WEBVTT 00:00:00.000 --> 00:00:02.419 Okay, let's unpack this. We've all seen AI as 00:00:02.419 --> 00:00:05.500 this brilliant conversationalist, right? A super 00:00:05.500 --> 00:00:07.500 smart chatbot, you ask something, and boom, you 00:00:07.500 --> 00:00:10.060 get this amazing answer. But the AI that's really 00:00:10.060 --> 00:00:12.019 going to change things, the one for the next 00:00:12.019 --> 00:00:14.279 decade, it doesn't just wait for you to ask. 00:00:14.380 --> 00:00:17.219 It's proactive. Imagine an AI checking flight 00:00:17.219 --> 00:00:19.219 prices for you, sees a big drop, checks your 00:00:19.219 --> 00:00:21.300 calendar, and just books your trip, all automatically, 00:00:21.460 --> 00:00:23.839 talking to different apps. That kind of independence, 00:00:24.140 --> 00:00:26.260 that takes some serious next -level engineering. 00:00:26.809 --> 00:00:28.769 Today, we're diving into the advanced stuff, 00:00:28.870 --> 00:00:30.890 the architectures that make that kind of autonomy 00:00:30.890 --> 00:00:33.609 actually work. Welcome, everyone, to the Deep 00:00:33.609 --> 00:00:36.509 Dive. So if you're already good with the basics, 00:00:36.909 --> 00:00:41.030 you know, LLMs, vector databases, ROG, then you 00:00:41.030 --> 00:00:43.770 are definitely ready for what's next. Our mission 00:00:43.770 --> 00:00:46.820 today. Go beyond those fundamentals. We're hitting 00:00:46.820 --> 00:00:50.000 10 critical advanced concepts. Stuff architects, 00:00:50.399 --> 00:00:52.520 developers, anyone planning AI strategy really 00:00:52.520 --> 00:00:54.359 needs to grasp. We're going to figure out how 00:00:54.359 --> 00:00:56.579 AI connects to the real world, how it actually 00:00:56.579 --> 00:00:58.899 reasons, learns on the fly, and super important, 00:00:58.979 --> 00:01:00.899 how we make it affordable and fast enough to 00:01:00.899 --> 00:01:03.500 deploy. We'll look at how AI takes action, how 00:01:03.500 --> 00:01:05.640 it gets smarter, and how engineers make it efficient. 00:01:06.040 --> 00:01:09.359 Let's start with how AI gets out of its virtual 00:01:09.359 --> 00:01:11.439 box. Yeah, that's the big hurdle, isn't it? And 00:01:11.439 --> 00:01:14.310 LLM, just by itself, it's kind of trapped. lives 00:01:14.310 --> 00:01:16.409 in its text world, how does it actually reach 00:01:16.409 --> 00:01:18.549 out and do something? Talk to an airline's booking 00:01:18.549 --> 00:01:21.230 system. Right. It needs a standard way to shake 00:01:21.230 --> 00:01:23.430 hands, basically. And that's where the model 00:01:23.430 --> 00:01:26.829 context protocol, MCP, comes into play. Think 00:01:26.829 --> 00:01:28.890 of it as the essential bridge, the agreed upon 00:01:28.890 --> 00:01:32.390 language for taking action. MCP sets up a structured 00:01:32.390 --> 00:01:34.549 way for the AI client that's usually the wrapper 00:01:34.549 --> 00:01:37.409 around the LLM to talk to outside systems, which 00:01:37.409 --> 00:01:39.849 we call MCP servers. It's like a really formalized 00:01:39.849 --> 00:01:42.590 API structure. Okay. So if the LLM figures out 00:01:43.019 --> 00:01:46.879 I need to book that flight. It uses MCP to frame 00:01:46.879 --> 00:01:50.019 a standard request, like call the Indigo booking 00:01:50.019 --> 00:01:53.680 server, use flight 1020, date X, something like 00:01:53.680 --> 00:01:56.239 that. Exactly like that. And this is huge because 00:01:56.239 --> 00:01:59.099 it brings in security and reliability. Standardizing 