WEBVTT 00:00:00.000 --> 00:00:02.080 If you don't change how you structure your work, 00:00:02.399 --> 00:00:05.660 AI just amplifies your sloppiness. Wow. I read 00:00:05.660 --> 00:00:07.280 that this morning and I had to put my coffee 00:00:07.280 --> 00:00:09.759 down. It felt less like tech advice and more 00:00:09.759 --> 00:00:11.800 like a personal attack. It's pretty aggressive, 00:00:11.839 --> 00:00:16.420 isn't it? But honestly, for 2026, it's the reality 00:00:16.420 --> 00:00:19.600 check I think a lot of us need. It is. Welcome 00:00:19.600 --> 00:00:22.899 back to the deep dive. It's Friday, January 23rd. 00:00:23.140 --> 00:00:25.710 Today we are. We're slowing things down. We're 00:00:25.710 --> 00:00:28.230 not talking about the newest apps. No new gadgets. 00:00:28.570 --> 00:00:31.250 No. We're looking at the invisible layer, the 00:00:31.250 --> 00:00:33.630 psychology of how we actually talk to these new 00:00:33.630 --> 00:00:35.909 kinds of intelligence. We're deconstructing a 00:00:35.909 --> 00:00:39.439 piece called AI problem solving, systems for 00:00:39.439 --> 00:00:41.700 reliable intelligence. Which is, you know, a 00:00:41.700 --> 00:00:44.100 very dry title for what is basically a manual 00:00:44.100 --> 00:00:45.859 on how to stop messing around with these tools. 00:00:45.939 --> 00:00:47.719 Right. And it feels like the timing is perfect. 00:00:47.820 --> 00:00:50.399 For the past few years, it's been all about experimenting, 00:00:50.780 --> 00:00:52.859 just, you know, throwing stuff at the wall. The 00:00:52.859 --> 00:00:55.619 let's see what happens phase. Exactly. But the 00:00:55.619 --> 00:00:57.560 whole idea here is that phase is over. It's time 00:00:57.560 --> 00:00:59.460 to build systems or you're just going to get 00:00:59.460 --> 00:01:01.820 buried in your own mess. That's the great divergence 00:01:01.820 --> 00:01:03.979 we're seeing. People who treat AI like a magic 00:01:03.979 --> 00:01:06.439 eight ball and people who treat it like an engineering 00:01:06.439 --> 00:01:08.780 problem. So here's the plan for today. First, 00:01:08.920 --> 00:01:11.439 we're going to dismantle that search box idea 00:01:11.439 --> 00:01:14.819 why that blinking cursor is actually a trap. 00:01:15.299 --> 00:01:17.420 Then we'll get into reliability, things like 00:01:17.420 --> 00:01:21.340 grounding. And the I don't know rule, which is 00:01:21.340 --> 00:01:23.579 way harder than it sounds. Then there's the LLM 00:01:23.579 --> 00:01:27.900 council. Sounds very grand. Very sci -fi. It 00:01:27.900 --> 00:01:30.739 does. We'll talk orchestration versus agents, 00:01:31.000 --> 00:01:33.640 and then spend some real time on a term I love, 00:01:34.140 --> 00:01:36.480 vibe coding. My favorite. It's a real shift in 00:01:36.480 --> 00:01:38.379 how we think. And finally, we'll talk about the 00:01:38.379 --> 00:01:40.719 last human bottleneck, things like judgment and 00:01:40.719 --> 00:01:43.400 curation. So let's start at the beginning, the 00:01:43.400 --> 00:01:46.040 search box trap. OK. The source argues our biggest 00:01:46.040 --> 00:01:48.859 problem is just. muscle memory. It's cognitive 00:01:48.859 --> 00:01:51.099 inertia. I mean, think about it. For what, 30 00:01:51.099 --> 00:01:53.659 years, the entire internet era, we've been trained 00:01:53.659 --> 00:01:55.879 to do one thing. You have a question? You type 00:01:55.879 --> 00:01:59.439 it in a box, you hit enter, and Google or whatever 00:01:59.439 --> 00:02:01.739 finds a list of answers that already exist. It's 00:02:01.739 --> 00:02:04.879 a retrieval task. You're a librarian, and you're 00:02:04.879 --> 00:02:07.260 just fetching a book from a shelf. The answer 00:02:07.260 --> 00:02:09.860 is already out there. Precisely. But here's the 00:02:09.860 --> 00:02:12.900 fundamental breakdown. An LLM is not a librarian. 