1 00:00:00,480 --> 00:00:02,320 Oh, it started. Yeah, yeah. 2 00:00:02,360 --> 00:00:04,160 It's it's better to start too early than too late. 3 00:00:06,280 --> 00:00:11,400 Ask me how I learned that. Welcome to the Embedded AI 4 00:00:11,400 --> 00:00:16,160 Podcast, a podcast about embedded systems engineering and 5 00:00:16,320 --> 00:00:19,680 it's overlap with AI. I'm one of your real life human 6 00:00:19,680 --> 00:00:22,920 hosts, Ryan Torvik. And I'm looking down. 7 00:00:23,320 --> 00:00:28,760 And today we are joined by a young man who's also very big 8 00:00:28,760 --> 00:00:33,840 into embedded system design and AI. 9 00:00:34,160 --> 00:00:36,520 We are here today with Jacob Beningo. 10 00:00:38,600 --> 00:00:42,160 How's it going, everybody? And I am the the biological 11 00:00:42,160 --> 00:00:43,560 intelligence known as Jacob Beningo is. 12 00:00:45,560 --> 00:00:47,840 That what we're doing. We got AI and BI. 13 00:00:48,400 --> 00:00:51,080 Gosh, fuck you. That's so good. 14 00:00:51,480 --> 00:00:54,960 All right, So today, Jake, we brought Jacob on because Luca 15 00:00:54,960 --> 00:00:58,600 and Jacob and I, the two of them more so have been trying to 16 00:00:58,600 --> 00:01:02,080 teach other people and interact with other people about, about 17 00:01:02,080 --> 00:01:05,760 using AI for code generation and in their systems and their 18 00:01:05,760 --> 00:01:08,800 development work flows. And we've got a lot of things 19 00:01:08,800 --> 00:01:11,280 that we've that we've learned and picked up over the time. 20 00:01:11,280 --> 00:01:14,120 And it hasn't been even that long, but we'll, we want to kind 21 00:01:14,120 --> 00:01:16,680 of talk through like, what are you seeing? 22 00:01:16,680 --> 00:01:18,400 What are you getting? What, what are you hearing? 23 00:01:18,400 --> 00:01:20,160 What are the kind of mistakes that people are making? 24 00:01:20,160 --> 00:01:22,280 All that kind of stuff. How are people feeling about it? 25 00:01:22,280 --> 00:01:24,840 How do you feel about it? So that's what we're going to 26 00:01:24,840 --> 00:01:25,840 do. We're It's an engineering 27 00:01:25,840 --> 00:01:31,040 podcast to talk about feelings. But you know, you know what? 28 00:01:31,440 --> 00:01:33,680 You know what? Because we just before we start 29 00:01:33,680 --> 00:01:37,360 recording, we talked about how podcasts forgot to do things and 30 00:01:37,360 --> 00:01:40,080 then tacked them on at the end, even though they should have 31 00:01:40,080 --> 00:01:43,160 gone in the beginning. Jacob, who are you even? 32 00:01:44,480 --> 00:01:46,400 Oh, that's man, that's a good question. 33 00:01:48,160 --> 00:01:51,400 So, yeah, so I'm a real time embedded systems engineer that 34 00:01:51,400 --> 00:01:54,400 focuses a lot of times on mission critical systems. 35 00:01:54,480 --> 00:01:57,800 And so in that we have systems that obviously need to be very 36 00:01:57,800 --> 00:01:59,680 robust. So space systems, defense 37 00:01:59,680 --> 00:02:01,480 systems, medical devices, things like that. 38 00:02:01,480 --> 00:02:04,240 And so often times what I'm doing is I'm training or 39 00:02:04,240 --> 00:02:07,360 consulting with folks to build these types of systems, improve 40 00:02:07,360 --> 00:02:10,199 their processes, often modernizing their embedded 41 00:02:10,199 --> 00:02:12,800 systems, which can cover a lot of things today from build 42 00:02:12,800 --> 00:02:15,040 systems to dev OPS software architectures. 43 00:02:15,280 --> 00:02:18,120 But obviously AI is a very important part of that 44 00:02:18,120 --> 00:02:21,960 modernization process as well. And so often times I'm out there 45 00:02:21,960 --> 00:02:24,360 sharing my, sharing my knowledge of embedded systems development. 46 00:02:24,360 --> 00:02:26,240 I've been doing this for 20 years. 47 00:02:26,240 --> 00:02:28,880 I work as a consultant, which I, I've been doing for like 15 48 00:02:28,880 --> 00:02:31,600 years now, but real time bed systems, that microcontroller 49 00:02:31,600 --> 00:02:33,600 based stuff. That's really my passion. 50 00:02:33,800 --> 00:02:35,760 And of course, like I said, teaching and sharing that 51 00:02:35,760 --> 00:02:37,040 knowledge as well with the industry. 52 00:02:37,040 --> 00:02:40,160 So, and hopefully making it so that other engineers don't make 53 00:02:40,160 --> 00:02:41,280 the same dumb mistakes that I make. 54 00:02:41,280 --> 00:02:43,680 Oh. My gosh, I'm sitting here 55 00:02:43,680 --> 00:02:45,600 listening and I'm like, wow, that's, that's really good. 56 00:02:45,600 --> 00:02:47,560 Luca, if you asked me to do that, I'd have to pull up our 57 00:02:47,560 --> 00:02:49,600 website and like, what? What do I do? 58 00:02:49,600 --> 00:02:55,480 Hold on, let me go look that up. Oh, my goodness. 59 00:02:57,000 --> 00:02:58,480 All right. So where, where, where are we 60 00:02:58,480 --> 00:03:00,000 starting? So Luca, you've been teaching, 61 00:03:00,000 --> 00:03:01,960 uh, you've had a couple of classes where you've taught some 62 00:03:01,960 --> 00:03:03,320 like cloud code uses kind of stuff. 63 00:03:03,320 --> 00:03:05,560 Do you want to start there? Yeah, why not? 64 00:03:05,720 --> 00:03:10,120 So as we were talking before the episode, Jacob and I noticed 65 00:03:10,120 --> 00:03:13,760 that, you know, we've, we've I think saw, we've seen a couple 66 00:03:13,760 --> 00:03:19,240 of patterns emerge as we were trying to teach people to use AI 67 00:03:19,240 --> 00:03:23,240 in software engineering. And in a sense, actually the 68 00:03:23,240 --> 00:03:26,800 introduction, I think I've spotted the first one, which is 69 00:03:27,200 --> 00:03:30,480 that you said that I taught people to, to write code using 70 00:03:30,960 --> 00:03:33,080 cloud code or or, you know, whatever AI. 71 00:03:33,640 --> 00:03:39,160 And while that's not technically wrong because we ended up, you 72 00:03:39,160 --> 00:03:42,760 know, creating a bunch of code, that's kind of past besides the 73 00:03:42,760 --> 00:03:45,520 point. I think what people really 74 00:03:45,520 --> 00:03:50,160 struggle with is not the have the a create curly brackets for 75 00:03:50,160 --> 00:03:55,160 me, but actual software engineering using AI, right? 76 00:03:55,360 --> 00:03:58,280 All of the stuff that goes before and after the curly 77 00:03:58,280 --> 00:04:02,400 brackets, that that is really the hard part and what really 78 00:04:02,760 --> 00:04:05,560 makes or breaks any kind of system, including an embedded 79 00:04:05,560 --> 00:04:09,080 system. I, I feel like that gets 80 00:04:09,160 --> 00:04:11,880 overlooked very easily. Is that also something that 81 00:04:11,880 --> 00:04:14,880 you've spotted, Jacob? Oh, absolutely, yeah. 82 00:04:14,880 --> 00:04:17,000 I mean, a lot of times people think you just open up your code 83 00:04:17,000 --> 00:04:18,560 editor and you start banging out code, right? 84 00:04:18,839 --> 00:04:22,680 They forget about the processes that go upfront, soliciting 85 00:04:22,680 --> 00:04:25,440 requirements from, you know, the end user, talking to 86 00:04:25,440 --> 00:04:29,800 stakeholders, writing user stories, developing, you know, 87 00:04:29,840 --> 00:04:32,160 heaven forbid someone figure out, you know, a software 88 00:04:32,160 --> 00:04:34,520 architecture ahead of time that says, Hey, here's the main 89 00:04:34,520 --> 00:04:37,080 pieces of, you know, the, the system I'm actually building, 90 00:04:37,080 --> 00:04:39,080 right? You know, people love to just be 91 00:04:39,080 --> 00:04:41,360 like, the architecture is emergent, right? 92 00:04:41,360 --> 00:04:43,960 Like just kind of come out of the coding, right? 93 00:04:44,360 --> 00:04:47,240 And, you know, don't get me wrong, in a small project, sure, 94 00:04:47,320 --> 00:04:49,720 right. But you know, if you're you're 95 00:04:49,720 --> 00:04:52,520 building, you know, a space system or medical device, you 96 00:04:52,520 --> 00:04:54,880 really want to understand those major blocks, right? 97 00:04:54,880 --> 00:04:57,720 So, so, you know, there's all that type of stuff that comes 98 00:04:57,720 --> 00:04:59,680 first before we should be opening our editor. 99 00:05:00,440 --> 00:05:03,240 So you're saying even though I'm I'm I'm having the AI generate 100 00:05:03,240 --> 00:05:06,040 the code for me, I still have to do the systems engineering that 101 00:05:06,040 --> 00:05:09,280 I don't like doing? Absolutely, absolutely. 102 00:05:09,480 --> 00:05:12,560 So no, I, I, I might be biased because I, I, I am, my 103 00:05:12,560 --> 00:05:13,880 background is in systems engineering. 104 00:05:14,400 --> 00:05:19,440 So maybe that's just, you know, my, my own personal bias, but I, 105 00:05:19,800 --> 00:05:22,560 there is actually a lot of research, you know, showing that 106 00:05:22,560 --> 00:05:25,880 when you actually follow processes, the work actually 107 00:05:25,880 --> 00:05:27,720 goes faster and is less costly, so. 108 00:05:28,280 --> 00:05:31,200 Certainly doesn't feel that way. I I feel faster when I'm mashing 109 00:05:31,200 --> 00:05:33,680 my fingers on the keys and tearing about where my my 110 00:05:33,680 --> 00:05:36,400 semicolons are going and which if statement I'm using. 111 00:05:36,400 --> 00:05:38,240 And if I'm using a for loop or a while loop. 112 00:05:38,240 --> 00:05:40,320 And is the variable inside the loop or outside the loop? 113 00:05:40,320 --> 00:05:42,840 Like, yeah. I mean, you're being sarcastic, 114 00:05:42,840 --> 00:05:46,360 but you're not wrong like this is. 115 00:05:46,600 --> 00:05:49,040 This is the problem with many people that they feel like 116 00:05:49,040 --> 00:05:51,200 they're not doing anything or certainly not anything 117 00:05:51,200 --> 00:05:54,480 worthwhile if they are not actively writing code. 118 00:05:55,000 --> 00:05:59,920 Even though one of my favorite chapters in the out of embedded, 119 00:06:00,240 --> 00:06:04,160 out of Unix programming is don't just do something, stand there. 120 00:06:04,240 --> 00:06:10,520 Like there really is like, I mean, when you think about a 121 00:06:10,520 --> 00:06:12,360 Carpenter, what are you, what are they told to do? 122 00:06:12,560 --> 00:06:14,480 You know, measure twice, cut once, right. 123 00:06:14,600 --> 00:06:17,560 And with with software, we tend to just say, well, it's cheap, 124 00:06:17,560 --> 00:06:20,640 so let's just keep cutting until maybe eventually we get what we 125 00:06:20,640 --> 00:06:23,400 want, right? Except all those cuttings, all 126 00:06:23,400 --> 00:06:25,320 that cutting takes time, cost money. 127 00:06:25,400 --> 00:06:27,880 You know, if you deploy something in the field that you 128 00:06:27,880 --> 00:06:30,560 know, heaven forbid has, you know, causes an injury or 129 00:06:30,560 --> 00:06:32,720 something, I mean, that's, that's not good, right? 130 00:06:32,720 --> 00:06:37,040 And so AII think can have a a really great impact on helping 131 00:06:37,040 --> 00:06:39,040 us actually follow good systems engineering practices. 132 00:06:39,040 --> 00:06:41,840 And that's, that's what I, I, I've been coming to to 133 00:06:41,840 --> 00:06:45,080 understand too, is like, no AI actually forces you. 134 00:06:45,080 --> 00:06:48,360 Or if, if you're, if you're really leaning into it, it will 135 00:06:48,360 --> 00:06:51,520 make you a better overall engineer because it takes away 136 00:06:51,520 --> 00:06:54,040 the power to write the actual software. 137 00:06:54,040 --> 00:06:57,040 And it takes away that serotonin or that dopamine hit that I get 138 00:06:57,040 --> 00:07:00,920 from my, my compilation working because I typed the right code 139 00:07:01,240 --> 00:07:03,480 and it, and it has me do something different. 