1 00:00:00,040 --> 00:00:02,200 Welcome to the deep dive. We are. 2 00:00:02,480 --> 00:00:05,000 We're really thrilled to have you here with us today because 3 00:00:05,280 --> 00:00:07,840 the mission ahead is just genuinely mind bending. 4 00:00:07,840 --> 00:00:09,960 It really is. I'm glad to be here to get into 5 00:00:09,960 --> 00:00:10,560 it. Yeah. 6 00:00:10,560 --> 00:00:13,000 And for those joining us, I'm your host, and I'm sitting here 7 00:00:13,000 --> 00:00:16,040 with our resident expert to look at this curated stack of 8 00:00:16,040 --> 00:00:18,160 insights. Right developer observations, 9 00:00:18,160 --> 00:00:20,240 Product updates. Exactly. 10 00:00:20,600 --> 00:00:24,080 And all of this is from early March 2026. 11 00:00:24,680 --> 00:00:27,760 And it all points to 1 unavoidable conclusion, the. 12 00:00:27,760 --> 00:00:31,000 Fundamental nature of building software is shifting right 13 00:00:31,000 --> 00:00:33,160 beneath our feet. It really is O. 14 00:00:33,160 --> 00:00:36,520 The goal for this deep dive is to figure out exactly what it 15 00:00:36,520 --> 00:00:39,920 means for you to transition from the traditional idea of being a 16 00:00:39,920 --> 00:00:43,880 typist who writes code to a completely new paradigm. 17 00:00:43,880 --> 00:00:46,520 Becoming an orchestrator of AI agents. 18 00:00:46,560 --> 00:00:47,600 Right. An orchestrator. 19 00:00:47,600 --> 00:00:49,040 Yeah. If you've ever wondered what the 20 00:00:49,040 --> 00:00:51,480 future of knowledge work actually looks like in practice, 21 00:00:51,720 --> 00:00:53,200 you were in the exact right place. 22 00:00:53,200 --> 00:00:54,960 Absolutely. OK, let's unpack this. 23 00:00:54,960 --> 00:00:58,240 The very first bombshell jumping out of this material comes from 24 00:00:58,240 --> 00:01:00,840 a 30 year software veteran. 30 years. 25 00:01:01,160 --> 00:01:03,560 Just pause and think about that timeline for a second. 26 00:01:04,239 --> 00:01:08,720 That means this person was hand typing code before the Internet 27 00:01:08,720 --> 00:01:11,160 was a household thing. They were there for all of it. 28 00:01:11,240 --> 00:01:13,040 Right. They have successfully built 29 00:01:13,040 --> 00:01:17,880 three entire companies, and yet for the last six months they 30 00:01:17,880 --> 00:01:20,520 haven't written a single line of code. 31 00:01:20,840 --> 00:01:22,520 Not one line. Not one. 32 00:01:22,760 --> 00:01:25,640 And the wildest part to me, they say they don't miss it at all. 33 00:01:25,840 --> 00:01:30,160 It really is a striking anecdote, and I think it serves 34 00:01:30,160 --> 00:01:32,840 as the perfect anchor for this entire shift we're exploring 35 00:01:32,840 --> 00:01:34,120 today. Yeah, what this veteran 36 00:01:34,120 --> 00:01:36,800 developer is doing instead is something they describe as 37 00:01:36,800 --> 00:01:40,320 managing 6 to 10 occasionally hallucinating chat bots. 38 00:01:40,400 --> 00:01:42,040 That is quite the job description. 39 00:01:42,040 --> 00:01:44,440 Right, that is their new daily reality. 40 00:01:44,600 --> 00:01:47,440 Instead of opening a blank file and typing out syntax, they're 41 00:01:47,440 --> 00:01:49,760 pointing specialized AI agents at a problem. 42 00:01:49,760 --> 00:01:52,520 Watching those agents generate entire architectures. 43 00:01:52,520 --> 00:01:55,120 Exactly. And then stepping in to review 44 00:01:55,200 --> 00:01:58,080 guide and course correct. And this developer claims their 45 00:01:58,080 --> 00:02:00,960 productivity has just absolutely skyrocketed. 46 00:02:00,960 --> 00:02:03,560 Yeah, I can imagine. But how does that actually work 47 00:02:03,560 --> 00:02:06,760 in a historical context? Because it feels like, I don't 48 00:02:06,760 --> 00:02:09,120 know, it feels like we're losing a fundamental craft. 49 00:02:09,160 --> 00:02:11,840 Well, we've seen this exact pattern before, just in 50 00:02:11,840 --> 00:02:13,440 different domains. Like what? 51 00:02:13,720 --> 00:02:18,000 Think about the transition from writing low level assembly code. 52 00:02:18,000 --> 00:02:20,920 Where you were manually managing computer memory. 53 00:02:20,920 --> 00:02:25,280 Yes, exactly. Moving from that to using a high 54 00:02:25,280 --> 00:02:28,440 level programming languages where the computer handles the 55 00:02:28,440 --> 00:02:30,160 memory for you. Right, OK. 56 00:02:30,360 --> 00:02:33,960 Or perhaps a more grounded analogy for you, the shift from 57 00:02:33,960 --> 00:02:37,280 using manual hand saws to using electric power tools. 58 00:02:37,360 --> 00:02:39,880 Oh, I like. That the individual granular 59 00:02:39,880 --> 00:02:43,400 actions of the human become less physical and less tedious. 60 00:02:43,520 --> 00:02:47,160 But the sheer volume of what one person can build scales up 61 00:02:47,160 --> 00:02:48,680 dramatically. Exactly. 