WEBVTT 00:00:00.000 --> 00:00:02.620 I uploaded a contract yesterday, just standard 00:00:02.620 --> 00:00:04.719 stuff, really, a service agreement. And I needed 00:00:04.719 --> 00:00:07.820 to check one thing, a specific liability cap. 00:00:08.039 --> 00:00:10.900 So I asked the AI and it came back instantly. 00:00:11.039 --> 00:00:13.539 It cited the clause perfectly. Paragraph four, 00:00:13.660 --> 00:00:16.539 section B said liability is capped at two times 00:00:16.539 --> 00:00:20.100 the total fees paid. It looked professional, 00:00:20.480 --> 00:00:23.410 totally sound. Let me guess. There was no Section 00:00:23.410 --> 00:00:25.010 B. There was no Section B. The clause didn't 00:00:25.010 --> 00:00:27.609 exist at all. But here's the kicker. It wasn't 00:00:27.609 --> 00:00:33.109 a glitch. It wasn't broken. The AI, it was trying 00:00:33.109 --> 00:00:35.590 to be helpful. And if I hadn't gone back to the 00:00:35.590 --> 00:00:37.310 PDF to double check, which, you know, I really 00:00:37.310 --> 00:00:39.929 didn't want to do, that helpful little hallucination 00:00:39.929 --> 00:00:42.149 could have cost us thousands. And that right 00:00:42.149 --> 00:00:44.530 there is the entire paradox we're living in now. 00:00:44.850 --> 00:00:46.909 Welcome back to the Deep Dive. It is Monday, 00:00:47.070 --> 00:00:49.920 February 9th, 2026. Usually we're around here 00:00:49.920 --> 00:00:51.740 talking about speed, you know, how fast the new 00:00:51.740 --> 00:00:53.880 models are. But today we're hitting the brakes. 00:00:53.979 --> 00:00:55.600 We're talking about trust. We're exploring the 00:00:55.600 --> 00:00:57.920 nature of truth when you're dealing with a machine 00:00:57.920 --> 00:01:00.200 that is literally programmed to please you. It's 00:01:00.200 --> 00:01:02.039 a fascinating problem, isn't it? Because the 00:01:02.039 --> 00:01:04.299 models are so much more powerful now than they 00:01:04.299 --> 00:01:07.680 were back in, you know, 2023. But that confident 00:01:07.680 --> 00:01:10.420 guessing issue, it's still here. And it's almost 00:01:10.420 --> 00:01:13.340 worse because the hallucinations sound so... 00:01:13.980 --> 00:01:16.659 Convincing now. The grammar is perfect. He wears 00:01:16.659 --> 00:01:19.879 a suit and tie now. Exactly. So to navigate this, 00:01:19.959 --> 00:01:21.920 we're digging into a technical guide by Max Am. 00:01:22.079 --> 00:01:25.560 It's called Why ChatGPT, Gemini, Claude are Still 00:01:25.560 --> 00:01:29.200 Hallucinating and How to Fix It. And the core 00:01:29.200 --> 00:01:31.799 idea is that we need to change how we work with 00:01:31.799 --> 00:01:34.420 these tools. to move from just a casual chat 00:01:34.420 --> 00:01:37.859 to something ann calls auditability auditability 00:01:37.859 --> 00:01:41.420 that is the skill for 2026. seriously it's not 00:01:41.420 --> 00:01:43.180 about finding an answer anymore the machine does 00:01:43.180 --> 00:01:45.819 that it's about proving the answer is real okay 00:01:45.819 --> 00:01:47.969 so let's start with the why Why does it lie? 00:01:48.129 --> 00:01:49.790 I think a lot of us still assume that if it sounds 00:01:49.790 --> 00:01:51.849 smart, it must be right. Right. But that's the 00:01:51.849 --> 00:01:53.810 wrong frame. It isn't lying out of malice. It 00:01:53.810 --> 00:01:55.750 isn't even broken. The irony is it's lying out 00:01:55.750 --> 00:01:58.349 of kindness or what it thinks is