WEBVTT 1 00:00:00.080 --> 00:00:04.280 Okay, let's dive in navigating the world of artificial intelligence AI. 2 00:00:05.160 --> 00:00:07.919 It can feel like trying to cross an ocean in 3 00:00:08.000 --> 00:00:11.080 a tea cup. Really, everywhere you look, there's just so 4 00:00:11.279 --> 00:00:16.120 much information, hype, expectations flying around. 5 00:00:16.239 --> 00:00:19.039 It's completely true. Yeah, you hear everything, don't you. From 6 00:00:19.120 --> 00:00:22.399 AI being I don't know, cute and fuzzy right all 7 00:00:22.440 --> 00:00:27.359 the way to being this potential mass murderer, it's hard 8 00:00:27.359 --> 00:00:29.320 to get a clear picture of what the tech can 9 00:00:29.320 --> 00:00:30.280 actually do right now, and. 10 00:00:30.320 --> 00:00:32.200 That's exactly what we're trying to do in this deep dive. 11 00:00:32.200 --> 00:00:34.079 We've got the stack of sources and they offer a 12 00:00:34.119 --> 00:00:36.799 really balanced view, kind of middle of the road, trying 13 00:00:36.840 --> 00:00:39.880 to cut through all that, you know, the crazy hype, 14 00:00:39.880 --> 00:00:41.399 but also the doom and gloom. 15 00:00:41.759 --> 00:00:43.759 Yeah, our goal here is really just to pull out 16 00:00:43.759 --> 00:00:46.479 the most important insights from all this material. We want 17 00:00:46.479 --> 00:00:50.039 to give you the listener, a clearer understanding of what 18 00:00:50.079 --> 00:00:54.320 AI actually is, how it works, maybe some surprising places it's. 19 00:00:54.240 --> 00:00:56.679 Used, and crucially where it just falls short. 20 00:00:56.520 --> 00:00:58.359 Right now, exactly where the limits are. 21 00:00:58.679 --> 00:01:01.000 So let's try in for at the big headlines for 22 00:01:01.039 --> 00:01:05.159 a bit and really dive into the foundations the practical 23 00:01:05.200 --> 00:01:06.680 side of AI. Let's unpack it. 24 00:01:06.840 --> 00:01:12.599 Okay, so where do we even start defining AI? It 25 00:01:12.640 --> 00:01:14.959 feels like everyone has a slightly different take. 26 00:01:15.359 --> 00:01:17.519 Well, yeah, and our sources point that out. How you 27 00:01:17.599 --> 00:01:19.840 define AI, it's kind of a mix, isn't it. It 28 00:01:19.879 --> 00:01:22.480 depends on your goals, the tech you're using, your perspective, 29 00:01:22.599 --> 00:01:24.959 your perspective exactly, which is why you get that whole 30 00:01:25.000 --> 00:01:28.439 spectrum from the cute and fuzzy view to the potential 31 00:01:28.439 --> 00:01:29.319 disaster end. 32 00:01:29.560 --> 00:01:32.680 Right. But the sources we're using mostly stick to that 33 00:01:32.760 --> 00:01:34.439 middle ground, trying to be objective. 34 00:01:34.480 --> 00:01:35.200 Here's just helpful. 35 00:01:35.439 --> 00:01:37.799 And they also remind us that, you know, human intelligence 36 00:01:37.879 --> 00:01:40.840 isn't just one single thing. There are multiple kinds, right. 37 00:01:40.719 --> 00:01:43.239 Like Gardener's multiple intelligences ideas. 38 00:01:43.120 --> 00:01:46.439 Sort of yeah, and computers or AI they can simulate 39 00:01:46.480 --> 00:01:47.120 some of those. 40 00:01:47.120 --> 00:01:51.280 Like physical movement skills exactly. 41 00:01:51.319 --> 00:01:55.680 That think about what sometimes called bodily kinesthetic intelligence. You know, 42 00:01:56.000 --> 00:01:59.400 the kind of surgeon uses or a really skilled crafts person. 43 00:02:00.439 --> 00:02:04.040 Robots are masters at that specific kind of thing, performing 44 00:02:04.079 --> 00:02:07.359 these repetitive physical taxts incredibly precisely. 45 00:02:07.400 --> 00:02:09.639 So it's a simulation of a human skill, but driven 46 00:02:09.680 --> 00:02:13.599 by programming. Often yes, impressive, but specific and then there's 47 00:02:13.639 --> 00:02:16.159 the whole acting humanly idea that brings up the Turing test, 48 00:02:16.159 --> 00:02:16.680 doesn't it? Right? 49 00:02:16.719 --> 00:02:18.479 The classic Turing test can you tell if you're talking 50 00:02:18.520 --> 00:02:21.960 to a machine or a person. And for an AI 51 00:02:22.080 --> 00:02:24.719 to even try that, it needs a bunch of things 52 00:02:25.319 --> 00:02:30.360 understanding language, representing knowledge, somehow reasoning with that knowledge. 53 00:02:30.000 --> 00:02:32.080 And learning right. Machine learning comes in there. 