WEBVTT 00:00:00.000 --> 00:00:04.299 Klarna just replaced 700 employees with a single 00:00:04.299 --> 00:00:07.980 AI chatbot. Yeah, it's a staggering shift, honestly. 00:00:08.119 --> 00:00:11.240 I mean, Duolingo just cut 10 % of its translation 00:00:11.240 --> 00:00:13.880 team. For the exact same reason. Right, exactly. 00:00:14.320 --> 00:00:16.399 And those are just the big names, you know. Goldman 00:00:16.399 --> 00:00:19.199 Sachs estimates AI could replace 300 million 00:00:19.199 --> 00:00:21.719 jobs globally. It's not a scare tactic. It's 00:00:21.719 --> 00:00:24.170 simply the current data. Absolutely. The people 00:00:24.170 --> 00:00:26.250 who survive this shift aren't necessarily the 00:00:26.250 --> 00:00:27.949 smartest ones. They're the ones who learn to 00:00:27.949 --> 00:00:30.129 work with AI. They figure it out before the rest 00:00:30.129 --> 00:00:32.609 of the industry catches up. Welcome to the Deep 00:00:32.609 --> 00:00:35.890 Dive. Today, our mission is intensely practical. 00:00:36.189 --> 00:00:39.229 We're unpacking a specific AI engineer roadmap. 00:00:39.750 --> 00:00:42.250 It's designed for you, even if you are starting 00:00:42.250 --> 00:00:44.850 from absolute zero. We are covering five distinct 00:00:44.850 --> 00:00:46.990 levels to future -proof your career. And the 00:00:46.990 --> 00:00:49.549 best part, you do not need a computer science 00:00:49.549 --> 00:00:51.969 degree for this. You don't need complex math 00:00:51.969 --> 00:00:54.200 theory either. You simply need to follow the 00:00:54.200 --> 00:00:57.520 path. Before we climb those five levels, we must 00:00:57.520 --> 00:00:59.579 define the game. What are we actually playing 00:00:59.579 --> 00:01:02.579 here? There is massive confusion out there. People 00:01:02.579 --> 00:01:05.120 constantly mix up data science and AI engineering. 00:01:05.519 --> 00:01:08.319 Oh, it happens all the time. That confusion sends 00:01:08.319 --> 00:01:11.120 a lot of people in the completely wrong direction. 00:01:12.099 --> 00:01:13.859 Here's a simple way to think about the difference. 00:01:14.340 --> 00:01:16.939 Let's use a house analogy. I like analogies. 00:01:17.120 --> 00:01:20.140 A software engineer builds the actual house. 00:01:20.459 --> 00:01:23.019 They pour the foundation. They build the walls, 00:01:23.200 --> 00:01:25.299 the roof, and the doors. They create the structural 00:01:25.299 --> 00:01:28.260 integrity. An AI engineer, though, makes that 00:01:28.260 --> 00:01:31.040 house smart. You add the sensors that turn on 00:01:31.040 --> 00:01:33.319 the lights. You add the voice assistant for the 00:01:33.319 --> 00:01:36.540 music. That is your specific layer. So you aren't 00:01:36.540 --> 00:01:39.450 building these algorithms from scratch. Instead, 00:01:39.489 --> 00:01:42.390 you take existing intelligence. You elegantly 00:01:42.390 --> 00:01:44.849 plug it into real -world products. You solve 00:01:44.849 --> 00:01:47.689 problems with practical tools. Exactly. You don't 00:01:47.689 --> 00:01:50.489 need to be a math genius inventing new algorithms. 00:01:50.670 --> 00:01:52.790 You just need to know how to pick the right tools. 00:01:53.049 --> 00:01:55.310 You make them work together seamlessly. It's 00:01:55.310 --> 00:01:57.510 like we are moving from building the engines 00:01:57.510 --> 00:01:59.849 to simply driving the cars. We are operating 00:01:59.849 --> 00:02:02.170 the machines now. We just need to know how to 00:02:02.170 --> 00:02:04.209 steer. That is the perfect way to look at it. 00:02:04.209 --> 00:02:06.849 You are the driver. The Frontier models, well, 00:02:06.870 --> 00:02:10.069 they're your engine. To sex silence? I still 00:02:10.069 --> 00:02:13.129 wrestle with prompt drift myself. The unpredictability 00:02:13.129 --> 00:02:15.550 of it all. It can be quite humbling. So we aren't 00:02:15.550 --> 00:02:17.729 building the brain, just the nervous system. 