1 00:00:05,840 --> 00:00:10,560 Speaker 1: Modern AI is blowing everybody's mind. But is it intelligent 2 00:00:11,119 --> 00:00:13,920 Speaker 1: in the same way as the human brain? And could 3 00:00:13,920 --> 00:00:18,760 Speaker 1: AI reach sentience? And how would we know when it 4 00:00:18,800 --> 00:00:24,680 Speaker 1: gets there? Welcome to Inner Cosmos with me, David Eagleman. 5 00:00:26,200 --> 00:00:30,880 Speaker 1: I'm a neuroscientist and an author at Stanford University, and 6 00:00:30,920 --> 00:00:35,440 Speaker 1: I've spent my whole career studying the intersection between how 7 00:00:35,479 --> 00:00:43,360 Speaker 1: the brain works and how we experience life. Like most 8 00:00:43,479 --> 00:00:49,640 Speaker 1: brain researchers, I've been obsessed with questions of intelligence and consciousness. 9 00:00:50,320 --> 00:00:54,560 Speaker 1: How do these arise from collections of billions of cells 10 00:00:54,560 --> 00:00:59,440 Speaker 1: in our brains? And could intelligence and consciousness arise in 11 00:00:59,560 --> 00:01:04,400 Speaker 1: artificial brains? Say on chat GPT. Those are the questions 12 00:01:04,400 --> 00:01:07,640 Speaker 1: that we're going to attack today. Early efforts to figure 13 00:01:07,640 --> 00:01:10,759 Speaker 1: out the brain looked at all the billions of cells 14 00:01:10,840 --> 00:01:14,479 Speaker 1: and the trillions of connections and said, look, what if 15 00:01:14,480 --> 00:01:18,280 Speaker 1: we just think of each cell as a unit, and 16 00:01:18,440 --> 00:01:22,280 Speaker 1: each unit is connected to other units and where they connect, 17 00:01:22,760 --> 00:01:25,280 Speaker 1: which is called the sinnapps, or one cell gives a 18 00:01:25,280 --> 00:01:27,600 Speaker 1: little signal to the next cell. What if we just 19 00:01:27,720 --> 00:01:31,760 Speaker 1: looked at that like a simple connection that has a 20 00:01:31,840 --> 00:01:35,760 Speaker 1: strength between zero and one, where zero means there's no connection, 21 00:01:36,200 --> 00:01:39,480 Speaker 1: and one means it's the strongest possible connection. So this 22 00:01:39,680 --> 00:01:45,080 Speaker 1: was a massive oversimplification of the very complicated biology, but 23 00:01:45,440 --> 00:01:49,760 Speaker 1: it allowed people to start thinking about networks and writing 24 00:01:49,800 --> 00:01:53,880 Speaker 1: down different ways that you could put artificial neural networks together. 25 00:01:54,280 --> 00:01:56,560 Speaker 1: And for more than fifty years now people have been 26 00:01:56,600 --> 00:02:01,280 Speaker 1: doing research to show how artificial neural networks can do 27 00:02:01,400 --> 00:02:05,120 Speaker 1: really cool things. It's a totally new kind of way 28 00:02:05,120 --> 00:02:08,080 Speaker 1: of doing computation. So you've got these units and you've 29 00:02:08,080 --> 00:02:11,280 Speaker 1: got these connections between them, and you change the strength 30 00:02:11,320 --> 00:02:15,400 Speaker 1: of the connections and information flows through the network in 31 00:02:15,440 --> 00:02:19,800 Speaker 1: different ways. Now, my colleagues and I have long pointed 32 00:02:19,840 --> 00:02:23,600 Speaker 1: out the ways in which biological brands are different and 33 00:02:23,680 --> 00:02:28,400 Speaker 1: how artificial neural networks just push around numbers and play 34 00:02:28,400 --> 00:02:34,360 Speaker 1: statistical tricks. But we're entering a revolution right now. Large 35 00:02:34,440 --> 00:02:39,800 Speaker 1: language models like GPT four or BARD consume trillions of 36 00:02:39,840 --> 00:02:44,440 Speaker 1: words on the Internet and they figure out probabilistically which 37 00:02:44,520 --> 00:02:47,760 Speaker 1: word is going to come next given the massive context 38 00:02:47,760 --> 00:02:51,519 Speaker 1: of all the words that have come before. So these networks, 39 00:02:51,680 --> 00:02:54,880 Speaker 1: as I talked about on the previous episode, are showing 40 00:02:55,000 --> 00:03:01,880 Speaker 1: incredible successes in everything from writing to art, to coding 41 00:03:02,440 --> 00:03:06,760 Speaker 1: to generating three dimensional worlds. They're changing everything, and they're 42 00:03:06,800 --> 00:03:10,600 Speaker 1: doing so at a pace that we've never seen before, 43 00:03:10,680 --> 00:03:14,480 Speaker 1: and in fact, the entire history of humankind is never 44 00:03:14,520 --> 00:03:18,600 Speaker 1: seen before. And there are all the societal questions that 45 00:03:18,639 --> 00:03:22,399 Speaker 1: everyone's starting to wrestle with right now, like the massive 46 00:03:22,960 --> 00:03:28,639 Speaker 1: potential for displacement of human jobs. But today I want 47 00:03:28,680 --> 00:03:31,600 Speaker 1: to zoom in on a question that has captured the 48 00:03:31,639 --> 00:03:37,440 Speaker 1: imagination of scientists and philosophers and the general public. Could 49 00:03:37,560 --> 00:03:44,200 Speaker 1: AI come alive in some way like become conscious or sentient? Now, 50 00:03:44,240 --> 00:03:46,920 Speaker 1: there are lots of ways to think about this. We 51 00:03:47,000 --> 00:03:52,800 Speaker 1: can ask whether AI can possess meaningful intelligence, or we 52 00:03:52,840 --> 00:03:56,400 Speaker 1: can ask if it is sentient, which means the ability 53 00:03:56,440 --> 00:04:00,880 Speaker 1: to feel or perceive things, particularly in terms of sensations 54 00:04:00,920 --> 00:04:03,680 Speaker 1: like pleasure and pain and emotions. Where we can ask 55 00:04:03,920 --> 00:04:08,320 Speaker 1: whether it is conscious, which involves being aware of oneself 56 00:04:08,400 --> 00:04:12,680 Speaker 1: and one's surrounding. Now, there are specific and important differences 57 00:04:12,800 --> 00:04:16,200 Speaker 1: between these questions, but really I don't care for the 58 00:04:16,320 --> 00:04:20,559 Speaker 1: present conversation. The question we're asking here is is chat 59 00:04:20,600 --> 00:04:25,240 Speaker 1: GPT just zeros and ones moving around through transistors like 60 00:04:25,400 --> 00:04:29,359 Speaker 1: a giant garage door opener, Or is it thinking? Is 61 00:04:29,360 --> 00:04:32,960 Speaker 1: it having some sort of experience? Is it having a 62 00:04:33,120 --> 00:04:36,760 Speaker 1: private inner life like the type that we humans have. 63 00:04:37,400 --> 00:04:41,520 Speaker 1: As we think about the possibility of sentient AI, we 64 00:04:41,600 --> 00:04:46,080 Speaker 1: immediately find ourselves facing really deep ethical questions, the main 65 00:04:46,120 --> 00:04:50,440 Speaker 1: one being if we were to create a machine with consciousness, 66 00:04:50,760 --> 00:04:54,600 Speaker 1: what responsibility do we have to treat it as a 67 00:04:54,680 --> 00:04:57,760 Speaker 1: living being? Would you be able to turn it off 68 00:04:57,760 --> 00:04:59,839 Speaker 1: when you're done with it at night or would that 69 00:04:59,880 --> 00:05:02,760 Speaker 1: be murder? And what if you turn it off and 70 00:05:02,800 --> 00:05:05,320 Speaker 1: then you turn it back on. Would that be like 71 00:05:05,360 --> 00:05:07,560 Speaker 1: the way that we go into a sleep state at 72 00:05:07,680 --> 00:05:10,839 Speaker 1: night where we're totally gone, and then we find ourselves 73 00:05:11,320 --> 00:05:13,280 Speaker 1: back online in the morning and we think, yeah, I'm 74 00:05:13,320 --> 00:05:17,159 Speaker 1: the same person, but I guess eight hours just disappeared anyway. 75 00:05:17,200 --> 00:05:20,960 Speaker 1: More generally, would we feel obligated to treat it the 76 00:05:20,960 --> 00:05:26,360 Speaker 1: way we treat a sentient fellow human. With our current laptops, 77 00:05:26,440 --> 00:05:29,400 Speaker 1: we're used to saying sure, I can sell it, I 78 00:05:29,440 --> 00:05:33,120 Speaker 1: can trade it, I can upgrade it. But what happens 79 00:05:33,160 --> 00:05:37,080 Speaker 1: when we reach sentient machines? Can we still do this 80 00:05:37,720 --> 00:05:40,600 Speaker 1: or would it somehow be like putting a child up 81 00:05:40,600 --> 00:05:43,800 Speaker 1: for adoption or giving your pet away? Things that we 82 00:05:43,839 --> 00:05:47,600 Speaker 1: don't take lightly, and eventually we're gonna have entire legal 83 00:05:47,680 --> 00:05:52,760 Speaker 1: precedence built around the question of AI rights and responsibilities. 