WEBVTT 1 00:00:00.160 --> 00:00:03.399 Welcome to the deep dive. We take stacks of research, articles, 2 00:00:03.439 --> 00:00:06.719 notes and boil it all down for you. Today, we're 3 00:00:06.719 --> 00:00:11.320 really digging into something that's well everywhere, generative AI. How 4 00:00:11.320 --> 00:00:15.400 it's changing content creation, sure, but also sparking these huge 5 00:00:15.480 --> 00:00:18.879 questions about the future of work, especially for jobs where 6 00:00:19.359 --> 00:00:22.719 you know, AI is getting surprisingly good at the core tasks. 7 00:00:22.839 --> 00:00:26.760 Yeah, exactly, And our focus today is this fascinating guide 8 00:00:26.800 --> 00:00:29.280 Coding with AI by Chris Minnick. We really want to 9 00:00:29.359 --> 00:00:32.439 unpack how AI isn't just like another tool for programmers. 10 00:00:32.560 --> 00:00:34.719 It could actually make you a way better, faster, more 11 00:00:34.719 --> 00:00:35.600 efficient coder. 12 00:00:35.840 --> 00:00:38.479 Right, So the big question we're tackling in this deep 13 00:00:38.520 --> 00:00:42.119 dive is, well, how does this stuff actually work? And 14 00:00:42.560 --> 00:00:44.600 maybe more importantly, how can you use it? You know, 15 00:00:44.759 --> 00:00:47.320 harness its power to write better code, understand things like 16 00:00:47.399 --> 00:00:50.719 machine learning deep learning, even handle the ethical side of 17 00:00:50.759 --> 00:00:53.799 working with AI, because honestly, whether you're just starting out 18 00:00:53.880 --> 00:00:56.320 or you've been coding for years, when it comes to 19 00:00:56.359 --> 00:00:59.439 generative AI for coding, the source puts it perfectly, We're 20 00:00:59.439 --> 00:01:00.679 all dummies at this point. 21 00:01:00.799 --> 00:01:04.200 It really feels that way sometimes. So in this deep 22 00:01:04.239 --> 00:01:07.599 dive we'll lay out the basics machine learning, generative AI. 23 00:01:08.040 --> 00:01:11.040 We'll look at actually working with tools like Copilot replet 24 00:01:11.480 --> 00:01:14.959 and see how AI helps with automating the boring stuff 25 00:01:15.000 --> 00:01:18.719 but also improving and maintaining your code down the line. 26 00:01:18.840 --> 00:01:21.879 Okay, let's start unpacking that. Because we hear all these terms, right, AI, 27 00:01:22.120 --> 00:01:25.680 machine learning, deep learning, generative AI. They get thrown around 28 00:01:25.680 --> 00:01:29.079 almost interchangeably sometimes, but the source makes it clear they're 29 00:01:29.120 --> 00:01:33.519 related but distinct. How do you usually break that down? 30 00:01:33.920 --> 00:01:36.439 That's a really good place to start. Think of AI 31 00:01:36.760 --> 00:01:40.400 artificial intelligence as the big dream the whole field making 32 00:01:40.439 --> 00:01:45.480 computers smart, basically solving problems that usually need human intelligence. Now, 33 00:01:45.640 --> 00:01:48.400 machine learning mL, that's kind of our first big success 34 00:01:48.439 --> 00:01:51.840 story with NAI. It's where we don't hardcode every single rule. 35 00:01:52.000 --> 00:01:54.840 Instead we let the computers learn from massive amounts of data. 36 00:01:54.959 --> 00:01:57.239 Okay, so mL is teaching them to learn from data. 37 00:01:57.280 --> 00:01:59.599 What about deep learning DL? That one seems to be 38 00:01:59.599 --> 00:02:00.400 everywhere lately. 39 00:02:00.599 --> 00:02:03.480 Yeah. Deep learning is a specific, really powerful type of 40 00:02:03.519 --> 00:02:06.599 machine learning. This is where you get those artificial neural networks. 41 00:02:06.599 --> 00:02:09.800 They're structured with many layers, hence the deep It's sort 42 00:02:09.800 --> 00:02:12.439 of inspired by how our brains are wired with neurons 43 00:02:12.439 --> 00:02:16.240 and connections and these deep networks are incredibly good finding 44 00:02:16.360 --> 00:02:19.639 complex patterns and huge data sets like really. 45 00:02:19.520 --> 00:02:23.159 Huge DATCHA and Generative AI GENAI, the one making all 46 00:02:23.199 --> 00:02:25.240 the headlines that fits under. 47 00:02:25.159 --> 00:02:28.719 Deep learning exactly. GENAI systems are powered by these deep 48 00:02:28.840 --> 00:02:33.479 learning models, and their big trick is creating new stuff text, images, code, 49 00:02:33.560 --> 00:02:36.159 you name it, based on all that data they learned from. 50 00:02:36.199 --> 00:02:39.039 The quality has just exploded recently, mostly thanks to those 51 00:02:39.080 --> 00:02:40.199 deep learning breakthroughs. 