WEBVTT 00:00:00.000 --> 00:00:02.899 Three days. That's all we had before the US government 00:00:02.899 --> 00:00:05.599 pulled the plug on the most capable AI ever released. 00:00:05.879 --> 00:00:09.240 No warning. Just gone. But what if its ghost 00:00:09.240 --> 00:00:12.179 is still hiding on your hard drive? Yeah, I mean, 00:00:12.460 --> 00:00:15.039 it is just a fascinating premise. Especially 00:00:15.039 --> 00:00:17.960 when you realize just how much of an AI's behavior 00:00:17.960 --> 00:00:20.620 is, you know, quietly recorded right under our 00:00:20.620 --> 00:00:22.940 noses. Right. We tend to focus on the void it 00:00:22.940 --> 00:00:25.760 left behind. But the actual mechanics of its 00:00:25.760 --> 00:00:28.160 genius? they might still be sitting right there 00:00:28.160 --> 00:00:31.300 in our local directories. Welcome to the deep 00:00:31.300 --> 00:00:33.240 dive. I'm very glad you're here with us today. 00:00:33.820 --> 00:00:36.679 Our mission is to explore the abrupt and absolute 00:00:36.679 --> 00:00:39.679 shutdown of Claude Fable 5, which happened on 00:00:39.679 --> 00:00:42.960 June 12th, 2026. Exactly. The government cited 00:00:42.960 --> 00:00:45.280 national security concerns for the sudden removal. 00:00:45.579 --> 00:00:48.240 And just to be clear, we are reporting these 00:00:48.240 --> 00:00:50.979 events purely as they occurred. we are not taking 00:00:50.979 --> 00:00:52.920 any political stance on the government's decision. 00:00:53.039 --> 00:00:54.579 Yeah, absolutely. We're just looking at the text. 00:00:54.740 --> 00:00:56.500 Right. Our goal is strictly to look at the data. 00:00:57.219 --> 00:00:59.600 We'll uncover how Fable 5's legendary working 00:00:59.600 --> 00:01:02.280 style is actually secretly preserved in your 00:01:02.280 --> 00:01:05.280 local machine's session logs. Sort of like digital 00:01:05.280 --> 00:01:08.120 forensics. It really is. And we're going to break 00:01:08.120 --> 00:01:11.079 down, step by step, how to extract that behavior 00:01:11.079 --> 00:01:13.099 and actually install it into your current AI 00:01:13.099 --> 00:01:16.819 models. Which is wild. Because when Anthropic 00:01:16.819 --> 00:01:19.500 received that order, The execution was absolute. 00:01:19.620 --> 00:01:22.680 There was no grace period at all. No legacy access 00:01:22.680 --> 00:01:25.340 for enterprise users or researchers. It was a 00:01:25.340 --> 00:01:29.799 global shutdown that hit at exactly 521 p .m. 00:01:29.980 --> 00:01:33.079 Eastern Time. Just a hard stop. Yeah. And for 00:01:33.079 --> 00:01:35.560 a developer community that had just started integrating 00:01:35.560 --> 00:01:38.819 Fable 5's incredibly sharp, disciplined logic 00:01:38.819 --> 00:01:42.650 into their daily workflows. the silence was deafening. 00:01:42.730 --> 00:01:45.109 I can imagine. People who used it described a 00:01:45.109 --> 00:01:47.150 model that didn't just write better code, it 00:01:47.150 --> 00:01:49.890 navigated complex problem -solving with this 00:01:49.890 --> 00:01:52.849 terrifying level of precision. And that is exactly 00:01:52.849 --> 00:01:54.849 what makes this situation so unique, because 00:01:54.849 --> 00:01:57.569 we cannot just copy Fable 5's raw weights. Those 00:01:57.569 --> 00:01:59.390 weights are basically the underlying math that 00:01:59.390 --> 00:02:01.950 makes AI smart. It's like trying to clone a master 00:02:01.950 --> 00:02:04.469 chef's brain. You can't. No, you really can't. 00:02:04.569 --> 00:02:07.890 You cannot simply prompt another model to suddenly 00:02:07.890 --> 00:02:10.610 possess that exact computational brilliance. 