WEBVTT 00:00:00.000 --> 00:00:03.020 Welcome back to the Deep Dive. So if you're like 00:00:03.020 --> 00:00:05.780 most people right now, AI isn't. I mean, it's 00:00:05.780 --> 00:00:08.080 not some far off concept anymore. It's right 00:00:08.080 --> 00:00:10.619 here. It's central to everything. A bit chaotic, 00:00:10.820 --> 00:00:12.580 right? You've got students using it. Managers 00:00:12.580 --> 00:00:14.339 are definitely assessing it. Oh, absolutely. 00:00:14.539 --> 00:00:16.559 Every professional is trying to figure out is 00:00:16.559 --> 00:00:19.679 this a tool or is this a thread? Exactly. And 00:00:19.679 --> 00:00:22.039 the speed of these systems, their competency. 00:00:22.860 --> 00:00:26.339 It's forcing us all to look at the fundamental 00:00:26.339 --> 00:00:28.980 work we do, whether that's learning a new skill 00:00:28.980 --> 00:00:32.039 or, you know, writing a big report. What is getting 00:00:32.039 --> 00:00:35.740 lost when things just become too efficient? And 00:00:35.740 --> 00:00:38.799 that is the mission for today. We're going to 00:00:38.799 --> 00:00:41.119 do a deep dive into this idea of black box learning. 00:00:41.620 --> 00:00:43.799 We're leaning on some analysis from Gary Ackerman, 00:00:43.859 --> 00:00:46.159 who was reading Brian Christian's book, The Alignment 00:00:46.159 --> 00:00:49.020 Problem. Our goal is to really define what this 00:00:49.020 --> 00:00:51.899 black box is, especially with modern AI, and 00:00:51.899 --> 00:00:53.899 then figure out why its efficiency can actually 00:00:53.899 --> 00:00:56.060 undermine the entire point of learning. OK. So 00:00:56.060 --> 00:00:58.500 let's dig in. The term black box, I feel like 00:00:58.500 --> 00:01:00.079 we all think we know what that means. Right. 00:01:00.079 --> 00:01:01.719 You put something in, you get something out. 00:01:02.000 --> 00:01:04.180 Exactly. You can see the input, a prompt, some 00:01:04.180 --> 00:01:07.400 data, and you see the output, an essay, a recommendation, 00:01:07.599 --> 00:01:09.659 whatever. The middle part. The middle part is 00:01:09.659 --> 00:01:12.959 mystery. The logic, the actual operations that 00:01:12.959 --> 00:01:16.040 turned A into B, they're hidden. But the sources 00:01:16.040 --> 00:01:17.719 we're looking at, they add a really interesting 00:01:17.719 --> 00:01:20.200 layer to this, a distinction. They do. They talk 00:01:20.200 --> 00:01:24.040 about two different kinds of opacity, really, 00:01:24.040 --> 00:01:26.079 when it comes to machine learning, two knowledge 00:01:26.079 --> 00:01:28.060 gaps. OK, so what's the first one? Well, the 00:01:28.060 --> 00:01:30.299 first is kind of the classic version. The logic 00:01:30.299 --> 00:01:35.099 is unknown to the user, but the designers, they 00:01:35.099 --> 00:01:38.650 know it. Right, like a trade secret. A proprietary 00:01:38.650 --> 00:01:41.090 algorithm? Precisely. The knowledge exists. It's 00:01:41.090 --> 00:01:43.250 just locked away. But then there's the second 00:01:43.250 --> 00:01:45.269 layer. And this is the one that's really central 00:01:45.269 --> 00:01:49.129 to modern AI. It is. This is where it gets, frankly, 00:01:49.209 --> 00:01:52.049 a bit unsettling. In a lot of these complex deep 00:01:52.049 --> 00:01:54.609 learning systems, the logic isn't just unknown 00:01:54.609 --> 00:01:57.689 to the user. Often, even the people who deploy 00:01:57.689 --> 00:02:00.370 the system don't fully understand the internal 00:02:00.370 --> 00:02:02.980 logic. Wait, really? Even the deployers? Even 00:02:02.980 --> 00:02:05.959 the deployers, they know the correlation. Input 00:02:05.959 --> 00:02:09.099 X gives you output Y with incredible accuracy, 00:02:09.620 --> 00:02:12.319 but they can't actually trace the step -by -step 00:02:12.319 --> 00:02:15.360 causal path inside the box. That's a huge distinction. 