Pełna transkrypcja odcinka ReadMUltiplex.com: A Simple Way to Make Grok Bot Run Faster and Cost Less.

--> 00:03:32.379 subscription tiers starting at $30 a month. And 00:03:32.379 --> 00:03:35.120 $30 a month. I mean, that sounds like a decent 00:03:35.120 --> 00:03:37.919 chunk of change for a consumer software subscription. 00:03:38.460 --> 00:03:40.520 Definitely. Most of us are used to paying $10 00:03:40.520 --> 00:03:44.120 or $15 for like Spotify or Netflix. But $30 is 00:03:44.120 --> 00:03:46.419 a drop in the bucket compared to the physical 00:03:46.419 --> 00:03:48.500 reality of what is happening on the back end 00:03:48.500 --> 00:03:50.719 every single time you hit enter. Right, because 00:03:50.719 --> 00:03:52.620 when you type a prompt into a generative AI, 00:03:52.900 --> 00:03:54.819 you aren't just triggering a simple database 00:03:54.819 --> 00:03:56.979 lookup. No, it's not like searching for a movie 00:03:56.979 --> 00:04:00.419 title on Netflix. You are forcing a massive neural 00:04:00.419 --> 00:04:03.560 network, we're talking often hundreds of billions 00:04:03.560 --> 00:04:06.259 if not trillions of parameters, to physically 00:04:06.259 --> 00:04:08.400 fire. Just think about the infrastructure for 00:04:08.400 --> 00:04:11.180 a second. SpaceX AI is carrying the cost of the 00:04:11.180 --> 00:04:13.659 energy to power thousands of server racks. Yeah, 00:04:13.719 --> 00:04:16.759 massive storage arrays, specialized cooling systems, 00:04:16.819 --> 00:04:20.240 and those incredibly expensive GPUs. And this 00:04:20.240 --> 00:04:23.310 brings us to the fundamental inefficiency. The 00:04:23.310 --> 00:04:25.350 real tragedy here is that most of the tokens 00:04:25.350 --> 00:04:28.290 spent on complex AI projects. Especially autonomous 00:04:28.290 --> 00:04:31.370 agents running in loops. Yes, they are completely 00:04:31.370 --> 00:04:33.810 wasted. They are wasted because generative models 00:04:33.810 --> 00:04:36.389 are being called upon to evaluate simple forks 00:04:36.389 --> 00:04:38.980 in a decision tree. Break that down a bit. So 00:04:38.980 --> 00:04:41.879 when you build an AI agent, by its very nature, 00:04:41.899 --> 00:04:44.420 it runs in a loop. It looks at a task and it 00:04:44.420 --> 00:04:46.860 asks itself a question like, should I do X or 00:04:46.860 --> 00:04:50.079 should I do Y? Okay. To answer that simple routing 00:04:50.079 --> 00:04:52.639 question, under the current paradigm, the agent 00:04:52.639 --> 00:04:55.560 passes the entire context to the massive generative 00:04:55.560 --> 00:04:58.660 model. The model has to process the query, generate 00:04:58.660 --> 00:05:01.420 a rationale word by word, and literally output 00:05:01.420 --> 00:05:05.379 text to say, you should do X. None of which actually 00:05:05.379 --> 00:05:08.490 required language generation. Exactly. It's generating 00:05:08.490 --> 00:05:12.250 text just to pick a lane. But wait, let me step 00:05:12.250 --> 00:05:14.009 into the shoes of the user for a second here. 00:05:14.110 --> 00:05:16.250 Go for

it. If you're listening to this and you're 00:05:16.250 --> 00:05:20.110 paying your $30 a month for Grok, you might be 00:05:20.110 --> 00:05:23.269 thinking, isn't the whole point of these massive 00:05:23.269 --> 00:05:26.350 models that they're omnipotent? Right. The universal 00:05:26.350 --> 00:05:31.100 tool. Yeah. Like if I am paying my fee. Why shouldn't 00:05:31.100 --> 00:05:33.720 I just force the big model to handle every tiny 00:05:33.720 --> 00:05:36.759 mundane decision? Why do I care if it's wasteful 00:05:36.759 --> 00:05:38.899 as long as it gets the job done? And that is 00:05:38.899 --> 00:05:41.220 a very natural reaction. I mean, we've been trained 00:05:41.220 --> 00:05:43.420 for a decade by consumer cloud software to just 00:05:43.420 --> 00:05:45.500 click a button and assume the invisible servers 00:05:45.500 --> 00:05:47.540 will handle it seamlessly. Out of sight, out 00:05:47.540 --> 00:05:50.660 of mind. Exactly. But that mindset ignores the 00:05:50.660 --> 00:05:53.240 physical limits of computation. You should care 00:05:53.240 --> 00:05:55.920 because bottlenecking a system doesn't just hurt 00:05:55.920 --> 00:05:58.579 the provider's server costs. It directly and 00:05:58.579 --> 00:06:01.699 negatively impacts you. Okay, how so? Like walk 00:06:01.699 --> 00:06:04.420 me through the practical impact on my daily workflow. 00:06:04.680 --> 00:06:06.959 It basically comes down to two things, latency 00:06:06.959 --> 00:06:10.560 and limits. Okay. Every time Grok has to spin 00:06:10.560 --> 00:06:13.519 up its massive generative engine just to decide, 00:06:13.560 --> 00:06:16.480 say, if an incoming email is a refund request 00:06:16.480 --> 00:06:19.139 or a tech support issue, it takes time. Right. 00:06:19.220 --> 00:06:21.980 We're talking hundreds of milliseconds, sometimes 00:06:21.980 --> 00:06:25.970 multiple seconds per step. Now, imagine you have 00:06:25.970 --> 00:06:28.149 an autonomous agent running an overnight task, 00:06:28.350 --> 00:06:30.529 maybe sorting through thousands of customer files. 00:06:30.730 --> 00:06:33.050 So that involves thousands of micro decisions. 00:06:33.430 --> 00:06:36.029 Exactly. And that latency compounds heavily. 00:06:36.449 --> 00:06:38.990 Suddenly, your overnight run takes three days 00:06:38.990 --> 00:06:41.709 to finish. Oh, wow. Because it's pausing to write 00:06:41.709 --> 00:06:43.750 a little essay to itself about every single file. 00:06:44.009 --> 00:06:47.810 Yeah, an essay you never even see. And more importantly, 00:06:48.050 --> 00:06:51.350 you have a token limit. Right. I think $30 only 00:06:51.350 --> 00:06:54.269 goes so far. If you burn all your allotted tokens 00:06:54.269 --> 00:06:57.069 having the Michelin star chef toast bagels, you 00:06:57.069 --> 00:06:59.709 hit a wall. When you actually need the chef to 00:06:59.709 --> 00:07:01.930 invent a new recipe, like when you need deep, 00:07:01.949 --> 00:07:04.649 nuanced, creative problem solving, you have no 00:07:04.649 --> 00:07:07.069 tokens left. You get rate limited. Exactly. Your 00:07:07.069 --> 00:07:10.029 workflow stops dead in its tracks. So it's really 00:07:10.029 --> 00:07:12.250 about protecting your own capacity to do the 00:07:12.250 --> 00:07:14.769 heavy lifting later. Yes. By being smart about 00:07:14.769 --> 00:07:17.069 the small stuff, you keep the big brain available 00:07:17.069 --> 00:07:19.149 for the hard stuff. That makes perfect sense. 00:07:20.149 --> 00:07:22.449 We've established that forcing a massive generative 00:07:22.449 --> 00:07:25.569 AI to make simple decisions is economically flawed 00:07:25.569 --> 00:07:28.490 for the provider, and it creates a slow, terrible 00:07:28.490 --> 00:07:30.910 experience for the user. Which brings us to the 00:07:30.910 --> 00:07:33.550 solution. Right. We need to transition away from 00:07:33.550 --> 00:07:37.129 a paradigm of generate to decide. And the insight 00:07:37.129 --> 00:07:39.649 we are exploring today introduces us to a highly 00:07:39.649 --> 00:07:42.709 practical way to do exactly that. Using a local 00:07:42.709 --> 00:07:44.750 solution that completely removes that expensive 00:07:44.750 --> 00:07:48.209 call to the cloud. Yes. And Herlea. This is an 00:07:48.209 --> 00:07:50.490 incredibly powerful tool made by Convint Innovations. 00:07:50.610 --> 00:07:52.910 It's an open source model released under an Apache 00:07:52.910 --> 00:07:56.550 2 .0 license, and it is relatively tiny. Tiny 00:07:56.550 --> 00:07:59.949 is right. It has just 421 million parameters. 00:08:00.410 --> 00:08:02.769 Just to pause on that for clarity, the Apache 00:08:02.769 --> 00:08:05.980 2 .0 license is crucial here. Oh, totally. When 00:08:05.980 --> 00:08:08.139 we say open source, we don't just mean you can 00:08:08.139 --> 00:08:11.240 look at the code. This specific license is highly 00:08:11.240 --> 00:08:13.779 permissive. It means developers, businesses, 00:08:14.000 --> 00:08:17.180 anyone can use it, modify it, and deploy it commercially. 00:08:17.480 --> 00:08:19.540 Without paying royalties or dealing with legal 00:08:19.540 --> 00:08:22.519 headaches. Exactly. It's essentially a gift to 00:08:22.519 --> 00:08:25.920 the engineering community. And the size, 421 00:08:25.920 --> 00:08:28.899 million parameters. Yeah. Like, to put that in 00:08:28.899 --> 00:08:32.100 perspective. The frontier generative models we've 00:08:32.100 --> 00:08:34.019 been talking about measure their parameters in 00:08:34.019 --> 00:08:36.639 the hundreds of billions. Some even in the trillions 00:08:36.639 --> 00:08:40.399 now. Right. So 421 million is an absolute featherweight. 00:08:40.460 --> 00:08:43.480 It is highly focused, highly specialized. And 00:08:43.480 --> 00:08:45.720 it's actually the practical local version of 00:08:45.720 --> 00:08:47.639 a concept that was first released by a group 00:08:47.639 --> 00:08:51.039 called TypeSafe back in