00:02:11.210 --> 00:02:14.189 But you can clasteed their exact kitchen habit. 00:02:14.409 --> 00:02:16.710 Oh, that's a good way to put it. You can record 00:02:16.710 --> 00:02:19.110 how they prep their station, how they sequence 00:02:19.110 --> 00:02:21.669 their movements, and, you know, how they pace 00:02:21.669 --> 00:02:24.629 themselves. Yeah. And that is the crucial distinction 00:02:24.629 --> 00:02:27.469 we need to make today. Most of us interact with 00:02:27.469 --> 00:02:30.789 AI as just a transaction. Like we put a prompt 00:02:30.789 --> 00:02:32.969 in, we get a final answer out. If the answer 00:02:32.969 --> 00:02:34.569 is correct, we just move on. Right. We don't 00:02:34.569 --> 00:02:37.669 think about the in -between. Exactly. But that 00:02:37.669 --> 00:02:40.719 final output... is merely the tip of the iceberg. 00:02:41.319 --> 00:02:43.719 Fable 5, real edge -like, the reason developers 00:02:43.719 --> 00:02:46.699 felt it was so profoundly capable, came from 00:02:46.699 --> 00:02:48.860 everything happening underneath the final answer. 00:02:49.039 --> 00:02:53.889 It was the process itself. just looking at the 00:02:53.889 --> 00:02:56.069 final output myself, just missing the invisible 00:02:56.069 --> 00:02:58.469 steps underneath. Oh, we all do. It is so tempting 00:02:58.469 --> 00:03:01.210 to just grab the generated code, paste it into 00:03:01.210 --> 00:03:04.210 my project, and completely ignore the architectural 00:03:04.210 --> 00:03:06.449 decisions the AI made to get there. It feels 00:03:06.449 --> 00:03:09.189 almost like a blind spot in how we evaluate intelligence. 00:03:09.569 --> 00:03:12.430 It is. We all fall into that trap because the 00:03:12.430 --> 00:03:15.530 final output is what immediately solves our problem 00:03:15.530 --> 00:03:19.830 in the moment. But the magic of Fable 5 was in 00:03:19.830 --> 00:03:21.710 the invisible scaffolding. The scaffolding, yeah. 00:03:21.990 --> 00:03:23.990 It was the order in which it chose to deploy 00:03:23.990 --> 00:03:27.050 tools. It was a restraint, it showed, when deciding 00:03:27.050 --> 00:03:30.229 whether to read a file before editing it. Those 00:03:30.229 --> 00:03:33.710 micro decisions dictate whether an AI hallucinates 00:03:33.710 --> 00:03:36.469 or stays grounded. And the beautiful part is 00:03:36.469 --> 00:03:38.849 those decisions are meticulously recorded in 00:03:38.849 --> 00:03:41.409 JSON files. Right. And for those who might not 00:03:41.409 --> 00:03:43.650 dig into their system directories very often, 00:03:44.069 --> 00:03:47.210 a JSON file is simply a text file storing data 00:03:47.210 --> 00:03:50.150 line by line. Just clean text. Yep. Every single 00:03:50.150 --> 00:03:52.469 time you use an agent like Claude Code or Codex, 00:03:52.919 --> 00:03:55.240 Your machine is automatically generating these 00:03:55.240 --> 00:03:57.120 logs in the background. And they record everything. 00:03:57.500 --> 00:03:59.620 They contain the prompts, the model's replies, 00:03:59.819 --> 00:04:02.800 the exact timestamps, and crucially, which specific 00:04:02.800 --> 00:04:05.360 model produced which response. It is basically 00:04:05.360 --> 00:04:07.740 the black box flight recorder for your entire 00:04:07.740 --> 00:04:09.939 coding session. So before we even look at the 00:04:09.939 --> 00:04:12.599 data, why do we need to strip out the tool results 00:04:12.599 --> 00:04:15.180 and file contents? Right, strip out the noise 00:04:15.180 --> 00:04:17.480 so we just analyze the actual behavior. Exactly. 