62 00:02:48,680 --> 00:02:50,720 Right, so you aren't sweating over the saw anymore, you're 63 00:02:50,720 --> 00:02:53,080 just guiding the blade? You're guiding the blade, yes, 64 00:02:53,120 --> 00:02:55,560 but we do have to be careful with that analogy because there 65 00:02:55,560 --> 00:02:59,480 is a massive new variable here which is a circular saw. 66 00:02:59,480 --> 00:03:01,520 Doesn't try to guess your intentions. 67 00:03:02,160 --> 00:03:04,480 AI agents, at least the ones we are looking at right now in 68 00:03:04,480 --> 00:03:07,480 these notes, they don't independently learn from their 69 00:03:07,480 --> 00:03:10,080 past mistakes without human intervention. 70 00:03:10,880 --> 00:03:14,680 They don't autonomously improve their underlying architecture on 71 00:03:14,680 --> 00:03:17,760 the fly. And, crucially, they don't push 72 00:03:17,760 --> 00:03:22,040 back with alternative better approaches the way a human 73 00:03:22,080 --> 00:03:25,240 junior engineer might. They just confidently execute. 74 00:03:25,320 --> 00:03:27,040 Exactly. They confidently execute 75 00:03:27,040 --> 00:03:28,440 whatever they think you want. Which? 76 00:03:28,440 --> 00:03:30,520 Brings up a really important friction point. 77 00:03:30,640 --> 00:03:34,160 If they just blindly execute, does that mean the traditional 78 00:03:34,160 --> 00:03:36,760 software developer is entirely obsolete? 79 00:03:36,880 --> 00:03:39,560 That's a big question. Or are we looking at a split in 80 00:03:39,560 --> 00:03:41,240 the industry? You're hitting on the exact 81 00:03:41,240 --> 00:03:43,560 friction point here. It's not a complete replacement, 82 00:03:43,560 --> 00:03:45,800 it's a bifurcation, a split, right? 83 00:03:45,920 --> 00:03:49,400 What we are seeing is that AI is incredibly capable if you need 84 00:03:49,400 --> 00:03:52,800 to churn out standard enterprise Crud's apps. 85 00:03:52,960 --> 00:03:55,120 And for those listening who might not live and breathe 86 00:03:55,120 --> 00:03:59,920 database architecture, CRUD stands for Create, Read, Update 87 00:03:59,920 --> 00:04:01,480 and Delete. The bread and butter. 88 00:04:01,640 --> 00:04:03,600 Exactly. It's essentially the basic 89 00:04:03,600 --> 00:04:06,160 framework of almost every standard business application. 90 00:04:06,400 --> 00:04:08,000 Like a customer management system. 91 00:04:08,000 --> 00:04:10,920 Yeah, where you add a user, read their profile, update their 92 00:04:10,920 --> 00:04:14,320 e-mail or delete their account. When the patterns are well 93 00:04:14,320 --> 00:04:17,959 established like that, or when maintaining legacy systems, the 94 00:04:18,160 --> 00:04:21,079 AI is a powerhouse. It's incredibly fast. 95 00:04:21,079 --> 00:04:24,840 But when you are dealing with novel problems, designing deep 96 00:04:24,840 --> 00:04:29,080 complex system architectures or conducting intense security 97 00:04:29,080 --> 00:04:31,240 analysis, where the requirements are murky. 98 00:04:31,240 --> 00:04:33,160 The AI stumbles. It stumbles hard. 99 00:04:33,160 --> 00:04:35,600 It confidently sprints in the wrong direction. 100 00:04:35,600 --> 00:04:38,320 So the bifurcation in the industry will look like what 101 00:04:38,400 --> 00:04:41,080 exactly? On one side you will have highly 102 00:04:41,080 --> 00:04:44,800 specialized engineers who still handwrite code for those hyper 103 00:04:44,800 --> 00:04:47,400 complex novel systems. And the other side. 104 00:04:47,400 --> 00:04:51,720 The vast majority will become AI, or people whose primary 105 00:04:51,720 --> 00:04:54,440 skill is managing and reviewing AI outputs. 106 00:04:54,440 --> 00:04:57,080 Which is very analogous to how senior engineers operate today 107 00:04:57,080 --> 00:04:57,880 anyway. Right. 108 00:04:58,000 --> 00:05:00,400 They spend far more time reviewing other people's pull 109 00:05:00,400 --> 00:05:02,960 requests than typing out the syntax themselves. 110 00:05:03,240 --> 00:05:06,840 The AI is just taking on the role of a very fast, slightly 111 00:05:06,840 --> 00:05:10,440 unpredictable junior developer. Speaking of fast and 112 00:05:10,440 --> 00:05:14,440 unpredictable, it's wild looking at the timeline of events in 113 00:05:14,440 --> 00:05:17,000 these notes. It really is moving at breakneck 114 00:05:17,000 --> 00:05:19,200 speed. Because Anthropic just announced 115 00:05:19,200 --> 00:05:22,480 a massive tooling update that fundamentally changes this 116 00:05:22,480 --> 00:05:24,040 dynamic. The auto mode. 117 00:05:24,240 --> 00:05:27,960 Yes, they are dropping a research preview for auto mode 118 00:05:28,200 --> 00:05:32,800 in clawed code, rolling out no earlier than March 12, 2026. 