00:01:59.099 --> 00:02:01.459 how they talk. That cuts down the risk of weird 00:02:01.459 --> 00:02:04.079 stuff like prompt injection or the model just 00:02:04.079 --> 00:02:06.599 making up API calls. It really shifts AI from 00:02:06.599 --> 00:02:08.800 just talking about tasks to reliably doing them 00:02:08.800 --> 00:02:11.060 in real systems. It's the plumbing you need for 00:02:11.060 --> 00:02:14.199 actual digital agents. Got it. So MCP handles 00:02:14.199 --> 00:02:17.120 talking to the outside world. But inside, the 00:02:17.120 --> 00:02:19.180 AI needs to remember what's going on, what we've 00:02:19.180 --> 00:02:21.039 talked about. That sounds like context engineering, 00:02:21.280 --> 00:02:23.379 which is way more than just writing one good 00:02:23.379 --> 00:02:26.360 prompt at the start. Oh, definitely. This is 00:02:26.360 --> 00:02:29.039 the really sophisticated art of managing that 00:02:29.039 --> 00:02:31.840 ongoing conversation history. You've got that 00:02:31.840 --> 00:02:35.340 limited context window, the model's short -term 00:02:35.340 --> 00:02:37.860 memory. You have to work around that and make 00:02:37.860 --> 00:02:41.580 it feel deeply personal. I mean, if you're interacting 00:02:41.580 --> 00:02:44.840 with an AI agent over weeks, it absolutely needs 00:02:44.840 --> 00:02:47.560 to remember what you like, what you didn't like 00:02:47.560 --> 00:02:50.219 last time, your specific rules. And keeping that 00:02:50.219 --> 00:02:53.060 straight. Honestly, I still wrestle with prompt 00:02:53.060 --> 00:02:55.800 drift myself sometimes. You craft this perfect 00:02:55.800 --> 00:02:58.370 starting instruction, but... 30 messages later, 00:02:58.509 --> 00:03:00.530 that AI seems to have completely forgotten it. 00:03:00.610 --> 00:03:02.409 Yeah, we fight that with some clever tricks. 00:03:02.669 --> 00:03:04.830 You can use a sliding window, just keeping the 00:03:04.830 --> 00:03:08.069 most recent chat history in focus. Or, more advanced, 00:03:08.330 --> 00:03:10.889 smart truncation. That's where you use a smaller, 00:03:10.969 --> 00:03:13.669 faster AI to kind of summarize the long history 00:03:13.669 --> 00:03:15.930 before feeding the important bits to the big 00:03:15.930 --> 00:03:18.189 model. It's all about creating this dynamic, 00:03:18.330 --> 00:03:21.009 hyper -relevant context on the fly. That sounds 00:03:21.009 --> 00:03:23.530 crucial for making a generic chatbot feel like 00:03:23.530 --> 00:03:25.129 a personal assistant that actually gets you. 00:03:25.270 --> 00:03:27.400 Absolutely. It's what delivers that coherence, 00:03:27.400 --> 00:03:29.180 that personalization, especially when conversations 00:03:29.180 --> 00:03:31.819 get really long and complex and twisty. Okay. 00:03:31.900 --> 00:03:34.060 Which brings us nearly to the big payoff here. 00:03:34.460 --> 00:03:36.659 Agents. These are the systems that really run 00:03:36.659 --> 00:03:39.000 on their own. We've given the AI access to the 00:03:39.000 --> 00:03:42.280 outside, MCP, and a memory, context engineering. 