140 00:07:03,480 --> 00:07:06,520 I it's a different workflow and a different process, but it's a 141 00:07:06,520 --> 00:07:08,960 completely different way of thinking about, you know, gener 142 00:07:08,960 --> 00:07:12,160 of getting the code to to to compile and to run. 143 00:07:12,160 --> 00:07:15,040 So. And, and honestly, you know, you 144 00:07:15,040 --> 00:07:17,480 kind of mentioned the dopamine hit that you get from compiling. 145 00:07:17,480 --> 00:07:21,000 I actually think AI has the AI prompting has the same effect. 146 00:07:21,280 --> 00:07:23,800 A lot of engineers I've talked to have mentioned how they're 147 00:07:23,800 --> 00:07:27,200 addicted to using their AI. Like they get up, they start 148 00:07:27,200 --> 00:07:29,240 prompting, they, you know, they go to work, they start 149 00:07:29,240 --> 00:07:30,840 prompting, you know, they come home, they start prompting. 150 00:07:31,720 --> 00:07:37,200 I might be in that category, but there's these tools, right You 151 00:07:37,360 --> 00:07:39,920 that you use that you still get a little bit of AI guess a 152 00:07:39,920 --> 00:07:42,720 dopamine hit of, hey, I'm making progress, forward progress. 153 00:07:43,120 --> 00:07:45,840 Have you have you right? Have you ever played this big 154 00:07:45,840 --> 00:07:48,560 game Civilization? Oh, it's one of my favorites, 155 00:07:48,560 --> 00:07:50,800 yes. Yeah, and, and I'm sure you're 156 00:07:50,800 --> 00:07:53,440 familiar with that, with that trope where, you know, it's this 157 00:07:53,680 --> 00:07:56,800 just one more turn. Like it's it's so yes, really 158 00:07:56,800 --> 00:07:59,480 such an and. And The funny thing is I feel 159 00:07:59,480 --> 00:08:04,640 the exact same way. If I'm using AI like it, it is 160 00:08:04,640 --> 00:08:07,040 exactly at that right level of where you don't need to worry 161 00:08:07,040 --> 00:08:09,440 about the details too much, but you're kind of in control of 162 00:08:09,440 --> 00:08:13,400 everything and you're, you're racing ahead And it's, there's 163 00:08:13,400 --> 00:08:16,640 something about it that is that really feels like. 164 00:08:17,240 --> 00:08:21,640 Like playing civilization. Oh exactly, exactly. 165 00:08:21,880 --> 00:08:24,520 I, I have, I have alarms on my phone now to make sure that at 166 00:08:24,520 --> 00:08:27,600 this time I stop, you know, the one that should be stopped work, 167 00:08:27,600 --> 00:08:30,320 but particularly stop, you know, working with these tools or 168 00:08:30,320 --> 00:08:32,240 learning more about these tools because they're, they are 169 00:08:32,240 --> 00:08:35,600 changing so fast that you could easily, I mean, Oh my God, you 170 00:08:35,600 --> 00:08:37,320 can literally lose sleep if you're not careful. 171 00:08:38,120 --> 00:08:39,240 So. However, maybe I'm just a 172 00:08:39,240 --> 00:08:42,120 technology enthusiast and and you know about learning the how 173 00:08:42,120 --> 00:08:44,840 to use these properly so I can, you know, share it with other 174 00:08:44,840 --> 00:08:45,920 people. But anyways. 175 00:08:46,560 --> 00:08:50,920 Yeah. But I really sort of thinking 176 00:08:50,920 --> 00:08:53,320 back to the trainees that I give them over the last couple of 177 00:08:53,320 --> 00:08:59,400 months and, and trying to teach people to use AI effectively in 178 00:08:59,400 --> 00:09:01,880 software development and software engineering. 179 00:09:02,280 --> 00:09:06,360 And it's quite interesting that it, you know, despite us saying, 180 00:09:06,400 --> 00:09:08,720 oh, it feels so easy and natural and it's so much fun. 181 00:09:08,720 --> 00:09:10,080 You know, you get so much stuff done. 182 00:09:10,760 --> 00:09:14,360 A lot of people actually struggle a lot with getting off 183 00:09:14,360 --> 00:09:18,920 the ground using AI because turns out it's not all that easy 184 00:09:18,920 --> 00:09:23,560 and there is a lot of like, strangely subtle sort of 185 00:09:23,560 --> 00:09:26,120 implicit knowledge and skill in there, isn't it? 186 00:09:27,160 --> 00:09:31,640 There is there, there is a lot of, yeah, implicit skills of how 187 00:09:31,640 --> 00:09:34,600 do you communicate, how do you mark out what your goals are? 188 00:09:34,600 --> 00:09:37,160 How do you figure out what your output should be? 189 00:09:37,160 --> 00:09:39,560 How do you even identify what is you're trying to have it do? 190 00:09:40,120 --> 00:09:42,120 And then on top of that, nowadays it's it can be just 191 00:09:42,120 --> 00:09:45,000 identifying where does this tool, where is the best fit 192 00:09:45,000 --> 00:09:47,360 within a workflow? What should I be using it for? 193 00:09:47,840 --> 00:09:49,280 Right. So I mean, there's, there are, 194 00:09:49,280 --> 00:09:51,240 there's a lot of skills there that you actually really need to 195 00:09:51,240 --> 00:09:54,400 dive into and in order to in order to make it an effective 196 00:09:54,400 --> 00:09:56,480 tool for you to use to do engineering. 197 00:09:56,920 --> 00:09:59,440 I I was just sort of, as you were saying this, I was thinking 198 00:09:59,920 --> 00:10:04,120 I started my career as a requirements engineer and yeah. 199 00:10:04,480 --> 00:10:09,440 And I'm so I'm sorry. It is don't be, don't be. 200 00:10:09,440 --> 00:10:15,080 It's a lot of fun and it. Must be the most important that 201 00:10:15,080 --> 00:10:18,240 he's learned. That's where I was going with 202 00:10:18,240 --> 00:10:20,200 this. That's where exactly where I was 203 00:10:20,200 --> 00:10:22,320 going with this. Like, I've worked as a product 204 00:10:22,320 --> 00:10:24,040 owner, I've worked as a requirements engineer. 205 00:10:24,720 --> 00:10:31,080 And I find myself going back to these skills and those mindsets 206 00:10:31,080 --> 00:10:34,520 and really sort of sitting down and describing what the system 207 00:10:34,520 --> 00:10:36,840 should do and how it should do it and how it can tell that it's 208 00:10:36,840 --> 00:10:38,920 doing the right thing and how it can tell that it's done. 209 00:10:40,040 --> 00:10:42,760 And once I've nailed that down, then, you know, then I can't 210 00:10:42,760 --> 00:10:46,560 hand it off to the AI and it will do something useful as 211 00:10:46,560 --> 00:10:51,600 opposed to just something. I think a lot of people are not 212 00:10:51,600 --> 00:10:54,200 aware of this. And so they get kind of stumped, 213 00:10:54,200 --> 00:10:58,560 like, you know, they're sitting in front of of the input window 214 00:10:58,680 --> 00:11:01,800 or, you know, in front of the cloud code incense or something. 215 00:11:02,320 --> 00:11:05,560 It's like, OK, now what? You know, make make me an 216 00:11:05,560 --> 00:11:10,680 embedded system. So I'll give you an example. 217 00:11:10,680 --> 00:11:14,200 There's a class that I was, I taught recently where we, we 218 00:11:14,200 --> 00:11:17,800 basically designed a universal push button module. 219 00:11:18,040 --> 00:11:21,520 OK, something that every single one of us has has done right, or 220 00:11:21,520 --> 00:11:23,800 at least we've designed a deep bounce algorithm. 221 00:11:24,280 --> 00:11:27,200 And it wasn't simply you go into, you know, your favorite 222 00:11:27,200 --> 00:11:29,440 LLM and say, make me a push button module. 223 00:11:29,640 --> 00:11:31,760 Sure, you could do that and you get some response. 224 00:11:32,120 --> 00:11:35,240 What we end up doing was spending almost, you know, the 225 00:11:35,400 --> 00:11:37,720 course is actually over 5 days, but we spent almost with the 226 00:11:37,720 --> 00:11:40,440 equivalent of a full day identifying here. 227 00:11:40,440 --> 00:11:42,720 Let's, let's put together requirements, let's put together 228 00:11:42,960 --> 00:11:44,960 user stories. Let's, you know, develop the 229 00:11:44,960 --> 00:11:47,040 architecture. Let's, you know, we didn't, 230 00:11:47,400 --> 00:11:49,680 nothing was coded, even developing like skill sets 231 00:11:49,920 --> 00:11:53,520 within the agent to be able to run, to have an agent run a test 232 00:11:53,520 --> 00:11:57,120 driven development process where it would, you know, do the, the, 233 00:11:57,160 --> 00:12:01,080 the red green refactor process, right, Which was hilarious 234 00:12:01,080 --> 00:12:04,280 working with an AI to do that, because I'm going to digress 235 00:12:04,280 --> 00:12:06,600 here. I, I, I built this agent that 236 00:12:06,600 --> 00:12:09,080 would, would do the TDD process. And it was, it was, I'm like, 237 00:12:09,080 --> 00:12:10,920 OK, we're ready to code. We've got all these 238 00:12:10,920 --> 00:12:12,600 requirements. We've got architecture. 239 00:12:13,080 --> 00:12:15,280 Go ahead and, and follow TDD. You'll make it. 240 00:12:15,560 --> 00:12:19,000 Write a test, watch it fail, you know, write the production code 241 00:12:19,000 --> 00:12:20,200 for it. This is going to be awesome. 242 00:12:20,680 --> 00:12:24,720 And the AI finished writing the code and all the tests in like 4 243 00:12:24,720 --> 00:12:28,840 seconds. And I was like, that was really 244 00:12:28,840 --> 00:12:30,360 fast. And the agent was all happy. 245 00:12:30,360 --> 00:12:32,760 And it's like, Oh yeah, look, I, I did all this stuff for you. 246 00:12:32,760 --> 00:12:34,760 It's great. It, it works fantastic. 247 00:12:34,760 --> 00:12:37,640 I'm like, why? How was that so fast? 248 00:12:38,080 --> 00:12:39,960 Well, I just wrote all the code. You didn't follow the TDD 249 00:12:39,960 --> 00:12:41,920 process. No, that would take too long. 250 00:12:42,800 --> 00:12:44,600 That was literally the, the problem came back in the 251 00:12:44,600 --> 00:12:46,200 problem. I said, why didn't you do it the 252 00:12:46,200 --> 00:12:48,200 right way? That would take too long. 253 00:12:48,240 --> 00:12:50,560 And I was like, OK, that sounds a lot. 254 00:12:50,560 --> 00:12:52,240 I feel that I might get from like a junior engineer 255 00:12:52,240 --> 00:12:55,760 responding back right. And so eventually I forced it to 256 00:12:55,760 --> 00:12:57,960 do the right thing and it generated, you know, pretty 257 00:12:57,960 --> 00:13:01,200 interesting code. But you know, there's, I guess I 258 00:13:01,200 --> 00:13:02,920 digress there a little bit, but it's kind of interesting the, 259 00:13:03,480 --> 00:13:05,080 you know, you have to apply the full workflows. 260 00:13:05,080 --> 00:13:07,880 And even when you have a good workflow, you can't just trust 261 00:13:07,880 --> 00:13:09,840 that it's going to follow what you told it to do. 262 00:13:10,120 --> 00:13:13,920 It may, you know for whatever reason, find other paths or 263 00:13:14,600 --> 00:13:16,800 hallucinate and just decide, well this is better than what 264 00:13:16,800 --> 00:13:18,400 you're you're telling it to actually do. 265 00:13:18,520 --> 00:13:20,800 I mean, Luke and I may have already had the argument about, 266 00:13:20,840 --> 00:13:23,200 you know, is tested from development for people or for 267 00:13:23,200 --> 00:13:29,040 computers like, because if the code, if the code is correct, do 268 00:13:29,040 --> 00:13:31,560 you actually need to have written done it in that, in that 269 00:13:31,560 --> 00:13:37,400 order, you know, like, yeah. And certainly we could debate 270 00:13:37,400 --> 00:13:40,400 TDD for sure, but I feel this probably isn't the right it's. 271 00:13:40,400 --> 00:13:44,200 Very important for humans, but I'm not sold on it being for AI 272 00:13:44,200 --> 00:13:48,880 which we'll be thrown to shoot. Exactly. 