00:02:13.120 --> 00:02:16.039 It's a poet. It's a poet. It's a generator. When 00:02:16.039 --> 00:02:18.719 you type in a prompt, it is not looking up an 00:02:18.719 --> 00:02:21.740 answer. It is calculating the most probable next 00:02:21.740 --> 00:02:24.620 word. So if I treat a generator like a retriever? 00:02:24.719 --> 00:02:27.479 You get hallucinations. You're asking a probability 00:02:27.479 --> 00:02:30.639 engine to act like a database. And because it's 00:02:30.639 --> 00:02:32.680 designed to be helpful, it will just confidently 00:02:32.680 --> 00:02:35.360 lie to you. It fills the gaps with noise that 00:02:35.360 --> 00:02:37.719 sounds plausible. And in a professional setting, 00:02:38.319 --> 00:02:41.860 legal medical, that's a huge liability. It's 00:02:41.860 --> 00:02:45.069 just dangerous. If your input is vague, if you 00:02:45.069 --> 00:02:47.509 just kind of toss a question out there, the AI's 00:02:47.509 --> 00:02:49.389 output is going to be just as vague. It defaults 00:02:49.389 --> 00:02:51.069 to the average of the internet. And the average 00:02:51.069 --> 00:02:54.250 of the internet is? Mediocrity. At best. So the 00:02:54.250 --> 00:02:56.930 goal is to shift from just getting an output 00:02:56.930 --> 00:02:59.050 to actually solving a problem. It's about building 00:02:59.050 --> 00:03:02.169 a workflow where errors are caught by the structure 00:03:02.169 --> 00:03:06.889 itself. You have to assume the model is a little 00:03:06.889 --> 00:03:09.729 bit drunk. And you have to build guardrails so 00:03:09.729 --> 00:03:11.689 that even if it wants to go off the rails, it 00:03:11.689 --> 00:03:14.389 can't. So let me pause on that. If that search 00:03:14.389 --> 00:03:17.409 box method is our default, what's the immediate 00:03:17.409 --> 00:03:20.330 consequence of using it for high stakes work? 00:03:20.650 --> 00:03:24.550 It creates fast but risky outputs. It gives you 00:03:24.550 --> 00:03:26.930 speed without reliability. Speed without reliability. 00:03:27.009 --> 00:03:28.750 That feels like the slogan for the last three 00:03:28.750 --> 00:03:31.449 years. So how do we engineer that reliability 00:03:31.449 --> 00:03:34.750 in? The source talks about grounding. We hear 00:03:34.750 --> 00:03:36.889 that term a lot. Explain it to me like I'm a 00:03:36.889 --> 00:03:41.689 skeptic. OK. So an ungrounded AI is improvising. 00:03:41.819 --> 00:03:44.439 It's just making things up based on its vast, 00:03:44.699 --> 00:03:47.080 messy training data. Grounding is like handing 00:03:47.080 --> 00:03:50.080 the AI a script. You're telling it, ignore everything 00:03:50.080 --> 00:03:52.539 you think you know. Look only at these three 00:03:52.539 --> 00:03:55.259 PDFs I just uploaded. Answer my question using 00:03:55.259 --> 00:03:57.199 only this information. So it's the difference 00:03:57.199 --> 00:03:59.539 between an essay from memory versus an open book 00:03:59.539 --> 00:04:01.560 test, where you have to cite your sources. It's 00:04:01.560 --> 00:04:03.659 even stricter, because with an open book test, 00:04:03.659 --> 00:04:05.719 you can still kind of fudge it. The magic trick 00:04:05.719 --> 00:04:08.259 with grounding is you have to add one more instruction. 00:04:08.319 --> 00:04:11.990 You have to explicitly tell the model. If the 00:04:11.990 --> 00:04:15.469 answer's not in this text, you must say, I don't 00:04:15.469 --> 00:04:20.550 know. That feels almost too simple. Does that 00:04:20.550 --> 00:04:23.170 really work? It changes everything. Because these 00:04:23.170 --> 00:04:25.529 models are people pleasers. They're optimized 00:04:25.529 --> 00:04:27.610 to give you an answer. They hate saying they 00:04:27.610 --> 00:04:29.750 don't know. It feels like a failure to them. 00:04:29.949 --> 00:04:32.410 Exactly. So by giving them permission to fail, 00:04:32.790 --> 