00:03:42.479 --> 00:03:44.740 What's the final piece? Initiative. That's the 00:03:44.740 --> 00:03:47.300 key difference. An agent is autonomous. It has 00:03:47.300 --> 00:03:50.419 planning skills. A chatbot just reacts to what 00:03:50.419 --> 00:03:53.120 you say now. An agent. It takes action based 00:03:53.120 --> 00:03:55.199 on a goal you might have set days, even weeks 00:03:55.199 --> 00:03:58.099 ago. So this agent has memory. It can use tools 00:03:58.099 --> 00:04:00.340 through MCP. And critically, it can break down 00:04:00.340 --> 00:04:02.539 a big, complex goal into lots and lots of smaller 00:04:02.539 --> 00:04:05.819 steps. It executes them one by one, checks if 00:04:05.819 --> 00:04:08.099 it's working, corrects course if needed. I love 00:04:08.099 --> 00:04:10.479 that autonomous travel agent idea. You tell it 00:04:10.479 --> 00:04:12.919 once, book my usual vacation when the price hits 00:04:12.919 --> 00:04:15.520 the sweet spot. And this agent just... running 00:04:15.520 --> 00:04:18.579 quietly in the background, 247 is watching fares, 00:04:18.579 --> 00:04:20.959 connects to the airline, the hotel, checks your 00:04:20.959 --> 00:04:23.079 calendar, and does the whole thing without you 00:04:23.079 --> 00:04:25.180 needing to nudge it again. It's that jump from 00:04:25.180 --> 00:04:28.079 just reactive help to being proactive, strategic. 00:04:28.439 --> 00:04:30.819 These things act like digital employees working 00:04:30.819 --> 00:04:32.959 around the clock. Right. So we've built a system 00:04:32.959 --> 00:04:35.839 that can act. Now, how do we make sure its actions 00:04:35.839 --> 00:04:39.519 are, well, good, aligned with what we want, and 00:04:39.519 --> 00:04:41.639 genuinely smart? Let's start with reinforcement 00:04:41.639 --> 00:04:44.819 learning, RL. RL is all about shaping behavior. 00:04:45.240 --> 00:04:48.620 You train the AI using a system of rewards and 00:04:48.620 --> 00:04:51.180 penalties, essentially, optimizing what it does 00:04:51.180 --> 00:04:54.040 based on human feedback. Think kind of like training 00:04:54.040 --> 00:04:56.680 a dog with treats. So the process is the model 00:04:56.680 --> 00:04:58.959 gives you a few possible responses, a human picks 00:04:58.959 --> 00:05:01.560 the better one, and that choice gets turned into 00:05:01.560 --> 00:05:03.399 like a mathematical score. If your choice was 00:05:03.399 --> 00:05:05.759 good, all the internal calculations the model 00:05:05.759 --> 00:05:07.879 used to get there get nudged in a positive direction. 00:05:08.120 --> 00:05:11.199 Bad choice. nudged negative. So you're steering 00:05:11.199 --> 00:05:13.019 the model's path through its huge possibility 00:05:13.019 --> 00:05:15.740 space towards outputs we find useful or helpful. 00:05:16.189 --> 00:05:18.750 Precisely. And it's really powerful for optimizing 00:05:18.750 --> 00:05:21.889 very complex behaviors. Like if you want an AI 00:05:21.889 --> 00:05:24.589 that's consistently helpful, polite, but also 00:05:24.589 --> 00:05:27.910 thorough, RL is great at shaping that kind of 00:05:27.910 --> 00:05:29.850 nuanced output, maybe better than just showing 00:05:29.850 --> 00:05:32.850 it examples. But there's catch. RL is great at 00:05:32.850 --> 00:05:35.149 optimizing the behavior, but it doesn't always 00:05:35.149 --> 00:05:37.670 build deep understanding. It can learn the pattern 00:05:37.670 --> 00:05:40.389 of what response makes the human happy without 00:05:40.389 --> 00:05:44.389 truly getting the underlying facts. Okay, that's 00:05:44.389 --> 00:05:47.029 a subtle but really important point. So to make 00:05:47.029 --> 00:05:49.589 sure the logic itself is sound, we need something 00:05:49.589 --> 00:05:52.420 like chain of thought. Say AT. This is about 00:05:52.420 --> 00:05:55.220 making the AI show its work, right? Exactly. 00:05:55.500 --> 00:05:57.839 Hardy forces the AI to break down the problem. 