119 00:05:32,800 --> 00:05:35,760 A huge shift. To give you some context on why 120 00:05:35,760 --> 00:05:38,960 this is such a big deal, you have to understand the current 121 00:05:38,960 --> 00:05:41,440 pain point of being an AI orchestrator. 122 00:05:41,520 --> 00:05:43,280 It is incredibly tedious right now. 123 00:05:43,760 --> 00:05:47,840 Right now, if you tell an AI to build a feature, it has to stop 124 00:05:47,840 --> 00:05:49,880 and ask for your permission to do literally anything. 125 00:05:49,880 --> 00:05:52,080 It asks to read a file. You click approach. 126 00:05:52,080 --> 00:05:53,160 Yeah. It asks to write a. 127 00:05:53,160 --> 00:05:56,160 File, you click approve, it has to run a test, you click 128 00:05:56,160 --> 00:05:57,200 approve. Over and over. 129 00:05:57,200 --> 00:06:00,000 Imagine trying to get into that magical flow state where you are 130 00:06:00,000 --> 00:06:04,120 zoned into a problem and every two minutes a little prompt pops 131 00:06:04,120 --> 00:06:06,520 up asking for your blessing it. Totally shatters your 132 00:06:06,520 --> 00:06:08,400 concentration. And makes long running 133 00:06:08,520 --> 00:06:10,920 autonomous tasks completely impossible. 134 00:06:11,120 --> 00:06:13,880 What's fascinating here is the underlying tension between 135 00:06:13,880 --> 00:06:17,280 safety and productivity that this update exposes. 136 00:06:17,720 --> 00:06:20,280 How so? Well, Auto mode allows Claude to 137 00:06:20,280 --> 00:06:23,160 just proceed with all these routine operations without 138 00:06:23,160 --> 00:06:26,640 pausing to ask for a human's permission every single step of 139 00:06:26,640 --> 00:06:28,880 the way. So you are trading granular 140 00:06:28,880 --> 00:06:31,080 manual control for immense speed? 141 00:06:31,080 --> 00:06:33,000 Exactly. Hold on, I'm struggling to see 142 00:06:33,000 --> 00:06:35,040 how that's a good idea from a safety perspective. 143 00:06:35,640 --> 00:06:38,520 Taking your hands completely off the steering wheel and letting 144 00:06:38,520 --> 00:06:41,640 an AI read and write files autonomously sounds like a 145 00:06:41,640 --> 00:06:45,040 recipe for a deleted hard drive. At first glance, it absolutely 146 00:06:45,040 --> 00:06:48,680 sounds riskier, right? But if you look closely at human 147 00:06:48,680 --> 00:06:52,560 behavior and the psychology of a light fatigue, the current 148 00:06:52,560 --> 00:06:54,960 system isn't actually as safe as it appears. 149 00:06:55,440 --> 00:06:56,760 Because people get tired of clicking. 150 00:06:57,240 --> 00:07:00,800 Yes, developers are experiencing intense fatigue from these 151 00:07:00,800 --> 00:07:04,120 constant pop up prompts. After the 50th prompt in an 152 00:07:04,120 --> 00:07:07,120 hour, they aren't reading the code the AI wants to execute 153 00:07:07,120 --> 00:07:08,800 anymore. They're just blindly clicking 154 00:07:08,800 --> 00:07:11,960 yes, approve, proceed just to keep the momentum going. 155 00:07:12,680 --> 00:07:16,240 So it's the illusion of safety. You feel like you're in control 156 00:07:16,240 --> 00:07:18,880 because you clicked a button, but you have no idea what you 157 00:07:18,880 --> 00:07:22,760 actually just authorized. Precisely so counterintuitively, 158 00:07:22,800 --> 00:07:26,160 explicit auto mode might actually be safer in practice. 159 00:07:26,360 --> 00:07:28,440 Provided it has the proper guardrails in place. 160 00:07:28,480 --> 00:07:30,360 Right. If the system knows it's running 161 00:07:30,360 --> 00:07:33,160 autonomously, and the human orchestrator knows they aren't 162 00:07:33,160 --> 00:07:36,320 going to be prompted for every little thing, they might rely on 163 00:07:36,320 --> 00:07:39,160 broader, more robust system level controls. 164 00:07:39,320 --> 00:07:43,360 Like explicitly restricting the AI to a specific sandbox 165 00:07:43,360 --> 00:07:45,520 directory. Rather than rely relying on a 166 00:07:45,520 --> 00:07:48,400 fatigued human blindly clicking, yes. 167 00:07:48,400 --> 00:07:51,480 Though it is worth noting that the technical details on exactly 168 00:07:51,480 --> 00:07:54,120 how those scope controls will work aren't fully fleshed out 169 00:07:54,120 --> 00:07:56,320 yet in these sources. And of course the competition is 170 00:07:56,320 --> 00:07:58,960 not sitting still. Yeah right as Anthropic is 171 00:07:58,960 --> 00:08:02,560 pushing auto mode, Open AI drops GPT 5.4. 172 00:08:02,560 --> 00:08:06,200 And the popular AI code editor cursor integrated it almost 173 00:08:06,200 --> 00:08:08,600 immediately. Almost overnight and this new 174 00:08:08,600 --> 00:08:11,360 model is apparently destroying all the benchmarks for reasoning 175 00:08:11,360 --> 00:08:13,960 and instruction following. But there's a very interesting 176 00:08:13,960 --> 00:08:17,960 catch that highlights what the day-to-day of an AI orchestrator 177 00:08:17,960 --> 00:08:21,000 actually feels like. It's not just about raw power. 