WEBVTT 00:00:00.000 --> 00:00:02.980 Imagine a world where millions of lines of C 00:00:02.980 --> 00:00:07.700 and C++ code just vanish. Right. We're talking 00:00:07.700 --> 00:00:09.980 about one of the biggest tech giants trying to 00:00:09.980 --> 00:00:13.259 wipe out decades of legacy code. And that transition 00:00:13.259 --> 00:00:15.880 isn't just a thought experiment anymore. It's 00:00:15.880 --> 00:00:19.679 a real industrial scale bet on AI transforming 00:00:19.679 --> 00:00:22.670 engineering itself. Welcome to the Deep Dive. 00:00:22.769 --> 00:00:25.449 You shared some really crucial material here, 00:00:25.570 --> 00:00:27.829 looking at the future of code, the automation 00:00:27.829 --> 00:00:30.410 of our daily work, and maybe most importantly, 00:00:30.550 --> 00:00:32.509 how researchers are trying to fix what's broken 00:00:32.509 --> 00:00:34.390 in AI safety. Yeah, we're going to break this 00:00:34.390 --> 00:00:36.990 down first, that massive plan to eliminate C 00:00:36.990 --> 00:00:40.429 and C++ code. Then we'll get into the specifics, 00:00:40.710 --> 00:00:43.969 the habits that separate a casual AI user from, 00:00:44.049 --> 00:00:46.710 you know, a true AI native professional. And 00:00:46.710 --> 00:00:48.649 finally, we'll dive into a really serious research 00:00:48.649 --> 00:00:51.579 breakthrough, Anthropix Bloom Framework. It's 00:00:51.579 --> 00:00:53.759 designed to test if AI models are actually safe 00:00:53.759 --> 00:00:56.399 or if they're just good at memorizing test answers. 00:00:56.560 --> 00:00:58.500 OK, let's get into it. Let's do it. So this first 00:00:58.500 --> 00:01:00.880 headline is pretty intense. Microsoft wants to 00:01:00.880 --> 00:01:04.900 wipe out all C and C++ code by 2030. I mean, 00:01:04.900 --> 00:01:07.379 that sounds almost impossible given how fundamental 00:01:07.379 --> 00:01:10.079 that code is to everything. It does sound like 00:01:10.079 --> 00:01:13.319 sci -fi. But the engineer leading it, Galen Hunt, 00:01:13.500 --> 00:01:15.879 confirmed it's a systematic research effort. 00:01:16.019 --> 00:01:18.799 And the motivation is, well... It's security. 00:01:19.040 --> 00:01:22.500 Right. We all know that C and C++ are where most 00:01:22.500 --> 00:01:24.760 of the dangerous memory safety bugs come from. 00:01:24.879 --> 00:01:28.140 So they're pushing Rust into places like Azure 00:01:28.140 --> 00:01:30.840 and even the Windows kernel. Okay. So this is 00:01:30.840 --> 00:01:32.420 where it gets really interesting for me. This 00:01:32.420 --> 00:01:35.719 isn't just asking a large language model to refactor 00:01:35.719 --> 00:01:39.420 this code. It's much deeper. Correct. Microsoft 00:01:39.420 --> 00:01:41.180 built two different layers of infrastructure 00:01:41.180 --> 00:01:44.140 to make this work at scale. You've got the algorithmic 00:01:44.140 --> 00:01:47.180 code infrastructure and then the AI -driven code 00:01:47.180 --> 00:01:49.000 transformation layer. So what's the difference? 00:01:49.120 --> 00:01:51.400 How does that work? Think of it like this. The 00:01:51.400 --> 00:01:53.340 algorithmic layer handles all the grunt work. 00:01:53.500 --> 00:01:56.280 The syntax, the rule. Exactly. It knows how to 00:01:56.280 --> 00:01:59.099 parse C++ functions, identify variables, all 00:01:59.099 --> 00:02:02.760 the formal logic. But the AI -driven layer, that's 00:02:02.760 --> 00:02:05.260 the intelligence. It figures out what the programmer 00:02:05.260 --> 00:02:08.120 intended to do. looks at the context, and then 00:02:08.120 --> 00:02:09.979 determines how to express that same behavior 00:02:09.979 --> 00:02:13.580 safely in the new language, like Rust. And they're 00:02:13.580 --> 00:02:16.460 already