00:02:18.090 --> 00:02:20.870 Precisely. You're wiring the connections. You 00:02:20.870 --> 00:02:23.490 let the massive pre -trained brains do the heavy 00:02:23.490 --> 00:02:26.129 lifting. You just direct their energy exactly 00:02:26.129 --> 00:02:29.090 where it needs to go. Exactly. We connect existing 00:02:29.090 --> 00:02:32.370 intelligence to real -world applications. Now 00:02:32.370 --> 00:02:35.219 we understand the job itself. But you cannot 00:02:35.219 --> 00:02:37.620 build a skyscraper on sand. No, you definitely 00:02:37.620 --> 00:02:40.240 cannot. We must lay the groundwork first. We 00:02:40.240 --> 00:02:43.500 have to do this before touching shiny AI tools. 00:02:43.719 --> 00:02:46.180 We use the SALT framework. Let's break down the 00:02:46.180 --> 00:02:49.060 SALT foundation. The S stands for software fluency. 00:02:49.280 --> 00:02:51.639 Specifically, you need to learn Python. Python 00:02:51.639 --> 00:02:54.000 is the language of AI. Yeah. Almost everything 00:02:54.000 --> 00:02:56.319 you do in this field runs on Python. It's actually 00:02:56.319 --> 00:02:58.419 one of the easiest languages to pick up. But 00:02:58.419 --> 00:03:00.439 beginners often fall into a nasty trap. They 00:03:00.439 --> 00:03:03.080 fall into tutorial hell. Yes. You watch a three 00:03:03.080 --> 00:03:05.099 -hour coding video. You feel like you completely 00:03:05.099 --> 00:03:07.719 get it. Then you open a blank file. Your mind 00:03:07.719 --> 00:03:10.139 goes completely empty. The fix is incredibly 00:03:10.139 --> 00:03:13.060 simple, but people resist it. You must learn 00:03:13.060 --> 00:03:16.740 by doing. You can't just watch passively. A tool 00:03:16.740 --> 00:03:19.900 like CAUTI .Tech forces you to write code immediately. 00:03:20.199 --> 00:03:21.860 It makes the concepts actually stick in your 00:03:21.860 --> 00:03:24.180 brain. You need to focus on four core things. 00:03:24.759 --> 00:03:27.319 Variables are first. Think of variables as digital 00:03:27.319 --> 00:03:30.620 buckets. When that massive AI brain sends back 00:03:30.620 --> 00:03:33.120 an answer, your program needs a bucket to catch 00:03:33.120 --> 00:03:36.659 it. Next, you need lists and dictionaries. That's 00:03:36.659 --> 00:03:39.699 how you organize lots of data. If the AI returns 00:03:39.699 --> 00:03:41.900 50 different estimate names, you need a structured 00:03:41.900 --> 00:03:43.860 way to hold them. Then you move to functions. 00:03:44.199 --> 00:03:46.599 This is how you write reusable code. You package 00:03:46.599 --> 00:03:48.939 a set of instructions together. You avoid repeating 00:03:48.939 --> 00:03:51.599 yourself. Finally, you have loops. Loops make 00:03:51.599 --> 00:03:54.259 the computer repeat a task automatically. If 00:03:54.259 --> 00:03:56.960 you have 100 emails, a loop processes them one 00:03:56.960 --> 00:03:59.439 by one. You write the code once. That covers 00:03:59.439 --> 00:04:02.620 the software fluency part. The A in salt is API 00:04:02.620 --> 00:04:05.060 architecture. Once you know Python, you need 00:04:05.060 --> 00:04:07.520 a way to talk to the AI. APIs are how models 00:04:07.520 --> 00:04:10.340 receive your requests. The easiest way to picture 00:04:10.340 --> 00:04:12.860 it is a waiter at a restaurant. You are the customer 00:04:12.860 --> 00:04:15.000 sitting at the table. And the kitchen in the 00:04:15.000 --> 00:04:17.339 back is the AI model. You don't walk into the 00:04:17.339 --> 00:04:19.439 kitchen yourself. You tell the waiter what you 00:04:19.439 --> 00:04:21.910 want. They relay your order to the kitchen, then 00:04:21.910 --> 00:04:24.170 they bring your cooked food back to your table. 