00:04:34.930 you eliminate all those forced hallucinations 00:04:34.930 --> 00:04:36.870 where they're trying to connect dots that just 00:04:36.870 --> 00:04:38.610 aren't there. So you're giving it permission 00:04:38.610 --> 00:04:41.660 to be useless. And that... paradoxically makes 00:04:41.660 --> 00:04:43.920 it incredibly useful. Because then when it does 00:04:43.920 --> 00:04:45.959 give you an answer, you know where it came from. 00:04:46.139 --> 00:04:48.920 And this connects to ARAG, right? Retrieval Augmented 00:04:48.920 --> 00:04:51.100 Generation. Right. That's just the plumbing for 00:04:51.100 --> 00:04:54.639 grounding at scale. If you have, say, 10 ,000 00:04:54.639 --> 00:04:56.819 company documents, you can't paste them all into 00:04:56.819 --> 00:04:59.399 the prompt. Of course not. ARAG is just the system 00:04:59.399 --> 00:05:01.480 that first searches your documents, finds the 00:05:01.480 --> 00:05:04.100 right page, and then hands just that one relevant 00:05:04.100 --> 00:05:07.350 page to the AI to ground its answer. So you move 00:05:07.350 --> 00:05:09.509 from an answer based on vibes to an answer based 00:05:09.509 --> 00:05:13.269 on evidence. Evidence is the word. You stop judging 00:05:13.269 --> 00:05:16.230 if the answer sounds right and you start checking 00:05:16.230 --> 00:05:19.389 its citations. So let's nail this down. What's 00:05:19.389 --> 00:05:22.410 the one critical instruction that turns a guessing 00:05:22.410 --> 00:05:26.490 AI into a grounded one? Explicitly permitting 00:05:26.490 --> 00:05:30.089 the model to say, I don't know, eliminates forced 00:05:30.089 --> 00:05:32.870 hallucinations. Yeah, I don't know rule. I feel 00:05:32.870 --> 00:05:34.790 like I need to apply that to my own life. No, 00:05:34.870 --> 00:05:38.350 we all. OK, let's scale this up. One model, even 00:05:38.350 --> 00:05:41.189 a grounded one, can still make a mistake. The 00:05:41.189 --> 00:05:43.990 source brings up this idea from André G. Carpathy, 00:05:44.790 --> 00:05:46.850 the LLM Council. That's the council, yes. It 00:05:46.850 --> 00:05:48.189 sounds like something out of Lord of the Rings. 00:05:48.670 --> 00:05:51.269 It really does. But it's actually just expensive 00:05:51.269 --> 00:05:53.230 and redundant computing. Walk me through it, 00:05:53.230 --> 00:05:54.949 because my first thought is, why would I need 00:05:54.949 --> 00:05:57.790 four different AIs to answer one question? Well, 00:05:57.949 --> 00:05:59.850 you wouldn't for a simple email. This is for 00:05:59.850 --> 00:06:02.670 high stakes work. Imagine you're analyzing a 00:06:02.670 --> 00:06:05.470 legal contract for a loophole. OK. If you just 00:06:05.470 --> 00:06:08.910 use one model, say GPT -5, you're a victim of 00:06:08.910 --> 00:06:12.149 its specific blind spots. Every model has a different 00:06:12.149 --> 00:06:14.470 personality, a different way of reasoning. A 00:06:14.470 --> 00:06:16.649 different flavor. Right. Some are more creative, 00:06:16.750 --> 00:06:19.250 some are more literal. The council approach is 00:06:19.250 --> 00:06:21.509 you write one clear prompt and you send it to 00:06:21.509 --> 00:06:23.529 three or four different models all at the same 00:06:23.529 --> 00:06:25.750 time. And then you compare the answers. Exactly. 00:06:26.069 --> 00:06:28.990 If three say this is safe and one says this is 00:06:28.990 --> 00:06:32.389 a risk, you stop. That disagreement is your signal. 00:06:32.569 --> 00:06:34.470 It's a smoke alarm telling you there's nuance 00:06:34.470 --> 00:06:36.629 you missed. And the source mentioned a meta step 00:06:36.629 --> 00:06:39.209 that I thought was brilliant, using another AI 00:06:39.209 --> 00:06:42.230 to judge the council's answers. The judge model. 