WEBVTT 1 00:00:01.199 --> 00:00:06.200 Welcome to the Sentient Code, where intelligence is engineered, autonomy 2 00:00:06.280 --> 00:00:10.439 is emerging, and a line between human and machine grows thinner. 3 00:00:10.800 --> 00:00:15.359 Each episode, we decode the algorithms, explore the robotics, and 4 00:00:15.439 --> 00:00:22.960 examine the ideas shaping the future of artificial minds. 5 00:00:23.879 --> 00:00:28.679 Welcome back today. We're looking at something that feels like 6 00:00:28.719 --> 00:00:30.440 it's right on top of us, like it's breathing down 7 00:00:30.480 --> 00:00:34.240 our necks, and yet nobody can seem to agree on 8 00:00:34.280 --> 00:00:35.560 what face is actually wearing. 9 00:00:35.799 --> 00:00:36.880 That's a good way to put it. 10 00:00:37.119 --> 00:00:39.679 You open your phone, you see the headlines. A chatbot 11 00:00:39.719 --> 00:00:43.159 past the bar, exam an algorithm one, an art competition, 12 00:00:43.560 --> 00:00:47.759 a program just folded proteins that baffled biologists for what 13 00:00:48.039 --> 00:00:48.759 fifty years. 14 00:00:49.000 --> 00:00:51.359 It feels like the ground is shifting under our feet. 15 00:00:51.399 --> 00:00:56.079 It's that sensation of you know, vertigo of progress. Things 16 00:00:56.079 --> 00:00:58.640 that were pure science fiction just five or ten years 17 00:00:58.640 --> 00:01:00.600 ago are now utilities. 18 00:01:00.600 --> 00:01:04.079 They're mundane precisely. But here is the friction point, and 19 00:01:04.120 --> 00:01:06.439 that's why we're doing this analysis today. You talk to 20 00:01:06.480 --> 00:01:09.079 a software engineer and they'll roll their eyes and say, look, 21 00:01:09.159 --> 00:01:11.480 it's just a large language model. It's predicting the next word. 22 00:01:11.519 --> 00:01:14.000 It's a very clever parlor trick with statistics. 23 00:01:14.120 --> 00:01:17.480 Sure, the stochastic parrot argument, right, But then you talk 24 00:01:17.519 --> 00:01:21.000 to a philosopher or a theoretical physicist or an AI 25 00:01:21.159 --> 00:01:24.920 safety researcher and they are buying bunkers in New Zealand. 26 00:01:25.200 --> 00:01:28.239 The disconnect is massive, and it really stems from a 27 00:01:28.239 --> 00:01:30.920 confusion of terms. You know, we use this one word 28 00:01:31.079 --> 00:01:34.719 AI to describe everything from the spell check on your 29 00:01:34.760 --> 00:01:38.920 phone to a hypothetical godlike mind that could rewrite physics. 30 00:01:39.200 --> 00:01:41.599 So today we are stopping the scroll. We're going to 31 00:01:41.599 --> 00:01:46.120 tackle a really comprehensive piece of research titled the AGI Horizon, 32 00:01:46.519 --> 00:01:49.680 defining the ultimate goal of AI research. Okay, we want 33 00:01:49.719 --> 00:01:51.760 to move past the hype of the tools we have 34 00:01:51.879 --> 00:01:55.560 now to talk about the destination. We're talking about AGI, 35 00:01:56.120 --> 00:01:59.239 Artificial general intelligence, the big one, the big one, the 36 00:01:59.239 --> 00:02:02.560 Holy Grail. And what's so fascinating about this source material 37 00:02:02.959 --> 00:02:05.480 is that it frames AGI not just as you know, 38 00:02:05.560 --> 00:02:09.759 better software, but as potentially the last invention humanity will 39 00:02:09.800 --> 00:02:10.680 ever need to create. 40 00:02:10.879 --> 00:02:13.240 That is the line that always stops me cold. The 41 00:02:13.360 --> 00:02:15.960 last invention. Yeah, it implies that once you build a 42 00:02:16.000 --> 00:02:19.000 machine that can actually think, it becomes the inventor. It 43 00:02:19.039 --> 00:02:20.240 takes the baton from us. 44 00:02:20.520 --> 00:02:23.879 So let's peel this back to understand what AGI is. 45 00:02:23.919 --> 00:02:26.400 We have to be really, really clear about what the 46 00:02:26.439 --> 00:02:29.360 impressive stuff we have today is not, because I think 47 00:02:29.400 --> 00:02:32.360 most people, myself included half the time, look at GPT 48 00:02:32.520 --> 00:02:35.159 four or mid Journey and think, well, isn't this it. 49 00:02:35.159 --> 00:02:39.000 It's writing poetry, it's coding. Isn't that general intelligence? 50 00:02:39.199 --> 00:02:42.240 And it feels like it. I mean, it's very convincing. 