178 00:08:21,120 --> 00:08:25,480 No GBT 5.4 is significantly more expensive, but more importantly, 179 00:08:25,680 --> 00:08:27,720 developers are actively complaining about its 180 00:08:27,720 --> 00:08:29,760 personality. That personality aspect is 181 00:08:29,760 --> 00:08:32,760 genuinely interesting. We often think of these tools as 182 00:08:32,760 --> 00:08:34,720 just calculators. But they have distinct 183 00:08:34,720 --> 00:08:36,440 communication styles. They really do. 184 00:08:36,440 --> 00:08:40,919 Right, the reports say GBT 5.4 has an issue with increased 185 00:08:40,919 --> 00:08:43,360 verbosity. It over explains it. 186 00:08:43,720 --> 00:08:46,560 Just talks and talks. If you just want a raw snippet 187 00:08:46,560 --> 00:08:50,040 of code, it gives you a 5 paragraph essay on the history 188 00:08:50,040 --> 00:08:51,400 of the programming language first. 189 00:08:51,400 --> 00:08:54,120 Which is exhausting. Meanwhile, developers are 190 00:08:54,120 --> 00:08:56,880 praising Claude for being incredibly concise. 191 00:08:57,520 --> 00:08:59,960 I love this one user observation we found. 192 00:09:00,200 --> 00:09:03,080 They describe Claude as feeling almost standoffish. 193 00:09:03,080 --> 00:09:05,280 Standoffish. Like a really busy senior 194 00:09:05,280 --> 00:09:08,800 engineer who has a million other people's questions to answer and 195 00:09:08,800 --> 00:09:11,960 zero time for small talk. Whereas Chachi PT feels like 196 00:09:11,960 --> 00:09:15,000 that overly Co worker who won't leave your desk. 197 00:09:15,000 --> 00:09:18,480 Exactly when you were working all day in an editor trying to 198 00:09:18,480 --> 00:09:21,440 maintain your own train of thought, that overly friendly, 199 00:09:21,440 --> 00:09:24,600 verbose tone can quickly become a massive hindrance. 200 00:09:24,600 --> 00:09:26,040 Do. You just want to output. 201 00:09:26,040 --> 00:09:29,800 Which naturally leads to a really practical strategy for 202 00:09:29,800 --> 00:09:32,480 you if you're managing these tools, which we can call the 203 00:09:32,480 --> 00:09:34,560 swap strategy I. Highly recommend this. 204 00:09:35,040 --> 00:09:38,520 The advice is simple. Do not use the newest, most 205 00:09:38,520 --> 00:09:41,280 expensive, most verbose models for everything. 206 00:09:41,800 --> 00:09:44,920 If you are just doing a routine task like a renaming a bunch of 207 00:09:44,920 --> 00:09:49,040 variables or writing a very basic API endpoint which is just 208 00:09:49,040 --> 00:09:52,640 the bridge that lets 2 pieces of software talk to each other, you 209 00:09:52,640 --> 00:09:55,240 should use the older, smaller, cheaper models. 210 00:09:55,360 --> 00:09:57,160 It's about using the right tool for the job. 211 00:09:57,400 --> 00:10:00,200 You wouldn't use a highly tuned sports car to haul a load of 212 00:10:00,200 --> 00:10:02,440 gravel, right? You want to reserve those heavy 213 00:10:02,440 --> 00:10:07,480 hitting powerful models like a GPT 5.4, the top tier clod for 214 00:10:07,480 --> 00:10:10,760 complex architectural decisions. Or gnarly debugging sessions. 215 00:10:10,760 --> 00:10:14,240 Yes, or writing out detailed project specifications and the 216 00:10:14,240 --> 00:10:16,160 current tooling fully supports this. 217 00:10:16,320 --> 00:10:19,320 You can literally swap models mid conversation in these 218 00:10:19,320 --> 00:10:21,360 editors. This strategy is critical, 219 00:10:21,360 --> 00:10:22,880 especially when you factor in auto mode. 220 00:10:22,880 --> 00:10:24,520 Because of the token cost, right? 221 00:10:24,640 --> 00:10:27,720 Tokens are essentially the currency or the unit of 222 00:10:27,720 --> 00:10:30,920 computing power that the AI company charges you for. 223 00:10:31,200 --> 00:10:34,960 Every word it reads and every word it writes costs tokens and. 224 00:10:35,160 --> 00:10:38,400 Autonomous operations can burn through thousands of tokens 225 00:10:38,520 --> 00:10:41,480 incredibly rapidly. If you let an expensive model 226 00:10:41,480 --> 00:10:45,520 run autonomously on a trivial task, you are just setting money 227 00:10:45,520 --> 00:10:47,680 on fire. OK, here's where it gets really 228 00:10:47,680 --> 00:10:51,000 interesting and honestly a little bit terrifying for the 229 00:10:51,000 --> 00:10:52,960 future of software. The danger zone. 230 00:10:53,480 --> 00:10:56,880 There is a massive red flag buried in these discussions. 