September 2026. Right. 00:08:51.100 --> 00:08:54.360 Under the name Jev. Yeah. Jev. That was the release 00:08:54.360 --> 00:08:55.899 that really got the engineering world talking 00:08:55.899 --> 00:08:59.110 about this concept of a clerk AI. And clerk is 00:08:59.110 --> 00:09:01.990 the perfect word for it. Laya acts as a local 00:09:01.990 --> 00:09:04.789 decision layer. And the most crucial aspect of 00:09:04.789 --> 00:09:07.789 this entire architecture is the cost. Which is 00:09:07.789 --> 00:09:10.730 zero. It costs absolutely nothing per decision. 00:09:10.870 --> 00:09:13.289 It runs locally on the physical hardware that 00:09:13.289 --> 00:09:16.110 Grokbot already owns in the cloud or, you know, 00:09:16.110 --> 00:09:18.289 right on your own personal laptop. Brian McNeil's 00:09:18.289 --> 00:09:20.590 insight uses three brilliant analogies to explain 00:09:20.590 --> 00:09:23.690 this. I love these analogies. They take something 00:09:23.690 --> 00:09:26.509 very abstract like model routing and make it 00:09:26.509 --> 00:09:29.080 instantly understandable. So let's walk through 00:09:29.080 --> 00:09:32.179 them. The first one is the writer versus the 00:09:32.179 --> 00:09:35.580 clerk. Right. Imagine you walk into a busy office. 00:09:35.720 --> 00:09:38.940 You hand a short, somewhat confusing note to 00:09:38.940 --> 00:09:41.940 a very careful, thoughtful writer. Okay. She 00:09:41.940 --> 00:09:45.659 reads the note, sits back, thinks for a while, 00:09:45.779 --> 00:09:48.299 and then she writes out a full, detailed paragraph 00:09:48.299 --> 00:09:50.600 explaining her thoughts on the note, ending with 00:09:50.600 --> 00:09:53.100 a final recommendation on what to do. Which is 00:09:53.100 --> 00:09:55.679 fantastic if you need deep, nuanced analysis 00:09:55.679 --> 00:09:59.210 of a complex problem. Sure. But it is incredibly 00:09:59.210 --> 00:10:02.730 slow and highly costly if all you actually needed 00:10:02.730 --> 00:10:05.990 was a single checkmark on a form. Right. So now 00:10:05.990 --> 00:10:08.409 imagine you hand that exact same note to a clerk. 00:10:08.710 --> 00:10:11.889 The clerk never writes paragraphs. It's not in 00:10:11.889 --> 00:10:13.549 his job description. He doesn't have the time. 00:10:13.730 --> 00:10:16.370 Exactly. He takes the note, looks at a form on 00:10:16.370 --> 00:10:18.809 his desk that already has specific boxes printed 00:10:18.809 --> 00:10:21.250 on it, checks the appropriate box, and hands 00:10:21.250 --> 00:10:23.269 the form back to you. And those checkmarks are 00:10:23.269 --> 00:10:25.450 essentially just numbers that the next step in 00:10:25.450 --> 00:10:28.259 your computer program can read directly. No sentences 00:10:28.259 --> 00:10:30.940 are produced. Nothing has to be parsed by a massive 00:10:30.940 --> 00:10:33.340 language model. And nothing extra is built. So 00:10:33.340 --> 00:10:35.899 the massive generative model is the careful writer, 00:10:36.039 --> 00:10:39.460 and Leia is the clerk. Precisely. The second 00:10:39.460 --> 00:10:41.519 analogy is the one I used at the very beginning 00:10:41.519 --> 00:10:44.320 of this deep dive. The chef versus the line cook. 00:10:44.399 --> 00:10:47.120 A classic. For years, software only had the chef 00:10:47.120 --> 00:10:49.460 available. The chef had to invent the daily special, 00:10:49.700 --> 00:10:51.759 but because they were the only one in the kitchen, 00:10:51.879 --> 00:10:54.480 they also had to read every single basic crumpled 00:10:54.480 --> 00:10:56.460 order ticket that came through the window. Which 00:10:56.460 --> 00:10:59.320 is such a waste of their talent. Exactly. Leia 00:10:59.320 --> 00:11:02.440 is the fast, highly consistent line cook. The 00:11:02.440 --> 00:11:04.100 line cook just looks at the ticket and yells, 00:11:04.200 --> 00:11:07.909 medium rare, sauce on the side. Both jobs matter 00:11:07.909 --> 00:11:10.649 immensely to a functioning restaurant, but only 00:11:10.649 --> 00:11:14.370 one needs the full creative pass. Yeah. And the 00:11:14.370 --> 00:11:17.490 third analogy is writing a letter versus sorting 00:11:17.490 --> 00:11:20.549 the mail. Generative AI invents the contents 00:11:20.549 --> 00:11:23.230 of the letter. Leia simply looks at the zip code 00:11:23.230 --> 00:11:25.389 and drops the envelope into pre -labeled bins. 00:11:25.899 --> 00:11:28.379 Bins representing the distinct options you define 00:11:28.379 --> 00:11:31.120 in your code. I actually want to add my own spin 00:11:31.120 --> 00:11:33.039 on this just to really drive the mechanics home 00:11:33.039 --> 00:11:34.820 for everyone listening. Oh, let's hear it. It's 00:11:34.820 --> 00:11:37.299 like using a massive supercomputer in a secret 00:11:37.299 --> 00:11:40.039 underground bunker to calculate a 15 % tip at 00:11:40.039 --> 00:11:42.960 a local diner. Okay, yeah. Sure, the supercomputer 00:11:42.960 --> 00:11:45.220 can absolutely do it. It will give you the right 00:11:45.220 --> 00:11:48.840 answer. But a $3 plastic pocket calculator does 00:11:48.840 --> 00:11:50.960 it instantly for free using a watch battery. 00:11:51.179 --> 00:11:54.379 I like that. Leia is the AI equivalent of that 00:11:54.379 --> 00:11:57.139 pocket calculator. It's purpose -built for a 00:11:57.139 --> 00:12:00.179 specific bounded kind of operation. And the underlying 00:12:00.179 --> 00:12:02.559 mechanics of how Leia achieves the speed are 00:12:02.559 --> 00:12:05.860 really fascinating. In machine learning terms, 00:12:06.080 --> 00:12:09.279 Leia executes what we call a single forward pass. 00:12:09.460 --> 00:12:11.500 Okay, let's unpack that. What does that mean 00:12:11.500 --> 00:12:14.879 compared to what Grok is normally doing? So generative 00:12:14.879 --> 00:12:18.159 models are autoregressive. When you ask Grok 00:12:18.159 --> 00:12:20.580 a question, it doesn't just... output the whole 00:12:20.580 -->

00:12:22.539 answer at once. Right. It kind of types it out. 00:12:22.639 --> 00:12:25.139 It predicts the first word. Then it feeds that 00:12:25.139 --> 00:12:27.500 word back into its massive network to predict 00:12:27.500 --> 00:12:29.700 the second word. And it loops like that word 00:12:29.700 --> 00:12:32.779 by word until it's finished. Wow. So it's looping 00:12:32.779 --> 00:12:35.000 through billions of parameters every single word. 00:12:35.100 --> 00:12:37.860 Over and over again. But Leia doesn't do that. 00:12:37.919 --> 00:12:40.500 It takes your input, passes it through its small 00:12:40.500 --> 00:12:43.139 network exactly once. That's the single forward 00:12:43.139 --> 00:12:45.759 pass. And it returns a label and a mathematical 00:12:45.759 --> 00:12:48.360 probability of how confident it is. And that's 00:12:48.360 --> 00:12:50.899 it. That's it. It does this in tens of milliseconds. 00:12:51.279 --> 00:12:54.559 No API key, no per -call charge draining your 00:12:54.559 --> 00:12:57.340 wallet, and crucially, no data leaving the computer. 00:12:57.600 --> 00:13:00.419 That last point cannot be overstated. When we 00:13:00.419 --> 00:13:02.600 start looking at the macro implications of privacy 00:13:02.600 --> 00:13:05.620 and security, keeping data on the local machine 00:13:05.620 --> 00:13:07.779 during these high -volume classification steps 00:13:07.779 --> 00:13:10.860 is a massive advantage. Oh, absolutely. We will 00:13:10.860 --> 00:13:12.899 dive deeper into the security implications shortly, 00:13:13.019 --> 00:13:15.789 for sure. It's game -changing. Okay, so we have 00:13:15.789 --> 00:13:19.389 this 421 million parameter clerk sitting on our 00:13:19.389 --> 00:13:22.490 machine. But how do we actually talk to the clerk? 