WEBVTT 00:00:00.000 --> 00:00:03.379 So we spend billions of dollars, I mean, literally 00:00:03.379 --> 00:00:06.839 billions every single year, trying to engineer 00:00:06.839 --> 00:00:10.439 human error entirely out of our technology. Right, 00:00:10.439 --> 00:00:12.699 because the cultural expectation is that modern 00:00:12.699 --> 00:00:15.259 engineering should lead us to this state of just, 00:00:15.259 --> 00:00:18.219 you know, sterile total perfection. Exactly. 00:00:18.320 --> 00:00:21.059 Like we want self -driving cars that never swerve 00:00:21.059 --> 00:00:24.399 or deterministic algorithms that can predict 00:00:24.399 --> 00:00:27.359 the stock market with cold. binary precision. 00:00:27.539 --> 00:00:29.579 Oh, and aviation systems too, calculating the 00:00:29.579 --> 00:00:32.020 absolute safest routes without a single hiccup. 00:00:32.399 --> 00:00:33.920 It's kind of the ultimate engineering dream, 00:00:33.939 --> 00:00:36.520 right? Yeah, a technological ecosystem that operates 00:00:36.520 --> 00:00:39.340 so flawlessly, it just simply does not need us 00:00:39.340 --> 00:00:41.299 anymore. Because when you look at any complex 00:00:41.299 --> 00:00:43.619 system, the human being is almost always the 00:00:43.619 --> 00:00:46.079 messiest, most unpredictable variable in the 00:00:46.079 --> 00:00:48.859 whole equation. Which is true. But, and this 00:00:48.859 --> 00:00:51.439 is wild, to make our most advanced artificial 00:00:51.439 --> 00:00:54.060 intelligence or our multi -million dollar flight 00:00:54.060 --> 00:00:56.700 simulators and even our autonomous defense systems 00:00:56.700 --> 00:00:59.100 actually function in the real world, engineers 00:00:59.100 --> 00:01:01.420 are actively doing the exact opposite. Yeah, 00:01:01.579 --> 00:01:04.420 they are intentionally injecting messy, panicking, 00:01:04.500 --> 00:01:06.700 unpredictable humans right back into the code. 00:01:07.099 --> 00:01:09.790 Which brings us to today. Welcome to our deep 00:01:09.790 --> 00:01:12.670 dive. Today we're pulling from a pretty comprehensive 00:01:12.670 --> 00:01:15.109 breakdown of a concept known as human in the 00:01:15.109 --> 00:01:21.250 loop or HITL. HITL. And our mission for you listening 00:01:21.250 --> 00:01:24.849 is to explore this massive fundamental irony 00:01:24.849 --> 00:01:28.700 like in an era completely obsessed with creating 00:01:28.700 --> 00:01:31.060 perfect autonomous computers, we're going to 00:01:31.060 --> 00:01:34.719 look at why deliberately introducing human friction 00:01:34.719 --> 00:01:37.379 into these systems is actually the secret to 00:01:37.379 --> 00:01:39.439 making them work at all. And the scope of this 00:01:39.439 --> 00:01:41.299 is just staggering. I mean, the frameworks we're 00:01:41.299 --> 00:01:43.900 exploring today span from the baseline parameters 00:01:43.900 --> 00:01:46.359 of commercial driving simulators to the absolute 00:01:46.359 --> 00:01:48.459 cutting edge of machine learning architecture. 00:01:48.659 --> 00:01:50.319 All the way up to lethal autonomous weapons, 00:01:50.340 --> 00:01:52.500 right? Exactly. All the way to deployment protocols 00:01:52.500 --> 00:01:54.819 for lethal autonomous weapons. And across all 00:01:54.819 --> 00:01:57.480 these domains, the mathematical of the machine 00:01:57.480 --> 00:02:00.980 is just kind of useless without the subjective 00:02:00.980 --> 00:02:03.060 imperfection of a human. Okay let's unpack this 00:02:03.060 --> 00:02:06.700 because to a tech literate audience deliberately 00:02:06.700 --> 00:02:09.219 adding human error back into a system sounds 00:02:09.219 --> 00:02:11.379 completely counterintuitive. Oh completely. Like 00:02:11.379 --> 00:02:14.740 if you're building a deterministic system introducing 00:02:14.740 --> 00:02:18.259 an inherently stochastic unpredictable variable 00:02:18.259 --> 00:02:21.759 feels like I don't know, intentionally throwing 00:02:21.759 --> 00:02:24.180 a wrench into a perfectly tuned engine. It does. 00:02:24.219 --> 00:02:26.939 It feels like sabotage at first glance. But to 00:02:26.939 --> 00:02:29.099 understand why engineers actually do this, we 00:02:29.099 --> 00:02:31.419 really have to look at how humans are used as 00:02:31.419 --> 00:02:34.500 the ultimate stress test. OK. So if we start 00:02:34.500 --> 00:02:36.800 in the realm of physical environments and modeling, 00:02:37.240 --> 00:02:40.099 there is this stark, really necessary contrast 00:02:40.099 --> 00:02:43.080 between a human out of the loop simulation and 00:02:43.080 --> 00:02:45.599 a human in the loop simulation. It really all 00:02:45.599 --> 00:02:47.400 comes down to the illusion of predictability. 