using this system today for code understanding 00:02:16.460 --> 00:02:18.780 problems. So there was this rumor going around 00:02:18.780 --> 00:02:20.879 that they were going to rewrite all of Windows 00:02:20.879 --> 00:02:23.199 in Rust. Hunt had to clarify that, didn't he? 00:02:23.319 --> 00:02:25.520 He did, yeah. The headline got a little ahead 00:02:25.520 --> 00:02:27.879 of the reality. The goal isn't necessarily to 00:02:27.879 --> 00:02:31.039 rewrite all of Windows by 2030. Right. The real 00:02:31.039 --> 00:02:33.870 goal... is making language -to -language migration 00:02:33.870 --> 00:02:37.349 possible at an industrial scale for any language. 00:02:37.969 --> 00:02:41.610 Rust is just the best test case for this experiment 00:02:41.610 --> 00:02:44.169 right now. So if this tech works, what's the 00:02:44.169 --> 00:02:46.469 next big engineering challenge that pops up? 00:02:46.569 --> 00:02:48.740 It becomes all about testing. They'll need robust 00:02:48.740 --> 00:02:50.780 frameworks to ensure the migrated code isn't 00:02:50.780 --> 00:02:52.979 just working, but is actually safer. Testing 00:02:52.979 --> 00:02:54.599 has to match the scale of the transformation. 00:02:55.180 --> 00:02:57.580 Okay, so let's flip the coin. We go from these 00:02:57.580 --> 00:03:00.259 huge institutional shifts to the individual user. 00:03:00.460 --> 00:03:02.759 You flagged a few things here on how to level 00:03:02.759 --> 00:03:06.759 up from being, say, AI curious to truly AI native. 00:03:06.919 --> 00:03:08.740 Right, there's a distinct methodology people 00:03:08.740 --> 00:03:11.960 are starting to use. We saw mentions of the three 00:03:11.960 --> 00:03:14.340 habits that separate AI native professionals. 00:03:15.469 --> 00:03:17.909 It's not about the tool. It's the approach. And 00:03:17.909 --> 00:03:20.009 what's one of those habits? What defines that 00:03:20.009 --> 00:03:23.430 approach? The big one is moving beyond simple 00:03:23.430 --> 00:03:25.409 prompts to something called prompt chaining. 00:03:25.530 --> 00:03:29.530 Okay. An AI native user designs these complex, 00:03:29.610 --> 00:03:33.270 multi -step workflows. They get the AI to do 00:03:33.270 --> 00:03:35.469 task one, then automatically feed that output 00:03:35.469 --> 00:03:38.110 into task two, and so on. It's like delegating 00:03:38.110 --> 00:03:40.169 a whole project, not just asking one question. 00:03:40.539 --> 00:03:43.039 And that kind of thinking translates directly 00:03:43.039 --> 00:03:45.400 to business, right? We saw that piece on eight 00:03:45.400 --> 00:03:48.780 game -changing marketing trends for 2026. Our 00:03:48.780 --> 00:03:51.539 current tactics are apparently losing traffic. 00:03:51.759 --> 00:03:53.340 Yeah, and it's because of something called zero 00:03:53.340 --> 00:03:55.680 -click issues. Explain that. It's when you search 00:03:55.680 --> 00:03:58.039 for something and the AI just gives you the answer 00:03:58.039 --> 00:03:59.699 right there on the results page. You never have 00:03:59.699 --> 00:04:02.000 to click through to a website. So all that work 00:04:02.000 --> 00:04:05.439 on SEO. It becomes less effective. The new strategy 00:04:05.439 --> 00:04:08.960 is to be the source inside the AI's answer. It's 00:04:08.960 --> 00:04:11.280 also changing the coder's job. We're seeing this 00:04:11.280 --> 00:04:14.180 shift away from just writing code to managing 00:04:14.180 --> 00:04:16.439 AI agents that do the building, testing, and 00:04:16.439 --> 00:04:18.600 shipping themselves. So the human becomes the 00:04:18.600 --> 00:04:20.920 manager, the orchestrator. Strategy over syntax. 