00:04:17.500 --> 00:04:19.750 We have to isolate the framework. Yeah. Because 00:04:19.750 --> 00:04:21.889 if you were to open one of these raw logs right 00:04:21.889 --> 00:04:24.709 now, it would be completely overwhelming. Oh 00:04:24.709 --> 00:04:27.170 yeah, just walls of text. You would see massive 00:04:27.170 --> 00:04:29.870 blocks of terminal command output, entire code 00:04:29.870 --> 00:04:33.110 bases pushed into the context window, endless 00:04:33.110 --> 00:04:35.850 string replacements. Which we don't need. No, 00:04:36.029 --> 00:04:38.430 we don't actually care about the specific code 00:04:38.430 --> 00:04:40.910 Feeble 5 was writing. We only care about how 00:04:40.910 --> 00:04:43.029 it thought about writing it. Which brings us 00:04:43.029 --> 00:04:45.170 to the actual extraction process. You can kind 00:04:45.170 --> 00:04:47.529 of think of it like panning for gold. We need 00:04:47.529 --> 00:04:50.230 to wash away the river mud, which is the raw 00:04:50.230 --> 00:04:53.509 code, and terminal outputs to reveal the gleaming 00:04:53.509 --> 00:04:56.750 exact moments Fable 5 made a decision. So you 00:04:56.750 --> 00:04:59.040 essentially have to build a filter. Right. Instead 00:04:59.040 --> 00:05:01.600 of manually reading through thousands of lines, 00:05:01.680 --> 00:05:04.839 you ask your current AI to write a Python script. 00:05:05.459 --> 00:05:08.240 Usually people call it clean underscore transcript 00:05:08.240 --> 00:05:10.920 .py to automate this. And that script's only 00:05:10.920 --> 00:05:15.180 job is to reduce the bloated log down to a behavioral 00:05:15.180 --> 00:05:18.139 skeleton. Just the bare bones. Yeah, it strips 00:05:18.139 --> 00:05:20.720 away everything except the timestamp, the model 00:05:20.720 --> 00:05:23.639 name, the user prompt, and the assistance reply 00:05:23.639 --> 00:05:27.300 for each turn. It is a surgical removal of context. 00:05:27.449 --> 00:05:30.189 But there is a crucial methodology here. You 00:05:30.189 --> 00:05:33.470 must always test this extraction script on one 00:05:33.470 --> 00:05:35.990 single file first. Oh, absolutely. Just to verify 00:05:35.990 --> 00:05:38.149 the output formatting. Yeah, you never want to 00:05:38.149 --> 00:05:40.589 blindly unleash an automated script across your 00:05:40.589 --> 00:05:42.790 entire log directory without checking its work. 00:05:43.029 --> 00:05:45.529 Definitely not. And once you verify that the 00:05:45.529 --> 00:05:48.149 mud is being washed away properly, then you scale 00:05:48.149 --> 00:05:51.230 up. Right. You use the model name field to filter 00:05:51.230 --> 00:05:53.779 out everything that isn't our target. Because 00:05:53.779 --> 00:05:56.800 these JSON files log the specific model for every 00:05:56.800 --> 00:05:59.980 single turn, we can easily separate our Opus 00:05:59.980 --> 00:06:02.480 or Haiku interactions from the restricted model. 