00:02:47.860 --> 00:02:50.439 So let's define those boundaries clearly. A human 00:02:50.439 --> 00:02:53.340 out of loop simulation is basically a closed 00:02:53.340 --> 00:02:56.400 deterministic model. You input a specific set 00:02:56.400 --> 00:02:58.539 of initial parameters, you execute the program, 00:02:58.900 --> 00:03:01.039 and you get a result. And if you reset the model 00:03:01.039 --> 00:03:03.599 and see that those exact same identical parameters, 00:03:03.840 --> 00:03:05.659 you're going to perfectly replicate the exact 00:03:05.659 --> 00:03:07.919 same results time and time again. The computer 00:03:07.919 --> 00:03:10.620 is basically just crunching the physics. Exactly, 00:03:11.020 --> 00:03:14.120 which is incredibly useful for, well, the early 00:03:14.120 --> 00:03:17.460 stages of project development. Tabletop simulations, 00:03:17.680 --> 00:03:19.719 you know, those closed -loop models, they allow 00:03:19.719 --> 00:03:22.520 engineers to set broad parameters, establish 00:03:22.520 --> 00:03:25.860 baseline data, and just confirm that the core 00:03:25.860 --> 00:03:28.460 physics of their design are sound. But they operate 00:03:28.460 --> 00:03:31.240 in a total vacuum. They do. Now, a human -in 00:03:31.240 --> 00:03:33.180 -the -loop simulation, on the other hand, is 00:03:33.180 --> 00:03:36.000 an interactive physical simulation. It requires 00:03:36.000 --> 00:03:39.120 a human operator to be structurally integrated 00:03:39.080 --> 00:03:42.219 into the simulation itself, constantly making 00:03:42.219 --> 00:03:44.500 decisions that actively alter the state of the 00:03:44.500 --> 00:03:47.000 model in real time. Yeah. I look at it this way. 00:03:47.439 --> 00:03:49.860 A human out of the loop simulation is like trying 00:03:49.860 --> 00:03:52.800 to solve a highly complex physics equation on 00:03:52.800 --> 00:03:55.259 a quiet whiteboard in an empty room. Right. You're 00:03:55.259 --> 00:03:56.879 going to arrive at the exact same answer every 00:03:56.879 --> 00:03:59.300 single time. But a human in the loop simulation 00:03:59.300 --> 00:04:01.560 is like trying to solve that exact same equation 00:04:01.560 --> 00:04:03.819 while riding a roller coaster. That is a great 00:04:03.819 --> 00:04:06.300 way to put it. You know. Yeah. You might drop 00:04:06.300 --> 00:04:08.819 your pen because of the g -force or you might 00:04:08.819 --> 00:04:11.259 misread a vital number because you're dizzy or 00:04:11.259 --> 00:04:14.539 you might just freeze in absolute panic. But, 00:04:14.620 --> 00:04:16.360 and here's where I kind of have to push back 00:04:16.360 --> 00:04:19.199 on the engineering logic, if the whole point 00:04:19.199 --> 00:04:22.420 of computer modeling is to achieve predictability, 00:04:23.120 --> 00:04:25.779 isn't inviting the roller coaster panic a massive 00:04:25.779 --> 00:04:28.899 step backward? Like, why corrupt clean data? 00:04:29.120 --> 00:04:31.240 Well, what's fascinating here is that capturing 00:04:31.240 --> 00:04:34.319 that exact physiological panic, the dizziness, 00:04:34.579 --> 00:04:37.040 the dropped pen on the roller coaster, that is 00:04:37.040 --> 00:04:40.600 the entire point. Really? Yeah. An intelligent, 00:04:40.899 --> 00:04:43.399 automated system can crunch the math perfectly, 00:04:43.399 --> 00:04:46.180 but it has no earthly idea how a human being 00:04:46.180 --> 00:04:48.800 will actually react when reality deviates from 00:04:48.800 --> 00:04:51.300 the mathematical plan. Oh, I see. So capturing 00:04:51.300 --> 00:04:53.680 human error isn't corrupting the data. It's actually 00:04:53.680 --> 00:04:56.160 discovering the system's blind spots before they 00:04:56.160 --> 00:04:58.199 happen in the real world. So it's about mapping 00:04:58.199 --> 00:05:00.600 the friction between human cognition and machine 00:05:00.600 --> 00:05:02.720 logic. Precisely. And if we look at the virtual 00:05:02.720 --> 00:05:05.300 simulation taxonomy, humans are brought into 00:05:05.300 --> 00:05:08.300 these loops to exercise three specific OK, what 00:05:08.300 --> 00:05:12.120 are they? First, motor control skills. So the 00:05:12.120 --> 00:05:14.579 physical act of manipulating an interface, like 00:05:14.579 --> 00:05:16.339 flying an airplane or steering a ship. Right. 