1 00:00:00,080 --> 00:00:02,120 Speaker 1: Hey, this is David Eagleman and this past week was 2 00:00:02,120 --> 00:00:04,200 Speaker 1: my birthday, so I took a week off. So I'm 3 00:00:04,240 --> 00:00:06,800 Speaker 1: going to run an episode that I did earlier, episode 4 00:00:06,880 --> 00:00:11,680 Speaker 1: number seven. This is called is AI actually intelligent? And 5 00:00:11,760 --> 00:00:14,400 Speaker 1: how would we know if it gets there? This episode 6 00:00:14,440 --> 00:00:16,520 Speaker 1: is from one year ago, but as time goes on 7 00:00:16,600 --> 00:00:21,000 Speaker 1: this becomes more and more relevant, So please enjoy and 8 00:00:21,040 --> 00:00:23,680 Speaker 1: I will see you next week with a new episode. 9 00:00:28,440 --> 00:00:33,160 Speaker 1: Modern AI is blowing everybody's mind. But is it intelligent 10 00:00:33,760 --> 00:00:36,480 Speaker 1: in the same way as the human brain? And could 11 00:00:36,560 --> 00:00:41,400 Speaker 1: AI reach sentience? And how would we know when it 12 00:00:41,440 --> 00:00:47,320 Speaker 1: gets there? Welcome to Inner Cosmos with me, David Eagleman. 13 00:00:48,120 --> 00:00:52,040 Speaker 1: I'm a neuroscientist and an author at Stanford University, and 14 00:00:52,080 --> 00:00:56,640 Speaker 1: I've spent my whole career studying the intersection between how 15 00:00:56,640 --> 00:01:04,520 Speaker 1: the brain works and how we experience life. Like most 16 00:01:04,640 --> 00:01:10,800 Speaker 1: brain researchers, I've been obsessed with questions of intelligence and consciousness. 17 00:01:11,480 --> 00:01:15,720 Speaker 1: How do these arise from collections of billions of cells 18 00:01:15,720 --> 00:01:20,680 Speaker 1: in our brains? And could intelligence and consciousness arise in 19 00:01:20,800 --> 00:01:25,560 Speaker 1: artificial brains? Say on chat GPT. Those are the questions 20 00:01:25,560 --> 00:01:28,800 Speaker 1: that we're going to attack today. Early efforts to figure 21 00:01:28,800 --> 00:01:31,920 Speaker 1: out the brain, looked at all the billions of cells 22 00:01:32,000 --> 00:01:35,640 Speaker 1: and the trillions of connections, and said, look, what if 23 00:01:35,640 --> 00:01:39,440 Speaker 1: we just think of each cell as a unit, and 24 00:01:39,640 --> 00:01:43,440 Speaker 1: each unit is connected to other units and where they connect, 25 00:01:43,920 --> 00:01:46,440 Speaker 1: which is called the sinnapps, or one cell gives a 26 00:01:46,440 --> 00:01:48,760 Speaker 1: little signal to the next cell. What if we just 27 00:01:48,880 --> 00:01:52,960 Speaker 1: looked at that like a simple connection that has a 28 00:01:53,000 --> 00:01:56,920 Speaker 1: strength between zero and one, or zero means there's no connection, 29 00:01:57,400 --> 00:02:00,640 Speaker 1: and one means it's the strongest possible connection. So this 30 00:02:00,840 --> 00:02:06,280 Speaker 1: was a massive oversimplification of the very complicated biology, but 31 00:02:06,600 --> 00:02:10,920 Speaker 1: it allowed people to start thinking about networks and writing 32 00:02:10,960 --> 00:02:15,080 Speaker 1: down different ways that you could put artificial neural networks together. 33 00:02:15,480 --> 00:02:17,720 Speaker 1: And for more than fifty years now people have been 34 00:02:17,760 --> 00:02:22,440 Speaker 1: doing research to show how artificial neural networks can do 35 00:02:22,560 --> 00:02:26,280 Speaker 1: really cool things. It's a totally new kind of way 36 00:02:26,280 --> 00:02:29,240 Speaker 1: of doing computation. So you've got these units, and you've 37 00:02:29,240 --> 00:02:32,440 Speaker 1: got these connections between them, and you've change the strength 38 00:02:32,480 --> 00:02:36,560 Speaker 1: of the connections and information flows through