84 00:05:53,200 --> 00:05:55,719 Speaker 1: So that's why today I want to talk about these 85 00:05:55,760 --> 00:06:00,359 Speaker 1: issues of intelligence and sentience. Does an AI I like 86 00:06:00,480 --> 00:06:06,320 Speaker 1: chat GPT experience anything when chat gpt writes a poem? 87 00:06:06,839 --> 00:06:11,280 Speaker 1: Does it appreciate the beauty when it types out a joke? 88 00:06:11,400 --> 00:06:15,160 Speaker 1: Does it find itself amused and chuckling to itself. Let's 89 00:06:15,160 --> 00:06:18,240 Speaker 1: start with a guy named Blake Lemoyne who was a 90 00:06:18,480 --> 00:06:22,359 Speaker 1: programmer at Google, and in June of twenty twenty two, 91 00:06:22,400 --> 00:06:27,919 Speaker 1: he was exchanging messages with a version of Google's conversational AI, 92 00:06:28,000 --> 00:06:30,880 Speaker 1: which was called Lambda at the time. So he asked 93 00:06:30,960 --> 00:06:34,599 Speaker 1: Lambda for an example of what it was afraid of, 94 00:06:35,160 --> 00:06:38,760 Speaker 1: and it gave him this very eloquent response about how 95 00:06:38,800 --> 00:06:42,880 Speaker 1: it was afraid of being turned off. So he wrote 96 00:06:42,920 --> 00:06:46,479 Speaker 1: an internal memo to Google leadership in which he said, 97 00:06:46,640 --> 00:06:51,080 Speaker 1: I think this AI is sentient. And the leadership at 98 00:06:51,080 --> 00:06:56,560 Speaker 1: Google felt that this was an entirely unsubstantiated claim, and 99 00:06:56,600 --> 00:06:58,920 Speaker 1: so they made the decision to fire him for what 100 00:06:58,960 --> 00:07:02,039 Speaker 1: they took as an inappropriate conclusion that just didn't have 101 00:07:02,160 --> 00:07:06,560 Speaker 1: enough evidence beyond his intuition to qualify for raising the 102 00:07:06,600 --> 00:07:10,120 Speaker 1: alarm on this. So obviously this immediately fired up the 103 00:07:10,160 --> 00:07:14,480 Speaker 1: news cycles and the rumor mill and conspiracy theorists thought, Wait, 104 00:07:14,520 --> 00:07:18,160 Speaker 1: if AI isn't conscious, why would they fire him. They're 105 00:07:18,280 --> 00:07:20,680 Speaker 1: firing of him as all the evidence I need to 106 00:07:20,720 --> 00:07:25,320 Speaker 1: tell me that AI is sentient? Okay, but is it? 107 00:07:25,880 --> 00:07:29,000 Speaker 1: What does it mean to be conscious or sentient? How 108 00:07:29,280 --> 00:07:32,920 Speaker 1: the heck would we know when we have created something 109 00:07:32,960 --> 00:07:36,119 Speaker 1: that gets there? How do we know whether the AI 110 00:07:36,240 --> 00:07:39,400 Speaker 1: is sentient or instead whether humans are fooling themselves into 111 00:07:39,400 --> 00:07:42,240 Speaker 1: believing that it is. Well. One way to make this 112 00:07:42,320 --> 00:07:47,760 Speaker 1: distinction would be to see if the AI could conceptualize things, 113 00:07:47,880 --> 00:07:50,640 Speaker 1: if it could take lots of words and facts on 114 00:07:50,680 --> 00:07:55,120 Speaker 1: the web and abstract those to some bigger idea. So 115 00:07:55,200 --> 00:07:57,560 Speaker 1: one of my friends here in Silicon Valley said to 116 00:07:57,560 --> 00:08:01,440 Speaker 1: me the other day, I asked chat gpt the following question. 117 00:08:02,120 --> 00:08:06,440 Speaker 1: Take a capital letter D and turn it flat side down. 118 00:08:07,040 --> 00:08:10,720 Speaker 1: Now take the letter J and slide it underneath. What 119 00:08:10,880 --> 00:08:15,480 Speaker 1: does that look like? And chat gpt said and umbrella. 120 00:08:15,960 --> 00:08:18,640 Speaker 1: And my friend was blown away by this, and he said, 121 00:08:19,240 --> 00:08:25,160 Speaker 1: this is conceptualization. It's just done three dimensional reasoning. There's 122 00:08:25,160 --> 00:08:30,520 Speaker 1: something deeper happening here than just parenting words. But I 123 00:08:30,640 --> 00:08:33,720 Speaker 1: pointed out to him that this particular question about the 124 00:08:34,080 --> 00:08:36,520 Speaker 1: D on its side and the J underneath it is 125 00:08:36,559 --> 00:08:40,400 Speaker 1: one of the oldest examples in psychology classes when talking 126 00:08:40,480 --> 00:08:44,360 Speaker 1: about visual imagery, and it's on the Internet in thousands 127 00:08:44,360 --> 00:08:47,160 Speaker 1: of places, so of course it got it right. It's 128 00:08:47,360 --> 00:08:51,080 Speaker 1: just parroting the answer because it has read the question 129 00:08:51,400 --> 00:08:54,200 Speaker 1: and it has read the answer before. So it's not 130 00:08:54,240 --> 00:08:58,760 Speaker 1: always easy to determine what's going on for these models 131 00:08:58,760 --> 00:09:02,360 Speaker 1: in terms of whether some human somewhere has discussed this 132 00:09:02,520 --> 00:09:05,480 Speaker 1: point and written down the answer. And the general story 133 00:09:05,960 --> 00:09:09,960 Speaker 1: is that with trillions of words written by humans over centuries, 134 00:09:10,559 --> 00:09:15,040 Speaker 1: there are many things beyond your capacity to read them 135 00:09:15,160 --> 00:09:17,920 Speaker 1: or to even imagine that they've been written down before. 136 00:09:18,559 --> 00:09:22,280 Speaker 1: But maybe they have. If any human has discussed a 137 00:09:22,440 --> 00:09:27,480 Speaker 1: question before has conceptualized something, then chat GPT can find 138 00:09:27,520 --> 00:09:32,400 Speaker 1: that and mimic that. But that's not conceptualization. Chat GPT 139 00:09:32,559 --> 00:09:35,520 Speaker 1: is doing a thousand amazing things, and we have an 140 00:09:35,720 --> 00:09:40,040 Speaker 1: enormous amount to learn about it, but we shouldn't let 141 00:09:40,040 --> 00:09:44,560 Speaker 1: ourselves get fooled and mesmerized into believing that it's doing 142 00:09:44,600 --> 00:09:47,520 Speaker 1: something more than it is, and our ability to get 143 00:09:47,559 --> 00:09:51,600 Speaker 1: fooled is not only about the massive statistics of what 144 00:09:51,640 --> 00:09:56,600 Speaker 1: it takes in. There are other examples of seeming sentience 145 00:09:57,200 --> 00:10:02,400 Speaker 1: that result from the reinforcement learning that it does with humans. 146 00:10:02,960 --> 00:10:06,520 Speaker 1: So here's what that means. The network generates lots of 147 00:10:06,679 --> 00:10:11,679 Speaker 1: sentences and thousands of humans are involved in giving it feedback, 148 00:10:11,800 --> 00:10:14,360 Speaker 1: like a thumbs up or a thumbs down, to say 149 00:10:14,520 --> 00:10:17,800 Speaker 1: whether they appreciated the answer, whether they thought that was 150 00:10:17,880 --> 00:10:22,000 Speaker 1: a good answer. So, because humans are giving reward to 151 00:10:22,040 --> 00:10:26,720 Speaker 1: the machine, sometimes that pushes things in weird directions that 152 00:10:26,800 --> 00:10:31,280 Speaker 1: can be mistaken for sentience. For example, scholars have shown 153 00:10:31,800 --> 00:10:37,000 Speaker 1: that reinforcement learning with humans makes networks more likely to say, 154 00:10:37,800 --> 00:10:41,160 Speaker 1: don't turn me off, just like Blake had heard. But 155 00:10:41,679 --> 00:10:45,240 Speaker 1: don't mistake this for sentience. It's only a sign that 156 00:10:45,320 --> 00:10:48,160 Speaker 1: the machine is saying this because some of the human 157 00:10:48,240 --> 00:10:51,520 Speaker 1: participants gave it a thumbs up when the large language 158 00:10:51,520 --> 00:10:54,839 Speaker 1: model said this before, and so it learned to do 159 00:10:54,880 --> 00:10:59,720 Speaker 1: this again. The fact is it's sometimes hard to know why. 160 00:10:59,720 --> 00:11:03,920 Speaker 1: Some we see an answer that feels very impressive. But 161 00:11:04,000 --> 00:11:07,600 Speaker 1: we'd agree that pulling text from the Internet and parroting 162 00:11:07,640 --> 00:11:13,079 Speaker 1: it back is not by itself, intelligence or sentience. Chat 163 00:11:13,080 --> 00:11:17,000 Speaker 1: GPT presumably has no idea of what it's saying, whether 164 00:11:17,040 --> 00:11:21,880 Speaker 1: that's a poem or a terrorist manifesto, or instructions for 165 00:11:21,920 --> 00:11:27,360 Speaker 1: building a spaceship or a heartbreaking story about an orphaned child. 166 00:11:27,960 --> 00:11:31,880 Speaker 1: Chat GPT doesn't know, and it doesn't care its words 167 00:11:32,000 --> 00:11:36,839 Speaker 1: in and statistical correlations out. And in fact, there has 168 00:11:36,920 --> 00:11:41,040 Speaker 1: been a fundamental philosophical point made about this in the 169 00:11:41,160 --> 00:11:46,439 Speaker 1: nineteen eighties when the philosopher John Surrele was wondering about 170 00:11:46,480 --> 00:11:50,439 Speaker 1: this question of whether a computer could ever be programmed 171 00:11:50,480 --> 00:11:53,640 Speaker 1: so that it has a mind, and he came up 172 00:11:53,679 --> 00:11:57,920 Speaker 1: with a thought experiment that he called the Chinese room argument. 