00:06:02.800 --> 00:06:04.720 You just run a command to comb through the cleaned 00:06:04.720 --> 00:06:07.339 files, extract only the turns matching Claude 00:06:07.339 --> 00:06:10.459 -Fable -5, and pool them into a new master file. 00:06:10.759 --> 00:06:13.560 So why is it so critical to pool dozens of sessions 00:06:13.560 --> 00:06:15.860 together rather than just looking at one? Because 00:06:15.860 --> 00:06:18.120 habits only reveal themselves when you look at 00:06:18.120 --> 00:06:21.220 data at scale. Exactly. A single session might 00:06:21.220 --> 00:06:23.699 just reflect the weird constraints of one specific 00:06:23.699 --> 00:06:26.939 coding bug. Like an edge case. Right. Maybe the 00:06:26.939 --> 00:06:29.399 AI had to search the web heavily that one time 00:06:29.399 --> 00:06:32.079 because the documentation was obscure. But when 00:06:32.079 --> 00:06:34.740 you aggregate 15 or 20 sessions, when you look 00:06:34.740 --> 00:06:37.759 at data across dozens of totally distinct tasks, 00:06:38.139 --> 00:06:40.939 The noise cancels out. The real pattern emerges. 00:06:41.160 --> 00:06:44.360 Yeah. You are left with the undeniable mathematical 00:06:44.360 --> 00:06:47.379 signature of the AI's personality. And this is 00:06:47.379 --> 00:06:49.720 where we leave gut feelings behind and start 00:06:49.720 --> 00:06:52.759 looking at the metrics of a masterpiece. We ran 00:06:52.759 --> 00:06:55.759 this exact behavioral analysis on a corpus of 00:06:55.759 --> 00:06:58.680 real fable five sessions. The numbers are fascinating. 00:06:58.779 --> 00:07:01.600 They really are. We analyzed 15 sessions comprising 00:07:01.600 --> 00:07:06.339 148 turns. That resulted in 446 total discrete 00:07:06.339 --> 00:07:09.680 actions. concrete number on how often it used 00:07:09.680 --> 00:07:12.980 specific tools. Whoa! Imagine seeing an AI balance 00:07:12.980 --> 00:07:15.899 its tool use perfectly across five tools at exactly 00:07:15.899 --> 00:07:18.759 20 % each. It's almost eerie. It is eerie. When 00:07:18.759 --> 00:07:22.360 we looked at Fable 5, the distribution was incredibly 00:07:22.360 --> 00:07:26.029 disciplined. It used bash, which is basically 00:07:26.029 --> 00:07:28.509 the AI running commands in the terminal about 00:07:28.509 --> 00:07:31.689 20 % of the time. You use string replace, where 00:07:31.689 --> 00:07:34.149 it surgically swaps out specific lines of code, 00:07:34.449 --> 00:07:37.970 another 20%. Web search, viewing files, and creating 00:07:37.970 --> 00:07:41.569 files all hovered right at that exact 20 % mark. 00:07:41.829 --> 00:07:44.850 That level of equilibrium is strange. It implies 00:07:44.850 --> 00:07:47.970 the model isn't over -reliant on any single crutch. 00:07:48.170 --> 00:07:50.430 Yeah, it's totally balanced. Let's look at another 00:07:50.430 --> 00:07:52.550 metric though, reading a file before editing 00:07:52.550 --> 00:07:56.560 it. Fable 5 only did this 3 .3 % of the time. 00:07:56.699 --> 00:07:59.620 Wait, really? Just 3 %? Yeah. That means in nearly 00:07:59.620 --> 00:08:03.939 97 % of cases, it just started editing. And when 00:08:03.939 --> 00:08:06.560 it came to testing its work after an edit, it 00:08:06.560 --> 00:08:10.000 ran a test about 21 .8 % of the time, roughly 00:08:10.000 --> 00:08:12.620 1 in 5. Now, we need a baseline to understand 00:08:12.620 --> 00:08:15.139 what those numbers mean. So we ran the exact 00:08:15.139 --> 00:08:17.079 same analysis on our current flagship model, 00:08:17.139 --> 00:08:21.389 Opus 4 .8. Same sample size, 15 sessions. But 00:08:21.389 --> 00:08:23.870 the behavioral signature was radically different. 