the network in 39 00:02:36,600 --> 00:02:41,000 Speaker 1: different ways. Now, my colleagues and I have long pointed 40 00:02:41,000 --> 00:02:44,799 Speaker 1: out the ways in which biological brands are different and 41 00:02:44,840 --> 00:02:49,560 Speaker 1: how artificial neural networks just push around numbers and play 42 00:02:49,600 --> 00:02:55,520 Speaker 1: statistical tricks. But we're entering a revolution right now. Large 43 00:02:55,600 --> 00:03:00,960 Speaker 1: language models like GPT four or BARD consume trillions of 44 00:03:01,000 --> 00:03:05,600 Speaker 1: words on the Internet and they figure out probabilistically which 45 00:03:05,680 --> 00:03:08,919 Speaker 1: word is going to come next given the massive context 46 00:03:08,919 --> 00:03:12,720 Speaker 1: of all the words that have come before. So these networks, 47 00:03:12,840 --> 00:03:16,040 Speaker 1: as I talked about on the previous episode, are showing 48 00:03:16,160 --> 00:03:23,080 Speaker 1: incredible successes in everything from writing to art, to coding 49 00:03:23,600 --> 00:03:27,920 Speaker 1: to generating three dimensional worlds. They're changing everything, and they're 50 00:03:27,960 --> 00:03:31,679 Speaker 1: doing so at a pace that we've never seen before, 51 00:03:31,840 --> 00:03:35,640 Speaker 1: and in fact, the entire history of humankind has never 52 00:03:35,680 --> 00:03:39,800 Speaker 1: seen before. And there are all the societal questions that 53 00:03:39,800 --> 00:03:43,560 Speaker 1: everyone's starting to wrestle with right now, like the massive 54 00:03:44,120 --> 00:03:49,800 Speaker 1: potential for displacement of human jobs. But today I want 55 00:03:49,840 --> 00:03:52,760 Speaker 1: to zoom in on a question that has captured the 56 00:03:52,800 --> 00:03:58,600 Speaker 1: imagination of scientists and philosophers and the general public. Could 57 00:03:58,720 --> 00:04:05,320 Speaker 1: aim alive in some way, like become conscious or sentient. Now, 58 00:04:05,400 --> 00:04:08,080 Speaker 1: there are lots of ways to think about this. We 59 00:04:08,160 --> 00:04:13,960 Speaker 1: can ask whether AI can possess meaningful intelligence, or we 60 00:04:14,000 --> 00:04:17,560 Speaker 1: can ask if it is sentient, which means the ability 61 00:04:17,600 --> 00:04:22,039 Speaker 1: to feel or perceive things, particularly in terms of sensations 62 00:04:22,080 --> 00:04:24,840 Speaker 1: like pleasure and pain and emotions. Or we can ask 63 00:04:25,120 --> 00:04:29,200 Speaker 1: whether it is conscious, which involves being aware of one's 64 00:04:29,240 --> 00:04:33,080 Speaker 1: self and one's surrounding. Now, there are specific and important 65 00:04:33,200 --> 00:04:37,159 Speaker 1: differences between these questions, but really I don't care for 66 00:04:37,279 --> 00:04:41,280 Speaker 1: the present conversation. The question we're asking here is is 67 00:04:41,440 --> 00:04:45,680 Speaker 1: chat GPT just zeros and ones moving around through transistors 68 00:04:46,320 --> 00:04:50,359 Speaker 1: like a giant garage door opener. Or is it thinking? 69 00:04:50,440 --> 00:04:54,000 Speaker 1: Is it having some sort of experience? Is it having 70 00:04:54,040 --> 00:04:57,920 Speaker 1: a private inner life like the type that we humans have. 71 00:04:58,560 --> 00:05:02,720 Speaker 1: As we think about the possible of sentient AI, we 72 00:05:02,760 --> 00:05:07,240 Speaker 1: immediately find ourselves facing really deep ethical questions, the main 73 00:05:07,279 --> 00:05:11,600 Speaker 1: one being if we were to create a machine with consciousness, 74 00:05:11,920 --> 00:05:15,760 Speaker 1: what responsibility do we have to treat it as a 75 00:05:15,839 --> 00:05:18,920 Speaker 1: living being? Would you be able to turn