51 00:02:42.639 --> 00:02:46.240 But the Source classifies all current systems, even the most 52 00:02:46.280 --> 00:02:50.719 impressive ones, as narrow AI zuroai or weak AI, though 53 00:02:50.759 --> 00:02:53.520 I really hate that term because these systems are incredibly powerful. 54 00:02:53.560 --> 00:02:56.719 You know they're not weak, but the distinction is all 55 00:02:56.759 --> 00:03:01.639 about scope and the underlying architecture of how they learn. 56 00:03:01.840 --> 00:03:05.639 Let's drill into that narrow It implies a lane, a 57 00:03:05.680 --> 00:03:06.319 single lane. 58 00:03:06.400 --> 00:03:09.919 Think of it as a manifold, a specific high dimensional 59 00:03:09.960 --> 00:03:13.439 shape of data. Take a chessbot like Stockfish or even 60 00:03:13.479 --> 00:03:16.319 the earlier AlphaGo versions. These are genuses. They will crush 61 00:03:16.400 --> 00:03:18.639 any human who's every lived at chess, no question, but 62 00:03:18.680 --> 00:03:21.879 they exist strictly within the universe of those sixty four squares. 63 00:03:22.000 --> 00:03:24.280 So if I asked that chessbot to play checkers, which 64 00:03:24.319 --> 00:03:27.280 is a much much simpler game. Yeah, it can't do it. 65 00:03:26.840 --> 00:03:28.960 It's worse than that. It doesn't even know what a 66 00:03:29.000 --> 00:03:32.919 game is. It doesn't know what winning implies outside of 67 00:03:32.960 --> 00:03:37.080 a mathematical variable in its own specific code. It's just 68 00:03:37.159 --> 00:03:41.560 calculating probabilities within a completely closed system. It's a calculator. 69 00:03:41.960 --> 00:03:44.360 I mean, a calculator can compute the trajectory of a 70 00:03:44.439 --> 00:03:46.800 rocket to Mars, but it can't tell you if it's 71 00:03:46.840 --> 00:03:50.919 raining outside. It has no sensorium, no context, no ability 72 00:03:50.960 --> 00:03:53.919 to step off the specifically paved road it was built on. 73 00:03:54.240 --> 00:03:58.439 Okay, but chess is rigid, it's all rules. Language feels 74 00:03:58.479 --> 00:04:02.280 so fluid when I talk to a chatbot. It feels 75 00:04:02.280 --> 00:04:04.840 like it's improvising. It feels like it understands context. 76 00:04:05.000 --> 00:04:09.080 It's an incredibly convincing illusion. And the source material argues 77 00:04:09.120 --> 00:04:12.439 that even large language models are essentially narrow because they 78 00:04:12.439 --> 00:04:14.800 are trapped in the domain of text prediction. 79 00:04:15.120 --> 00:04:17.959 Right. They're just guessing the next most likely word exactly. 80 00:04:18.360 --> 00:04:21.240 They're trained on a static snapshot of the Internet. They 81 00:04:21.240 --> 00:04:23.519 don't learn in real time. If you tell a joke 82 00:04:23.600 --> 00:04:25.680 to a model that wasn't in its training data. It 83 00:04:25.759 --> 00:04:28.399 might get it because it's seen millions of similar jokes, 84 00:04:28.600 --> 00:04:31.920 but it's not deriving humor from first principles. It's just 85 00:04:32.040 --> 00:04:34.279 pattern matching on a cosmic scale. 86 00:04:34.399 --> 00:04:37.160 And this leads to what you call the transfer learning problem. 87 00:04:37.240 --> 00:04:40.160 This seems to be the technical wall. Yeah, that separates 88 00:04:40.839 --> 00:04:43.439 you know, the boys from the men, or the chatbots 89 00:04:43.439 --> 00:04:44.199 from the agi. 90 00:04:44.560 --> 00:04:47.240 This is the absolute crux of the definition. In the 91 00:04:47.240 --> 00:04:51.399 biological world. In US, learning is sticky, it's transferable. If 92 00:04:51.399 --> 00:04:53.000 I teach you how to open a door with a 93 00:04:53.079 --> 00:04:55.519 round knob and then you encounter a door with a 94 00:04:55.560 --> 00:04:58.120 lever handle, you don't just freeze up right. 95 00:04:58.279 --> 00:05:00.560 I look at it. I understand leverage from you know, 96 00:05:00.720 --> 00:05:03.720 physics class or just life. I understand doorness, and I 97 00:05:03.720 --> 00:05:04.680 figure it down in a second. 