00:06:42.410 --> 00:06:44.949 It turns out models are often better at critiquing 00:06:44.949 --> 00:06:48.649 than creating. It's easier to spot a flaw than 00:06:48.649 --> 00:06:51.029 have an original idea. So you just feed the four 00:06:51.029 --> 00:06:53.290 answers into a fifth model. And you say, rank 00:06:53.290 --> 00:06:56.350 these. Who's right? Who missed the point? It's 00:06:56.350 --> 00:06:58.850 like peer review at the speed of light. It's 00:06:58.850 --> 00:07:01.029 an insurance policy against a single point of 00:07:01.029 --> 00:07:04.389 failure. OK, so the probing question is, why 00:07:04.389 --> 00:07:07.589 go through all that trouble and expense of consulting 00:07:07.589 --> 00:07:11.250 multiple brains for one problem? It exposes blind 00:07:11.250 --> 00:07:14.509 spots. If the models disagree, you know you need 00:07:14.509 --> 00:07:17.870 to verify. Find spots. Okay. So that's how we 00:07:17.870 --> 00:07:20.009 get reliable answers. Now let's talk about getting 00:07:20.009 --> 00:07:23.389 work done. The source draws a really clear line 00:07:23.389 --> 00:07:26.410 between two things, orchestration and agents. 00:07:26.949 --> 00:07:29.250 The train versus the taxi. This is the mental 00:07:29.250 --> 00:07:31.529 model everyone needs. Break it down. What's the 00:07:31.529 --> 00:07:34.740 train? The train is orchestration. It's on rails. 00:07:34.779 --> 00:07:36.980 It goes from station A to station B to station 00:07:36.980 --> 00:07:40.639 C. It's rigid. It can't improvise. Not at all. 00:07:40.779 --> 00:07:45.160 It's brittle, but it's 100 % predictable. Orchestration 00:07:45.160 --> 00:07:47.839 is for when you map out a boring, repetitive 00:07:47.839 --> 00:07:51.819 task. Step one, summarize this text. Step two, 00:07:52.160 --> 00:07:55.410 extract the key dates. Step three, put them in 00:07:55.410 --> 00:07:57.610 a calendar invite. And if something changes in 00:07:57.610 --> 00:08:00.290 step one? The whole train derails. But as long 00:08:00.290 --> 00:08:02.170 as the track is clear, it works perfectly every 00:08:02.170 --> 00:08:04.089 single time. OK, so that's the train. What's 00:08:04.089 --> 00:08:06.170 the taxi? The taxi is an agent. When you get 00:08:06.170 --> 00:08:08.430 into a taxi, you don't give it step by step directions. 00:08:08.730 --> 00:08:10.790 No, you give it a destination. Right. Get me 00:08:10.790 --> 00:08:13.410 to the airport. The driver, the agent, figures 00:08:13.410 --> 00:08:15.189 out the route. It adapts. If there's traffic 00:08:15.189 --> 00:08:17.829 on one street, it takes another. Agents are goal 00:08:17.829 --> 00:08:20.170 -oriented, not step -oriented. And you give them 00:08:20.170 --> 00:08:23.829 tools like access to a browser or your calendar. 00:08:24.269 --> 00:08:26.370 Exactly. You tell it schedule a meeting with 00:08:26.370 --> 00:08:28.810 Sarah for next week. You don't tell it how to 00:08:28.810 --> 00:08:31.250 do that. It figures it out. See, that's the part 00:08:31.250 --> 00:08:33.029 that makes me and I think a lot of people a little 00:08:33.029 --> 00:08:36.509 nervous giving an AI that much agency. What if 00:08:36.509 --> 00:08:38.909 it decides the fastest way to the airport is 00:08:38.909 --> 00:08:41.850 through a park? That is the core alignment problem. 00:08:41.929 --> 00:08:45.070 and it's why real agents are only just now becoming 00:08:45.070 --> 00:08:48.250 practical. The risk was too high. The train is 00:08:48.250 --> 00:08:51.250 safe because you built the track yourself. The 00:08:51.250 --> 00:08:54.929 taxi requires trust. So you use trains for boring, 00:08:55.169 --> 00:08:58.470 predictable tasks and agents for messy problems. 