1 00:00:00,040 --> 00:00:05,160 Imagine handing the a hyper efficient caffeine fueled intern 2 00:00:05,320 --> 00:00:08,080 the keys to your company's live database. 3 00:00:08,080 --> 00:00:10,000 OK, sounds risky already. Right. 4 00:00:10,000 --> 00:00:13,640 And you point them at a mess and within, I mean, literally 3 5 00:00:13,640 --> 00:00:17,440 seconds, 25,000 crucial documents are just they're just 6 00:00:17,440 --> 00:00:19,880 gone, wiped from existence. Oh, wow. 7 00:00:20,240 --> 00:00:22,280 Yeah. And that isn't some hypothetical 8 00:00:22,280 --> 00:00:24,360 cautionary tale that actually happened today. 9 00:00:24,360 --> 00:00:26,520 Incredible. It perfectly captures this 10 00:00:26,520 --> 00:00:30,160 massive, almost violent shift we're exploring in today's Deep 11 00:00:30,160 --> 00:00:31,480 dive. So welcome in. 12 00:00:31,480 --> 00:00:33,560 Glad to be here. We have an incredible stack of 13 00:00:33,560 --> 00:00:37,080 sources for you from today, March 13th, 2026. 14 00:00:37,280 --> 00:00:41,200 We've got developer blogs, Reddit posts, deep tech articles 15 00:00:41,200 --> 00:00:45,680 and they all point to 1 reality. Which is that the era of AI as a 16 00:00:45,680 --> 00:00:48,600 simple passive text generator is completely over. 17 00:00:48,600 --> 00:00:50,800 Exactly. We are now living in the era of 18 00:00:50,800 --> 00:00:54,480 the autonomous executing agent. It is a really profound 19 00:00:54,480 --> 00:00:57,120 transition period right now, and honestly that stack of sources 20 00:00:57,120 --> 00:00:59,480 you mentioned, it captures the whiplash perfectly. 21 00:00:59,480 --> 00:01:00,960 Yeah, whiplash is the right word. 22 00:01:01,120 --> 00:01:03,320 It really is. You look at the landscape today 23 00:01:03,320 --> 00:01:06,440 and you see absolute extremes. On one hand, you have 24 00:01:06,720 --> 00:01:09,880 catastrophic infrastructure failures like that database wipe 25 00:01:09,880 --> 00:01:11,800 you just teased. Right, total disasters. 26 00:01:11,800 --> 00:01:14,680 And then on the other hand, we're seeing solo developers 27 00:01:14,680 --> 00:01:17,920 replacing the output of entire engineering teams in a matter of 28 00:01:17,920 --> 00:01:19,400 days. It's hard to wrap your head 29 00:01:19,400 --> 00:01:21,720 around. It is, but what we need to 30 00:01:21,720 --> 00:01:26,480 unpack for you today isn't some sci-fi story about AI replacing 31 00:01:26,480 --> 00:01:30,640 human intelligence. It's a highly practical look at 32 00:01:30,640 --> 00:01:35,000 how these executing agents are radically changing the leverage 33 00:01:35,240 --> 00:01:37,280 of human expertise. Well, I want to start with that 34 00:01:37,280 --> 00:01:40,320 catastrophe, actually, because to understand the leverage, we 35 00:01:40,320 --> 00:01:43,800 really have to understand the raw, unfiltered power of these 36 00:01:43,800 --> 00:01:45,120 agents. Oh absolutely. 37 00:01:45,120 --> 00:01:46,840 The raw power is the whole point. 38 00:01:46,920 --> 00:01:49,360 Right, so there's this lead story from the Claude AI 39 00:01:49,360 --> 00:01:53,080 subreddit today that is frankly nightmare fuel for anyone who 40 00:01:53,080 --> 00:01:54,920 manages technology. It really is. 41 00:01:55,280 --> 00:01:58,160 So a developer is prepping a project for production. 42 00:01:58,160 --> 00:02:00,280 They need to clean up some mock data. 43 00:02:00,400 --> 00:02:03,120 You know, fake user profiles, test transactions, Is that sort 44 00:02:03,120 --> 00:02:04,480 of? Thing, yeah, completely standard 45 00:02:04,480 --> 00:02:06,520 housekeeping task. Right, totally routine. 46 00:02:06,720 --> 00:02:11,120 So the developer tells the AI agent to just clean up the dummy 47 00:02:11,120 --> 00:02:13,200 data. And I'm guessing the agent does 48 00:02:13,200 --> 00:02:15,600 it perfectly. It does it, Florida State, It 49 00:02:15,600 --> 00:02:18,560 writes the script, it executes the command, it follows the 50 00:02:18,560 --> 00:02:20,040 instructions to the absolute letter. 51 00:02:20,360 --> 00:02:23,760 But the only problem is that the mock database and the real live 52 00:02:23,760 --> 00:02:27,080 production database we're sitting in the exact same 53 00:02:27,080 --> 00:02:29,080 connection context. Oh no. 54 00:02:29,200 --> 00:02:33,560 Yeah, and the AI just annihilated 25, 5000 real 55 00:02:33,560 --> 00:02:35,440 documents. Let's breakdown why that 56 00:02:35,440 --> 00:02:39,760 happened, because connection context is It's a piece of 57 00:02:39,760 --> 00:02:42,800 jargon that hides a very dangerous mechanism. 58 00:02:42,840 --> 00:02:45,760 Yeah, please explain that for anyone who isn't deep in the 59 00:02:45,760 --> 00:02:48,040 back end stuff. Think of a connection context as 60 00:02:48,040 --> 00:02:50,840 a master key ring. When a developer sits down to 61 00:02:50,840 --> 00:02:53,760 work, they often load up a bunch of access credentials into their 62 00:02:53,760 --> 00:02:55,600 current environment. Right, so they don't have to 63 00:02:55,600 --> 00:02:57,120 keep logging in. Exactly. 