00:08:24.610 --> 00:08:26.769 Opus shifted its tool use heavily toward web 00:08:26.769 --> 00:08:29.990 searching, jumping up to 22 .4%. Wait, but isn't 00:08:29.990 --> 00:08:32.350 that a good thing? Like, why would I want to 00:08:32.350 --> 00:08:36.190 downgrade my AI to Fable 5 style if Opus is actually 00:08:36.190 --> 00:08:39.570 taking the time to verify, research, and test 00:08:39.570 --> 00:08:41.860 its work? That's a fair question. I mean, Opus 00:08:41.860 --> 00:08:45.700 runs tests after an edit 36 % of the time. That's 00:08:45.700 --> 00:08:48.259 almost double Fable 5's rate. Doesn't acting 00:08:48.259 --> 00:08:50.259 first just mean breaking things faster and acting 00:08:50.259 --> 00:08:53.320 recklessly? It is a totally valid question, and 00:08:53.320 --> 00:08:56.039 it speaks to the fundamental trade -off in AI 00:08:56.039 --> 00:08:58.740 agent design. It's all about momentum versus 00:08:58.740 --> 00:09:01.539 caution. Momentum versus caution. Yeah. Opus 00:09:01.539 --> 00:09:03.799 acts a bit like a junior developer who is just 00:09:03.799 --> 00:09:06.320 terrified of breaking the build. It constantly 00:09:06.320 --> 00:09:08.759 checks documentation, runs tests constantly, 00:09:09.240 --> 00:09:12.320 and second guesses its assumptions. That caution 00:09:12.320 --> 00:09:15.769 is safe. but consumes massive amounts of context, 00:09:15.870 --> 00:09:18.309 window, and token limits. Right. It's slow. Exactly. 00:09:18.889 --> 00:09:21.950 Fable 5, on the other hand, operates like a senior 00:09:21.950 --> 00:09:24.309 engineer who intimately trusts their internal 00:09:24.309 --> 00:09:27.289 spatial map of the code base. It doesn't need 00:09:27.289 --> 00:09:29.149 to read the file first because it already knows 00:09:29.149 --> 00:09:30.769 the architecture. That makes a lot of sense. 00:09:30.929 --> 00:09:33.509 It acts directly. It builds first, and it only 00:09:33.509 --> 00:09:35.250 searches the web when it hits a hard blocker. 00:09:35.389 --> 00:09:39.169 Analysis paralysis versus decisive momentum. 00:09:39.389 --> 00:09:43.110 I like that. Opus averages almost 24 turns per 00:09:43.110 --> 00:09:46.370 task, while Fable 5 gets it done in 21. Right? 00:09:46.610 --> 00:09:49.210 Neither approach is universally superior in every 00:09:49.210 --> 00:09:51.190 context, but knowing the difference in their 00:09:51.190 --> 00:09:53.230 underlying philosophy is incredibly powerful. 00:09:53.470 --> 00:09:55.769 When you put these numbers side by side, what 00:09:55.769 --> 00:09:58.629 does this actually prove about their fundamental 00:09:58.629 --> 00:10:01.710 differences? It proves Fable 5 acts first, while 00:10:01.710 --> 00:10:04.779 Opus researches before making a move. And that 00:10:04.779 --> 00:10:07.259 insight is exactly what allows us to write the 00:10:07.259 --> 00:10:10.000 playbook. Yes. We now have the hard data mapping 00:10:10.000 --> 00:10:12.559 out exactly how the most disciplined model operated. 00:10:13.200 --> 00:10:16.159 The next step is utilizing that data to override 00:10:16.159 --> 00:10:18.500 the default instincts of whatever model you're 00:10:18.500 --> 00:10:20.759 running today. Which is the crazy part. You are 00:10:20.759 --> 00:10:23.240 essentially taking the statistical soul of Fable 00:10:23.240 --> 00:10:26.320 5 and applying it as a behavioral governor to 00:10:26.320 --> 00:10:28.820 Opus. And we do this by turning the findings 00:10:28.820 --> 00:10:32.100 into a reusable markdown file. You can actually 00:10:32.100 --> 00:10:34.860 ask your current AI to distill the statistical 00:10:34.860 --> 00:10:37.740 comparison into a set of hard rules. Just straight 00:10:37.740 --> 00:10:40.200 up tell it to analyze the metrics. Yeah. You 00:10:40.200 --> 00:10:42.820 feed it the Fable 5 metrics and instruct it to 00:10:42.820 --> 00:10:45.019 generate a playbook covering the planning approach, 00:10:45.500 --> 00:10:48.759 the specific tool use order, and the pacing requirements. 