231 00:10:57,160 --> 00:11:00,280 A true danger zone for anyone doing this kind of work. 232 00:11:01,120 --> 00:11:03,080 The realization from the notes is this. 233 00:11:03,840 --> 00:11:08,840 Your AI generated tests have the exact same blind spots as your 234 00:11:08,920 --> 00:11:11,320 AI generated code. If you take nothing else away 235 00:11:11,320 --> 00:11:14,480 from this deep dive today, write this concept on a sticky note 236 00:11:14,480 --> 00:11:16,360 and put it on your monitor. Is that important? 237 00:11:16,440 --> 00:11:18,960 This is the core challenge of the Orchestrator era. 238 00:11:19,000 --> 00:11:22,200 Think about the underlying logic if an AI model fundamentally 239 00:11:22,200 --> 00:11:24,920 misunderstands the requirements you've given it when it writes 240 00:11:24,920 --> 00:11:27,200 the application carry. It is going to misunderstand 241 00:11:27,200 --> 00:11:29,880 those exact same requirements identically when you ask it to 242 00:11:29,880 --> 00:11:31,640 write the safety tests for that code. 243 00:11:32,000 --> 00:11:35,160 It's like a student taking a test and then grading their own 244 00:11:35,160 --> 00:11:37,680 homework. Of course they're going to give 245 00:11:37,680 --> 00:11:40,240 themselves an A+. They think they got the answers 246 00:11:40,240 --> 00:11:41,000 right. Right. 247 00:11:41,120 --> 00:11:44,600 And this creates A phenomenon called false confidence. 248 00:11:44,600 --> 00:11:47,480 Which is terrifying. You look at your dashboard, you 249 00:11:47,480 --> 00:11:51,000 see 100 green check marks saying all the tests passed and you 250 00:11:51,000 --> 00:11:52,600 think your application is bulletproof. 251 00:11:52,600 --> 00:11:56,160 But it's completely broken. And this is especially dangerous 252 00:11:56,280 --> 00:11:58,760 with a new trend called vibe coding. 253 00:11:59,160 --> 00:12:03,960 Vibe coding is a fascinating and highly risky byproduct of this 254 00:12:03,960 --> 00:12:05,840 era. What is it exactly? 255 00:12:05,880 --> 00:12:08,720 It's when a human developer doesn't actually understand the 256 00:12:08,720 --> 00:12:11,280 code the AI generated under the hood. 257 00:12:11,280 --> 00:12:14,760 They just prompt the AI, look at the visual output of the app and 258 00:12:14,760 --> 00:12:17,520 say, yeah, the vibe feels right. It seems to work. 259 00:12:17,560 --> 00:12:20,880 Wow, so they are tweaking things based on surface level 260 00:12:20,880 --> 00:12:23,160 appearances without understanding the underlying 261 00:12:23,160 --> 00:12:25,240 mechanics. Yes, if you don't understand the 262 00:12:25,240 --> 00:12:28,240 code and your automated tests share the same blind spots as 263 00:12:28,240 --> 00:12:30,520 the code, you are flying completely blind. 264 00:12:31,040 --> 00:12:34,560 So if the AI's generated code is slawed, and generating tests 265 00:12:34,560 --> 00:12:37,720 from that code just perfectly duplicates the flaws, how do we 266 00:12:37,720 --> 00:12:40,560 break the cycle? There are three very actionable 267 00:12:40,560 --> 00:12:43,240 strategies that emerge from the notes to combat this false 268 00:12:43,240 --> 00:12:45,080 confidence. OK, what's the first one? 269 00:12:45,400 --> 00:12:49,200 The first strategy is to massively over specify edge 270 00:12:49,200 --> 00:12:52,040 cases and failure modes in your initial prompts. 271 00:12:52,600 --> 00:12:54,000 What does that look like in practice? 272 00:12:54,080 --> 00:12:58,520 Don't just ask the AI to build a checkout cart, explicitly tell 273 00:12:58,520 --> 00:13:01,640 it to handle the dark corners. The weird scenario, right? 274 00:13:02,040 --> 00:13:05,720 Ask it what happens if the user inputs A null value? 275 00:13:05,840 --> 00:13:09,680 What happens if their Internet router unplugs right as they hit 276 00:13:09,680 --> 00:13:12,000 the buy button, causing a network timeout? 277 00:13:12,400 --> 00:13:15,760 Oh, or what happens if 2 users try to buy the last item at the 278 00:13:15,760 --> 00:13:18,280 exact same millisecond, creating a race condition? 279 00:13:18,280 --> 00:13:22,200 Yes, you have to force the AI to look at the edge cases it would 280 00:13:22,200 --> 00:13:24,840 otherwise gloss over. OK, so that hardens the code 281 00:13:24,840 --> 00:13:26,680 itself, but what about the tests? 282 00:13:26,800 --> 00:13:31,120 How do we stop the AI from just grading its own flawed homework? 283 00:13:31,120 --> 00:13:33,880 That brings us to the second strategy, which is brilliant in 284 00:13:33,880 --> 00:13:35,120 its simplicity. I'm ready. 285 00:13:35,280 --> 00:13:37,880 Generate your tests from your original human written 286 00:13:37,880 --> 00:13:41,720 specifications and requirements, not from the code the AI just 287 00:13:41,720 --> 00:13:43,440 wrote. Oh, that makes so much sense. 288 00:13:43,440 --> 00:13:46,280 If you generate from the spec, you have an independent check. 