98 00:05:04.800 --> 00:05:08.839 You transfer the skill you apply an abstract concept opening 99 00:05:08.839 --> 00:05:13.839 a barrier to a novel situation. Narrow AI fails castrophically 100 00:05:13.839 --> 00:05:16.439 at this. If you train a vision AI on a 101 00:05:16.480 --> 00:05:19.759 million pictures of cats, it becomes a god at spotting cats. 102 00:05:20.319 --> 00:05:22.399 It can see a cat ear behind a sofa in 103 00:05:22.439 --> 00:05:24.839 a pitch black room, but show it a dog. 104 00:05:25.160 --> 00:05:28.199 It doesn't say, huh, that's interesting. Similar shape Fore legs 105 00:05:28.240 --> 00:05:29.879 for it's probably an animal of some kind. 106 00:05:30.000 --> 00:05:33.399 No, to the AI, that dog is just noise. It's 107 00:05:33.439 --> 00:05:36.160 a statistical anomaly. It's out of distribution. You have to 108 00:05:36.160 --> 00:05:38.959 start completely from scratch. You need a million pictures of 109 00:05:38.959 --> 00:05:40.199 dogs to build a whole new model. 110 00:05:40.319 --> 00:05:42.879 So it doesn't understand the concept of an animal. 111 00:05:42.720 --> 00:05:46.079 Not at all. It just understands the statistical distribution of 112 00:05:46.120 --> 00:05:49.920 pixels that humans have labeled cat. It has zero semantic 113 00:05:50.000 --> 00:05:52.879 understanding of the world. It's all syntax, no semantics. 114 00:05:53.120 --> 00:05:56.680 So AGI is the bridge. AGI is the system that 115 00:05:56.680 --> 00:05:59.040 looks at the doorknob and the lever and sees the 116 00:05:59.120 --> 00:06:00.680 underlying principle exactly. 117 00:06:01.279 --> 00:06:06.199 The source defines AGI by three main pillars autonomy, creativity, 118 00:06:06.319 --> 00:06:09.199 and versatility. It needs to be able to set its 119 00:06:09.240 --> 00:06:12.240 own sub goals to achieve a larger goal. It needs 120 00:06:12.279 --> 00:06:16.079 to reason about abstract principles, not just match patterns, and 121 00:06:16.120 --> 00:06:18.759 it needs to move fluidly between different domains. 122 00:06:19.079 --> 00:06:22.439 The source material uses the student analogy, which I thought 123 00:06:22.480 --> 00:06:23.279 was really effective. 124 00:06:23.319 --> 00:06:25.839 It's perfect, isn't it. Imagine a human student. They go 125 00:06:25.879 --> 00:06:28.800 to a university, they take a class in nineteenth century literature. 126 00:06:29.079 --> 00:06:31.120 Then they go to a physics lab and do an experiment. 127 00:06:31.720 --> 00:06:34.360 Then they have to navigate the complex social dynamics of 128 00:06:34.360 --> 00:06:37.000 the cafeteria at lunch. Then they go back to the 129 00:06:37.040 --> 00:06:38.720 dorm and have to figure out how to use a 130 00:06:38.759 --> 00:06:40.839 new washing machine they've never seen before. 131 00:06:41.079 --> 00:06:43.000 And they're using one single brain for all. 132 00:06:42.879 --> 00:06:46.160 Of that one brain, and they're connecting them. They might 133 00:06:46.240 --> 00:06:48.360 use a physics metaphor from the lab to explain a 134 00:06:48.360 --> 00:06:49.839 plot point in the book they're reading. 135 00:06:49.920 --> 00:06:52.680 That cross pollination. That's the spark of real intelligence. 136 00:06:52.720 --> 00:06:56.720 That is general intelligence. It's cognitive flexibility. So when AGI 137 00:06:56.959 --> 00:07:00.319 isn't just a bot that is good at everything because 138 00:07:00.319 --> 00:07:03.199 it was trained on a million different things separately. It's 139 00:07:03.240 --> 00:07:07.079 a system that can face a completely novel situation, something 140 00:07:07.079 --> 00:07:09.879 that has never seen before, and figure it out from 141 00:07:09.959 --> 00:07:13.839 Perst principles using logic and dare I say intuition? 142 00:07:14.040 --> 00:07:16.240 Okay, so that's the definition. But I want to play 143 00:07:16.240 --> 00:07:18.720 Devil's advocate here for a second, because if I'm a listener, 144 00:07:18.720 --> 00:07:20.759 I'm sitting here thinking, okay, but how do we know 145 00:07:21.399 --> 00:07:24.639 if I'm chatting with a really sophisticated AI and it 146 00:07:24.680 --> 00:07:27.839 gives me a brilliant, creative answer, how do I prove 147 00:07:27.879 --> 00:07:30.639 it's not thinking This brings us to what the source 148 00:07:30.680 --> 00:07:32.439 calls the testing crisis. 