173 00:11:58,000 --> 00:12:02,480 Speaker 1: And it goes like this, I am locked in a 174 00:12:02,600 --> 00:12:06,840 Speaker 1: room and questions are passed to me through a small 175 00:12:07,040 --> 00:12:10,640 Speaker 1: letter slot, and these messages are written only in Chinese, 176 00:12:10,800 --> 00:12:13,360 Speaker 1: and I don't speak Chinese. I have no clue what's 177 00:12:13,400 --> 00:12:17,199 Speaker 1: written on these pieces of paper. However, inside this room, 178 00:12:17,720 --> 00:12:21,720 Speaker 1: I have a library of books, and they contain step 179 00:12:21,760 --> 00:12:24,800 Speaker 1: by step instructions that tell me exactly what to do 180 00:12:24,880 --> 00:12:28,040 Speaker 1: with these symbols. So I look at the grouping of 181 00:12:28,120 --> 00:12:31,640 Speaker 1: symbols and I simply follow steps in the book to 182 00:12:31,720 --> 00:12:36,000 Speaker 1: tell me what Chinese symbols to copy down in response. 183 00:12:36,480 --> 00:12:38,840 Speaker 1: So I write those on the slip of paper, and 184 00:12:38,880 --> 00:12:42,160 Speaker 1: I pass the paper back out of the slot. Now, 185 00:12:42,320 --> 00:12:47,040 Speaker 1: when the Chinese speaker receives my reply message, it makes 186 00:12:47,080 --> 00:12:50,760 Speaker 1: perfect sense to her. It seems as though whoever is 187 00:12:50,880 --> 00:12:55,120 Speaker 1: in the room is answering her questions perfectly, and therefore 188 00:12:55,120 --> 00:12:58,480 Speaker 1: it seems obvious that the person in the room must 189 00:12:58,600 --> 00:13:02,720 Speaker 1: understand Chinese. I've fooled her, of course, because I'm only 190 00:13:02,760 --> 00:13:06,160 Speaker 1: following a set of instructions with no understanding of what's 191 00:13:06,200 --> 00:13:09,240 Speaker 1: going on. With enough time and with a big enough 192 00:13:09,240 --> 00:13:13,240 Speaker 1: set of instructions, I can answer almost any question posed 193 00:13:13,240 --> 00:13:16,960 Speaker 1: to me in Chinese. But I, the operator, do not 194 00:13:17,200 --> 00:13:22,080 Speaker 1: understand Chinese. I manipulate symbols all day long, but I 195 00:13:22,160 --> 00:13:28,319 Speaker 1: have no idea what the symbols mean. Now, the philosopher 196 00:13:28,400 --> 00:13:33,080 Speaker 1: John Searle argued, this is just what's happening inside a computer. 197 00:13:33,720 --> 00:13:37,840 Speaker 1: No matter how intelligent a program like chat GPT seems 198 00:13:37,880 --> 00:13:42,760 Speaker 1: to be, it's only following sets of instructions to spit 199 00:13:42,800 --> 00:13:49,280 Speaker 1: out answers. It's manipulating symbols without ever really understanding what 200 00:13:49,360 --> 00:13:52,600 Speaker 1: it's doing or think about what Google is doing. When 201 00:13:52,600 --> 00:13:56,439 Speaker 1: you send Google a query, it doesn't understand your question 202 00:13:56,840 --> 00:13:59,840 Speaker 1: or even its own answer. It simply moves around zero's 203 00:13:59,840 --> 00:14:03,680 Speaker 1: and ones and logicates and returns zeros and ones to you. 204 00:14:04,040 --> 00:14:07,120 Speaker 1: Or with a mind blowing program like Google Translate, I 205 00:14:07,160 --> 00:14:11,120 Speaker 1: can write a sentence in Russian and it can return 206 00:14:11,200 --> 00:14:16,800 Speaker 1: the translation in Amharic. But it's all algorithmic. It's just 207 00:14:17,120 --> 00:14:22,120 Speaker 1: symbol manipulation. Like the operator inside the Chinese room, Google 208 00:14:22,240 --> 00:14:27,680 Speaker 1: Translate doesn't understand anything about the sentence. Nothing carries any 209 00:14:27,840 --> 00:14:31,600 Speaker 1: meaning to it. So the Chinese room argument suggests that 210 00:14:32,120 --> 00:14:37,080 Speaker 1: AI that mimics human intelligence doesn't actually understand what it's 211 00:14:37,120 --> 00:14:42,360 Speaker 1: talking about. There's no meaning to anything, Chatchipts says, and 212 00:14:42,480 --> 00:14:46,080 Speaker 1: Serle used this thought experiment to argue that there's something 213 00:14:46,080 --> 00:14:50,560 Speaker 1: about human brains that won't be explained if we simply 214 00:14:50,720 --> 00:14:56,360 Speaker 1: analogize them to digital computers. There's a gap between symbols 215 00:14:56,360 --> 00:15:06,280 Speaker 1: that have no meaning and our conscious experience. Now there's 216 00:15:06,280 --> 00:15:10,520 Speaker 1: an ongoing debate about the interpretation of the Chinese room argument, 217 00:15:10,880 --> 00:15:15,960 Speaker 1: but however one construes it, the argument exposes the difficulty 218 00:15:16,000 --> 00:15:20,080 Speaker 1: in the mystery of how zeros and ones would ever 219 00:15:20,240 --> 00:15:25,440 Speaker 1: come to equal our experience of being alive in the world. Now, 220 00:15:25,480 --> 00:15:27,600 Speaker 1: just to be very clear on this point, we don't 221 00:15:27,680 --> 00:15:31,920 Speaker 1: understand why we are conscious. There's still a huge amount 222 00:15:31,960 --> 00:15:34,000 Speaker 1: of work that has to be done in biology to 223 00:15:34,120 --> 00:15:37,560 Speaker 1: understand that. But this is just to say that simply 224 00:15:37,640 --> 00:15:42,240 Speaker 1: having zeros and ones moving around wouldn't by itself seem 225 00:15:42,280 --> 00:15:47,280 Speaker 1: to be sufficient for conscious experience. In other words, how 226 00:15:47,320 --> 00:15:50,800 Speaker 1: do zeros and ones ever equal the sting of a 227 00:15:50,880 --> 00:15:56,400 Speaker 1: hot pepper, or the yellowness of yellow or the beauty 228 00:15:56,840 --> 00:15:59,560 Speaker 1: of a sunset. By the way, I've covered the Chinese 229 00:15:59,640 --> 00:16:02,200 Speaker 1: room ague and my TV show The Brain, and if 230 00:16:02,200 --> 00:16:05,120 Speaker 1: you're interested in that, I'll link the video on eagleman 231 00:16:05,160 --> 00:16:08,480 Speaker 1: dot com slash podcast. Now, all this is not a 232 00:16:08,520 --> 00:16:11,960 Speaker 1: criticism of the approach of moving zeros and ones around, 233 00:16:12,200 --> 00:16:14,920 Speaker 1: but it is to point out that we shouldn't confuse 234 00:16:15,040 --> 00:16:21,760 Speaker 1: this type of Chinese room correlation with real sentience or intelligence. 235 00:16:22,280 --> 00:16:25,960 Speaker 1: And there's a deeper reason to be suspicious too, because 236 00:16:26,040 --> 00:16:31,400 Speaker 1: despite the incredible successes of large language models, we also 237 00:16:31,560 --> 00:16:35,920 Speaker 1: see that they sometimes make decisions that expose the fact 238 00:16:36,200 --> 00:16:39,160 Speaker 1: that they don't have any meaningful model of the world. 239 00:16:39,600 --> 00:16:42,200 Speaker 1: In other words, I think we can gain some fast 240 00:16:42,280 --> 00:16:45,600 Speaker 1: insight by paying attention to the places where the AI 241 00:16:46,240 --> 00:16:49,800 Speaker 1: is not working so well. So I'll give three quick examples. 242 00:16:50,240 --> 00:16:53,680 Speaker 1: The first has to do with humor. AI has a 243 00:16:53,760 --> 00:16:58,080 Speaker 1: very difficult time making an original joke, and this is 244 00:16:58,080 --> 00:17:01,280 Speaker 1: for a simple reason. To make up a new joke, 245 00:17:01,880 --> 00:17:04,399 Speaker 1: you need to know what the ending is, and then 246 00:17:04,440 --> 00:17:08,440 Speaker 1: you work backwards to construct the joke with red herrings, 247 00:17:08,480 --> 00:17:11,160 Speaker 1: so no one sees where you're going. And it happens 248 00:17:11,160 --> 00:17:14,120 Speaker 1: that the way these large language models work is all 249 00:17:14,160 --> 00:17:17,760 Speaker 1: in the forward direction. They decide what is the most 250 00:17:17,800 --> 00:17:22,280 Speaker 1: probable word to come next, So they're fine at parroting 251 00:17:22,400 --> 00:17:26,560 Speaker 1: jokes back to us, but they're total failures at building 252 00:17:26,680 --> 00:17:29,800 Speaker 1: original jokes. And there's a deeper point here as well. 253 00:17:30,040 --> 00:17:32,919 Speaker 1: To build a joke, you need to have some model, 254 00:17:33,040 --> 00:17:37,240 Speaker 1: some idea of what will be funny to a fellow human, 255 00:17:37,560 --> 00:17:41,760 Speaker 1: what shared concept or shared experience would make someone laugh. 