00:05:57.920 --> 00:05:59.819 We're not just asking for the final number. We're 00:05:59.819 --> 00:06:02.300 telling it, show me the steps. Make it explicit, 00:06:02.420 --> 00:06:04.959 like how a person would solve it on paper. So 00:06:04.959 --> 00:06:07.019 if you ask it to calculate, say, a tricky sales 00:06:07.019 --> 00:06:09.360 commission with different tiers and taxes, it 00:06:09.360 --> 00:06:10.860 won't just give you the dollar amount. It has 00:06:10.860 --> 00:06:13.060 to spell out. First, convert the percentage, 00:06:13.300 --> 00:06:15.540 calculate tier one, calculate tier two, add them 00:06:15.540 --> 00:06:18.060 up, maybe round it off. That focus on showing 00:06:18.060 --> 00:06:21.060 the intermediate steps seems vital for trust 00:06:21.060 --> 00:06:24.220 and for fixing things, especially in fields like, 00:06:24.240 --> 00:06:27.379 I don't know, engineering or finance, where a 00:06:27.379 --> 00:06:30.079 mistake in step two messes everything else up. 00:06:30.160 --> 00:06:32.019 You can see where it went wrong. It absolutely 00:06:32.019 --> 00:06:34.139 cuts down on those multi -step errors because 00:06:34.139 --> 00:06:36.240 the model is kind of checking its own work as 00:06:36.240 --> 00:06:38.800 it goes. It makes the LLM a much more reliable 00:06:38.800 --> 00:06:42.319 reasoning tool. But Coty is still kind of following 00:06:42.319 --> 00:06:45.720 a known recipe, even if it shows the steps. The 00:06:45.720 --> 00:06:48.290 real frontier. reasoning models that's where 00:06:48.290 --> 00:06:50.389 the ai starts to figure out the recipe itself 00:06:50.389 --> 00:06:52.550 for new problems that's exactly it that's the 00:06:52.550 --> 00:06:54.550 cutting edge right now these models are built 00:06:54.550 --> 00:06:56.389 to figure out how to tackle problems they've 00:06:56.389 --> 00:06:58.930 never seen before not just applying patterns 00:06:58.930 --> 00:07:01.069 they memorize during training they use really 00:07:01.069 --> 00:07:03.269 sophisticated strategies things like tree of 00:07:03.269 --> 00:07:05.569 thought where the ai explores multiple possible 00:07:05.569 --> 00:07:07.810 solution paths like branches of a tree before 00:07:07.810 --> 00:07:10.550 picking the best one or graph of thought which 00:07:10.550 --> 00:07:12.629 can handle problems where steps aren't just linear 00:07:12.629 --> 00:07:15.500 they depend on each other in complex ways Okay, 00:07:15.560 --> 00:07:17.420 so if Chain of Thought is like following a marked 00:07:17.420 --> 00:07:20.639 trail, reasoning models are like expert navigators 00:07:20.639 --> 00:07:23.259 charting a course through totally unknown territory, 00:07:23.579 --> 00:07:26.279 maybe grabbing different tools as needed, picking 00:07:26.279 --> 00:07:28.540 the best strategy for something completely novel. 00:07:28.740 --> 00:07:31.300 It's about cognitive flexibility. You see things 00:07:31.300 --> 00:07:35.079 like OpenAI's O1 or DeepSeek R1 really pushing 00:07:35.079 --> 00:07:38.060 here. It's aiming for that true strategic thinking 00:07:38.060 --> 00:07:40.560 needed for, say, scientific breakthroughs or 00:07:40.560 --> 00:07:42.579 designing complex systems. That kind of thinking 00:07:42.579 --> 00:07:44.579 power probably needs more than just text data. 00:07:45.050 --> 00:07:47.550 Let's shift gears to data types with multimodal 00:07:47.550 --> 00:07:50.050 models. Right, because the world isn't just words. 