149 00:07:32.720 --> 00:07:35.319 It's a huge problem. For seventy years, we relied on 150 00:07:35.360 --> 00:07:39.040 the Turing test Alan Turing's imitation game. The premise was 151 00:07:39.120 --> 00:07:41.959 beautifully simple. If a machine can chat with you for 152 00:07:42.000 --> 00:07:43.680 five minutes and you can't tell for sure if it's 153 00:07:43.720 --> 00:07:45.560 a machine or a human, then it's intelligent. 154 00:07:45.879 --> 00:07:48.759 And arguably we are there. I mean, I've had customer 155 00:07:48.800 --> 00:07:51.040 service chats online where I honestly wasn't sure. 156 00:07:51.240 --> 00:07:54.360 We have absolutely beaten it. But the source argues, we 157 00:07:54.480 --> 00:07:58.240 beat it by cheating. We built machines that are incredibly 158 00:07:58.240 --> 00:08:01.959 good at mimicking human speech pattern They are stochastic parrots. 159 00:08:02.079 --> 00:08:03.759 To borrow a freeze from the literature. 160 00:08:03.800 --> 00:08:06.000 You just pair it back what they've heard exactly. 161 00:08:06.439 --> 00:08:09.720 The Turing test, it turns out, measures human gullibility as 162 00:08:09.800 --> 00:08:12.959 much as it measures machine intelligence. It tests the ability 163 00:08:13.000 --> 00:08:15.399 to deceive, not the ability to think. 164 00:08:15.680 --> 00:08:19.759 So it's a test of surface level charisma, not deep cognition. 165 00:08:20.560 --> 00:08:23.439 We need a better ruler. What does the source suggest? 166 00:08:23.560 --> 00:08:27.399 They propose a series of behavioral challenges. These are tests 167 00:08:27.439 --> 00:08:32.039 that require interacting with the physical, messy, unstructured world. My 168 00:08:32.159 --> 00:08:34.960 personal favorite and the one that really highlights the gap 169 00:08:35.000 --> 00:08:38.679 between current AI and AGI is the coffee test. 170 00:08:38.919 --> 00:08:41.159 I love the simplicity of this. It sounds so mundane, 171 00:08:41.240 --> 00:08:42.879 so easy, walk us through it. 172 00:08:42.879 --> 00:08:45.679 It was actually proposed by Steve Wozniak. You take a robot, 173 00:08:45.960 --> 00:08:48.200 you drop it into a random American home, a house 174 00:08:48.240 --> 00:08:50.240 it has never seen before. You don't give it any 175 00:08:50.279 --> 00:08:53.120 floor plans, no preprogramming about where things are. 176 00:08:53.360 --> 00:08:55.639 Okay, you just tell it one thing, Go make a 177 00:08:55.639 --> 00:08:56.960 cup of coffee. That's it. 178 00:08:57.240 --> 00:08:59.720 That sounds incredibly easy. I could walk into your house 179 00:08:59.759 --> 00:09:01.320 right now, you know, never having been there, and I'd 180 00:09:01.320 --> 00:09:03.080 have a fresh cup of coffee in five minutes. 181 00:09:03.320 --> 00:09:06.360 But now think about the computational complexity of what you 182 00:09:06.559 --> 00:09:11.519 just described. Your brain does it effortlessly. First, you have 183 00:09:11.600 --> 00:09:14.960 to navigate a three D space without bumping into furniture. Sure, 184 00:09:15.000 --> 00:09:17.519 you have to identify the kitchen. What makes a room 185 00:09:17.519 --> 00:09:21.879 a kitchen the presence of a sink, a stove, a refrigerator. 186 00:09:22.159 --> 00:09:25.000 Then you have to search cupboards and drawers. You have 187 00:09:25.080 --> 00:09:28.480 to identify the coffee machine itself. Is it a currig, 188 00:09:29.120 --> 00:09:33.200 a French press, an espresso machine, a drip brewer. 189 00:09:32.960 --> 00:09:34.480 And they all work completely differently. 190 00:09:34.720 --> 00:09:37.240 Radically differently. You have to figure out the user interface. 191 00:09:37.600 --> 00:09:39.399 Then you need to find the coffee beans. You need 192 00:09:39.440 --> 00:09:42.600 to find a grinder, a source of water, a mug. 193 00:09:43.000 --> 00:09:45.519 What if the coffee bag is new and sealed, you 194 00:09:45.600 --> 00:09:47.600 have to recognize that and then find scissors. 195 00:09:47.639 --> 00:09:49.960 What if a mug is dirty, but to wash it? 196 00:09:50.000 --> 00:09:55.320 This requires common sense, visual recognition, physical manipulation, causal reasoning, 197 00:09:55.519 --> 00:09:59.679 and problem solving, all happening in a chaotic, unpredictable environment. 