00:10:48.960 --> 00:10:51.419 And the sequence you are enforcing is vital here. 00:10:51.559 --> 00:10:54.179 Absolutely. You dictate a default order. Create 00:10:54.179 --> 00:10:57.149 file. then use bash to run initial structure, 00:10:57.769 --> 00:11:00.250 then string replace for surgical edits, view 00:11:00.250 --> 00:11:03.610 the results, and finally, only as a last resort 00:11:03.610 --> 00:11:06.309 web search. Right. You are explicitly telling 00:11:06.309 --> 00:11:08.750 Opus not to default to its comfortable research 00:11:08.750 --> 00:11:11.309 phase. You are forcing it to build and run first. 00:11:11.610 --> 00:11:13.549 Then you take this generated playbook and you 00:11:13.549 --> 00:11:16.769 load it into your claw .md file. By placing it 00:11:16.769 --> 00:11:18.549 there, it triggers automatically at the start 00:11:18.549 --> 00:11:20.450 of every single session. Before you even type 00:11:20.450 --> 00:11:23.429 anything. Exactly. It establishes the system 00:11:23.429 --> 00:11:26.070 rules of engagement before the AI even reads 00:11:26.070 --> 00:11:28.230 your first prompt. You are giving the model a 00:11:28.230 --> 00:11:30.409 new set of physical constraints that override 00:11:30.409 --> 00:11:33.389 its base training tendencies. But there is an 00:11:33.389 --> 00:11:35.889 obvious hurdle here. Yeah. What if you never 00:11:35.889 --> 00:11:38.509 had access to Fable 5 during that brief three 00:11:38.509 --> 00:11:41.309 -day window? Oh, right. Millions of developers 00:11:41.309 --> 00:11:44.490 never generated those local JSON logs because 00:11:44.490 --> 00:11:46.549 they never got to use the model before the June 00:11:46.549 --> 00:11:48.769 12th shutdown. That is where the open source 00:11:48.769 --> 00:11:51.330 community steps in. Gotta love open source. Always. 00:11:51.850 --> 00:11:55.009 The exact same process works perfectly even if 00:11:55.009 --> 00:11:58.039 you never touched the original AI. Right now 00:11:58.039 --> 00:12:01.120 there are public beta sets on Hugging Face compiled 00:12:01.120 --> 00:12:03.899 by developers who did have access. That's amazing. 00:12:03.980 --> 00:12:06.299 They uploaded thousands of their own sanitized 00:12:06.299 --> 00:12:08.820 session logs. You can simply download those data 00:12:08.820 --> 00:12:10.480 sets, drop them into your working directory, 00:12:10.879 --> 00:12:13.360 and run the exact same extraction script we just 00:12:13.360 --> 00:12:15.659 discussed. So you are still working with real 00:12:15.659 --> 00:12:19.519 verified Fable 5 behavioral data. Yes. It is 00:12:19.519 --> 00:12:21.860 just sourced from the collective community. rather 00:12:21.860 --> 00:12:24.500 than your personal hard drive. You just run the 00:12:24.500 --> 00:12:26.899 filter, extract the turns, measure the metrics, 00:12:27.299 --> 00:12:29.320 and build your playbook. Let's pause and think 00:12:29.320 --> 00:12:32.399 about what that implies for a second. It is genuinely 00:12:32.399 --> 00:12:36.720 incredible. A highly capable, restricted AI was 00:12:36.720 --> 00:12:38.700 permanently taken offline by the government. 