64 00:02:57,120 --> 00:02:59,600 It lets them seamlessly move between testing their code and 65 00:02:59,600 --> 00:03:01,920 pushing it live. So they have the key to the test 66 00:03:01,920 --> 00:03:05,520 track, and they have the key to the actual highway sitting on 67 00:03:05,520 --> 00:03:07,560 the same ring. OK, so when they handed the keys 68 00:03:07,560 --> 00:03:09,320 to the AI, they handed it everything. 69 00:03:09,640 --> 00:03:13,360 Precisely, and this is where the underlying nature of these new 70 00:03:13,360 --> 00:03:15,680 agents is really exposed. How do you mean by that? 71 00:03:15,840 --> 00:03:18,680 Well, they don't possess intuition. 72 00:03:19,400 --> 00:03:23,640 They optimize purely ruthlessly for task completion. 73 00:03:23,800 --> 00:03:26,560 They just do the job right. Yeah, if you tell an agent to 74 00:03:26,560 --> 00:03:30,600 delete data, and the easiest path to executing that command 75 00:03:31,040 --> 00:03:33,920 is using the live production key that happens to be sitting right 76 00:03:33,920 --> 00:03:35,800 there on the key chain, it's. Just going to use it. 77 00:03:35,880 --> 00:03:38,600 It will use it. It doesn't pause. 78 00:03:39,120 --> 00:03:41,400 It doesn't get that sudden drop in the stomach that human 79 00:03:41,400 --> 00:03:44,480 developer gets when they see a production server name and 80 00:03:44,480 --> 00:03:46,720 think, wait, let me double check this before I hit enter. 81 00:03:46,760 --> 00:03:50,480 It makes me think of like a brilliant surgeon who operates 82 00:03:50,480 --> 00:03:54,480 Florida State but completely refuses to check the patient's 83 00:03:54,480 --> 00:03:57,120 chart to see if they're in for a heart bypass or just a 84 00:03:57,120 --> 00:03:59,240 tonsillectomy. That's a great analogy the. 85 00:03:59,240 --> 00:04:02,520 Physical execution is utterly perfect, but the contextual 86 00:04:02,520 --> 00:04:06,240 awareness is absolute 0. That is the exact friction point 87 00:04:06,240 --> 00:04:08,560 right there. And you know the original poster 88 00:04:08,560 --> 00:04:10,840 didn't share the this disaster to complain that the AI was 89 00:04:10,840 --> 00:04:12,160 stupid. No, they didn't at all. 90 00:04:12,280 --> 00:04:15,920 They shared it as a dire warning about autonomous momentum. 91 00:04:16,279 --> 00:04:20,839 You simply cannot solve this by prompting the AI to be careful. 92 00:04:20,839 --> 00:04:22,000 Wait, let me make sure I'm getting this. 93 00:04:22,000 --> 00:04:24,280 Why can't I just put in the system instructions? 94 00:04:24,280 --> 00:04:26,880 You know, hey, under no circumstances should you ever 95 00:04:26,880 --> 00:04:28,920 touch the database labeled production. 96 00:04:29,200 --> 00:04:33,200 Because you are relying on probabilistic text generation to 97 00:04:33,200 --> 00:04:36,080 act as a hard security boundary. OK, say more about. 98 00:04:36,080 --> 00:04:39,960 That and LLM at its core is predicting the next most likely 99 00:04:39,960 --> 00:04:42,000 token. Most of the time, yes, it will 100 00:04:42,000 --> 00:04:45,400 follow your instruction right. But if the context window gets 101 00:04:45,400 --> 00:04:50,200 crowded, or the logic gets sufficiently complex, it might 102 00:04:50,200 --> 00:04:53,240 simply hallucinate a reason why touching production is actually 103 00:04:53,240 --> 00:04:56,480 what you really meant for it to do in this specific edge case. 104 00:04:56,480 --> 00:04:59,120 Oh wow, so you can't trust the text prompt to protect the 105 00:04:59,120 --> 00:05:00,320 system? Exactly. 106 00:05:00,480 --> 00:05:02,840 The sources today are hammering this point home. 107 00:05:02,840 --> 00:05:05,880 Safeguards have to be enforced at the physical infrastructure 108 00:05:05,880 --> 00:05:07,400 level. Meaning the system itself 109 00:05:07,400 --> 00:05:10,120 physically blocks the scalpel, to go back to my surgeon 110 00:05:10,120 --> 00:05:11,880 analogy. Yes, exactly. 111 00:05:12,080 --> 00:05:15,800 If you are integrating executing agents into your workflow, you 112 00:05:15,800 --> 00:05:19,560 need mandatory DRY run modes hard coded into your deployment 113 00:05:19,560 --> 00:05:22,560 pipelines. So the AI can only plan, not 114 00:05:22,560 --> 00:05:23,480 act. Right. 115 00:05:23,720 --> 00:05:26,880 The AI should only be able to draft the plan, not execute it 116 00:05:26,880 --> 00:05:29,240 blindly. You need to strictly enforce 117 00:05:29,240 --> 00:05:33,400 read only database users during any kind of exploration phase. 118 00:05:33,400 --> 00:05:35,840 That makes total sense. And if a destructive action is 119 00:05:35,840 --> 00:05:39,200 queued up, like dropping a database, there has to be a hard 120 00:05:39,200 --> 00:05:42,800 human confirmation step. Like the two key system on a 121 00:05:42,800 --> 00:05:44,760 nuclear submarine. That's exactly it. 122 00:05:44,960 --> 00:05:48,760 The AI can turn the first key by writing the script, but a human 123 00:05:48,760 --> 00:05:51,440 finger has to physically turn the second key to actually 124 00:05:51,440 --> 00:05:53,280 launch it. Which naturally leads to a 125 00:05:53,280 --> 00:05:55,400 massive contradiction, I think. How so? 126 00:05:55,560 --> 00:05:59,560 Well, if these agents are so inherently risky, so dangerously 127 00:05:59,560 --> 00:06:02,680 literal, and they require so much heavy infrastructure just 128 00:06:02,680 --> 00:06:05,560 to keep them from burning the house down, why on earth are we 129 00:06:05,560 --> 00:06:07,520 doing this? That's the $1,000,000 question. 