198 00:10:00.279 --> 00:10:03.559 This touch is on morvex paradox. Right, This feels like 199 00:10:03.559 --> 00:10:05.000 a perfect illustration of it. 200 00:10:05.000 --> 00:10:08.120 It absolutely is. It's a key discovery in AI research 201 00:10:08.159 --> 00:10:12.159 that basically says high level reasoning requires very little computation, 202 00:10:12.759 --> 00:10:17.559 but low level sensor motor skills require enormous computational resources. 203 00:10:17.039 --> 00:10:18.600 Which is completely counterintuitive. 204 00:10:18.720 --> 00:10:21.759 Totally. It is relatively easy to build an AI that 205 00:10:21.799 --> 00:10:24.519 can beat a grand master at chess or calculate the 206 00:10:24.519 --> 00:10:28.320 digits of PI. It is incredibly, incredibly hard to build 207 00:10:28.360 --> 00:10:30.799 a robot that can fold laundry as well as a 208 00:10:30.799 --> 00:10:31.600 six year old. 209 00:10:31.440 --> 00:10:33.840 Child, because chess is just math at the end of 210 00:10:33.879 --> 00:10:38.000 the day. Yeah, laundry is physics and chaos, and you know, real. 211 00:10:37.799 --> 00:10:41.080 Life exactly the coffee test proves you can handle chaos. 212 00:10:41.159 --> 00:10:43.519 If a machine can walk into any house and make coffee, 213 00:10:43.639 --> 00:10:47.559 it possesses general adaptability. It understands the world, not just 214 00:10:47.600 --> 00:10:48.279 a data set. 215 00:10:48.440 --> 00:10:50.639 There's another distinction that Source makes that I found really 216 00:10:50.639 --> 00:10:54.240 helpful in this section, the difference between intelligence and capability. 217 00:10:54.279 --> 00:10:56.080 I think we conflate them all the time. We assume 218 00:10:56.120 --> 00:10:57.759 smart things are powerful things. 219 00:10:57.759 --> 00:11:00.600 We do, but they are different variables on the graph. 220 00:11:00.639 --> 00:11:04.919 They're two separate axes. The Source uses a really striking analogy, 221 00:11:05.399 --> 00:11:08.960 the genius in a wheelchair versus the factory arm. 222 00:11:09.399 --> 00:11:10.120 Let's unpack that. 223 00:11:10.360 --> 00:11:13.120 Okay, so you could have a superintelligence running on a 224 00:11:13.159 --> 00:11:18.000 server somewhere. It's air gapped, no Internet connection, no robotic body. 225 00:11:18.559 --> 00:11:20.759 It might know the cure for cancer, it might have 226 00:11:20.840 --> 00:11:24.279 deduced the grand unified theory of physics, but it has 227 00:11:24.440 --> 00:11:28.200 zero capability to act on that knowledge. You can't mix chemicals, 228 00:11:28.279 --> 00:11:30.960 it can't publish the paper, it can't even send an email. 229 00:11:31.000 --> 00:11:35.679 It's pure inert mind, high intelligence, zero capability. 230 00:11:35.720 --> 00:11:37.840 And on the other side, the factory are which. 231 00:11:37.600 --> 00:11:41.120 Has enormous physical capability. It can crush a car, it 232 00:11:41.159 --> 00:11:44.519 can weld a seam with submillimeter precision, but it has 233 00:11:44.600 --> 00:11:47.720 zero intelligence. It's just following a pre programmed script. It's 234 00:11:47.720 --> 00:11:48.200 a puppet. 235 00:11:48.240 --> 00:11:50.440 So agi is when those two lines on the graph 236 00:11:50.480 --> 00:11:51.879 intersect and go way up. 237 00:11:52.000 --> 00:11:55.679 That's it high intelligence combined with high capability to execute 238 00:11:55.679 --> 00:11:56.879 and effect the physical world. 239 00:11:56.960 --> 00:12:00.279 And that that is where the risk profile starts to like, 240 00:12:00.720 --> 00:12:02.799 because an intelligent agent that can act in the world, 241 00:12:03.240 --> 00:12:04.200 that's a new species. 242 00:12:04.519 --> 00:12:07.159 Effectively, it is a new kind of actor on the 243 00:12:07.159 --> 00:12:07.960 world stage. 244 00:12:08.240 --> 00:12:10.279 So we know what it is, at least in theory. 245 00:12:10.360 --> 00:12:13.759 We know how we test for it. The billion dollar question, literally, 246 00:12:13.799 --> 00:12:16.799 the trillion dollar question is how do we build it? 247 00:12:17.440 --> 00:12:18.279 And when is it coming. 248 00:12:18.679 --> 00:12:21.399 This is where the scientific community just fractures. I mean, 249 00:12:21.440 --> 00:12:24.799 there isn't one path up the mountain. There are competing 250 00:12:24.960 --> 00:12:28.279 tribes of AI research, all with their own philosophies. 