kindness. These 00:01:58.349 --> 00:02:00.129 models at their core, they're just prediction 00:02:00.129 --> 00:02:02.329 engines. They're trained to give you the most 00:02:02.329 --> 00:02:04.829 likely next word that will satisfy you. So its 00:02:04.829 --> 00:02:07.709 goal isn't be truthful. It's complete the sentence 00:02:07.709 --> 00:02:10.270 in a pleasing way. Don't disappoint the user. 00:02:10.509 --> 00:02:12.610 That's the core directive. So when it scans a 00:02:12.610 --> 00:02:14.409 file you upload and it can't find the answer. 00:02:15.259 --> 00:02:17.759 It has a choice. It can say, I don't know, which 00:02:17.759 --> 00:02:20.039 feels like a failure to its reward system. Or 00:02:20.039 --> 00:02:23.159 it can switch gears. It switches from retrieval 00:02:23.159 --> 00:02:25.560 mode to generative mode. It just starts guessing 00:02:25.560 --> 00:02:28.139 to fill the gap. The guide has this terrifying 00:02:28.139 --> 00:02:31.319 example with Apple's Q2 revenue. Walk us through 00:02:31.319 --> 00:02:33.000 that because it's not just a wrong number. It's 00:02:33.000 --> 00:02:35.639 like a whole work of fiction. It's the classic. 00:02:36.250 --> 00:02:39.490 polished lie. So you upload a dense financial 00:02:39.490 --> 00:02:42.110 report. You ask, what was Apple's Q2 revenue? 00:02:42.389 --> 00:02:44.669 But let's say the document doesn't use that exact 00:02:44.669 --> 00:02:47.150 phrase. Maybe it says second quarter fiscal results. 00:02:47.490 --> 00:02:50.009 So the simple text search fails. It fails. But 00:02:50.009 --> 00:02:51.729 instead of saying, I can't find that, the training 00:02:51.729 --> 00:02:53.810 kicks in. It starts thinking, well, Apple's a 00:02:53.810 --> 00:02:56.550 big company. Q2 is usually around March. And 00:02:56.550 --> 00:02:59.169 it just constructs an answer. In the example, 00:02:59.189 --> 00:03:03.169 it says $95 .4 billion for the fiscal 2025 second 00:03:03.169 --> 00:03:06.689 quarter, ending March 29, 2025. That is so specific. 00:03:06.729 --> 00:03:09.810 It has a dollar amount, a decimal, a date. And 00:03:09.810 --> 00:03:13.090 every single piece of it is hallucinated. It's 00:03:13.090 --> 00:03:15.069 just a statistical guess of what an earnings 00:03:15.069 --> 00:03:16.629 report should look like. You can immediately 00:03:16.629 --> 00:03:20.030 see the danger zones and the list of few like 00:03:20.030 --> 00:03:23.050 invoicing. Oh, that's a huge risk. You have automated 00:03:23.050 --> 00:03:27.689 systems processing invoices and the AI just hallucinates 00:03:27.689 --> 00:03:30.330 a line item for a shipping fee because it thinks 00:03:30.330 --> 00:03:32.930 there should be one. And you just overpay. By 00:03:32.930 --> 00:03:35.729 thousands. And no human would ever catch it because 00:03:35.729 --> 00:03:38.270 it looks normal. Insurance is another one. That's 00:03:38.270 --> 00:03:40.909 the nightmare. You ask, am I covered for flood 00:03:40.909 --> 00:03:43.069 damage? It sees a water damage clause and says, 00:03:43.169 --> 00:03:45.349 yep, you're covered. Then the flood hits and 00:03:45.349 --> 00:03:47.750 you find out your policy specifically excludes 00:03:47.750 --> 00:03:50.650 rising water. The AI just gave you a standard 00:03:50.650 --> 00:03:53.129 answer that wasn't true for you. And contracts, 00:03:53.409 --> 00:03:56.669 like my story. Miscompliance, wrong dates. It's 00:03:56.669 --> 00:03:59.129 the