54 00:02:31.879 --> 00:02:34.919 And machine learning absolutely to adapt and improve. But that 55 00:02:34.960 --> 00:02:38.120 whole Turing test focus, it's just one angle on AI. 56 00:02:38.280 --> 00:02:41.560 Really, it feels like a big leap from say, those 57 00:02:41.599 --> 00:02:42.800 older expert systems. 58 00:02:42.840 --> 00:02:45.879 What were they exactly expert systems? They were totally a 59 00:02:45.879 --> 00:02:49.439 different approach. The idea was to capture knowledge from human 60 00:02:49.479 --> 00:02:53.639 experts and then encode it directly as rules like if 61 00:02:53.680 --> 00:02:57.520 then statements, Oh right, like some sophisticated grammar checkers. 62 00:02:57.560 --> 00:02:58.800 Maybe do those still count? 63 00:02:59.240 --> 00:03:03.800 Our source actually says yeah, fundamentally, things like complex grammar 64 00:03:03.879 --> 00:03:07.240 checkers still work like that. They're built on heaps of 65 00:03:07.319 --> 00:03:09.680 linguistic rules defined by experts. 66 00:03:10.159 --> 00:03:12.280 I remember using things like that, and they could be 67 00:03:12.360 --> 00:03:14.520 helpful but also really rigid. 68 00:03:14.599 --> 00:03:18.199 That was the trade off exactly. The big plus was transparency. 69 00:03:18.599 --> 00:03:20.840 You could literally see the rules it was following, so. 70 00:03:20.800 --> 00:03:22.960 You could understand why it made a decision. 71 00:03:22.680 --> 00:03:26.240 Precisely, and you could tweak the rules to improve it. 72 00:03:26.439 --> 00:03:29.439 But yeah, brittle is the word. Anything slightly outside the rules, 73 00:03:29.520 --> 00:03:31.879 it breaks, It just breaks. Yeah. Hard to build, hard 74 00:03:31.919 --> 00:03:35.400 to maintain, especially for complex problems. But they're still around, 75 00:03:35.439 --> 00:03:37.879 you know, for things like credit scoring, where you need 76 00:03:37.879 --> 00:03:38.759 that transparency. 77 00:03:38.960 --> 00:03:42.879 Okay, So if expert systems ran on human defined rules, 78 00:03:43.080 --> 00:03:46.400 what's fueling most modern AI Ah, Well. 79 00:03:46.240 --> 00:03:49.159 That's data, loads and loads of data, big data. 80 00:03:49.199 --> 00:03:51.199 And it's not just about having lots of data, is it. 81 00:03:51.199 --> 00:03:51.879 There's more to. 82 00:03:51.800 --> 00:03:54.840 It, oh much more. It's the volume, sure, but also 83 00:03:54.879 --> 00:03:57.240 the speed it comes in at the variety. It's the 84 00:03:57.280 --> 00:04:02.280 scale that modern tech enables. Computers, sensors, smartphones, the internet. 85 00:04:02.560 --> 00:04:05.719 More's law enabling all that data handling absolutely connected. 86 00:04:05.759 --> 00:04:08.759 It lets us do fundamentally new things that just weren't 87 00:04:08.800 --> 00:04:09.639 possible before. 88 00:04:09.719 --> 00:04:11.719 And where's all the data actually coming from? Is it 89 00:04:11.759 --> 00:04:12.599 all automated? 90 00:04:12.960 --> 00:04:16.120 A surprising amount still comes from us from human input, 91 00:04:16.560 --> 00:04:20.680 even in automated systems. Every click online, every purchase, every comment, 92 00:04:20.959 --> 00:04:23.319 that's all data. It's all data. And then of course 93 00:04:23.360 --> 00:04:27.879 you have sensors and systems collecting vast amounts automatically, even 94 00:04:27.920 --> 00:04:30.480 little things like drop down lists on websites. Yeah, those 95 00:04:30.480 --> 00:04:33.439 are data input aids designed to make the data cleaner, 96 00:04:33.759 --> 00:04:35.360 more reliable when it's collected. 97 00:04:35.720 --> 00:04:38.360 So we've got this mountain of data. What are the 98 00:04:38.360 --> 00:04:40.160 big headaches in dealing with it? 99 00:04:40.480 --> 00:04:43.240 Well, missing data is a huge one. Always. Sometimes it's 100 00:04:43.240 --> 00:04:46.079 just random gaps. Sometimes the whole sequence is gone, and 101 00:04:46.120 --> 00:04:49.199 you have to decide do we drop that field entirely, 102 00:04:49.680 --> 00:04:52.279 try to calculate the missing values, find them. 103 00:04:52.199 --> 00:04:55.040 Somewhere else, or maybe even change the question you're asking 104 00:04:55.079 --> 00:04:57.319 based on what data you do have exactly. 105 00:04:57.439 --> 00:04:59.120 That's a real strategic decision. 