00:02:15.479 --> 00:02:18.360 This isn't just a secret. It's... It's genuine 00:02:18.360 --> 00:02:21.939 opacity. And that brings us right to causality 00:02:21.939 --> 00:02:24.580 versus correlation. Because if that internal 00:02:24.580 --> 00:02:27.439 part is a total mystery, you can still get amazing 00:02:27.439 --> 00:02:29.800 correlations. You can predict what will happen 00:02:29.800 --> 00:02:32.719 with stunning accuracy. But you cannot understand 00:02:32.719 --> 00:02:36.060 causality. You don't know how that input was 00:02:36.060 --> 00:02:38.699 caused to become that output. And that difference 00:02:38.699 --> 00:02:42.080 is just, it's everything. It's vital. Especially 00:02:42.080 --> 00:02:43.699 when the stakes are high. Okay, let's use an 00:02:43.699 --> 00:02:45.840 example here before we get to education. Say 00:02:45.840 --> 00:02:48.419 a bank uses an AI tool for loan applications, 00:02:48.719 --> 00:02:52.039 and it's 99 % accurate. Great correlation. Fantastic 00:02:52.039 --> 00:02:54.639 correlation. But then it starts flagging applicants 00:02:54.639 --> 00:02:57.020 from a certain neighborhood, a specific zip code. 00:02:57.080 --> 00:02:59.340 And because it's a black box. The bank can't 00:02:59.340 --> 00:03:02.560 say why. They have no rationale for why those 00:03:02.560 --> 00:03:04.560 inputs led to that output. They just know it 00:03:04.560 --> 00:03:06.800 happens. They can't check for bias. They can't 00:03:06.800 --> 00:03:08.300 explain it to a regulator. And they can't fix 00:03:08.300 --> 00:03:11.340 it. You can't fix it. How do you debug something 00:03:11.340 --> 00:03:14.039 when you can't see the code? You can't point 00:03:14.039 --> 00:03:16.000 to the specific weight or decision that caused 00:03:16.000 --> 00:03:18.560 the problem. So the data looks perfect, but the 00:03:18.560 --> 00:03:20.979 knowledge is a dead end. Exactly. The knowledge 00:03:20.979 --> 00:03:23.960 is capped. Causality is what lets you fix things, 00:03:24.120 --> 00:03:26.360 extend things, make them better. Without it, 00:03:26.360 --> 00:03:29.000 you're just stuck. That is the perfect transition. 00:03:29.060 --> 00:03:31.099 So if that's the risk in finance or medicine, 00:03:31.500 --> 00:03:34.259 let's bring this black box problem into, well, 00:03:34.759 --> 00:03:36.919 into learning. Right. And the source material 00:03:36.919 --> 00:03:39.939 argues that no matter the subject, The real work 00:03:39.939 --> 00:03:43.460 of learning is just answering questions. Even 00:03:43.460 --> 00:03:45.259 if they look like other things. Yeah, even if 00:03:45.259 --> 00:03:48.000 they're disguised as essays or, you know, proofs 00:03:48.000 --> 00:03:50.879 or business plans, the process is always converting 00:03:50.879 --> 00:03:53.599 an input the assignment into a graded output. 00:03:53.900 --> 00:03:56.379 And here's the trap. Machine learning tools are 00:03:56.379 --> 00:03:58.840 unbelievably good at this. They're answer machines. 00:03:59.060 --> 00:04:01.979 They are built to maximize the correlation between 00:04:01.979 --> 00:04:04.139 a prompt and what a good answer should look like. 00:04:04.219 --> 00:04:06.740 Which is, let's be honest, incredibly tempting 00:04:06.740 --> 00:04:10.250 for a student. Oh, it's a seductive power. You're 00:04:10.250 --> 00:04:13.389 faced with this huge complex assignment, and 00:04:13.389 --> 00:04:15.870 the AI black box just hands you the perfect output 00:04:15.870 --> 00:04:18.850 almost instantly. So here's the pushback I imagine. 