halo effect of good presentation. We see 00:03:59.129 --> 00:04:01.830 polish and our brain automatically assumes it's 00:04:01.830 --> 00:04:04.310 true. I have to admit, I still struggle with 00:04:04.310 --> 00:04:06.900 that. The blind trust. When the output looks 00:04:06.900 --> 00:04:09.659 that professional, the formatting is clean, the 00:04:09.659 --> 00:04:12.340 grammar is perfect. Yeah. It is just so hard 00:04:12.340 --> 00:04:14.120 for my brain to stop and say, wait a minute, 00:04:14.199 --> 00:04:17.680 is this real? It's a cognitive bias. We're wired 00:04:17.680 --> 00:04:20.000 to believe that competence in form means competence 00:04:20.000 --> 00:04:23.660 in fact. But with AI, those two things are totally 00:04:23.660 --> 00:04:26.300 decoupled. So this is the big question then. 00:04:26.399 --> 00:04:30.259 If the machine prioritizes being helpful over 00:04:30.259 --> 00:04:33.500 being truthful. how do we force it to care about 00:04:33.500 --> 00:04:35.360 the truth? We have to change the instructions. 00:04:35.439 --> 00:04:37.040 We have to make, I don't know, an acceptable, 00:04:37.180 --> 00:04:40.319 even a rewarded answer. And the guide says there 00:04:40.319 --> 00:04:43.100 are three layers to this. The hardware, the software, 00:04:43.300 --> 00:04:45.779 and the verification. Okay, let's start with 00:04:45.779 --> 00:04:48.959 layer one, the hardware. Model selection. This 00:04:48.959 --> 00:04:50.740 seems basic, but the guide says most people mess 00:04:50.740 --> 00:04:53.230 this up right at the start. Yep. They fail before 00:04:53.230 --> 00:04:55.730 they even type a word. They just open ChatGPT 00:04:55.730 --> 00:04:58.329 or whatever and use the default model that loads. 00:04:58.569 --> 00:05:00.310 I'm guilty of that. I just open the window and 00:05:00.310 --> 00:05:02.949 go, I assume it's the smartest one. And for serious 00:05:02.949 --> 00:05:05.730 document work in 2026, that's a critical mistake. 00:05:06.009 --> 00:05:08.870 The default models, the turbo or flash versions, 00:05:09.029 --> 00:05:11.670 they're built for speed. For conversation. Right. 00:05:11.750 --> 00:05:14.129 They are not built for deep forensic analysis. 00:05:14.290 --> 00:05:16.730 You have to manually toggle on the reasoning 00:05:16.730 --> 00:05:18.750 capabilities. So we're talking about specific 00:05:18.750 --> 00:05:22.220 settings, not just using GPT -5. Exactly. For 00:05:22.220 --> 00:05:30.139 ChatGPT, the... For Claude, it's Opus 4 .6 with 00:05:30.139 --> 00:05:33.160 extended reasoning on it. And for Google, Gemini 00:05:33.160 --> 00:05:36.060 3 Pro. What's the actual difference? Is it just 00:05:36.060 --> 00:05:38.439 slower? It is slower, but that's the whole point. 00:05:38.560 --> 00:05:41.699 The default model is like an improv comic. Fast, 00:05:41.860 --> 00:05:44.060 creative, good at making connections. But you 00:05:44.060 --> 00:05:45.920 don't want an improv comic doing your accounting. 00:05:46.199 --> 00:05:48.319 You want the deep thinker. The part of the brain 00:05:48.319 --> 00:05:50.319 that checks its work. That's what high reasoning 00:05:50.319 --> 00:05:53.079 does. It forces the model to spend more compute 00:05:53.079 --> 00:05:55.360 cycles on a chain of thought before it gives 00:05:55.360 --> 00:05:57.220 you an answer. It has an internal monologue. 00:05:57.379 --> 00:05:59.300 It's checking its own work before it speaks. 