00:08:58.570 --> 00:09:01.129 For dynamic, messy problems. Like, do some research 00:09:01.129 --> 00:09:03.009 on this new company and find a good person to 00:09:03.009 --> 00:09:05.269 contact. You can't script that. Every website 00:09:05.269 --> 00:09:07.470 is different. You need an agent that can try 00:09:07.470 --> 00:09:09.629 something, fail, back up, and try a different 00:09:09.629 --> 00:09:11.490 approach. It's creating its own map as it goes. 00:09:11.629 --> 00:09:13.909 Yes. but you have to put it in a sandbox. You 00:09:13.909 --> 00:09:15.690 have to give it clear boundaries. So what's the 00:09:15.690 --> 00:09:17.809 fundamental distinction here between an agent 00:09:17.809 --> 00:09:21.309 and just a standard automation workflow? Automations 00:09:21.309 --> 00:09:24.509 follow rigid steps. Agents follow a goal and 00:09:24.509 --> 00:09:27.409 adapt their path. Adaptability. Okay. I want 00:09:27.409 --> 00:09:30.269 to shift to the term that I admit made me roll 00:09:30.269 --> 00:09:34.529 my eyes at first. Vibe coding. Ah, don't do that. 00:09:34.590 --> 00:09:36.830 It's important. I'm trying not to. It just sounds 00:09:36.830 --> 00:09:38.669 like something you'd overhear at a skate park. 00:09:39.049 --> 00:09:41.909 But the source treats it as this huge Revolution. 00:09:42.090 --> 00:09:44.149 It is a revolution. Just forget the slang for 00:09:44.149 --> 00:09:45.970 a second. Think of it as natural language programming. 00:09:45.970 --> 00:09:49.529 Go on. In the old days, meaning like 2023, if 00:09:49.529 --> 00:09:52.049 you wanted to build software, you were a translator. 00:09:52.450 --> 00:09:54.509 You had a human idea in your head and you had 00:09:54.509 --> 00:09:57.269 to painstakingly translate it into the machine's 00:09:57.269 --> 00:09:59.809 language. All the semicolons and brackets. Right. 00:09:59.950 --> 00:10:02.169 Syntax was the barrier. Yeah. Miss one comma, 00:10:02.169 --> 00:10:04.970 the whole thing breaks. Vibe coding is the shift 00:10:04.970 --> 00:10:07.289 where you stop being a translator and start being 00:10:07.289 --> 00:10:09.789 an architect. You just describe the intent. So 00:10:09.789 --> 00:10:13.190 instead of set background color to hex code FFF. 00:10:13.370 --> 00:10:15.750 You say, make the landing page feel clean and 00:10:15.750 --> 00:10:18.389 minimalist. The AI handles the implementation. 00:10:19.090 --> 00:10:21.629 It translates clean into the right hex codes 00:10:21.629 --> 00:10:24.230 and CSS. And this matters because it completely 00:10:24.230 --> 00:10:27.049 changes who can build things. Totally. You don't 00:10:27.049 --> 00:10:29.129 need to know Python anymore. You need to know 00:10:29.129 --> 00:10:32.220 how to clearly describe a problem and what a 00:10:32.220 --> 00:10:34.360 solution should feel like. So instead of being 00:10:34.360 --> 00:10:36.480 stuck with a messy spreadsheet, you can just... 00:10:36.480 --> 00:10:39.220 You can just say to the AI, build me a simple 00:10:39.220 --> 00:10:42.200 web page that lets me search and filter the rows 00:10:42.200 --> 00:10:45.159 in this spreadsheet. You describe the function, 00:10:45.279 --> 00:10:47.919 the vibe of the tool, and it generates the code. 00:10:48.039 --> 00:10:50.179 But that requires a totally different skill set, 00:10:50.299 --> 00:10:53.320 right? If it's not syntax, what is it? It's systems 00:10:53.320 --> 00:10:56.720 thinking. It's clarity of communication. If your 00:10:56.720 --> 00:10:59.259 description of the problem is sloppy, the tool 00:10:59.259 --> 00:11:02.120 the AI builds will be sloppy. We're seeing this 00:11:02.120 --> 00:11:05.440 weird thing where people with, say, a background 00:11:05.440 --> 00:11:09.000 in literature are becoming great vibe coders 00:11:09.000 --> 00:11:10.940 because they're masters of describing things 00:11:10.940 --> 00:11:13.120 with precision. That's a wild thought. The entire 00:11:13.120 --> 00:11:15.340 skill stack is flipping upside down. It's the 00:11:15.340 --> 00:11:17.779 revenge of the liberal arts. I love that. So 00:11:17.779 --> 00:11:20.860 the probing question. What's the main skill for 00:11:20.860 --> 00:11:24.279 vibe coding if it isn't programming? The ability 00:11:24.279 --> 00:11:27.399 to clearly describe a problem and a desired functional 00:11:27.399 --> 00:11:30.779 outcome. Describing the outcome. Sounds so simple, 00:11:31.000 --> 00:11:33.279 but it's incredibly hard. We're going to take 00:11:33.279 --> 00:11:35.679 a very quick break when we get back. If AI is 00:11:35.679 --> 00:11:38.440 handling all this, what's actually left for us 00:11:38.440 --> 00:11:42.139 to do? The good stuff. Stay with us, sponsor. 