1 00:00:00,040 --> 00:00:05,160 Imagine you run a massive Fortune 500 company and you just 2 00:00:05,160 --> 00:00:08,320 fired your entire software development team, not to save 3 00:00:08,320 --> 00:00:13,040 money, but because starting today, every single employee in 4 00:00:13,040 --> 00:00:16,280 your company is generating their own bespoke software 5 00:00:16,280 --> 00:00:18,640 applications just on the fly. Yeah. 6 00:00:18,640 --> 00:00:20,720 Which sounds like absolute chaos. 7 00:00:20,800 --> 00:00:22,720 Exactly. A total management nightmare. 8 00:00:22,960 --> 00:00:25,320 But I mean, looking at the developments from just this past 9 00:00:25,320 --> 00:00:28,240 week, the infrastructure to make this happen isn't science 10 00:00:28,240 --> 00:00:30,200 fiction anymore. It actually exists. 11 00:00:30,200 --> 00:00:32,360 It does. It is a profound structural 12 00:00:32,360 --> 00:00:36,160 change in computing. Today is Friday, April 17th, 13 00:00:36,200 --> 00:00:39,920 2026, and we are tracking a transition that fundamentally 14 00:00:39,920 --> 00:00:41,680 redefines your relationship with software. 15 00:00:42,160 --> 00:00:45,880 The data we're analyzing today from the new Stack Latent Space 16 00:00:45,880 --> 00:00:48,920 Import AI, along with some really fresh survey data from 17 00:00:48,920 --> 00:00:51,760 Stack Overflow, it all points to the exact same conclusion. 18 00:00:51,760 --> 00:00:54,040 It's just that we're moving away from the era of the AI 19 00:00:54,040 --> 00:00:56,320 assistant, right? Precisely, we are entering the 20 00:00:56,320 --> 00:00:59,960 era of the autonomous AI staff. For you, the listener, the 21 00:00:59,960 --> 00:01:02,880 fundamental shift here is moving your role from being a human in 22 00:01:02,880 --> 00:01:04,760 the loop to becoming a human on the loop. 23 00:01:05,080 --> 00:01:06,440 OK, let's unpack this for a second. 24 00:01:06,800 --> 00:01:10,640 Because the human in the loop, that's when you ask an AI to 25 00:01:10,640 --> 00:01:13,680 say, write a specific recipe, you check it's work, and then 26 00:01:13,680 --> 00:01:16,240 you go cook the meal yourself. Right, you are micromanaging the 27 00:01:16,240 --> 00:01:17,160 process. Exactly. 28 00:01:17,560 --> 00:01:21,320 But being on the loop, it feels like we're moving to having an 29 00:01:21,480 --> 00:01:24,280 AI staff completely manage your kitchen. 30 00:01:24,360 --> 00:01:26,400 I mean, they are monitoring the inventory, ordering the 31 00:01:26,400 --> 00:01:29,360 groceries, prepping the stations, doing the dishes, and 32 00:01:29,360 --> 00:01:32,720 you are just overseeing the menu and tasting the final product. 33 00:01:33,120 --> 00:01:36,120 But to actually pull that off, to let software run your kitchen 34 00:01:36,120 --> 00:01:38,920 autonomously, you need insanely capable models. 35 00:01:39,640 --> 00:01:41,560 Which brings us to yesterday's massive launch. 36 00:01:41,560 --> 00:01:45,080 Oh yes, anthropic dropping clawed Opus 4.7. 37 00:01:45,080 --> 00:01:48,200 Yeah, yesterday, April 16th. And the timing of that release 38 00:01:48,200 --> 00:01:50,320 is really the catalyst for this entire discussion. 39 00:01:50,400 --> 00:01:52,960 Late in Space published a breakdown almost immediately, I 40 00:01:52,960 --> 00:01:55,080 think the headline called Opus 4.7. 41 00:01:55,080 --> 00:01:57,320 Literally one step better in every dimension. 42 00:01:57,320 --> 00:01:58,960 Wow. Yeah, they highlighted the 43 00:01:58,960 --> 00:02:02,400 significantly upgraded vision processing, the expanded state, 44 00:02:02,400 --> 00:02:05,880 full memory and just a drastic improvement in complex 45 00:02:05,880 --> 00:02:08,440 instruction following over version 4.6. 46 00:02:08,720 --> 00:02:11,200 And developers, I mean they didn't even wait for the dust to 47 00:02:11,200 --> 00:02:13,880 settle. Simon Willison pushed an update 48 00:02:13,880 --> 00:02:18,520 to his philanthropic plug in version Viro .25 on the exact 49 00:02:18,520 --> 00:02:20,760 same day. He moves incredibly fast. 50 00:02:20,800 --> 00:02:22,920 He does. And he didn't just, you know, 51 00:02:23,080 --> 00:02:26,400 swap the API endpoint. He unlocked the new reasoning 52 00:02:26,400 --> 00:02:28,960 parameters, specifically that thinking effort setting. 53 00:02:29,320 --> 00:02:31,680 He allowed users to Max it out at XI. 54 00:02:32,960 --> 00:02:35,520 Yeah, plus he added these boolean toggles for thinking 55 00:02:35,520 --> 00:02:38,880 Display and thinking adaptive. He even increased the default 56 00:02:38,880 --> 00:02:40,760 Max tokens to the absolute maximum. 