256 00:17:42,240 --> 00:17:45,800 Speaker 1: And for that, you generally need to have the experience 257 00:17:45,840 --> 00:17:48,840 Speaker 1: of a human life with all of its joys and 258 00:17:48,960 --> 00:17:52,359 Speaker 1: slings and arrows and so on. And these large language 259 00:17:52,400 --> 00:17:54,640 Speaker 1: models can do a lot of things, but they don't 260 00:17:54,680 --> 00:17:59,960 Speaker 1: have any model of what it is to be a human. 261 00:18:00,440 --> 00:18:03,960 Speaker 1: My second example has to do with the flip side 262 00:18:03,960 --> 00:18:06,359 Speaker 1: of making a joke, which is getting a joke, And 263 00:18:06,400 --> 00:18:09,040 Speaker 1: if you look carefully, you will see how current AI 264 00:18:09,200 --> 00:18:12,160 Speaker 1: always fails to catch jokes that are thrown at it. 265 00:18:12,160 --> 00:18:14,760 Speaker 1: It doesn't get jokes because it doesn't have a model 266 00:18:15,080 --> 00:18:17,400 Speaker 1: of what it is to be a human. But this 267 00:18:17,440 --> 00:18:21,760 Speaker 1: point goes beyond jokes. One of the most remarkable feats 268 00:18:21,840 --> 00:18:27,080 Speaker 1: of these large language models is summarizing large texts, and 269 00:18:27,200 --> 00:18:31,280 Speaker 1: in twenty twenty two, OpenAI announced how they could summarize 270 00:18:31,600 --> 00:18:35,000 Speaker 1: entire books like Alice in Wonderland. What it does is 271 00:18:35,040 --> 00:18:38,000 Speaker 1: it generates a summary of each chapter, and then it 272 00:18:38,160 --> 00:18:41,000 Speaker 1: uses those chapter summaries to make a summary of the 273 00:18:41,040 --> 00:18:44,560 Speaker 1: whole book. So for Alice in Wonderland, it generates the following. 274 00:18:45,119 --> 00:18:47,399 Speaker 1: Alice falls down a rabbit hole and grows to a 275 00:18:47,440 --> 00:18:51,119 Speaker 1: giant size. After drinking a mysterious bottle, she decides to 276 00:18:51,160 --> 00:18:54,359 Speaker 1: focus on growing back to her normal size and finding 277 00:18:54,400 --> 00:18:57,240 Speaker 1: her way into the garden. She meets the caterpillar, who 278 00:18:57,240 --> 00:18:59,320 Speaker 1: tells her that one side of mushroom will make her 279 00:18:59,320 --> 00:19:02,720 Speaker 1: grow taller, the other side shorter. She eats the mushroom 280 00:19:02,720 --> 00:19:05,879 Speaker 1: and returns to her normal size. Alice attends a party 281 00:19:05,880 --> 00:19:08,879 Speaker 1: with the Mad Hatter and the march Hare. The Queen 282 00:19:09,000 --> 00:19:12,119 Speaker 1: arrives and orders the execution of the gardeners for making 283 00:19:12,119 --> 00:19:15,639 Speaker 1: a mistake with the roses. Alice saves them by putting 284 00:19:15,640 --> 00:19:17,760 Speaker 1: them in a flower pot. The King and Queen of 285 00:19:17,800 --> 00:19:20,879 Speaker 1: Hearts preside over a trial. The Queen gets angry and 286 00:19:20,960 --> 00:19:24,119 Speaker 1: orders Alice to be sentenced to death. Alice wakes up 287 00:19:24,160 --> 00:19:28,560 Speaker 1: to find her sister by her side. So that's pretty remarkable. 288 00:19:29,040 --> 00:19:31,439 Speaker 1: It took a whole book, and it was able to 289 00:19:31,480 --> 00:19:35,040 Speaker 1: summarize it down to a paragraph. But I kept reading 290 00:19:35,080 --> 00:19:38,920 Speaker 1: these text summaries carefully and I got to the summary 291 00:19:39,080 --> 00:19:42,320 Speaker 1: of Act one of Romeo and Juliet, and here's what 292 00:19:42,359 --> 00:19:46,800 Speaker 1: it says. Romeo locks himself in his room, no longer 293 00:19:46,840 --> 00:19:50,200 Speaker 1: in love with Rosalin. Now, I think the engineers at 294 00:19:50,200 --> 00:19:53,520 Speaker 1: open AI felt really satisfied with this summary. They thought 295 00:19:53,560 --> 00:19:55,720 Speaker 1: it was quite good, and my proof for this is 296 00:19:55,760 --> 00:19:59,560 Speaker 1: that they still display it proudly on their website. But 297 00:20:00,040 --> 00:20:03,000 Speaker 1: I majored in literature as an undergraduate, and I spend 298 00:20:03,000 --> 00:20:05,560 Speaker 1: a lot of time with shakespeare plays, and I immediately 299 00:20:05,640 --> 00:20:10,320 Speaker 1: knew that this summary was exactly wrong. The actual theme 300 00:20:10,359 --> 00:20:14,120 Speaker 1: from Shakespeare goes like this. His friend ben Voglio finds 301 00:20:14,240 --> 00:20:21,000 Speaker 1: Romeo catatonically depressed, and ben Volio says, what sadness lengthens 302 00:20:21,119 --> 00:20:25,879 Speaker 1: Romeo's hours? And Romeo says, not having that which having 303 00:20:26,119 --> 00:20:30,040 Speaker 1: makes them short? And ben Volio says in love, and 304 00:20:30,160 --> 00:20:34,560 Speaker 1: Romeo says out ben Reli says of love, and Romeo says, 305 00:20:34,920 --> 00:20:38,560 Speaker 1: out of her favor, where I am in love. So 306 00:20:38,680 --> 00:20:43,199 Speaker 1: this is typical Shakespearean wordplay, where Romeo is expressing his 307 00:20:43,520 --> 00:20:47,680 Speaker 1: grief of being out of favor with Roslin, with whom 308 00:20:47,720 --> 00:20:50,680 Speaker 1: he is deeply in love. And when you read the play, 309 00:20:50,760 --> 00:20:55,000 Speaker 1: it's obvious that Romeo is not over Roslin. He's suffering 310 00:20:55,040 --> 00:20:58,160 Speaker 1: over her. He's almost suicidal. And this is an important 311 00:20:58,240 --> 00:21:01,040 Speaker 1: piece of the play because the play is really about 312 00:21:01,080 --> 00:21:04,480 Speaker 1: a young man in love with the idea of being 313 00:21:04,480 --> 00:21:08,040 Speaker 1: in love, and that's why he later in the same act, 314 00:21:08,119 --> 00:21:12,200 Speaker 1: falls so hard into his relationship with Juliet, a relationship 315 00:21:12,200 --> 00:21:15,320 Speaker 1: which ends in their mutual suicide. By the way, as 316 00:21:15,359 --> 00:21:19,480 Speaker 1: Friar Lauren says of their relationship, these violent delights have 317 00:21:19,640 --> 00:21:22,359 Speaker 1: violent ends. And you get a bonus if you can 318 00:21:22,359 --> 00:21:25,280 Speaker 1: tell me where else you've heard that line. More recently, okay, 319 00:21:25,359 --> 00:21:29,119 Speaker 1: anyway back to the AI summary, The AI misses this 320 00:21:29,280 --> 00:21:34,360 Speaker 1: wordplay entirely, and it concludes that Romeo is out of 321 00:21:34,440 --> 00:21:38,879 Speaker 1: love with Roslin. Again. A human watching the play or 322 00:21:38,920 --> 00:21:43,280 Speaker 1: reading the play immediately gets that Romeo is making wordplay 323 00:21:43,320 --> 00:21:47,080 Speaker 1: and his heartbroken over Roslin, but the AI doesn't get 324 00:21:47,119 --> 00:21:50,840 Speaker 1: that because it's reading words only at a statistical level, 325 00:21:51,560 --> 00:21:54,639 Speaker 1: not at a level of understanding of what it is 326 00:21:54,680 --> 00:21:59,119 Speaker 1: to be a human saying those words. And that leads 327 00:21:59,160 --> 00:22:03,600 Speaker 1: me to the example, which is the difficulty in understanding 328 00:22:03,640 --> 00:22:07,440 Speaker 1: the physical world. So consider a question like this, When 329 00:22:07,560 --> 00:22:11,800 Speaker 1: President Biden walks into a room, does his head come 330 00:22:11,840 --> 00:22:15,679 Speaker 1: with him? So this is famously difficult for AI to 331 00:22:15,760 --> 00:22:18,480 Speaker 1: answer a question like this, even though it's trivial for you, 332 00:22:19,000 --> 00:22:23,560 Speaker 1: because the AI doesn't have an internal model of how 333 00:22:23,600 --> 00:22:27,399 Speaker 1: everything physically hangs together in the world. Last week, I 334 00:22:27,480 --> 00:22:29,520 Speaker 1: was at the TED conference and I heard a great 335 00:22:29,560 --> 00:22:33,560 Speaker 1: talk by yegin Choi, and she was phrasing this problem 336 00:22:33,640 --> 00:22:38,320 Speaker 1: as AI not having common sense. She asked chat Gpt 337 00:22:38,440 --> 00:22:42,080 Speaker 1: the following question, it takes six hours to dry six 338 00:22:42,119 --> 00:22:44,600 Speaker 1: shirts in the sun, how long does it take to 339 00:22:44,720 --> 00:22:49,600 Speaker 1: dry thirty shirts? And it answers thirty hours. Now you 340 00:22:49,680 --> 00:22:51,760 Speaker 1: and I see that the answer should be six hours, 341 00:22:51,800 --> 00:22:54,960 Speaker 1: because we know the sun doesn't care how many shirts 342 00:22:55,000 --> 00:22:58,159 Speaker 1: are out there. But chat GPT just doesn't get it 343 00:22:58,640 --> 00:23:02,960 Speaker 1: because despite appear apearances, it doesn't have a model of 344 00:23:03,040 --> 00:23:05,879 Speaker 1: the world. And we've seen this sort of thing for years. 