00:11:42.799 --> 00:11:45.340 And we are back on the deep dive. Before the 00:11:45.340 --> 00:11:47.720 break, we covered the search box, the LLM council, 00:11:47.980 --> 00:11:51.159 trains, taxis, and the rise of vibe coding. Now 00:11:51.159 --> 00:11:53.360 let's talk about the human in the loop. Where 00:11:53.360 --> 00:11:55.860 do we fit in all this? The source points to two 00:11:55.860 --> 00:11:58.679 main roles, debugging and curation. Let's start 00:11:58.679 --> 00:12:00.259 with debugging. Well, it makes sense, right? 00:12:00.279 --> 00:12:02.519 If you're vibe coding an app or using an agent, 00:12:03.000 --> 00:12:04.639 things are going to go wrong, the AI is going 00:12:04.639 --> 00:12:06.700 to misunderstand you, the agent will get stuck. 00:12:06.899 --> 00:12:09.419 So we stop being writers and we become mechanics. 00:12:09.559 --> 00:12:11.659 We become diagnosticians. Yeah. And the source 00:12:11.659 --> 00:12:14.320 warns against what it calls the retry loop. You 00:12:14.320 --> 00:12:16.139 know, when you get a bad answer from the AI and 00:12:16.139 --> 00:12:18.700 you just hit regenerate over and over, hoping 00:12:18.700 --> 00:12:21.379 for a better one. I do that all the time. We 00:12:21.379 --> 00:12:23.860 all do. It's the definition of insanity. The 00:12:23.860 --> 00:12:26.419 new skill is to look at the failure and ask why. 00:12:26.759 --> 00:12:29.600 Was my context bad? Did I assume the AI knew 00:12:29.600 --> 00:12:31.820 something it didn't? The source mentioned a technique 00:12:31.820 --> 00:12:34.799 called stress testing. Oh, I use this constantly. 00:12:34.919 --> 00:12:36.980 After you get an answer you like, you ask the 00:12:36.980 --> 00:12:39.759 AI a follow -up. Where is this output likely 00:12:39.759 --> 00:12:42.340 to break? Or what assumptions did you make to 00:12:42.340 --> 00:12:44.980 get here? You're asking it to audit its own work. 00:12:45.220 --> 00:12:47.960 You're forcing it to reveal its own weak points. 00:12:48.220 --> 00:12:51.000 And that requires a human. to understand the 00:12:51.000 --> 00:12:53.679 context of the real world, which the AI just 00:12:53.679 --> 00:12:55.820 doesn't have. And then there's the second role, 00:12:56.139 --> 00:13:00.559 curation. This line really hit me. By 2026, creation 00:13:00.559 --> 00:13:04.320 is trivial. It's basically free. The cost to 00:13:04.320 --> 00:13:06.340 generate a thousand words or a piece of code 00:13:06.340 --> 00:13:09.740 or an image is trending to zero. And in economics, 00:13:09.960 --> 00:13:12.799 when supply is infinite, value collapses. So 00:13:12.799 --> 00:13:15.860 being able to write generic text isn't a valuable 00:13:15.860 --> 00:13:18.620 skill anymore. No. The value moves up the stack. 00:13:18.700 --> 00:13:20.779 It moves to judgment. It moves to the role of 00:13:20.779 --> 00:13:22.659 the editor. I like the magazine editor analogy. 00:13:22.840 --> 00:13:25.500 It's perfect. It used to be that the hard part 00:13:25.500 --> 00:13:27.200 was getting the articles written. Now you can 00:13:27.200 --> 00:13:29.399 generate 1 ,000 articles in a second. The hard 00:13:29.399 --> 00:13:31.919 part, the valuable part, is knowing which one 00:13:31.919 --> 00:13:35.379 is true, which one is important, and which 999 00:13:35.379 --> 00:13:38.059 to throw away. So you stop asking the AI to give 00:13:38.059 --> 00:13:41.360 me 10 ideas. And you start asking it to filter 00:13:41.360 --> 00:13:44.039 these 50 ideas. You give it 100 pages and say, 00:13:44.220 --> 00:13:47.200 reduce this to the single most important paragraph. 