106 00:04:59.199 --> 00:05:02.839 And or intuitively, perhaps having too much data can also 107 00:05:02.879 --> 00:05:03.480 be a problem. 108 00:05:03.560 --> 00:05:07.839 It absolutely can. Just drowning in data isn't helpful. AI 109 00:05:08.079 --> 00:05:11.560 needs well, just enough of the right data to solve 110 00:05:11.560 --> 00:05:15.160 the problem efficiently. Managing the flood is almost as important 111 00:05:15.199 --> 00:05:16.680 as getting the data in the first place. 112 00:05:16.959 --> 00:05:19.839 Now, this is where one of our sources gets really fascinating. 113 00:05:19.879 --> 00:05:22.959 It talks about the five mistruths in data, things like 114 00:05:23.120 --> 00:05:27.720 errors of commission omission, problems with perspective bias from of reference. 115 00:05:27.839 --> 00:05:30.600 Oh, this is such a crucial point because we humans, 116 00:05:30.600 --> 00:05:34.720 we deal with subjective stuff, opinions, even lies all the time. Right, 117 00:05:35.040 --> 00:05:37.480 We often have a gut feeling or use our experience, 118 00:05:37.480 --> 00:05:40.800 our imagination to see when data is skewed or just 119 00:05:40.920 --> 00:05:41.519 plain wrong. 120 00:05:41.680 --> 00:05:43.040 But a computer can't do that. 121 00:05:43.319 --> 00:05:47.079 Not really. To an AI, data is just data. It 122 00:05:47.079 --> 00:05:51.160 doesn't inherently distinguish between something factual and something that's well 123 00:05:51.240 --> 00:05:51.920 a mistruth. 124 00:05:52.120 --> 00:05:54.560 So where we might try to work with messy data 125 00:05:54.600 --> 00:05:56.920 trying to figure out the underlying truth, and. 126 00:05:57.040 --> 00:05:59.480 AI is more likely to see that messy data point 127 00:05:59.519 --> 00:06:02.160 as an outlme liar, something weird to be filtered out 128 00:06:02.199 --> 00:06:06.680 ignored our ability to handle ambiguity to make these intuitive 129 00:06:06.759 --> 00:06:10.600 leaps or even spot a fib based on context that's 130 00:06:10.639 --> 00:06:14.160 fundamentally different the source puts it nicely. AI is stuck 131 00:06:14.199 --> 00:06:15.319 in reality and. 132 00:06:15.240 --> 00:06:18.879 Those mistruths of commission, like someone just typing the wrong 133 00:06:18.959 --> 00:06:22.600 thing or passing on hearsay human spot that often we. 134 00:06:22.600 --> 00:06:25.319 Do, or we at least question it. AI struggles much 135 00:06:25.360 --> 00:06:27.600 more to navigate that kind of thing authentically. It just 136 00:06:27.600 --> 00:06:28.360 sees the input. 137 00:06:28.480 --> 00:06:31.480 Okay, so we've got this imperfect ocean of data. How 138 00:06:31.519 --> 00:06:33.639 do we actually use it to solve problems? 139 00:06:33.720 --> 00:06:37.399 That's algorithms right, Yeah, algorithms. At its core, an algorithm 140 00:06:37.439 --> 00:06:39.160 is just to set us steps to find a solution. 141 00:06:39.800 --> 00:06:42.839 AI algorithms are special because they tackle problems we usually 142 00:06:42.839 --> 00:06:46.720 think need human intelligence, like complex scheduling or finding the 143 00:06:46.720 --> 00:06:48.759 best route, or recognizing patterns. 144 00:06:48.800 --> 00:06:51.600 And the source mentioned these problems can be insanely complex 145 00:06:52.199 --> 00:06:53.720 AI complete. 146 00:06:53.360 --> 00:06:55.879 Right, those are the really tricky ones technically called n 147 00:06:56.000 --> 00:06:58.959 key complete in computer science. It means you can't just 148 00:06:59.000 --> 00:07:02.720 brute force it try every single possible combination. It would 149 00:07:02.759 --> 00:07:06.879 take forever too many possibilities, way too many think about 150 00:07:06.959 --> 00:07:10.639 scheduling deliveries for a huge fleet or playing a complex 151 00:07:10.720 --> 00:07:14.600 game like Go. AI algorithms need smarter strategies. 152 00:07:14.920 --> 00:07:16.879 Can you give an example what kind of strategies? 