00:13:22.690 --> 00:13:25.110 Ah, yes. Because I imagine it requires a very 00:13:25.110 --> 00:13:27.470 different approach than prompting a generative 00:13:27.470 --> 00:13:30.730 bot. Like, you can't just type, hey, Leah, take 00:13:30.730 --> 00:13:32.110 a look at this email and let me know what you 00:13:32.110 --> 00:13:34.129 think I should do. No, you absolutely cannot. 00:13:34.289 --> 00:13:37.149 That is a guaranteed recipe for failure. Using 00:13:37.149 --> 00:13:39.710 a local decision layer requires us to completely 00:13:39.710 --> 00:13:42.070 change our mental model of prompting. Right. 00:13:42.110 --> 00:13:44.149 We have to learn the art of the bounded prompt. 00:13:44.580 --> 00:13:46.940 Leia is designed to answer exactly three types 00:13:46.940 --> 00:13:49.500 of questions. It cannot invent outside of these 00:13:49.500 --> 00:13:51.519 boundaries. Let's go through those. The three 00:13:51.519 --> 00:13:54.080 flavors of Leia's logic. What is flavor number 00:13:54.080 --> 00:13:57.179 one? The first is choice. This is where you ask 00:13:57.179 --> 00:14:00.039 Leia to pick exactly one label from a predefined 00:14:00.039 --> 00:14:02.679 list that you provide. You give it the bins and 00:14:02.679 --> 00:14:05.539 it picks the best fit. Simple enough. Number 00:14:05.539 --> 00:14:08.039 two is score. You ask Leia to place something 00:14:08.039 --> 00:14:10.740 on a scale that you describe. Right. For example, 00:14:10.820 --> 00:14:13.039 you might ask it to read the urgency of an email 00:14:13.039 --> 00:14:15.799 from zero to two. Where zero is spam and two 00:14:15.799 --> 00:14:18.100 is an angry customer threatening to cancel. And 00:14:18.100 --> 00:14:20.720 the third is what the documentation calls a now 00:14:20.720 --> 00:14:24.299 all question. N -O -U -L. Which is essentially 00:14:24.299 --> 00:14:27.139 a yes or no binary, right? Exactly. You ask it 00:14:27.139 --> 00:14:29.279 to return the probability that a specific statement 00:14:29.279 --> 00:14:32.340 is true. Is this email asking for a refund? Yes 00:14:32.340 --> 00:14:36.399 or no? Choice, score, and now all. Now, the real 00:14:36.399 --> 00:14:39.059 secret to making layout work and work well right 00:14:39.059 --> 00:14:40.820 out of the box without you having to spend weeks 00:14:40.820 --> 00:14:43.419 fine tuning it on your own private data. Which 00:14:43.419 --> 00:14:45.299 the industry calls zero -shot performance, by 00:14:45.299 --> 00:14:47.840 the way. Right, zero -shot. The secret to that 00:14:47.840 --> 00:14:50.360 is how you shape the options you give it. This 00:14:50.360 --> 00:14:52.659 is the most critical skill for anyone implementing 00:14:52.659 --> 00:14:55.980 this. You cannot just give Laya bare one -word 00:14:55.980 --> 00:14:58.360 labels. What do you mean? Well, let's say you 00:14:58.360 --> 00:15:01.539 are building an agent for inbox triage. You can't 00:15:01.539 --> 00:15:03.940 just give the clerk the options, billing, technical, 00:15:04.120 --> 00:15:06.779 other. Why not? That seems simple enough. Because 00:15:06.779 --> 00:15:09.799 a 421 million parameter model simply doesn't 00:15:09.799 --> 00:15:12.179 have the vast generalized world knowledge to 00:15:12.179 --> 00:15:14.700 reliably guess what you specifically mean by 00:15:14.700 --> 00:15:17.899 those bare words in every possible context. It 00:15:17.899 --> 00:15:19.740 doesn't have the life experience, so to speak. 00:15:19.899 --> 00:15:24.039 Right. It needs semantic anchors. So you have 00:15:24.039 --> 00:15:27.059 to provide plain language criteria. You have 00:15:27.059 --> 00:15:29.259 to tell the clerk exactly what the boxes on the 00:15:29.259 --> 00:15:32.639 form mean. Yes. So instead of just writing billing 00:15:32.639 --> 00:15:36.429 in your code, you need to write billing. Invoices, 00:15:36.570 --> 00:15:39.190 missing payments, refund requests, credit card 00:15:39.190 --> 00:15:41.509 updates. You spell it out. And instead of just 00:15:41.509 --> 00:15:44.450 technical, you write technical. Bugs, server 00:15:44.450 --> 00:15:47.789 outages, login errors, password resets. Exactly. 00:15:47.809 -->

00:15:50.009 You are giving the models internal and quota 00:15:50.009 --> 00:15:52.490 something rich and concrete to match the incoming 00:15:52.490 --> 00:15:55.370 text against. And there is a hard physical constraint 00:15:55.370 --> 00:15:57.370 here, which the insight refers to as the rule 00:15:57.370 --> 00:16:00.470 of 10. Yes, the rule of 10. Your option lists, 00:16:00.590 --> 00:16:02.929 like the number of bins you give the clerk, must 00:16:02.929 --> 00:16:06.399 be kept under 10 items. I want to dig into that 00:16:06.399 --> 00:16:08.519 because it sounds kind of arbitrary. Why 10? 00:16:08.779 --> 00:16:11.000 It's a good question. Like what happens at item 00:16:11.000 --> 00:16:13.059 11? Does the clerk just get frustrated and quit? 00:16:13.299 --> 00:16:16.039 It's not arbitrary at all. It's actually a mathematical 00:16:16.039 --> 00:16:18.820 reality of how small models map information. 00:16:19.179 --> 00:16:21.740 Okay, how so? Think of the model's brain as a 00:16:21.740 --> 00:16:25.299 giant 3D room, a vector space where concepts 00:16:25.299 --> 00:16:28.720 live. In a massive model like Grok, that room 00:16:28.720 --> 00:16:31.139 is the size of a galaxy. Huge. It can easily 00:16:31.139 --> 00:16:33.559 keep a thousand different concepts far apart 00:16:33.559 --> 00:16:36.460 from each other. But Leia's room is small. Right. 00:16:36.519 --> 00:16:38.740 If you put three or four distinct categories 00:16:38.740 --> 00:16:40.860 into that small room, it can easily tell them 00:16:40.860 --> 00:16:44.200 apart. But if you shove 20 or 30 categories in 00:16:44.200 --> 00:16:46.240 there, the boundaries between the concepts start 00:16:46.240 --> 00:16:50.159 to overlap. The semantic distance shrinks. So 00:16:50.159 --> 00:16:53.600 if you have billing, invoices, payments, and 00:16:53.600 --> 00:16:56.679 accounting, all as separate bins, the small model 00:16:56.679 --> 00:16:58.740 gets confused because they're all huddled in 00:16:58.740 --> 00:17:01.279 the exact same corner of its tiny room. Exactly. 00:17:01.580 --> 00:17:04.109 Its cognitive load is exceeded. The statistical 00:17:04.109 --> 00:17:06.549 confidence of the model drops significantly and 00:17:06.549 --> 00:17:08.710 it starts making mistakes. It puts the mail in 00:17:08.710 --> 00:17:10.710 the wrong bin. Yeah. And when you have more than 00:17:10.710 --> 00:17:12.869 10 options, the system is fundamentally designed 00:17:12.869 --> 00:17:16.349 to fail over. It routes the task back to the 00:17:16.349 --> 00:17:18.970 expensive generative model. Which completely 00:17:18.970 --> 00:17:20.630 defeats the purpose of having the clerk in the 00:17:20.630 --> 00:17:23.849 first place. Completely. If you give it 70 labels, 00:17:24.009 --> 00:17:25.490 it's just going to shrug, throw its hands up 00:17:25.490 --> 00:17:27.349 and send the ticket to the Michelin star chef 00:17:27.349 --> 00:17:30.450 anyway. Precisely. And it's not just the number 00:17:30.450 --> 00:17:32.769 of options that you have to control. It's the 00:17:32.769 --> 00:17:35.210 size of the input you give it. Right, because 00:17:35.210 --> 00:17:38.289 of the context window. Yes. Lalia requires short 00:17:38.289 --> 00:17:41.809 inputs. A subject line and a short email body 00:17:41.809 --> 00:17:45.089 work perfectly. A sprawling thread of 20 back 00:17:45.089 --> 00:17:47.009 and forth customer service messages will just 00:17:47.009 --> 00:17:49.809 overwhelm its active memory. It physically cannot 00:17:49.809 --> 00:17:52.650 hold that much text at once. So you have to design 00:17:52.650 --> 00:17:54.410 your workflow to extract just the most recent 