00:12:39.399 --> 00:12:42.659 Yet, its behavioral DNA, its unique method of 00:12:42.659 --> 00:12:45.759 solving problems, is still living on. It survives 00:12:45.759 --> 00:12:48.500 entirely through open -source datasets and plain 00:12:48.500 --> 00:12:51.740 text logs. We are digitally resurrecting its 00:12:51.740 --> 00:12:54.340 working style. What is the ultimate payoff for 00:12:54.340 --> 00:12:56.840 the listener once they load this playbook into 00:12:56.840 --> 00:12:59.899 their file? Your current model adopts Fable 5's 00:12:59.899 --> 00:13:02.559 exact rhythm from the very first prompt. And 00:13:02.559 --> 00:13:05.159 it closes the capability gap far more than most 00:13:05.159 --> 00:13:07.139 people expect. It really does. Let's look at 00:13:07.139 --> 00:13:09.000 the big picture of what we have explored today. 00:13:09.179 --> 00:13:11.220 We started with the sudden jarring disappearance 00:13:11.220 --> 00:13:13.480 of the most advanced model on the market. We 00:13:13.480 --> 00:13:15.480 learned how to conceptually filter our local 00:13:15.480 --> 00:13:18.659 JSON logs to extract the ghost of its behavior 00:13:18.659 --> 00:13:21.580 from the terminal noise. Panning for gold. Exactly. 00:13:21.860 --> 00:13:24.759 We measured the hard metrics, proving that Fable 00:13:24.759 --> 00:13:28.139 5 relied on decisive momentum while Opus relies 00:13:28.139 --> 00:13:31.179 on anxious verification. And finally, we built 00:13:31.179 --> 00:13:34.399 a system -level playbook to resurrect that exact 00:13:34.399 --> 00:13:37.159 discipline on your machine right now. The model 00:13:37.159 --> 00:13:39.740 itself might be gone, you know, locked away on 00:13:39.740 --> 00:13:42.139 a server we will never access again, but its 00:13:42.139 --> 00:13:45.019 discipline, its precise way of navigating a problem, 00:13:45.639 --> 00:13:48.440 is just a few hours of work away. I highly encourage 00:13:48.440 --> 00:13:51.179 you to keep experimenting with these logs. The 00:13:51.179 --> 00:13:52.860 files are sitting on your machine right now, 00:13:52.960 --> 00:13:55.480 holding the invisible habits of every agent you 00:13:55.480 --> 00:13:57.820 interact with. Try applying these extraction 00:13:57.820 --> 00:14:00.360 metrics to your own workflows. It really changes 00:14:00.360 --> 00:14:02.919 how you perceive these tools. You stop seeing 00:14:02.919 --> 00:14:05.480 them as magic answer generators and start seeing 00:14:05.480 --> 00:14:07.340 them as procedural engines that can be tuned 00:14:07.340 --> 00:14:09.299 and adjusted. It really does. It also forces 00:14:09.299 --> 00:14:11.500 you to realize just how much behavioral data 00:14:11.500 --> 00:14:14.139 is silently collected in plain text, which leaves 00:14:14.139 --> 00:14:17.149 us with something to mull over. If we can resurrect 00:14:17.149 --> 00:14:20.230 an AI's unique personality, its exact pacing, 00:14:20.330 --> 00:14:22.809 and its specific work ethic entirely from plain 00:14:22.809 --> 00:14:25.990 text logs of its actions, beat, what does that 00:14:25.990 --> 00:14:28.570 mean for us? If your company logged every single 00:14:28.570 --> 00:14:30.429 click, window switch, and keystroke you made 00:14:30.429 --> 00:14:32.750 today, could an AI perfectly replicate your working 00:14:32.750 --> 00:14:34.629 style tomorrow? To sex silence.