00:04:24.670 --> 00:04:26.889 APIs work exactly the same way. In practice, 00:04:26.889 --> 00:04:29.430 you use Python's requests library. It acts as 00:04:29.430 --> 00:04:31.930 your waiter. It sends those digital messages 00:04:31.930 --> 00:04:34.810 for you behind the scenes. Now, for the L, that 00:04:34.810 --> 00:04:38.110 stands for life cycle and version control. Specifically, 00:04:38.230 --> 00:04:40.810 we're talking about a tool called Git. Every 00:04:40.810 --> 00:04:43.850 professional project involves mistakes. Code 00:04:43.850 --> 00:04:46.430 breaks all the time. Features get deleted by 00:04:46.430 --> 00:04:49.639 accident. A tiny typo can crash an entire server. 00:04:49.819 --> 00:04:52.360 Panic sets in. Git is the safety net. It lets 00:04:52.360 --> 00:04:54.980 you roll back to any previous version. It tracks 00:04:54.980 --> 00:04:57.660 every single change you make over time. It is 00:04:57.660 --> 00:05:00.139 basically an undo button for your entire code 00:05:00.139 --> 00:05:02.339 base. GitHub is the website where those saves 00:05:02.339 --> 00:05:05.160 live online. Every single tech company uses it 00:05:05.160 --> 00:05:08.139 today. Finally, the T is the text deck. This 00:05:08.139 --> 00:05:10.560 is building end -to -end. You need a database, 00:05:10.800 --> 00:05:13.519 like MongoDB, to store user information permanently. 00:05:13.800 --> 00:05:16.980 You need a backend, like Flask or FastAPI. That 00:05:16.980 --> 00:05:19.879 handles the unseen logic. And you need a frontend, 00:05:20.120 --> 00:05:22.839 like React, for the visual user interface. Can't 00:05:22.839 --> 00:05:25.019 we just skip to the AI stuff and learn Python 00:05:25.019 --> 00:05:27.389 later? You really can't. Without Python, you 00:05:27.389 --> 00:05:29.250 have no way to communicate with the models. You 00:05:29.250 --> 00:05:31.629 would just be staring at APIs with no way to 00:05:31.629 --> 00:05:34.389 connect them to anything. No. Python is the absolute 00:05:34.389 --> 00:05:37.629 necessary foundation for everything else. Now 00:05:37.629 --> 00:05:39.990 that the foundation is poured, we move to level 00:05:39.990 --> 00:05:43.750 two. It is time to invite the AI inside. This 00:05:43.750 --> 00:05:46.649 level is controlled intelligence. You are finally 00:05:46.649 --> 00:05:49.350 using the best models available. Let's start 00:05:49.350 --> 00:05:52.389 with the OpenAI API. You have probably used the 00:05:52.389 --> 00:05:55.500 ChatGPT website. But as an engineer, you interact 00:05:55.500 --> 00:05:57.959 with it completely differently. You write code 00:05:57.959 --> 00:06:01.230 to automate it. The difference is massive. Instead 00:06:01.230 --> 00:06:04.029 of manually typing into a chat box, your script 00:06:04.029 --> 00:06:07.370 runs things automatically. It operates at superhuman 00:06:07.370 --> 00:06:10.870 speeds. Think about a script processing 100 customer 00:06:10.870 --> 00:06:14.509 emails. It flags complaints instantly. It extracts 00:06:14.509 --> 00:06:16.810 the emotional sentiment in seconds. It sorts 00:06:16.810 --> 00:06:19.649 them into positive, neutral, or negative categories. 