00:17:54.410 --> 00:17:56.750 human message or maybe your short summary before 00:17:56.750 --> 00:18:00.029 you hand it to the clerk. Yep. Okay, here's where 00:18:00.029 --> 00:18:01.809 it gets really interesting for me, but also a 00:18:01.809 --> 00:18:03.809 little bit concerning as someone who built these 00:18:03.809 --> 00:18:07.049 tools. Uh -oh, what's the concern? If Lalia is 00:18:07.049 --> 00:18:10.549 so small and it has this limited context window, 00:18:10.769 --> 00:18:14.349 how do we actually trust it? What happens when 00:18:14.349 --> 00:18:17.490 a user's input is incredibly vague? Yeah, that 00:18:17.490 --> 00:18:21.349 happens a lot. Or worse, sarcastic. Like someone 00:18:21.349 --> 00:18:23.690 emails tech support and says, great job, guys. 00:18:23.789 --> 00:18:26.900 You billed me twice and deleted my account. Absolutely 00:18:26.900 --> 00:18:29.660 thrilled. Oh, a classic angry customer. Does 00:18:29.660 --> 00:18:31.640 this tiny clerk just look at the word thrilled 00:18:31.640 --> 00:18:34.200 and confidently toss it into the customer praise 00:18:34.200 --> 00:18:37.599 bin? And that is the exact vulnerability of small 00:18:37.599 --> 00:18:41.039 models. They lack the deep, worldly nuance to 00:18:41.039 --> 00:18:43.940 detect dry sarcasm. So it would miscategorize 00:18:43.940 --> 00:18:46.279 it. But this is where the architecture of the 00:18:46.279 --> 00:18:49.180 system saves us. Leia doesn't just return a label. 00:18:49.279 --> 00:18:52.039 It returns a label and a probability score. Right. 00:18:52.099 --> 00:18:55.460 The math score. It tells you mathematically how 00:18:55.460 --> 00:18:58.359 confident it is in its choice. OK, so in that 00:18:58.359 --> 00:19:01.420 sarcastic scenario, it might say, I think this 00:19:01.420 --> 00:19:03.400 is customer praise, but my confidence is only 00:19:03.400 --> 00:19:07.380 0 .3 out of 1 .0. Exactly. And when you are writing 00:19:07.380 --> 00:19:10.140 the code to run this, you set a threshold. You 00:19:10.140 --> 00:19:12.640 tell your system, if Leah's confidence is below 00:19:12.640 --> 00:19:16.089 0 .7, do not trust the clerk. If the confidence 00:19:16.089 --> 00:19:18.309 drops below that line, it acts as a tripwire. 00:19:18.349 --> 00:19:20.670 It

triggers a failsafe. So the system immediately 00:19:20.670 --> 00:19:23.410 routes that confusing, sarcastic ticket back 00:19:23.410 --> 00:19:25.869 up to the chef. The generative model takes over, 00:19:26.009 --> 00:19:27.990 reads the sarcasm, and handles it appropriately. 00:19:28.289 --> 00:19:31.210 That is just brilliant. It handles the 80 % of 00:19:31.210 --> 00:19:33.869 mundane, clear -cut stuff for free, and it only 00:19:33.869 --> 00:19:36.250 escalates the weird, tricky 20 % to the expensive 00:19:36.250 --> 00:19:38.630 brain. It's the perfect synthesis of capabilities, 00:19:39.009 --> 00:19:41.710 really. So we've discussed the theory. We know 00:19:41.710 --> 00:19:43.849 the three types of questions and we understand 00:19:43.849 --> 00:19:47.190 the vector space limitations. Now, how does a 00:19:47.190 --> 00:19:49.549 user actually get this running? This is the fun 00:19:49.549 --> 00:19:52.470 part. Because let's be honest, when you start 00:19:52.470 --> 00:19:55.009 talking about local AI models, Python dependencies 00:19:55.009 --> 00:19:57.990 and virtual environments, a lot of people's eyes 00:19:57.990 --> 00:20:00.450 start to glaze over. It can be highly intimidating 00:20:00.450 --> 00:20:03.289 for sure. But the deployment method here highlights 00:20:03.289 --> 00:20:06.750 the absolute magic of modern AI. The bot can 00:20:06.750 --> 00:20:09.619 actually install its own upgrade. This is what 00:20:09.619 --> 00:20:12.420 it truly means to be living in the future. Let's 00:20:12.420 --> 00:20:14.259 walk through this step by step because the source 00:20:14.259 --> 00:20:17.119 insight provides the literal playbook for hands 00:20:17.119 --> 00:20:20.420 -free deployment. Right. First, we need to understand 00:20:20.420 --> 00:20:22.880 the environment. When you are using Grokbot, 00:20:23.059 --> 00:20:25.799 you aren't just typing into a dumb chat window 00:20:25.799 --> 00:20:28.779 on a website. No, not at all. You are interacting 00:20:28.779 --> 00:20:31.480 with an agent that actually has access to a terminal, 00:20:31.599 --> 00:20:34.440 a file system, and a Python interpreter on its 00:20:34.440 --> 00:20:38.130 own cloud machine. Right. The computer that Grokbot 00:20:38.130 --> 00:20:40.769 is running on stays on after you close your browser 00:20:40.769 --> 00:20:43.349 tab. You don't need to open a remote desktop 00:20:43.349 --> 00:20:46.309 connection or type complex command line instructions 00:20:46.309 --> 00:20:49.670 yourself. You literally just paste a highly specific 00:20:49.670 --> 00:20:52.210 detailed prompt into the Grokbot chat window. 00:20:52.450 --> 00:20:54.730 And the insight provides the exact prompt to 00:20:54.730 --> 00:20:57.750 use. Yeah, you basically tell Grok, inspect what 00:20:57.750 --> 00:21:00.450 already exists on your machine, keep my current 00:21:00.450 --> 00:21:03.950 bots running, create a Python virtual environment 00:21:03.950 --> 00:21:08.500 named leia -env, Activate it. Run pip install 00:21:08.500 --> 00:21:11.279 leia. And the automation goes even further than 00:21:11.279 --> 00:21:13.660 that. The prompt instructs Grok to write a test 00:21:13.660 --> 00:21:17.720 script, name it testleia .py, load the English 00:21:17.720 --> 00:21:20.779 checkpoint from Coven Innovations, and run a 00:21:20.779 --> 00:21:23.339 sample prediction on a dummy email about a duplicate 00:21:23.339 --> 00:21:26.039 charge. It builds the room, puts the clerk in 00:21:26.039 --> 00:21:28.339 the room, slides a test note under the door, 00:21:28.359 --> 00:21:31.079 and checks the answer. Yes. And then once that 00:21:31.079 --> 00:21:33.740 test succeeds, the prompt tells Grok to write 00:21:33.740 --> 00:21:36.559 a local skill that intercepts all future decisions, 00:21:36.819 --> 00:21:39.400 running them through Leia in what we call shadow 00:21:39.400 --> 00:21:42.579 mode. Shadow mode. We will definitely explore 00:21:42.579 --> 00:21:44.920 shadow mode deeply in a moment because it's critical 00:21:44.920 --> 00:21:47.299 for safety. Absolutely. But OK, I have to interject 00:21:47.299 --> 00:21:50.400 here because I know how notoriously finicky Python 00:21:50.400 --> 00:21:52.380 environments can be. Oh, they can be a nightmare. 00:21:52.599 --> 00:21:54.859 We've all been trapped in dependency hell where 00:21:54.859 --> 00:21:57.740 installing one package breaks three others. Is 00:21:57.740 --> 00:22:00.599 this hands -free installation really as seamless 00:22:00.599 --> 00:22:03.700 as it sounds? Or are we kind of glossing over 00:22:03.700 --> 00:22:07.039 the messy reality of debugging code? It's a very 00:22:07.039 --> 00:22:10.599 valid skepticism. Python environments can and 00:22:10.599 --> 00:22:13.440 do break. But that is precisely why the prompt 00:22:13.440 --> 00:22:16.000 explicitly commands Grok to create a virtual 00:22:16.000 --> 00:22:18.920 environment first. Isolating it? Exactly. By 00:22:18.920 --> 00:22:22.299 isolating Leia inside a folder called leia -env, 00:22:22.339 --> 00:22:25.380 it creates a walled garden. It prevents the installation 00:22:25.380 --> 00:22:27.779 of Leia's dependencies from conflicting with 00:22:27.779 --> 00:22:30.660 any of Grok's existing delicate internal tools. 00:22:30.900 --> 00:22:33.619 You are giving Grok a safe sandbox to play in. 00:22:33.880 --> 00:22:36.640 Because the task is highly bounded, installing 00:22:36.640 --> 00:22:39.380 a specific package and running a basic test script, 00:22:39.819 --> 00:22:42.200 Frontier models are actually incredibly reliable 00:22:42.200 --> 00:22:44.460 at executing this successfully today. That makes 00:22:44.460 --> 00:22:47.309 a lot of sense. walled gardens save lives in 00:22:47.309 --> 00:22:50.329 software development they really do but what 00:22:50.329 --> 00:22:52.569 if i don't want to run this on grok's cloud computer 00:22:52.569 --> 00:22:54.829 what if i'm a developer and i want