00:07:50.269 --> 00:07:52.410 These models are trained to handle multiple kinds 00:07:52.410 --> 00:07:55.529 of data at the same time, text plus images or 00:07:55.529 --> 00:07:59.170 video or audio. And the big advantage, it's how 00:07:59.170 --> 00:08:01.910 they learn. Imagine an AI that hasn't just read 00:08:01.910 --> 00:08:04.350 millions of sentences about cats, but has also 00:08:04.350 --> 00:08:06.629 seen millions of pictures and videos of cats. 00:08:06.850 --> 00:08:10.230 It builds a much richer, deeper, almost multisensory 00:08:10.230 --> 00:08:13.290 understanding of catness than a text -only model 00:08:13.290 --> 00:08:15.759 ever could. Yeah, that deeper understanding seems 00:08:15.759 --> 00:08:18.139 key for applications where different data types 00:08:18.139 --> 00:08:21.160 merge. Like analyzing medical scans alongside 00:08:21.160 --> 00:08:24.000 doctor's notes or creating marketing campaigns 00:08:24.000 --> 00:08:26.480 where the images and text really work together 00:08:26.480 --> 00:08:28.819 seamlessly. Exactly. It's not just linking different 00:08:28.819 --> 00:08:31.420 data. It's fusing the concepts together at a 00:08:31.420 --> 00:08:33.120 deeper level. Okay, now let's talk efficiency. 00:08:33.360 --> 00:08:36.279 Because not every task needs a planet -sized 00:08:36.279 --> 00:08:38.500 AI. There's a big move towards small language 00:08:38.500 --> 00:08:41.419 models, SLMs too. A huge move. Yeah, we're talking 00:08:41.419 --> 00:08:44.470 about much more focused AIs. maybe 3 million 00:08:44.470 --> 00:08:46.889 parameters, up to a few hundred million. You 00:08:46.889 --> 00:08:49.409 trade that broad general knowledge of a giant 00:08:49.409 --> 00:08:52.570 LLM for incredibly sharp expert level skill on 00:08:52.570 --> 00:08:55.509 one specific narrow task. And the benefits are 00:08:55.509 --> 00:08:56.950 massive. They're way cheaper to run, they're 00:08:56.950 --> 00:08:59.470 much faster, and you get tighter control, especially 00:08:59.470 --> 00:09:01.730 over your own private data. You can fine tune 00:09:01.730 --> 00:09:04.269 an SLM on your company's specific jargon or processes 00:09:04.269 --> 00:09:07.190 and get amazing performance just for that niche. 00:09:07.529 --> 00:09:09.889 So instead of paying for the giant general practitioner 00:09:09.889 --> 00:09:13.759 LLM for everything, you deploy a a cost -effective 00:09:13.759 --> 00:09:17.259 specialist, SLM. Maybe one for customer service 00:09:17.259 --> 00:09:20.600 queries, another for summarizing legal docs that 00:09:20.600 --> 00:09:23.399 does its one job brilliantly and cheaply. That's 00:09:23.399 --> 00:09:25.820 the play. It's the smart way to specialize and 00:09:25.820 --> 00:09:28.759 scale AI across very specific business functions 00:09:28.759 --> 00:09:30.860 without breaking the bank. But what if you want 00:09:30.860 --> 00:09:33.659 the smarts of the big model, but need the speed 00:09:33.659 --> 00:09:36.519 and cost of the small one? That's where distillation 00:09:36.519 --> 00:09:38.860 comes in. Precisely. Distillation is this cool 00:09:38.860 --> 00:09:42.320 teacher -student process. You take a huge, knowledgeable 00:09:42.320 --> 00:09:45.580 teacher model and essentially compress its wisdom 00:09:45.580 --> 00:09:49.500 into a smaller, faster student model. How? Well, 00:09:49.539 --> 00:09:51.639 you feed the same prompts to both models. Then 00:09:51.639 --> 00:09:53.340 you train the... student model, not just to get 00:09:53.340 --> 00:09:55.440 the right answer, but to mimic the way the teacher 00:09:55.440 --> 00:09:58.000 model arrives at its answer, matching its internal 00:09:58.000 --> 00:10:00.600 patterns and probability outputs. So you're basically 00:10:00.600 --> 00:10:02.639 downloading the teacher's expertise, or most 00:10:02.639 --> 00:10:05.139 of it, into a lean, production -ready student. 