00:04:19.209 --> 00:04:22.470 If the AI output is genuinely good, like 95 % 00:04:22.470 --> 00:04:24.689 of what the teacher wanted, why should an efficiency 00:04:24.689 --> 00:04:27.050 win? Why should a student struggle for five hours 00:04:27.050 --> 00:04:29.459 on something an AI does in five seconds? That 00:04:29.459 --> 00:04:31.779 is the critical question. Yeah. And it's because 00:04:31.779 --> 00:04:34.360 the student's job isn't just to produce the output. 00:04:34.680 --> 00:04:36.660 The problem is that when the student uses the 00:04:36.660 --> 00:04:40.480 black box, they get the data of learning the 00:04:40.480 --> 00:04:43.600 finished essay, but they completely bypass building 00:04:43.600 --> 00:04:45.600 the internal logic that was supposed to produce 00:04:45.600 --> 00:04:48.019 it. So they get the perfect cake, but they didn't 00:04:48.019 --> 00:04:50.079 measure the flour or crack an egg. They didn't 00:04:50.079 --> 00:04:52.819 even read the recipe. Not a single step. They 00:04:52.819 --> 00:04:54.600 found a correlation between the prompt and the 00:04:54.600 --> 00:04:56.319 answer, but they learned nothing about the causal 00:04:56.319 --> 00:04:59.639 mechanism and that mechanism. That is what we 00:04:59.639 --> 00:05:01.459 call learning. So let's talk about the student's 00:05:01.459 --> 00:05:03.540 actual job description then. Yeah. According 00:05:03.540 --> 00:05:06.019 to the sources, a student's real job is to take 00:05:06.019 --> 00:05:08.779 that input, the question, and use their own logic. 00:05:08.920 --> 00:05:11.100 Their own developing internal operations. Yeah. 00:05:11.220 --> 00:05:13.800 To generate the output. It's an internal manufacturing 00:05:13.800 --> 00:05:15.899 process. You're building mental scaffolding. 00:05:16.100 --> 00:05:18.259 You're laying the wiring in your own brain. So 00:05:18.259 --> 00:05:20.259 for a math problem, it's not the final number. 00:05:20.379 --> 00:05:22.860 It's building the sequence of steps that cause 00:05:22.860 --> 00:05:26.019 that number. Exactly. Or for a history essay, 00:05:26.339 --> 00:05:29.089 it's constructing the argument. the causal chain 00:05:29.089 --> 00:05:31.709 of evidence that leads to your conclusion. And 00:05:31.709 --> 00:05:34.769 the black box just lets you skip all of that 00:05:34.769 --> 00:05:39.110 heavy lifting. The AI path is fast, efficient. 00:05:39.329 --> 00:05:41.089 And the learning path is the total opposite. 00:05:41.310 --> 00:05:44.449 It's slow, it's difficult, it can be really frustrating. 00:05:45.149 --> 00:05:47.550 But that inefficiency is the entire point. But 00:05:47.550 --> 00:05:50.350 the pressure is real. The student gets the grade 00:05:50.350 --> 00:05:53.250 now by using the box, so besides just... you 00:05:53.250 --> 00:05:55.569 know, not learning, what's the actual consequence? 00:05:55.850 --> 00:05:57.769 There are two big ones. First, like we said, 00:05:57.769 --> 00:05:59.709 you haven't built the internal model, so you 00:05:59.709 --> 00:06:01.810 can't apply that knowledge to a new problem later. 