00:05:16.819 --> 00:05:19.500 Second, decision making skills, such as, say, 00:05:19.579 --> 00:05:21.759 committing fire control resources in a high pressure 00:05:21.759 --> 00:05:25.579 scenario. And third, communication skills. Just 00:05:25.579 --> 00:05:27.759 acting as a functional member of a command team 00:05:27.759 --> 00:05:30.259 where information has to be relayed verbally 00:05:30.259 --> 00:05:33.420 or textually. Which perfectly maps onto the wide 00:05:33.420 --> 00:05:35.740 variety of applications we see today. I mean, 00:05:35.879 --> 00:05:37.819 the frameworks cover everything from flight, 00:05:38.000 --> 00:05:43.040 driving, and marine simulators to complex video 00:05:43.040 --> 00:05:45.519 games, supply chain management models, and even 00:05:45.519 --> 00:05:47.860 digital puppetry for animation. Oh yeah, it's 00:05:47.860 --> 00:05:50.279 everywhere. And in all of these, the benefits 00:05:50.279 --> 00:05:52.560 seem to run in two parallel directions. You really 00:05:52.560 --> 00:05:54.720 do. The first benefit is mainly for the human, 00:05:54.879 --> 00:05:56.800 which is positive skill transfer. Like training. 00:05:57.079 --> 00:05:59.029 Exactly. The immersion of a human -in -the -loop 00:05:59.029 --> 00:06:01.910 system allows a trainee pilot in a flight simulator 00:06:01.910 --> 00:06:04.209 to actively change the outcome of a simulated 00:06:04.209 --> 00:06:07.129 engine failure. The system responds to their 00:06:07.129 --> 00:06:09.230 specific inputs, literally preparing their neural 00:06:09.230 --> 00:06:12.290 pathways for a real cockpit emergency. But the 00:06:12.290 --> 00:06:14.910 second benefit, which is arguably far more vital 00:06:14.910 --> 00:06:17.870 for the engineers themselves, is the identification 00:06:17.870 --> 00:06:21.209 of latent system flaws. OK, let's ground this 00:06:21.209 --> 00:06:24.449 with a concrete example for you listening. Imagine 00:06:24.449 --> 00:06:27.399 you are an air traffic controller. The Federal 00:06:27.399 --> 00:06:30.079 Aviation Administration, the FAA, is rolling 00:06:30.079 --> 00:06:33.319 out a brand new, highly advanced automation procedure 00:06:33.319 --> 00:06:35.800 designed to perfectly route commercial jets. 00:06:36.029 --> 00:06:38.410 And let's say the math behind this new system 00:06:38.410 --> 00:06:41.209 is totally flawless. Right. The tabletop closed 00:06:41.209 --> 00:06:44.009 loop simulations show zero midair collisions. 00:06:44.189 --> 00:06:46.670 But the FAA doesn't just push that code live 00:06:46.670 --> 00:06:49.470 based on the math. They build a human in the 00:06:49.470 --> 00:06:52.209 loop simulation. They put experienced air traffic 00:06:52.209 --> 00:06:54.470 controllers in front of identical radar screens, 00:06:55.050 --> 00:06:57.689 feed them simulated air traffic, and instruct 00:06:57.689 --> 00:07:00.110 them to use the new automated procedures to manage 00:07:00.110 --> 00:07:02.569 a simulated crisis. And this is where the rubber 00:07:02.569 --> 00:07:05.189 meets the road. Because what if the new system 00:07:05.189 --> 00:07:07.579 is technically mathematically perfect, but the 00:07:07.579 --> 00:07:10.600 user interface is an absolute nightmare. What 00:07:10.600 --> 00:07:13.839 if, during a simulated blizzard, the screen is 00:07:13.839 --> 00:07:16.740 so cluttered with automated alerts that the human 00:07:16.740 --> 00:07:19.639 controller misinterprets the data and routes 00:07:19.639 --> 00:07:22.720 two planes into the same airspace? The closed 00:07:22.720 --> 00:07:25.220 -loop math said the system was safe. But the 00:07:25.220 --> 00:07:28.459 human in the loop proved it was lethal. The system 00:07:28.459 --> 00:07:30.779 simply must account for human cognitive load, 00:07:31.360 --> 00:07:33.339 subjective judgment, and interface friction. 