153 00:07:17.000 --> 00:07:19.600 Well, the source talks about things like states based search 154 00:07:19.639 --> 00:07:23.959 for games, figuring out moves ahead, but also local search 155 00:07:24.000 --> 00:07:26.920 and heuristics. That's more about starting with a guess, maybe 156 00:07:26.920 --> 00:07:30.360 not a perfect solution, and then using rules of thumb 157 00:07:30.360 --> 00:07:33.839 heuristics to explore nearby solutions, trying to improve it step 158 00:07:33.839 --> 00:07:37.199 by step, Like a robot navigating a room. Yeah, it 159 00:07:37.279 --> 00:07:40.519 uses sensors and heuristics to decide its next move, avoid obstacles, 160 00:07:40.600 --> 00:07:43.160 find its goal without needing a perfect map of the 161 00:07:43.160 --> 00:07:47.079 whole room beforehand. It's making local informed guesses, got it. 162 00:07:47.639 --> 00:07:51.079 So, taking all this data, these clever algorithms that leads 163 00:07:51.160 --> 00:07:53.600 us to machine learning, which feels like the core of 164 00:07:53.720 --> 00:07:54.839 modern AI hype. 165 00:07:54.959 --> 00:07:58.639 It's definitely central. mL is often seen as, yeah, the 166 00:07:58.680 --> 00:08:02.560 pinnacle of data now today. What makes it so powerful 167 00:08:02.639 --> 00:08:05.519 is that it learns directly from the data. It doesn't 168 00:08:05.600 --> 00:08:10.240 need humans to explicitly program every single rule beforehand, like. 169 00:08:10.160 --> 00:08:13.600 For things we just do naturally but can't explain step by. 170 00:08:13.519 --> 00:08:17.480 Step exactly, Like recognizing faces. We all do it, but 171 00:08:17.519 --> 00:08:19.680 could you write down the exact rules for how you 172 00:08:19.720 --> 00:08:23.759 recognize your friend? Probably not, no way. mL algorithms are 173 00:08:23.800 --> 00:08:27.639 good at figuring out those kinds of patterns from examples. 174 00:08:27.439 --> 00:08:29.519 So instead of being told the rules, it sort of 175 00:08:30.879 --> 00:08:32.279 deduces them from the data. 176 00:08:32.399 --> 00:08:35.120 Essentially, Yeah, it's trying to find a mathematical function a 177 00:08:35.200 --> 00:08:38.320 representation that connects the input data it sees to the 178 00:08:38.320 --> 00:08:42.320 correct output or label. The learning is really that mathematical 179 00:08:42.360 --> 00:08:44.960 process of finding the best fit. That's why it's called training. 180 00:08:45.240 --> 00:08:48.279 You train the algorithm to make the right associations. 181 00:08:47.679 --> 00:08:50.399 But it doesn't actually understand the meaning behind it. 182 00:08:50.559 --> 00:08:53.240 Not in the human sense. No, it doesn't grasp why 183 00:08:53.279 --> 00:08:56.480 that input means that output, just that they correlate strongly 184 00:08:56.480 --> 00:08:57.320 in the data it saw. 185 00:08:57.399 --> 00:09:00.720 And machine learning is already kind of everywhere, often hidden. 186 00:09:00.519 --> 00:09:04.159 Oh absolutely, it's running in our phones, our cars, filtering spam, 187 00:09:04.440 --> 00:09:08.960 recommending products, doing predictive analysis and businesses way faster than 188 00:09:09.039 --> 00:09:11.919 humans could. It's often invisible, but very wise. 189 00:09:11.679 --> 00:09:14.279 Bread Okay, but the source material is really clear on this. 190 00:09:14.919 --> 00:09:18.759 Machine learning, powerful as it is, has some serious limitations. 191 00:09:18.879 --> 00:09:22.919 Yes, crucially, it's a tool, a very sophisticated tool for 192 00:09:23.000 --> 00:09:26.519 analysis and finding patterns in data. It needs those algorithms, 193 00:09:26.519 --> 00:09:30.519 it needs those huge data sets. But it cannot think, 194 00:09:30.840 --> 00:09:34.559 it cannot feel. It has no self awareness, no consciousness, 195 00:09:34.559 --> 00:09:35.279 no free will. 196 00:09:35.519 --> 00:09:38.080 So it can spot patterns and medical scans. 197 00:09:37.720 --> 00:09:40.919 Maybe it can highlight areas that look statistically unusual. Yes, 198 00:09:41.519 --> 00:09:43.919 but a human doctor has to take that information, combine 199 00:09:43.919 --> 00:09:46.759 it with the patient's history, symptoms, other. 200 00:09:46.639 --> 00:09:50.279 Tests, and then make the actual diagnosis and treatment plan exactly. 201 00:09:50.440 --> 00:09:54.200 The human brings the context, the judgment, the ethical considerations. 202 00:09:54.679 --> 00:09:58.080 mL provides analysis, not understanding or decision making. In that 203 00:09:58.120 --> 00:10:01.480 brighter sense, it's the learning compon but it's miles away 204 00:10:01.480 --> 00:10:03.480 from the sensing AI of science fiction. 