52 00:02:40.400 --> 00:02:43.360 So how does that magic actually happen with language or code? 53 00:02:43.439 --> 00:02:45.319 I mean, how does the machine go from data to 54 00:02:46.120 --> 00:02:47.840 like writing a function? 55 00:02:48.199 --> 00:02:50.680 Well, it comes down to those neural networks. Again, our 56 00:02:50.719 --> 00:02:54.639 source explains they're trained on just enormous data sets. Think 57 00:02:54.680 --> 00:02:57.240 about recognizing a hot dog. You feed it millions of 58 00:02:57.280 --> 00:03:01.159 pictures labeled hot dog or not hot dog. It learns 59 00:03:01.199 --> 00:03:04.719 the patterns right, bun cylinder, shape, et cetera. For language, 60 00:03:04.759 --> 00:03:08.199 it's way more complex. They don't understand words like we do. 61 00:03:08.360 --> 00:03:12.439 Your input, like text or code, gets broken down into numbers. 62 00:03:12.520 --> 00:03:15.199 We call them tokens. Think of them like pixels for text. 63 00:03:15.520 --> 00:03:18.080 These tokens go through the network layers and the model 64 00:03:18.080 --> 00:03:20.719 predicts the next most likely token and the next building 65 00:03:20.800 --> 00:03:21.479 up a response. 66 00:03:22.080 --> 00:03:25.159 Okay, tokens make sense numbers instead of words, But how 67 00:03:25.199 --> 00:03:27.560 does it get context? How does it know bank means 68 00:03:27.560 --> 00:03:30.879 a riverbank here, but a financial bank there. Older model 69 00:03:30.879 --> 00:03:33.520 struggled with that, just going word by word exactly. 70 00:03:33.599 --> 00:03:36.639 That was a huge limitation. The real revolution, the thing 71 00:03:36.719 --> 00:03:40.080 that made Jennai take off for language was the transformer model. 72 00:03:40.400 --> 00:03:44.719 Transformers introduce this thing called a self attention mechanism. Instead 73 00:03:44.719 --> 00:03:47.599 of just processing things in order one after another, they 74 00:03:47.639 --> 00:03:50.840 can look at the entire input sequence, a whole sentence, 75 00:03:50.919 --> 00:03:53.479 a whole block of code, all at once and figure 76 00:03:53.479 --> 00:03:57.520 out how different parts relate to each other, which words 77 00:03:57.719 --> 00:03:59.639 or tokens are most important to each other. 78 00:03:59.719 --> 00:04:02.080 Wow, okay, so it's not just sequential anymore. It's like 79 00:04:02.120 --> 00:04:06.280 it grasps the whole context instantly. Yeah, that's a huge lead, 80 00:04:06.400 --> 00:04:08.439 and you can immediately see how that helps with the 81 00:04:09.280 --> 00:04:11.800 really repetitive, annoying coding tasks. 82 00:04:11.879 --> 00:04:14.599 Oh. Absolutely, take boilerplate code, you know, the stuff you 83 00:04:14.680 --> 00:04:17.040 have to write but nobody really enjoys. Like setting up 84 00:04:17.040 --> 00:04:18.839 a basic HTML page structure. 85 00:04:19.360 --> 00:04:22.600 Yes, necessary, but it's such a drag. You just want 86 00:04:22.639 --> 00:04:24.639 to get to the interesting part. So how does AI 87 00:04:24.759 --> 00:04:26.120 help there? What's the example. 88 00:04:26.319 --> 00:04:28.399 The source gives a great one. Chris Minnick just asks 89 00:04:28.480 --> 00:04:31.439 chat sheeppt for an HTML template. He specified like a 90 00:04:31.480 --> 00:04:36.040 three column flexbox layout top, navbar, footter, a pretty standard 91 00:04:36.079 --> 00:04:37.240 but non trivial setup. 92 00:04:37.519 --> 00:04:42.240 And did it like actually generate usable flexbox CSS? Yeah, 93 00:04:42.240 --> 00:04:44.879 because I've spent hours tweaking flexbox before it did. 94 00:04:44.720 --> 00:04:47.680 It gave him a full, ready to go HTML file 95 00:04:47.879 --> 00:04:50.600 CSS included that did exactly what he asked. It looked 96 00:04:50.639 --> 00:04:52.519 pretty much like what a human would write, but done 97 00:04:52.519 --> 00:04:56.120 in seconds. Compare that to finding snippets are starting from scratch. 98 00:04:56.600 --> 00:04:57.560 Huge time saver. 99 00:04:57.800 --> 00:05:00.439 That really is impressive for just a natural language. Quest 100 00:05:00.680 --> 00:05:06.120 makes you think about other common stuff. Yeah, like CRU operations, create, read, update, delete. 