00:06:02.250 --> 00:06:04.269 And second, it messes with the production of 00:06:04.269 --> 00:06:06.389 data that the instructor needs to see. Ah, the 00:06:06.389 --> 00:06:09.750 breadcrumbs. The breadcrumbs, exactly. A student's 00:06:09.750 --> 00:06:11.870 messy draft, the things they cross out, their 00:06:11.870 --> 00:06:15.009 step -by -step work. That visible trail is the 00:06:15.009 --> 00:06:17.110 only way an instructor knows that the student's 00:06:17.110 --> 00:06:19.170 own brain is developing. So if you just hand 00:06:19.170 --> 00:06:21.410 in the perfect cake with no flour on your hands. 00:06:21.470 --> 00:06:24.060 Right, there's no proof of work. And the instructor 00:06:24.060 --> 00:06:26.839 loses their ability to diagnose where you're 00:06:26.839 --> 00:06:29.660 struggling or what you've mastered. That output 00:06:29.660 --> 00:06:32.939 becomes incredibly hard to accept as proof of 00:06:32.939 --> 00:06:35.699 real learning. So it comes down to this. Causing 00:06:35.699 --> 00:06:37.660 the logic and the operations that become your 00:06:37.660 --> 00:06:40.819 output. That is hard work. It requires struggle. 00:06:41.500 --> 00:06:44.199 But that struggle is the job of being a student. 00:06:44.459 --> 00:06:47.220 If you bypass that, you bypass the learning itself. 00:06:47.759 --> 00:06:49.959 That struggle is the only way to go from knowing 00:06:49.959 --> 00:06:53.019 what to really knowing how and why. Fantastic. 00:06:53.160 --> 00:06:55.579 So let's quickly consolidate this. In this deep 00:06:55.579 --> 00:06:57.980 dive, we define the black box input, output, 00:06:58.519 --> 00:07:01.500 but the causal logic inside is a mystery. And 00:07:01.500 --> 00:07:03.860 sometimes a mystery even to its creators. Right. 00:07:04.079 --> 00:07:06.560 And we saw that AI is incredibly efficient at 00:07:06.560 --> 00:07:08.800 producing answers, but that efficiency is a trap. 00:07:09.000 --> 00:07:11.459 It directly undermines the student's real job. 00:07:11.680 --> 00:07:14.439 Which is to build and strengthen their own internal 00:07:14.439 --> 00:07:17.040 logic, their own understanding. And this goes 00:07:17.040 --> 00:07:20.300 way beyond the classroom. This is the core takeaway 00:07:20.300 --> 00:07:23.139 for you, for the learner. Absolutely. Understanding 00:07:23.139 --> 00:07:26.180 how knowledge is made, the causality, the mechanism. 00:07:26.800 --> 00:07:29.100 That's the only way you can ever truly advance 00:07:29.100 --> 00:07:31.639 it or fix its flaws or adapt it to something 00:07:31.639 --> 00:07:34.800 new in your career. If you only rely on correlation, 00:07:35.319 --> 00:07:37.240 you're accepting a ceiling on your own potential. 00:07:37.399 --> 00:07:39.519 And that is the essential insight we want to 00:07:39.519 --> 00:07:41.480 leave you with. If the goal of real learning 00:07:41.480 --> 00:07:44.220 isn't just predicting the right answer correlation, 00:07:44.459 --> 00:07:48.120 but understanding the why, the causation, then 00:07:48.120 --> 00:07:51.279 here's one final thought. How can you apply this 00:07:51.279 --> 00:07:54.500 black box thinking to your own work? Where are 00:07:54.500 --> 00:07:56.779 you maybe accepting a correlation without demanding 00:07:56.779 --> 00:07:58.920 to understand the cause? Where are you using 00:07:58.920 --> 00:08:01.740 an answer or relying on a system's output without 00:08:01.740 --> 00:08:03.519 knowing the logic that made it? Because if you 00:08:03.519 --> 00:08:06.189 don't know the logic, You can't innovate on it. 00:08:06.370 --> 00:08:08.610 You can't correct it. So demand the rationale. 00:08:08.790 --> 00:08:11.170 Demand the causality. That's the work of a true 00:08:11.170 --> 00:08:13.750 expert. We'll see you next time on the Deep Dive.