00:06:20.009 --> 00:06:22.329 You do not touch a single thing manually. Your 00:06:22.329 --> 00:06:24.870 script handles the entire inbox. The prompt you 00:06:24.870 --> 00:06:27.029 write in the code is so simple. You tell the 00:06:27.029 --> 00:06:29.110 model to act as a helpful assistant. You ask 00:06:29.110 --> 00:06:31.180 it for the sentiment. You ask it for the main 00:06:31.180 --> 00:06:34.319 problem in one sentence. You wrap that in a Python 00:06:34.319 --> 00:06:37.620 loop. It processes an entire inbox instantly. 00:06:38.000 --> 00:06:40.639 Once you understand how to automate OpenAI, you 00:06:40.639 --> 00:06:44.220 explore hugging face. Hugging Face is an incredible 00:06:44.220 --> 00:06:46.980 resource. Think of it as the app store for AI 00:06:46.980 --> 00:06:49.540 models. The best part is that most of it is completely 00:06:49.540 --> 00:06:52.639 free. You don't have to rely entirely on expensive, 00:06:52.819 --> 00:06:55.379 closed models. It handles so many different tasks. 00:06:55.680 --> 00:06:57.699 It handles language translation. It gives you 00:06:57.699 --> 00:06:59.759 image descriptions. It handles sound transcription. 00:07:00.060 --> 00:07:02.360 You just pick pre -trained models off the shelf. 00:07:02.540 --> 00:07:04.740 You run them locally. If you want a structured 00:07:04.740 --> 00:07:06.899 way to practice all this, look at DataCam. They 00:07:06.899 --> 00:07:09.519 have an associate AI engineer track. It covers 00:07:09.519 --> 00:07:12.199 exactly what you need to master these tools. 00:07:12.839 --> 00:07:15.639 Is Hugging Face just a repository for text generation 00:07:15.639 --> 00:07:18.420 models? Not at all. Text is just the beginning. 00:07:18.579 --> 00:07:21.000 It has thousands of models for analyzing images, 00:07:21.420 --> 00:07:23.680 processing audio files, and translating spoken 00:07:23.680 --> 00:07:26.060 languages in real time. No, it offers audio, 00:07:26.199 --> 00:07:28.120 vision, and language translation models too. 00:07:28.300 --> 00:07:31.459 Sponsor. Generic models are great. But in 2026, 00:07:31.860 --> 00:07:34.620 companies don't want a generic AI. They want 00:07:34.620 --> 00:07:37.899 an AI that deeply knows their specific data. 00:07:38.199 --> 00:07:41.319 That brings us to level three. Intelligent systems 00:07:41.319 --> 00:07:44.920 and RRAG. This is the heart of the roadmap. This 00:07:44.920 --> 00:07:46.800 is what companies are actually paying big money 00:07:46.800 --> 00:07:50.779 for. Let's define this jargon. What is ARG? Giving 00:07:50.779 --> 00:07:53.800 an AI your custom data to read before it answers. 00:07:53.879 --> 00:07:56.759 That is the core idea. Imagine you have a 500 00:07:56.759 --> 00:07:59.899 page policy manual for a company. A standard 00:07:59.899 --> 00:08:03.199 model, well, it does not know what is inside 00:08:03.199 --> 00:08:05.360 that specific manual. If a customer asks about 00:08:05.360 --> 00:08:08.120 a niche return policy, the model might just guess. 00:08:08.459 --> 00:08:11.740 It hallucinates. Argi fixes this problem entirely. 