101 00:05:06.560 --> 00:05:08.199 Every app with data needs that right. 102 00:05:08.199 --> 00:05:11.000 Pretty much, and it's often tedious. The source shows an 103 00:05:11.000 --> 00:05:14.079 example of generating no JS code for CRU with Mango 104 00:05:14.199 --> 00:05:17.560 dB using Mongoose. The interesting part is the process. The 105 00:05:17.639 --> 00:05:21.040 first attempt, with a vague prompt to bing or chat GPT, 106 00:05:21.399 --> 00:05:24.439 gave back code that was okay ish, but it had 107 00:05:24.439 --> 00:05:27.000 this funny fly. It tried to use a simple DAVASCRIPT 108 00:05:27.040 --> 00:05:29.120 array as if it were a real database. 109 00:05:29.240 --> 00:05:32.959 Ah. Classic AI oversimplification. When it doesn't have enough context, 110 00:05:33.319 --> 00:05:35.120 so refining the prompt was key. 111 00:05:35.360 --> 00:05:38.720 Absolutely, This really drives home the importance of prompt engineering. 112 00:05:39.199 --> 00:05:43.839 When the prompt was updated to clearly say no JS, Mango, dB, Mongoose, 113 00:05:44.160 --> 00:05:47.720 the result was way better, impressively human like code actually 114 00:05:48.199 --> 00:05:52.680 latest dramascript, syntax, error handling, even comments. Then, of course 115 00:05:52.680 --> 00:05:54.600 you still have to do the actual setup and PM 116 00:05:54.639 --> 00:05:57.319 and it install Mondose, run the script, check the database, 117 00:05:57.560 --> 00:05:59.680 but the code generation part was solid. 118 00:06:00.000 --> 00:06:02.800 It shows that iteration, that back and forth is crucial. 119 00:06:02.959 --> 00:06:04.519 You guide the AI exactly. 120 00:06:04.600 --> 00:06:05.439 It's a partnership. 121 00:06:05.680 --> 00:06:09.360 So AI handled the gruntwork freeze up brain space. That's massive. 122 00:06:09.879 --> 00:06:13.079 But what about when you're in the code wrestling with syntax, 123 00:06:13.519 --> 00:06:16.959 trying to remember some specific API method. AI is an 124 00:06:17.000 --> 00:06:18.399 intelligent partner, right. 125 00:06:18.240 --> 00:06:20.839 Definitely, Just remembering syntax or looking it up is a 126 00:06:20.920 --> 00:06:23.839 huge chunk of programming time. We've had code completion for 127 00:06:23.920 --> 00:06:27.480 ages like Microsofts and Tellisons back in ninety six, useful, but. 128 00:06:27.639 --> 00:06:30.879 Limited, right, suggesting the next word or function name. Jenny 129 00:06:30.920 --> 00:06:32.240 I takes this way further. 130 00:06:32.279 --> 00:06:37.360 I assume, oh totally. Tools like getthub, Copilot, Amazon code Whisper. 131 00:06:38.040 --> 00:06:40.720 They don't just suggest the next word. They can suggest 132 00:06:41.000 --> 00:06:44.720 whole lines, entire functions, even blocks of code. They learn 133 00:06:44.759 --> 00:06:47.480 from massive code data sets, but also look at your 134 00:06:47.519 --> 00:06:51.160 current project, imported libraries, other open files for context. The 135 00:06:51.199 --> 00:06:54.639 source mentions copilot helping with the arguments for array dot reduce. 136 00:06:55.120 --> 00:06:56.800 That's a common one that trips people up. 137 00:06:56.839 --> 00:06:59.000 Oh yeah, reduce always takes me a second look. But 138 00:06:59.079 --> 00:07:02.040 what about learning news syntax? Can these tools actually teach 139 00:07:02.079 --> 00:07:03.160 you or just remind you? 140 00:07:03.240 --> 00:07:05.560 They can definitely help you learn. The chat interfaces are 141 00:07:05.600 --> 00:07:08.160 great for asking how do I do X in language? Why? 142 00:07:08.879 --> 00:07:10.839 But and this is a big butt from the source, 143 00:07:10.879 --> 00:07:12.839 you have to watch out for the training data cutoff 144 00:07:12.920 --> 00:07:16.279 dates Like GPT three doesn't know about JavaScript features added 145 00:07:16.319 --> 00:07:19.720 after September twenty twenty one. GBT four is much more current, 146 00:07:19.720 --> 00:07:21.040 but it's still a moving target. 147 00:07:21.199 --> 00:07:24.079 Always verify good point, so it's generating helping with syntax. 