57 00:02:40,760 --> 00:02:43,000 That thinking adaptive parameter, though, That is 58 00:02:43,000 --> 00:02:45,680 particularly revealing about where the technology is heading 59 00:02:45,680 --> 00:02:46,880 right now. How so? 60 00:02:46,960 --> 00:02:50,560 Well, we are no longer forcing a model to use a static amount of 61 00:02:50,560 --> 00:02:53,360 compute for every query. When you toggle Thinking 62 00:02:53,360 --> 00:02:56,640 Adaptive, you're allowing the model to evaluate the perplexity 63 00:02:56,640 --> 00:02:58,640 of your prompt in real time. Oh, I see. 64 00:02:58,640 --> 00:03:00,160 So it judges how hard the question is. 65 00:03:00,200 --> 00:03:03,000 Exactly. If it hits a comp complex high 66 00:03:03,000 --> 00:03:07,280 entropy logic puzzle, it dynamically allocates a massive 67 00:03:07,280 --> 00:03:10,440 budget of hidden reasoning tokens to literally think 68 00:03:10,440 --> 00:03:13,560 through the problem before generating a single word of 69 00:03:13,560 --> 00:03:15,760 output. But if the prompt is simple, it 70 00:03:15,760 --> 00:03:18,920 just bypasses that layer entirely to save you latency and 71 00:03:18,920 --> 00:03:20,960 cost. But here is where it gets really 72 00:03:20,960 --> 00:03:23,880 interesting though. So you have Opus 4.7, this 73 00:03:23,880 --> 00:03:27,960 massive, highly sophisticated frontier cloud model that can 74 00:03:27,960 --> 00:03:29,720 dynamically budget its own reasoning. 75 00:03:30,080 --> 00:03:33,840 And yet Simon Willison runs what calls the Pelican benchmark. 76 00:03:35,040 --> 00:03:37,520 Yes, the Pelican bench. It's just so funny. 77 00:03:37,520 --> 00:03:41,440 It's his quirky informal image generation test where he asks a 78 00:03:41,440 --> 00:03:44,080 model to generate an image of a Pelican riding a bicycle. 79 00:03:44,080 --> 00:03:47,520 Which tests spatial reasoning, anatomical alignment, and prompt 80 00:03:47,520 --> 00:03:50,280 adherence for just totally absurd concepts. 81 00:03:50,280 --> 00:03:52,040 Exactly. So he runs this benchmark on the 82 00:03:52,040 --> 00:03:56,040 powerhouse Opus 4.7 and then he runs it on a local Alibaba model 83 00:03:56,040 --> 00:03:57,680 running on just a standard laptop. 84 00:03:57,680 --> 00:04:02,200 The Quinn 3.635 BA 3B and the local laptop bound Quin model 85 00:04:02,400 --> 00:04:05,800 actually generated A noticeably better Pelican than Opus 4.7. 86 00:04:05,800 --> 00:04:08,800 It did, and what's fascinating here is that this result forces 87 00:04:08,800 --> 00:04:11,720 a complete reevaluation of model deployment strategies. 88 00:04:12,120 --> 00:04:15,320 To understand why a local model outperformed a frontier model on 89 00:04:15,320 --> 00:04:18,079 spatial reasoning, you really have to look at the architecture 90 00:04:18,079 --> 00:04:21,440 indicated by that 35 BA 3B naming convention. 91 00:04:21,560 --> 00:04:24,840 The MO we architecture. Yes, it is a mixture of experts 92 00:04:24,840 --> 00:04:27,560 model. It has 35 billion parameters in 93 00:04:27,560 --> 00:04:31,400 total, but only three billionaire active for any given 94 00:04:31,400 --> 00:04:33,160 token. I know we hear about mixture 95 00:04:33,160 --> 00:04:36,000 experts a lot, but the delta here is fascinating. 96 00:04:36,480 --> 00:04:41,200 Opus 4.7 has been heavily fine-tuned and RLHFD. 97 00:04:41,680 --> 00:04:44,720 You know, reinforcement learning from human feedback for safety, 98 00:04:45,080 --> 00:04:48,040 coding standards, and rigorous logical reasoning. 99 00:04:48,040 --> 00:04:50,440 Yes, very heavily guarded. Which creates a sort of 100 00:04:50,440 --> 00:04:53,120 generalized intelligence tax when you ask it to draw a 101 00:04:53,120 --> 00:04:56,120 Pelican on a bicycle. That request is filtered through 102 00:04:56,120 --> 00:04:58,000 layers of enterprise safety and logic checks. 103 00:04:58,560 --> 00:05:02,440 But the Quinn model, it's MO E router instantly identifies the 104 00:05:02,440 --> 00:05:05,760 spatial rendering request and just fires it directly to a 105 00:05:05,760 --> 00:05:09,400 highly specialized 3 billion parameter subnetwork that is 106 00:05:09,400 --> 00:05:12,720 aggressively over indexed on visual alignments and absurdity. 