00:07:33.939 --> 00:07:36.500 If a perfectly calculated automation causes a 00:07:36.500 --> 00:07:39.420 human to make a fatal error, the system is fundamentally 00:07:39.420 --> 00:07:42.399 a failure. It's broken. Right. The human's unpredictable 00:07:42.399 --> 00:07:44.939 messiness is really the only variable that can 00:07:44.939 --> 00:07:47.660 genuinely validate the safety of the entire architecture. 00:07:48.079 --> 00:07:50.680 That transition from closed loop to open reality 00:07:50.680 --> 00:07:53.819 is so vital. We've seen how humans act as the 00:07:53.819 --> 00:07:57.000 ultimate real -world stress test, you know, basically 00:07:57.000 --> 00:07:59.180 experiencing a simulated system to locate its 00:07:59.180 --> 00:08:00.860 breaking points. Yeah. But what happens when 00:08:00.860 --> 00:08:03.100 we move upstream? Let's shift from humans experiencing 00:08:03.100 --> 00:08:05.579 systems to humans actually teaching the algorithms 00:08:05.579 --> 00:08:08.019 themselves long before those systems are ever 00:08:08.019 --> 00:08:11.060 compiled and run. Ah, so now we are stepping 00:08:11.060 --> 00:08:13.079 into the architecture of machine learning and 00:08:13.079 --> 00:08:16.100 artificial intelligence. And the dynamic fundamentally 00:08:16.100 --> 00:08:19.449 shifts here from tester to tutor. In a machine 00:08:19.449 --> 00:08:21.870 learning context, a human in the loop framework 00:08:21.870 --> 00:08:24.529 means that humans are actively aiding the computer 00:08:24.529 --> 00:08:27.470 in making correct decisions during the actual 00:08:27.470 --> 00:08:30.319 building and training of the model. And to appreciate 00:08:30.319 --> 00:08:32.419 the mechanical significance of this, we really 00:08:32.419 --> 00:08:34.700 have to look at the default alternative, which 00:08:34.700 --> 00:08:37.899 is random sampling. Right. Unsupervised learning. 00:08:38.360 --> 00:08:41.399 Exactly. A standard unsupervised machine learning 00:08:41.399 --> 00:08:44.500 algorithm operates by absorbing massive, totally 00:08:44.500 --> 00:08:46.779 unstructured data sets completely at random. 00:08:47.179 --> 00:08:50.059 It relies on sheer volume and processing power 00:08:50.059 --> 00:08:52.820 to find statistical correlations and slowly, 00:08:52.840 --> 00:08:55.509 slowly build a predictive model. Here's where 00:08:55.509 --> 00:08:57.669 it gets really interesting because the frameworks 00:08:57.669 --> 00:09:00.750 we are exploring explicitly detail that human 00:09:00.750 --> 00:09:02.629 -in -the -loop active learning exponentially 00:09:02.629 --> 00:09:05.149 improves machine learning over random sampling. 00:09:05.370 --> 00:09:07.389 It's night and day. And the mechanism behind 00:09:07.389 --> 00:09:10.289 how it does this is brilliant. Let's imagine 00:09:10.289 --> 00:09:12.889 traditional random sampling as a student trying 00:09:12.889 --> 00:09:15.970 to prepare for a massive final exam by blindly 00:09:15.970 --> 00:09:18.289 reading every single book in a million -volume 00:09:18.289 --> 00:09:22.200 library. Sounds exhausting. Right. They are absorbing 00:09:22.200 --> 00:09:25.240 terabytes of information, but they have absolutely 00:09:25.240 --> 00:09:28.500 no idea what is actually important and what is 00:09:28.500 --> 00:09:31.389 just noise. Because that is a massive computational 00:09:31.389 --> 00:09:34.289 search space. The algorithm is basically wasting 00:09:34.289 --> 00:09:36.830 millions of cycles processing perfectly standard 00:09:36.830 --> 00:09:39.429 data points that teach it absolutely nothing 00:09:39.429 --> 00:09:42.950 new. Exactly. Now, utilizing human in the loop 00:09:42.950 --> 00:09:45.669 is like assigning a master tutor to that student. 00:09:46.230 --> 00:09:48.429 The tutor taps the student on the shoulder, walks 00:09:48.429 --> 00:09:51.090 them past thousands of useless books, and pulls 00:09:51.090 --> 00:09:54.429 out the exact three obscure texts that contain 00:09:54.429 --> 00:09:56.850 the most difficult, nuanced questions that will 00:09:56.850 --> 00:09:59.570 definitely be on the exam. this to the bigger 00:09:59.570 --> 00:10:02.330 picture, that analogy perfectly isolates the 00:10:02.330 --> 00:10:04.889 mathematical advantage. The human tutor isn't 00:10:04.889 --> 00:10:07.529 taking the test for the algorithm. They are drastically 00:10:07.529 --> 00:10:09.809 reducing the search space by manually selecting 00:10:09.809 --> 00:10:11.909 the most critical high -friction data needed 00:10:11.909 --> 00:10:14.230 to refine the model's weights. Makes total sense. 