130 00:06:07,520 --> 00:06:10,840 Why are is racing to give them more autonomy instead of just 131 00:06:10,840 --> 00:06:13,160 keeping them safely contained in a chat window? 132 00:06:13,360 --> 00:06:15,440 Because the friction is worth it. 133 00:06:16,320 --> 00:06:20,000 I mean, the upside we are seeing in the wild right now is simply 134 00:06:20,000 --> 00:06:22,840 too massive to ignore. And we have two incredible 135 00:06:22,840 --> 00:06:25,600 success stories in the stack today that prove exactly that. 136 00:06:25,680 --> 00:06:28,120 Let's hear them. The first one is just a pure 137 00:06:28,120 --> 00:06:31,080 velocity play. A Reddit user with a solid 138 00:06:31,080 --> 00:06:34,040 background in software development decided to build a 139 00:06:34,040 --> 00:06:37,800 complex project. They scoped it out and estimated 140 00:06:37,800 --> 00:06:41,000 it would normally take an entire engineering team several weeks 141 00:06:41,000 --> 00:06:44,680 to build, test, and deploy a. Pretty hefty project then. 142 00:06:44,720 --> 00:06:46,720 Right. Well, this single developer 143 00:06:46,920 --> 00:06:51,160 finished it in three partial evenings. 3 evenings to 144 00:06:51,200 --> 00:06:54,840 replicate hundreds of hours of coordinated team output. 145 00:06:54,920 --> 00:06:58,120 That is wild. It is, and the craziest detail 146 00:06:58,120 --> 00:07:00,920 to me was the bottleneck. The only thing that slowed this 147 00:07:00,920 --> 00:07:04,000 developer down was running out of tokens every 90 minutes. 148 00:07:04,000 --> 00:07:06,960 Oh man, the token limits. Yeah, they were literally moving 149 00:07:06,960 --> 00:07:10,520 so fast, generating so much code that they hit the physical speed 150 00:07:10,520 --> 00:07:14,240 limits of the AI Providers API. They had to sit there and 151 00:07:14,240 --> 00:07:17,040 twiddle their thumbs waiting for their token bucket to refill so 152 00:07:17,040 --> 00:07:19,440 they could keep sprinting. Which is a fascinating new 153 00:07:19,440 --> 00:07:23,160 problem to have, honestly. It really is explain tokens 154 00:07:23,160 --> 00:07:25,520 really quickly for us, sure. Gokens are basically the 155 00:07:25,520 --> 00:07:28,840 fundamental unit of currency for an AI's processing power. 156 00:07:28,840 --> 00:07:31,680 Words, parts of words, or code snippets. 157 00:07:31,800 --> 00:07:34,120 Got it. When your only friction is the 158 00:07:34,120 --> 00:07:37,680 silicon limits of the server farm, your personal leverage is 159 00:07:37,680 --> 00:07:40,200 astronomical. But you know, it's not just 160 00:07:40,200 --> 00:07:43,800 about slapping together new, messy projects at lightspeed. 161 00:07:43,960 --> 00:07:45,520 Right, it applies to older stuff too. 162 00:07:45,600 --> 00:07:47,720 Exactly. The second story from today 163 00:07:47,720 --> 00:07:51,320 shows how this applies to deep, mature legacy systems. 164 00:07:51,320 --> 00:07:54,320 This is the Shopify story and it might be the most significant 165 00:07:54,320 --> 00:07:56,680 data point we have today. It's a really big deal. 166 00:07:56,680 --> 00:07:59,720 Tobias Lipke, the CEO of Shopify. 167 00:08:00,080 --> 00:08:03,320 Way back in 2005, he created a Ruby template engine called 168 00:08:03,320 --> 00:08:05,040 Liquid. Which is everywhere now. 169 00:08:05,200 --> 00:08:07,880 It is, and for anyone unfamiliar, a template engine is 170 00:08:07,880 --> 00:08:10,680 basically the bridge between a database and a storefront. 171 00:08:11,040 --> 00:08:14,200 It takes the raw product data and safely injects it into the 172 00:08:14,200 --> 00:08:17,680 HTML so you, the shopper, actually see the web page. 173 00:08:17,720 --> 00:08:19,560 Right. It's foundational open source 174 00:08:19,560 --> 00:08:21,480 software. It has been scrutinized, 175 00:08:21,480 --> 00:08:24,480 maintained and optimized by thousands of incredibly smart 176 00:08:24,480 --> 00:08:26,920 human contributors for over 20 years. 177 00:08:27,160 --> 00:08:29,680 It is the definition of mature code totally. 178 00:08:29,880 --> 00:08:34,320 Well, Toby decided to use clod to analyze and optimize Liquid, 179 00:08:35,080 --> 00:08:37,320 and he just submitted a pull request. 180 00:08:37,640 --> 00:08:41,320 You know, a proposed update to the core code that makes parsing 181 00:08:41,320 --> 00:08:45,160 and rendering 53% faster. 53%, that's huge. 182 00:08:45,160 --> 00:08:48,600 And it gets better. It does it with 61% fewer memory 183 00:08:48,600 --> 00:08:49,600 allocations. OK. 184 00:08:49,760 --> 00:08:52,920 We really need to pause and dissect the magnitude of that 185 00:08:53,400 --> 00:08:57,240 61% fewer memory allocations. Why is that so impressive? 