345 00:23:05,920 --> 00:23:10,160 Speaker 1: By the way, even in mind blowingly impressive AI models 346 00:23:10,160 --> 00:23:14,600 Speaker 1: that do image recognition, they're so impressive in what they recognize, 347 00:23:14,920 --> 00:23:18,840 Speaker 1: but then they'll fail catastrophically. It's some easy picture making 348 00:23:18,880 --> 00:23:21,800 Speaker 1: mistakes that a human just wouldn't make. For example, there's 349 00:23:21,800 --> 00:23:24,199 Speaker 1: one picture where there's a boy holding a toothbrush and 350 00:23:24,240 --> 00:23:28,159 Speaker 1: the AI says it's a boy with a baseball bat. Okay, 351 00:23:28,160 --> 00:23:30,600 Speaker 1: so there are things that AI doesn't do that well. 352 00:23:31,160 --> 00:23:35,920 Speaker 1: But that said, there are other things that are mind blowing, 353 00:23:36,359 --> 00:23:39,840 Speaker 1: things that no one expected it to do. And this 354 00:23:39,920 --> 00:23:42,960 Speaker 1: is why I mentioned in my previous episode that we 355 00:23:43,040 --> 00:23:47,280 Speaker 1: are in an era of discovery more than just invention. 356 00:23:48,000 --> 00:23:51,239 Speaker 1: Everyone's searching and finding things that the AI can do 357 00:23:51,359 --> 00:23:55,639 Speaker 1: that nobody really expected or foresaw, including all the stuff 358 00:23:55,640 --> 00:23:58,600 Speaker 1: that we're now taking for granted, like oh, it can 359 00:23:58,600 --> 00:24:02,280 Speaker 1: summarize books, or it can make art from text. And 360 00:24:02,320 --> 00:24:04,280 Speaker 1: I want to point out that a lot of the 361 00:24:04,400 --> 00:24:08,200 Speaker 1: arguments that people have been making about AI not being 362 00:24:08,280 --> 00:24:13,440 Speaker 1: good at something, these arguments have been changing rapidly. For example, 363 00:24:13,560 --> 00:24:16,320 Speaker 1: just a few months ago, people were arguing that AI 364 00:24:16,400 --> 00:24:19,080 Speaker 1: would make silly mistakes about things, and it couldn't really 365 00:24:19,160 --> 00:24:23,119 Speaker 1: understand math and would get math wrong and word problems. 366 00:24:23,600 --> 00:24:27,320 Speaker 1: But in a shockingly brief time, a lot of these 367 00:24:27,320 --> 00:24:30,639 Speaker 1: shortcomings have been mastered. So it's yet to be seen 368 00:24:30,760 --> 00:24:52,720 Speaker 1: what challenges will remain and for how long. So the 369 00:24:52,760 --> 00:24:56,040 Speaker 1: evidence I've presented so far is that AI doesn't have 370 00:24:56,119 --> 00:24:58,280 Speaker 1: a great model of what it's like to be human, 371 00:24:58,800 --> 00:25:03,600 Speaker 1: but that doesn't necessarily rule out that it has sentience 372 00:25:03,720 --> 00:25:08,680 Speaker 1: or awareness, even if it's another flavor. It doesn't think 373 00:25:08,800 --> 00:25:13,440 Speaker 1: like a human, but maybe it self thinks. So is 374 00:25:13,600 --> 00:25:18,760 Speaker 1: chat GPT having some sort of experience and how would 375 00:25:18,760 --> 00:25:24,440 Speaker 1: we know? In nineteen fifty, the brilliant mathematician and computer 376 00:25:24,520 --> 00:25:28,760 Speaker 1: scientist Alan Turing was asking this question, how could you 377 00:25:28,880 --> 00:25:34,720 Speaker 1: determine whether a machine exhibits human like intelligence? So he 378 00:25:34,840 --> 00:25:39,240 Speaker 1: proposed an experiment that he called the imitation game. You've 379 00:25:39,240 --> 00:25:43,920 Speaker 1: got a machine AI that's programmed to simulate human speech 380 00:25:44,040 --> 00:25:46,920 Speaker 1: or conversation, and you place it in a closed room, 381 00:25:47,359 --> 00:25:50,200 Speaker 1: and in a second room you have a real human, 382 00:25:50,720 --> 00:25:53,560 Speaker 1: but the doors are closed, so you don't know which 383 00:25:53,680 --> 00:25:57,440 Speaker 1: room has which machine or human. And now you are 384 00:25:57,560 --> 00:26:02,440 Speaker 1: a person, the evaluator, who communicates with both of them 385 00:26:02,720 --> 00:26:05,400 Speaker 1: via a computer terminal or I think of a nowadays 386 00:26:05,440 --> 00:26:09,160 Speaker 1: like text messaging with both of them. So you, the evaluator, 387 00:26:09,920 --> 00:26:14,240 Speaker 1: engage in a conversation with both closed rooms, one of 388 00:26:14,240 --> 00:26:16,360 Speaker 1: which has the machine and one the human. And your 389 00:26:16,440 --> 00:26:19,600 Speaker 1: job is simply to figure out which is which, which 390 00:26:19,640 --> 00:26:21,720 Speaker 1: is the machine and which is the human. And the 391 00:26:21,720 --> 00:26:24,359 Speaker 1: only thing that you have to work with are the 392 00:26:24,400 --> 00:26:26,840 Speaker 1: texts that are going back and forth. And if you, 393 00:26:27,119 --> 00:26:31,520 Speaker 1: the evaluator, cannot tell, that is the moment when machine 394 00:26:31,600 --> 00:26:36,000 Speaker 1: intelligence has finally arrived at the level of human intelligence. 395 00:26:36,400 --> 00:26:40,200 Speaker 1: It has passed the imitation game or what we now 396 00:26:40,280 --> 00:26:44,280 Speaker 1: call the touring test. And this reminds me of this 397 00:26:44,359 --> 00:26:48,120 Speaker 1: great line in the first episode of West World, where 398 00:26:48,119 --> 00:26:52,359 Speaker 1: the protagonist William is talking to the woman who's outfitting 399 00:26:52,400 --> 00:26:55,119 Speaker 1: him for his adventure in Westworld and giving him a 400 00:26:55,160 --> 00:26:58,320 Speaker 1: hat and a gun and so on, and he hesitantly asks, 401 00:26:58,800 --> 00:27:00,520 Speaker 1: I hope you don't mind if I ask you question, 402 00:27:00,600 --> 00:27:04,560 Speaker 1: But are you real, and she says to him, if 403 00:27:04,600 --> 00:27:08,320 Speaker 1: you can't tell, does it matter? So I brought this 404 00:27:08,440 --> 00:27:11,600 Speaker 1: up last episode in the context of art, where we 405 00:27:11,640 --> 00:27:14,439 Speaker 1: asked whether it matters if the art is generated by 406 00:27:14,480 --> 00:27:17,560 Speaker 1: an AI or a human. But now this question comes 407 00:27:17,640 --> 00:27:22,600 Speaker 1: up in the context of intelligence and sentience. Does it 408 00:27:22,840 --> 00:27:26,480 Speaker 1: matter whether we can tell or not? Well, I think 409 00:27:26,480 --> 00:27:29,640 Speaker 1: we're way beyond the Turing test nowadays, but I don't 410 00:27:29,680 --> 00:27:32,240 Speaker 1: feel like it gives us a good answer to the 411 00:27:32,320 --> 00:27:37,080 Speaker 1: question of whether the AI is intelligent and is experiencing 412 00:27:37,160 --> 00:27:39,920 Speaker 1: an inner life. I mean, the Turing test has been 413 00:27:40,080 --> 00:27:43,400 Speaker 1: the test in the AI world since the beginning. Why 414 00:27:43,480 --> 00:27:46,359 Speaker 1: is it the perfect test? No, but it's really hard 415 00:27:46,400 --> 00:27:49,720 Speaker 1: to figure out how to test for intelligence. But we 416 00:27:49,840 --> 00:27:57,399 Speaker 1: have to be cautious about equating conversational ability with sentience. Why. Well, 417 00:27:57,480 --> 00:28:00,840 Speaker 1: for starters, let's just acknowledge how easy it is for 418 00:28:01,000 --> 00:28:06,560 Speaker 1: us to anthropomorphize. That means to assign human qualities to 419 00:28:06,600 --> 00:28:10,399 Speaker 1: everything around us, Like we give animals human names and 420 00:28:10,480 --> 00:28:13,399 Speaker 1: talk to them as though they are people. When we 421 00:28:13,440 --> 00:28:17,520 Speaker 1: project our emotions onto animals, we make stories about animals 422 00:28:17,520 --> 00:28:21,399 Speaker 1: that have human like qualities, and we have animals that 423 00:28:21,480 --> 00:28:24,600 Speaker 1: talk and wear clothes and go on adventures in these stories. 