148 00:07:24.319 --> 00:07:27.120 Can it also check your code? Like linting? That seems 149 00:07:27.120 --> 00:07:27.839 like a natural fit. 150 00:07:28.120 --> 00:07:31.439 It is, so linting for anyone maybe not familiar, is 151 00:07:31.480 --> 00:07:35.439 like a superpowered spell checker and grammar checker for your code. 152 00:07:35.439 --> 00:07:40.160 Static analysis it finds potential errors, style issues, bugs without 153 00:07:40.160 --> 00:07:41.199 even running the code. 154 00:07:41.439 --> 00:07:45.000 And generative AI tools like Bard or whatever can basically 155 00:07:45.079 --> 00:07:47.759 act like a linter now pointing out typos or awkward 156 00:07:47.759 --> 00:07:48.720 style exactly. 157 00:07:48.839 --> 00:07:51.399 Or you have tools like eslin AI that integrate with 158 00:07:51.480 --> 00:07:54.399 existing linters. They don't just flag the issue, they can 159 00:07:54.439 --> 00:07:57.279 give you a natural language explanation of why it's considered 160 00:07:57.360 --> 00:07:58.279 bad practice. 161 00:07:58.399 --> 00:08:00.600 This all leads to a big question about to code, 162 00:08:00.639 --> 00:08:04.439 doesn't it? Is AI a crutch or is it like 163 00:08:04.519 --> 00:08:07.319 having a tutor. There's that whole debate in education. 164 00:08:07.759 --> 00:08:10.519 Yeah, the source leans towards AI becoming a normal and 165 00:08:10.639 --> 00:08:13.920 essential tool for learning, much like Google or textbooks are now, 166 00:08:14.319 --> 00:08:17.560 but it stresses, really stresses that there's no substitute for 167 00:08:17.600 --> 00:08:20.600 gaining experience through writing code or from interacting with more 168 00:08:20.600 --> 00:08:24.639 experienced programmers. AI can assist, it can explain, but it 169 00:08:24.680 --> 00:08:27.319 doesn't replace actually doing it and learning from humans. 170 00:08:27.639 --> 00:08:30.920 That makes sense, which brings us to AI pair programming. 171 00:08:31.720 --> 00:08:35.240 Sounds futuristic, but it's happening. The model is you're the 172 00:08:35.360 --> 00:08:39.519 navigator doing the high level thinking, setting direction with comments 173 00:08:39.600 --> 00:08:43.679 and code, and the AI is the driver or assistant, 174 00:08:43.919 --> 00:08:47.240 suggesting code snippets as you go. What are the pros 175 00:08:47.279 --> 00:08:48.000 and cons there? 176 00:08:48.440 --> 00:08:51.320 Well, the pros are pretty clear less time hunting for syntax, 177 00:08:51.720 --> 00:08:55.679 less typing, boilerplate. Your AI partner is always there, super 178 00:08:55.720 --> 00:08:57.960 fast and knows tons of different ways to code things 179 00:08:57.960 --> 00:09:00.840 from its training. No waiting for a human partner. 180 00:09:01.240 --> 00:09:03.679 But the cons I mean a human pair programmer does 181 00:09:03.720 --> 00:09:06.639 more than just type, right. They challenge ideas, spot bigger 182 00:09:06.759 --> 00:09:07.679 architectural issues. 183 00:09:07.840 --> 00:09:11.399 Exactly, you lose that deeper knowledge sharing and critical feedback 184 00:09:11.440 --> 00:09:14.480 loop you get with a human. Plus the AI suggestions 185 00:09:14.559 --> 00:09:18.320 might be subtly wrong, outdated, or even introduce security vulnerability 186 00:09:18.399 --> 00:09:21.519 is a human would catch. And here's the key takeaway 187 00:09:21.519 --> 00:09:24.519 from the source on this AI pair programming works best 188 00:09:24.519 --> 00:09:27.080 for coders who know their language and have experienced writing 189 00:09:27.080 --> 00:09:29.919 code without the use of AI. The AI, as the 190 00:09:29.960 --> 00:09:33.279 source puts, it, speaks confidently but doesn't know anything about programming. 191 00:09:33.440 --> 00:09:35.600 It's pattern matching, not true understanding. 192 00:09:35.919 --> 00:09:39.039 Right, It's a tool for someone who understands, not a replacement. 193 00:09:39.480 --> 00:09:42.399 The source walks through building a trivia game with Copilot 194 00:09:42.559 --> 00:09:43.559 to show this, right. 195 00:09:43.679 --> 00:09:46.799 Yeah, it's a good practical example. Start with index dot html, 196 00:09:47.120 --> 00:09:50.759 write a detailed comment explaining the game. Copilot generates some 197 00:09:50.840 --> 00:09:54.240 starter HTML. Then you refine it. Tell it you want 198 00:09:54.320 --> 00:09:58.440 JavaScript data from a specific file datatrivia dot js, tell 199 00:09:58.440 --> 00:10:02.200 it to use radio buttons. Rance Copilot generates the trivia 200 00:10:02.279 --> 00:10:06.039 Questions dot js data structure. Then, because it sees both 201 00:10:06.039 --> 00:10:09.080 index dot html and Trivia Questions dot js are open, 202 00:10:09.440 --> 00:10:12.360 it uses that context to help write the functions in 203 00:10:12.440 --> 00:10:14.360 the main trivia dot js file. 204 00:10:14.440 --> 00:10:17.000 So it's that iterative cycle again. You write some code, 205 00:10:17.000 --> 00:10:20.480 write a comment, accept or modify a suggestion, test it out. 206 00:10:20.399 --> 00:10:22.919 Exactly, and after maybe twenty minutes or so on the 207 00:10:22.919 --> 00:10:26.600 second try, you end up with a somewhat functional trivia game. 208 00:10:26.679 --> 00:10:29.320 Not perfect, but a solid start built collaboratively. 