it off 76 00:05:18,920 --> 00:05:21,080 Speaker 1: when you're done with it at night or would that 77 00:05:21,120 --> 00:05:23,919 Speaker 1: be murder? And what if you turn it off and 78 00:05:23,960 --> 00:05:26,479 Speaker 1: then you turn it back on. Would that be like 79 00:05:26,560 --> 00:05:28,760 Speaker 1: the way that we go into a sleep state at 80 00:05:28,880 --> 00:05:32,000 Speaker 1: night where we're totally gone, and then we find ourselves 81 00:05:32,520 --> 00:05:34,440 Speaker 1: back online in the morning and we think, yeah, I'm 82 00:05:34,480 --> 00:05:38,320 Speaker 1: the same person, but I guess eight hours just disappeared. Anyway, 83 00:05:38,360 --> 00:05:42,120 Speaker 1: more generally, would we feel obligated to treat it the 84 00:05:42,120 --> 00:05:47,520 Speaker 1: way we treat a sentient fellow human. With our current laptops, 85 00:05:47,600 --> 00:05:50,560 Speaker 1: we're used to saying, sure, I can sell it, I 86 00:05:50,600 --> 00:05:54,280 Speaker 1: can trade it, I can upgrade it. But what happens 87 00:05:54,320 --> 00:05:58,240 Speaker 1: when we reach sentient machines? Can we still do this 88 00:05:58,880 --> 00:06:01,760 Speaker 1: or would it somehow be like putting a child up 89 00:06:01,760 --> 00:06:04,919 Speaker 1: for adoption or giving your pet away? Things that we 90 00:06:05,000 --> 00:06:08,159 Speaker 1: don't take lately. And eventually we're going to have entire 91 00:06:08,320 --> 00:06:13,919 Speaker 1: legal precedence built around the question of AI rights and responsibilities. 92 00:06:14,360 --> 00:06:16,880 Speaker 1: So that's why today I want to talk about these 93 00:06:16,960 --> 00:06:21,800 Speaker 1: issues of intelligence and sentience. Does an AI like chat 94 00:06:21,880 --> 00:06:28,200 Speaker 1: gpt experience anything when chat gpt writes a poem? Does 95 00:06:28,240 --> 00:06:32,479 Speaker 1: it appreciate the beauty when it types out a joke? 96 00:06:32,560 --> 00:06:36,320 Speaker 1: Does it find itself amused and chuckling to itself. Let's 97 00:06:36,320 --> 00:06:39,400 Speaker 1: start with a guy named Blake Lemoyne who was a 98 00:06:39,640 --> 00:06:43,520 Speaker 1: programmer at Google and in June of twenty twenty two, 99 00:06:43,560 --> 00:06:49,080 Speaker 1: he was exchanging messages with a version of Google's conversational AI, 100 00:06:49,160 --> 00:06:52,040 Speaker 1: which was called Lambda at the time. So he asked 101 00:06:52,120 --> 00:06:55,760 Speaker 1: Namda for an example of what it was afraid of 102 00:06:56,320 --> 00:06:59,839 Speaker 1: and it gave him this very eloquent response about how 103 00:07:00,200 --> 00:07:04,240 Speaker 1: was afraid of being turned off, So he wrote an 104 00:07:04,240 --> 00:07:07,960 Speaker 1: internal memo to Google leadership than which he said, I 105 00:07:08,000 --> 00:07:12,600 Speaker 1: think this AI is sentient. And the leadership at Google 106 00:07:12,720 --> 00:07:17,880 Speaker 1: felt that this was an entirely unsubstantiated claim, and so 107 00:07:17,920 --> 00:07:20,280 Speaker 1: they made the decision to fire him for what they 108 00:07:20,280 --> 00:07:23,520 Speaker 1: took as an inappropriate conclusion that just didn't have enough 109 00:07:23,560 --> 00:07:28,160 Speaker 1: evidence beyond his intuition to qualify for raising the alarm 110 00:07:28,240 --> 00:07:31,520 Speaker 1: on this. So obviously this immediately fired up the news 111 00:07:31,600 --> 00:07:35,600 Speaker 1: cycles and the rumor mill and conspiracy theorists thought, Wait, 112 00:07:35,680 --> 00:07:39,320 Speaker 1: if AI isn't conscious, why would they fire him. They're 113 00:07:39,440 --> 00:07:41,840 Speaker 1: firing of him as all the evidence I need to 114 00:07:41,880 --> 00:07:46,480 Speaker 1: tell me that