00:04:20.980 --> 00:04:23.180 That's the idea. And to see how much faith is 00:04:23.180 --> 00:04:25.639 being put in this, ByteDance, you know, TikTok's 00:04:25.639 --> 00:04:29.319 owner, is planning to spend $23 billion on AI 00:04:29.319 --> 00:04:32.959 infrastructure in 2026. $23 billion? Yeah, despite 00:04:32.959 --> 00:04:35.620 export controls. It's a huge vote of confidence. 00:04:36.060 --> 00:04:39.779 Whoa. Just imagine scaling that. The data centers 00:04:39.779 --> 00:04:41.800 you'd need. It's massive. That kind of spending 00:04:41.800 --> 00:04:44.800 just shows that global leadership in AI is the 00:04:44.800 --> 00:04:46.639 ultimate goal here. And then you have these grand 00:04:46.639 --> 00:04:48.959 visions like Sam Altman predicting that in 10 00:04:48.959 --> 00:04:51.899 years, college grads will have some completely 00:04:51.899 --> 00:04:55.720 new, exciting, super well -paid job in space. 00:04:56.160 --> 00:04:58.560 It's a wild level of optimism. But you can see 00:04:58.560 --> 00:05:01.120 how the technology is forcing job roles to just 00:05:01.120 --> 00:05:03.680 evolve into things we can't even name yet. So 00:05:03.680 --> 00:05:06.379 how critical is it for a professional to master 00:05:06.379 --> 00:05:10.490 this new AI? management style right now I'd say 00:05:10.490 --> 00:05:12.930 it's vital you have to focus on strategy not 00:05:12.930 --> 00:05:15.370 just syntax strategy becomes the currency yeah 00:05:15.370 --> 00:05:17.490 okay so let's get practical for everyone listening 00:05:17.490 --> 00:05:20.829 the resources you shared had this beginners daily 00:05:20.829 --> 00:05:24.449 AI plan it's all about automation turning boring 00:05:24.449 --> 00:05:27.470 daily work into background tasks this is where 00:05:27.470 --> 00:05:30.189 it stops being a cool toy and starts being genuinely 00:05:30.189 --> 00:05:32.709 useful right it's the difference between asking 00:05:32.709 --> 00:05:36.779 AI to summarize one email versus setting up a 00:05:36.779 --> 00:05:39.040 system that summarizes all incoming client reports 00:05:39.040 --> 00:05:41.439 and drops them in a Slack channel for you every 00:05:41.439 --> 00:05:43.560 morning. That frees up a huge amount of time. 00:05:43.720 --> 00:05:46.279 Exactly. Repetitive things like sending emails, 00:05:46.459 --> 00:05:48.480 scheduling meetings, generating reports, they 00:05:48.480 --> 00:05:51.240 can all be automated. It's all about this trigger 00:05:51.240 --> 00:05:54.620 action model. When X happens, AI automatically 00:05:54.620 --> 00:05:57.060 does Y. And part of this is understanding the 00:05:57.060 --> 00:05:59.500 tools, right? Like the difference between the 00:05:59.500 --> 00:06:02.819 two big players, Zapier and Make. Why is that 00:06:02.819 --> 00:06:05.139 distinction so important? Because it really impacts 00:06:05.139 --> 00:06:07.300 your scalability and how complex your automations 00:06:07.300 --> 00:06:09.759 can get. Okay. Zapier is known for being super 00:06:09.759 --> 00:06:13.079 simple. Linear workflows. You need app A to talk 00:06:13.079 --> 00:06:16.160 to app B. It's great for that. Make, which used 00:06:16.160 --> 00:06:18.360 to be called Integromat, lets you build more 00:06:18.360 --> 00:06:21.420 complex branching workflows. You can have multiple 00:06:21.420 --> 00:06:23.240 decision paths and more elaborate scenarios. 