00:13:47.399 --> 00:13:50.120 You use it to prune, not just to generate. The 00:13:50.120 --> 00:13:53.340 source calls this cognitive offloading. We offload 00:13:53.340 --> 00:13:55.539 the busy work, but never the final judgment. 00:13:55.879 --> 00:13:58.659 That is the bright red line. You can outsource 00:13:58.659 --> 00:14:01.539 labor. You can outsource summarization. You can 00:14:01.539 --> 00:14:04.620 not outsource the decision of what actually matters. 00:14:04.759 --> 00:14:06.840 The moment you do that, You're not using the 00:14:06.840 --> 00:14:10.179 tool anymore. The tool is using you. Wow. That's 00:14:10.179 --> 00:14:12.639 a powerful distinction. So the final probing 00:14:12.639 --> 00:14:16.279 question. In a world of infinite AI generation, 00:14:16.940 --> 00:14:19.799 what becomes the most valuable human contribution? 00:14:19.960 --> 00:14:22.580 Curation. The ability to decide what is worth 00:14:22.580 --> 00:14:25.039 keeping and what to ignore. Deciding what to 00:14:25.039 --> 00:14:26.879 ignore. That might be the most important skill 00:14:26.879 --> 00:14:28.899 of the century. I really think it is. Let's bring 00:14:28.899 --> 00:14:30.860 this all together. We have covered a ton of ground 00:14:30.860 --> 00:14:33.019 today. We really have. We started by tearing 00:14:33.019 --> 00:14:35.919 down the search box idea. Then we built reliability 00:14:35.919 --> 00:14:39.279 with grounding and the LLM council. We look at 00:14:39.279 --> 00:14:41.879 the actual systems. orchestration for the predictable 00:14:41.879 --> 00:14:44.440 stuff, agents for the messy stuff. And we talked 00:14:44.440 --> 00:14:46.840 about the big shift in our roles from being just 00:14:46.840 --> 00:14:50.379 users to being architects with vibe coding and 00:14:50.379 --> 00:14:54.200 editors through curation. The single theme connecting 00:14:54.200 --> 00:14:58.149 all of this seems to be. That's it. The people 00:14:58.149 --> 00:15:00.769 who are succeeding with AI are the ones who treat 00:15:00.769 --> 00:15:03.169 it like an engineering discipline. They aren't 00:15:03.169 --> 00:15:05.850 just chatting with it. They're designing systems 00:15:05.850 --> 00:15:07.769 for it to operate within. It's the difference 00:15:07.769 --> 00:15:10.429 between just experimenting and actually operating. 00:15:10.649 --> 00:15:12.610 That's the whole game right there. So our challenge 00:15:12.610 --> 00:15:15.169 for you this week is to pick one boring, repetitive 00:15:15.169 --> 00:15:17.830 task you do. And don't just ask AI to do it. 00:15:17.929 --> 00:15:20.350 Try to orchestrate it. Write down the steps A, 00:15:20.389 --> 00:15:23.049 then B, then C. See if you can build a reliable 00:15:23.049 --> 00:15:25.309 track for it. And try the, I don't know, rule. 00:15:25.409 --> 00:15:27.370 Just add that one sentence to your instructions. 00:15:27.750 --> 00:15:30.149 See how much better and quieter your outputs 00:15:30.149 --> 00:15:32.389 become. I want to leave you with one final thought 00:15:32.389 --> 00:15:34.350 from the source, a bit of a tough one to sit 00:15:34.350 --> 00:15:38.370 with. If AI exposes how you think and your AI 00:15:38.370 --> 00:15:41.610 outputs are messy, what does that say about your 00:15:41.610 --> 00:15:45.210 current thinking process? Oof. Yeah. That's the 00:15:45.210 --> 00:15:47.080 one that'll keep you up at night. Something to 00:15:47.080 --> 00:15:49.139 reflect on. Thanks for diving in with us. Always 00:15:49.139 --> 00:15:50.480 a pleasure. We'll see you next time.