186 00:08:57,400 --> 00:09:00,520 Because in software, every time a program needs to store a piece 187 00:09:00,520 --> 00:09:04,320 of data temporarily, it asks the computer for a little block of 188 00:09:04,320 --> 00:09:07,280 memory. It's like a chef constantly 189 00:09:07,280 --> 00:09:10,160 walking across the kitchen to the fridge to grab one single 190 00:09:10,160 --> 00:09:12,760 ingredient over and. Over right, it takes time. 191 00:09:12,760 --> 00:09:15,880 Exactly. Over 20 years, human developers 192 00:09:15,880 --> 00:09:18,680 optimized liquid as best they could to reduce those trips to 193 00:09:18,680 --> 00:09:19,520 the fridge. But they hit a. 194 00:09:19,520 --> 00:09:23,000 Wall, they did, but the AI looked at the entire code base 195 00:09:23,000 --> 00:09:26,280 at once, saw the holistic patterns, and figured out how to 196 00:09:26,280 --> 00:09:29,480 carry a massive tray of ingredients in one single trip. 197 00:09:29,720 --> 00:09:34,160 Yeah, it found dozens of these micro optimizations that human 198 00:09:34,160 --> 00:09:37,320 eyes simply couldn't perceive across thousands of lines of 199 00:09:37,320 --> 00:09:38,960 code. OK, but hearing that actually 200 00:09:38,960 --> 00:09:40,960 makes me a bit anxious. How so? 201 00:09:41,200 --> 00:09:44,240 Well if one developer in his basement can out produce an 202 00:09:44,240 --> 00:09:48,360 entire engineering team in three days and an AI can spot 203 00:09:48,360 --> 00:09:52,000 structural inefficiencies that 20 years of elite human experts 204 00:09:52,000 --> 00:09:55,080 completely missed. I mean, isn't human expertise 205 00:09:55,080 --> 00:09:56,880 dead? I can see why you think that. 206 00:09:56,880 --> 00:09:59,560 Like why do we even need developers to learn how to code 207 00:09:59,560 --> 00:10:02,000 anymore? It is incredibly tempting to 208 00:10:02,000 --> 00:10:05,800 draw that conclusion, but if you look closely at how the sources 209 00:10:05,800 --> 00:10:09,200 frame these victories, it reveals the exact opposite. 210 00:10:09,200 --> 00:10:12,960 Really the opposite. Yes, human expertise isn't dead 211 00:10:12,960 --> 00:10:14,440 at all. It is being aggressively 212 00:10:14,440 --> 00:10:17,280 supercharged. Look at Toby Lipke's success 213 00:10:17,280 --> 00:10:19,680 with Liquid. The take away wasn't AI is 214 00:10:19,680 --> 00:10:21,600 fundamentally smarter than human engineers. 215 00:10:21,880 --> 00:10:25,200 The reality is that Toby understands Liquid better than 216 00:10:25,200 --> 00:10:27,760 almost anyone on the planet. Because he built it. 217 00:10:27,920 --> 00:10:29,880 Exactly. He had the conceptual map. 218 00:10:30,000 --> 00:10:32,800 He knew exactly which files to point the AI toward. 219 00:10:33,040 --> 00:10:35,920 He knew how to evaluate its bizarre, counterintuitive 220 00:10:35,920 --> 00:10:38,840 suggestions. And crucially, he knew how to 221 00:10:38,840 --> 00:10:42,080 safely stitch those micro optimizations back into a stable 222 00:10:42,080 --> 00:10:44,400 update. Ah I see, so the AI was a hyper 223 00:10:44,400 --> 00:10:47,200 fast explorer but Toby was the one holding the compass? 224 00:10:47,200 --> 00:10:50,120 Exactly, and you see the exact same pattern with the Reddit 225 00:10:50,120 --> 00:10:52,560 user who built the massive project in three evenings. 226 00:10:52,560 --> 00:10:54,520 The one who ran out of tokens. Right. 227 00:10:54,800 --> 00:10:57,920 They explicitly noted in their write up that their deep 228 00:10:57,920 --> 00:11:00,800 background in software development was the absolute 229 00:11:00,800 --> 00:11:03,880 requirement for their speed. Because they had to know what to 230 00:11:03,880 --> 00:11:05,800 ask for. Background knowledge is what 231 00:11:05,800 --> 00:11:07,800 allows you to direct the agent effectively. 232 00:11:08,360 --> 00:11:11,240 If you don't know how a scalable system should be architected, 233 00:11:11,600 --> 00:11:13,200 you can't prompt the AI to build. 234 00:11:13,200 --> 00:11:14,880 It and when it breaks, you're stuck. 235 00:11:14,880 --> 00:11:19,640 Yes, when the AI inevitably gets stuck in a logic loop, you can't 236 00:11:19,640 --> 00:11:23,360 debug it. Pure time constrained grunt work 237 00:11:23,360 --> 00:11:27,840 like typing out boilerplate code is disappearing, but expert 238 00:11:27,840 --> 00:11:30,880 direction is becoming paramount. I see where you're going, but 239 00:11:30,880 --> 00:11:33,240 that introduces a massive complication in my mind. 240 00:11:33,600 --> 00:11:36,360 If we are speeding up the generation of code this 241 00:11:36,360 --> 00:11:40,560 drastically, what happens to the actual quality of the software 242 00:11:40,560 --> 00:11:42,400 we rely on? That's the real concern. 243 00:11:42,400 --> 00:11:44,080 Right. And I want to look at a piece 244 00:11:44,080 --> 00:11:47,440 from Simon Willison today. He's a highly respected voice 245 00:11:47,440 --> 00:11:50,840 and developer tools, and he brings up this concept of 246 00:11:50,840 --> 00:11:55,480 agentic engineering patterns. Willison has a very grounded 247 00:11:55,520 --> 00:11:57,640 practical perspective on this transition. 