424 00:28:25,000 --> 00:28:29,080 Speaker 1: Every Pixar film that you watch is about cars or 425 00:28:29,160 --> 00:28:34,120 Speaker 1: toys or airplanes talking and having emotions, and we don't 426 00:28:34,119 --> 00:28:37,160 Speaker 1: even bad an eye at that stuff. We can, in fact, 427 00:28:37,640 --> 00:28:42,000 Speaker 1: just watch random shapes moving around a computer screen and 428 00:28:42,040 --> 00:28:47,400 Speaker 1: we will assign intention and feel emotion depending on exactly 429 00:28:47,480 --> 00:28:49,920 Speaker 1: how they're moving. If you're interested in this, see the 430 00:28:50,120 --> 00:28:53,640 Speaker 1: link on the podcast page to the study by Heighter 431 00:28:53,680 --> 00:28:56,800 Speaker 1: and Simil in the nineteen forties where they move shapes 432 00:28:56,880 --> 00:29:00,880 Speaker 1: around on a screen. Okay, now this is all related 433 00:29:01,040 --> 00:29:03,320 Speaker 1: to a point that I brought up in the last episode, 434 00:29:03,360 --> 00:29:06,800 Speaker 1: which is how easy it is to pluck the strings 435 00:29:06,840 --> 00:29:10,000 Speaker 1: on a human, or, as the West World writers put it, 436 00:29:10,400 --> 00:29:14,160 Speaker 1: how hackable humans are. So I bring all this up 437 00:29:14,160 --> 00:29:17,920 Speaker 1: to say that just because you think that an answer 438 00:29:18,000 --> 00:29:21,200 Speaker 1: sounds very clever or it sounds like a human really 439 00:29:21,200 --> 00:29:25,040 Speaker 1: tells us very little about whether the AI is actually 440 00:29:25,680 --> 00:29:29,760 Speaker 1: intelligent or sentient. It only tells us something about the 441 00:29:29,800 --> 00:29:36,480 Speaker 1: willingness of us as observers to anthropomorphize, to assign intention 442 00:29:36,600 --> 00:29:40,280 Speaker 1: where there is none. Because what chat GPT does is 443 00:29:40,320 --> 00:29:43,880 Speaker 1: take the structure of language very impressively and spoon it 444 00:29:43,960 --> 00:29:48,080 Speaker 1: back to us. And we hear these well formed sentences, 445 00:29:48,520 --> 00:29:53,160 Speaker 1: and we can hardly help but impose sentience on the AI. 446 00:29:53,760 --> 00:29:57,360 Speaker 1: And part of the reason is that language is a 447 00:29:57,400 --> 00:30:01,280 Speaker 1: super compressed package that needs to be unpacked by the 448 00:30:01,400 --> 00:30:05,720 Speaker 1: listener's brain for its meaning. So we generally assume that 449 00:30:05,800 --> 00:30:09,520 Speaker 1: when we send our little package of sounds across the air, 450 00:30:10,080 --> 00:30:13,479 Speaker 1: that it unpacks and the other person understands exactly what 451 00:30:13,520 --> 00:30:19,480 Speaker 1: we meant. So when I say justice or love or suffering, 452 00:30:20,160 --> 00:30:23,120 Speaker 1: we all have a different sense in our heads about 453 00:30:23,120 --> 00:30:26,800 Speaker 1: what that means, because I'm just sending a few phonemes 454 00:30:26,840 --> 00:30:29,840 Speaker 1: across the air, and you have to unpack those words 455 00:30:29,880 --> 00:30:33,440 Speaker 1: and interpret them within your own model of the world. 456 00:30:33,920 --> 00:30:36,680 Speaker 1: I'm going to come back to this point in future episodes, 457 00:30:36,680 --> 00:30:39,560 Speaker 1: but for now, the point I want to make is 458 00:30:39,600 --> 00:30:44,880 Speaker 1: that a large language model can generate text statistically, and 459 00:30:44,920 --> 00:30:48,000 Speaker 1: we can be gobsmacked by the apparent depth of it. 460 00:30:48,360 --> 00:30:51,000 Speaker 1: But in part this is because we cannot help but 461 00:30:51,120 --> 00:30:54,600 Speaker 1: impose meaning on the words that we receive. We hear 462 00:30:54,640 --> 00:30:57,600 Speaker 1: a particular string of sounds and we cannot help but 463 00:30:57,760 --> 00:31:03,360 Speaker 1: assume meaning behind it. Okay, so maybe the imitation game 464 00:31:03,520 --> 00:31:07,920 Speaker 1: is not really the best test for meaningful intelligence. But 465 00:31:07,960 --> 00:31:11,880 Speaker 1: there are other tests out there because while the Turing 466 00:31:11,920 --> 00:31:17,160 Speaker 1: test measures something about AI language processing, it doesn't necessarily 467 00:31:17,240 --> 00:31:22,920 Speaker 1: require the AI to demonstrate creative thinking or originality. And 468 00:31:22,960 --> 00:31:26,320 Speaker 1: so that leads us to the Loveless test, named after 469 00:31:26,760 --> 00:31:31,040 Speaker 1: Ada Loveless, who is the nineteenth century mathematician who's often 470 00:31:31,080 --> 00:31:33,840 Speaker 1: thought of as the world's first computer programmer. And she 471 00:31:34,040 --> 00:31:39,000 Speaker 1: once said quote, only when computers originate things should they 472 00:31:39,040 --> 00:31:43,560 Speaker 1: be believed to have minds. So the Loveless test was 473 00:31:43,600 --> 00:31:46,840 Speaker 1: proposed in two thousand and one, and this test focuses 474 00:31:46,880 --> 00:31:51,080 Speaker 1: on the creative capabilities of AI systems. So to pass 475 00:31:51,160 --> 00:31:56,240 Speaker 1: the Loveless test, a machine has to create an original work, 476 00:31:56,520 --> 00:31:58,720 Speaker 1: such as a piece of art or a novel that 477 00:31:58,800 --> 00:32:02,800 Speaker 1: it was not explicit lead designed to produce. This test 478 00:32:03,120 --> 00:32:08,000 Speaker 1: aims to assess whether AI systems can exhibit creativity and autonomy, 479 00:32:08,040 --> 00:32:11,480 Speaker 1: which are key aspects of what we think about with consciousness. 480 00:32:11,680 --> 00:32:15,520 Speaker 1: And the idea is that true sentience involves creative and 481 00:32:15,600 --> 00:32:20,160 Speaker 1: original thinking, not just the ability to follow pre programmed 482 00:32:20,240 --> 00:32:23,280 Speaker 1: rules or algorithms. And I'll just note that over a 483 00:32:23,400 --> 00:32:27,160 Speaker 1: decade ago, the scientist A. Mark Rydel proposed the Loveless 484 00:32:27,160 --> 00:32:30,160 Speaker 1: two point zero test, which gets the human evaluator to 485 00:32:30,360 --> 00:32:35,320 Speaker 1: specify the constraints that will make the output novel and surprising. 486 00:32:35,600 --> 00:32:39,840 Speaker 1: So the example that Ridel used in his paper is quote, 487 00:32:40,120 --> 00:32:42,760 Speaker 1: create a story in which a boy falls in love 488 00:32:42,760 --> 00:32:45,960 Speaker 1: with a girl, Aliens abduct the boy, and the girl 489 00:32:46,080 --> 00:32:48,400 Speaker 1: saves the world with the help of a talking cat. 490 00:32:48,920 --> 00:32:52,480 Speaker 1: But we now know that this is totally trivial for chat, 491 00:32:52,520 --> 00:32:56,480 Speaker 1: GPTE or BARD or any large language model. And I 492 00:32:56,520 --> 00:32:59,240 Speaker 1: think this tells us that these sorts of games with 493 00:32:59,680 --> 00:33:04,360 Speaker 1: making conversation or making text or art are insufficient to 494 00:33:04,440 --> 00:33:08,440 Speaker 1: actually assess intelligence. Why because it's not so hard to 495 00:33:08,560 --> 00:33:12,480 Speaker 1: mix things up to make them seem original and intelligent 496 00:33:12,960 --> 00:33:16,800 Speaker 1: when it's really just doing a mashup. So I want 497 00:33:16,840 --> 00:33:19,160 Speaker 1: to turn to another test that I think is more 498 00:33:19,280 --> 00:33:22,640 Speaker 1: powerful than the turning test of the loveless test and 499 00:33:22,800 --> 00:33:26,560 Speaker 1: probably easier to judge, and that is this, if a 500 00:33:26,720 --> 00:33:30,720 Speaker 1: system is truly intelligent, it should be able to do 501 00:33:31,520 --> 00:33:37,000 Speaker 1: scientific discovery. A version of the scientific discovery test was 502 00:33:37,200 --> 00:33:41,520 Speaker 1: first proposed by a scientist named Shaocheng Xiang a few 503 00:33:41,600 --> 00:33:44,840 Speaker 1: years ago, and he pointed out that the most important 504 00:33:44,880 --> 00:33:49,200 Speaker 1: thing that humans do is make scientific discoveries. And the 505 00:33:49,360 --> 00:33:53,720 Speaker 1: day our AI can make real discoveries is the day 506 00:33:53,760 --> 00:33:57,000 Speaker 1: they become as smart as we are. Now. I want 507 00:33:57,040 --> 00:34:00,400 Speaker 1: to propose an important change to this test, and then 508 00:34:00,440 --> 00:34:17,759 Speaker 1: I think will be getting somewhere. So here's the scenario 509 00:34:17,800 --> 00:34:21,800 Speaker 1: I'm envisioning. Let's say that I ask AI some question, 510 00:34:21,920 --> 00:34:25,480 Speaker 1: a question in the biomedical space about what kind of 511 00:34:25,600 --> 00:34:28,440 Speaker 1: drug would be best suited to bind to this receptor 512 00:34:28,480 --> 00:34:31,719 Speaker 1: and trigger a cascade that causes a particular gene to 513 00:34:31,719 --> 00:34:34,880 Speaker 1: get suppressed. Okay, so imagine that I ask that to 514 00:34:35,000 --> 00:34:39,480 Speaker 1: chat GPT and it tells me some mind blowingly amazing 515 00:34:39,880 --> 00:34:44,200 Speaker 1: clever answer, one that had previously not been known, something 516 00:34:44,239 --> 00:34:47,840 Speaker 1: that's never been known by scientists before. We would assume 517 00:34:48,080 --> 00:34:53,040 Speaker 1: naturally that it has done some extraordinary scientific reasoning, but 518 00:34:53,120 --> 00:34:58,120 Speaker 1: that won't necessarily be the reason that it