107 00:05:12,720 --> 00:05:14,160 It doesn't pay the generalized tax at. 108 00:05:14,240 --> 00:05:16,440 All exactly. It just executes the specific 109 00:05:16,440 --> 00:05:18,040 task. It's like, well, it's like 110 00:05:18,040 --> 00:05:20,880 hiring a Michelin star chef to make you a quick Taco when the 111 00:05:20,880 --> 00:05:23,000 street vendor on the corner actually does it better and 112 00:05:23,000 --> 00:05:25,480 cheaper. That's a perfect analogy, and 113 00:05:25,480 --> 00:05:27,920 the implication of that architectural advantage is 114 00:05:27,960 --> 00:05:30,840 profound. It proves that the human on the 115 00:05:30,840 --> 00:05:34,000 loop kitchen cannot run efficiently if you wrote every 116 00:05:34,000 --> 00:05:36,600 single task to a massive cloud model. 117 00:05:36,600 --> 00:05:40,360 Yeah, it's just too expensive. Right cost optimization and task 118 00:05:40,360 --> 00:05:42,640 latency. They demand a hybrid approach. 119 00:05:43,080 --> 00:05:45,960 You need the heavy frontier models for complex back end 120 00:05:45,960 --> 00:05:49,480 reasoning, but you need these highly specialized, incredibly 121 00:05:49,480 --> 00:05:52,520 efficient local models for narrow execution. 122 00:05:52,960 --> 00:05:55,840 But wait, how do we decide which model gets which task? 123 00:05:56,280 --> 00:05:59,160 Because there was a fatal flaw in relying on an army of 124 00:05:59,160 --> 00:06:02,360 different models, especially scrappy local ones, for 125 00:06:02,360 --> 00:06:04,960 enterprise tasks. And that flaw is trust. 126 00:06:05,240 --> 00:06:07,920 Yes, trust. If we are dynamically routing 127 00:06:07,920 --> 00:06:11,560 complex tasks down to local 3 billion parameter specialists 128 00:06:11,560 --> 00:06:14,400 instead of Claude, we suddenly have a zero trust environment. 129 00:06:15,120 --> 00:06:17,800 I mean, if a local Quinn model decides to rewrite a core 130 00:06:17,800 --> 00:06:21,240 database, or an autonomous agent initiates a financial 131 00:06:21,240 --> 00:06:23,920 transaction, how do we know it hasn't been hijacked? 132 00:06:24,160 --> 00:06:26,560 We can't just cross our fingers and hope the local model isn't 133 00:06:26,560 --> 00:06:29,320 compromised. We can, and the industry 134 00:06:29,320 --> 00:06:32,800 recognizes that vulnerability, which is why the second major 135 00:06:32,800 --> 00:06:35,960 development in our sources tackles identity verification 136 00:06:35,960 --> 00:06:38,200 head on. The new Stack reports that 137 00:06:38,200 --> 00:06:41,560 Anthropic is rolling out a dedicated identity verification 138 00:06:41,560 --> 00:06:44,840 layer for Claude, and they are being explicitly clear about 139 00:06:44,840 --> 00:06:46,640 this. This is not for casual consumer 140 00:06:46,640 --> 00:06:47,880 use. No, definitely not. 141 00:06:47,880 --> 00:06:50,720 This is engineered specifically for high stakes enterprise 142 00:06:50,720 --> 00:06:54,120 environments to establish verifiable audit trails. 143 00:06:54,120 --> 00:06:56,640 Let's drill into the how of that, because audit trails can 144 00:06:56,640 --> 00:06:59,600 sound like, I don't know, middle management buzzwords, right? 145 00:06:59,840 --> 00:07:02,760 We aren't talking about creating AI bouncers that check a 146 00:07:02,760 --> 00:07:04,400 password at a door. No, not at. 147 00:07:04,600 --> 00:07:07,360 All we are talking about cryptographic software identity. 148 00:07:07,880 --> 00:07:11,640 When an autonomous clawed agent touches A repository or triggers 149 00:07:11,640 --> 00:07:16,000 an API, it isn't just leaving a text log, it is attaching A 150 00:07:16,000 --> 00:07:19,360 cryptographically signed execution token to the payload. 151 00:07:19,360 --> 00:07:21,520 Yes. So if an enterprise security 152 00:07:21,520 --> 00:07:24,640 team needs to know exactly which model altered a piece of code, 153 00:07:24,960 --> 00:07:27,000 they can verify the signature mathematically. 154 00:07:27,160 --> 00:07:30,400 And that level of mathematical certainty is really the only way 155 00:07:30,400 --> 00:07:35,200 to calm enterprise security teams, who, rightfully so, view 156 00:07:35,200 --> 00:07:38,000 agentic AI as a terrifying black box. 157 00:07:38,000 --> 00:07:39,880 Oh, totally. But identity verification is 158 00:07:39,880 --> 00:07:42,520 just the defensive baseline. We are also seeing the 159 00:07:42,520 --> 00:07:45,160 deployment of models specifically designed for 160 00:07:45,160 --> 00:07:47,560 offensive security, right? Claude Mythos. 161 00:07:47,600 --> 00:07:51,200 Yes, the notes highlight Claude Mythos, which is a cybersecurity 162 00:07:51,200 --> 00:07:54,160 focused model. The UKAI Safety Institute 163 00:07:54,200 --> 00:07:56,480 recently published an independent evaluation of 164 00:07:56,480 --> 00:07:59,800 Mythos, concluding it is quote exceptionally effective at 165 00:07:59,800 --> 00:08:02,520 identifying 0 day vulnerabilities and logic. 