209 00:10:29.440 --> 00:10:31.639 That's pretty cool. And for people who prefer coding right 210 00:10:31.679 --> 00:10:33.159 in the browser, there's Replet. 211 00:10:33.399 --> 00:10:36.759 Yeah. Replet's an interesting platform. It's an online ID, so 212 00:10:36.840 --> 00:10:40.279 no local setup needed. It's got collaboration built in deployment, 213 00:10:40.519 --> 00:10:45.159 and its own suite of AI tools things like generate, debug, explain, 214 00:10:45.360 --> 00:10:48.000 modify code, plus the copilot. 215 00:10:47.559 --> 00:10:50.639 Style completion, and the source shows using it to spin 216 00:10:50.720 --> 00:10:53.000 up a quick website, even for a fake band called 217 00:10:53.039 --> 00:10:53.759 Grapefruit Pulp. 218 00:10:53.879 --> 00:10:55.679 Right. It just shows how quickly you can get a 219 00:10:55.720 --> 00:10:59.679 project moving using these integrated AI features in an online environment. 220 00:11:00.159 --> 00:11:03.799 So we've seen AI generate code, help and syntax act 221 00:11:03.840 --> 00:11:06.879 as a pair programmer. But what about improving the code 222 00:11:06.919 --> 00:11:10.039 we already have refactoring, testing, maintenance. 223 00:11:10.240 --> 00:11:13.399 Yeah, that's another huge area where AI is making inroads. 224 00:11:13.679 --> 00:11:16.720 Let's talk refactoring first. That's improving the code's internal structure 225 00:11:16.720 --> 00:11:19.360 and making it cleaner, more efficient, easier to understand without 226 00:11:19.480 --> 00:11:20.759 changing what it actually does. 227 00:11:21.000 --> 00:11:24.600 And the source uses that great term code smells like 228 00:11:24.639 --> 00:11:28.480 something's a bit off, indicating a deeper potential problem. Right, 229 00:11:28.720 --> 00:11:30.000 good lead to code rot if you. 230 00:11:29.960 --> 00:11:33.559 Ignore it exactly. It's a perfect analogy. And these smells 231 00:11:33.559 --> 00:11:37.240 fall into types dispensables like duplicate code or comments that 232 00:11:37.360 --> 00:11:41.600 just state the obvious. Bloaters like massive functions or classes 233 00:11:41.679 --> 00:11:46.360 doing too much, abusers, maybe using language features incorrectly, couplers 234 00:11:46.399 --> 00:11:48.559 where different parts of the code are way too dependent 235 00:11:48.639 --> 00:11:51.600 on each other, and change preventors things that make future 236 00:11:51.679 --> 00:11:52.799 updates really hard. 237 00:11:53.039 --> 00:11:57.720 So can AI like copilot chat actually detect these smells? 238 00:11:57.840 --> 00:11:59.679 Can it sniff out a bloater function? 239 00:12:00.039 --> 00:12:02.679 I can try? In the tiktac toe example from the source, 240 00:12:02.720 --> 00:12:06.559 copilot Chat did identify several smells a function that was 241 00:12:06.600 --> 00:12:09.240 too long magic numbers just unexplained numbers in the code, 242 00:12:09.519 --> 00:12:13.200 also inconsistent naming, not enough comments, and using global data 243 00:12:13.240 --> 00:12:14.840 when it could have been better encapsulated. 244 00:12:14.919 --> 00:12:17.200 Okay, so it found the smells, did it suggest how 245 00:12:17.240 --> 00:12:19.559 to fix them, like actual refactoring steps. 246 00:12:19.879 --> 00:12:23.159 It did make suggestions, for instance, moving event handlers out 247 00:12:23.159 --> 00:12:26.200 of the HTML and into the JavaScript, or wrapping the 248 00:12:26.240 --> 00:12:29.919 game state in a class to avoid global variables. However, 249 00:12:30.080 --> 00:12:32.759 the source also cautions that trying to get Copilot to 250 00:12:32.799 --> 00:12:37.240 automatically fix the code sometimes introduce new errors, and for 251 00:12:37.360 --> 00:12:40.320 more complex things like improving the tic tac toe AI 252 00:12:40.320 --> 00:12:44.000 player itself, the suggestions weren't always great, right, So. 253 00:12:44.039 --> 00:12:47.759 Again, it's a helpful assistant for identifying issues and suggesting approaches, 254 00:12:47.919 --> 00:12:51.159 but you absolutely need to review and test its suggestions carefully. 