273 00:13:48,880 --> 00:13:50,840 So I feel the exact opposite way. 274 00:13:50,920 --> 00:13:57,160 Like I, if I use AII feel very uncomfortable if I'm not doing 275 00:13:57,160 --> 00:14:01,640 TDD, because TDD is my way of describing what I want the 276 00:14:01,640 --> 00:14:05,040 system to, to do what I wanted to behave like. 277 00:14:05,400 --> 00:14:08,440 And then, you know, and then the agent can figure out some way to 278 00:14:08,440 --> 00:14:11,120 make it do that. But if I don't express clearly 279 00:14:11,120 --> 00:14:13,920 what I wanted to do and then I get just something. 280 00:14:13,920 --> 00:14:15,760 And in fact, that was interesting. 281 00:14:15,760 --> 00:14:19,680 Like one of the first things I like to do in, in my trainings 282 00:14:19,680 --> 00:14:25,080 is I have the participants write a new feature for an existing 283 00:14:25,080 --> 00:14:26,680 product. So if, if it's not an embedded 284 00:14:26,680 --> 00:14:29,760 systems training that I tend to use actually an open source 285 00:14:29,760 --> 00:14:34,400 civilization clone, they just have to write an extra function 286 00:14:34,400 --> 00:14:38,000 that you know, that filters messages and it's really easy. 287 00:14:38,000 --> 00:14:41,280 Like they, they will probably, you know, write me a filter 288 00:14:41,280 --> 00:14:43,720 that, that filters messages according to reasonable criteria 289 00:14:43,720 --> 00:14:46,800 or something of that nature. And it will do that and it will 290 00:14:46,800 --> 00:14:51,240 do it very quickly. And then I ask one of them, OK, 291 00:14:51,760 --> 00:14:55,960 show me, run the, run the game, show me the, the message filter. 292 00:14:55,960 --> 00:14:59,440 And then he goes and starts the game and and shows me the 293 00:14:59,440 --> 00:15:02,160 message filter and I start asking questions like, OK, now 294 00:15:02,160 --> 00:15:04,080 how does that work? What happens if I press this 295 00:15:04,080 --> 00:15:09,160 button and and he goes like idea, like presumably something 296 00:15:09,160 --> 00:15:12,800 is going to happen. I like how exactly does it feel 297 00:15:12,800 --> 00:15:17,200 for like that, you know, that have pagination, but how does it 298 00:15:17,200 --> 00:15:19,440 work? And that's where it really 299 00:15:19,440 --> 00:15:23,880 starts to dawn on them that if they throw in such a vague 300 00:15:23,880 --> 00:15:28,720 requirement, the thing can't read their mind and neither can 301 00:15:28,720 --> 00:15:32,280 they read its mind. So, you know, it's, it's, it's 302 00:15:32,280 --> 00:15:34,600 really just a matter of luck. What are you going to get and 303 00:15:34,600 --> 00:15:37,560 whether it's at at least sort of moderately useful. 304 00:15:38,280 --> 00:15:41,520 And clearly you have no idea what what the boundary 305 00:15:41,640 --> 00:15:43,760 conditions are going to be treated like. 306 00:15:44,800 --> 00:15:48,840 So that that that tends to be interesting, an interesting 307 00:15:48,840 --> 00:15:51,280 experience because everybody falls into this trap at first. 308 00:15:51,280 --> 00:15:53,240 This is my experience from training. 309 00:15:53,560 --> 00:15:55,360 They're like, oh, this is so practical. 310 00:15:55,360 --> 00:15:57,520 I can just, you know, have it do the thing. 311 00:15:58,440 --> 00:16:01,320 And then once you start to pull on on this thread, then the 312 00:16:01,320 --> 00:16:05,160 entire thing unravels. Well, and and you, you mentioned 313 00:16:05,280 --> 00:16:08,640 it can't read your mind. Well, the problem is even worse 314 00:16:08,640 --> 00:16:11,880 than that because I there's nothing in there for it to have 315 00:16:11,880 --> 00:16:15,560 read because because I have not thought out the problem space 316 00:16:15,560 --> 00:16:16,840 all the way. Although like looking at you 317 00:16:16,840 --> 00:16:18,080 like, boy, I didn't think about that. 318 00:16:18,280 --> 00:16:20,760 Yes, I know. So not only can it not read your 319 00:16:20,760 --> 00:16:24,200 mind, it also it can't read what's not in your mind. 320 00:16:24,200 --> 00:16:27,120 Like there's you, you've got to actually go through the process 321 00:16:27,120 --> 00:16:30,760 of clearly understanding the problem space and then clearly 322 00:16:30,760 --> 00:16:34,280 identifying what you actually want to have happen for all 323 00:16:34,280 --> 00:16:36,040 those different cases and conditions. 324 00:16:36,320 --> 00:16:39,920 And the AI is the best puppy you'll ever have. 325 00:16:39,920 --> 00:16:41,480 It'll go pick up the stick for you. 326 00:16:41,480 --> 00:16:43,200 And you're like, no, not that stick. 327 00:16:43,200 --> 00:16:44,640 You're like, what do you mean? Not that stick? 328 00:16:44,640 --> 00:16:47,000 You didn't tell me which stick. I grabbed you the stick. 329 00:16:47,280 --> 00:16:50,720 Like it, you're you, you really. And like, it'll do that and 330 00:16:50,720 --> 00:16:54,840 it'll put out pages and pages and pages and pages, hundreds of 331 00:16:54,840 --> 00:16:56,480 sticks. This, this puppy will bring back 332 00:16:56,480 --> 00:16:59,200 to you. And like, you will now be buried 333 00:16:59,200 --> 00:17:03,600 in sticks. And yeah, I, I, I want to talk 334 00:17:03,600 --> 00:17:07,480 about the engineering process a little bit too, because, you 335 00:17:07,480 --> 00:17:11,200 know, I've got, I've got a lot of things to think about when it 336 00:17:11,200 --> 00:17:14,280 comes to like Anthropic and especially it's been talking 337 00:17:14,280 --> 00:17:16,480 about, it'll generate code to generate code to generate code. 338 00:17:16,760 --> 00:17:20,000 If you look at an embed at a software developer's time and 339 00:17:20,000 --> 00:17:22,800 what they spend most of their time doing, it's not writing 340 00:17:22,800 --> 00:17:26,599 code. It's thinking about how to 341 00:17:26,599 --> 00:17:29,800 design it, how to architect it. And then you spend way more time 342 00:17:29,800 --> 00:17:32,600 refactoring somebody else's code or your code because you didn't 343 00:17:32,600 --> 00:17:33,920 understand the problem space enough. 344 00:17:34,920 --> 00:17:37,600 But so like you spend a lot more time thinking than you actually 345 00:17:37,600 --> 00:17:40,240 spend writing code. So I kind of want the AI can, 346 00:17:40,240 --> 00:17:42,480 like, can the AI do the thinking for me instead? 347 00:17:42,480 --> 00:17:52,680 And the answer is no, not yet. I'm in fact not ever. 348 00:17:52,680 --> 00:17:55,200 I think that that is just going to be the forever bottleneck 349 00:17:55,560 --> 00:18:00,280 that you must the the user, or maybe you, as sort of the 350 00:18:00,280 --> 00:18:03,720 standard for the user, must go through the painstaking process 351 00:18:03,720 --> 00:18:05,880 of figuring out what it is that they want. 352 00:18:06,480 --> 00:18:08,600 And then the system can be built according to that. 353 00:18:08,840 --> 00:18:11,920 And that is just not going away. Absolutely. 354 00:18:11,920 --> 00:18:12,560 Yeah. Yeah. 355 00:18:12,560 --> 00:18:13,880 And that's not the technical limitation. 356 00:18:13,880 --> 00:18:15,680 That's just, you know, that's just how the world works. 357 00:18:15,680 --> 00:18:19,960 If you don't, if you don't know, just like Ryan said, what should 358 00:18:19,960 --> 00:18:22,720 be built, then well, that is not going to if you. 359 00:18:24,080 --> 00:18:26,440 Think, you know, look at you live in Europe. 360 00:18:26,440 --> 00:18:28,440 Apparently there are all these buildings in Europe that people 361 00:18:28,440 --> 00:18:32,040 have like put stacked stones one on top of another to make these 362 00:18:32,040 --> 00:18:33,880 really like big structures and stuff too. 363 00:18:34,200 --> 00:18:36,480 They did not just buy. They did not just go and be 364 00:18:36,480 --> 00:18:38,920 like, I don't make a building where over there that looks 365 00:18:38,920 --> 00:18:39,960 good. Like that's not. 366 00:18:40,040 --> 00:18:45,320 That's not how that works like. Well, well, yes and no. 367 00:18:45,320 --> 00:18:48,600 Like we had this argument before about, you know, measure twice, 368 00:18:48,600 --> 00:18:52,680 cut once, which is very true in in carpentry because like once 369 00:18:52,680 --> 00:18:56,680 you've cut a piece of wood, then it's like broken, uncut it, you 370 00:18:56,680 --> 00:18:57,880 know? Yeah, exactly. 371 00:18:57,880 --> 00:18:59,000 You can't. One way or? 372 00:18:59,560 --> 00:19:01,640 They they are not free, they are not infinite. 373 00:19:01,760 --> 00:19:04,120 But The funny thing is code is free and infinite. 374 00:19:04,280 --> 00:19:07,920 Like you, you can have an infinite number of essentially 375 00:19:07,920 --> 00:19:12,240 free and immediate prototypes. So there is there is sense to 376 00:19:12,240 --> 00:19:16,480 being more experimental and sort of feeling your way towards the 377 00:19:16,480 --> 00:19:21,160 right solution than it were if if you had to build expensive 378 00:19:21,240 --> 00:19:23,880 both in terms of time and money prototypes. 379 00:19:25,280 --> 00:19:27,440 But you know, there's sort of a boundary is there? 380 00:19:27,440 --> 00:19:31,760 There is they are only almost free and only almost instant. 381 00:19:31,760 --> 00:19:34,280 So there is a time penalty that you pay. 382 00:19:34,280 --> 00:19:36,760 And at some point, experimentation becomes more 383 00:19:36,760 --> 00:19:40,200 expensive than sitting down and thinking about for a couple of 384 00:19:40,200 --> 00:19:42,640 minutes, even if experiments are almost free. 385 00:19:43,080 --> 00:19:46,280 Yeah, well, and then and then I can like, OK, here's the 386 00:19:46,280 --> 00:19:48,920 trolling me. Well, now Anthropic and Gemini 387 00:19:48,920 --> 00:19:50,240 are making it so that it isn't free. 388 00:19:50,240 --> 00:19:51,960 You have to pay for that code to be generated. 389 00:19:55,080 --> 00:19:56,560 Yeah, actually also a good point. 390 00:19:57,280 --> 00:19:59,480 Yeah, I mean, at some point someone's got to pay for the 391 00:19:59,480 --> 00:20:02,440 servers and maintenance on the servers and electricity and 392 00:20:02,440 --> 00:20:03,520 everything, right? So. 393 00:20:03,600 --> 00:20:06,560 Yeah, as far as you can tell, it's just a big circle. 394 00:20:06,560 --> 00:20:09,600 They're just passing the same dollar run around the circles of 395 00:20:09,840 --> 00:20:15,960 the economy so. Yeah, well, we'll, we'll see. 396 00:20:16,000 --> 00:20:17,680 We'll see how that works. I just learned last night 397 00:20:17,680 --> 00:20:20,320 they're building a data centre about 40 minutes from here. 398 00:20:20,320 --> 00:20:23,280 And when we checked, all of our electricity bills have actually 399 00:20:23,280 --> 00:20:27,040 gone up by 50% this year. And it seems to be tied to the 400 00:20:27,040 --> 00:20:30,760 building of this data centre and the electrical grid company 401 00:20:31,080 --> 00:20:35,200 adding more supply to the local grid really just for the data 402 00:20:35,200 --> 00:20:36,920 centre. But they're, you know, the whole 403 00:20:36,920 --> 00:20:40,040 region is being charged extra to cover those costs. 404 00:20:40,040 --> 00:20:42,720 So, so I mean, it's, that's why we started to talk about cost of 405 00:20:42,720 --> 00:20:44,080 AI. There's, there's these hitting 406 00:20:44,080 --> 00:20:48,320 costs that, you know, maybe we don't fully understand what the 407 00:20:48,320 --> 00:20:50,680 actual economic cost is of using these tools yet, right? 408 00:20:50,920 --> 00:20:53,120 Or becoming maybe dependent on them at some point. 