00:08:11.959 --> 00:08:13.879 There are four clear steps here to make it work. 00:08:13.959 --> 00:08:16.519 Step one, you break that massive manual into 00:08:16.519 --> 00:08:18.819 small, manageable chunks. Step two, you turn 00:08:18.819 --> 00:08:20.899 those pieces into embeddings. Embeddings are 00:08:20.899 --> 00:08:22.980 just coordinates. It mathematically turns words 00:08:22.980 --> 00:08:25.459 into numbers. Exactly. The AI understands those 00:08:25.459 --> 00:08:28.279 numbers deeply. If two concepts mean similar 00:08:28.279 --> 00:08:30.259 things, their numerical coordinates sit close 00:08:30.259 --> 00:08:32.399 together. Step three, you store those numbers 00:08:32.399 --> 00:08:34.840 in a vector database. Tools like Pinecone or 00:08:34.840 --> 00:08:36.860 Weaviate are built specifically to hold these 00:08:36.860 --> 00:08:40.000 coordinates. Step four is the magic. A user asks 00:08:40.000 --> 00:08:42.860 a question. Your system rapidly searches the 00:08:42.860 --> 00:08:45.379 database. It finds the most relevant pieces of 00:08:45.379 --> 00:08:48.039 information. It feeds those specific pieces to 00:08:48.039 --> 00:08:50.779 the AI. The AI reads them. Then it answers the 00:08:50.779 --> 00:08:53.000 user. It sounds like an absolute expert on your 00:08:53.000 --> 00:08:54.879 company. Here is a prompt example you would write 00:08:54.879 --> 00:08:58.240 in the code. You tell the AI, use this context, 00:08:58.299 --> 00:09:00.580 don't make things up. You pass it the chunk containing 00:09:00.580 --> 00:09:03.820 a 30 -day return policy. The user asks, can I 00:09:03.820 --> 00:09:06.980 return this laptop after 40 days? The AI confidently 00:09:06.980 --> 00:09:09.960 says no. It bases that answer purely on the reference 00:09:09.960 --> 00:09:12.299 notes you just handed it. It doesn't guess. It 00:09:12.299 --> 00:09:15.159 is brilliant. We also have LandGraph and MCP 00:09:15.159 --> 00:09:17.840 in this level. LandGraph allows the AI to think 00:09:17.840 --> 00:09:21.139 in logical steps. It is fascinating. The AI can 00:09:21.139 --> 00:09:23.750 decide to search the web, check a database, and 00:09:23.750 --> 00:09:25.549 then write a summary. It follows a workflow. 00:09:25.889 --> 00:09:29.049 MCP stands for Model Context Protocol. It provides 00:09:29.049 --> 00:09:31.809 strict rules for external tools. It tells the 00:09:31.809 --> 00:09:34.190 AI exactly how to use a calculator safely. It 00:09:34.190 --> 00:09:35.610 creates boundaries. Yeah. You don't want the 00:09:35.610 --> 00:09:39.330 AI doing whatever it wants. MCP acts as the supervisor 00:09:39.330 --> 00:09:41.789 for the tools it accesses. Two secs silence. 00:09:42.269 --> 00:09:45.429 Whoa. Imagine stacking Lego blocks of data, scaling 00:09:45.429 --> 00:09:47.970 that up to a billion queries across an entire 00:09:47.970 --> 00:09:50.490 company's history. It's truly staggering. Does 00:09:50.490 --> 00:09:53.889 Arsh physically alter the AI model's core training? 00:09:54.230 --> 00:09:56.309 It does not touch the core training at all. The 00:09:56.309 --> 00:09:58.950 underlying model remains exactly the same. You 00:09:58.950 --> 00:10:01.110 are simply sliding a cheat sheet across the desk 00:10:01.110 --> 00:10:03.769 before it speaks. No, it just gives the AI reference 00:10:03.769 --> 00:10:06.870 notes before answering. We now have a brilliant 00:10:06.870 --> 00:10:10.129 customized AI running perfectly on our laptop. 00:10:10.529 --> 00:10:13.629 We arrive at level four. Putting apps online. 00:10:13.970 --> 00:10:16.950 How do you let 10 ,000 people use your app without 00:10:16.950 --> 00:10:19.889 it instantly crashing? That is a completely different 00:10:19.889 --> 00:10:22.370 engineering problem. We start with Docker. Every 00:10:22.370 --> 00:10:24.889 developer knows the pain of sharing code. You 00:10:24.889 --> 00:10:27.389 say, it works perfectly on my machine. But you 00:10:27.389 --> 00:10:29.679 send it away and it breaks immediately. The client 00:10:29.679 --> 00:10:32.039 has an older version of Python. The libraries 00:10:32.039 --> 00:10:35.139 clash. Docker fixes this massive headache. It 00:10:35.139 --> 00:10:37.580 wraps the app in a protective container. It packs 00:10:37.580 --> 00:10:39.940 everything the app needs to run inside that bubble. 