00:10:14.549 --> 00:10:17.009 Because AI possesses staggering raw processing 00:10:17.009 --> 00:10:19.269 efficiency, but it completely lacks baseline 00:10:19.269 --> 00:10:21.549 contextual judgment. It doesn't inherently know 00:10:21.549 --> 00:10:23.250 the difference between a vital edge case and 00:10:23.250 --> 00:10:25.840 just irrelevant background static. Like, okay, 00:10:26.179 --> 00:10:28.659 let's say you're training a computer vision model 00:10:28.659 --> 00:10:32.759 for a self -driving car. If you use random sampling, 00:10:33.379 --> 00:10:35.840 the computer spends days analyzing millions of 00:10:35.840 --> 00:10:39.159 photos of perfectly clear, unobstructed stop 00:10:39.159 --> 00:10:40.820 signs. And it learns what a perfect stop sign 00:10:40.820 --> 00:10:43.360 looks like. Yeah. But what happens when it encounters 00:10:43.360 --> 00:10:46.639 a stop sign that is, like, 90 % covered by a 00:10:46.639 --> 00:10:49.019 snowbank in a tree branch? Well... The random 00:10:49.019 --> 00:10:51.700 sampling model likely just fails to recognize 00:10:51.700 --> 00:10:53.779 it because it hasn't seen enough statistical 00:10:53.779 --> 00:10:56.299 evidence of that highly specific degradation. 00:10:57.240 --> 00:11:00.559 But in a human -in -the -loop system, human labelers 00:11:00.559 --> 00:11:03.039 intentionally seek out those bizarre, confusing 00:11:03.039 --> 00:11:06.639 edge cases, the snowy hidden stop signs, and 00:11:06.639 --> 00:11:09.200 manually feed them to the algorithm, explicitly 00:11:09.200 --> 00:11:11.679 labeling them. So we're actively curating the 00:11:11.679 --> 00:11:15.250 curriculum. There is this misconception out there 00:11:15.250 --> 00:11:17.929 that forcing a computer to wait for a human to 00:11:17.929 --> 00:11:20.570 manually intervene and select data would create 00:11:20.570 --> 00:11:22.529 a bottleneck and actually slow the machine down. 00:11:22.610 --> 00:11:24.710 Yeah, people think that a lot. But structurally, 00:11:24.830 --> 00:11:27.820 the reality is the exact opposite. It accelerates 00:11:27.820 --> 00:11:30.559 the timeline dramatically. By aggressively filtering 00:11:30.559 --> 00:11:33.360 out the useless data and forcing the algorithm 00:11:33.360 --> 00:11:36.080 to process only the most difficult, ambiguous 00:11:36.080 --> 00:11:39.820 examples, human intervention exponentially speeds 00:11:39.820 --> 00:11:42.360 up the model's overall accuracy. Because it's 00:11:42.360 --> 00:11:44.799 focused. Right. It actively prevents the computer 00:11:44.799 --> 00:11:47.360 from wasting expensive compute cycles learning 00:11:47.360 --> 00:11:50.379 the wrong lessons from irrelevant data. And the 00:11:50.379 --> 00:11:52.879 technical term for the symbiosis is humanistic 00:11:52.879 --> 00:11:55.559 intelligence. Yes. It's intelligence that literally 00:11:55.559 --> 00:11:58.940 arises directly from having the human being embedded 00:11:58.940 --> 00:12:02.000 inside the computational feedback loop. The machine 00:12:02.000 --> 00:12:04.840 supplies the tireless brute force raw processing 00:12:04.840 --> 00:12:07.919 power and the human supplies the subjective judgment, 00:12:08.340 --> 00:12:11.059 the context and like the instinct for edge cases. 00:12:11.379 --> 00:12:14.000 It is the ultimate master and apprentice dynamic. 00:12:14.200 --> 00:12:16.299 The human is carefully shaping the neural pathways 00:12:16.299 --> 00:12:18.379 so the machine can eventually operate completely 00:12:18.379 --> 00:12:20.940 independently. But that brings us to the inevitable 00:12:20.940 --> 00:12:23.539 friction point. We have discussed how humans 00:12:23.539 --> 00:12:26.519 test simulated systems to find their flaws. We 00:12:26.519 --> 00:12:29.019 have explored how humans act as master tutors 00:12:29.019 --> 00:12:31.360 to aggressively refine machine learning algorithms. 