this running 00:22:54.829 --> 00:22:58.230 locally on my own personal laptop or my company's 00:22:58.230 --> 00:23:00.170 internal server then it's just standard python 00:23:00.170 --> 00:23:02.630 commands you open your terminal you create the 00:23:02.630 --> 00:23:05.150 virtual environment and you run pip install leia 00:23:05.150 --> 00:23:07.650 and the first time you run it it obviously has 00:23:07.650 --> 00:23:10.029 to download the model itself right like leia 00:23:10.029 --> 00:23:12.430 has to physically live on your hard drive yep 00:23:12.430 --> 00:23:15.309 how big is this thing It downloads the English 00:23:15.309 --> 00:23:17.589 checkpoint from Hugging Face, which is, you know, 00:23:17.609 --> 00:23:20.509 the standard open source repository. The total 00:23:20.509 --> 00:23:24.450 download size is roughly 800 megabytes. 800 megabytes. 00:23:24.569 --> 00:23:27.809 That is astonishingly small. Barely takes up 00:23:27.809 --> 00:23:30.210 space. That's less than a single high -definition 00:23:30.210 --> 00:23:31.869 movie file. You can keep that on your laptop's 00:23:31.869 --> 00:23:33.269 hard drive without even noticing it's there. 00:23:33.490 --> 00:23:36.470 And that small footprint is the key to its immense 00:23:36.470 --> 00:23:38.950 speed and flexibility. If you are running it 00:23:38.950 --> 00:23:41.609 on a machine with an NVIDIA GPU, it automatically 00:23:41.609 --> 00:23:44.450 supports CUDA, which means it uses parallel processing 00:23:44.450 --> 00:23:47.450 to run even faster. Nice. If you are on a modern 00:23:47.450 --> 00:23:50.029 Mac, it leverages Apple Silicon's MPS framework. 00:23:50.309 --> 00:23:53.170 But here is the best part. What's that? Even 00:23:53.170 --> 00:23:55.150 if you are just running it on a standard CPU 00:23:55.150 --> 00:23:58.369 on an older laptop, the call still finishes in 00:23:58.369 --> 00:24:00.970 tens of milliseconds. It is so small that it 00:24:00.970 --> 00:24:03.109 doesn't require specialized hardware to be useful. 00:24:03.410 --> 00:24:05.630 And once that 800 megabyte file is downloaded? 00:24:06.190 --> 00:24:08.809 It runs completely offline. No internet connection 00:24:08.809 --> 00:24:11.049 required to make a decision. Which brings us 00:24:11.049 --> 00:24:13.609 back to what we touched on earlier. Yes. Let's 00:24:13.609 --> 00:24:15.529 circle back to that because the security implications 00:24:15.529 --> 00:24:18.970 here are massive. If we look at corporate data 00:24:18.970 --> 00:24:22.130 privacy, this architecture is a total game changer. 00:24:22.329 --> 00:24:24.829 Imagine a hospital network or a massive financial 00:24:24.829 --> 00:24:28.430 firm. They want to use AI to route highly sensitive 00:24:28.430 --> 00:24:31.130 internal data, patient medical records, or raw 00:24:31.130 --> 00:24:33.920 financial transactions. Under the old paradigm, 00:24:34.079 --> 00:24:36.000 they would have to send all that sensitive data 00:24:36.000 --> 00:24:38.799 out over the internet to a third -party API like 00:24:38.799 --> 00:24:42.079 Grok or OpenAI just to figure out which department 00:24:42.079 --> 00:24:44.539 should handle the file. Which is a massive compliance 00:24:44.539 --> 00:24:47.140 nightmare. You can't just send patient records 00:24:47.140 --> 00:24:50.460 to OpenAI. Exactly. With Laia, they can perform 00:24:50.460 --> 00:24:53.180 complex semantic classification and routing locally. 00:24:53.380 --> 00:24:55.660 The data never leaves the physical building. 00:24:55.859 --> 00:24:58.220 It never hits an external API during the decision 00:24:58.220 --> 00:25:01.400 -making phase. The privacy is absolute. So, we 00:25:01.400 --> 00:25:03.220 have the system installed. We have our super 00:25:03.220 --> 00:25:05.480 fast, highly secure clerk sitting at a desk, 00:25:05.579 --> 00:25:08.279 ready to work. Ready to sort. What are the exact 00:25:08.279 --> 00:25:10.960 practical use cases where this saves the most 00:25:10.960 --> 00:25:13.559 money and time? Let's move from the theoretical 00:25:13.559 --> 00:25:16.319 setup to the practical execution. This is where 00:25:16.319 --> 00:25:18.759 we get into what I consider a master class in 00:25:18.759 --> 00:25:21.359 modern software architecture. We are going to 00:25:21.359 --> 00:25:23.839 explore seven specific playbook patterns for 00:25:23.839 --> 00:25:26.299 local AI. Seven patterns, okay. These patterns 00:25:26.299 --> 00:25:28.940 detail the exact boundaries where a clerk thrives. 00:25:29.259 --> 00:25:31.480 And they contrast sharply with the open -ended 00:25:31.480 --> 00:25:34.119 research tasks where the clerk fails. Let's dive 00:25:34.119 --> 00:25:36.059 deep into these seven patterns because this is 00:25:36.059 --> 00:25:38.059 where the rubber meets the road for anyone actually 00:25:38.059 --> 00:25:41.019 building these systems. Pattern number one, inbox 00:25:41.019 --> 00:25:44.079 triage. This is a classic bottleneck. Imagine 00:25:44.079 --> 00:25:46.400 a mid -sized e -commerce company on a Monday 00:25:46.400 --> 00:25:49.380 morning. They have received 5 ,000 customer service 00:25:49.380 --> 00:25:51.519 emails over the weekend. That's a lot of reading. 00:25:51.740 --> 00:25:54.180 Under the old system, an AI agent reads every 00:25:54.180 --> 00:25:56.380 single email, thinks about it, and writes a response. 00:25:56.819 --> 00:26:00.349 Massive token burn. Huge. But with Leia, the 00:26:00.349 --> 00:26:02.589 bot reads each new message, but it only hands 00:26:02.589 --> 00:26:04.769 the subject line and the body to the local Leia 00:26:04.769 --> 00:26:07.869 skill. It asks Leia three highly specific questions. 00:26:08.210 --> 00:26:11.089 Right. It says, look at this email. Which category 00:26:11.089 --> 00:26:14.029 fits best from this short list? Billing, technical, 00:26:14.289 --> 00:26:17.369 account management, or other? Next, score the 00:26:17.369 -->

00:26:19.710 urgency of this email on a scale of zero to two. 00:26:19.849 --> 00:26:22.430 And finally, yes or no, is this person asking 00:26:22.430 --> 00:26:25.750 for a refund? The local model processes all 5 00:26:25.750 --> 00:26:29.000 ,000 emails in minutes. for free, marking those 00:26:29.000 --> 00:26:31.299 virtual boxes. And then the system looks at the 00:26:31.299 --> 00:26:34.019 results. If an email's urgency score comes back 00:26:34.019 --> 00:26:37.579 below 1 .2, and the category is anything except 00:26:37.579 --> 00:26:41.180 other, the bot knows it's a routine issue. So 00:26:41.180 --> 00:26:44.079 it bypasses the generative draft entirely. Exactly. 00:26:44.259 --> 00:26:46.819 It just places the message in a low -priority 00:26:46.819 --> 00:26:49.099 queue for a human, or it triggers a pre -written 00:26:49.099 --> 00:26:52.259 macro response. The expensive generative model 00:26:52.259 --> 00:26:54.599 is only ever called for the emails that cross 00:26:54.599 --> 00:26:57.619 the high urgency threshold. Or the ones labeled 00:26:57.619 --> 00:27:00.359 other that genuinely need a custom thoughtful 00:27:00.359 --> 00:27:04.019 reply. It completely filters out the noise. Yeah. 00:27:04.119 --> 00:27:07.079 Okay. Pattern number two is research gating. 00:27:07.299 --> 00:27:09.859 I absolutely love this one. Because anyone who's 00:27:09.859 --> 00:27:12.779 built AI agents knows they have a terrible habit 00:27:12.779 --> 00:27:15.279 of going down endless expensive rabbit holes. 00:27:15.619 --> 00:27:18.440 Oh, they absolutely do. Yeah. You might ask an 00:27:18.440 --> 00:27:20.940 agent to research a competitor's new product 00:27:20.940 --> 00:27:23.769 feature. Right. It will open a browser. Read 00:27:23.769 --> 00:27:26.690 a page, click a link, read another page, get 00:27:26.690 --> 00:27:29.170 distracted by a Wikipedia article about 19th 00:27:29.170 --> 00:27:32.069 century trade routes. And endlessly search the 00:27:32.069 --> 00:27:34.430 web. Yes. And every time it reads a new page, 00:27:34.549 --> 00:27:37.069 it is stuffing massive amounts of text into its 00:27:37.069 --> 00:27:39.450 context window, burning tokens at an alarming 00:27:39.450 --> 00:27:42.549 rate. So how does the clerk stop that? Before 00:27:42.549 --> 00:27:44.569 the agent is allowed to open a new web page, 00:27:44.789 --> 00:27:48.089 it has to ask Leah two questions. First, is there 00:27:48.089 --> 00:27:50.710 a cached summary from the last hour that already 00:27:50.710 --> 00:27:53.250 answers the core question? Okay. And second, 00:27:53.470 --> 00:27:56.369 based on the title of this next link, is opening 00:27:56.369 --> 00:27:58.450 it actually likely to change our final conclusion? 