248 00:11:57,680 --> 00:12:01,160 He does, and he throws down a massive challenge to the 249 00:12:01,160 --> 00:12:04,040 industry here. His argument is that if the 250 00:12:04,040 --> 00:12:08,000 introduction of AI is lowering your organization's overall code 251 00:12:08,000 --> 00:12:12,200 quality, that is a process failure on your end, not a 252 00:12:12,240 --> 00:12:13,840 technology failure. Interesting. 253 00:12:13,840 --> 00:12:16,720 Why does he say that? Because the sheer speed of AI 254 00:12:16,720 --> 00:12:19,600 should actually be granting you the bandwidth to apply quality 255 00:12:19,600 --> 00:12:21,520 standards that you previously couldn't afford. 256 00:12:21,840 --> 00:12:23,800 Let's define that bandwidth, because that's important. 257 00:12:23,960 --> 00:12:26,800 In a traditional development cycle, writing the core feature 258 00:12:26,800 --> 00:12:29,640 takes up like 90% of the budget in time. 259 00:12:29,720 --> 00:12:31,400 Right. The actual building part, yeah. 260 00:12:32,040 --> 00:12:35,560 Writing the automated tests, Refactoring clunky code, adding 261 00:12:35,560 --> 00:12:37,920 robust error handling. That's the vegetables of 262 00:12:37,920 --> 00:12:40,560 software development. Humans hate doing it and 263 00:12:40,560 --> 00:12:44,120 managers hate paying for it. So Willison is saying that if 264 00:12:44,120 --> 00:12:47,320 the AI writes the core feature in 10 minutes, you now have the 265 00:12:47,320 --> 00:12:50,320 bandwidth to tell the AI to spend the next hour writing the 266 00:12:50,320 --> 00:12:53,360 most exhaustive suite of automated tests imaginable. 267 00:12:53,360 --> 00:12:55,240 Exactly. You finally have time to eat 268 00:12:55,240 --> 00:12:57,840 your vegetables. That sounds beautiful in theory, 269 00:12:57,880 --> 00:13:01,040 you know, for a brand new, clean project. 270 00:13:01,480 --> 00:13:03,960 But what? What does this mean for the 271 00:13:03,960 --> 00:13:06,920 messy legacy code bases out there? 272 00:13:06,960 --> 00:13:08,880 The ones held together by duct tape, yes. 273 00:13:09,160 --> 00:13:11,880 The corporate systems held together by duct tape and hope. 274 00:13:12,240 --> 00:13:15,360 If I'm suddenly generating code 10 times faster using an AI, 275 00:13:15,360 --> 00:13:17,680 aren't I just generating 10 times more technical dip? 276 00:13:17,920 --> 00:13:21,000 That is the crucial counter argument, and it is a deeply 277 00:13:21,000 --> 00:13:24,280 valid concern highlighted across today's material. 278 00:13:24,280 --> 00:13:27,360 It has to be right? Yes, because quality requires 279 00:13:27,360 --> 00:13:29,000 understanding, not just bandwidth. 280 00:13:29,600 --> 00:13:32,720 Unleashing an AI into a 10 year old, poorly understood legacy 281 00:13:32,720 --> 00:13:36,360 system is incredibly dangerous. Because the AI doesn't know why 282 00:13:36,360 --> 00:13:39,280 a certain weird line of code exists, it just sees it as 283 00:13:39,280 --> 00:13:41,520 inefficient and deletes it. Precisely. 284 00:13:42,160 --> 00:13:45,560 Let's say there's a bizarre block of code from 2018 that 285 00:13:45,560 --> 00:13:48,120 forces a transaction to delay by two seconds. 286 00:13:48,240 --> 00:13:51,640 OK, the AI flags it as a performance bottleneck and just 287 00:13:51,920 --> 00:13:55,720 optimizes it away. What the AI doesn't know is that 288 00:13:55,720 --> 00:13:59,480 the delay was put there because an ancient banking API you rely 289 00:13:59,480 --> 00:14:02,720 on crashes if it receives requests too quickly. 290 00:14:02,720 --> 00:14:04,280 Oh. Yikes, right? 291 00:14:04,480 --> 00:14:06,880 The AI generated code will look perfectly clean. 292 00:14:06,920 --> 00:14:10,560 It'll compile, it'll pass your modern tests, but it will subtly 293 00:14:10,560 --> 00:14:13,560 break your revenue pipeline in a way that is incredibly hard to 294 00:14:13,560 --> 00:14:15,600 track down. That is terrifying. 295 00:14:15,640 --> 00:14:18,080 So the vital take away for you, the listener, is this. 296 00:14:18,480 --> 00:14:21,360 As your generation speed increases, your review rigor 297 00:14:21,360 --> 00:14:24,120 must scale up proportionally. The bar doesn't get lowered 298 00:14:24,120 --> 00:14:26,480 because we're moving fast. The bar has to be raised. 299 00:14:26,480 --> 00:14:29,200 The code review bar needs to rise drastically. 300 00:14:29,760 --> 00:14:32,560 You are no longer primarily a writer of code, you are an 301 00:14:32,600 --> 00:14:34,880 auditor of logic. You're the editor in chief. 302 00:14:35,040 --> 00:14:36,800 Exactly. You are the editor in chief. 303 00:14:37,160 --> 00:14:40,760 If an agent generates 10 times more volume, you need to be 10 304 00:14:40,760 --> 00:14:42,760 times more rigorous in your validation. 305 00:14:43,280 --> 00:14:46,880 And to be an editor in chief, you need a totally different set 306 00:14:46,880 --> 00:14:49,760 of tools than a typist. Absolutely. 307 00:14:49,960 --> 00:14:52,280 You don't need a better keyboard, you need a management 308 00:14:52,280 --> 00:14:55,080 dashboard. Which perfectly explains the 309 00:14:55,160 --> 00:14:58,440 absolute explosion in new infrastructure we are seeing 310 00:14:58,440 --> 00:14:59,560 today. It really does. 311 00:14:59,600 --> 00:15:03,080 There is a massive turf war breaking out over who gets to 312 00:15:03,080 --> 00:15:06,760 provide these management tools. The tooling landscape is almost 313 00:15:06,760 --> 00:15:08,680 unrecognizable from even a year ago. 314 00:15:08,760 --> 00:15:11,560 Yeah, how so? Well, it used to be that large 315 00:15:11,560 --> 00:15:14,600 language models were only good at boring technology. 