166 00:08:02,520 --> 00:08:04,840 Flaws and Open AI is pushing into that exhibit. 167 00:08:04,920 --> 00:08:09,680 Exact same space with GPT 5.4 Cyber, which they explicitly 168 00:08:09,680 --> 00:08:12,680 market as being fine-tuned for defensive cybersecurity. 169 00:08:12,680 --> 00:08:14,800 The arms race is definitely on. It is. 170 00:08:14,960 --> 00:08:17,400 We also have partnerships forming like the Cloudflare 171 00:08:17,400 --> 00:08:20,240 Agent Cloud integrating with Open AI, and that is 172 00:08:20,240 --> 00:08:23,560 specifically to offer AI agents that come with hard, verifiable 173 00:08:23,560 --> 00:08:25,360 security guarantees right out-of-the-box. 174 00:08:25,360 --> 00:08:28,320 And the underlying philosophy here is captured perfectly by a 175 00:08:28,320 --> 00:08:30,280 quote from Simon Willison in our sources. 176 00:08:30,280 --> 00:08:32,559 Oh I love this quote. He notes cybersecurity looks 177 00:08:32,559 --> 00:08:35,200 like proof of work now. Yeah, what he means is that the 178 00:08:35,200 --> 00:08:38,480 traditional heuristic models of cybersecurity are essentially 179 00:08:38,480 --> 00:08:41,919 dead. In an agentic ecosystem, trust 180 00:08:41,919 --> 00:08:44,920 is established by a model's ability to systematically break 181 00:08:44,920 --> 00:08:47,640 things. The models that can autonomously 182 00:08:47,640 --> 00:08:51,000 discover vulnerabilities and execute complex adversarial 183 00:08:51,000 --> 00:08:54,520 attacks are really the only ones we trust to write the defense 184 00:08:54,520 --> 00:08:58,920 protocols. Import AI Issue 453 dedicated a 185 00:08:58,920 --> 00:09:01,760 massive deep dive to this exact concept. 186 00:09:02,200 --> 00:09:05,440 You have to subject these agentic systems to relentless 187 00:09:05,440 --> 00:09:08,520 adversarial testing just to understand their attack 188 00:09:08,520 --> 00:09:10,360 surfaces. I have to push back on this a 189 00:09:10,360 --> 00:09:11,680 little bit though. OK, go ahead. 190 00:09:11,920 --> 00:09:15,840 When we talk about cryptographic execution tokens, adversarial 191 00:09:15,840 --> 00:09:18,920 proof of work, and strict enterprise identity layers, 192 00:09:19,200 --> 00:09:22,680 doesn't this entirely kill the Wild West frictionless fun of 193 00:09:22,680 --> 00:09:24,800 generative AI? It does change the vibe, 194 00:09:24,800 --> 00:09:26,520 certainly. I mean, the whole appeal of this 195 00:09:26,520 --> 00:09:28,520 technology for the last few years was that it was 196 00:09:28,520 --> 00:09:31,280 unconstrained. You just type a prompt and watch 197 00:09:31,280 --> 00:09:33,880 it run. Are we just creating AI bouncers 198 00:09:33,880 --> 00:09:36,080 to check the ID's of other AI models? 199 00:09:36,200 --> 00:09:38,880 Now we're burying it in bureaucratic security protocols. 200 00:09:39,040 --> 00:09:42,120 I understand the nostalgia for that frictionless environment, I 201 00:09:42,120 --> 00:09:44,920 really do. But consumer experimentation and 202 00:09:44,920 --> 00:09:47,480 enterprise infrastructure operate under completely 203 00:09:47,480 --> 00:09:50,040 different physics. Yeah, unconstrained generation 204 00:09:50,040 --> 00:09:52,720 is fantastic if you are brainstorming marketing copy. 205 00:09:53,360 --> 00:09:55,840 But if you are a healthcare provider using an autonomous 206 00:09:55,840 --> 00:09:59,600 agent to route patient data or a financial institution updating 207 00:09:59,600 --> 00:10:03,040 transaction ledgers, unconstrained is basically 208 00:10:03,040 --> 00:10:05,880 synonymous with uninsurable. Wow, uninsurable. 209 00:10:06,240 --> 00:10:09,560 The why behind these rigid protocols is sheer survival. 210 00:10:10,160 --> 00:10:13,960 When models execute tasks autonomously, a single 211 00:10:13,960 --> 00:10:18,160 hallucination or hijacked prompt can corrupt an entire database. 212 00:10:18,720 --> 00:10:21,560 If we connect this to the bigger picture, you just can't have a 213 00:10:21,560 --> 00:10:23,280 genetic workflows without identity. 214 00:10:23,480 --> 00:10:26,240 You cannot scale them without replacing blind trust with 215 00:10:26,240 --> 00:10:29,880 cryptographic verification. Fairpoint, if the AI kitchen 216 00:10:29,880 --> 00:10:32,480 staff is going to operate the deep fryer without me watching, 217 00:10:32,480 --> 00:10:35,240 I want mathematical proof that they know this AC protocol. 