00:12:31.720 --> 00:12:34.919 But what happens when the tutor steps away? We 00:12:34.919 --> 00:12:37.279 train these systems in highly controlled environments, 00:12:37.639 --> 00:12:40.100 but the real world moves infinitely faster than 00:12:40.100 --> 00:12:42.620 our training data. We have to look at what happens 00:12:42.620 --> 00:12:45.240 when these systems graduate and deploy into reality 00:12:45.240 --> 00:12:48.610 specifically. When the stakes are quite literally 00:12:48.610 --> 00:12:51.090 life and death. Yeah, this takes us out of the 00:12:51.090 --> 00:12:53.950 laboratory and right into the realm of lethal 00:12:53.950 --> 00:12:57.269 autonomous weapons. This is where the theoretical 00:12:57.269 --> 00:12:59.289 paradigms of control we've been discussing just 00:12:59.289 --> 00:13:03.250 completely collide with incredibly grave real 00:13:03.250 --> 00:13:06.500 -world consequences. We can look to a widely 00:13:06.500 --> 00:13:09.759 cited 2012 report published by Human Rights Watch, 00:13:10.279 --> 00:13:12.259 specifically authored by Bonnie Doherty, which 00:13:12.259 --> 00:13:14.740 actually builds upon Department of Defense frameworks 00:13:14.740 --> 00:13:18.279 from 1998. Okay. And it establishes three distinct 00:13:18.279 --> 00:13:20.580 classifications for the degree of human control 00:13:20.580 --> 00:13:23.379 over autonomous weapon systems. These classifications 00:13:23.379 --> 00:13:25.779 are vital because they strictly define the boundary 00:13:25.779 --> 00:13:28.960 between human agency and machine autonomy in 00:13:28.960 --> 00:13:31.259 combat scenarios. Let's break them down. The 00:13:31.259 --> 00:13:33.759 first classification is human in the loop. In 00:13:33.759 --> 00:13:36.179 this architecture, a human must instigate the 00:13:36.179 --> 00:13:39.019 action of the weapon. The system is not fully 00:13:39.019 --> 00:13:41.480 autonomous. The machine's radar might handle 00:13:41.480 --> 00:13:43.480 the aiming, the tracking, and the ballistics 00:13:43.480 --> 00:13:47.120 math, but a human operator must physically pull 00:13:47.120 --> 00:13:50.700 the trigger or press the button to initiate the 00:13:50.700 --> 00:13:54.080 lethal action. Right. And then the second classification 00:13:54.080 --> 00:13:56.279 systematically steps back, the human involvement. 00:13:56.500 --> 00:13:58.620 This is human on the loop. On the loop. On the 00:13:58.620 --> 00:14:00.799 loop. Under this design, the weapon system has 00:14:00.799 --> 00:14:04.740 the autonomy to identify targets, lock on, and 00:14:04.740 --> 00:14:07.100 actually initiate the lethal action entirely 00:14:07.100 --> 00:14:10.600 on its own. However, a human operator is actively 00:14:10.600 --> 00:14:13.039 monitoring the system's decisions in real time 00:14:13.039 --> 00:14:15.720 and retains the mechanical ability to override 00:14:15.720 --> 00:14:18.139 or abort the action before it is completed. And 00:14:18.139 --> 00:14:20.580 then the third classification removes the biological 00:14:20.580 --> 00:14:22.919 element entirely, human out of the loop. Right. 00:14:23.179 --> 00:14:25.259 In this framework, once the weapon system is 00:14:25.259 --> 00:14:27.980 activated, no further human action or oversight 00:14:27.980 --> 00:14:30.779 is involved whatsoever. It sweeps for targets, 00:14:31.059 --> 00:14:33.259 selects them, and engages them autonomously. 00:14:33.480 --> 00:14:35.820 Entirely on its own. So what does this all mean 00:14:35.820 --> 00:14:38.700 for the reality of human control? To visualize 00:14:38.700 --> 00:14:42.820 this practically is in the loop, essentially 00:14:42.820 --> 00:14:44.860 like driving a car yourself and pressing the 00:14:44.860 --> 00:14:47.360 gas pedal, while on the loop is being a passenger 00:14:47.360 --> 00:14:49.580 in a self -driving car who just happens to have 00:14:49.580 --> 00:14:51.559 their hand hovering over the emergency brake. 