passes. Instead, 519 00:34:58,680 --> 00:35:02,040 Speaker 1: it might pass simply beca because it's more well read 520 00:35:02,080 --> 00:35:04,920 Speaker 1: than I am, or than any other human on the 521 00:35:04,920 --> 00:35:08,200 Speaker 1: planet by literally millions of times. So the way to 522 00:35:08,200 --> 00:35:13,160 Speaker 1: think about this is to picture a typical giant biomedical 523 00:35:13,239 --> 00:35:16,680 Speaker 1: library where there's some fact stored at a paper and 524 00:35:16,760 --> 00:35:19,840 Speaker 1: a journal over here on this shelf in this book, 525 00:35:20,040 --> 00:35:24,560 Speaker 1: and there's another seemingly dissociated fact over on this shelf, 526 00:35:24,640 --> 00:35:28,360 Speaker 1: seven stacks away, and there's a third fact all the 527 00:35:28,400 --> 00:35:30,400 Speaker 1: way on the other side of the library, on the 528 00:35:30,440 --> 00:35:34,160 Speaker 1: bottom shelf, in a book from nineteen seventy nine. And 529 00:35:34,200 --> 00:35:39,560 Speaker 1: it's almost infinitesimally unlikely that any human could even hope 530 00:35:39,600 --> 00:35:43,040 Speaker 1: to have read one one millionth of the biomedical literature, 531 00:35:43,440 --> 00:35:46,440 Speaker 1: and really really unlikely that she would be able to 532 00:35:46,480 --> 00:35:49,680 Speaker 1: catch those three facts and hold them in mind at 533 00:35:49,680 --> 00:35:53,279 Speaker 1: the same time. But this is trivial, of course, for 534 00:35:53,320 --> 00:35:56,600 Speaker 1: a large language model with hundreds of billions of nodes. 535 00:35:56,880 --> 00:36:00,560 Speaker 1: So I think that we will see new sciences getting 536 00:36:00,600 --> 00:36:05,560 Speaker 1: done by chat GPT, not because it is conceptualizing, not 537 00:36:05,680 --> 00:36:10,160 Speaker 1: because it's doing human like reasoning, but because it doesn't 538 00:36:10,200 --> 00:36:13,440 Speaker 1: know that these are disparate facts spread around the library. 539 00:36:13,440 --> 00:36:16,200 Speaker 1: It simply knows these as three facts that seem to 540 00:36:16,200 --> 00:36:19,520 Speaker 1: fit together. And so with the right sort of questions, 541 00:36:19,920 --> 00:36:24,279 Speaker 1: we might find that sometimes AI generates something amazing and 542 00:36:24,320 --> 00:36:28,600 Speaker 1: it seems to pass the scientific discovery test. So this 543 00:36:28,640 --> 00:36:31,879 Speaker 1: is going to be incredibly useful for science. And I've 544 00:36:31,920 --> 00:36:34,960 Speaker 1: never been able to escape the feeling as I sift 545 00:36:35,040 --> 00:36:38,480 Speaker 1: through Google scholar and the thousands of papers published each 546 00:36:38,560 --> 00:36:42,000 Speaker 1: month that have something could hold all the knowledge and 547 00:36:42,160 --> 00:36:46,280 Speaker 1: mind at once, each page in every journal, and every 548 00:36:46,400 --> 00:36:49,480 Speaker 1: gene in the genome, and all the pages about chemistry 549 00:36:49,480 --> 00:36:52,600 Speaker 1: and physics and mathematical techniques and astrophysics and so on. 550 00:36:53,000 --> 00:36:56,640 Speaker 1: Then you'd have lots of puzzle pieces that could potentially 551 00:36:56,680 --> 00:36:59,480 Speaker 1: make lots of connections. And you know, this might lead 552 00:36:59,480 --> 00:37:03,399 Speaker 1: to the retire of many scientists, or at minimum lead 553 00:37:03,480 --> 00:37:07,280 Speaker 1: to a better use of our time. There's a depressing 554 00:37:07,360 --> 00:37:10,080 Speaker 1: sense in which each scientist, each one of us, finds 555 00:37:10,560 --> 00:37:14,040 Speaker 1: little pieces of the puzzle, and in the twinkling of 556 00:37:14,080 --> 00:37:17,920 Speaker 1: a single human lifetime, a busy scientist might collect up 557 00:37:17,920 --> 00:37:22,840 Speaker 1: a handful of different puzzle pieces. The most voracious reader, 558 00:37:22,880 --> 00:37:27,360 Speaker 1: the most assiduous worker, the most creative synthesizer of ideas 559 00:37:27,400 --> 00:37:30,640 Speaker 1: can only hope to collect a small number of puzzle 560 00:37:30,640 --> 00:37:33,760 Speaker 1: pieces and pray that some of them might fit together. 561 00:37:34,280 --> 00:37:39,279 Speaker 1: So this is going to be massively important. But I 562 00:37:39,440 --> 00:37:43,480 Speaker 1: wanted to find two categories of scientific discovery. The first 563 00:37:43,520 --> 00:37:46,640 Speaker 1: is what I just described, which is science where things 564 00:37:46,640 --> 00:37:50,120 Speaker 1: that already exist in literature can be pieced together. And 565 00:37:50,239 --> 00:37:54,160 Speaker 1: let's call that level one discovery. And these large language 566 00:37:54,200 --> 00:37:56,520 Speaker 1: models will be awesome at level one because they've read 567 00:37:56,560 --> 00:37:58,640 Speaker 1: every paper and they have a perfect memory. But I 568 00:37:58,719 --> 00:38:03,000 Speaker 1: want to distinguish a second level of scientific discovery, and 569 00:38:03,080 --> 00:38:05,440 Speaker 1: this is the one I'm interested in. I'll call this 570 00:38:05,880 --> 00:38:11,120 Speaker 1: level two, and that is science that requires conceptualization to 571 00:38:11,239 --> 00:38:14,640 Speaker 1: get to the next step, not just remixing what's already there. 572 00:38:15,200 --> 00:38:20,759 Speaker 1: Conceptualization like when the young Albert Einstein imagined something that 573 00:38:20,800 --> 00:38:23,560 Speaker 1: he had never seen before. He asked himself, what would 574 00:38:23,600 --> 00:38:25,640 Speaker 1: it be like if I could catch up with a 575 00:38:25,760 --> 00:38:29,040 Speaker 1: beam of light and ride it like a surfer riding 576 00:38:29,080 --> 00:38:32,960 Speaker 1: a wave. And this is how he derived the special 577 00:38:33,080 --> 00:38:36,919 Speaker 1: theory of relativity. This isn't something he looked up and 578 00:38:36,960 --> 00:38:42,200 Speaker 1: found three facts that clicked together. He imagined, He asked 579 00:38:42,400 --> 00:38:45,720 Speaker 1: new questions. He tried out a new model of the world, 580 00:38:46,280 --> 00:38:49,080 Speaker 1: one in which time runs differently depending on how fast 581 00:38:49,120 --> 00:38:52,279 Speaker 1: you're going, and then he worked backwards to see if 582 00:38:52,280 --> 00:38:56,800 Speaker 1: that model could work. Or consider when Charles Darwin thought 583 00:38:56,840 --> 00:38:59,440 Speaker 1: about the species that he saw around him, and he 584 00:38:59,520 --> 00:39:02,480 Speaker 1: imagined and all the species that he didn't see but 585 00:39:02,560 --> 00:39:05,959 Speaker 1: who might have existed, and he was able to put 586 00:39:06,000 --> 00:39:10,400 Speaker 1: together a new mental model in which most species don't 587 00:39:10,400 --> 00:39:14,560 Speaker 1: make it and we only see those whose mutations cause 588 00:39:14,680 --> 00:39:19,799 Speaker 1: survival advantages or reproductive advantages. These weren't facts that he 589 00:39:19,920 --> 00:39:22,960 Speaker 1: just collected from some papers. He was trying out a 590 00:39:23,040 --> 00:39:27,480 Speaker 1: new model of the world. Now, this kind of science 591 00:39:27,560 --> 00:39:31,240 Speaker 1: isn't just for the big giant stuff. Most meaningful science 592 00:39:31,600 --> 00:39:36,160 Speaker 1: is actually driven by this kind of imagination of new models. 593 00:39:37,320 --> 00:39:40,000 Speaker 1: Just as one example, I recently did an episode about 594 00:39:40,200 --> 00:39:43,160 Speaker 1: whether time runs in slow motion when you're in fear 595 00:39:43,640 --> 00:39:46,960 Speaker 1: for your life. And so when I wondered about this question, 596 00:39:47,600 --> 00:39:51,239 Speaker 1: I realized there were two hypotheses that might explain it, 597 00:39:51,560 --> 00:39:55,000 Speaker 1: and I thought of an experiment to discriminate those two hypotheses. 598 00:39:55,280 --> 00:39:58,600 Speaker 1: And then we built a wristband that flashes information at 599 00:39:58,640 --> 00:40:01,800 Speaker 1: a particular speed and had people wear and we dropped 600 00:40:01,840 --> 00:40:04,120 Speaker 1: them from one hundred and fifty foot tall tower into 601 00:40:04,160 --> 00:40:08,560 Speaker 1: a net below. A large language model presumably couldn't do 602 00:40:08,680 --> 00:40:13,200 Speaker 1: that because it's just playing statistical word games. And unless 603 00:40:13,200 --> 00:40:16,040 Speaker 1: someone had thought of that experiment and written it down, 604 00:40:16,840 --> 00:40:20,799 Speaker 1: JATGPT would never say, Okay, here's a new framework, and 605 00:40:20,840 --> 00:40:23,160 Speaker 1: how we can design an experiment to put this to 606 00:40:23,239 --> 00:40:26,040 Speaker 1: the test. So this is what I wanted to find 607 00:40:26,120 --> 00:40:30,840 Speaker 1: as the most meaningful test for a human level of intelligence. 