218 00:10:35,480 --> 00:10:38,880 So we solve the routing problem by mixing cloud and local 219 00:10:38,880 --> 00:10:41,080 models. We solve the trust problem with 220 00:10:41,080 --> 00:10:43,280 cryptographic identity and adversarial testing. 221 00:10:43,760 --> 00:10:46,680 But that brings us to the communication bottleneck. 222 00:10:46,680 --> 00:10:48,960 A big one. Yeah, if I have a secure Opus 223 00:10:48,960 --> 00:10:52,680 model, a secure Quinn model, secure quad Methos model, all 224 00:10:52,680 --> 00:10:55,720 working on my behalf, how do they actually share state? 225 00:10:55,720 --> 00:10:58,640 How do they pass data back and forth without me basically 226 00:10:58,640 --> 00:11:00,680 acting as the middleman? And that bottleneck was 227 00:11:00,680 --> 00:11:04,560 addressed head on at the MCP Summit in New York City on April 228 00:11:04,560 --> 00:11:08,160 16th, right? The Model Context Protocol, or 229 00:11:08,160 --> 00:11:11,440 MCP, has been simmering in the background for a while, but the 230 00:11:11,440 --> 00:11:14,680 new Stacks coverage of the Summit signals a massive turning 231 00:11:14,680 --> 00:11:17,240 point. Clara Liguori, a senior 232 00:11:17,240 --> 00:11:20,680 principal software engineer at AWS, delivered a presentation 233 00:11:20,680 --> 00:11:23,760 that functionally committed Amazon to MCP as the universal 234 00:11:23,760 --> 00:11:26,120 standard. This is such a critical shift 235 00:11:26,240 --> 00:11:28,680 because initially MCP was designed just for tool calling, 236 00:11:28,720 --> 00:11:29,200 right? Exactly. 237 00:11:29,200 --> 00:11:32,560 It was the protocol on LLM used to securely query a local 238 00:11:32,560 --> 00:11:34,600 database. Or, you know, ping a weather 239 00:11:34,600 --> 00:11:38,080 API. But the summit positioned MCP as 240 00:11:38,080 --> 00:11:39,960 an agent to agent communication standard. 241 00:11:40,240 --> 00:11:42,680 And to understand why that matters we have to look at how 242 00:11:42,680 --> 00:11:46,200 these models used to interact. Historically, if one model 243 00:11:46,200 --> 00:11:48,960 needed to pass work to another, you basically had to have model 244 00:11:49,040 --> 00:11:52,920 A spit out raw text and model B had to scrape and interpret that 245 00:11:52,920 --> 00:11:55,960 text. Which was incredibly fragile and 246 00:11:55,960 --> 00:11:59,920 prone to massive context loss. MCP changes the underlying 247 00:11:59,920 --> 00:12:03,080 mechanics completely. It provides a universal Jason 248 00:12:03,080 --> 00:12:07,160 schema for context injection and tool calling it standardizes the 249 00:12:07,160 --> 00:12:08,800 payload. So it's not guessing anymore. 250 00:12:09,080 --> 00:12:12,240 Right, so a local kin model working on a visual layout 251 00:12:12,240 --> 00:12:16,760 doesn't just pass text to Opus 4.7, it passes a structured 252 00:12:16,760 --> 00:12:20,800 machine readable semantic state directly into Opus context 253 00:12:20,800 --> 00:12:22,720 window. They share a unified 254 00:12:22,720 --> 00:12:25,000 understanding of the tasks current environment. 255 00:12:25,240 --> 00:12:27,640 So what does this all mean? If these models are all 256 00:12:27,640 --> 00:12:31,320 different workers in a factory, MCP is like the universal UN 257 00:12:31,320 --> 00:12:33,960 translation headset that lets them seamlessly talk to each 258 00:12:33,960 --> 00:12:36,880 other to finish your project. That is exactly what it is and 259 00:12:36,880 --> 00:12:38,320 we are already seeing this in the wild. 260 00:12:38,440 --> 00:12:41,480 Cloud Code natively supports this multi agent MCP 261 00:12:41,480 --> 00:12:44,480 coordination right now and the developer community is 262 00:12:44,480 --> 00:12:46,360 aggressively upscaling to match it. 263 00:12:46,720 --> 00:12:50,080 GitHub just dropped a series of secure code game challenges 264 00:12:50,080 --> 00:12:54,000 focused entirely on a AI security vulnerabilities. 265 00:12:54,000 --> 00:12:56,920 Basically teaching how these models interact over these 266 00:12:56,920 --> 00:13:00,400 protocols without exposing vectors for injection attacks. 267 00:13:00,400 --> 00:13:03,000 Right, exactly. And over 10,000 developers have 268 00:13:03,000 --> 00:13:04,320 already cleared those challenges. 