00:27:58.769 --> 00:28:01.390 So if the answer to both is no, the agent is 00:28:01.390 --> 00:28:03.670 forced to stop. It just uses the information 00:28:03.670 --> 00:28:06.529 it already has. Right. And if it decides it does 00:28:06.529 --> 00:28:09.410 need to search, you use Leia to enforce a hard 00:28:09.410 --> 00:28:12.670 limit, say three pages. Oh, I see. After three 00:28:12.670 --> 00:28:15.730 pages, Leia acts as a gatekeeper and says, Stop 00:28:15.730 --> 00:28:18.369 browsing and summarize what you have. It prevents 00:28:18.369 --> 00:28:20.369 the agent from infinitely looping and burning 00:28:20.369 --> 00:28:22.289 through your entire monthly token allowance while 00:28:22.289 --> 00:28:24.369 you sleep. It's literally like putting a leash 00:28:24.369 --> 00:28:27.109 on an overly enthusiastic puppy. It really is. 00:28:27.609 --> 00:28:31.390 Pattern number three is retry control. We all 00:28:31.390 --> 00:28:34.930 know that in software, APIs fail all the time. 00:28:35.549 --> 00:28:38.029 A server times out, a database connection drops 00:28:38.029 --> 00:28:40.710 momentarily. Happens every day. And AI agents, 00:28:40.829 --> 00:28:43.069 when they encounter an error, usually just keep 00:28:43.069 --> 00:28:44.940 hitting their heads against the wall. They try 00:28:44.940 --> 00:28:48.259 the exact same action over and over, culling 00:28:48.259 --> 00:28:50.400 the generative model each time to ask, did it 00:28:50.400 --> 00:28:53.200 work yet? Why did it fail? Which is incredibly 00:28:53.200 --> 00:28:56.500 inefficient. So instead, when an error occurs, 00:28:56.819 --> 00:28:59.819 you hand the raw error text and the current retry 00:28:59.819 --> 00:29:02.920 count to the local skill. Okay. You ask Leah, 00:29:03.099 --> 00:29:05.740 has this exact same error message appeared three 00:29:05.740 --> 00:29:08.779 times in a row? And is making another attempt 00:29:08.779 --> 00:29:11.900 actually worth the API cost, or is this a hard 00:29:11.900 --> 00:29:14.859 failure? If it's the exact same error three times, 00:29:15.160 --> 00:29:17.500 Leia returns a yes, and the system cuts power 00:29:17.500 --> 00:29:20.079 to the loop. It stops the agent from hallucinating 00:29:20.079 --> 00:29:22.039 that a fourth try will magically fix a broken 00:29:22.039 --> 00:29:25.279 server. Pattern four is transaction sorting. 00:29:25.579 --> 00:29:29.079 Imagine you work in finance or data analytics, 00:29:29.259 --> 00:29:33.039 and you have a massive spreadsheet with 50 ,000 00:29:33.039 --> 00:29:36.119 rows of raw transaction data. Under the old brute 00:29:36.119 --> 00:29:38.630 force paradigm, you would send... every single 00:29:38.630 --> 00:29:41.309 row to the generative AI, asking it to write 00:29:41.309 --> 00:29:43.990 a sentence classifying the transaction. The token 00:29:43.990 --> 00:29:45.910 burn would be astronomical, and it would take 00:29:45.910 --> 00:29:49.009 hours. With Lawya, you run the whole batch locally. 00:29:49.250 --> 00:29:52.109 The local skill returns just a label, like software 00:29:52.109 --> 00:29:54.549 subscription or travel expense, and a confidence 00:29:54.549 --> 00:29:57.349 score for each

individual row. No sentences generated 00:29:57.349 --> 00:30:00.529 per row. Just a label and a math score, flying 00:30:00.529 --> 00:30:02.670 through thousands of rows per second. Exactly. 00:30:03.210 --> 00:30:05.589 And your code simply says, anything that falls 00:30:05.589 --> 00:30:08.430 under a 0 .7 confidence score, flag it in red 00:30:08.430 --> 00:30:11.400 for human review. And everything else? The 90 00:30:11.400 --> 00:30:14.140 % of routine transactions, they are routed and 00:30:14.140 --> 00:30:17.000 categorized automatically, instantly, and for 00:30:17.000 --> 00:30:20.420 free. Amazing. Now, pattern five is crucial. 00:30:20.619 --> 00:30:23.480 It's called action gates. This is basically the 00:30:23.480 --> 00:30:26.839 safety mechanism. Very important. Before an autonomous 00:30:26.839 --> 00:30:29.880 agent is allowed to take a destructive or consequential 00:30:29.880 --> 00:30:33.640 action, moving money. deleting a database table, 00:30:33.799 --> 00:30:36.940 sending a mass email to 10 ,000 customers. It 00:30:36.940 --> 00:30:40.940 must pause. It asks Slaya, is this action reversible 00:30:40.940 --> 00:30:44.240 and does the impact of this action stay entirely 00:30:44.240 --> 00:30:47.500 inside our internal sandbox? It acts as a semantic 00:30:47.500 --> 00:30:50.960 circuit breaker. Yes. If the answer to either 00:30:50.960 --> 00:30:53.940 of those questions is no, if the action is irreversible 00:30:53.940 --> 00:30:56.599 or affects the outside world, the bot stops completely 00:30:56.599 --> 00:30:59.470 and demands human approval. You're using the 00:30:59.470 --> 00:31:02.349 cheap local model to ensure the expensive powerful 00:31:02.349 --> 00:31:04.589 model doesn't accidentally burn your company 00:31:04.589 --> 00:31:08.930 down. Exactly. Pattern six is support ticket 00:31:08.930 --> 00:31:11.970 batching. Let's say Amazon Web Services goes 00:31:11.970 --> 00:31:14.910 down and a company gets a flood of 500 support 00:31:14.910 --> 00:31:16.950 tickets in 10 minutes, all complaining about 00:31:16.950 --> 00:31:20.509 the exact same login error. Normally, the generative 00:31:20.509 --> 00:31:22.910 AI would look at the queue and try to thoughtfully 00:31:22.910 --> 00:31:26.609 read and write 500 unique apologies. It would 00:31:26.609 --> 00:31:29.029 take forever and cost a fortune. Right. Instead, 00:31:29.250 --> 00:31:31.289 the system sends the cluster of tickets to LIA 00:31:31.289 --> 00:31:34.250 and asks two questions. Does this ticket match 00:31:34.250 --> 00:31:37.390 a known ongoing incident pattern? And is the 00:31:37.390 --> 00:31:39.970 severity of this individual user's issue high 00:31:39.970 --> 00:31:42.390 enough to wake up an engineer? So if it matches 00:31:42.390 --> 00:31:46.049 the known AWS outage, LIA groups all 500 tickets 00:31:46.049 --> 00:31:48.930 together into a single batch. It allows a single 00:31:48.930 --> 00:31:51.369 human manager to look at the batch, click approve, 00:31:51.730 --> 00:31:54.269 and close all 500 tickets at once with a single 00:31:54.269 --> 00:31:58.069 templated response. And finally, pattern seven. 00:31:58.210 --> 00:32:00.369 Yeah. Overnight go or no go? This one is great. 00:32:01.049 --> 00:32:04.009 Imagine you have scheduled a bot to start a massive 00:32:04.009 --> 00:32:05.930 eight -hour data migration routine at midnight. 00:32:06.930 --> 00:32:09.369 Before it fires up that massive compute engine, 00:32:09.589 --> 00:32:13.529 it asks LIA to perform three quick checks. Is 00:32:13.529 --> 00:32:15.750 the overall goal still the same as it was yesterday? 00:32:16.269 --> 00:32:18.769 Have any new high -priority emergencies arrived 00:32:18.769 --> 00:32:20.670 in our system that should pause this migration? 