11:54.019 --> 00:11:57.240 at the deepest architectural level. MoE is dividing 00:11:57.240 --> 00:12:00.120 an AI into specialized subnetworks to save computing 00:12:00.120 --> 00:12:02.750 power. It doesn't use the whole brain for every 00:12:02.750 --> 00:12:06.210 single task. Exactly. The AI brain is divided 00:12:06.210 --> 00:12:09.370 into specialized expert groups. So there is a 00:12:09.370 --> 00:12:12.350 dedicated math expert hidden inside the model. 00:12:12.490 --> 00:12:14.830 There might be a coding expert or a translation 00:12:14.830 --> 00:12:17.710 expert too. Right. When you ask a simple math 00:12:17.710 --> 00:12:20.789 question, only the math expert wakes up to answer 00:12:20.789 --> 00:12:23.450 it. The rest of the massive model stays resting. 00:12:23.649 --> 00:12:26.009 That stops the AI from using giant computing 00:12:26.009 --> 00:12:29.139 power for basic simple questions. Yes. It saves 00:12:29.139 --> 00:12:32.080 massive amounts of electricity and valuable computing 00:12:32.080 --> 00:12:35.440 time. Models like Mistral and DeepSeq use this 00:12:35.440 --> 00:12:38.440 heavily. It makes everything much faster and 00:12:38.440 --> 00:12:40.940 drastically cheaper to run. Why divide the AI 00:12:40.940 --> 00:12:43.379 into specialized expert groups? I'm just going 00:12:43.379 --> 00:12:45.860 to claim that question. Why divide the AI into 00:12:45.860 --> 00:12:48.360 specialized expert groups? It drastically speeds 00:12:48.360 --> 00:12:51.460 up response times and saves massive amounts of 00:12:51.460 --> 00:12:53.320 computing energy. Beep beep. OK, we're going 00:12:53.320 --> 00:12:55.309 to take a really quick break. Mid -roll sponsor, 00:12:55.570 --> 00:12:57.769 Reed Placeholder. Welcome back. We have covered 00:12:57.769 --> 00:13:00.789 a staggering amount of ground today. Let's synthesize 00:13:00.789 --> 00:13:03.490 this massive journey. We really have. If there's 00:13:03.490 --> 00:13:06.490 one big takeaway, it's that AI is not magic. 00:13:06.990 --> 00:13:09.830 It is highly advanced mechanics. It's predicting 00:13:09.830 --> 00:13:13.429 tokens with LLMs. It's looping actions with agents. 00:13:13.669 --> 00:13:16.990 It's plugging directly into your data using MCP. 00:13:17.309 --> 00:13:20.289 It's fetching verified accurate facts with RG. 00:13:20.429 --> 00:13:23.009 And it's using specialized brains through MoE 00:13:23.009 --> 00:13:26.470 to stay incredibly efficient. Exactly. The final 00:13:26.470 --> 00:13:28.570 piece of advice from the handbook is vital here. 00:13:28.990 --> 00:13:31.169 Do not get overwhelmed by all of this at once. 00:13:31.350 --> 00:13:34.470 It's a lot to take in. Start small. Yes. Start 00:13:34.470 --> 00:13:36.889 with the very first concepts. Really understand 00:13:36.889 --> 00:13:40.450 LLMs and tokens. Once you firmly grasp how AI 00:13:40.450 --> 00:13:43.450 reads and talks, the rest becomes much easier 00:13:43.450 --> 00:13:45.710 to digest. You don't need to be an expert in 00:13:45.710 --> 00:13:47.629 everything today. Just take it one step at a 00:13:47.629 --> 00:13:49.779 time. Play around with the models. Observe how 00:13:49.779 --> 00:13:53.000 they react to your specific inputs. In 2026, 00:13:53.299 --> 00:13:55.500 you shouldn't just be using AI. You need to be 00:13:55.500 --> 00:13:57.720 the one who controls its environment. If you 00:13:57.720 --> 00:13:59.860 curate the context, you control the machine. 00:13:59.980 --> 00:14:01.840 That is the absolute truth. You hold the steering 00:14:01.840 --> 00:14:03.860 wheel now. Thank you for taking this deep dive 00:14:03.860 --> 00:14:05.899 with us. We hope these concepts help you build 00:14:05.899 --> 00:14:08.279 amazing things. Out T -Row music.