205 00:10:03.759 --> 00:10:06.320 Within mL, the big buzz seems to be around neural 206 00:10:06.360 --> 00:10:07.600 networks definitely. 207 00:10:08.080 --> 00:10:11.240 They're key to this recent AI renaissance, as people call it. 208 00:10:12.039 --> 00:10:16.200 Their computing systems inspired loosely by the structure of the brain, 209 00:10:17.039 --> 00:10:20.120 interconnected nodes or neurons, And how do. 210 00:10:20.120 --> 00:10:21.960 They actually learn it sounds complicated. 211 00:10:22.200 --> 00:10:26.200 The core mechanism for many is called backpropagation. It's basically 212 00:10:26.279 --> 00:10:29.559 a clever mathematical way to do trial and error. The 213 00:10:29.600 --> 00:10:33.039 network makes a guess based on the input. If the 214 00:10:33.080 --> 00:10:37.600 guess is wrong, backpropagation figures out how much each connection 215 00:10:37.759 --> 00:10:40.840 contributed to the error ah, and then it adjusts the 216 00:10:40.840 --> 00:10:43.879 weights or strengths of those connections slightly so next time 217 00:10:43.919 --> 00:10:45.840 it's hopefully a bit closer to the right answer. 218 00:10:45.919 --> 00:10:47.919 And that needs a lot of computing power, a. 219 00:10:47.840 --> 00:10:51.000 Lot, especially for large networks. That's where things like GPUs, 220 00:10:51.120 --> 00:10:53.879 those graphics chips come in. They're very good at the 221 00:10:53.919 --> 00:10:55.480 kind of parallel calculations needed. 222 00:10:55.600 --> 00:10:58.759 And deep learning. Is that just bigger neural networks or 223 00:10:58.799 --> 00:11:00.440 something fundamentally different. 224 00:11:00.720 --> 00:11:05.080 Well, that's debated. Generally, Yes, deep learning refers to neural 225 00:11:05.120 --> 00:11:08.960 networks with many layers deep layers, and they often run 226 00:11:09.000 --> 00:11:12.519 on even more powerful hardware. But our sources point out 227 00:11:12.519 --> 00:11:16.120 some issues. Public perception is often way ahead of reality. 228 00:11:16.639 --> 00:11:19.879 Even the experts developing these systems don't always fully understand 229 00:11:20.080 --> 00:11:24.639 why a specific deep network gives a particular answer. Can 230 00:11:24.679 --> 00:11:26.240 be a bit of a black box. 231 00:11:26.279 --> 00:11:29.960 And fundamentally it still doesn't understand anything correct. 232 00:11:30.639 --> 00:11:35.960 It's incredibly sophisticated pattern matching based on statistical correlations in 233 00:11:36.000 --> 00:11:39.159 the training data, it's not understanding concepts or meaning. 234 00:11:39.440 --> 00:11:41.840 So it can recognize picture of a cat because it's 235 00:11:41.879 --> 00:11:43.840 seen millions of cat pictures. 236 00:11:43.440 --> 00:11:45.919 And learn the pixel patterns that correlate strongly with the 237 00:11:46.000 --> 00:11:46.879 labeled cat. 238 00:11:46.919 --> 00:11:48.759 But it doesn't know what a cat is in any 239 00:11:48.799 --> 00:11:51.360 real sense exactly. What about some of these more advanced 240 00:11:51.440 --> 00:11:54.320 techniques like transfer learning that sounds useful. 241 00:11:54.080 --> 00:11:57.000 It is very transfer learning is a smart shortcut. You 242 00:11:57.039 --> 00:11:59.120 take a massive network that's already been trained on a 243 00:11:59.159 --> 00:12:03.200 general task like recognizing thousands of different objects and photos, 244 00:12:03.240 --> 00:12:06.120 and you reuse a large part of that network's learned 245 00:12:06.200 --> 00:12:11.360 knowledge for a new but related task, maybe identifying specific 246 00:12:11.360 --> 00:12:13.879 types of plants, using much less new data. 247 00:12:13.960 --> 00:12:16.960 So you sort of freeze the early layers that learn 248 00:12:17.039 --> 00:12:19.240 general features and just retrain the final layers. 249 00:12:19.279 --> 00:12:22.639 That's the basic idea. Yeah, it leverages the previous learning, 250 00:12:22.960 --> 00:12:25.360 saving a lot of time and data, like teaching it 251 00:12:25.399 --> 00:12:28.919 dogs and cats, then fine tuning it for macaroni versus cheese, 252 00:12:29.000 --> 00:12:31.279 as the source amusingly puts it, huh. 253 00:12:31.320 --> 00:12:34.279 And there are specialized networks too for different kinds of 254 00:12:34.360 --> 00:12:36.279 data like images versus text. 255 00:12:36.480 --> 00:12:41.480 Yes, Convolutional neural networks CNNs are stars. For image processing, 256 00:12:41.960 --> 00:12:45.200 they have specific structures like convolutional layers that are good 257 00:12:45.240 --> 00:12:49.759 at finding visual features edges, textures, shapes regardless of where 258 00:12:49.759 --> 00:12:50.639 they are in the image. 