289 00:13:46,600 --> 00:13:50,120 You then run those spec based tests against the generated 290 00:13:50,120 --> 00:13:51,920 code. It breaks the echo chamber. 291 00:13:51,960 --> 00:13:54,480 And the third strategy? The third strategy is the most 292 00:13:54,480 --> 00:13:58,200 fundamental shift in your job description as an orchestrator. 293 00:13:58,440 --> 00:14:01,200 You have to actually read the generated tests. 294 00:14:01,560 --> 00:14:05,960 But wait, if we are moving away from reading code, why are we 295 00:14:05,960 --> 00:14:09,440 suddenly reading tests? Isn't that just as tedious? 296 00:14:09,440 --> 00:14:11,320 Not at all, and this is a crucial distinction. 297 00:14:11,560 --> 00:14:14,640 Application code is generally imperative. 298 00:14:14,680 --> 00:14:19,600 Imperative, yes, It is a massive tangled maze of logic explaining 299 00:14:19,600 --> 00:14:22,240 the how. Go to the database, fetch this 300 00:14:22,240 --> 00:14:24,600 variable, run this loop, calculate this math. 301 00:14:24,600 --> 00:14:27,840 It's incredibly dense. Tests, on the other hand, are 302 00:14:28,040 --> 00:14:30,200 declared. They just state the what it's 303 00:14:30,200 --> 00:14:32,440 like a restaurant menu versus the actual recipe in the 304 00:14:32,440 --> 00:14:34,720 kitchen. Oh, that is a perfect analogy. 305 00:14:34,720 --> 00:14:37,080 The menu just says you'll get a steak with a side of potatoes. 306 00:14:37,120 --> 00:14:40,560 That's declarative. The imperative recipe is the 3 307 00:14:40,560 --> 00:14:43,480 pages of instructions on how to chop, season and sear 308 00:14:43,480 --> 00:14:45,760 everything. Yes, tests just say if I put an 309 00:14:45,920 --> 00:14:48,760 XI should get Y. Because they are declarative, 310 00:14:49,000 --> 00:14:52,600 they are significantly easier and faster for a human to skim. 311 00:14:53,040 --> 00:14:56,480 Reading the tests is a highly efficient way to spot if the AI 312 00:14:56,480 --> 00:14:59,680 has fundamentally misunderstood the core requirements of your 313 00:14:59,680 --> 00:15:03,480 application, without having to decipher the tangled imperative 314 00:15:03,480 --> 00:15:07,200 logic of the app itself. OK, so we know how to test the 315 00:15:07,320 --> 00:15:10,720 AI, but how do we talk to it in the first place to avoid these 316 00:15:10,720 --> 00:15:14,280 messes? There is a really revealing user 317 00:15:14,280 --> 00:15:18,120 question here from the sources about prompt engineering. 318 00:15:18,440 --> 00:15:19,960 Oh, the one about making no mistakes. 319 00:15:19,960 --> 00:15:24,160 Yeah, someone asked if telling the AI to make no mistakes is a 320 00:15:24,160 --> 00:15:26,520 good strategy. And the answer to that is a 321 00:15:26,520 --> 00:15:29,480 resounding no. Telling an AI to make no 322 00:15:29,480 --> 00:15:31,720 mistakes is like telling a driver don't crash. 323 00:15:31,920 --> 00:15:34,280 It doesn't actually give them any useful information on how to 324 00:15:34,280 --> 00:15:36,640 navigate the road. Right, because all these models 325 00:15:36,640 --> 00:15:39,000 are already statistically weighted to try to provide the 326 00:15:39,000 --> 00:15:41,520 most accurate answer. So if we aren't just yelling do 327 00:15:41,520 --> 00:15:43,960 a good job at the screen, what is the alternative? 328 00:15:43,960 --> 00:15:46,960 The alternative is a much better framework which we can call the 329 00:15:46,960 --> 00:15:49,960 Checkpoint framework. The core philosophy here is to 330 00:15:49,960 --> 00:15:53,080 be hyperspecific about the process, not just the outcome. 331 00:15:53,320 --> 00:15:56,200 Treat the AI like a brand new junior, your team member. 332 00:15:56,320 --> 00:15:59,720 Exactly, you wouldn't tell a new hire to build an entire 333 00:15:59,720 --> 00:16:02,480 corporate system over the weekend with 0 check insurance, 334 00:16:02,480 --> 00:16:03,520 right? You'd want to see their 335 00:16:03,520 --> 00:16:06,840 architecture plan first before they spend 40 hours doing the 336 00:16:06,840 --> 00:16:09,000 wrong thing. So your prompt should look 337 00:16:09,000 --> 00:16:11,680 something like this. First, read all the files to 338 00:16:11,680 --> 00:16:15,400 understand the architecture. Second, propose a step by step 339 00:16:15,400 --> 00:16:20,960 plan. 3rd wait for my explicit approval. 4th write the code 340 00:16:20,960 --> 00:16:24,440 with declarative tests and finally, if you are unsure about 341 00:16:24,440 --> 00:16:26,480 anything, ask me. Do not guess. 342 00:16:26,760 --> 00:16:30,040 That part right there. Propose a plan and wait for 343 00:16:30,040 --> 00:16:32,360 approval. It feels like the magic bullet. 