00:10:40.000 --> 00:10:42.480 It is entirely self -sufficient. It runs identically 00:10:42.480 --> 00:10:44.559 anywhere in the world. It is completely consistent. 00:10:44.759 --> 00:10:47.100 You eliminate the environmental variables that 00:10:47.100 --> 00:10:49.279 cause random crashes. Then we have cloud hosting. 00:10:49.620 --> 00:10:52.500 You cannot run a popular app from your bedroom 00:10:52.500 --> 00:10:55.639 laptop. You use AWS or Google Cloud. You put 00:10:55.639 --> 00:10:58.240 your containerized app on their massive, powerful 00:10:58.240 --> 00:11:01.179 servers. This is how you dynamically scale up 00:11:01.179 --> 00:11:03.879 during sudden traffic spikes. If a thousand people 00:11:03.879 --> 00:11:06.539 log on at once, the cloud handles the load gracefully. 00:11:06.759 --> 00:11:09.480 It spins up more resources automatically. But 00:11:09.480 --> 00:11:13.100 there is a catch. Every API call to an AI model 00:11:13.100 --> 00:11:16.360 costs actual money. That brings us to Redis Caching. 00:11:16.730 --> 00:11:19.710 This is the secret weapon of professional engineers. 00:11:19.970 --> 00:11:22.730 It is pure economics. Every time you ask the 00:11:22.730 --> 00:11:25.570 AI a question, a computer somewhere burns energy. 00:11:25.769 --> 00:11:28.610 You pay for that energy. Imagine 1000 people 00:11:28.610 --> 00:11:31.590 asking for the exact same return policy. Without 00:11:31.590 --> 00:11:34.750 caching, you pay for 1000 separate AI generations. 00:11:34.929 --> 00:11:37.909 The bill adds up incredibly fast. Redis solves 00:11:37.909 --> 00:11:40.549 this beautifully. It stores the answer the very 00:11:40.549 --> 00:11:42.929 first time it is generated. It returns that saved 00:11:42.929 --> 00:11:46.769 answer instantly the next 999 times. Cost for 00:11:46.769 --> 00:11:49.129 those repeat answers is absolutely zero. It is 00:11:49.129 --> 00:11:51.090 the difference between a hobbyist leaking money 00:11:51.090 --> 00:11:53.909 and a viable tech business. Is Docker literally 00:11:53.909 --> 00:11:56.649 just a heavy virtual machine? It is very different 00:11:56.649 --> 00:11:59.570 from a virtual machine. A virtual machine loads 00:11:59.570 --> 00:12:02.769 an entire heavy operating system. Docker just 00:12:02.769 --> 00:12:04.889 isolates the app and its immediate dependencies, 00:12:05.350 --> 00:12:07.669 making it incredibly fast. It's much lighter, 00:12:07.870 --> 00:12:11.149 packing only exactly what the app needs. We finally 00:12:11.149 --> 00:12:13.769 arrive at level five, managing a professional 00:12:13.769 --> 00:12:17.970 roadmap. also known as LLMOAPES. The app is live. 00:12:18.590 --> 00:12:20.350 Millions are potentially using it. It is cheap 00:12:20.350 --> 00:12:23.870 to run. But is it actually giving good answers? 00:12:24.200 --> 00:12:26.700 Or is it embarrassing the company? You must shift 00:12:26.700 --> 00:12:28.500 your mindset completely here. You go from being 00:12:28.500 --> 00:12:30.320 a builder to being a business owner. You have 00:12:30.320 --> 00:12:32.960 to monitor the outcomes constantly. Models hallucinate. 00:12:33.080 --> 00:12:35.639 They get things wildly wrong. You cannot manually 00:12:35.639 --> 00:12:38.200 read 10 ,000 automated customer service chats. 00:12:38.740 --> 00:12:41.360 We use DeepView for this. It runs automated quality 00:12:41.360 --> 00:12:44.340 checks continuously. It flags inaccurate answers. 