259 00:12:50.639 --> 00:12:54.080 And for sequences like words in a sentence or video frames, that's. 260 00:12:54.000 --> 00:12:56.720 Where recurrent neural networks are and NS come in. They 261 00:12:56.720 --> 00:12:58.759 have connections that loop back, giving them a sort of 262 00:12:58.799 --> 00:13:02.440 memory to consider previous elements in the sequence when processing 263 00:13:02.480 --> 00:13:05.840 the current one, crucial for language or time series data. 264 00:13:05.919 --> 00:13:08.200 So going back to the Turing test for a second, 265 00:13:08.879 --> 00:13:12.320 all this stuff mL, deep learning, CNNs, R and NS 266 00:13:12.519 --> 00:13:14.799 this is all still considered weak AI. 267 00:13:15.120 --> 00:13:18.080 Yes, absolutely, our sources are clear on that it's AI 268 00:13:18.159 --> 00:13:22.559 that can perform specific intelligent tasks, sometimes even better than humans. 269 00:13:23.200 --> 00:13:28.759 But it lacks consciousness, self awareness, genuine understanding, or consistent personality. 270 00:13:28.879 --> 00:13:31.799 So a strong AI, one that could truly pass for 271 00:13:31.960 --> 00:13:35.320 human in a deep conversation, would need much more. 272 00:13:35.559 --> 00:13:38.840 It would need to integrate context, maybe have consistent beliefs 273 00:13:38.919 --> 00:13:44.039 or personality, understand nuance, things that are way beyond current capabilities, 274 00:13:44.279 --> 00:13:45.519 still very much theoretical. 275 00:13:45.600 --> 00:13:48.159 Okay, let's shift gears a bit moving from the software 276 00:13:48.320 --> 00:13:51.519 the AI brains to the physical world. AI and robots 277 00:13:51.519 --> 00:13:52.559 are often lumped together. 278 00:13:52.639 --> 00:13:54.720 They are, but the source makes a really important distinction, 279 00:13:54.799 --> 00:13:57.879 which is AI is the software, the intelligence, the problem 280 00:13:57.960 --> 00:14:01.240 solving part. Robotics is the hardware, the physical machine that 281 00:14:01.279 --> 00:14:02.000 acts in the world. 282 00:14:02.080 --> 00:14:04.440 So AI can be the brain. Robotics is the body. 283 00:14:04.600 --> 00:14:06.120 That's a good way to put it. You can have 284 00:14:06.159 --> 00:14:08.480 AI without a robot body, and you can have robots 285 00:14:08.480 --> 00:14:10.440 a very simple or no AI. 286 00:14:10.360 --> 00:14:14.279 And robots themselves. They have a history way beyond modern AI. Right. 287 00:14:14.559 --> 00:14:16.080 Even the word robot. 288 00:14:15.759 --> 00:14:20.360 Oh yeah, the word comes from Zech robota, meaning forced labor. 289 00:14:20.720 --> 00:14:23.720 It was popularized by a play in the nineteen twenties. 290 00:14:23.480 --> 00:14:26.279 The one that introduced the idea of robots rising up. 291 00:14:26.399 --> 00:14:29.279 That's the one Carol Apex, Are you are but the 292 00:14:29.320 --> 00:14:33.279 idea of automated machines automata goes back way further, even 293 00:14:33.320 --> 00:14:38.360 into Greece machines following predetermined steps physical algorithms essentially. 294 00:14:38.120 --> 00:14:41.240 And the first industrial robot that was the Unimit back 295 00:14:41.279 --> 00:14:44.759 in nineteen sixty one, basically a programmable arm for doing 296 00:14:44.840 --> 00:14:47.440 dangerous jobs and factories like handl in hot metal. 297 00:14:47.840 --> 00:14:49.120 A long way from sci fi. 298 00:14:49.039 --> 00:14:52.799 Humanoids, and today industrial robots are still the biggest category 299 00:14:52.919 --> 00:14:53.360 by far. 300 00:14:53.480 --> 00:14:59.080 They're the backbone of modern manufacturing industry four point zero welding, painting, assembling, packaging, 301 00:14:59.440 --> 00:15:03.039 especially in tasks that are dangerous, repetitive, or need high precision. 