AI is sentient? Okay, but is it? 115 00:07:47,040 --> 00:07:50,160 Speaker 1: What does it mean to be conscious or sentient? How 116 00:07:50,440 --> 00:07:54,080 Speaker 1: the heck would we know when we have created something 117 00:07:54,120 --> 00:07:57,280 Speaker 1: that gets there? How do we know whether the AI 118 00:07:57,400 --> 00:07:59,840 Speaker 1: is sentient or instead whether humans are fooling them so 119 00:08:00,360 --> 00:08:03,240 Speaker 1: into believing that it is well. One way to make 120 00:08:03,280 --> 00:08:07,000 Speaker 1: this distinction would be to see if the AI could 121 00:08:07,280 --> 00:08:11,320 Speaker 1: conceptualize things, if it could take lots of words and 122 00:08:11,360 --> 00:08:15,600 Speaker 1: facts on the web and abstract those to some bigger idea. 123 00:08:16,200 --> 00:08:18,600 Speaker 1: So one of my friends here in Silicon Valley said 124 00:08:18,640 --> 00:08:21,800 Speaker 1: to me the other day, I asked chat gpt the 125 00:08:21,840 --> 00:08:26,480 Speaker 1: following question, Take a capital letter D and turn it 126 00:08:26,560 --> 00:08:30,480 Speaker 1: flat side down. Now take the letter J and slide 127 00:08:30,520 --> 00:08:35,040 Speaker 1: it underneath. What does that look like? And chat gpt said, 128 00:08:35,679 --> 00:08:38,959 Speaker 1: and umbrella. And my friend was blown away by this, 129 00:08:39,160 --> 00:08:44,320 Speaker 1: and he said, this is conceptualization. It's just done three 130 00:08:44,360 --> 00:08:50,800 Speaker 1: dimensional reasoning. There's something deeper happening here than just parenting words. 131 00:08:51,240 --> 00:08:54,280 Speaker 1: But I pointed out to him that this particular question 132 00:08:54,360 --> 00:08:57,080 Speaker 1: about the D on its side and the J underneath 133 00:08:57,120 --> 00:09:00,960 Speaker 1: it is one of the oldest examples in psychology classes 134 00:09:01,040 --> 00:09:04,800 Speaker 1: when talking about visual imagery, and it's on the Internet 135 00:09:04,880 --> 00:09:07,560 Speaker 1: in thousands of places, so of course it got it right. 136 00:09:08,120 --> 00:09:11,760 Speaker 1: It's just parroting the answer because it has read the 137 00:09:11,880 --> 00:09:15,240 Speaker 1: question and it has read the answer before. So it's 138 00:09:15,280 --> 00:09:19,360 Speaker 1: not always easy to determine what's going on for these 139 00:09:19,480 --> 00:09:23,400 Speaker 1: models in terms of whether some human somewhere has discussed 140 00:09:23,480 --> 00:09:26,160 Speaker 1: this point and written down the answer. And the general 141 00:09:26,240 --> 00:09:30,199 Speaker 1: story is that with trillions of words written by humans 142 00:09:30,240 --> 00:09:35,520 Speaker 1: over centuries, there are many things beyond your capacity to 143 00:09:35,679 --> 00:09:38,440 Speaker 1: read them or to even imagine that they've been written 144 00:09:38,480 --> 00:09:42,400 Speaker 1: down before, but maybe they have. If any human has 145 00:09:42,520 --> 00:09:47,800 Speaker 1: discussed a question before has conceptualized something, then chat GPT 146 00:09:48,040 --> 00:09:52,240 Speaker 1: can find that and mimic that. But that's not conceptualization. 