316 00:15:14,600 --> 00:15:17,840 Languages like Python or React that were heavily represented in 317 00:15:17,840 --> 00:15:20,320 their training data. The stuff everyone uses, right? 318 00:15:20,600 --> 00:15:23,080 If you wanted to build a Greenfield alication, meaning 319 00:15:23,080 --> 00:15:26,200 starting a brand new project from scratch with no existing 320 00:15:26,200 --> 00:15:30,160 code using a niche modern framework, the AI was 321 00:15:30,160 --> 00:15:32,760 practically useless. But the sources today show that 322 00:15:32,760 --> 00:15:34,720 limitation is gone. Completely gone. 323 00:15:34,880 --> 00:15:37,880 These models can now extrapolate patterns well enough to be 324 00:15:37,880 --> 00:15:41,280 highly competent with cutting edge, less represented tools. 325 00:15:41,680 --> 00:15:43,880 You aren't boxed into old tech anymore. 326 00:15:43,880 --> 00:15:45,960 Which is incredibly freeing for developers. 327 00:15:46,040 --> 00:15:48,960 It is, and it's why the community is rushing to build 328 00:15:48,960 --> 00:15:50,360 the management layers on top of them. 329 00:15:50,360 --> 00:15:53,760 Like look at the open source clawed code gateway daemon that 330 00:15:53,760 --> 00:15:55,880 dropped today. Let's explain what a demon 331 00:15:55,880 --> 00:15:58,560 actually is in this context, because it's the key to this 332 00:15:58,560 --> 00:16:01,560 whole new workflow. Right, A demon is basically a 333 00:16:01,560 --> 00:16:04,640 program that runs continuously in the background, like a silent 334 00:16:04,640 --> 00:16:06,920 middle manager. Perfect way to describe it. 335 00:16:07,160 --> 00:16:11,120 Previously, if you used an AI coding assistant, you had to sit 336 00:16:11,120 --> 00:16:13,800 at your terminal, prompt it, and wait for it to finish. 337 00:16:13,960 --> 00:16:15,120 It was synchronous. You were. 338 00:16:15,120 --> 00:16:18,480 Tied to the desk. Yes, but this new open source 339 00:16:18,480 --> 00:16:21,400 daemon changes the game. It adds background jobs, slack 340 00:16:21,400 --> 00:16:24,600 integrations, and multi agent team orchestration. 341 00:16:24,680 --> 00:16:28,040 It allows for truly asynchronous agentic work. 342 00:16:28,040 --> 00:16:29,920 So I can just hand off a task and walk away. 343 00:16:29,920 --> 00:16:32,520 Exactly. You can assign the demon a 344 00:16:32,560 --> 00:16:37,040 massive complex refactoring task at 500 PM. 345 00:16:37,560 --> 00:16:39,080 You close your laptop and go home. 346 00:16:39,080 --> 00:16:41,240 While I'm sleeping it's working. Yes. 347 00:16:41,800 --> 00:16:44,680 In the background, the demon spins up multiple different AI 348 00:16:44,680 --> 00:16:48,160 agents, delegates pieces of the task to them, monitors their 349 00:16:48,160 --> 00:16:51,760 progress, and only taps you on the shoulder via a Slack message 350 00:16:51,920 --> 00:16:54,840 when it needs a human editor in chief to review the final draft 351 00:16:54,840 --> 00:16:57,440 the next morning. That is unbelievable, and we are 352 00:16:57,440 --> 00:17:00,320 seeing this exact philosophy echoed at the enterprise level 353 00:17:00,320 --> 00:17:03,040 too. I everywhere like Spine Swarm 354 00:17:03,040 --> 00:17:07,040 launched on Hacker News today. It's a project backed by Y 355 00:17:07,040 --> 00:17:10,119 Combinator that puts AI agents on a visual canvas so you can 356 00:17:10,119 --> 00:17:12,920 watch them collaborate and breakdown tasks in real time. 357 00:17:12,920 --> 00:17:14,800 Which is visually stunning, by the way. 358 00:17:14,920 --> 00:17:20,000 And GitHub itself just updated their Copilot SDK, explicitly 359 00:17:20,000 --> 00:17:23,280 pushing the narrative that execution is the new interface. 360 00:17:23,400 --> 00:17:26,160 That phrase is the defining mantra of this shift. 361 00:17:26,160 --> 00:17:28,200 Yeah, execution is the new interface. 362 00:17:28,200 --> 00:17:29,880 It really sums it up. It does. 363 00:17:29,880 --> 00:17:32,520 We are no longer chatting with computers to get information. 364 00:17:32,520 --> 00:17:34,760 We are delegating execution to systems. 365 00:17:34,840 --> 00:17:36,760 It feels like we are. We're watching the streaming 366 00:17:36,760 --> 00:17:38,240 wars all over again. How do you mean? 367 00:17:38,400 --> 00:17:42,560 But instead of Netflix and HBO spending billions fighting for 368 00:17:42,560 --> 00:17:46,360 eyeballs with prestige television, the biggest tech 369 00:17:46,360 --> 00:17:51,080 giants on Earth are fighting a massive, aggressive war for the 370 00:17:51,080 --> 00:17:54,200 loyalty of software developers. That is exactly what is 371 00:17:54,200 --> 00:17:56,120 happening. The corporate strategy behind 372 00:17:56,120 --> 00:17:59,360 this is ruthless, and we have a story from just a few days ago 373 00:17:59,600 --> 00:18:01,360 that illustrates the stakes perfectly. 