269 00:13:04,320 --> 00:13:07,320 That's a lot of buy in. It is the rapid adoption by 270 00:13:07,320 --> 00:13:10,320 developers highlights that standardized communication was 271 00:13:10,320 --> 00:13:13,760 really the final missing piece required to unlock the ultimate 272 00:13:13,760 --> 00:13:16,000 benefit for the listener. Which is customized personal 273 00:13:16,000 --> 00:13:18,600 tools, yes. We have highly capable models 274 00:13:18,600 --> 00:13:21,680 for reasoning and execution. We have the cryptographic 275 00:13:21,680 --> 00:13:25,880 identity layers to secure them, and now with MCP we have the 276 00:13:25,880 --> 00:13:28,840 universal schema allowing them to coordinate autonomously. 277 00:13:29,120 --> 00:13:31,560 When you stack those 3 pillars together, the resulting 278 00:13:31,560 --> 00:13:34,480 structure changes the very definition of software itself. 279 00:13:34,680 --> 00:13:37,160 Which brings us to the core thesis of our deep dive today, 280 00:13:37,440 --> 00:13:41,240 the rise of Personal software. On April 15th, The New Stack 281 00:13:41,240 --> 00:13:44,400 published a piece arguing that the ultimate end game of AI 282 00:13:44,560 --> 00:13:48,160 isn't about helping startups build the next massive 1,000,000 283 00:13:48,160 --> 00:13:50,040 user saws application. No, not at. 284 00:13:50,200 --> 00:13:52,960 All the end game is eliminating the overhead of scale, 285 00:13:52,960 --> 00:13:56,520 deployment and maintenance so entirely that you, the 286 00:13:56,520 --> 00:13:59,480 individual listener, can build complex software applications 287 00:13:59,840 --> 00:14:02,400 just for yourself. And the economic implications of 288 00:14:02,400 --> 00:14:04,320 personal software are just staggering. 289 00:14:04,480 --> 00:14:07,640 Traditional software development is defined by compromises. 290 00:14:07,760 --> 00:14:09,640 Always. You have to build features that 291 00:14:09,640 --> 00:14:12,840 appeal to a mass market to justify the enormous cost of 292 00:14:12,840 --> 00:14:15,680 engineering, deploying and maintaining the code base. 293 00:14:16,240 --> 00:14:19,640 But if a coordinated team of AI agents can write, secure and 294 00:14:19,640 --> 00:14:23,600 deploy an application at near 0 marginal cost, the need for mass 295 00:14:23,600 --> 00:14:26,440 market appeal simply vanishes. Yeah, the sources site, this 296 00:14:26,640 --> 00:14:29,680 brilliant example of a GitHub engineer who used the GitHub 297 00:14:29,680 --> 00:14:33,560 Copilot CLI to build a hyper specific organizational command 298 00:14:33,560 --> 00:14:36,360 center, right. This wasn't a product designed 299 00:14:36,360 --> 00:14:38,880 for onboarding new users and no marketing site. 300 00:14:39,240 --> 00:14:43,360 It was a tool built to perfectly map to the exact neurological 301 00:14:43,360 --> 00:14:45,880 quirks of that single engineers workflow. 302 00:14:46,240 --> 00:14:48,360 It is software with an audience of 1. 303 00:14:48,640 --> 00:14:51,520 And we are seeing major enterprise platforms tear their 304 00:14:51,520 --> 00:14:55,040 own products down to the studs just to prepare for this shift. 305 00:14:55,040 --> 00:14:58,000 Oh, like Notion? Yes, if you listen to the Latent 306 00:14:58,000 --> 00:15:00,540 Space podcast from earlier this week featuring Notions Co 307 00:15:00,540 --> 00:15:03,920 founder Simon Last and their head of AI, Sarah Saks, they 308 00:15:03,920 --> 00:15:07,080 detail what they call the software factory Future. 309 00:15:07,120 --> 00:15:09,720 Yeah. To prepare for a world where AI 310 00:15:09,720 --> 00:15:13,320 agents build personal software, Notion went through five 311 00:15:13,320 --> 00:15:17,160 complete architecture rebuilds and utilized over 100 internal 312 00:15:17,160 --> 00:15:18,600 tools. I've rebuilds. 313 00:15:18,600 --> 00:15:21,040 That is wild. It's a massive undertaking. 314 00:15:21,360 --> 00:15:24,720 I mean, you can't just slap an AI chatbot on top of a legacy 315 00:15:24,720 --> 00:15:26,560 database and call it a software factory. 316 00:15:26,800 --> 00:15:30,200 If I want AI Agent to instantly spin up a custom dashboard for 317 00:15:30,200 --> 00:15:33,360 my project, the underlying architecture has to support that 318 00:15:33,360 --> 00:15:36,600 dynamic manipulation natively. That is exactly what Notion 319 00:15:36,600 --> 00:15:39,200 realized. They had to transition away from 320 00:15:39,200 --> 00:15:44,040 rigid static data schemas and build a hyper flexible block 321 00:15:44,040 --> 00:15:46,880 based graph architecture. They had to expose their 322 00:15:46,880 --> 00:15:49,800 primitive API's so that autonomous agents could 323 00:15:49,800 --> 00:15:53,440 programmatically assemble, manipulate and tear down user 324 00:15:53,440 --> 00:15:57,360 interfaces on the fly without corrupting the underlying data 325 00:15:57,360 --> 00:15:59,000 layer. They basically rebuilt their 326 00:15:59,000 --> 00:16:02,640 entire foundation so that the AI AI could be the developer. 