302 00:15:03.240 --> 00:15:06.240 They're faster, often more accurate, and work twenty four to 303 00:15:06.240 --> 00:15:06.600 seven and. 304 00:15:06.559 --> 00:15:07.240 They're showing it more. 305 00:15:07.240 --> 00:15:11.600 In medicine too, definitely, robotic systems as cissurgeons allowing for 306 00:15:11.759 --> 00:15:16.720 much greater precision and minimally invasive procedures, and AI is 307 00:15:16.720 --> 00:15:20.799 also helping to make medical equipment smaller, smarter, easier to use. 308 00:15:21.480 --> 00:15:22.559 The source had that great. 309 00:15:22.360 --> 00:15:25.159 Example the heart disease diagnosis and Kenya. 310 00:15:24.919 --> 00:15:29.480 Exactly using AI to analyze data from relatively simple portable 311 00:15:29.480 --> 00:15:33.000 equipment to screen children for romatic heart disease. In places 312 00:15:33.200 --> 00:15:37.039 with limited access to specialists. That's a huge potential impact. 313 00:15:37.320 --> 00:15:39.840 But operating in the real physical world must be much 314 00:15:39.879 --> 00:15:43.159 harder than just processing data, right, What are the challenges 315 00:15:43.200 --> 00:15:44.919 for robots, even smart ones? 316 00:15:45.000 --> 00:15:49.679 Oh? Absolutely, the real world is messy, unpredictable. Robots have 317 00:15:49.720 --> 00:15:53.120 issues with latency delays in sensing or acting, timing problems, 318 00:15:53.120 --> 00:15:57.039 but the biggest challenge is probably environmental uncertainty. Things change, 319 00:15:57.080 --> 00:16:01.960 obstacles appear, people move unexpectedly, controlled lab environment not at all, 320 00:16:02.039 --> 00:16:05.480 And that's where AI's learning capability becomes really crucial, helping 321 00:16:05.559 --> 00:16:08.240 robots adapt on the fly respond to things they weren't 322 00:16:08.279 --> 00:16:12.200 explicitly programmed for dealing with other unpredictable agents like people 323 00:16:12.279 --> 00:16:15.240 or other robots, the multi agent problem is super challenging. 324 00:16:15.399 --> 00:16:18.639 What about specialty robots like drones? They seem to be everywhere. 325 00:16:18.720 --> 00:16:24.360 Drones, Yeah, UAVs. They're a huge area, initially for military surveillance, 326 00:16:24.759 --> 00:16:29.200 now used for everything from agriculture monitoring and delivery to filmmaking. 327 00:16:30.000 --> 00:16:33.039 AI is needed to give them more advanced abilities navigating 328 00:16:33.080 --> 00:16:38.200 indoors without GPS, identifying specific targets, coordinating actions, and. 329 00:16:38.159 --> 00:16:41.519 They're putting serious AI onto these small drones now they. 330 00:16:41.399 --> 00:16:44.399 Are developing nimble deep learning networks that can run in 331 00:16:44.399 --> 00:16:47.480 the limited processing power available on a small drone. That's 332 00:16:47.519 --> 00:16:48.759 a key research area. 333 00:16:48.799 --> 00:16:50.840 And coordinating them drones swarms. 334 00:16:50.919 --> 00:16:54.480 That's another big focus, especially for military applications, but also 335 00:16:54.519 --> 00:16:57.320 thinking ahead to future crowded skies with delivery drones and 336 00:16:57.320 --> 00:17:01.240 air taxis. MIT had an algorithm exam for preventing. 337 00:17:00.799 --> 00:17:04.279 Collisions, and regulation is playing catchup always. 338 00:17:03.840 --> 00:17:06.680 It seems, things like requiring operators to keep drones in 339 00:17:06.720 --> 00:17:10.000 their line of sight. That's partly because less sophisticated drones 340 00:17:10.039 --> 00:17:12.920 can become erratic or unpredictable if they lose their connection. 341 00:17:13.599 --> 00:17:15.559 Balancing safety and innovation is tough. 342 00:17:15.839 --> 00:17:19.119 The other big physical AI application we hear constantly about 343 00:17:19.200 --> 00:17:21.680 is self driving cars, the ultimate robot. 344 00:17:21.960 --> 00:17:25.200 Well, they're certainly seen as having a potentially massive impact 345 00:17:25.279 --> 00:17:29.400 on society, the economy, how cities are designed, but again 346 00:17:29.559 --> 00:17:34.559 our source stresses realism. What we have today are mostly prototypes, 347 00:17:34.759 --> 00:17:37.960 pilot projects in specific areas. We're not about to see 348 00:17:38.240 --> 00:17:41.440 fully autonomous cars handling all conditions everywhere overnight. 349 00:17:41.519 --> 00:17:44.240 It's going to be gradual, more assistant systems first. 