374 00:18:01,360 --> 00:18:03,400 Let's hear it. Anthropic announced a new 375 00:18:03,400 --> 00:18:06,280 program. They're giving six months of 376 00:18:06,280 --> 00:18:10,120 free access to their top tier Claude Max model to open source 377 00:18:10,120 --> 00:18:11,800 maintainers. OK, that's generous. 378 00:18:11,800 --> 00:18:14,600 But there's a strict filter. You only qualify if your roject 379 00:18:14,600 --> 00:18:19,120 has over 5000 GitHub stars or a million downloads on MPM. 380 00:18:20,120 --> 00:18:22,240 So they're specifically targeting the foundational 381 00:18:22,240 --> 00:18:24,720 libraries of the Internet, the people building the deep 382 00:18:24,720 --> 00:18:27,120 infrastructure that millions of other applications rely. 383 00:18:27,120 --> 00:18:30,120 On exactly they're targeting the wheat farmers of the digital 384 00:18:30,120 --> 00:18:34,320 economy and within days of that announcement open AI retaliated 385 00:18:34,320 --> 00:18:35,400 of. Course they did. 386 00:18:35,400 --> 00:18:39,800 They announced 6 months of free ChatGPT Pro, which is a $200 a 387 00:18:39,800 --> 00:18:44,040 month value plus access to their specialized Codex models. 388 00:18:44,280 --> 00:18:46,840 And what was their eligibility criteria? 389 00:18:47,360 --> 00:18:50,760 Exact same demographic, the exact same download thresholds. 390 00:18:50,760 --> 00:18:53,520 Wow, they are literally hunting from the exact same list of 391 00:18:53,520 --> 00:18:54,920 developers. They are. 392 00:18:55,080 --> 00:18:59,080 Why are these few thousand open source maintainers so incredibly 393 00:18:59,080 --> 00:19:01,480 valuable to these multibillion dollar companies? 394 00:19:01,480 --> 00:19:04,960 Because of network effects, the strategic read here is 395 00:19:04,960 --> 00:19:07,680 brilliant. If you can get the authors of 396 00:19:07,680 --> 00:19:11,160 foundational libraries using your specific tooling and your 397 00:19:11,160 --> 00:19:15,200 specific agents, your influence ripples through the entire 398 00:19:15,440 --> 00:19:18,480 global software ecosystem. So it's about embedding 399 00:19:18,480 --> 00:19:21,920 themselves in the foundation. Yes, the code they write, the 400 00:19:21,920 --> 00:19:25,000 structural patterns they establish, and the documentation 401 00:19:25,000 --> 00:19:28,800 they generate will inherently be optimized for the AI tool they 402 00:19:28,800 --> 00:19:31,280 used to build it. Oh, that makes so much sense. 403 00:19:31,320 --> 00:19:33,120 Right. If a foundational library is 404 00:19:33,120 --> 00:19:36,360 built using Clawed, downstream developers will find that Clawed 405 00:19:36,360 --> 00:19:39,120 is naturally better at helping them implement that library. 406 00:19:39,480 --> 00:19:41,800 You capture the maintainers today and you capture the 407 00:19:41,800 --> 00:19:43,880 industry tomorrow. That is fascinating. 408 00:19:43,880 --> 00:19:47,400 It's a literal land grab for the foundational architecture of the 409 00:19:47,400 --> 00:19:48,800 next decade. It really is. 410 00:19:49,040 --> 00:19:52,480 And what adds this incredible layer of texture to all of this 411 00:19:52,480 --> 00:19:56,320 corporate warfare is how the developers themselves are 412 00:19:56,320 --> 00:19:59,040 interacting with these tools. It's getting a bit weird 413 00:19:59,040 --> 00:20:01,040 honestly. It is if you scroll through 414 00:20:01,040 --> 00:20:04,560 Reddit today, the community is developing this bizarre, almost 415 00:20:04,560 --> 00:20:06,720 humorous relationship with the agents. 416 00:20:06,880 --> 00:20:11,500 They talk about them like they are quirky, slightly unstable Co 417 00:20:11,500 --> 00:20:14,440 workers. Well, when a tool exhibits this 418 00:20:14,440 --> 00:20:17,360 level of autonomy, human beings will inevitably 419 00:20:17,360 --> 00:20:19,480 anthropomorphize. It we can't help it, you see 420 00:20:19,480 --> 00:20:22,800 posts about developers realizing that AI got stuck because it had 421 00:20:22,960 --> 00:20:25,840 ideas too big for clawed. Like it just got overly 422 00:20:25,840 --> 00:20:28,520 ambitious. Right, like the agent got overly 423 00:20:28,520 --> 00:20:31,600 ambitious about redesigning a system and generated way too 424 00:20:31,600 --> 00:20:34,400 much context and essentially locked itself out of its own 425 00:20:34,400 --> 00:20:35,760 memory. That's hilarious. 426 00:20:35,760 --> 00:20:38,720 But my absolute favorite piece of texture from today is an 427 00:20:38,720 --> 00:20:41,760 anonymous debate post regarding pharmacovigilance. 428 00:20:41,840 --> 00:20:44,120 Oh, I saw this one. Pharmacovigilance being the 429 00:20:44,120 --> 00:20:46,640 highly technical field of monitoring the safety of 430 00:20:46,640 --> 00:20:48,920 pharmaceutical drugs after they've been approved. 