00:14:51.740 --> 00:14:54.299 This raises an important question about the illusion 00:14:54.299 --> 00:14:57.299 of agency, and your analogy perfectly captures 00:14:57.299 --> 00:15:00.360 the profound psychological and mechanical shift 00:15:00.360 --> 00:15:03.559 occurring across these three tiers. Okay. Moving 00:15:03.559 --> 00:15:05.860 from in the loop to on the loop is a massive 00:15:05.860 --> 00:15:09.539 transition. The human downgrades from being the 00:15:09.539 --> 00:15:11.879 active instigator of the event to a basically 00:15:11.879 --> 00:15:15.019 passive reactive monitor. And finally, without 00:15:15.019 --> 00:15:17.580 of the loop, the human is rendered into an obsolete 00:15:17.580 --> 00:15:20.940 bystander. Let's apply this to a specific historical 00:15:20.940 --> 00:15:23.059 example to see how the emergency brake scenario 00:15:23.059 --> 00:15:25.139 actually plays out. We could look at these YMO 00:15:25.139 --> 00:15:27.700 -104 Patriot missile systems. Right. The Patriot 00:15:27.700 --> 00:15:30.019 missile system is a surface -to -air defense 00:15:30.019 --> 00:15:32.440 architecture designed to operate with a human 00:15:32.440 --> 00:15:35.580 on the loop. When set to automatic mode, its 00:15:35.580 --> 00:15:38.220 radar can detect an incoming threat, classify 00:15:38.220 --> 00:15:41.259 it, arm an interceptor, and fire all autonomously. 00:15:41.580 --> 00:15:44.659 The human operator sits at a console, monitoring 00:15:44.659 --> 00:15:46.799 the engagements, theoretically holding the power 00:15:46.799 --> 00:15:49.039 to hit an abort button, if the computer targets 00:15:49.039 --> 00:15:51.840 a friendly aircraft or a civilian anomaly. But 00:15:51.840 --> 00:15:54.820 there is a terrifying physiological flaw in relying 00:15:54.820 --> 00:15:57.899 on that human emergency brake. The documentation 00:15:57.899 --> 00:16:00.399 notes that systems operating under these parameters 00:16:00.399 --> 00:16:02.600 have actually posed serious threats to friendly 00:16:02.600 --> 00:16:05.240 forces. And when we look at the mechanics of 00:16:05.240 --> 00:16:08.379 why, it's not a failure of the machine's code, 00:16:08.519 --> 00:16:11.370 it's a failure of human biology. Exactly. And 00:16:11.370 --> 00:16:13.769 when we analyze these classifications objectively, 00:16:14.129 --> 00:16:16.769 totally independently of any specific military 00:16:16.769 --> 00:16:19.769 doctrine or policy, the inescapable reality is 00:16:19.769 --> 00:16:22.509 the delta in processing speed. A modern automated 00:16:22.509 --> 00:16:25.470 defense system computes in microseconds. It can 00:16:25.470 --> 00:16:28.389 detect a radar signature, classify its trajectory, 00:16:28.889 --> 00:16:31.370 and launch a supersonic interceptor in a fraction 00:16:31.370 --> 00:16:34.610 of a second. The biological limit of human reaction 00:16:34.610 --> 00:16:37.429 time Just for the optic nerve to send a signal 00:16:37.429 --> 00:16:39.850 to the visual cortex, for the brain to process 00:16:39.850 --> 00:16:42.149 the anomaly, and for the finger to physically 00:16:42.149 --> 00:16:46.269 depress an abort button is roughly 250 milliseconds. 00:16:46.529 --> 00:16:48.669 Meaning the emergency brake is entirely useless 00:16:48.669 --> 00:16:50.929 if the car is moving faster than your brain can 00:16:50.929 --> 00:16:54.230 even perceive the road? Yes. Even if you're hypervigilant, 00:16:54.570 --> 00:16:57.289 staring directly at the screen, the machine acts 00:16:57.289 --> 00:16:59.669 before you can cognitively formulate the command 00:16:59.669 --> 00:17:02.019 to stop it. The human operator is technically 00:17:02.019 --> 00:17:04.440 present, perfectly satisfying a bureaucratic 00:17:04.440 --> 00:17:06.980 requirement for human oversight, but functionally, 00:17:07.099 --> 00:17:09.579 the sheer speed of the legal action renders their 00:17:09.579 --> 00:17:12.380 judgment completely moot. It is a stark proof 00:17:12.380 --> 00:17:14.460 that simply hardwiring a human being into the 00:17:14.460 --> 00:17:16.559 architecture of an on -the -loop system does 00:17:16.559 --> 00:17:18.519 not mathematically guarantee that unintended 00:17:18.519 --> 00:17:20.880 threats are mitigated. The friction of human 00:17:20.880 --> 00:17:23.420 interaction cannot save a situation if the machine 00:17:23.420 --> 00:17:27.140 operates faster than human thought. Wow. It is 00:17:27.140 --> 00:17:29.819 a deeply sobering reality check on the limits 00:17:29.819 --> 00:17:33.779 of our engineering hubris. I mean, we spend billions 00:17:33.779 --> 00:17:35.960 building these deterministic systems. We thoroughly 00:17:35.960 --> 00:17:38.579 test them in closed simulators. We act as master 00:17:38.579 --> 00:17:41.619 tutors, feeding them perfectly curated edge case 00:17:41.619 --> 00:17:44.480 data. But once they're deployed into the chaotic 00:17:44.480 --> 00:17:47.099 high speed reality of the physical world, our 00:17:47.099 --> 00:17:50.539 position in the loop or on the loop becomes incredibly 00:17:50.539 --> 00:17:53.460 precarious. Oh, absolutely. We are entirely outpaced 00:17:53.460 --> 00:17:55.599 by the very perfection we tried to build. It 00:17:55.599 --> 00:17:58.819 really forces us to confront a structural reality. 