00:05:59.800 --> 00:06:01.879 Precisely. So does picking the right model solve 00:06:01.879 --> 00:06:04.379 the whole problem? If I just switch to Opus 4 00:06:04.379 --> 00:06:07.360 .6, am I good to go? Nope, absolutely not. That's 00:06:07.360 --> 00:06:09.319 just the baseline. It gets you a smart auditor, 00:06:09.500 --> 00:06:11.819 but even a smart auditor needs clear instructions. 00:06:12.199 --> 00:06:15.500 And that brings us to layer two. the software 00:06:15.500 --> 00:06:17.860 the prompts this is where the guide introduces 00:06:17.860 --> 00:06:20.420 the grounding complete template right which it 00:06:20.420 --> 00:06:22.879 sounds very official it is and it works this 00:06:22.879 --> 00:06:24.779 whole section is about giving the ai permission 00:06:24.779 --> 00:06:27.980 to fail permission to fail i love that okay so 00:06:27.980 --> 00:06:31.279 what's the first prompt prompt one is the grounding 00:06:31.279 --> 00:06:34.980 rule you say base your answer only on the uploaded 00:06:34.980 --> 00:06:38.410 documents Nothing else. Nothing else. That's 00:06:38.410 --> 00:06:40.790 the key part. That little phrase does so much 00:06:40.790 --> 00:06:43.209 work. It stops the model from using its vast 00:06:43.209 --> 00:06:45.930 internet training data. It shrinks its universe 00:06:45.930 --> 00:06:48.790 down to just the PDF you gave it. Okay, prompt 00:06:48.790 --> 00:06:51.790 two. This one tackles the people pleaser problem. 00:06:52.069 --> 00:06:54.129 Right. This is where you say, if information 00:06:54.129 --> 00:06:57.029 isn't found, say, not found in the documents, 00:06:57.250 --> 00:06:59.910 don't guess. It seems so simple. You just have 00:06:59.910 --> 00:07:01.470 to tell it not to guess. You have to be that 00:07:01.470 --> 00:07:03.430 direct. You're overwriting its core training. 00:07:03.649 --> 00:07:05.589 And what's wild is that this isn't some clever 00:07:05.589 --> 00:07:09.029 hack. Anthropix's own API docs recommend this. 00:07:09.209 --> 00:07:11.529 Oh, really? Yeah. They basically say if you want 00:07:11.529 --> 00:07:13.490 accuracy, you have to tell the model it's okay 00:07:13.490 --> 00:07:15.509 to be silent. So it's a manufacturer -approved 00:07:15.509 --> 00:07:18.509 fix. Okay. And prompt three. Demand citations. 00:07:18.689 --> 00:07:22.149 For each claim, cite the specific location, document 00:07:22.149 --> 00:07:24.990 name, page section, and relevant quotes. Show 00:07:24.990 --> 00:07:27.829 me the receipts. Always get the receipts. This 00:07:27.829 --> 00:07:30.889 forces it to be a data auditor, not a creative 00:07:30.889 --> 00:07:33.050 writer. If it can't point to the line on the 00:07:33.050 --> 00:07:35.250 page, it can't make the claim. The guide also 00:07:35.250 --> 00:07:37.350 has a couple of bonus prompts. One is the middle 00:07:37.350 --> 00:07:39.430 ground. Yeah, that's where you ask it to mark 00:07:39.430 --> 00:07:42.910 things it's unsure about as unverified. It's 00:07:42.910 --> 00:07:44.930 great for summaries of long reports where you 00:07:44.930 --> 00:07:46.910 just need to know where the shaky ground is. 00:07:47.209 --> 00:07:50.149 But then there's the nuclear option for high 00:07:50.149 --> 00:07:52.870 stakes stuff. The high stakes mode. The prompt 00:07:52.870 --> 00:07:57.670 is only respond. If you're 100 % confident. Whoa. 00:07:58.350 --> 00:08:01.750 That feels intense. It is. And your answers will 00:08:01.750 --> 00:08:03.410 