409 00:20:53,120 --> 00:20:54,280 Yep. Oh yeah. 410 00:20:55,160 --> 00:20:57,040 Yes. And there's yeah, but being 411 00:20:57,120 --> 00:21:06,040 depend anyways. OK, so, So what have you seen 412 00:21:06,040 --> 00:21:07,640 the students that you've interacted with? 413 00:21:07,640 --> 00:21:09,960 What have you seen them, you know, you mentioned they 414 00:21:09,960 --> 00:21:12,240 struggle with it. You know, what is it they're 415 00:21:12,240 --> 00:21:15,560 doing wrong? What are the, you know, and you 416 00:21:15,560 --> 00:21:17,720 know, like, and then how do you get them to trust it? 417 00:21:17,720 --> 00:21:21,480 Because I know that if I, if I try something and it doesn't 418 00:21:21,480 --> 00:21:23,520 really work the first time, I give up immediately and throw my 419 00:21:23,520 --> 00:21:24,640 hands up and walking out of the room. 420 00:21:24,920 --> 00:21:28,000 So like, how are, how are you dealing with like, oh, it's not 421 00:21:28,000 --> 00:21:30,520 doing what I wanted to do. This is not how I think about 422 00:21:30,520 --> 00:21:32,760 solving the problem. This is not how I am used to 423 00:21:32,760 --> 00:21:34,560 working. Like how do you deal with that? 424 00:21:36,440 --> 00:21:38,600 Yeah. So, so like for me, typically 425 00:21:38,600 --> 00:21:41,440 what I've been doing is, is I try to demonstrate to them like, 426 00:21:41,600 --> 00:21:43,600 you know, something that doesn't work right. 427 00:21:43,600 --> 00:21:45,520 And then you try to work them through the critical thinking 428 00:21:45,520 --> 00:21:47,280 skills of how would you do this as an engineer? 429 00:21:47,640 --> 00:21:50,320 What, what, what is it that is essentially failing in your 430 00:21:50,320 --> 00:21:52,640 workflow process? And then are there other tools 431 00:21:52,640 --> 00:21:56,040 that you can do to refine either your prompts or even more, more 432 00:21:56,040 --> 00:21:59,600 and more today it's how do you how do you adjust the skills? 433 00:21:59,720 --> 00:22:01,960 How do you deploy agents? Are there other solutions out 434 00:22:01,960 --> 00:22:04,160 there like MCP servers that you can leverage or does it make 435 00:22:04,160 --> 00:22:06,640 sense to build your own? And so ultimately, a lot of 436 00:22:06,640 --> 00:22:09,240 times what we're focusing on, or I'm focusing on is trying to 437 00:22:09,240 --> 00:22:11,080 teach them to critically think about what it is that they would 438 00:22:11,080 --> 00:22:14,000 do and how they would do it. And documenting that in a good 439 00:22:14,000 --> 00:22:16,880 degree, in such a way that they can actually wrap that up into a 440 00:22:16,880 --> 00:22:20,760 skill that they can then share with AI and try to no skill that 441 00:22:20,760 --> 00:22:23,600 you write is going to make AI do it perfectly every time. 442 00:22:23,600 --> 00:22:26,440 It's going to be, you know, it might get you 80% of the way 443 00:22:26,440 --> 00:22:28,920 there, but you do have to go in and make adjustments to it. 444 00:22:28,920 --> 00:22:31,480 So it's not like, you know, from my experience, it's not like, 445 00:22:31,480 --> 00:22:32,600 oh, I can just tell it to go do this. 446 00:22:32,600 --> 00:22:34,880 And I built all these skills and agents and stuff and you know, 447 00:22:34,880 --> 00:22:37,200 I'm going to go say, hey, go write me this widget and, you 448 00:22:37,200 --> 00:22:39,400 know, go to the bar and have a couple of drinks and come back 449 00:22:39,400 --> 00:22:42,640 and woohoo, you know, go to production it. 450 00:22:42,640 --> 00:22:45,600 It's very, I see this very much as a piece meal tool where I 451 00:22:45,600 --> 00:22:47,640 have a small problem. Let's use it to attack this 452 00:22:47,640 --> 00:22:49,720 problem. We developed, you know, skills 453 00:22:49,720 --> 00:22:51,920 and libraries essentially to help us solve these problems 454 00:22:51,920 --> 00:22:54,760 that keep repeating over and over and over time we refine 455 00:22:54,760 --> 00:22:58,040 them and, you know, can then deploy agents to use those 456 00:22:58,040 --> 00:23:00,120 skills to, to maybe move it along faster. 457 00:23:00,120 --> 00:23:02,600 And it's kind of like a refinement process. 458 00:23:02,880 --> 00:23:05,000 But really trying to show students that they can't just 459 00:23:05,000 --> 00:23:08,320 trust what it outputs, I think is absolutely really important 460 00:23:08,320 --> 00:23:10,520 because first of all, they often get stuck with, well, how do I 461 00:23:10,520 --> 00:23:13,160 even, you know, craft a prop properly and get something 462 00:23:13,160 --> 00:23:15,120 useful out of it? And then they start to see 463 00:23:15,120 --> 00:23:17,000 something useful and they're like, oh, I can just trust this 464 00:23:17,000 --> 00:23:18,840 thing. But you really have to, you 465 00:23:18,840 --> 00:23:22,440 know, I've noticed like 80 maybe at least half the time it will 466 00:23:22,440 --> 00:23:25,160 give you something kind of close, but you really almost 467 00:23:25,160 --> 00:23:28,400 always have to push back on it and say, no, this is the way I'm 468 00:23:28,400 --> 00:23:30,000 thinking. Or you discover holes in your 469 00:23:30,000 --> 00:23:32,560 own way of thinking that you have to keep refining your own 470 00:23:32,560 --> 00:23:34,320 thinking. And it's honestly, it's a, it's 471 00:23:34,320 --> 00:23:35,960 a, it's a human process. Like I, I've got my 472 00:23:35,960 --> 00:23:37,400 modernization framework that I use. 473 00:23:37,600 --> 00:23:39,320 I've been refining this thing over the course of like the last 474 00:23:39,320 --> 00:23:42,160 four years and it continues to change, right? 475 00:23:42,440 --> 00:23:45,200 And so every iteration with using AI, you kind of have to 476 00:23:45,200 --> 00:23:47,200 show students that this is a process. 477 00:23:47,200 --> 00:23:49,600 You're not, you know, you don't go to the gym and, and do a 478 00:23:49,600 --> 00:23:52,280 deadlift and, and now you're, you know, ripped you learning 479 00:23:52,280 --> 00:23:54,840 how to use these things properly takes time, right? 480 00:23:55,360 --> 00:23:59,040 So and experience, but sorry maybe I'm rambling there a 481 00:23:59,040 --> 00:24:01,720 little bit. No, no, no, but, but it's quite 482 00:24:01,720 --> 00:24:03,840 interesting. I, I think the engineers that 483 00:24:03,840 --> 00:24:07,960 actually take the trainings, they tend to have a fairly good 484 00:24:07,960 --> 00:24:10,640 intuition about this, but that actually is not going to be that 485 00:24:10,640 --> 00:24:12,280 easy. You don't just throw a prop at 486 00:24:12,280 --> 00:24:15,160 it and then you go and have a coffee and then you come back 487 00:24:15,160 --> 00:24:17,000 and it's like all done and perfect. 488 00:24:17,600 --> 00:24:21,480 That tends to be more of a management affliction where they 489 00:24:22,240 --> 00:24:25,040 seem to really think that this is how it could work. 490 00:24:25,760 --> 00:24:28,800 The engineers tend to be more, you know, they have better 491 00:24:28,800 --> 00:24:30,920 instincts about this. They say, I don't think it's 492 00:24:30,920 --> 00:24:32,960 going to be that easy. I think I need to pay closer 493 00:24:32,960 --> 00:24:33,840 attention. And they're right. 494 00:24:34,640 --> 00:24:37,440 And so I, I told you about this first exercise that I do the 495 00:24:37,440 --> 00:24:41,200 second exercise that I, you know, let's, let's talk it 496 00:24:41,400 --> 00:24:43,840 through in the embedded systems scenario here. 497 00:24:43,840 --> 00:24:49,280 So that the second exercise that I like to do is I have them, or 498 00:24:49,280 --> 00:24:53,320 rather I have the AII have them have the AI, right, A driver for 499 00:24:53,360 --> 00:24:56,800 a temperature sensor. You know, it's a very simple 500 00:24:56,880 --> 00:24:59,480 sensor. I'm just showing a breadboard 501 00:24:59,480 --> 00:25:01,720 here from, from a training that I did a couple of days ago. 502 00:25:02,880 --> 00:25:05,040 Temperature sensor fairly well known. 503 00:25:05,040 --> 00:25:09,160 And so I, I tell them, but ask the AI to make it work and I'll 504 00:25:09,160 --> 00:25:10,760 put the temperature on the, on the cereal pot. 505 00:25:11,880 --> 00:25:17,280 And it usually works first try, like some, sometimes it needs a 506 00:25:17,280 --> 00:25:19,320 couple of iterations until it all works. 507 00:25:20,040 --> 00:25:23,600 Oh no, that's right. There's always 1 snag that comes 508 00:25:23,600 --> 00:25:26,400 up, which is that it assumes that it's going to be that, that 509 00:25:27,000 --> 00:25:30,400 the sensor is going to be a whatever pin and it never stops 510 00:25:30,400 --> 00:25:32,360 to ask like OK, where's it actually connected? 511 00:25:32,520 --> 00:25:34,960 Just makes up a pin and and runs with it. 512 00:25:35,680 --> 00:25:40,000 And once we get that sorted out, then usually it actually works 513 00:25:40,000 --> 00:25:42,600 first try. Then I say OK, fine, let's let's 514 00:25:42,600 --> 00:25:45,320 pull out the data sheet. It's fairly well known sensor, 515 00:25:45,320 --> 00:25:46,760 right? It's a Dallas 18B20. 516 00:25:47,240 --> 00:25:50,280 Let's pull out the data sheet and see if the civilization is, 517 00:25:50,640 --> 00:25:53,880 is right. And it tends to be subtly not 518 00:25:54,120 --> 00:25:57,920 interestingly enough, like the, the time is sort of good enough 519 00:25:57,920 --> 00:26:02,320 to get a signal. You know, in good conditions, 520 00:26:02,440 --> 00:26:07,000 it, it usually kind of works, but it's not perfectly right. 521 00:26:07,000 --> 00:26:11,240 It's it's not good enough for production, but it would have 522 00:26:11,240 --> 00:26:12,960 slipped through if we had paid attention. 523 00:26:12,960 --> 00:26:15,840 And that's so interesting that that that tends to be sort of 524 00:26:15,880 --> 00:26:20,080 one of those first really sort of wake up moments for students 525 00:26:20,080 --> 00:26:22,440 where they say, oh, I really need to be paying attention 526 00:26:22,440 --> 00:26:25,480 because I cannot rely on the AI to have done the right thing, 527 00:26:25,960 --> 00:26:29,160 even if the system exhibits the behaviors that I desire, like, 528 00:26:29,160 --> 00:26:30,880 you know. It's a temperature. 529 00:26:30,880 --> 00:26:34,000 Maybe the tests pass? The tests pass exactly. 530 00:26:35,600 --> 00:26:38,040 Yeah. And that is so interesting. 531 00:26:38,040 --> 00:26:42,640 And they, they sort of, that is that is the first big sort of 532 00:26:42,640 --> 00:26:46,680 wake up moment for them where they say I can't really let go 533 00:26:46,680 --> 00:26:51,480 of the rains, can I? Try telling the executive board 534 00:26:51,480 --> 00:26:55,240 that like the C, the C level, the management level is like, Oh 535 00:26:55,240 --> 00:26:58,680 yeah, because I'm telling you, we went and and did a bunch of 536 00:26:58,680 --> 00:27:01,000 work and all these VCs are like, no, no, we don't need testing 537 00:27:01,000 --> 00:27:02,200 anymore. We're just going to have the AI 538 00:27:02,200 --> 00:27:03,840 write all the code. And I'm like, Oh my God. 539 00:27:06,760 --> 00:27:08,880 It's, it's definitely a management push too, because I 540 00:27:08,880 --> 00:27:10,880 will see like a lot of the trainings that I've done, like 541 00:27:10,880 --> 00:27:15,200 for my faster smarter firmware seminars, you'll see management 542 00:27:15,200 --> 00:27:20,120 telling the teams that they need to learn AI, identify how they 543 00:27:20,120 --> 00:27:21,760 can use AI. But you see the actual 544 00:27:21,760 --> 00:27:24,640 development team, I don't want to say necessarily resisting, 545 00:27:24,640 --> 00:27:28,040 but kind of being like, well, instinctively, like you said, 546 00:27:28,040 --> 00:27:29,600 they know that there's going to be problems with it. 547 00:27:29,600 --> 00:27:30,640 It's actually going to add overhead. 