344 00:16:32,440 --> 00:16:36,960 It forces the AI to lay out its assumptions on the table before 345 00:16:36,960 --> 00:16:38,960 it generates thousands of lines of code. 346 00:16:39,120 --> 00:16:41,600 It catches those blind spots right at the starting line. 347 00:16:41,680 --> 00:16:44,160 By imposing these process constraints, you maintain 348 00:16:44,160 --> 00:16:47,560 ultimate control while still giving the AI the autonomy to do 349 00:16:47,560 --> 00:16:49,640 the heavy lifting. This ties directly back to 350 00:16:49,640 --> 00:16:52,640 making auto mode safer. You get the speed of automation, 351 00:16:52,680 --> 00:16:54,840 but with strategically placed decision points. 352 00:16:55,080 --> 00:16:58,920 Now, when you introduce tools powerful enough to autonomously 353 00:16:58,920 --> 00:17:02,160 write complex software and reason to through complex logic, 354 00:17:02,680 --> 00:17:05,640 it's inevitably going to impact more than just our IDE's and 355 00:17:05,640 --> 00:17:08,280 code bases it. Bleeds into human sychology. 356 00:17:08,280 --> 00:17:12,400 And eventually, Global scale and the community discussions 357 00:17:12,400 --> 00:17:15,760 highlight this beautifully. For instance, outside of pure 358 00:17:15,760 --> 00:17:18,839 coding, users are praising Claude's conversational 359 00:17:18,839 --> 00:17:21,560 reasoning for helping them navigate intense life 360 00:17:21,560 --> 00:17:23,599 complexities. Like providing structured 361 00:17:23,599 --> 00:17:26,560 support and frameworks for an ADHD diagnosis. 362 00:17:26,560 --> 00:17:29,480 It's a reminder that these are reasoning engines, not just 363 00:17:29,480 --> 00:17:31,480 coding calculators. And they are engines with 364 00:17:31,480 --> 00:17:34,200 distinct alignment tuning. They aren't blank slates that 365 00:17:34,200 --> 00:17:35,720 just say yes to everything, right? 366 00:17:36,000 --> 00:17:39,080 The discussions note that Claude in particular will occasionally 367 00:17:39,080 --> 00:17:41,280 flat out refuse a user's request. 368 00:17:41,360 --> 00:17:44,160 Responding with things like, no, I don't think that's a good idea 369 00:17:44,640 --> 00:17:47,560 if it determines the request violates its safety protocols or 370 00:17:47,560 --> 00:17:51,000 logic constraints. Which brings us to perhaps the 371 00:17:51,000 --> 00:17:53,960 most jarring curveball in all of this source material. 372 00:17:54,480 --> 00:17:57,680 The sheer scale and capability of these tools are triggering 373 00:17:57,680 --> 00:18:00,320 massive geopolitical ripples. They really are. 374 00:18:00,520 --> 00:18:03,800 Now, to be absolutely clear to you, the listener, we are 375 00:18:03,800 --> 00:18:06,160 strictly imparting the information as it appears in 376 00:18:06,160 --> 00:18:07,960 these Community and developer notes. 377 00:18:08,200 --> 00:18:11,120 We are not endorsing any viewpoints, validating rumors or 378 00:18:11,120 --> 00:18:14,080 taking any political sides. We are just looking at the 379 00:18:14,080 --> 00:18:16,160 reality of the ecosystem as reported. 380 00:18:17,000 --> 00:18:19,920 According to the notes, the Pentagon has officially labeled 381 00:18:19,920 --> 00:18:22,200 anthropic as a supply chain risk. 382 00:18:22,600 --> 00:18:24,280 That is a staggering development. 383 00:18:24,560 --> 00:18:27,840 When a Defense Department labels a foundational AI company as a 384 00:18:27,840 --> 00:18:31,080 supply chain risk, it carries immense geopolitical 385 00:18:31,080 --> 00:18:33,640 implications for how these tools will be regulated. 386 00:18:33,840 --> 00:18:37,360 Regulated, procured and integrated into national 387 00:18:37,360 --> 00:18:39,800 infrastructure. It apparently prompted a direct 388 00:18:39,800 --> 00:18:42,800 statement from Anthropic CEO Dario Amede. 389 00:18:42,960 --> 00:18:46,200 Though the granular details that statement aren't fully unpacked 390 00:18:46,200 --> 00:18:48,280 in the immediate developer discussions we're looking at, 391 00:18:48,320 --> 00:18:49,640 right? And just to illustrate how 392 00:18:49,640 --> 00:18:53,360 chaotic the flow of information this space can be, right next to 393 00:18:53,360 --> 00:18:57,440 that massive geopolitical news is an entry that honestly looks 394 00:18:57,440 --> 00:18:59,760 like a glitch in the Matrix. It really does. 395 00:18:59,920 --> 00:19:04,000 It's categorized as a fun fact and claims that Hideo Kojima, 396 00:19:04,000 --> 00:19:07,760 the legendary and very human video game director, was 397 00:19:07,760 --> 00:19:11,720 apparently named Anthropic CEO. The sources explicitly state the 398 00:19:11,720 --> 00:19:13,960 context for this is totally unclear. 399 00:19:14,640 --> 00:19:17,160 I don't know if that's a bizarre inside joke from a developer 400 00:19:17,160 --> 00:19:19,560 forum. Or a hallucination generated by 401 00:19:19,560 --> 00:19:23,200 an AI summarizing the news. Or just a wild Internet rumor 402 00:19:23,200 --> 00:19:25,400 that got caught in the data dragnet? 