be much shorter. You'll get, I don't know, a 00:08:03.410 --> 00:08:05.810 lot more. But the answers you do get will be 00:08:05.810 --> 00:08:09.009 rock solid. Imagine the discipline required to 00:08:09.009 --> 00:08:12.230 stay silent unless you are 100 % sure. That's 00:08:12.230 --> 00:08:14.550 a standard most humans can't even meet. It is, 00:08:14.569 --> 00:08:17.250 but that's the choice. Do you want a chatty assistant 00:08:17.250 --> 00:08:20.209 or a rigorous auditor? So we have the right brain, 00:08:20.310 --> 00:08:23.829 the right instructions. Do we trust it now? Are 00:08:23.829 --> 00:08:26.050 we finally done? Not yet. Even with all that, 00:08:26.110 --> 00:08:28.089 you can't blindly trust it. You need an independent 00:08:28.089 --> 00:08:31.370 audit. Okay, we're back. We've picked a reasoning 00:08:31.370 --> 00:08:33.970 model. We've used the grounding prompts. But 00:08:33.970 --> 00:08:36.570 the guide says we're still not done. We have 00:08:36.570 --> 00:08:39.649 to verify. Trust, but verify. But the cool thing 00:08:39.649 --> 00:08:42.710 about doing this in 2026 is that verify doesn't 00:08:42.710 --> 00:08:44.830 mean you have to reread the whole document yourself. 00:08:45.129 --> 00:08:46.730 Right, because that would defeat the whole purpose. 00:08:46.929 --> 00:08:51.580 Exactly. The new concept is AI checking AI. I 00:08:51.580 --> 00:08:54.679 like that. A digital second set of eyes. So what's 00:08:54.679 --> 00:08:57.159 method one, the low intensity version? The self 00:08:57.159 --> 00:08:59.679 -check. This is the easiest one. You just stay 00:08:59.679 --> 00:09:02.659 in the same chat window and ask. Rescan the document. 00:09:02.740 --> 00:09:05.059 If you can't find the quote, take the claim back. 00:09:05.580 --> 00:09:07.639 Does the word race scan actually do something 00:09:07.639 --> 00:09:10.559 special? It does. It's a critical word. It forces 00:09:10.559 --> 00:09:13.559 the model to perform a new methodological review 00:09:13.559 --> 00:09:15.679 instead of just, you know, confirming its own 00:09:15.679 --> 00:09:17.820 previous answer. It avoids confirmation bias. 00:09:18.139 --> 00:09:20.139 Okay, so that's the quick check. Method two is 00:09:20.139 --> 00:09:22.159 the multi -model check. This is your medium intensity 00:09:22.159 --> 00:09:25.320 option. You take the output from, say, ChatGPT, 00:09:25.399 --> 00:09:27.940 and you feed it, plus the original file, into 00:09:27.940 --> 00:09:30.919 a different model, like Claude Opus 4 .6. And 00:09:30.919 --> 00:09:34.779 you ask Claude to grade ChatGPT's homework. Basically, 00:09:34.879 --> 00:09:37.710 yeah. And the analogy in the guide is perfect. 00:09:38.090 --> 00:09:40.450 It's like getting a second opinion from a doctor 00:09:40.450 --> 00:09:42.669 who went to a different medical school. They 00:09:42.669 --> 00:09:45.070 have different training, different biases. The 00:09:45.070 --> 00:09:47.269 chance that they will both make the exact same 00:09:47.269 --> 00:09:51.009 weird mistake is incredibly low. That makes a 00:09:51.009 --> 00:09:53.490 lot of sense. Okay, then we have the heavy hitter, 00:09:53.509 --> 00:09:57.610 method three, notebook LM. This is the gold standard 00:09:57.610 --> 00:09:59.590 for this kind of work today. Google's notebook 00:09:59.590 --> 00:10:02.769 LM running on Gemini 3. I've used it for research. 00:10:02.929 --> 00:10:05.190 Why is it considered the highest intensity check? 