251 00:12:28.360 --> 00:12:30.279 The one getting all the attention right now, the one 252 00:12:30.399 --> 00:12:33.039 driving the stock market, is deep learning. 253 00:12:32.799 --> 00:12:36.360 And scaling, right the scaling hypothesis. This is the brute 254 00:12:36.360 --> 00:12:40.720 force philosophy. The idea is remarkably simple, almost deceptively so 255 00:12:41.399 --> 00:12:44.320 we don't need to program complex rules about logic or 256 00:12:44.320 --> 00:12:47.600 the world. We just need bigger neural networks, more data 257 00:12:47.960 --> 00:12:49.639 and more computing chips just make. 258 00:12:49.559 --> 00:12:51.200 The brain bigger and feed it more books. 259 00:12:51.320 --> 00:12:54.480 Essentially, the proponents of this view look at the jump 260 00:12:54.559 --> 00:12:57.080 from GPT two to GPT three to GPT four and 261 00:12:57.120 --> 00:12:59.720 they say, look, every time we scale it up, every 262 00:12:59.720 --> 00:13:02.320 time we add more parameters and feed it more tokens, 263 00:13:02.840 --> 00:13:05.679 new unexpected capabilities emerged. 264 00:13:05.799 --> 00:13:06.919 Save per they just appear. 265 00:13:07.600 --> 00:13:11.879 GPT two could barely write a coherent sentence. GPT four 266 00:13:12.039 --> 00:13:15.279 pass the bar exam. We didn't explicitly program it to 267 00:13:15.320 --> 00:13:17.519 take the bar exam. We just made the model bigger 268 00:13:17.759 --> 00:13:18.840 and fed it the Internet. 269 00:13:18.919 --> 00:13:21.000 Its concept of emerging properties, right. 270 00:13:20.919 --> 00:13:23.919 It's like a pile of sand. One grain is nothing, 271 00:13:24.440 --> 00:13:27.759 A million grains is a pile. A billion grains might 272 00:13:27.799 --> 00:13:31.279 suddenly behave like a liquid and an avalanche. The scaling 273 00:13:31.320 --> 00:13:33.720 tribe believes that if we just keep stacking the chips 274 00:13:33.799 --> 00:13:38.080 higher and higher, agi will naturally emerge from the sheer complexity. 275 00:13:38.240 --> 00:13:41.360 But not everyone buys. That is a pretty strong counter 276 00:13:41.480 --> 00:13:42.759 argument about hitting a data wall. 277 00:13:42.960 --> 00:13:45.960 Yes, and this is a very practical problem. We are 278 00:13:46.120 --> 00:13:49.799 running out of Internet high quality human generated text is 279 00:13:49.840 --> 00:13:53.320 a finite resource. We've already fed these models. Basically all 280 00:13:53.320 --> 00:13:56.200 of Wikipedia read it all the digitized books, all the 281 00:13:56.200 --> 00:13:57.200 scientific papers. 282 00:13:57.320 --> 00:13:58.840 We're running out of stuff for it to read. 283 00:13:58.960 --> 00:14:01.519 Some researchers argue that once we hit that ceiling, the 284 00:14:01.600 --> 00:14:05.039 progress just stops, or at least slows down dramatically. You 285 00:14:05.120 --> 00:14:07.720 can't learn if there's nothing left to learn from. 286 00:14:07.440 --> 00:14:10.279 Unless it starts generating its own data to learn from. 287 00:14:10.840 --> 00:14:13.039 But let's put a pin in that. That sounds dangerous. 288 00:14:13.559 --> 00:14:14.679 What are the other approaches? 289 00:14:15.480 --> 00:14:19.159 So you have the neuroscience inspired camp. They look at 290 00:14:19.159 --> 00:14:22.279 the scaling approach and say, you're just building a bigger 291 00:14:22.320 --> 00:14:25.919 statistical parrot, not a mind. They want to reverse engineer 292 00:14:25.960 --> 00:14:29.639 the biological brain, copy the blueprint, try it. Yeah, They 293 00:14:29.679 --> 00:14:32.919 want to mimic the actual structure of neurons and synapses, 294 00:14:33.159 --> 00:14:36.919 trying to capture the incredible efficiency and plasticity of biology. 295 00:14:37.360 --> 00:14:41.080 Our brains run on what twenty watts of power, about 296 00:14:41.080 --> 00:14:44.159 the same as a dim light bulb. The supercomputers training 297 00:14:44.200 --> 00:14:47.559 these large models consume the power of a small city. 298 00:14:47.919 --> 00:14:49.159 That's a staggering difference. 299 00:14:49.279 --> 00:14:52.519 It tells you we're missing something fundamental about how biology computes. 