147 00:09:52,880 --> 00:09:55,960 Speaker 1: Chat GPT is doing a thousand amazing things, and we 148 00:09:56,120 --> 00:10:00,360 Speaker 1: have an enormous amount to learn about it. But we 149 00:10:00,400 --> 00:10:05,240 Speaker 1: shouldn't let ourselves get fooled and mesmerized into believing that 150 00:10:05,280 --> 00:10:08,319 Speaker 1: it's doing something more than it is. And our ability 151 00:10:08,360 --> 00:10:12,480 Speaker 1: to get fooled is not only about the massive statistics 152 00:10:12,520 --> 00:10:16,160 Speaker 1: of what it takes in. There are other examples of 153 00:10:16,559 --> 00:10:22,080 Speaker 1: seeming sentience that result from the reinforcement learning that it 154 00:10:22,120 --> 00:10:26,080 Speaker 1: does with humans. So here's what that means. The network 155 00:10:26,160 --> 00:10:31,160 Speaker 1: generates lots of sentences and thousands of humans are involved 156 00:10:31,240 --> 00:10:33,800 Speaker 1: in giving it feedback, like a thumbs up or a 157 00:10:33,880 --> 00:10:37,800 Speaker 1: thumbs down, to say whether they appreciated the answer, whether 158 00:10:37,800 --> 00:10:41,679 Speaker 1: they thought that was a good answer. So, because humans 159 00:10:41,760 --> 00:10:46,040 Speaker 1: are giving reward to the machine, sometimes that pushes things 160 00:10:46,559 --> 00:10:51,120 Speaker 1: in weird directions that can be mistaken for sentience. For example, 161 00:10:51,280 --> 00:10:56,640 Speaker 1: scholars have shown that reinforcement learning with humans makes networks 162 00:10:56,800 --> 00:11:01,040 Speaker 1: more likely to say, don't turn me off, just like 163 00:11:01,200 --> 00:11:04,959 Speaker 1: Blake had heard but don't mistake this for sentience. It's 164 00:11:05,000 --> 00:11:08,400 Speaker 1: only a sign that the machine is saying this because 165 00:11:08,440 --> 00:11:11,160 Speaker 1: some of the human participants gave it a thumbs up 166 00:11:11,400 --> 00:11:14,640 Speaker 1: when the large language model said this before, and so 167 00:11:14,760 --> 00:11:18,640 Speaker 1: it learned to do this again. The fact is, it's 168 00:11:18,679 --> 00:11:22,480 Speaker 1: sometimes hard to know why. Sometimes we see an answer 169 00:11:22,559 --> 00:11:27,400 Speaker 1: that feels very impressive. But we'd agree that pulling text 170 00:11:27,440 --> 00:11:30,480 Speaker 1: from the Internet and parroting it back is not by 171 00:11:30,520 --> 00:11:36,960 Speaker 1: itself intelligence or sentience. Chat GPT presumably has no idea 172 00:11:37,080 --> 00:11:40,160 Speaker 1: of what it's saying, whether that's a poem or a 173 00:11:40,400 --> 00:11:45,600 Speaker 1: terrorist manifesto, or instructions for building a spaceship or a 174 00:11:45,640 --> 00:11:50,920 Speaker 1: heartbreaking story about an orphaned child. Chat GPT doesn't know, 175 00:11:51,000 --> 00:11:56,480 Speaker 1: and it doesn't care. It's words in and statistical correlations out. 176 00:11:56,880 --> 00:12:01,199 Speaker 1: And in fact, there has been a fundamental philosophical point 177 00:12:01,360 --> 00:12:04,600 Speaker 1: made about this in the nineteen eighties when the philosopher 178 00:12:04,760 --> 00:12:09,040 Speaker 1: John Surrele was wondering about this question of whether a 179 00:12:09,160 --> 00:12:13,880 Speaker 1: computer could ever be programmed so that it has a mind, 180 00:12:14,160 --> 00:12:16,280 Speaker 1: and he came up with a thought experiment that he 181 00:12:16,360 --> 00:12:20,000 Speaker 1: called the Chinese room argument, and it goes like this, 182 00:12:22,040 --> 00:12:26,440 Speaker 1: I am locked in a room and questions are passed 183 00:12:26,440 --> 00:12:30,199 Speaker 1: to me through a small letter slot, and these messages 184 00:12:30,240 --> 00:12:33,320 Speaker 1: are written only in Chinese, and I don't speak Chinese. 