403 00:19:25,400 --> 00:19:28,080 It's a perfect encapsulation of the current landscape. 404 00:19:28,120 --> 00:19:30,640 You have hyper advanced technical breakthroughs, 405 00:19:30,920 --> 00:19:33,680 profound geopolitical maneuvering, and absolute 406 00:19:33,680 --> 00:19:37,080 Internet chaos all swirling together in the exact same 407 00:19:37,080 --> 00:19:38,880 forums. But if we pull back from the 408 00:19:38,880 --> 00:19:41,480 rumors and focus purely on the technical realities we've 409 00:19:41,480 --> 00:19:45,520 unraveled today, a very clear, very defining picture emerges. 410 00:19:45,600 --> 00:19:48,560 So what does this all mean? If we synthesize everything 411 00:19:48,560 --> 00:19:51,240 we've talked about, from the 30 year veteran who put down their 412 00:19:51,240 --> 00:19:54,520 keyboard, to the sheer speed of auto mode, to the terrifying 413 00:19:54,520 --> 00:19:56,520 reality of AI greeting its own homework. 414 00:19:56,520 --> 00:19:59,240 It means we are officially moving out of the era of syntax 415 00:19:59,240 --> 00:20:01,360 and typing. We are fully entering the era of 416 00:20:01,360 --> 00:20:04,800 orchestration, logic checking and high level system design. 417 00:20:05,000 --> 00:20:08,360 That is the ultimate So what for you as a listener today? 418 00:20:08,600 --> 00:20:12,200 Whether you are a professional software developer, a project 419 00:20:12,200 --> 00:20:15,520 manager, or really just anyone trying to navigate a workplace 420 00:20:15,520 --> 00:20:19,320 that is rapidly saturating with AI tools. 421 00:20:19,440 --> 00:20:21,720 The core skill set has fundamentally changed. 422 00:20:21,720 --> 00:20:25,400 The real value you bring to the table is no longer doing the 423 00:20:25,400 --> 00:20:28,960 rote, repetitive execution. Your value is in knowing how to 424 00:20:28,960 --> 00:20:32,680 define strict constraints, how to build checkpoint frameworks, 425 00:20:32,840 --> 00:20:35,480 how to rigorously review the output. 426 00:20:35,480 --> 00:20:39,320 And above all, how to avoid the deadly trap of false confidence? 427 00:20:39,320 --> 00:20:41,560 You have to be the critical thinker in the loop. 428 00:20:41,680 --> 00:20:44,400 Have to be the one who actually reads the declarative tests. 429 00:20:44,400 --> 00:20:46,240 Exactly. Which leads me to a final 430 00:20:46,240 --> 00:20:48,160 thought for you to Mull over as we wrap up today. 431 00:20:48,880 --> 00:20:51,720 We've established that the ultimate solution to AI blind 432 00:20:51,720 --> 00:20:55,000 spots right now is for human orchestrators to read those 433 00:20:55,000 --> 00:20:58,440 declarative menus, those tests, to ensure the AI actually 434 00:20:58,440 --> 00:21:00,800 understood the assignment. That's the safeguard. 435 00:21:00,960 --> 00:21:03,680 But think about the trajectory we are on with things like auto 436 00:21:03,680 --> 00:21:06,640 mode and GPT 5.4. The systems are getting bigger. 437 00:21:06,840 --> 00:21:10,240 Much bigger. What happens in 3-5 or ten years 438 00:21:10,360 --> 00:21:13,680 when these systems and the applications they build become 439 00:21:13,680 --> 00:21:18,600 so massively complex that even the declarative tests are simply 440 00:21:18,600 --> 00:21:22,520 too vast for any single human mind to read and comprehend? 441 00:21:22,520 --> 00:21:25,800 That is a daunting thought. Will we inevitably reach a point 442 00:21:25,800 --> 00:21:28,600 where we need a second, completely separate AI model 443 00:21:28,760 --> 00:21:31,960 whose only job is to audit the work of the primary orchestrator 444 00:21:32,160 --> 00:21:34,120 AI? And if we do cross that 445 00:21:34,120 --> 00:21:36,240 threshold, who is going to audit the auditor? 446 00:21:36,680 --> 00:21:39,440 It raises A profound question about the upper limits of human 447 00:21:39,440 --> 00:21:43,080 oversight in an age of hyper scaled autonomous AI 448 00:21:43,080 --> 00:21:45,840 orchestration. It is a puzzle the industry will 449 00:21:45,840 --> 00:21:47,880 inevitably and quickly have to solve. 450 00:21:48,200 --> 00:21:50,600 Thank you for walking through this massive paradigm shift with 451 00:21:50,600 --> 00:21:52,600 us. We hope this deep dive has given 452 00:21:52,600 --> 00:21:54,920 you a clearer, more practical map of where software 453 00:21:54,920 --> 00:21:58,040 development and really the future of all knowledge work is 454 00:21:58,040 --> 00:21:59,240 heading. Thanks for having me. 455 00:21:59,360 --> 00:22:00,720 Keep questioning your. Assumptions. 456 00:22:00,720 --> 00:22:03,480 Keep checking those AI blind spots, and above all, stay 457 00:22:03,480 --> 00:22:05,440 curious. We will catch you on the next 458 00:22:05,440 --> 00:22:05,960 deep dive.