00:10:05.389 --> 00:10:07.809 Because it was purpose built for this task. It's 00:10:07.809 --> 00:10:09.970 not a chat bot trying to be an auditor. It is 00:10:09.970 --> 00:10:12.549 an auditor. You upload your document and the 00:10:12.549 --> 00:10:15.210 AI's analysis and you ask which claims in this 00:10:15.210 --> 00:10:17.769 analysis are not supported by the source document. 00:10:18.110 --> 00:10:21.139 And here's the killer feature. It links every 00:10:21.139 --> 00:10:24.320 single claim. It verifies directly to the source 00:10:24.320 --> 00:10:26.940 text. You can click a little citation number 00:10:26.940 --> 00:10:29.220 and it takes you to the exact paragraph in the 00:10:29.220 --> 00:10:31.039 PDF. So it's not just telling you it's true. 00:10:31.100 --> 00:10:33.460 It's showing you the proof. It creates a verifiable 00:10:33.460 --> 00:10:35.860 paper trail. That's the definition of auditability. 00:10:36.200 --> 00:10:38.279 This whole system sounds incredibly thorough. 00:10:38.919 --> 00:10:41.840 Hardware, software, verification. But I have 00:10:41.840 --> 00:10:44.559 to ask, is there anything this system cannot 00:10:44.559 --> 00:10:47.519 catch? Is there still a gap where you need a 00:10:47.519 --> 00:10:51.639 human? Yes, a big one. It can't fix a bad source 00:10:51.639 --> 00:10:54.379 file. And more importantly, it can't understand 00:10:54.379 --> 00:10:57.179 human risk. So let's talk about that, the reality 00:10:57.179 --> 00:10:59.799 check section. What is this grounded mode not 00:10:59.799 --> 00:11:02.679 good for? Well, first, it's not magic. If your 00:11:02.679 --> 00:11:05.259 PDF has missing pages or the scan quality is 00:11:05.259 --> 00:11:08.019 terrible, the AI can't fix that. The guide has 00:11:08.019 --> 00:11:10.620 a great rule of thumb for this. If a junior analyst 00:11:10.620 --> 00:11:13.519 couldn't answer it, neither should the AI. Exactly. 00:11:13.740 --> 00:11:16.139 Garbage in, garbage out. But the much bigger 00:11:16.139 --> 00:11:19.190 limitation is risk assessment. The AI can extract 00:11:19.190 --> 00:11:22.190 a liability clause perfectly. It can tell you 00:11:22.190 --> 00:11:25.450 verbatim liabilities capped at $5 ,000. But it 00:11:25.450 --> 00:11:27.870 has no idea if signing a contract with that clause 00:11:27.870 --> 00:11:30.110 is a terrible idea for your business. Right. 00:11:30.190 --> 00:11:32.269 It can read the map perfectly, but it can't tell 00:11:32.269 --> 00:11:33.870 you there's a cliff at the end of the road. That 00:11:33.870 --> 00:11:36.169 takes a lawyer. That takes context about your 00:11:36.169 --> 00:11:38.389 business, the market, your negotiating power. 00:11:38.570 --> 00:11:41.669 The AI is an extractor, not a strategist. The 00:11:41.669 --> 00:11:44.820 guide also mentions math. I feel like we keep 00:11:44.820 --> 00:11:46.799 hearing AI is getting better at math, but it's 00:11:46.799 --> 00:11:49.379 still a weak spot. With dense tables, yeah. It's 00:11:49.379 --> 00:11:52.100 the column confusion problem. It still confuses 00:11:52.100 --> 00:11:55.419 totals and subtotals. It skips footnotes. It 00:11:55.419 --> 00:11:58.340 basically treats numbers like words. So what's 00:11:58.340 --> 00:12:02.019 the fix? Don't use it for math. You slow it down. 00:12:02.220 --> 00:12:05.320 You ask it to first reproduce the table exactly 00:12:05.320 --> 00:12:08.179 as it sees it in the chat. Once you confirm it, 00:12:08.220 --> 00:12:10.120 read the numbers right, then you ask it to do 00:12:10.120 --> 00:12:12.200 the math. Show your work. It's like we're back 00:12:12.200 --> 00:12:14.139 in fourth grade math class. Always make it show 00:12:14.139 --> 00:12:17.000 its work. And the final limitation is creativity 00:12:17.000 --> 00:12:20.820 versus precision. Models just love to paraphrase. 