00:32:21.349 --> 00:32:24.950 And is the expected runtime still under our budget 00:32:24.950 --> 00:32:28.230 limit? If all three answers are yes, the migration 00:32:28.230 --> 00:32:31.710 runs. If any answer is no, it leaves a log note 00:32:31.710 --> 00:32:34.569 and stops. It doesn't ask the massive generative 00:32:34.569 --> 00:32:37.829 model to reread a 50 -page project plan just 00:32:37.829 --> 00:32:40.690 to confirm that nothing has changed. If we look 00:32:40.690 --> 00:32:43.769 at these seven patterns holistically, they really 00:32:43.769 --> 00:32:46.009 represent a fundamental shift in how we build 00:32:46.009 --> 00:32:48.450 software. They do. We are moving from a world 00:32:48.450 --> 00:32:51.470 where we use deep intelligence to generate answers 00:32:51.470 --> 00:32:54.869 to every problem to a world where we use lightweight 00:32:54.869 --> 00:32:58.220 intelligence to classify and route data. Listening 00:32:58.220 --> 00:33:00.400 to this, it sounds exactly like setting up a 00:33:00.400 --> 00:33:02.960 smart home routine with physical tripwires. How 00:33:02.960 --> 00:33:05.160 so? You know, if the motion sensor on the porch 00:33:05.160 --> 00:33:08.160 detects movement, turn on the light. But instead 00:33:08.160 --> 00:33:11.200 of a dumb physical motion sensor, Lya is a semantic 00:33:11.200 --> 00:33:14.440 sensor. Ah, I love that. It is checking the actual 00:33:14.440 --> 00:33:17.119 meaning, the context of the data, before it decides 00:33:17.119 --> 00:33:19.980 to wake up the big, expensive, generative AI. 00:33:20.299 --> 00:33:22.559 That is a brilliant way to conceptualize it. 00:33:22.809 --> 00:33:25.930 It is a semantic tripwire. But this brings us 00:33:25.930 --> 00:33:29.410 to a very critical practical challenge. Which 00:33:29.410 --> 00:33:32.390 is? If we are handing off thousands of daily 00:33:32.390 --> 00:33:36.450 micro decisions to a 421 million parameter clerk, 00:33:36.609 --> 00:33:39.569 if this clerk is acting as our tripwire, how 00:33:39.569 --> 00:33:41.549 do we ensure the clerk isn't making mistakes? 00:33:42.109 --> 00:33:44.950 Right. How do we know it isn't quietly ruining

00:33:44.950 --> 00:33:47.670 a business by sorting legitimate refund requests 00:33:47.670 --> 00:33:50.579 into a spam folder? We have to monitor it. We 00:33:50.579 --> 00:33:52.799 have to supervise the clerk, and that brings 00:33:52.799 --> 00:33:55.859 us to the concept of shadow logging. It's a deeply 00:33:55.859 --> 00:33:59.539 rigorous way to trust but verify this local model. 00:33:59.779 --> 00:34:02.480 The process introduces the concept of shadow 00:34:02.480 --> 00:34:05.799 mode. When you first install Lya and set up these 00:34:05.799 --> 00:34:07.940 routing rules, say for your customer service 00:34:07.940 --> 00:34:11.039 inbox, you do not let Lya actually take any action. 00:34:11.280 --> 00:34:13.239 No, it's like putting the new clerk at a desk, 00:34:13.400 --> 00:34:16.519 having them fill out the sorting forms, but intercepting 00:34:16.519 --> 00:34:18.380 those forms before they actually go into the 00:34:18.380 --> 00:34:21.210 filing cabinet. Right. For a full day, or maybe 00:34:21.210 --> 00:34:23.829 a full week, the bot simply logs everything. 00:34:24.070 --> 00:34:26.130 It records the exact question it asked Leia, 00:34:26.269 --> 00:34:29.170 the answer Leia gave, the probability score Leia 00:34:29.170 --> 00:34:32.050 returned, the timestamp, and what action Leia 00:34:32.050 --> 00:34:34.010 would have taken if it were in charge. Meanwhile, 00:34:34.190 --> 00:34:36.969 the expensive generative model or a human team 00:34:36.969 --> 00:34:39.969 continues to actually do the work. Then, the 00:34:39.969 --> 00:34:42.179 next morning, you pull up the daily report. You 00:34:42.179 --> 00:34:44.340 can see exactly how many decisions Leia would 00:34:44.340 --> 00:34:46.880 have routed successfully, what the latency was, 00:34:47.059 --> 00:34:50.019 and most importantly... You compare Leia's decisions 00:34:50.019 --> 00:34:52.980 to what the human or the big AI actually did. 00:34:53.219 --> 00:34:56.380 Yes. You are looking for the match rate. Are 00:34:56.380 --> 00:34:59.579 they agreeing 95 % of the time? 99 %? And you 00:34:59.579 --> 00:35:01.679 don't just do this during setup. Once you turn 00:35:01.679 --> 00:35:04.780 Leia on for real, you establish a weekly honesty 00:35:04.780 --> 00:35:08.099 check. I find this concept so clever. You literally 00:35:08.099 --> 00:35:10.980 write a script that says... Show me the 10 Leia 00:35:10.980 --> 00:35:13.260 decisions from the past week that had the absolute 00:35:13.260 --> 00:35:15.380 lowest confidence scores. And tell me what the 00:35:15.380 --> 00:35:17.280 final outcome of those specific tickets was. 00:35:17.579 --> 00:35:20.380 This is how you debug the system without debugging 00:35:20.380 --> 00:35:23.539 complex Python code. Exactly. You look at those 00:35:23.539 --> 00:35:26.679 10 low confidence failures. Why did Leia struggle? 00:35:27.099 --> 00:35:29.860 Almost always, it's not a bug in the AI. It's 00:35:29.860 --> 00:35:32.579 because your form was confusing. Right. The criteria 00:35:32.579 --> 00:35:35.360 you wrote for the billing bin wasn't clear enough. 00:35:35.829 --> 00:35:37.969 Or there was a new type of customer complaint 00:35:37.969 --> 00:35:40.409 that didn't fit neatly into any of your predefined 00:35:40.409 --> 00:35:43.150 bins. So if you are looking at a shadow log and 00:35:43.150 --> 00:35:44.829 seeing low confidence on a bunch of tickets, 00:35:45.030 --> 00:35:47.309 you just rewrite the plain English criteria. 00:35:47.670 --> 00:35:51.289 It's that simple. But realistically, how do we 00:35:51.289 --> 00:35:53.550 avoid falling into a never -ending game of whack 00:35:53.550 --> 00:35:55.289 -a -mole with prompt engineering? What do you 00:35:55.289 --> 00:35:58.269 mean? Let's say I fix the wording for the billing 00:35:58.269 --> 00:36:01.630 bin. But in doing so, I accidentally blur the 00:36:01.630 --> 00:36:03.570 lines and break the routing for the technical 00:36:03.570 --> 00:36:06.219 bin. Do we just sit there tweaking plain English 00:36:06.219 --> 00:36:08.500 definitions all day long? Well, that is the inherent 00:36:08.500 --> 00:36:11.300 risk of zero -shot prompting. You are relying 00:36:11.300 --> 00:36:14.179 on the model's general knowledge. For most standard 00:36:14.179 --> 00:36:16.960 businesses, carefully rewording the criteria 00:36:16.960 --> 00:36:20.760 solves the problem entirely. But if you are operating 00:36:20.760 --> 00:36:24.159 in a highly specific domain, let's say specialized 00:36:24.159 --> 00:36:27.940 medical billing or aerospace engineering, where 00:36:27.940 --> 00:36:30.179 there is a massive amount of highly specific, 00:36:30.300 --> 00:36:33.420 bizarre vocabulary. Zero -shot prompting might 00:36:33.420 --> 00:36:36.480 hit a hard ceiling. Exactly. The clerk just doesn't 00:36:36.480 --> 00:36:38.960 know the jargon well enough to sort it accurately, 00:36:39.119 --> 00:36:41.260 no matter how well you write the form. And what 00:36:41.260 --> 00:36:43.099 do you do when you hit that ceiling? Do you just 00:36:43.099 --> 00:36:44.840 give up and go back to the generative model? 00:36:45.380 --> 00:36:48.360 No, you fine -tune the model. Because LEIA is 00:36:48.360 --> 00:36:50.519 open source and very small, you can actually 00:36:50.519 --> 00:36:53.840 train it on your specific company data. You can 00:36:53.840 --> 00:36:56.719 use a process called the RLCD loop, which stands 00:36:56.719 --> 00:36:59.739 for Reinforcement Learning from Contrastive Distillation. 00:36:59.920 --> 00:37:01.619 Okay, let's break that down for the audience. 00:37:02.099 --> 00:37:04.380 What does that actually mean in practice? Essentially, 00:37:04.440 --> 00:37:06.460 it's a way of tweaking the internal weights, 00:37:06.639 --> 00:37:10.360 the knobs and dials inside the 421 million parameters 00:37:10.360 --> 00:37:13.199 by showing the model examples of right and wrong 00:37:13.199 --> 00:37:15.760