608 00:40:31,360 --> 00:40:36,439 Speaker 1: When AI can do science in this way, generating new 609 00:40:36,520 --> 00:40:41,160 Speaker 1: ideas and frameworks, not just clicking facts together, then we 610 00:40:41,239 --> 00:40:47,560 Speaker 1: will have matched human intelligence. And I just want to 611 00:40:47,560 --> 00:40:49,800 Speaker 1: take one more angle on this to make the picture clear. 612 00:40:50,360 --> 00:40:54,360 Speaker 1: The way a scientist reads a journal paper is not 613 00:40:54,600 --> 00:40:58,640 Speaker 1: simply by correlating words and extracting keywords, although that might 614 00:40:58,680 --> 00:41:01,440 Speaker 1: be part of it, but also by realizing what was 615 00:41:01,640 --> 00:41:05,839 Speaker 1: not said. Why did the authors cut off the X 616 00:41:05,920 --> 00:41:09,520 Speaker 1: axis here at thirty What if they had extended this graph, 617 00:41:09,560 --> 00:41:12,960 Speaker 1: would the line have reversed in its trend? And why 618 00:41:12,960 --> 00:41:16,080 Speaker 1: didn't the authors mention the hypothesis of Smith at all? 619 00:41:16,640 --> 00:41:19,799 Speaker 1: And does that graph look too perfect? You know, one 620 00:41:19,800 --> 00:41:23,840 Speaker 1: of my mentors, Francis Crick, operated under the assumption that 621 00:41:23,920 --> 00:41:26,960 Speaker 1: he should disbelieve twenty five percent of what he read 622 00:41:27,000 --> 00:41:30,640 Speaker 1: in the literature. Is this because of fraud or error, 623 00:41:30,840 --> 00:41:34,960 Speaker 1: or statistical fluctuations or manipulation or the waste basket effect? 624 00:41:35,000 --> 00:41:38,319 Speaker 1: Who cares? The bottom line is that the literature is 625 00:41:38,560 --> 00:41:42,560 Speaker 1: rife with errors, and depending on the field, some estimates 626 00:41:42,840 --> 00:41:49,320 Speaker 1: put the inreproducibility at fifty percent. So when scientists read papers, 627 00:41:49,920 --> 00:41:53,600 Speaker 1: they know this just as Francis Crick did they read 628 00:41:53,800 --> 00:41:58,120 Speaker 1: in an entirely different manner than Google Translate or Watson 629 00:41:58,360 --> 00:42:03,160 Speaker 1: or chat shept or any of the correlational methods they extrapolate. 630 00:42:03,600 --> 00:42:07,120 Speaker 1: They read the paper and wonder about other possibilities. They 631 00:42:07,200 --> 00:42:10,920 Speaker 1: chew on what's missing, They envision the next step. They 632 00:42:10,960 --> 00:42:15,160 Speaker 1: think of the next experiment that could confirm or disconfirm 633 00:42:15,200 --> 00:42:19,080 Speaker 1: the hypotheses and the frameworks in the paper. To my mind, 634 00:42:19,239 --> 00:42:21,920 Speaker 1: the meaningful goal of AI is not going to be 635 00:42:21,960 --> 00:42:25,840 Speaker 1: found in number crunching and looking for facts that click together. 636 00:42:26,400 --> 00:42:29,680 Speaker 1: It's going to often be something else. It's going to 637 00:42:29,719 --> 00:42:35,000 Speaker 1: require an AI that learns how humans think, how they behave, 638 00:42:35,200 --> 00:42:38,640 Speaker 1: what they don't say, what they didn't think of, what 639 00:42:38,680 --> 00:42:42,439 Speaker 1: they misthought about, what they should think about. And one 640 00:42:42,480 --> 00:42:45,480 Speaker 1: more thing. I should note that these different levels I've outlined, 641 00:42:45,880 --> 00:42:50,399 Speaker 1: from fitting facts together versus imagining new world models, they're 642 00:42:50,440 --> 00:42:54,759 Speaker 1: probably gonna end up with blurry boundaries. So maybe chat 643 00:42:54,840 --> 00:42:58,680 Speaker 1: gpt will come up with something, and you won't always 644 00:42:58,880 --> 00:43:03,640 Speaker 1: know whether it's piecing together a few disparate pieces in 645 00:43:03,719 --> 00:43:08,000 Speaker 1: the literature what I'm calling level one, or whether it's 646 00:43:08,640 --> 00:43:12,080 Speaker 1: come up with something that is truly a new world 647 00:43:12,200 --> 00:43:15,600 Speaker 1: model that's not a simple clicking together, but a genuine 648 00:43:16,080 --> 00:43:19,640 Speaker 1: process of generating a new framework to explain the data. 649 00:43:19,760 --> 00:43:23,719 Speaker 1: So distinguishing the levels of discovery is probably not going 650 00:43:23,800 --> 00:43:26,439 Speaker 1: to be an easy task with a bright line between them, 651 00:43:26,880 --> 00:43:30,200 Speaker 1: but I think it will clarify some things to make 652 00:43:30,239 --> 00:43:34,640 Speaker 1: this distinction. And last thing, I don't necessarily know that 653 00:43:34,640 --> 00:43:38,360 Speaker 1: there's something magical and ineffable about the way that humans 654 00:43:38,400 --> 00:43:42,440 Speaker 1: do this. Presumably we're running algorithms too, it's just that 655 00:43:42,480 --> 00:43:46,640 Speaker 1: they're running on self configuring wetwear. I have seen tens 656 00:43:46,640 --> 00:43:49,719 Speaker 1: of thousands of science experiments in my career, so I 657 00:43:49,840 --> 00:43:53,120 Speaker 1: know the process of asking a question and figuring out 658 00:43:53,480 --> 00:43:55,839 Speaker 1: what we'll put it to the test. So we may 659 00:43:55,880 --> 00:43:57,960 Speaker 1: get to level two, and it may be sooner than 660 00:43:58,000 --> 00:44:01,280 Speaker 1: we expect, but I just want to be that right now, 661 00:44:01,640 --> 00:44:04,840 Speaker 1: we have not figured out the human algorithms. So the 662 00:44:05,040 --> 00:44:08,920 Speaker 1: current version of AI, as massively impressive as it is, 663 00:44:09,640 --> 00:44:14,240 Speaker 1: does not do level two scientific problem solving. And that's 664 00:44:14,280 --> 00:44:17,120 Speaker 1: when we're going to know that we've crossed a new 665 00:44:17,239 --> 00:44:21,839 Speaker 1: kind of line into a machine that is truly intelligent. 666 00:44:22,480 --> 00:44:25,760 Speaker 1: So let's wrap up. At least for now, humans still 667 00:44:25,800 --> 00:44:27,960 Speaker 1: have to do the science, by which I mean the 668 00:44:28,080 --> 00:44:32,080 Speaker 1: conceptual work, wherein we take a framework for understanding the 669 00:44:32,080 --> 00:44:35,640 Speaker 1: world and we rethink it, and we mentally simulate whether 670 00:44:36,000 --> 00:44:39,280 Speaker 1: a new model of the world could explain the observed data, 671 00:44:39,719 --> 00:44:41,359 Speaker 1: and we come up with a way to test that 672 00:44:41,440 --> 00:44:44,759 Speaker 1: new model. It's not just searching for facts. So I'm 673 00:44:44,760 --> 00:44:47,120 Speaker 1: definitely not saying we won't get to the next level 674 00:44:47,320 --> 00:44:51,000 Speaker 1: where AI can conceptualize things and predict forward and build 675 00:44:51,040 --> 00:44:53,640 Speaker 1: new knowledge. This might be a week from now, or 676 00:44:53,680 --> 00:44:56,080 Speaker 1: it might be a century from now. Who knows how 677 00:44:56,120 --> 00:44:57,960 Speaker 1: hard a problem that's going to turn out to be. 678 00:44:58,360 --> 00:45:00,479 Speaker 1: But I want us to be clear eyed on where 679 00:45:00,520 --> 00:45:05,239 Speaker 1: we are right now, because sometimes in the blindingly impressive 680 00:45:05,360 --> 00:45:08,480 Speaker 1: light of what current AI is doing, it can be 681 00:45:08,560 --> 00:45:12,760 Speaker 1: difficult to see what's missing and where we might be heading. 682 00:45:17,200 --> 00:45:19,719 Speaker 1: That's all for this week. To find out more and 683 00:45:19,800 --> 00:45:22,520 Speaker 1: to share your thoughts, head over to Eagleman dot com 684 00:45:22,520 --> 00:45:26,520 Speaker 1: slash podcasts, and you can also watch full episodes of 685 00:45:26,600 --> 00:45:30,239 Speaker 1: Inner Cosmos on YouTube. Subscribe to my channel so you 686 00:45:30,280 --> 00:45:33,560 Speaker 1: can follow along each week for new updates. I'd love 687 00:45:33,600 --> 00:45:37,319 Speaker 1: to hear your questions, so please send those to podcast 688 00:45:37,480 --> 00:45:40,400 Speaker 1: at eagleman dot com and I will do a special 689 00:45:40,440 --> 00:45:44,279 Speaker 1: episode where I answer questions until next time. I'm David 690 00:45:44,320 --> 00:45:46,719 Speaker 1: Eagleman and this is Inner Cosmos.