00:10:05.399 --> 00:10:08.419 That's the idea. You want a smaller, faster model 00:10:08.419 --> 00:10:10.759 that's really optimized for running millions 00:10:10.759 --> 00:10:13.000 of times in production, saving time and money. 00:10:13.470 --> 00:10:16.330 you usually accept a tiny, almost negligible 00:10:16.330 --> 00:10:19.549 loss in nuance, but gain massively in efficiency. 00:10:19.909 --> 00:10:22.049 And the final optimization trick, really down 00:10:22.049 --> 00:10:24.929 on the weeds technically, quantization, shrinking 00:10:24.929 --> 00:10:26.950 the memory footprint. Yeah, this one's purely 00:10:26.950 --> 00:10:28.970 about efficiency. It's about reducing the precision 00:10:28.970 --> 00:10:31.169 of the numbers the model uses for its internal 00:10:31.169 --> 00:10:33.590 weight. So instead of using super precise, like 00:10:33.590 --> 00:10:36.779 32 -bit... floating point numbers you can press 00:10:36.779 --> 00:10:39.419 them down maybe to simpler 8 -bit integers it's 00:10:39.419 --> 00:10:41.779 kind of like saving a huge high -res photo file 00:10:41.779 --> 00:10:44.580 as a smaller jpeg you lose a microscopic bit 00:10:44.580 --> 00:10:47.299 of fidelity maybe but the file size reduction 00:10:47.299 --> 00:10:50.379 is enormous quantization often slashes the memory 00:10:50.379 --> 00:10:53.460 needed by like 75 percent And that smaller memory 00:10:53.460 --> 00:10:56.700 size, even with almost no noticeable change in 00:10:56.700 --> 00:10:59.639 output quality, drastically cuts the cost of 00:10:59.639 --> 00:11:02.120 running the model and lets powerful AI run on 00:11:02.120 --> 00:11:04.740 way less powerful hardware. Exactly. It's what 00:11:04.740 --> 00:11:07.340 makes it feasible to run impressive AI on your 00:11:07.340 --> 00:11:11.320 phone or on small sensors, edge devices. Whoa. 00:11:11.779 --> 00:11:14.419 Imagine scaling that efficiency across like... 00:11:14.960 --> 00:11:18.059 a billion queries a day globally. That's how 00:11:18.059 --> 00:11:20.720 this tech becomes truly everywhere. Wow. Okay. 00:11:20.799 --> 00:11:22.919 We've covered a lot of ground with these 10 concepts 00:11:22.919 --> 00:11:25.139 really fast. Let's try and weave it all together 00:11:25.139 --> 00:11:27.860 now. How do these pieces fit into a complete 00:11:27.860 --> 00:11:30.700 modern AI system? Okay. Let's trace a complex 00:11:30.700 --> 00:11:33.990 request. Input comes in, gets tokenized, the 00:11:33.990 --> 00:11:36.409 system needs context, it grabs internal knowledge 00:11:36.409 --> 00:11:40.330 using RG, and crucially, uses MCP to pull in 00:11:40.330 --> 00:11:42.909 real -time external data or trigger actions out 00:11:42.909 --> 00:11:45.470 in the world. Then the brain kicks in, the reasoning 00:11:45.470 --> 00:11:48.210 core. The transformer architecture chews on all 00:11:48.210 --> 00:11:50.129 that info. It might use chain of thought to ensure 00:11:50.129 --> 00:11:52.309 the steps are logical and transparent, or even 00:11:52.309 --> 00:11:54.330 advanced reasoning models to figure out the best 00:11:54.330 --> 00:11:56.730 strategy on the fly. And if there's images or 00:11:56.730 --> 00:11:59.309 audio involved, multimodal capabilities handle 00:11:59.309 --> 00:12:01.330 that. And this whole thing isn't just a one -off 00:12:01.330 --> 00:12:03.309 process. It's likely running as an intelligent, 00:12:03.450 --> 00:12:06.590 proactive AI agent. That agent is using sophisticated 00:12:06.590 --> 00:12:09.210 context engineering to manage the long conversation 00:12:09.210 --> 00:12:12.230 or task, remembering what happened before. Right. 