answers. Oh, like flashcards. Yeah, kind of. 00:37:16.219 --> 00:37:18.760 You take a few hundred of your specific medical 00:37:18.760 --> 00:37:21.260 tickets that have already been perfectly sorted 00:37:21.260 --> 00:37:24.920 by humans. You feed them to Laya. When Laya gets 00:37:24.920 --> 00:37:27.219 it right, you mathematically reward it. When 00:37:27.219 --> 00:37:29.420 it gets it wrong, you penalize it. Makes sense. 00:37:29.920 --> 00:37:32.659 Because the model is small, you can do this training 00:37:32.659 --> 00:37:36.380 on a free T4 GPU using cloud platforms like Kaggle 00:37:36.380 --> 00:37:39.599 or Google Colab in a matter of hours. The model 00:37:39.599 --> 00:37:42.480 learns your exact vocabulary. But the beauty 00:37:42.480 --> 00:37:44.159 of the system is most people will never even 00:37:44.159 --> 00:37:46.699 need to do that. The out -of -the -box zero -shot 00:37:46.699 --> 00:37:48.500 capabilities with a well -written prompt are 00:37:48.500 --> 00:37:51.360 usually more than enough. And when the system 00:37:51.360 --> 00:37:54.260 is implemented correctly, the macro impact is... 00:37:54.539 --> 00:37:58.440 There is a published incident report that perfectly 00:37:58.440 --> 00:38:00.760 encapsulates this win -win scenario. Oh, the 00:38:00.760 --> 00:38:04.099 server outage one. Yes. A company was using an 00:38:04.099 --> 00:38:06.380 incident bot with this local decision layer. 00:38:06.679 --> 00:38:09.820 A massive server outage occurred. The bot cleared 00:38:09.820 --> 00:38:13.760 204 support tickets in about 10 minutes. That's 00:38:13.760 --> 00:38:17.099 wild. And it required only a single human approval 00:38:17.099 --> 00:38:19.440 step at the very end to authorize the batch closure. 00:38:20.300 --> 00:38:23.769 204 highly specific tickets in 10 minutes. The 00:38:23.769 --> 00:38:26.150 generative model was still utilized for the final 00:38:26.150 --> 00:38:29.630 writing of the apology and for the nuanced judgment 00:38:29.630 --> 00:38:32.389 call of assessing the overall outage. But LIA 00:38:32.389 --> 00:38:34.650 was placed in front of all the intermediate forks. 00:38:34.889 --> 00:38:37.969 So LIA checked if the incoming tickets matched 00:38:37.969 --> 00:38:41.070 the known outage pattern. It checked if the severity 00:38:41.070 --> 00:38:43.789 was high. It checked if the database retries 00:38:43.789 --> 00:38:46.530 had failed three times. It stripped all of that 00:38:46.530 --> 00:38:49.070 token cost away from the generative model. And 00:38:49.070 --> 00:38:51.650 this is where the economics for massive AI providers 00:38:51.650 --> 00:38:55.179 completely shift. If a company like SpaceX AI 00:38:55.179 --> 00:38:58.380 adopts this two -tiered architecture, if they 00:38:58.380 --> 00:39:01.500 place a local layer like Laya in front of every 00:39:01.500 --> 00:39:04.960 single Grokbot agent, their marginal cost for 00:39:04.960 --> 00:39:07.760 these millions of daily decision steps plummets 00:39:07.760 --> 00:39:10.599 to essentially zero. Their server load drops 00:39:10.599 --> 00:39:13.159 precipitously. Their massive data centers run 00:39:13.159 --> 00:39:16.280 cooler. The energy grids get a break. And they 00:39:16.280 --> 00:39:18.980 stop burning money on bagels. And yet... You, 00:39:19.059 --> 00:39:21.179 the user, perceive the product as being significantly 00:39:21.179 --> 00:39:24.099 faster, more responsive, and vastly more efficient. 00:39:24.420 --> 00:39:26.679 It is that rarest of things in the technology 00:39:26.679 --> 00:39:30.619 sector. An actual, tangible free lunch. The user 00:39:30.619 --> 00:39:33.400 saves their $30 a month token allowance for the 00:39:33.400 --> 00:39:36.019 actual creative, heavy lifting work they bought 00:39:36.019 --> 00:39:39.139 the tool for. And the provider saves millions 00:39:39.139 --> 00:39:42.760 in unsustainably high compute costs. It really 00:39:42.760 --> 00:39:45.250 is the only mathematical way forward. If we want 00:39:45.250 --> 00:39:47.610 to move into a future where billions of autonomous 00:39:47.610 --> 00:39:50.469 agents are running simultaneously, managing our 00:39:50.469 --> 00:39:52.989 schedules, our finances, our software, we cannot 00:39:52.989 --> 00:39:55.750 do it with the current brute force method. It 00:39:55.750 --> 00:39:57.550 would melt down our data centers. We absolutely 00:39:57.550 --> 00:39:59.889 would. We have to separate the creative generation 00:39:59.889 --> 00:40:02.710 from the mundane classification. By giving an 00:40:02.710 --> 00:40:05.750 agent a local clerk, we achieve incredible speed 00:40:05.750 --> 00:40:08.050 and complete data privacy for those high volume 00:40:08.050 --> 00:40:11.280 sorting tasks. And the deployment. thanks to 00:40:11.280 --> 00:40:13.360 the self -executing capabilities of modern bots, 00:40:13.579 --> 00:40:16.099 can be largely automated right inside your existing 00:40:16.099 --> 00:40:19.400 chat interface. Yeah. If you are building workflows, 00:40:19.639 --> 00:40:22.300 managing AI agents, or if you are just trying 00:40:22.300 --> 00:40:24.260 to stretch your software subscription fee as 00:40:24.260 --> 00:40:27.059 far as it will go, delegating bounded decisions 00:40:27.059 --> 00:40:30.699 to local small models is the ultimate hack. It 00:40:30.699 --> 00:40:33.059 is how you actually live in the future efficiently. 00:40:33.380 --> 00:40:35.519 But as we look at this architecture, it raises 00:40:35.519 --> 00:40:37.820 a rather provocative thought about the long -term 00:40:37.820 --> 00:40:40.380 future of the AI industry. Okay, lay it on me. 00:40:40.739 --> 00:40:43.360 We've just spent this time outlining how local 00:40:43.360 --> 00:40:45.659 offline machines will eventually get so good 00:40:45.659 --> 00:40:49.320 at handling 90 % of the cognitive sorting work 00:40:49.320 --> 00:40:52.380 silently, securely, and for free. That's the 00:40:52.380

--> 00:40:55.619 trajectory we are on, yes? If that happens, if 00:40:55.619 --> 00:40:58.420 the local clerk becomes that efficient and that 00:40:58.420 --> 00:41:01.300 ubiquitous, will we soon find that our massive 00:41:01.300 --> 00:41:03.539 billion -dollar cloud models are largely sitting 00:41:03.539 --> 00:41:06.340 idle? Oh, wow. Will they just be waiting days 00:41:06.340 --> 00:41:09.380 at a time for a problem actually worthy of their 00:41:09.380 --> 00:41:11.809 immense intellect? And if so, what happens to 00:41:11.809 --> 00:41:14.289 the entire subscription business model of the 00:41:14.289 --> 00:41:17.670 giant AI labs when users realize they only actually 00:41:17.670 --> 00:41:20.329 need to tap into the cloud once a week? Wow. 00:41:21.000 --> 00:41:24.579 That is a massive existential question for the 00:41:24.579 --> 00:41:27.119 industry to ponder. Are we building the ultimate 00:41:27.119 --> 00:41:30.219 tools just to put them out of a job? Time will 00:41:30.219 --> 00:41:32.239 tell. Well, thank you for joining us on this 00:41:32.239 --> 00:41:35.320 incredible exploration. As a reminder, this has 00:41:35.320 --> 00:41:38.650 been the ReadMultiplex .com. deep dive into a 00:41:38.650 --> 00:41:41.730 truly classic Brian Romilly insight. The architecture 00:41:41.730 --> 00:41:43.909 of the future isn't just about building bigger 00:41:43.909 --> 00:41:46.369 brains. It's about being profoundly smarter in 00:41:46.369 --> 00:41:49.610 how we delegate tasks. Absolutely. Please remember 00:41:49.610 --> 00:41:51.650 to support Brian Rollman and his groundbreaking 00:41:51.650 --> 00:41:54.909 work by becoming a member or buying him a coffee 00:41:54.909 --> 00:41:58.269 over at multiplex .tips. Because the next time 00:41:58.269 --> 00:42:00.869 you look at a glowing AI chat box, just remember, 00:42:01.050 --> 00:42:03.250 you really shouldn't be asking the Michelin star 00:42:03.250 --> 00:42:07.329 chef to toast your bagel. Until next time. Welcome 00:42:07.329 --> 00:42:11.250 to the multiplex, where the future comes alive. 00:42:11.929 --> 00:42:16.510 Past, present, future tech, open up your eyes. 00:42:17.260 --> 00:42:20.639 Brian Romella leads the way with research deep 00:42:20.639 --> 00:42:25.300 and true. Technology and history, he connects 00:42:25.300 --> 00:42:28.860 them all for you. Brian Romella leads the way 00:42:28.860 --> 00:42:31.019 with research deep and true. Technology and history, 00:42:31.260 --> 00:42:33.039 he connects them all for you. Ancient sparks 00:42:33.039 --> 00:42:36.340 to modern code, he traces every track of exploration. 00:42:37.230 --> 00:43:13.769 But humanity moves on, love equation for AI Zero 00:43:13.989 --> 00:43:20.070 human companies built from the start Thermodynamic 00:43:20.070 --> 00:43:24.789 wages earning the key Secrets we will read, experiment 00:43:24.789 --> 00:43:28.690 in publishing Real researches and spending every 00:43:28.690 --> 00:43:34.389 month Rate more, flex, go