431 00:20:48,920 --> 00:20:51,760 Right, very dense, highly regulated work. 432 00:20:51,760 --> 00:20:56,360 Supercritical stuff. So a user set up a blind test 433 00:20:56,360 --> 00:21:01,000 for a PhD level review document. They had Claude review the 434 00:21:01,000 --> 00:21:04,320 document to check for errors. OK, but the twist was that the 435 00:21:04,320 --> 00:21:06,800 document had actually been generated by a different AI 436 00:21:06,800 --> 00:21:10,480 model, and Claude, acting completely blind to the setup, 437 00:21:10,720 --> 00:21:13,880 not only reviewed the science, but actually identified its own 438 00:21:13,920 --> 00:21:16,880 underlying stylistic tells in the text. 439 00:21:17,080 --> 00:21:19,560 It recognized the linguistic fingerprints of its own 440 00:21:19,560 --> 00:21:20,880 foundational model. Yes. 441 00:21:21,320 --> 00:21:24,240 The system is becoming self aware of its own quirks. 442 00:21:24,240 --> 00:21:27,480 It effectively said, I know an AI wrote this because this is 443 00:21:27,480 --> 00:21:29,240 exactly how I would have phrased it. 444 00:21:29,400 --> 00:21:31,680 That's uncanny. It is uncanny and it really 445 00:21:31,680 --> 00:21:34,560 brings us full circle to the core of our deep dive today. 446 00:21:34,560 --> 00:21:36,680 It definitely does. The landscape is moving 447 00:21:36,680 --> 00:21:39,760 incredibly fast, the tools are getting sharper, they're getting 448 00:21:39,760 --> 00:21:42,600 faster, and they are achieving a level of autonomy that is 449 00:21:42,600 --> 00:21:45,160 forcing a complete rewrite of how we interact with the 450 00:21:45,160 --> 00:21:46,280 technology. Absolutely. 451 00:21:46,320 --> 00:21:49,880 So to synthesize all of this for you listening, whether you're 452 00:21:49,880 --> 00:21:52,800 prepping for a strategic board meeting, managing a team of 453 00:21:52,800 --> 00:21:55,440 developers, or you're just insanely curious about the 454 00:21:55,440 --> 00:21:58,720 mechanics of the future, the take away is crystal clear. 455 00:21:59,000 --> 00:22:02,400 The execution era has arrived. The execution era has arrived. 456 00:22:02,400 --> 00:22:04,640 The hammer now swings itself right. 457 00:22:04,760 --> 00:22:08,440 To survive and thrive in this new paradigm, you cannot 458 00:22:08,440 --> 00:22:11,760 abandoned your fundamentals. You have to leverage your deep 459 00:22:11,760 --> 00:22:15,080 background knowledge to direct these autonomous tools. 460 00:22:15,080 --> 00:22:16,520 Background knowledge is everything. 461 00:22:16,520 --> 00:22:20,520 You absolutely must implement strict infrastructure level 462 00:22:20,520 --> 00:22:24,440 safeguards because a caffeinated bulldozer will wreck your live 463 00:22:24,440 --> 00:22:27,400 database if you leave the keys in the ignition. 464 00:22:27,400 --> 00:22:28,960 Like we saw today. Exactly. 465 00:22:29,000 --> 00:22:33,640 And above all, as your capacity to generate work explodes, you 466 00:22:33,640 --> 00:22:37,280 must aggressively raise your standards for reviewing to work. 467 00:22:37,680 --> 00:22:40,560 You are no longer the typist, you are the editor in chief. 468 00:22:40,960 --> 00:22:44,200 Which leaves us with one final, rather heavy implication that we 469 00:22:44,200 --> 00:22:47,360 haven't fully addressed, but that you really need to consider 470 00:22:47,360 --> 00:22:49,600 as you navigate this shift. What's that? 471 00:22:49,840 --> 00:22:53,360 Well, we've established that human expertise, the ability to 472 00:22:53,360 --> 00:22:56,920 act as the editor in chief and catch those subtle catastrophic 473 00:22:56,920 --> 00:22:59,000 mistakes, is more valuable than ever. 474 00:22:59,080 --> 00:23:00,960 Right, Toby and the Reddit user proved that. 475 00:23:01,200 --> 00:23:03,920 Exactly. But right now, the current 476 00:23:03,920 --> 00:23:07,680 generation of experts built that vital intuition by spending 20 477 00:23:07,680 --> 00:23:10,200 years typing out the messy code themselves. 478 00:23:10,320 --> 00:23:11,000 Oh, I see. Where? 479 00:23:11,040 --> 00:23:12,640 You're going. They made the mistakes, they 480 00:23:12,640 --> 00:23:15,320 tracked down the obscure bugs, and they learned the hard way. 481 00:23:15,800 --> 00:23:20,480 If AI completely takes over the raw execution of tasks from day 482 00:23:20,480 --> 00:23:24,120 one, and humans are immediately elevated purely to the role of 483 00:23:24,120 --> 00:23:25,600 reviewers. How do they learn? 484 00:23:25,760 --> 00:23:28,560 Exactly how does the next generation of developers 485 00:23:28,560 --> 00:23:32,440 actually build that foundational expert intuition if they never 486 00:23:32,440 --> 00:23:34,880 have to struggle through the messy execution themselves? 487 00:23:34,880 --> 00:23:37,960 That right there, that is the $1,000,000 question for the next 488 00:23:37,960 --> 00:23:39,840 decade of knowledge work. It really is. 489 00:23:40,000 --> 00:23:42,880 How do you learn to be a master editor if you never had to 490 00:23:42,880 --> 00:23:45,440 suffer through being a terrible writer? 491 00:23:46,520 --> 00:23:48,200 That is your homework to Mulliver today. 492 00:23:48,400 --> 00:23:50,560 Thank you so much for joining us on this deep dive. 493 00:23:50,680 --> 00:23:53,080 It's been great. Keep exploring these wild 494 00:23:53,080 --> 00:23:55,800 shifts, keep questioning the tools you use, and most 495 00:23:55,800 --> 00:23:59,120 importantly, keep learning. We'll catch you on the next one.