327 00:16:02,800 --> 00:16:05,800 Because when the AI is the developer, the signal to noise 328 00:16:05,800 --> 00:16:07,960 ratio hits absolute zero. Yes. 329 00:16:08,240 --> 00:16:09,960 Think about how software usually works. 330 00:16:10,120 --> 00:16:12,720 You have a product manager guessing what a user might need, 331 00:16:12,720 --> 00:16:16,440 a developer trying to interpret the product managers spec, and 332 00:16:16,440 --> 00:16:19,280 then the actual user who just has to adapt their workflow to 333 00:16:19,280 --> 00:16:21,000 whatever the developer finally ships. 334 00:16:21,000 --> 00:16:22,960 Which is so much friction. Exactly. 335 00:16:23,800 --> 00:16:26,400 With personal software, the operator, the user, and the 336 00:16:26,400 --> 00:16:28,680 domain expert are all the exact same person. 337 00:16:28,720 --> 00:16:31,200 It's you. You become the CEO of your own 338 00:16:31,200 --> 00:16:34,640 software factory, and this transition is actively being 339 00:16:34,640 --> 00:16:36,840 measured right now. By the Stack Overflow survey. 340 00:16:37,080 --> 00:16:40,600 Yes, the survey that dropped on the 15th specifically asked 341 00:16:40,600 --> 00:16:43,440 developers to classify their current daily workflows. 342 00:16:43,760 --> 00:16:46,080 Are they a human in the loop or a human on the loop? 343 00:16:46,200 --> 00:16:49,160 And the results. The metrics show a massive 344 00:16:49,160 --> 00:16:52,760 accelerating migration from developers identifying as a 345 00:16:52,760 --> 00:16:56,000 human in the loop writing the code to a human on the loop 346 00:16:56,040 --> 00:16:57,960 orchestrating the agents that write the code. 347 00:16:58,040 --> 00:17:00,200 And if you are stepping into that human on the loop 348 00:17:00,200 --> 00:17:03,520 management role this week, we actually have an incredibly 349 00:17:03,520 --> 00:17:06,440 actionable pro tip pulled directly from the developer 350 00:17:06,440 --> 00:17:09,119 dispatches in our sources. This is a really good one. 351 00:17:09,160 --> 00:17:11,680 Yeah, so if you are running a genetic clawed code session 352 00:17:11,680 --> 00:17:14,119 specifically where you have turned off the safety 353 00:17:14,119 --> 00:17:18,040 confirmations to let the AI act completely autonomously, you 354 00:17:18,040 --> 00:17:21,560 need to implement a strict oversight rule in your claw EE 355 00:17:21,560 --> 00:17:24,920 dot MD file. The rule is simple, but it is 356 00:17:24,920 --> 00:17:27,520 critical for maintaining architectural integrity. 357 00:17:27,760 --> 00:17:30,920 You instruct the model that before executes any action that 358 00:17:30,920 --> 00:17:34,320 touches more than two files simultaneously, it must write a 359 00:17:34,320 --> 00:17:36,720 one line intense summary as a code comment. 360 00:17:36,960 --> 00:17:40,320 Let's explain why that specific mechanism is so powerful. 361 00:17:41,400 --> 00:17:44,880 By forcing the model to articulate its intent before it 362 00:17:44,880 --> 00:17:48,040 executes a multi file change, you are essentially functioning 363 00:17:48,160 --> 00:17:50,080 as your own enterprise security team. 364 00:17:50,200 --> 00:17:52,920 Exactly. First off, you force the LLM to 365 00:17:52,920 --> 00:17:55,360 commit its reasoning to the context window, which 366 00:17:55,360 --> 00:17:58,120 significantly reduces the chance of hallucinations during 367 00:17:58,120 --> 00:18:00,560 execution. It has to think before it acts. 368 00:18:01,240 --> 00:18:05,040 Second, you're generating a machine readable audit trail. 369 00:18:05,440 --> 00:18:08,560 Because the intent summaries are uniformly formatted, they are 370 00:18:08,560 --> 00:18:10,640 highly graphical for post session review. 371 00:18:10,720 --> 00:18:12,800 It makes oversight incredibly easy. 372 00:18:12,840 --> 00:18:14,960 Totally. When you step back into your 373 00:18:14,960 --> 00:18:17,880 oversight role at the end of the day, you can run a simple search 374 00:18:17,880 --> 00:18:21,920 query and instantly review every major architectural decision the 375 00:18:21,920 --> 00:18:23,720 autonomous agents made while you were away. 376 00:18:24,120 --> 00:18:26,760 You maintain total visibility over the factory floor. 