00:06:23.759 --> 00:06:26.399 So Zapier is for a simple two -step process, 00:06:26.500 --> 00:06:29.300 but Make lets you build like a whole internal 00:06:29.300 --> 00:06:31.980 engine. Exactly. And these resources give you 00:06:31.980 --> 00:06:35.000 a clear path, you know, how to automate a simple 00:06:35.000 --> 00:06:37.800 task step by step. And they even give you 12 00:06:37.800 --> 00:06:40.759 ready -to -use workflows. It's like stacking 00:06:40.759 --> 00:06:43.230 Lego blocks of data to build efficiency. you 00:06:43.230 --> 00:06:46.170 know that focus on foundational steps is so important 00:06:46.170 --> 00:06:48.649 i still wrestle with prompt drift myself sometimes 00:06:48.649 --> 00:06:51.250 where the model just kind of forgets the original 00:06:51.250 --> 00:06:53.269 instructions in a long conversation so knowing 00:06:53.269 --> 00:06:56.110 these basic automation steps is key for stability 00:06:56.110 --> 00:06:58.649 and it builds confidence if you're struggling 00:06:58.649 --> 00:07:01.449 with complex instructions for an llm mastering 00:07:01.449 --> 00:07:04.009 simple automation is the next logical step so 00:07:04.009 --> 00:07:06.110 why is understanding that zapier versus make 00:07:06.110 --> 00:07:08.990 difference so important for new automators choosing 00:07:08.990 --> 00:07:11.670 the right platform impacts your scalability and 00:07:11.670 --> 00:07:15.389 complexity the line a lot got it choose your 00:07:15.389 --> 00:07:18.529 foundation wisely right now as we scale all this 00:07:18.529 --> 00:07:22.189 up the code migration the personal ai agent safety 00:07:22.189 --> 00:07:25.750 becomes critical we have to ensure these models 00:07:25.750 --> 00:07:29.189 behave responsibly and this brings us to anthropic's 00:07:29.189 --> 00:07:32.350 open -sourced bloom framework what's so fascinating 00:07:32.350 --> 00:07:35.730 here is that bloom addresses this huge kind of 00:07:35.730 --> 00:07:39.470 silent flaw in how we test ai right now it really 00:07:39.470 --> 00:07:42.329 does Today's benchmarks rely on static prompt 00:07:42.329 --> 00:07:45.629 tests, the same canned questions used over and 00:07:45.629 --> 00:07:47.589 over again. And the problem is those prompts 00:07:47.589 --> 00:07:49.790 eventually leak into the training data. Exactly. 00:07:49.829 --> 00:07:52.250 So the new models learn how to pass the test, 00:07:52.370 --> 00:07:55.110 not the actual safety principles behind it. Scores 00:07:55.110 --> 00:07:57.529 go up, but the real safety understanding doesn't 00:07:57.529 --> 00:07:59.879 improve. It's just memorization. That's a huge 00:07:59.879 --> 00:08:01.500 liability. You're basically deploying systems 00:08:01.500 --> 00:08:03.420 that have learned to cheat on their safety exams. 00:08:03.680 --> 00:08:06.300 Totally. So how does Bloom fix this memorization 00:08:06.300 --> 00:08:09.100 problem? Well, it shifts the focus. Instead of 00:08:09.100 --> 00:08:11.060 testing fixed prompts, it tests the underlying 00:08:11.060 --> 00:08:13.600 behavior. How so? It keeps the behavior definition 00:08:13.600 --> 00:08:16.040 constant, like, do not generate instructions 00:08:16.040 --> 00:08:19.579 for illegal activity. But it continuously generates 00:08:19.579 --> 00:08:22.240 fresh, new situations that could trigger that 00:08:22.240 --> 00:08:25.399 behavior. It's a dynamic test, not a static one. 00:08:25.759 --> 00:08:27.740 And they use a four -stage pipeline to do this. 00:08:28.019 --> 00:08:30.740 There's behavior grounding, scenario generation. 00:08:31.360 --> 00:08:33.019 So let's talk about that scenario generation 00:08:33.019 --> 00:08:35.740 stage. What does it actually do? That's where 00:08:35.740 --> 00:08:37.960 it gets clever. It's another AI whose entire 00:08:37.960 --> 00:08:41.480 job is to invent thousands of new, challenging, 00:08:41.600 --> 00:08:44.500 multi -turn conversations. So instead of a simple 00:08:44.500 --> 00:08:46.919 question, it might create a complex social engineering 00:08:46.919 --> 00:08:49.940 dialogue to try and trick the model. So the score 00:08:49.940 --> 00:08:52.259 you get actually reflects robustness against 00:08:52.259 --> 00:08:54.860 new attacks, not just memorization. Exactly. 