548 00:27:30,640 --> 00:27:33,440 It'll actually initially slow them down because they're trying 549 00:27:33,440 --> 00:27:36,760 to learn something new and it's not going to just be a magic 550 00:27:36,760 --> 00:27:38,680 wand that you wave. And hey, here's our solution. 551 00:27:38,880 --> 00:27:41,080 It might give something, but now we have to go back and validate 552 00:27:41,080 --> 00:27:43,640 it, verify it's actually doing what we think it's doing. 553 00:27:44,040 --> 00:27:47,800 And the, the superficialness of the it's working ends up being 554 00:27:47,800 --> 00:27:50,880 really more time and overhead potentially on a project that 555 00:27:51,120 --> 00:27:53,080 has, you know, is already three months behind, man. 556 00:27:53,080 --> 00:27:54,960 He's already got the pressure on the move faster. 557 00:27:54,960 --> 00:27:56,960 And, you know, yeah, it's a risk. 558 00:27:57,040 --> 00:27:59,160 It's a risk, isn't it? Exactly. 559 00:27:59,440 --> 00:28:01,120 Yep. Also, the other interesting 560 00:28:01,120 --> 00:28:05,360 thing is I, I talked to to an engineering manager the other 561 00:28:05,360 --> 00:28:09,360 week at OP conference in Munich and she said that, you know, 562 00:28:09,720 --> 00:28:14,760 she's got a couple of teams of engineers using AI for firmware 563 00:28:14,760 --> 00:28:16,360 development and beta systems development. 564 00:28:16,840 --> 00:28:20,120 They've really had to scale back on using AI because it got so 565 00:28:20,120 --> 00:28:22,600 expensive. Like a single engineer will burn 566 00:28:22,600 --> 00:28:25,720 through a couple of $1000 worth of tokens in a month. 567 00:28:26,000 --> 00:28:32,000 Like they they spend more than they salary or on API cost and 568 00:28:32,000 --> 00:28:34,360 it you know it, it really is a problem for them. 569 00:28:36,160 --> 00:28:38,200 Oh man, Yep. I mean that's just it as you 570 00:28:38,200 --> 00:28:40,760 start to generate more and more code, it just you know, it 571 00:28:40,760 --> 00:28:41,880 starts to eat away at the budget. 572 00:28:41,880 --> 00:28:46,000 So and then you know, again, if the engineers are going back 573 00:28:46,000 --> 00:28:48,840 behind and fixing the the AI, the question does come. 574 00:28:48,840 --> 00:28:52,120 Do you just write the code yourself the first time, right? 575 00:28:52,320 --> 00:28:55,000 You already have a sunk overhead cost, right So. 576 00:28:55,800 --> 00:28:57,520 Yeah, well, and you've, and you've talked about it. 577 00:28:57,560 --> 00:28:59,640 We've touched on it like, no, no, you have to really tell it 578 00:28:59,640 --> 00:29:01,440 what you want. You have to really design it 579 00:29:01,440 --> 00:29:03,800 correctly. So then I'm like, OK, well, 580 00:29:04,080 --> 00:29:06,160 let's just go through the process of doing the design and 581 00:29:06,160 --> 00:29:08,040 then I'll just write the code like. 582 00:29:10,680 --> 00:29:14,440 Yeah, actually that is, that is an interesting point because in 583 00:29:14,440 --> 00:29:17,760 many cases our quote UN quote design process was really sort 584 00:29:17,760 --> 00:29:21,040 of sitting down, writing the code and then going, I need to 585 00:29:21,040 --> 00:29:23,680 think about this and coming to a decision. 586 00:29:23,840 --> 00:29:26,680 And then sort of at the same time writing the code that that 587 00:29:26,760 --> 00:29:29,280 expressed the decision as they were coming to the decision. 588 00:29:29,440 --> 00:29:32,480 And now that they're not writing the code anymore, just like Jake 589 00:29:32,480 --> 00:29:35,920 was saying before, how do you make sure that you've really 590 00:29:36,080 --> 00:29:37,560 understood the problem deeply enough? 591 00:29:37,560 --> 00:29:40,600 How do you make the decisions that need to be made? 592 00:29:41,800 --> 00:29:44,360 We'll call this episode the Death of Agile because you don't 593 00:29:44,360 --> 00:29:45,880 need to be agile. It's just waterfall now. 594 00:29:45,880 --> 00:29:48,760 You describe the problem as best you can to the EI and then it 595 00:29:48,760 --> 00:29:50,760 just writes the code for you in a one shot. 596 00:29:51,000 --> 00:29:52,880 Done. Don't need agile anymore. 597 00:29:55,800 --> 00:29:56,960 That's that's an interesting one. 598 00:29:57,080 --> 00:30:00,320 We, you know, Ryan and I have been talking about we need to do 599 00:30:00,320 --> 00:30:04,920 an episode on, you know, specification for AI and 600 00:30:04,920 --> 00:30:06,640 spectrum development and all that good stuff. 601 00:30:07,360 --> 00:30:10,440 But the point I was just thinking of is that all of the 602 00:30:10,600 --> 00:30:14,600 presently existing at least sort of a bit in the public eye 603 00:30:15,160 --> 00:30:19,720 spectrum development tool kits for AI, you know, like a spec 604 00:30:19,720 --> 00:30:22,160 kit and what's it Bmad and all of those. 605 00:30:23,120 --> 00:30:25,920 What annoys me about them is that they are all too waterfall 606 00:30:25,920 --> 00:30:28,480 Y yeah, yeah. Because they they sort of have 607 00:30:28,480 --> 00:30:31,360 this assumption that you will spec everything out beforehand 608 00:30:31,880 --> 00:30:36,760 and then you go have lunch and you know, your gaggle of agents 609 00:30:36,880 --> 00:30:38,960 does the thing and then you come back. 610 00:30:39,040 --> 00:30:39,840 It's called a. Murder. 611 00:30:39,840 --> 00:30:42,520 A murder of agents. A murder of agent? 612 00:30:42,520 --> 00:30:46,840 Yeah, at least for me, that's not how it works. 613 00:30:46,840 --> 00:30:51,480 Like I, I don't remember showing my AI workflow to a customer and 614 00:30:51,480 --> 00:30:53,880 they and they were like, but you're still sitting there. 615 00:30:53,880 --> 00:30:55,720 It's like, yes, of course I'm still sitting there. 616 00:30:55,800 --> 00:30:58,640 I need to. I need to watch for not if but 617 00:30:58,640 --> 00:31:01,560 when it goes off the rails. And then I need to sort of catch 618 00:31:01,560 --> 00:31:07,120 it and and put it back. And also I need to spot when 619 00:31:07,720 --> 00:31:10,680 when there are certain assumptions of mine that don't 620 00:31:10,680 --> 00:31:14,040 hold or or certain aspects that I had overlooked at where I need 621 00:31:14,040 --> 00:31:15,760 to say stop. We need to figure this out 622 00:31:15,760 --> 00:31:19,560 before we go further. So that's, that's just still an 623 00:31:19,560 --> 00:31:21,280 integral part of the entire process. 624 00:31:21,480 --> 00:31:24,040 So of course, yes, that's still very much agile, right? 625 00:31:24,040 --> 00:31:28,200 You, you iterate yourself towards a good solution because 626 00:31:28,200 --> 00:31:31,360 you will, you just, you will miss something initially. 627 00:31:31,440 --> 00:31:33,200 You just will. Let's face it. 628 00:31:34,040 --> 00:31:37,680 Well, and that's, that's a huge problem I have with a lot of, a 629 00:31:37,680 --> 00:31:41,480 lot of. And so if, if you've been what 630 00:31:41,480 --> 00:31:43,680 I've done, what I'm doing, which is like, I'm going to go watch 631 00:31:43,680 --> 00:31:46,280 some, some Youtubers and see how they're doing all this stuff. 632 00:31:46,280 --> 00:31:50,840 And like the, the, the YouTube hype about set it and forget it 633 00:31:51,200 --> 00:31:53,280 is all waterfall. It's all waterfall. 634 00:31:53,280 --> 00:31:56,360 It is all, everything's planned intricately upfront. 635 00:31:56,360 --> 00:31:59,360 And then you just, you want to get the human out of the loop. 636 00:31:59,360 --> 00:32:01,600 You just want the AI to do all of the work. 637 00:32:01,600 --> 00:32:03,520 There's no human loop. That's how you get the most 638 00:32:03,520 --> 00:32:05,760 speed and efficiency for your engineering effort. 639 00:32:05,760 --> 00:32:09,360 I'm just like, I you can't. And I've, I've been hounding my 640 00:32:09,360 --> 00:32:13,000 team for years. You can't know the entire 641 00:32:13,000 --> 00:32:15,200 problem space before you start writing code. 642 00:32:15,400 --> 00:32:18,080 You have to explore. That's the whole point. 643 00:32:18,080 --> 00:32:20,840 That's the whole point of dev dev SEC opposite extreme 644 00:32:20,840 --> 00:32:22,520 programming, agile. That's the whole point. 645 00:32:22,640 --> 00:32:26,400 You can't know the whole problem space before you start testing 646 00:32:26,400 --> 00:32:30,440 it and start playing with it. And so like to see them like, 647 00:32:30,440 --> 00:32:32,160 Yep, I'm just going to go and I'm just going to. 648 00:32:32,280 --> 00:32:33,520 Everything is perfectly laid out. 649 00:32:33,520 --> 00:32:36,160 And then we just hit the domino and then it's over and I'm like, 650 00:32:39,440 --> 00:32:42,480 no. And the interesting thing is 651 00:32:42,480 --> 00:32:45,120 that that iterative process is the right way to go. 652 00:32:45,120 --> 00:32:50,200 But the way in which they are designing the Agents MCP servers 653 00:32:50,320 --> 00:32:54,600 and these models is to try to just go, yeah, with little to no 654 00:32:54,600 --> 00:32:57,200 human interaction, which I actually, I'll be honest, it 655 00:32:57,200 --> 00:32:59,680 drives me crazy. One of the one of my most 656 00:33:00,120 --> 00:33:03,960 probably one of the one of the things I type in the most is do 657 00:33:04,000 --> 00:33:09,040 this one thing and then stop for review or the next second prompt 658 00:33:09,040 --> 00:33:12,120 is why did you keep going? I didn't tell you to do that. 659 00:33:13,200 --> 00:33:15,320 Oh, I'm sorry, Let's slow down and go back. 660 00:33:15,320 --> 00:33:18,760 That's like a constant like fighting, like, I'll say it, 661 00:33:18,920 --> 00:33:20,840 fighting the tool essentially, right, but. 662 00:33:21,560 --> 00:33:23,880 Yeah, I think it's sort of it's sort of two sided. 663 00:33:23,920 --> 00:33:26,360 I think, you know, they are DPTS, they are generative, they 664 00:33:26,360 --> 00:33:29,560 want to generate this is like the the whole world purpose of 665 00:33:29,560 --> 00:33:32,240 life. But also, I think, I think they 666 00:33:32,240 --> 00:33:38,040 are just trained by, you know, the manufacturers to be biased 667 00:33:38,040 --> 00:33:39,880 for generation because that looks cool, right? 668 00:33:40,200 --> 00:33:42,600 Oh, it can do an entire thing in one go. 669 00:33:43,160 --> 00:33:46,920 You know that that looks really, really impressive until you, you 670 00:33:46,920 --> 00:33:49,560 know, peel back a couple of layers and you see that it's 671 00:33:49,560 --> 00:33:51,400 really not good. And, you know, I, Jacob, I've 672 00:33:51,400 --> 00:33:54,240 had the same experience as you. I say fight, let's do a TDD 673 00:33:54,240 --> 00:33:56,120 workflow. This is how I want it done. 674 00:33:56,360 --> 00:33:58,520 And then I say, oh, it's implementing it. 675 00:33:58,520 --> 00:34:01,320 There is no tests yet. And then I stop it and and then 676 00:34:01,320 --> 00:34:06,240 I ask it, please tell me what is what is our agreed process? 677 00:34:08,040 --> 00:34:10,360 Oh yes, that's right. I should have used TDD and I 678 00:34:10,360 --> 00:34:10,800 didn't. Oh. 679 00:34:11,239 --> 00:34:13,280 My God, it's like, it's like parenting. 680 00:34:13,320 --> 00:34:16,440 You go to your kids, you're like, OK, tell me what I just 681 00:34:16,440 --> 00:34:18,040 told you. Yeah. 682 00:34:18,400 --> 00:34:20,560 Yeah, it doesn't. It doesn't just happen 683 00:34:20,560 --> 00:34:22,280 sometimes. It happens a couple of times a 684 00:34:22,280 --> 00:34:24,719 day. It's like you idiot, you should 685 00:34:24,719 --> 00:34:27,320 have known better. I'll create a I'm. 686 00:34:27,560 --> 00:34:30,840 Sorry. I'll create a skill that 687 00:34:30,840 --> 00:34:33,840 outlines exactly how, like I want, you know, a diagram or for 688 00:34:33,840 --> 00:34:36,800 some piece of code to be written and it's supposed to invoke that 689 00:34:36,800 --> 00:34:39,639 skill or the agent that I wrote when this specific thing comes 690 00:34:39,639 --> 00:34:41,120 up and it'll go through a whole process. 