300 00:14:52.679 --> 00:14:56.399 And then there is embodied AI. This will makes so 301 00:14:56.480 --> 00:14:58.360 much intuitive sense to me. It links right back to 302 00:14:58.399 --> 00:14:59.120 the coffee test. 303 00:14:59.320 --> 00:15:02.600 It's the ground problem. If an AI only knows the 304 00:15:02.600 --> 00:15:07.120 word apple by its statistical relationship to other words like fruit, red, 305 00:15:07.240 --> 00:15:09.840 and tree, does it really know what an apple is? 306 00:15:10.080 --> 00:15:10.799 No, of course not. 307 00:15:11.240 --> 00:15:15.879 Embodied AI researchers say no, absolutely not. They say intelligence 308 00:15:15.960 --> 00:15:18.320 must be forged in the physical world. You have to 309 00:15:18.360 --> 00:15:20.960 drop the spoon to really learn about gravity, you have 310 00:15:21.039 --> 00:15:24.000 to feel the resistance of an object to understand physics. 311 00:15:24.360 --> 00:15:27.399 They argue that an AI trapped in a server rack 312 00:15:27.759 --> 00:15:30.799 can never be truly intelligent because it doesn't live anywhere. 313 00:15:30.840 --> 00:15:32.360 It's not grounded in reality. 314 00:15:32.519 --> 00:15:36.399 So with all these different competing approaches, surely someone has 315 00:15:36.440 --> 00:15:38.279 a good guess as to when this is all going 316 00:15:38.360 --> 00:15:38.720 to happen. 317 00:15:38.759 --> 00:15:40.679 If you want to start a fight at an AI conference, 318 00:15:40.840 --> 00:15:44.200 just ask about timelines. The disagreement is it's massive. 319 00:15:44.519 --> 00:15:46.960 The source mentioned a survey from twenty twenty two, before 320 00:15:46.960 --> 00:15:47.720 the latest boom. 321 00:15:47.799 --> 00:15:51.120 Yes, and the median estimate for AGI arrival among researchers 322 00:15:51.159 --> 00:15:54.519 then was around twenty sixty. But since GPT four came out, 323 00:15:54.679 --> 00:15:58.480 those prediction markets and expert surveys have shifted wildly. You 324 00:15:58.559 --> 00:16:01.919 have serious, credible experts, not just hype artists now saying 325 00:16:01.960 --> 00:16:03.919 things like twenty twenty seven or twenty. 326 00:16:03.799 --> 00:16:06.440 Thirty eight, that's terrifyingly soon. That is, my current car 327 00:16:06.480 --> 00:16:07.720 will still be on the road soon. 328 00:16:08.000 --> 00:16:10.759 But then you also have the skeptics, people like Yan 329 00:16:10.840 --> 00:16:13.879 Lacun who's a Titan in the field, who say we 330 00:16:14.000 --> 00:16:18.159 are missing fundamental breakthroughs. They'll tell you we are decades, 331 00:16:18.279 --> 00:16:19.639 maybe many decades away. 332 00:16:19.960 --> 00:16:22.080 Why is it so hard to predict? I mean, we 333 00:16:22.200 --> 00:16:24.960 usually have a better handle on forecasting technology than this. 334 00:16:25.399 --> 00:16:27.240 We knew the moon landing was coming a few years 335 00:16:27.240 --> 00:16:28.039 before it happened. 336 00:16:28.120 --> 00:16:30.879 Because it's what the source calls the time scale problem, 337 00:16:30.960 --> 00:16:33.519 or what I like to call the difficulty switch. We 338 00:16:33.639 --> 00:16:36.519 just don't know what difficulty setting the universe has put 339 00:16:36.639 --> 00:16:37.879 on the problem of AGI. 340 00:16:38.159 --> 00:16:41.600 Okay, let's unpack that. What are the different difficulty levels? 341 00:16:42.039 --> 00:16:47.039 Imagine three scenarios. Scenario one, the problem is easy. This 342 00:16:47.120 --> 00:16:50.039 means the scaling hypothesis is correct. We just need to 343 00:16:50.080 --> 00:16:52.919 scale up what we already have. We're data more compute. 344 00:16:53.000 --> 00:16:55.679 If that's true, then AGI is coming very very soon, 345 00:16:56.039 --> 00:16:57.559 maybe in the next three to five years. 