00:12:44.500 --> 00:12:47.299 It catches unhelpful or off -tone responses before 00:12:47.299 --> 00:12:49.799 PR disasters happen. It is your automated quality 00:12:49.799 --> 00:12:52.379 assurance team. Then we have analytics. Tools 00:12:52.379 --> 00:12:55.139 like PostFog track actual user behavior inside 00:12:55.139 --> 00:12:56.980 the app. You see where users get frustrated. 00:12:57.279 --> 00:12:59.120 You see exactly where they drop off entirely. 00:12:59.500 --> 00:13:01.379 The data tells you exactly what parts of your 00:13:01.379 --> 00:13:04.039 app need fixing. Finally, we have model routing. 00:13:04.500 --> 00:13:07.019 This is incredibly important for long -term efficiency. 00:13:07.279 --> 00:13:09.700 You don't need a supercomputer for every single 00:13:09.700 --> 00:13:13.200 question. You send simple FAQs to a fast sheet 00:13:13.200 --> 00:13:17.419 model, like Claude Haiku. You only send complex, 00:13:17.539 --> 00:13:20.870 nuanced tasks to a powerful model. Like Claude 00:13:20.870 --> 00:13:22.950 Opus. Model routing is like having an ecosystem. 00:13:23.049 --> 00:13:25.549 You don't ask the village elder to chop wood. 00:13:25.909 --> 00:13:27.970 You don't ask the lumberjack for cosmic wisdom. 00:13:28.190 --> 00:13:30.970 You match the task appropriately. That is a very 00:13:30.970 --> 00:13:32.710 philosophical way to view it. It makes perfect 00:13:32.710 --> 00:13:35.090 sense. Isn't routing between different models 00:13:35.090 --> 00:13:37.409 over complicating the architecture? It adds a 00:13:37.409 --> 00:13:40.429 routing step, sure, but the payoff is undeniable. 00:13:40.830 --> 00:13:43.350 The cheaper models respond faster and they cost 00:13:43.350 --> 00:13:45.889 a fraction of the frontier models. No, it saves 00:13:45.889 --> 00:13:48.009 massive amounts of computing money and time. 00:13:48.250 --> 00:13:50.769 This brings us to the end of our roadmap. Let's 00:13:50.769 --> 00:13:53.470 do a big idea recap. We covered a lot of crucial 00:13:53.470 --> 00:13:56.549 ground today. Becoming an AI engineer in 2026 00:13:56.549 --> 00:13:59.309 isn't about being a math genius. It is about 00:13:59.309 --> 00:14:01.870 consistency. It's about laying a solid, salt 00:14:01.870 --> 00:14:04.610 foundation. You connect APIs cleanly. You ground 00:14:04.610 --> 00:14:07.169 your data strictly with a rag. You deploy safely 00:14:07.169 --> 00:14:10.009 in containers. You monitor your operations constantly. 00:14:10.669 --> 00:14:13.110 A smart doorbell is totally useless if the walls 00:14:13.110 --> 00:14:16.350 fall down. Start with level one today. Learn 00:14:16.350 --> 00:14:19.269 Python. Build a simple website first. The high 00:14:19.269 --> 00:14:21.389 salaries exist because it is hard, persistent 00:14:21.389 --> 00:14:24.389 work, but it is deeply achievable. You just have 00:14:24.389 --> 00:14:26.769 to follow the roadmap step by step. Don't skip 00:14:26.769 --> 00:14:29.350 the basics. Put the time in. You will be miles 00:14:29.350 --> 00:14:31.169 ahead of the curve before the rest of your industry 00:14:31.169 --> 00:14:33.960 even wakes up. Neat. We've spent this time talking 00:14:33.960 --> 00:14:36.720 about how AI engineers connect tools to automate 00:14:36.720 --> 00:14:39.840 tasks. But as models get smarter at coding, how 00:14:39.840 --> 00:14:42.220 long until the AI engineer relies on an AI to 00:14:42.220 --> 00:14:44.879 automate the AI engineering roadmap itself? Something 00:14:44.879 --> 00:14:46.220 to ponder. Keep exploring.