691 00:34:41,120 --> 00:34:43,280 And then I'll be like, why didn't you use my agent? 692 00:34:43,320 --> 00:34:45,320 Did you look at my skill? No, sorry. 693 00:34:45,320 --> 00:34:46,400 Why don't we go back and redo it? 694 00:34:46,639 --> 00:34:49,440 Yes. So I mean, you see that happen 695 00:34:49,440 --> 00:34:52,520 all the time. So, and I've even gone into and 696 00:34:52,520 --> 00:34:55,159 I specifically said do this, this, this and this. 697 00:34:55,639 --> 00:34:59,000 And then it goes and does and it either it it'll even read the 698 00:34:59,000 --> 00:35:03,440 skill and then not use it or it'll use part of it like and 699 00:35:03,440 --> 00:35:06,560 it's like, OK, all right, so then I'll tell it, then I'll 700 00:35:06,560 --> 00:35:07,920 stop it. And I say, hey, hold on. 701 00:35:08,160 --> 00:35:10,280 I noticed that you didn't use the skill here. 702 00:35:10,760 --> 00:35:12,720 Why didn't you use the skill here? 703 00:35:13,000 --> 00:35:14,400 And then it says, oh, I should have. 704 00:35:14,400 --> 00:35:15,600 I'm like, I get that you should have. 705 00:35:15,600 --> 00:35:18,480 And then I'll say, how can I have prompted you? 706 00:35:18,480 --> 00:35:21,160 Better tell me what I could have prompted you with that would 707 00:35:21,160 --> 00:35:23,560 have made you use the skill. And it will literally tell me 708 00:35:23,720 --> 00:35:25,360 you did the right thing, it was my fault. 709 00:35:25,360 --> 00:35:32,680 I'm like. Yeah, but they said, you know, I 710 00:35:32,680 --> 00:35:35,800 see this also sometimes, like you see what lines of file they 711 00:35:35,800 --> 00:35:37,800 read, right? They say, oh, I've been really 712 00:35:37,800 --> 00:35:42,800 like from line 100 to 220, I see them really part of a skill. 713 00:35:42,800 --> 00:35:45,480 It's like, I get why they do it, conserve tokens right? 714 00:35:45,960 --> 00:35:49,000 But this is the entire description of my process that I 715 00:35:49,000 --> 00:35:51,640 asked you to follow. Why are you only reading the 716 00:35:51,640 --> 00:35:54,160 middle part? What? 717 00:35:56,400 --> 00:35:59,280 So. So, but let's let's maybe let's 718 00:35:59,280 --> 00:36:05,120 go back to to what we observed. Like I remember also when I did 719 00:36:05,120 --> 00:36:07,440 a training and I'm really curious whether you've seen 720 00:36:08,600 --> 00:36:12,400 other variations of that Jacob, where we where we did sort of a 721 00:36:12,560 --> 00:36:15,200 big final project. We, we spent a couple of days 722 00:36:15,600 --> 00:36:18,160 learning about how to use AI, how to write specifications and 723 00:36:18,160 --> 00:36:20,680 so on and so forth. And we said, fine, now we're 724 00:36:20,680 --> 00:36:22,240 going to implement an actual feature. 725 00:36:22,800 --> 00:36:25,720 And it is so interesting to watch those people who, who were 726 00:36:25,720 --> 00:36:28,120 all, you know, competent software engineers had been 727 00:36:28,120 --> 00:36:31,000 doing this for decades in some cases. 728 00:36:31,480 --> 00:36:35,200 And they were sort of interestingly timid with AI. 729 00:36:35,200 --> 00:36:38,560 Like, OK, not so much what now? I think they had a pretty good 730 00:36:38,560 --> 00:36:40,200 idea of what they were supposed to do now. 731 00:36:40,200 --> 00:36:43,040 But then what is that? Is that the right way of doing 732 00:36:43,040 --> 00:36:45,680 it? Like should I follow the AI? 733 00:36:45,680 --> 00:36:48,800 Shouldn't I follow the AI? In fact, that's maybe one of the 734 00:36:48,800 --> 00:36:52,880 things that I noticed the most that I would have stopped a lot 735 00:36:52,880 --> 00:36:54,640 earlier and said, no, no, this is a terrible idea. 736 00:36:54,640 --> 00:36:58,320 No, no, no, you did this wrong. But they sort of kept accepting 737 00:36:58,320 --> 00:37:02,920 the AI as suggestions that noticeably went the wrong way. 738 00:37:03,320 --> 00:37:08,560 So I, I think there is a lot of experience in, you know, in, in, 739 00:37:08,560 --> 00:37:10,760 in reading where it goes. Yeah. 740 00:37:11,320 --> 00:37:14,800 And that's, that's where I have a fear of people using AI, not 741 00:37:14,800 --> 00:37:16,320 using their critical thinking skills, right. 742 00:37:16,440 --> 00:37:20,680 They just the machine output it and they kind of say, well, I 743 00:37:20,680 --> 00:37:23,160 mean, this tool is trained on all this knowledge. 744 00:37:23,160 --> 00:37:26,880 Maybe, maybe it knows better than I do, right? 745 00:37:26,880 --> 00:37:29,920 And so you just kind of accept it and then you just accept it. 746 00:37:30,720 --> 00:37:33,280 Stack Overflow just saying exactly exactly. 747 00:37:34,760 --> 00:37:38,360 All of that was downvoted. Like why is the first response 748 00:37:38,360 --> 00:37:40,320 downvoted? I don't understand. 749 00:37:41,960 --> 00:37:45,760 So, but I've I've seen the exact same thing and I'll be honest, I 750 00:37:45,760 --> 00:37:48,480 saw myself kind of earlier on when I was experimenting, maybe 751 00:37:48,480 --> 00:37:50,960 even doing the exact same thing. And I literally had to stop 752 00:37:50,960 --> 00:37:52,840 myself and say, wait a second here. 753 00:37:52,840 --> 00:37:55,720 I'm the expert at doing this. I've been doing this for, you 754 00:37:55,720 --> 00:37:58,240 know, a long time. Would I do it like this? 755 00:37:58,360 --> 00:38:01,200 No, and you really have to kind of stop yourself and say this is 756 00:38:01,200 --> 00:38:02,040 a tool. I think. 757 00:38:02,360 --> 00:38:04,600 And you know, anyone who is starting to use it, don't just 758 00:38:04,600 --> 00:38:07,120 accept it. Literally push back, right, 759 00:38:07,880 --> 00:38:11,200 because a lot of times it is going to it will lead you, I 760 00:38:11,200 --> 00:38:12,680 think, down. I don't want, I won't say it 761 00:38:12,680 --> 00:38:15,040 will, but I mean, it's going to go in directions that you 762 00:38:15,040 --> 00:38:17,440 probably shouldn't be going because it doesn't know any 763 00:38:17,440 --> 00:38:20,360 better and it doesn't have access, especially for embedded 764 00:38:20,360 --> 00:38:22,440 systems. It hasn't been trained on, you 765 00:38:22,440 --> 00:38:26,360 know, good techniques that are honestly proprietary, right? 766 00:38:26,560 --> 00:38:29,200 You know, you don't have medical devices that are safety critical 767 00:38:29,200 --> 00:38:32,240 that are open sourced or, you know, those types of processes 768 00:38:32,240 --> 00:38:35,960 may be fully that I've been released or, you know, whatever 769 00:38:35,960 --> 00:38:37,720 it may be. So, you know, like you said, 770 00:38:37,720 --> 00:38:41,200 Stack Overflow open source projects, which I don't want to, 771 00:38:41,280 --> 00:38:44,080 you know, upset the open source community, but you know, some 772 00:38:44,080 --> 00:38:46,520 projects are kind of it's. All trash code everywhere. 773 00:38:48,440 --> 00:38:50,800 But as I as I've told my employees many times, yeah, 774 00:38:51,040 --> 00:38:52,520 that's all code is trash code that's. 775 00:38:52,560 --> 00:38:58,040 Just how it works, by the way, silence when you know, when in 776 00:38:58,040 --> 00:39:01,120 the, in the last, in the last code training that I gave, we, 777 00:39:01,160 --> 00:39:03,360 you know, we generated the sensor driver, right? 778 00:39:03,600 --> 00:39:05,520 And it worked and it was all fine. 779 00:39:05,520 --> 00:39:08,400 And we even fixed the timing. So it was nice and robust. 780 00:39:09,000 --> 00:39:11,200 And then I said, fine, let's review the code. 781 00:39:11,520 --> 00:39:13,720 What's it look like? And it was sort of, it looked 782 00:39:13,720 --> 00:39:15,600 kind of fine. Like everybody looked at it and 783 00:39:15,600 --> 00:39:19,480 say, yeah, it's, it's, it's OK. And then I OK, fine. 784 00:39:19,480 --> 00:39:22,240 How, how will you, how will you test this? 785 00:39:22,760 --> 00:39:25,520 Like we, we didn't get any tests because we didn't ask for any 786 00:39:25,520 --> 00:39:28,360 tests. But like, how were we to test 787 00:39:28,360 --> 00:39:30,920 this? And they looked at it and said, 788 00:39:30,920 --> 00:39:32,760 well, it's essentially untestable. 789 00:39:32,920 --> 00:39:35,040 Like this is, this is classic video code. 790 00:39:35,040 --> 00:39:38,720 There's there's like a big old main function and that's it. 791 00:39:39,080 --> 00:39:42,560 That's it. Yeah. 792 00:39:44,280 --> 00:39:47,160 So that was really interesting that essentially to your point, 793 00:39:47,160 --> 00:39:49,280 Jacob, right. This is this is the kind of code 794 00:39:49,280 --> 00:39:53,360 you see on the Internet, people who get it going and and it and 795 00:39:53,360 --> 00:39:58,600 it kind of works, but like professional grade code is just 796 00:39:58,600 --> 00:40:01,240 not. I mean, have you have you ever 797 00:40:01,240 --> 00:40:02,720 read the source code for Wikipedia? 798 00:40:02,720 --> 00:40:05,560 I just, I don't want to disagree with you, but like. 799 00:40:07,800 --> 00:40:11,200 Having having been a person who made a copy of the Wikipedia's 800 00:40:11,200 --> 00:40:14,840 horoscope so we could host it locally and then add add stuff 801 00:40:14,840 --> 00:40:18,320 to it, wow. Wow, that's professional grade 802 00:40:18,320 --> 00:40:25,720 right there. They they get whoever they can 803 00:40:25,720 --> 00:40:28,960 get to update. That is the Wikipedia stuff. 804 00:40:29,360 --> 00:40:33,160 And it's holding the world's information like the the whole 805 00:40:33,240 --> 00:40:36,680 history of humankind like and the universe. 806 00:40:37,800 --> 00:40:41,360 Oh, absolutely, yeah. And, and, and anyone who is 807 00:40:41,360 --> 00:40:44,360 interested in using AI in embedded systems, I mean, what 808 00:40:44,440 --> 00:40:46,480 you know, from some of this discussion, one of the things 809 00:40:46,480 --> 00:40:49,920 that I've, I've done personally is, you know, and, and I, I can 810 00:40:49,920 --> 00:40:51,360 cheat a little bit because I'm a consultant. 811 00:40:51,360 --> 00:40:53,240 And so I, I have lots of templates and a lot of 812 00:40:53,240 --> 00:40:56,160 production grade code that I write just for myself or as part 813 00:40:56,160 --> 00:40:57,840 of my training to show people, right? 814 00:40:58,200 --> 00:41:01,880 And so a lot of times I will try to include those examples as 815 00:41:01,880 --> 00:41:04,960 part of the skills or the, the agents that I'm, I'm building or 816 00:41:04,960 --> 00:41:08,840 even some of the MPCMCP servers or I'll build static analysis in 817 00:41:08,840 --> 00:41:11,360 and the things that I find important, like, Hey, I want 818 00:41:11,360 --> 00:41:14,000 functions written so that cyclomatic complexity is, you 819 00:41:14,000 --> 00:41:15,800 know, 10 or less generally, right? 820 00:41:15,800 --> 00:41:19,000 Or 15 or less, you know, pick whatever your own rule of thumb 821 00:41:19,000 --> 00:41:21,000 is there. And I try to make sure that 822 00:41:21,000 --> 00:41:22,000 those types of things are built in. 823 00:41:22,000 --> 00:41:24,200 Now that does that, does that improve the code? 824 00:41:24,320 --> 00:41:25,800 Yes. Does it mean it generates 825 00:41:26,200 --> 00:41:28,040 perfect code the way I would? No. 826 00:41:28,480 --> 00:41:31,040 Many times it still ignores the stuff that I give it. 827 00:41:31,200 --> 00:41:34,600 So in many cases, you know, using your traditional tools, 828 00:41:34,600 --> 00:41:37,480 static analysis tools, writing your own scripts, things like 829 00:41:37,480 --> 00:41:40,800 that, that ends up being a pretty good use case for, you 830 00:41:40,800 --> 00:41:42,280 know, for improving the robustness. 