00:17:59.119 --> 00:18:01.339 Our technological capability to automate action 00:18:01.339 --> 00:18:04.559 is advancing significantly faster than our cognitive 00:18:04.559 --> 00:18:06.980 ability to effectively supervise it. And bringing 00:18:06.980 --> 00:18:08.859 this entire deep dive back to you, the listener, 00:18:09.359 --> 00:18:11.539 you don't need to be a fighter pilot in a simulator 00:18:11.539 --> 00:18:14.519 or an AI engineer labeling snowy stop signs to 00:18:14.519 --> 00:18:16.359 be part of this dynamic. Not at all. Whether 00:18:16.359 --> 00:18:18.740 you consciously realize it or not, you are actively 00:18:18.740 --> 00:18:21.119 participating as a human in the loop every single 00:18:21.119 --> 00:18:23.240 day. It is basically the invisible architecture 00:18:23.240 --> 00:18:25.779 of modern digital life. Think about it. When 00:18:25.779 --> 00:18:27.940 you are scrolling through your social media feed, 00:18:28.539 --> 00:18:31.000 hovering over a video for an extra three seconds, 00:18:31.519 --> 00:18:34.359 you are actively tutoring the algorithm. Yep. 00:18:34.759 --> 00:18:37.619 When you play a complex video game and find an 00:18:37.619 --> 00:18:40.000 unexpected shortcut, or when you are testing 00:18:40.000 --> 00:18:42.180 a new workflow software at your office and click 00:18:42.180 --> 00:18:45.240 the wrong menu entirely, you are the human in 00:18:45.240 --> 00:18:48.309 the loop. The massive data centers running these 00:18:48.309 --> 00:18:51.190 tools desperately rely on your unpredictable, 00:18:51.190 --> 00:18:53.809 highly subjective judgment. They need it. The 00:18:53.809 --> 00:18:55.750 algorithms need to know what you click on by 00:18:55.750 --> 00:18:58.769 mistake, what interface you find confusing, and 00:18:58.769 --> 00:19:01.589 what you choose to ignore. Your human error, 00:19:01.809 --> 00:19:04.329 your biological messiness, is not a bug to be 00:19:04.329 --> 00:19:07.490 squashed out by engineers. It is a vital load 00:19:07.490 --> 00:19:09.650 -bearing feature of the entire technological 00:19:09.650 --> 00:19:12.650 ecosystem. It is the compass the machine uses 00:19:12.650 --> 00:19:15.410 to navigate the messiness of human reality. Which 00:19:15.410 --> 00:19:17.750 leaves us with a rather profound paradox to consider 00:19:17.750 --> 00:19:20.829 as we wrap up. Oh, yeah. Yeah. If the ultimate 00:19:20.829 --> 00:19:23.490 engineering goal of utilizing human -in -the 00:19:23.490 --> 00:19:26.230 -loop systems, whether in machine learning, aviation, 00:19:26.410 --> 00:19:29.869 or simulation, is to constantly refine, correct, 00:19:30.029 --> 00:19:33.009 and perfect these artificial models. Wait, sorry. 00:19:33.009 --> 00:19:35.049 Let me restart that thought. If the goal is to 00:19:35.049 --> 00:19:37.529 refine and perfect these models, are we essentially 00:19:37.529 --> 00:19:40.009 using our temporary vital position in the loop 00:19:40.009 --> 00:19:42.730 today to successfully design ourselves completely 00:19:42.730 --> 00:19:45.900 out of the loop for the future? Wow. It's a question 00:19:45.900 --> 00:19:48.660 that brings us right back to the beginning. We 00:19:48.660 --> 00:19:50.960 are the panic on the roller coaster, doing everything 00:19:50.960 --> 00:19:53.099 we possibly can to teach the machine how to be 00:19:53.099 --> 00:19:56.160 perfectly still. But once the machine finally 00:19:56.160 --> 00:19:58.660 learns to be perfect and no longer needs our 00:19:58.660 --> 00:20:00.500 friction, what happens to the roller coaster? 00:20:01.539 --> 00:20:03.160 Something to think about the next time a website 00:20:03.160 --> 00:20:05.160 asks you to click on all the squares with traffic 00:20:05.160 --> 00:20:07.440 lights to prove you're not a robot. Thanks for 00:20:07.440 --> 00:20:09.079 joining us on this deep dive. We'll catch you 00:20:09.079 --> 00:20:09.460 next time.