350 00:17:44.160 --> 00:17:48.039 Exactly, a progressive introduction of automation. Humans will likely be 351 00:17:48.079 --> 00:17:50.920 involved overseeing or ready to take over for a long time. 352 00:17:51.880 --> 00:17:55.480 The main driver, initially at least, is safety, using AI 353 00:17:55.559 --> 00:17:57.920 to assist human drivers and prevent accidents. 354 00:17:58.559 --> 00:18:00.839 So what are the main parts of US driving system? 355 00:18:01.039 --> 00:18:01.640 How does it work? 356 00:18:01.680 --> 00:18:06.160 Basically, you need several core systems working together. First, perception 357 00:18:06.279 --> 00:18:08.680 and localization, figuring out what's around the car and exactly 358 00:18:08.720 --> 00:18:09.440 where the car. 359 00:18:09.319 --> 00:18:10.640 Is seeing the world. 360 00:18:10.359 --> 00:18:13.720 Seeing the world yeah. Then planning and decision making, predicting 361 00:18:13.799 --> 00:18:16.640 what other cars or pedestrians might do and planning the 362 00:18:16.680 --> 00:18:20.240 car's own path and actions the thinking part, the thinking part, 363 00:18:20.599 --> 00:18:27.240 and finally control an actuation. Actually executing the plan through steering, braking, accelerating. 364 00:18:26.839 --> 00:18:28.200 And redundancy is key, right. 365 00:18:28.240 --> 00:18:33.079 Having backups absolutely critical, Multiple types of sensors, multiple computing systems, 366 00:18:33.160 --> 00:18:37.519 all cross checking. The goal is extreme reliability, aiming for 367 00:18:37.680 --> 00:18:40.920 zero errors. Because the stakes are so high, even the 368 00:18:40.920 --> 00:18:43.480 best AI systems can sometimes be fooled or make mistakes, 369 00:18:43.480 --> 00:18:44.599 so backups are essential. 370 00:18:44.799 --> 00:18:47.359 What kind of sensors do they use to see. 371 00:18:47.039 --> 00:18:51.319 A whole suite? Cameras are obviously vital providing rich visual information. 372 00:18:52.039 --> 00:18:59.119 AI uses vision processing pattern matching to identify objects like cars, pedestrians, signs. 373 00:18:58.880 --> 00:19:01.440 But cameras struggle in bad weather or darkness. 374 00:19:01.559 --> 00:19:04.880 They do. That's why you also have radar uses radio 375 00:19:04.920 --> 00:19:07.920 waves bounces them off objects to determine distance and speed. 376 00:19:08.279 --> 00:19:09.599 It's much less affected by. 377 00:19:09.480 --> 00:19:11.039 Weather, but maybe less detailed. 378 00:19:11.119 --> 00:19:14.640 Generally, yes, lower resolution than cameras, and sometimes there's trouble 379 00:19:14.640 --> 00:19:16.119 with stationary objects. 380 00:19:15.720 --> 00:19:18.839 And light r that's the spinning thing you sometimes see often. 381 00:19:19.000 --> 00:19:22.519 Yeah, lighter uses laser pulses. It creates a very detailed 382 00:19:22.640 --> 00:19:25.920 three D map of the surroundings, great resolution, but it 383 00:19:25.920 --> 00:19:28.440 can also struggle in heavy rain or fog, and it's 384 00:19:28.480 --> 00:19:32.039 traditionally been quite expensive, though costs are coming down, so 385 00:19:32.119 --> 00:19:34.319 you use them all together fusing the data. 386 00:19:34.400 --> 00:19:37.680 Okay, let's shift the focus slightly again. How is AI 387 00:19:37.799 --> 00:19:40.799 interacting more directly with us with human capabilities and how 388 00:19:40.839 --> 00:19:41.920 we interact with each other. 389 00:19:42.279 --> 00:19:44.839 Well, one big area is just making us more efficient, right, 390 00:19:45.160 --> 00:19:47.839 taking over the boring, repetitive parts of. 391 00:19:47.880 --> 00:19:50.920 Jobs, freeing people up for more interesting stuff hopefully. 392 00:19:50.640 --> 00:19:53.039 Yeah, making work less tedious, more engaging. 393 00:19:53.079 --> 00:19:56.920 And AI can also simulate interaction like talking to Alexa 394 00:19:57.079 --> 00:19:57.880 or Google Home. 395 00:19:58.039 --> 00:20:00.960 It can and our source mentions this ease assistance can 396 00:20:01.000 --> 00:20:05.039 handle mundane tasks, fine info control smart devices, but it 397 00:20:05.079 --> 00:20:08.240 also touches on the idea that this simulated interaction might 398 00:20:08.279 --> 00:20:10.519 help some people feel less lonely or bored. 399 00:20:10.680 --> 00:20:15.119 Interesting. Can AI actually make us physically or mentally more 400 00:20:15.160 --> 00:20:17.240 capable enhance us? 401 00:20:17.400 --> 00:20:21.079