255 00:12:51.679 --> 00:12:53.360 Don't accept a suggestion blindly. 256 00:12:53.679 --> 00:12:57.519 That's the constant refrain and it's crucial. Now, let's shift testing. 257 00:12:57.799 --> 00:13:00.480 The source gives that fun anecdote about Admiral Grace Hopper 258 00:13:00.519 --> 00:13:03.600 and the first literal computer bug A moth. Testing is 259 00:13:03.679 --> 00:13:04.600 just fundamental. 260 00:13:04.759 --> 00:13:06.639 Absolutely, Yeah, I got to make sure the code does 261 00:13:06.639 --> 00:13:08.480 what you think it does. How does AI help with 262 00:13:08.519 --> 00:13:11.720 the testing process itself? Planning writing tests. 263 00:13:11.440 --> 00:13:15.200 Both actually for planning. AI can analyze requirements documents to 264 00:13:15.240 --> 00:13:18.279 suggest test cases or scenarios. It can help identify which 265 00:13:18.320 --> 00:13:20.279 parts of the code are riskiest and need more testing. 266 00:13:20.600 --> 00:13:23.759 It can even simulate user journeys to find potential problems. 267 00:13:23.960 --> 00:13:27.840 That's pretty advanced and for writing the actual tests like 268 00:13:28.000 --> 00:13:31.639 unit tests. The source mentions using jest with copilot chat. 269 00:13:31.879 --> 00:13:34.279 Yeah, once you have just set up, you can literally 270 00:13:34.360 --> 00:13:38.240 prompt Copilot like write just test for the checkwin function 271 00:13:38.320 --> 00:13:41.399 in this file, and it will generate test cases trying 272 00:13:41.399 --> 00:13:43.879 to cover different scenarios. And you can even ask for 273 00:13:43.919 --> 00:13:44.639 coverage reports. 274 00:13:44.720 --> 00:13:47.559 But there's a catch, right, I've heard AI can sometimes 275 00:13:47.639 --> 00:13:50.000 write tests that pass even if the code is wrong. 276 00:13:50.279 --> 00:13:54.120 Yes, that's a massive pitfall, the source highlights. Just because 277 00:13:54.159 --> 00:13:57.000 all the AI generated tests passed does not mean the 278 00:13:57.039 --> 00:13:59.679 code is correct. There's an example with a simple sum 279 00:13:59.759 --> 00:14:03.559 number function that had a bug. Copilot initially wrote tests 280 00:14:03.559 --> 00:14:05.840 that confirmed the buggy behavior. You have to be really 281 00:14:05.879 --> 00:14:08.440 specific in your prompts, and more importantly, you need to 282 00:14:08.519 --> 00:14:10.279 understand the code well enough to know if the tests 283 00:14:10.279 --> 00:14:13.639 are actually validating the intended logic, not just whatever logic 284 00:14:13.639 --> 00:14:16.399 happens to be there. Human oversight is non negotiable. 285 00:14:16.480 --> 00:14:21.039 Got it tricky? What about TDD test driven development? Writing 286 00:14:21.039 --> 00:14:24.480 the failing test first, then the code can AI fit 287 00:14:24.519 --> 00:14:25.360 into that workflow? 288 00:14:25.639 --> 00:14:28.000 It can help you write the failing test. Then you 289 00:14:28.000 --> 00:14:30.120 can ask copilot to help write the code to make 290 00:14:30.159 --> 00:14:33.519 that specific test pass. But the source author found that 291 00:14:33.600 --> 00:14:37.399 sometimes in the TIC TAC TOE TDD process, copilot got 292 00:14:37.399 --> 00:14:41.000 weirdly stubborn and refused to answer programming questions, forcing them 293 00:14:41.000 --> 00:14:43.519 to write the con manually. It seems like it might 294 00:14:43.519 --> 00:14:46.960 struggle with that specific constrained TDD cycle. 295 00:14:46.679 --> 00:14:50.240 Sometimes interesting, So it's not quite seamless there yet. Okay. 296 00:14:50.360 --> 00:14:55.399 Beyond writing and testing, documentation and maintenance huge parts of 297 00:14:55.440 --> 00:14:56.360 the software life cycle. 