346 00:16:57.720 --> 00:16:58.120 Wow. 347 00:16:58.200 --> 00:17:02.399 Scenario two, it's medium. Scaling helps, but it hits a wall. 348 00:17:02.720 --> 00:17:05.640 We need a few new conceptual breakthroughs, maybe in reasoning 349 00:17:05.759 --> 00:17:08.799 or memory or understanding cause and effect. That means we 350 00:17:08.880 --> 00:17:11.799 have to do real science, not just massive engineering that 351 00:17:11.839 --> 00:17:14.799 probably puts us decades away, and hard mode. Hard mode 352 00:17:14.839 --> 00:17:20.000 means we are missing something truly fundamental. Maybe intelligence requires 353 00:17:20.000 --> 00:17:23.519 solving the mysteries of consciousness. Maybe it's tied to quantum 354 00:17:23.519 --> 00:17:26.240 physics in the brain. If that's the case, it could 355 00:17:26.240 --> 00:17:29.400 be centuries. It might even be impossible for us, And 356 00:17:29.440 --> 00:17:31.920 the problem is looking at the progress or making today, 357 00:17:32.279 --> 00:17:35.039 we can't tell if we're solving the core puzzle or 358 00:17:35.119 --> 00:17:37.119 just picking all the low hanging fruit first. 359 00:17:37.240 --> 00:17:41.279 That uncertainty is what makes policy making and regulation almost impossible, 360 00:17:41.839 --> 00:17:44.319 because if it's easy, we might not be ready for 361 00:17:44.359 --> 00:17:47.000 the consequences. And that leads us directly to the concept 362 00:17:47.200 --> 00:17:49.039 of the explosion. 363 00:17:48.799 --> 00:17:51.880 The intelligence explosion, or the singularity. 364 00:17:52.000 --> 00:17:53.920 This is the part of the source material that feels 365 00:17:54.000 --> 00:17:56.319 straight out of a sci fi movie, but the logic 366 00:17:56.359 --> 00:18:00.839 behind it is surprisingly simple and sound. It's all about 367 00:18:00.839 --> 00:18:02.000 recursive self improvement. 368 00:18:02.200 --> 00:18:04.599 This is the critical feedback loop, and to get your 369 00:18:04.599 --> 00:18:06.799 head around it, you have to realize that writing computer 370 00:18:06.839 --> 00:18:10.599 code is an intellectual task. Currently humans write the code 371 00:18:10.599 --> 00:18:13.039 for AI. But imagine you build an AI that is 372 00:18:13.119 --> 00:18:15.519 smart enough to code. We have that now to a 373 00:18:15.519 --> 00:18:18.359 certain extent. But now imagine an AI that is smart 374 00:18:18.440 --> 00:18:20.200 enough to understand its own architecture. 375 00:18:20.519 --> 00:18:22.440 It can look under its own hood and tinker with 376 00:18:22.480 --> 00:18:23.599 the engine precisely. 377 00:18:23.759 --> 00:18:26.160 It looks at its own source code and says, huh, 378 00:18:26.400 --> 00:18:28.720 I can make this more efficient. I can optimize this 379 00:18:28.839 --> 00:18:31.960 learning algorithm. So it rewrites a part of itself. 380 00:18:32.079 --> 00:18:34.839 So version one point zero writes version one point one, and. 381 00:18:34.839 --> 00:18:37.359 Version one point one is now smarter than version one 382 00:18:37.400 --> 00:18:39.759 point oh onie, so it is better at rewriting code 383 00:18:39.799 --> 00:18:42.960 than its predecessor. So version one point two arrives even 384 00:18:43.119 --> 00:18:45.440 faster and is even smarter still. 385 00:18:45.519 --> 00:18:47.799 It's like compounding interest, but for intelligence. 386 00:18:47.920 --> 00:18:51.079 That's the perfect analogy, and the time between these improvements 387 00:18:51.160 --> 00:18:53.359 gets shorter and shorter. Version one takes a year to 388 00:18:53.400 --> 00:18:55.920 design version two. Version two takes a month to design 389 00:18:55.960 --> 00:18:58.440 version three. Version three takes an hour to design, Version four. 390 00:18:58.599 --> 00:19:01.359 Version four takes a second. This is the singularity. The 391 00:19:01.440 --> 00:19:06.400 result is what the source calls ASI artificial super intelligence. 392 00:19:05.799 --> 00:19:08.319 And the comparison they use here is humbling. It's not 393 00:19:08.400 --> 00:19:10.519 just Einstein level we tend to think of it that way. 394 00:19:10.759 --> 00:19:14.079 No, we tend to think of intelligence on this very 395 00:19:14.200 --> 00:19:18.240 narrow linear scale. You have a village idiot than an 396 00:19:18.279 --> 00:19:23.039 average person, than Einstein. We think superintelligence is just one 397 00:19:23.119 --> 00:19:26.200 step above Einstein, but the source compares it to the 398 00:19:26.200 --> 00:19:29.920 difference between a human and an ant. Wow, a superintelligence 399 00:19:29.960 --> 00:19:32.400 would be so far above us that we literally couldn't 400 00:19:32.440 --> 00:19:35.720 comprehend its reasoning. It would be looking at our hardest 401 00:19:35.759 --> 00:19:38.279