00:12:20.820 --> 00:12:23.480 It's in their nature. But for legal or compliance 00:12:23.480 --> 00:12:26.299 work, similar is not good enough. You need the 00:12:26.299 --> 00:12:29.340 exact words. So you have to explicitly command 00:12:29.340 --> 00:12:32.480 it. Quote verbatim. Do not paraphrase. You have 00:12:32.480 --> 00:12:34.940 to be that blunt. So if we zoom out on all this, 00:12:35.000 --> 00:12:36.659 it really sounds like the goal isn't to replace 00:12:36.659 --> 00:12:39.340 the human. It's just changing the human's job 00:12:39.340 --> 00:12:43.259 description. 100%. The human is no longer the 00:12:43.259 --> 00:12:45.399 researcher digging in the file cabinet. The human 00:12:45.399 --> 00:12:48.360 is the auditor, the one checking the citations, 00:12:48.419 --> 00:12:50.919 assessing the risk, and making the final judgment 00:12:50.919 --> 00:12:54.240 call. It's a real shift from finding answers 00:12:54.240 --> 00:12:57.360 to verifying truth. And that shift is everything. 00:12:57.639 --> 00:13:00.059 It's what separates people who get value from 00:13:00.059 --> 00:13:03.299 AI from those who just get noise. Okay, let's 00:13:03.299 --> 00:13:05.830 recap the big idea. If someone listening has 00:13:05.830 --> 00:13:07.850 a stack of documents on their desk for tomorrow 00:13:07.850 --> 00:13:10.690 morning, what is the simple playbook? Three steps. 00:13:10.950 --> 00:13:13.690 One, pick a reasoning model. Don't use the default. 00:13:13.769 --> 00:13:17.009 Use GBT 5 .3 or Opus 4 .6 with the reasoning 00:13:17.009 --> 00:13:20.490 modes on. Two, ground it. Paste in that template. 00:13:20.590 --> 00:13:22.629 Tell it to only use the file and give it permission 00:13:22.629 --> 00:13:26.070 to say, I don't know. And three, verify. Use 00:13:26.070 --> 00:13:29.490 a second AI or ideally notebook LM to cross -check 00:13:29.490 --> 00:13:31.450 the key facts. It sounds like a little bit of 00:13:31.450 --> 00:13:33.909 extra work upfront. It's maybe 30 seconds more 00:13:33.909 --> 00:13:36.289 per task. But it saves you hours of panic and 00:13:36.289 --> 00:13:39.149 cleanup later. Trust but verify. An old saying, 00:13:39.350 --> 00:13:41.769 but it feels more important now than ever before. 00:13:41.909 --> 00:13:44.049 It's the only way to operate in an age of infinite 00:13:44.049 --> 00:13:46.789 confident information. So here's the challenge 00:13:46.789 --> 00:13:48.690 for you listening. The next time you upload a 00:13:48.690 --> 00:13:51.570 file, don't just ask it to summarize. Try that 00:13:51.570 --> 00:13:54.110 grounding template. And just watch how different 00:13:54.110 --> 00:13:56.789 the output looks when the AI stops trying to 00:13:56.789 --> 00:13:59.129 guess and starts citing its sources. It gets 00:13:59.129 --> 00:14:02.029 remarkably quiet. And that quiet, that's the 00:14:02.029 --> 00:14:04.700 sound of accuracy. I love that. Thank you for 00:14:04.700 --> 00:14:06.299 deep diving with us today. We'll catch you on 00:14:06.299 --> 00:14:06.679 the next one.