298 00:14:56.480 --> 00:15:00.559 Absolutely. Documentation breaks down into internal for other devs like 299 00:15:00.600 --> 00:15:03.679 code comments, design docs, and external for users like read 300 00:15:03.679 --> 00:15:05.919 deems tutorials, API docs. 301 00:15:05.759 --> 00:15:08.879 Right and for comments. There are tools like mintlified doc 302 00:15:08.919 --> 00:15:11.960 writer in vs code that use AI to generate them. 303 00:15:12.000 --> 00:15:12.679 How good are they? 304 00:15:13.000 --> 00:15:15.519 They can be useful for generating standard stuff like jas 305 00:15:15.600 --> 00:15:18.759 dot comments explaining parameters and return types, but the source 306 00:15:18.759 --> 00:15:21.799 compares AI comments to human written ones for some underscore 307 00:15:21.840 --> 00:15:25.679 dot JS functions. The verdict AI can be detailed, but 308 00:15:25.799 --> 00:15:28.960 sometimes overly wordy, and it often misses the nuance. The 309 00:15:29.000 --> 00:15:31.960 why behind the code that a human developer would include 310 00:15:32.240 --> 00:15:32.639 makes sense. 311 00:15:32.639 --> 00:15:34.919 It explains the what, but maybe not the why. What 312 00:15:34.960 --> 00:15:37.840 about visual docs diagrams? 313 00:15:37.960 --> 00:15:41.080 Yeah, tools like draw dot io can now use AI 314 00:15:41.200 --> 00:15:44.320 to generate diagrams from text prompts. You could ask it 315 00:15:44.399 --> 00:15:47.360 to make a sequence diagram for making an apple pie, 316 00:15:47.879 --> 00:15:51.039 or describe your software components and it'll try to visualize it. 317 00:15:51.200 --> 00:15:53.360 Pretty handy for quick diagrams. 318 00:15:52.879 --> 00:15:58.679 And API documentation that's often the pain point. Using OpenAPI Specification. 319 00:15:58.279 --> 00:16:02.799 OAS, huge timesaver Copilot can help generate that OAS file, 320 00:16:02.960 --> 00:16:07.120 usually in eamal format, describing your API endpoints. Then tools 321 00:16:07.159 --> 00:16:10.720 like read me can ingest that file and automatically create interactive, 322 00:16:10.879 --> 00:16:13.480 user friendly API documentation websites. 323 00:16:13.600 --> 00:16:16.159 That's slick and the source mentions. Create your own chatbot 324 00:16:16.159 --> 00:16:18.759 for your API docs using chat GPT. 325 00:16:18.360 --> 00:16:21.399 Plus, Yeah, you can create a custom GPT upload your 326 00:16:21.399 --> 00:16:23.840 OAS file as its knowledge base. Then you or your 327 00:16:23.919 --> 00:16:26.159 users can ask it questions like how do I authenticate? 328 00:16:26.399 --> 00:16:28.440 Or show me how to create a user with Python, 329 00:16:28.600 --> 00:16:31.720 and it'll answer based on your specific API docs. Still 330 00:16:31.799 --> 00:16:34.000 might make mistakes, but very cool. 331 00:16:33.799 --> 00:16:38.080 Potential, very cool, indeed okay. Final, big area software maintenance 332 00:16:38.159 --> 00:16:40.279 keeping the code alive and healthy. 333 00:16:40.240 --> 00:16:43.360 Right and the source breaks maintenance into four types. Corrective 334 00:16:43.559 --> 00:16:47.600 fixing bugs, adaptive, changing code because the environment changed, like 335 00:16:47.639 --> 00:16:51.399 a new OS version, perfective adding new features or removing 336 00:16:51.440 --> 00:16:56.039 old ones. And preventative making changes now to prevent problems later, 337 00:16:56.240 --> 00:16:58.200 like refactoring or improving security. 338 00:16:58.480 --> 00:17:01.039 So how does AI help across those four types? 339 00:17:01.080 --> 00:17:05.039 In lots of ways. Automated error detection and reporting, Predictive 340 00:17:05.079 --> 00:17:08.519 maintenance like analyzing logs to guess where failures might happen, 341 00:17:08.759 --> 00:17:12.640 helping with refactoring as we discussed, managing dependencies, checking for 342 00:17:12.680 --> 00:17:17.119 outdated libraries, analyzing user feedback from bug reports or feature requests, 343 00:17:17.240 --> 00:17:21.400 and a big one security Vulnerability scanning tools like Snick 344 00:17:21.519 --> 00:17:24.759 can scan your code and dependencies, find known vulnerabilities, and 345 00:17:24.839 --> 00:17:28.160 sometimes even automatically create pull requests with the fixes. 346 00:17:27.880 --> 00:17:30.920 Wow automatically opening pole requests. That's proactive as all ties 347 00:17:30.920 --> 00:17:32.720 into code quality and managing. 348 00:17:32.440 --> 00:17:36.160 Technical debt right exactly. Code quality isn't just does it 349 00:17:36.240 --> 00:17:40.079 work functional quality, but also is it well built? Easy 350 00:17:40.160 --> 00:17:44.440 to change? Structural quality. Technical debt is that idea that 351 00:17:44.480 --> 00:17:47.839 taking shortcuts now means you'll pay more later in terms 352 00:17:47.880 --> 00:17:50.240 of time and effort to fix or refactor things. 