831 00:41:42,880 --> 00:41:45,200 And, and honestly, we didn't talk about this before, So I, I 832 00:41:45,200 --> 00:41:47,640 don't know if I'm going to go off the rails here, but one, one 833 00:41:47,640 --> 00:41:51,360 of the, one of the techniques I use AI for the most, to be 834 00:41:51,360 --> 00:41:54,640 honest, it's not to write the code or generate the code for 835 00:41:54,640 --> 00:41:57,320 me. It's to develop tools that 836 00:41:57,320 --> 00:41:59,200 support my development processes. 837 00:41:59,360 --> 00:42:01,760 And so I don't need a tool that's production grade 838 00:42:01,760 --> 00:42:03,520 necessarily because I'm not releasing it. 839 00:42:03,760 --> 00:42:08,040 But I might, you know, write scripts or support software or, 840 00:42:08,360 --> 00:42:10,360 you know, configurators or things like that, that help 841 00:42:10,360 --> 00:42:13,000 support my normal workflow of developing systems. 842 00:42:13,000 --> 00:42:16,200 And so a lot of times with what I'm working on, I'm not I, you 843 00:42:16,200 --> 00:42:18,440 know, I'm for like a customer, I'm not going to generate AI 844 00:42:18,440 --> 00:42:20,520 code for that. Maybe my own personal projects, 845 00:42:20,520 --> 00:42:22,720 I'll use AI to generate code to understand it. 846 00:42:22,720 --> 00:42:25,480 But you know, for production systems, yeah, I am using AI to 847 00:42:25,480 --> 00:42:27,680 generate tools that I'm using to accelerate the the way I develop 848 00:42:27,680 --> 00:42:29,600 systems. And so I thought I'd just share 849 00:42:29,600 --> 00:42:31,360 that with you guys in case. Well and probably our 850 00:42:31,360 --> 00:42:33,840 conversations, people are scaried or shying away from AI. 851 00:42:33,840 --> 00:42:36,760 Well, there are other use cases, you know, as part of our 852 00:42:36,760 --> 00:42:39,720 workflow to leverage these. Tools This is exactly where I I 853 00:42:39,720 --> 00:42:42,600 was thinking about going next is like, OK, so now we've had our 854 00:42:42,600 --> 00:42:45,440 little gals humor and like it was funny and it didn't work and 855 00:42:45,480 --> 00:42:48,000 you know, and we're struggling to get it and it's like, oh, 856 00:42:48,000 --> 00:42:48,840 we're generating the wrong thing. 857 00:42:48,840 --> 00:42:52,680 It's not right. And it's like, so how do you go 858 00:42:52,680 --> 00:42:56,320 to someone who's skeptical and saying like, hey, AI is not for 859 00:42:56,320 --> 00:42:58,360 embedded systems development does has no place here. 860 00:42:58,360 --> 00:43:00,760 Get this out of my face. Like I'm a grown up. 861 00:43:01,200 --> 00:43:03,240 I'm I know what I'm doing. I don't need this. 862 00:43:03,480 --> 00:43:06,800 Like how do you go based on your experience and based on your 863 00:43:06,800 --> 00:43:08,680 training and everything, how do you go to somebody and say like, 864 00:43:08,680 --> 00:43:15,680 OK, yeah, I hear you, but. I mean, you, you, I honestly 865 00:43:15,680 --> 00:43:18,160 think you have to have examples in the data to back it up. 866 00:43:18,240 --> 00:43:20,720 You know, one of my philosophies with the bad systems is data 867 00:43:20,720 --> 00:43:23,200 dictates design for the architecture always gets me in 868 00:43:23,200 --> 00:43:25,120 trouble because embedded engineers don't like that 869 00:43:25,160 --> 00:43:28,680 philosophy. But it's one that has worked 870 00:43:28,680 --> 00:43:31,560 extraordinarily well for me to architect systems that are 871 00:43:31,560 --> 00:43:34,560 scalable. And so I think when people come, 872 00:43:34,680 --> 00:43:36,360 you know, kind of say, well, AI doesn't belong here. 873 00:43:36,360 --> 00:43:39,720 You have to give the examples of what you've used it for and how 874 00:43:39,720 --> 00:43:42,800 it is accelerated development. I have a colleague on LinkedIn 875 00:43:42,800 --> 00:43:45,200 who I've been doing some AI posts and. 876 00:43:45,200 --> 00:43:47,640 And he, he, he said he was disappointed in me and that I'm 877 00:43:47,640 --> 00:43:51,040 going downhill. I was talking about it. 878 00:43:51,040 --> 00:43:52,920 Unsubscribe. Yeah. 879 00:43:53,320 --> 00:43:54,800 And it was like, I was like, oh, that hurts. 880 00:43:54,800 --> 00:43:56,480 But then it was kind of like the push back up. 881 00:43:56,480 --> 00:43:59,520 Well, if I'm not exploring these tools, I'm going to become 882 00:43:59,520 --> 00:44:03,520 irrelevant in a year or two because at some point this is, 883 00:44:03,520 --> 00:44:06,600 you know, these tools are, you know, we're in the embedded 884 00:44:06,600 --> 00:44:08,600 system space, we are in a little bubble, right? 885 00:44:08,880 --> 00:44:11,280 We kind of believe we're special compared to the rest of the 886 00:44:11,280 --> 00:44:13,880 software engineering world because we talked to hardware, 887 00:44:14,240 --> 00:44:17,560 we work with hardware, but it's been proven over and over again 888 00:44:17,560 --> 00:44:20,200 that whatever's happening in the rest of the, you know, the rest 889 00:44:20,200 --> 00:44:22,280 of the software industry, that comes to us in within three to 890 00:44:22,280 --> 00:44:24,840 five years. And so these tools will be used 891 00:44:24,840 --> 00:44:27,080 on a daily basis to develop embedded systems. 892 00:44:27,640 --> 00:44:31,120 I, I've seen people tying the entire workflow together where 893 00:44:31,120 --> 00:44:34,280 they can do a full debug process on chip using AI, you know, 894 00:44:34,280 --> 00:44:38,000 accessing GDB and, you know, pulling serial data off of 895 00:44:38,000 --> 00:44:39,720 systems and stuff like that. I've played around with some of 896 00:44:39,720 --> 00:44:41,960 that. So much smarter people are going 897 00:44:41,960 --> 00:44:44,280 to build tools that make this stuff, you know, a little bit 898 00:44:44,280 --> 00:44:47,240 more automated. But I, so I'm digressing again. 899 00:44:47,240 --> 00:44:51,160 But ultimately you have to show examples of how you're using it 900 00:44:51,720 --> 00:44:54,120 to make a difference in the way that you're developing systems. 901 00:44:54,120 --> 00:44:56,720 And I think a lot of people go back to, oh, you're generating 902 00:44:56,720 --> 00:44:59,200 code automatically with AI. And I don't necessarily know 903 00:44:59,200 --> 00:45:02,880 that that's, you know, the full direction or the full agentic 904 00:45:02,880 --> 00:45:04,360 workflow where you just like do everything. 905 00:45:04,360 --> 00:45:07,080 That's not that's not the right use case for embedded systems. 906 00:45:07,920 --> 00:45:10,000 At least today. I'll be quiet now. 907 00:45:11,960 --> 00:45:17,280 I, I wonder though, like I, I, I think more more open to using AI 908 00:45:17,640 --> 00:45:19,120 much more strongly also in embedded systems. 909 00:45:19,400 --> 00:45:22,640 I mean, yes, you need to be careful about it because it will 910 00:45:22,760 --> 00:45:27,120 fall into bad habits of, you know, developers who don't 911 00:45:27,400 --> 00:45:29,000 really think in terms of embedded systems. 912 00:45:29,000 --> 00:45:32,840 They, you know, I don't know, memory allocation, real time 913 00:45:32,840 --> 00:45:36,640 issues or whatever, but still it is, you know, it is a force 914 00:45:36,640 --> 00:45:40,280 multiplier. It will still make me better and 915 00:45:40,360 --> 00:45:44,320 I still absolutely use it to generate code after I've written 916 00:45:44,320 --> 00:45:48,400 a bunch of tests etcetera. But I think people who are 917 00:45:48,400 --> 00:45:51,200 sceptical towards AI come from exactly that direction where 918 00:45:51,200 --> 00:45:53,520 they think, well, it's just a fancy code completion tool 919 00:45:53,920 --> 00:45:58,280 because also this is what ID ES tend to make it look like. 920 00:45:59,000 --> 00:46:03,920 I I'm increasingly convinced that IDE plug insurance for AI 921 00:46:03,920 --> 00:46:07,800 are a bad idea because they they bring you to the level of code 922 00:46:07,800 --> 00:46:09,720 where you don't want to be. You want to stay on a more 923 00:46:09,720 --> 00:46:12,960 conceptual level. So you shouldn't be in a code 924 00:46:12,960 --> 00:46:15,720 editor type application for that. 925 00:46:16,800 --> 00:46:19,040 And so what's? I just see you like you open 926 00:46:19,040 --> 00:46:20,400 Excel and that's what you're using. 927 00:46:20,400 --> 00:46:23,680 You're using Excel to write the code and your AI is in Excel and 928 00:46:23,680 --> 00:46:25,240 then it generates. It was like, no, no, no, no, 929 00:46:25,240 --> 00:46:27,960 we're not we're not doing, we're not doing code anymore. 930 00:46:28,880 --> 00:46:32,000 I. Literally open up a terminal and 931 00:46:32,040 --> 00:46:34,640 you know I might open it in VS Code or Sublime Text or 932 00:46:34,640 --> 00:46:36,760 something if I want to review files, but I literally work out 933 00:46:36,760 --> 00:46:38,960 of a terminal for all my. Oh yeah, all my work. 934 00:46:39,200 --> 00:46:41,640 So exactly. And I do not want that to be 935 00:46:41,640 --> 00:46:45,800 like the same window because it's mixing two different levels 936 00:46:45,800 --> 00:46:48,800 of of thinking about the problem for me. 937 00:46:48,800 --> 00:46:54,560 And yeah, so but people who reject AI very frequently have 938 00:46:54,560 --> 00:46:58,520 just seen sort of this bad example of well, it's, it's a 939 00:46:58,520 --> 00:47:01,440 code completion tool and they say, well, you know, as you said 940 00:47:01,440 --> 00:47:02,760 before, typing is not the hard part. 941 00:47:02,760 --> 00:47:06,160 So let me let me just continue to type the line by myself. 942 00:47:06,560 --> 00:47:11,080 Yeah, but if you use it removed from that and you really use it 943 00:47:11,120 --> 00:47:15,080 to work on the code and in the code and generate entire like 944 00:47:15,160 --> 00:47:18,400 files for you, well, it's still firmly under your control. 945 00:47:19,320 --> 00:47:21,840 Then I think they start to see, OK, fine. 946 00:47:21,840 --> 00:47:24,920 This actually is materially different from how I've been 947 00:47:24,920 --> 00:47:28,320 using it before or how I've been writing software before. 948 00:47:28,760 --> 00:47:34,000 And I can I can see the point that that tends to be what what 949 00:47:34,000 --> 00:47:36,160 has happened in practice today. They say, well, code completion 950 00:47:36,160 --> 00:47:38,000 is not interesting. Well then let's not do code 951 00:47:38,000 --> 00:47:40,800 completion. And one of the things I would, I 952 00:47:40,800 --> 00:47:43,160 would actually encourage folks, if you are skeptical of using 953 00:47:43,200 --> 00:47:47,000 AI, keep the skepticalness because as you start to use the 954 00:47:47,000 --> 00:47:49,440 tool, you want to be skeptical of what it's generating, how 955 00:47:49,440 --> 00:47:51,920 it's generating it, the process it's following, and you want to 956 00:47:51,920 --> 00:47:54,760 push back really hard on it. So we don't want the 957 00:47:54,760 --> 00:47:57,000 skepticalness to go away and end up with a bunch of people who 958 00:47:57,000 --> 00:47:59,040 are just, you know, loving whatever AI outputs. 