377 00:18:26,920 --> 00:18:30,120 It is a perfect example of how the human on the loop interacts 378 00:18:30,120 --> 00:18:32,640 with the system. You aren't writing the code, but 379 00:18:32,640 --> 00:18:35,240 you are strictly defining the operational parameters. 380 00:18:35,320 --> 00:18:38,920 So if I can instantly spin up a perfectly customized app just 381 00:18:38,920 --> 00:18:42,440 for my own daily workflow, does this mean the death of mass 382 00:18:42,440 --> 00:18:45,000 market sauce apps? Well, this raises an important 383 00:18:45,000 --> 00:18:46,760 question about the future of the industry. 384 00:18:46,760 --> 00:18:49,680 The traditional mass market sauce model will face an 385 00:18:49,680 --> 00:18:51,680 existential threat at the interface level. 386 00:18:51,680 --> 00:18:55,320 Absolutely right. A generic off the shelf project 387 00:18:55,320 --> 00:18:58,880 management tool simply cannot compete with a bespoke tool 388 00:18:58,880 --> 00:19:02,320 tailored to perfectly match your specific cognitive habits. 389 00:19:02,480 --> 00:19:05,480 It's no contest. However, mass market saws won't 390 00:19:05,480 --> 00:19:09,360 disappear completely, it will be forced to pivot and retreat down 391 00:19:09,360 --> 00:19:11,280 the stack. What do you mean Companies will 392 00:19:11,280 --> 00:19:15,480 pivot away from selling user interfaces and focus entirely on 393 00:19:15,480 --> 00:19:18,640 selling highly secure, heavily compliant back end 394 00:19:18,640 --> 00:19:21,480 infrastructure. The API's, the data lakes, the 395 00:19:21,480 --> 00:19:23,640 raw materials. The back end will remains 396 00:19:23,640 --> 00:19:26,840 standardized but the front end interface and logic will be 397 00:19:26,840 --> 00:19:30,000 generated on the family as hyper personalized software. 398 00:19:30,000 --> 00:19:31,760 Wow. OK, to quickly bring this all 399 00:19:31,760 --> 00:19:33,760 together for you, we were watching a fundamental 400 00:19:33,760 --> 00:19:36,520 reorganization of computing. We started the week with 401 00:19:36,520 --> 00:19:40,160 Anthropic launching Opus 4.7, only to learn that specialized 402 00:19:40,160 --> 00:19:44,280 local models like Quinn 3.6 can wrap specific spatial tasks 403 00:19:44,440 --> 00:19:47,240 faster and better. We saw the deployment of AI 404 00:19:47,240 --> 00:19:50,840 bouncers, checking ID's, cryptographic identity layers 405 00:19:51,120 --> 00:19:54,360 and adversarial proof of work to ensure we can actually trust 406 00:19:54,360 --> 00:19:55,920 these agents to operate securely. 407 00:19:55,920 --> 00:19:58,840 Crucial infrastructure. And we watched AWS cement the 408 00:19:58,840 --> 00:20:02,280 MCP translation headset standard, giving these disparate 409 00:20:02,280 --> 00:20:05,200 models a unified schema to communicate seamlessly. 410 00:20:06,000 --> 00:20:08,120 And the result of all that infrastructure is that you are 411 00:20:08,120 --> 00:20:11,240 no longer just a user of software, you are the domain 412 00:20:11,240 --> 00:20:14,960 expert overseeing a vast, capable digital workforce 413 00:20:14,960 --> 00:20:16,800 building tools specifically for you. 414 00:20:17,160 --> 00:20:20,800 It is an incredibly empowering position to be in, but as we 415 00:20:20,800 --> 00:20:24,040 transition into this era of hyper personalized software 416 00:20:24,040 --> 00:20:27,480 tailored exactly to our own brains and workflows, I want to 417 00:20:27,480 --> 00:20:29,040 leave you with a final thought to Mull over. 418 00:20:29,640 --> 00:20:32,360 If we all build custom dashboards, unique workflows, 419 00:20:32,360 --> 00:20:35,520 and bespoke interfaces, do we lose the shared language? 420 00:20:35,600 --> 00:20:37,400 Language of work? Oh wow, When you and your 421 00:20:37,400 --> 00:20:40,600 colleagues no longer use the same standardized tools, how 422 00:20:40,600 --> 00:20:42,080 will we collaborate in the future? 423 00:20:42,200 --> 00:20:44,600 That is, yeah, that is a fascinating problem. 424 00:20:45,000 --> 00:20:49,000 If everyone is the head chef of their own completely customized 425 00:20:49,120 --> 00:20:52,080 AI run kitchen, how do we ever coordinate a cook a meal 426 00:20:52,080 --> 00:20:54,920 together, somebody to chew on as you start spinning up your own 427 00:20:54,920 --> 00:20:56,480 autonomous software factory this weekend. 428 00:20:56,760 --> 00:20:59,000 Thank you so much for joining us for this deep dive. 429 00:20:59,000 --> 00:21:00,000 We will see you next time.