353 00:17:50.319 --> 00:17:53.240 And tools like code Climate automate checking for that. Yeah, 354 00:17:53.240 --> 00:17:54.680 giving your code a grade. 355 00:17:54.480 --> 00:17:57.519 Yep, you connected to your repo. It analyzes the code 356 00:17:57.519 --> 00:18:01.559 for maintainability issues, complexity, duplicate. It gives you a grade 357 00:18:01.759 --> 00:18:04.240 AF and estimates how long it would take to fix 358 00:18:04.279 --> 00:18:05.799 the identified technical debt. 359 00:18:06.160 --> 00:18:09.359 The source mentions actually beating code Climate's estimate for fixing 360 00:18:09.400 --> 00:18:11.680 things in the signapage dot js file. 361 00:18:11.640 --> 00:18:14.960 Right, which shows human intuition and focus can still sometimes 362 00:18:15.000 --> 00:18:18.519 outperform the automated estimate, but it also highlights that fixing 363 00:18:18.559 --> 00:18:22.200 one thing often uncovered other bugs. It's a continuous cycle, 364 00:18:22.599 --> 00:18:25.400 achieving that a gray doesn't mean you're done forever. Maintenance 365 00:18:25.440 --> 00:18:26.079 is ongoing. 366 00:18:26.200 --> 00:18:28.720 Okay, that's a lot we've covered. Bringing it all together, 367 00:18:29.000 --> 00:18:32.480 what are the really crucial takeaways for you listening right now? First, 368 00:18:32.599 --> 00:18:35.759 maybe obviously, don't stop coding yourself. You need to lead 369 00:18:36.440 --> 00:18:38.880 use AI as a tool, but rely on your own 370 00:18:38.960 --> 00:18:40.440 skills first and foremost. 371 00:18:40.720 --> 00:18:43.799 Absolutely, and when the AI gives you something weird or unexpected, 372 00:18:44.079 --> 00:18:47.799 ask follow up questions, probe it, don't just accept confusing output, 373 00:18:48.119 --> 00:18:51.880 and always double check against official documentation. AI models can 374 00:18:51.880 --> 00:18:53.519 be out of date or just make things up. They 375 00:18:53.559 --> 00:18:54.400 call it hallucination. 376 00:18:54.720 --> 00:18:58.279 Verify facts, good point and be specific with your prompts. 377 00:18:58.559 --> 00:19:01.559 Give clear context, providing examples if you can. The better 378 00:19:01.640 --> 00:19:04.880 your input, the better the AI's output. Garbage in garbage 379 00:19:04.880 --> 00:19:09.440 out still applies, and security. Be careful anonymized sensitive data 380 00:19:09.480 --> 00:19:12.359 before pasting it into a prompt. Don't feed the AI 381 00:19:12.480 --> 00:19:14.319 your company's secret sauce or user. 382 00:19:14.119 --> 00:19:18.480 PII critically important. Also, keep your tools updated. This whole 383 00:19:18.480 --> 00:19:22.519 field is moving incredibly fast, new models, new features, improved accuracy, 384 00:19:22.799 --> 00:19:26.200 Stay current and remember the limits. AI isn't magic. It 385 00:19:26.279 --> 00:19:29.039 learns from existing data. If you're doing something truly novel, 386 00:19:29.079 --> 00:19:31.680 something way outside its training data, it probably won't be 387 00:19:31.759 --> 00:19:32.160 much help. 388 00:19:32.440 --> 00:19:36.480 Right. And finally, that classic advice from Donald Knuth maybe 389 00:19:36.519 --> 00:19:40.079 more relevant than ever with AI tempting us to tweak endlessly. 390 00:19:40.759 --> 00:19:46.319 Avoid premature optimization. Spending hours to shave off a few microseconds, 391 00:19:46.359 --> 00:19:50.839 potentially introducing bugs usually not worth it. Get it working 392 00:19:50.880 --> 00:19:53.759 correctly and clearly first, then optimize if needed. 393 00:19:53.920 --> 00:19:56.519 So true, we've really peeled back the layers here, haven't we? 394 00:19:56.680 --> 00:20:00.599 From AI automating the boring bits to being a code partner, 395 00:20:01.000 --> 00:20:04.559 to helping refactor, test, and maintain software. We've seen the 396 00:20:04.599 --> 00:20:08.599 amazing potential, but also hammered home that need for human oversight, 397 00:20:09.160 --> 00:20:11.880 critical thinking, and just continuous learning. Yeah. 398 00:20:11.960 --> 00:20:14.240 Chris Minnick, the author of our main source notes. He 399 00:20:14.279 --> 00:20:16.720 wrote the book about a year after chat GPT burst 400 00:20:16.759 --> 00:20:19.000 onto the scene. He says, these tools have begun to 401 00:20:19.079 --> 00:20:20.960