00:12:12.289 --> 00:12:15.009 And its overall behavior, the way it responds 00:12:15.009 --> 00:12:17.649 and acts, has been fine -tuned using reinforcement 00:12:17.649 --> 00:12:20.029 learning to make sure it aligns with what users 00:12:20.029 --> 00:12:22.399 actually want and find helpful. And then finally, 00:12:22.519 --> 00:12:24.779 before it ever gets deployed to millions of users, 00:12:25.000 --> 00:12:27.480 that whole architecture has been squeezed for 00:12:27.480 --> 00:12:29.899 efficiency. Its core knowledge might have been 00:12:29.899 --> 00:12:32.320 compressed using distillation from a bigger model, 00:12:32.440 --> 00:12:35.039 and its final memory size shrunk drastically 00:12:35.039 --> 00:12:38.159 via quantization. Yeah, getting comfortable with 00:12:38.159 --> 00:12:40.019 this whole vocabulary, it gives you a massive 00:12:40.019 --> 00:12:42.240 strategic edge. You could design smarter systems, 00:12:42.299 --> 00:12:44.759 make better choices like, do I need a specialized 00:12:44.759 --> 00:12:48.240 SLM here or a big LM hooked up with MCP and RG 00:12:48.240 --> 00:12:50.559 and cut through the hype? Knowing this stuff 00:12:50.559 --> 00:12:53.659 is potential power, but using it, that's real 00:12:53.659 --> 00:12:56.980 power. You now have the language for this next 00:12:56.980 --> 00:12:59.950 wave of AI. So what should you do? Three things. 00:13:00.070 --> 00:13:02.990 First, start using this language. Put terms like 00:13:02.990 --> 00:13:05.789 agent, MCP, COT into your notes, your discussions, 00:13:05.889 --> 00:13:08.250 your project plans. Show you understand what's 00:13:08.250 --> 00:13:10.730 under the hood. Second, look at the AI tools 00:13:10.730 --> 00:13:13.429 you already use. If something acts proactively, 00:13:13.610 --> 00:13:16.649 ask. How does it remember things? Is that context 00:13:16.649 --> 00:13:18.769 engineering? Is it acting like an agent? Try 00:13:18.769 --> 00:13:21.720 to deconstruct it. Don't try to master everything 00:13:21.720 --> 00:13:24.299 at once. Pick maybe two or three concepts that 00:13:24.299 --> 00:13:26.860 feel most relevant to what you do. Maybe it's 00:13:26.860 --> 00:13:29.519 agents, RIG, and multimodal if you're building 00:13:29.519 --> 00:13:32.519 user -facing apps. Go deep on those. Yeah, because 00:13:32.519 --> 00:13:34.279 understanding these architectural choices means 00:13:34.279 --> 00:13:36.779 you're not just reacting to AI trends. You're 00:13:36.779 --> 00:13:39.559 actually equipped to lead the next phase of building 00:13:39.559 --> 00:13:42.220 genuinely smart, useful systems. Because really, 00:13:42.340 --> 00:13:44.480 these concepts show the future isn't just about 00:13:44.480 --> 00:13:46.899 making models bigger and bigger. It's about smart 00:13:46.899 --> 00:13:49.139 integration, making them efficient, helping them 00:13:49.139 --> 00:13:51.379 learn, continue. continuously, building systems 00:13:51.379 --> 00:13:53.259 that can actually connect, think and act for 00:13:53.259 --> 00:13:54.399 us in the real world.