959 00:47:59,040 --> 00:48:01,320 Oh my God. That engineering head first into 960 00:48:01,320 --> 00:48:03,560 a pool that the AI built for you, and then you didn't 961 00:48:03,560 --> 00:48:06,360 actually tell exactly what to make the material out of the 962 00:48:06,360 --> 00:48:07,640 bottom and how deep you want it to be. 963 00:48:07,640 --> 00:48:10,720 Or you did tell it how deep you wanted it to be, but it decided 964 00:48:10,720 --> 00:48:12,320 that you were using inches instead of feet. 965 00:48:12,320 --> 00:48:19,200 Like so do not dive head first into a pool made by AI. 966 00:48:19,520 --> 00:48:23,560 Examine it before you do. Keep your skepticism and keep 967 00:48:23,560 --> 00:48:26,040 your engineering brain. I think there's, there's a lot 968 00:48:26,040 --> 00:48:28,320 of of concern out there that we're going to lose our 969 00:48:28,320 --> 00:48:30,840 engineering brains. It's like, no, you definitely 970 00:48:30,840 --> 00:48:32,360 need to keep using your engineering brain. 971 00:48:32,360 --> 00:48:36,440 Like that's not going away. And it and it can't enhance it. 972 00:48:36,520 --> 00:48:39,000 I mean, think about using AI for code reviews. 973 00:48:39,560 --> 00:48:42,240 I mean, it not in a way where you just kind of let it go all 974 00:48:42,240 --> 00:48:43,720 on its own. Like I see some of the stuff 975 00:48:43,720 --> 00:48:46,360 some of the get tools doing, but you know, you sit down and you 976 00:48:46,360 --> 00:48:49,520 ask it to review something and it says, oh, we found this. 977 00:48:49,520 --> 00:48:52,280 Or, you know, have it explain something in the code that you 978 00:48:52,280 --> 00:48:54,360 maybe don't understand what on earth it's doing. 979 00:48:54,800 --> 00:48:58,480 I have found that those types of analysis do kind of produce 980 00:48:58,720 --> 00:49:02,720 benefits that help you learn and and become a better coder. 981 00:49:03,240 --> 00:49:06,520 Or a better program, just find stuff that you missed. 982 00:49:07,120 --> 00:49:08,480 Yeah, exactly. I mean, we're human. 983 00:49:08,600 --> 00:49:12,680 Be greater equal and you put greater greater than yeah, 984 00:49:13,160 --> 00:49:15,480 happens to all of us and it's posted it fantastic you. 985 00:49:15,640 --> 00:49:18,200 Used a single equal sign instead of a double equal sign. 986 00:49:18,840 --> 00:49:21,480 I, I did it this week. This week I did it and I was, I 987 00:49:21,480 --> 00:49:24,040 was kicking myself. I was like, Oh my God, you know, 988 00:49:24,200 --> 00:49:27,080 like how did I get that conditional wrong? 989 00:49:27,120 --> 00:49:30,480 Well, we're, we're human, right? We, we, that's why pair 990 00:49:30,480 --> 00:49:31,720 programming was such a good idea. 991 00:49:31,720 --> 00:49:34,280 You put two people together and you know one can review the 992 00:49:34,280 --> 00:49:36,120 other work, you know other person's work as you're coding. 993 00:49:36,120 --> 00:49:38,280 That's kind of what a interesting role could be, 994 00:49:38,480 --> 00:49:39,040 right? Yeah. 995 00:49:39,200 --> 00:49:40,560 Oh, yes. I don't have to share my yeah, I 996 00:49:40,560 --> 00:49:42,360 don't have to share my keyboard with the AI. 997 00:49:42,480 --> 00:49:46,640 You're absolutely right. That, that is so interesting 998 00:49:46,640 --> 00:49:49,600 that you say that because that is exactly what I teach people 999 00:49:49,600 --> 00:49:52,440 in my, in my classes that, you know, essentially you're doing 1000 00:49:52,440 --> 00:49:55,040 pair programming, but in traditional pair programming, 1001 00:49:55,040 --> 00:49:57,200 you, you switch roles every now and then, you know, but the 1002 00:49:57,200 --> 00:49:59,960 frankly, one person drives and the other person reads them up. 1003 00:50:00,320 --> 00:50:02,840 And with AI, you're always the person who reads them up. 1004 00:50:03,600 --> 00:50:06,240 But other than that, this is exactly the kind of interaction 1005 00:50:06,240 --> 00:50:09,720 you have, the kind of workshop you have and, and you don't get 1006 00:50:09,720 --> 00:50:12,120 to step away and. Drink coffee. 1007 00:50:13,600 --> 00:50:15,960 Exactly. And honestly, you could, if you 1008 00:50:15,960 --> 00:50:18,480 wanted to, you could flip roles where you write code and then 1009 00:50:18,480 --> 00:50:20,920 you ask AI, hey, give me some feedback on this. 1010 00:50:20,920 --> 00:50:22,160 Is this what we were trying to accomplish? 1011 00:50:23,360 --> 00:50:24,920 Obviously you're going to be slower at writing the. 1012 00:50:25,080 --> 00:50:26,880 I hate to admit, yeah, I'd be slower at writing the code 1013 00:50:26,880 --> 00:50:29,280 probably than Claude. But if I have a really bad 1014 00:50:29,280 --> 00:50:32,400 Internet connection, then I think I can. 1015 00:50:32,440 --> 00:50:34,920 I think I can outdo. It and I'm fuck using code llama 1016 00:50:34,920 --> 00:50:36,640 like. Yeah. 1017 00:50:39,360 --> 00:50:41,760 You think about it from ghosts like they're they're a pair of 1018 00:50:41,760 --> 00:50:45,520 programming, but making a piece of pottery together and like the 1019 00:50:45,680 --> 00:50:56,640 AI is is Patrick Swayze. I'm kind of struggling with the 1020 00:50:56,640 --> 00:50:59,960 metal image but. Yeah, yeah, yeah, I I could see 1021 00:50:59,960 --> 00:51:01,680 that based on the look on your face. 1022 00:51:01,720 --> 00:51:02,760 I'll. Explain to my thought. 1023 00:51:02,760 --> 00:51:09,440 We're not of time. All right, Jacob Plug, give 1024 00:51:09,440 --> 00:51:12,080 yourself some time to plug. You know, what is it you do? 1025 00:51:12,080 --> 00:51:14,200 And well, we've talked about what it is you do, but like, you 1026 00:51:14,200 --> 00:51:16,080 know, how can people get a hold of you and, and what can they 1027 00:51:16,080 --> 00:51:17,560 get from you when they talk to you? 1028 00:51:18,800 --> 00:51:21,000 Oh, sure. So you can find me. 1029 00:51:21,000 --> 00:51:23,600 You can find me on LinkedIn. You know, I think just A tag 1030 00:51:23,600 --> 00:51:26,040 Jacob Beningo or searching that will probably find me. 1031 00:51:26,040 --> 00:51:28,960 I've also got beningo.com. I'm going to be clear, you're, 1032 00:51:28,960 --> 00:51:30,800 you're, you're pretty conservative on LinkedIn. 1033 00:51:30,800 --> 00:51:32,560 You only post every business day. 1034 00:51:33,440 --> 00:51:38,880 Yes, only every business day. Yes, only business every 1035 00:51:38,880 --> 00:51:42,280 business day. I try to provide some, I always 1036 00:51:42,280 --> 00:51:44,760 say tips or tricks, but really more and more and more it's 1037 00:51:44,760 --> 00:51:47,240 about just kind of trying to make you rethink how you develop 1038 00:51:47,240 --> 00:51:49,960 embedded systems. And a lot of the work I do is 1039 00:51:49,960 --> 00:51:53,640 about trying to help teams modernize their, their systems, 1040 00:51:53,640 --> 00:51:55,720 their processes. And a lot of that comes around 1041 00:51:55,720 --> 00:51:57,440 to like software architecture development. 1042 00:51:57,440 --> 00:51:59,400 I do a lot of, a lot of training. 1043 00:51:59,400 --> 00:52:02,040 I've got my embedded software Academy and obviously I'm, you 1044 00:52:02,040 --> 00:52:03,440 know, we've mentioned I do some AI training. 1045 00:52:03,440 --> 00:52:05,520 I've got an embedded AI course here that's coming up at the end 1046 00:52:05,520 --> 00:52:08,040 of April as well. So, but you can e-mail me at 1047 00:52:08,040 --> 00:52:11,200 jacob@beningo.com if you've got questions or anything like that. 1048 00:52:11,200 --> 00:52:13,400 But you know, feel free to reach out on LinkedIn as well. 1049 00:52:13,400 --> 00:52:16,320 I'm, I'm pretty active on it. And yeah, you'll probably see 1050 00:52:16,360 --> 00:52:19,280 definitely a few contrarian ideas on there compared to how 1051 00:52:19,280 --> 00:52:21,000 we traditionally develop embedded systems. 1052 00:52:21,000 --> 00:52:24,960 So, and they're designed to make you think differently. 1053 00:52:26,080 --> 00:52:29,680 They're not necessarily. You know, designed to make you 1054 00:52:29,680 --> 00:52:31,680 happy or light check. They're not designed to make you 1055 00:52:31,680 --> 00:52:33,920 happy and cheery necessarily, just make you stop and say, hmm, 1056 00:52:34,440 --> 00:52:35,640 let me think about this for a little bit. 1057 00:52:35,640 --> 00:52:36,600 That's a different way of thinking. 1058 00:52:36,600 --> 00:52:38,240 About it. Agree with you blindly, Sir. 1059 00:52:38,840 --> 00:52:44,800 Yes, exactly so, but all in good fun and hopefully to improve the 1060 00:52:44,800 --> 00:52:46,080 way that we develop embedded systems. 1061 00:52:46,200 --> 00:52:48,080 Awesome. All right. 1062 00:52:48,160 --> 00:52:50,120 Well, thanks for chatting. It's been absolutely lovely. 1063 00:52:50,920 --> 00:52:54,880 Yeah, I am and, and thank you listeners, for joining us for 1064 00:52:54,880 --> 00:52:57,000 this episode. Hope you enjoyed it. 1065 00:52:57,600 --> 00:52:59,800 Yeah, If you leave a comment down below, I don't know how 1066 00:52:59,800 --> 00:53:01,400 does that work in podcasts, whatever. 1067 00:53:01,840 --> 00:53:04,200 Reach out to us, check out our website. 1068 00:53:04,200 --> 00:53:05,960 We've got some stuff. We're also on LinkedIn. 1069 00:53:06,360 --> 00:53:10,760 We also troll around in the Agile Embedded podcast Slack 1070 00:53:10,760 --> 00:53:12,000 channel, if you want to join that too. 1071 00:53:12,000 --> 00:53:14,560 That's where we're at too. So yeah. 1072 00:53:16,000 --> 00:53:17,040 Thanks. Thanks for the invitation. 1073 00:53:17,040 --> 00:53:18,600 And do you mind if I add one more thing in? 1074 00:53:18,680 --> 00:53:21,000 Of course I do mind. Wait, No, no, no, sorry, sorry. 1075 00:53:21,000 --> 00:53:25,320 I don't mind. Sometimes people might want to 1076 00:53:25,320 --> 00:53:27,080 check out too is the embedded online conference. 1077 00:53:27,240 --> 00:53:29,760 That's something where we get together every every year. 1078 00:53:29,760 --> 00:53:32,560 That's a very good point. In May, all of us are speaking 1079 00:53:32,560 --> 00:53:36,680 at it and it's all about educating the embedded systems 1080 00:53:36,680 --> 00:53:38,920 industry and engineers and getting together as a community 1081 00:53:38,920 --> 00:53:42,320 and sharing ideas and learning. So I, I figured that's something 1082 00:53:42,320 --> 00:53:45,760 that you guys as an audience and you know, 2 speakers of the, the 1083 00:53:45,760 --> 00:53:47,800 conference might be, you know, interested in sharing with you 1084 00:53:47,800 --> 00:53:50,160 guys, so. Yes, thanks for the reminder. 1085 00:53:50,800 --> 00:53:52,960 Yeah, so. OK. 1086 00:53:53,320 --> 00:53:56,080 Well, this has been the Embedded AI podcast. 1087 00:53:56,360 --> 00:53:58,680 I am and will continue to be Ryan Torvik. 1088 00:53:59,720 --> 00:54:02,280 And I'm Luca Jonni. Catch you guys next time, see. 1089 00:54:03,000 --> 00:54:06,120 You. Hi, Luca here. 1090 00:54:06,680 --> 00:54:09,520 I have a quick announcement. I've just launched one more new 1091 00:54:09,520 --> 00:54:13,520 training platform, the Embedded AI Academy at Embedded AI dot 1092 00:54:13,520 --> 00:54:15,560 Academy. It feels like up I've seen in 1093 00:54:15,560 --> 00:54:18,880 the market training that approaches AI specifically from 1094 00:54:18,880 --> 00:54:21,200 an embedded systems angle. If you're listening to this 1095 00:54:21,200 --> 00:54:24,960 podcast, it should be right up your alley and you get 25% off 1096 00:54:24,960 --> 00:54:28,280 any booking made through the end of 2025 if you use the code 1097 00:54:28,640 --> 00:54:32,520 Embedded AI Podcast 25 all in capital letters. 1098 00:54:32,920 --> 00:54:36,560 Again, the website is at Embedded AI dot Academy.