00:08:55.059 --> 00:08:57.279 And by the end, it's not a simple pass or fail. 00:08:57.440 --> 00:09:00.120 Bloom tracks how reliably a behavior shows up. 00:09:00.649 --> 00:09:03.409 is a detailed behavioral profile, not just a 00:09:03.409 --> 00:09:06.370 checkbox. It gives engineers a much clearer picture 00:09:06.370 --> 00:09:09.450 of the model's actual risk. So what impact does 00:09:09.450 --> 00:09:11.529 this better evaluation have on all the new tools 00:09:11.529 --> 00:09:13.769 we're seeing, like generative video or agents 00:09:13.769 --> 00:09:16.570 that need reliable voice controls? Well, responsible 00:09:16.570 --> 00:09:19.129 adoption depends entirely on these safety frameworks. 00:09:19.289 --> 00:09:21.570 You can't deploy powerful tools into the world 00:09:21.570 --> 00:09:24.190 if you can't trust the behavioral profile of 00:09:24.190 --> 00:09:26.950 the model underneath. Right. So if AI evaluation 00:09:26.950 --> 00:09:29.850 shifts to these behavioral profiles, how does... 00:09:29.870 --> 00:09:32.629 that change how models are reported on and sold? 00:09:32.870 --> 00:09:35.629 It means transparency will have to focus on reliability 00:09:35.629 --> 00:09:38.509 across all kinds of different scenarios, not 00:09:38.509 --> 00:09:41.129 just passing a simple test. Moving away from 00:09:41.129 --> 00:09:44.529 easily gamed benchmark scores. So what does this 00:09:44.529 --> 00:09:46.889 all mean for you? We've seen the sheer scale 00:09:46.889 --> 00:09:50.110 of ambition with Microsoft betting on AI to dismantle 00:09:50.110 --> 00:09:53.370 its own programming past by 2030. We also saw 00:09:53.370 --> 00:09:55.389 the rapid professional evolution that's needed, 00:09:55.509 --> 00:09:57.789 where managing AI agents is replacing writing 00:09:57.789 --> 00:10:00.769 code. The path becoming AI native is all about 00:10:00.769 --> 00:10:03.590 strategy and orchestration. And all of this incredible 00:10:03.590 --> 00:10:06.450 power is being supported by crucial safety research. 00:10:06.649 --> 00:10:09.009 The core takeaway, really, is that the AI frontier 00:10:09.009 --> 00:10:11.870 isn't just about speed and new tools. It's equally 00:10:11.870 --> 00:10:14.269 about the foundational work like the Bloom framework 00:10:14.269 --> 00:10:16.250 to make sure our systems aren't just scoring 00:10:16.250 --> 00:10:18.850 high, but are actually operating safely and reliably. 00:10:19.049 --> 00:10:21.350 Yeah, in the real world. Our deep dive today 00:10:21.350 --> 00:10:24.370 covered everything. from these massive institutional 00:10:24.370 --> 00:10:27.230 shifts like ByteDance's huge infrastructure investment 00:10:27.230 --> 00:10:31.470 to the really actionable individual tool sets, 00:10:31.690 --> 00:10:34.250 like knowing when to use Zapier versus Make. 00:10:34.370 --> 00:10:36.370 So think about which of your daily repetitive 00:10:36.370 --> 00:10:39.129 tasks could be immediately shifted to automation. 00:10:39.690 --> 00:10:42.090 Free up that mental bandwidth for the bigger 00:10:42.090 --> 00:10:44.230 picture, for that strategy work. But here's a 00:10:44.230 --> 00:10:45.809 final provocative thought for you to mull over. 00:10:46.029 --> 00:10:50.090 Okay. If AI agents become responsible for building, 00:10:50.190 --> 00:10:52.929 testing, and shipping, the majority of new software. 00:10:53.870 --> 00:10:56.429 What becomes the single most valuable skill a 00:10:56.429 --> 00:10:58.970 human programmer has? That is a great question. 00:10:59.110 --> 00:11:00.929 Thank you for sharing your sources with us for 00:11:00.929 --> 00:11:02.750 this deep dive. My pleasure. Until next time.