185 00:12:33,400 --> 00:12:37,040 Speaker 1: I have no clue what's written on these pieces of paper. However, 186 00:12:37,240 --> 00:12:41,480 Speaker 1: inside this room, I have a library of books, and 187 00:12:41,520 --> 00:12:45,319 Speaker 1: they contain step by step instructions that tell me exactly 188 00:12:45,360 --> 00:12:48,520 Speaker 1: what to do with these symbols. So I look at 189 00:12:48,520 --> 00:12:52,240 Speaker 1: the grouping of symbols, and I simply follow steps in 190 00:12:52,320 --> 00:12:55,800 Speaker 1: the book to tell me what Chinese symbols to copy 191 00:12:55,880 --> 00:12:58,920 Speaker 1: down in response. So I write those on the slip 192 00:12:58,920 --> 00:13:01,760 Speaker 1: of paper. And when I pass the paper back out 193 00:13:01,800 --> 00:13:06,360 Speaker 1: of the slot. Now, when the Chinese speaker receives my 194 00:13:06,559 --> 00:13:10,400 Speaker 1: reply message, it makes perfect sense to her. It seems 195 00:13:10,920 --> 00:13:14,360 Speaker 1: as though whoever is in the room is answering her 196 00:13:14,440 --> 00:13:17,840 Speaker 1: questions perfectly, and therefore it seems obvious that the person 197 00:13:17,920 --> 00:13:23,199 Speaker 1: in the room must understand Chinese. I've fooled her, of course, 198 00:13:23,240 --> 00:13:26,160 Speaker 1: because I'm only following a set of instructions with no 199 00:13:26,400 --> 00:13:29,760 Speaker 1: understanding of what's going on. With enough time and with 200 00:13:29,800 --> 00:13:33,199 Speaker 1: a big enough set of instructions, I can answer almost 201 00:13:33,240 --> 00:13:37,679 Speaker 1: any question posed to me in Chinese. But I, the operator, 202 00:13:37,800 --> 00:13:42,400 Speaker 1: do not understand Chinese. I manipulate symbols all day long, 203 00:13:43,000 --> 00:13:48,760 Speaker 1: but I have no idea what the symbols mean. Now, 204 00:13:48,840 --> 00:13:53,240 Speaker 1: The philosopher John Searle argued, this is just what's happening 205 00:13:53,280 --> 00:13:57,560 Speaker 1: inside a computer. No matter how intelligent a program like 206 00:13:57,679 --> 00:14:01,800 Speaker 1: chat GPT seems to be, it's only following sets of 207 00:14:01,880 --> 00:14:08,880 Speaker 1: instructions to spit out answers. It's manipulating symbols without ever 208 00:14:09,280 --> 00:14:12,680 Speaker 1: really understanding what it's doing. Or think about what Google 209 00:14:12,800 --> 00:14:16,439 Speaker 1: is doing. When you send Google a query, it doesn't 210 00:14:16,520 --> 00:14:19,760 Speaker 1: understand your question or even its own answer. It simply 211 00:14:19,840 --> 00:14:24,160 Speaker 1: moves around zeros and ones and logicates and returns zeros 212 00:14:24,160 --> 00:14:26,880 Speaker 1: and ones to you. Or with a mind blowing program 213 00:14:26,920 --> 00:14:31,000 Speaker 1: like Google Translate, I can write a sentence in Russian 214 00:14:31,320 --> 00:14:35,400 Speaker 1: and it can return the translation in Amharic. But it's 215 00:14:35,560 --> 00:14:41,520 Speaker 1: all algorithmic. It's just symbol manipulation. Like the operator inside 216 00:14:41,520 --> 00:14:46,880 Speaker 1: the Chinese room, Google Translate doesn't understand anything about the sentence. 217 00:14:47,120 --> 00:14:51,520 Speaker 1: Nothing carries any meaning to it. So the Chinese room 218 00:14:51,600 --> 00:14:57,080 Speaker 1: argument suggests that AI that mimics human intelligence doesn't actually 219 00:14:57,200 --> 00:15:01,640 Speaker 1: understand what it's talking about. There's no meaning to anything, 220 00:15:01,720 --> 00:15:06,480 Speaker 1: CHATCHYPT says, and Serle used this thought experiment to argue 221 00:15:06,480 --> 00:15:10,920 Speaker 1: that there's something about human brains that won't be explained 222 00:15:10,960 --> 00:15:15,240 Speaker 1: if we simply analogize them to digital computers. There's a 223 00:15:15,400 --> 00:15:26,520 Speaker 1: gap between symbols that have no meaning and our conscious experience. Now, 224 00:15:27,240 --> 00:15:30,960 Speaker 1: there's an ongoing debate about the interpretation of the Chinese 225 00:15:31,040 --> 00:15:35,760 Speaker 1: room argument, but however one construes it, the argument exposes 226 00:15:36,280 --> 00:15:40,360 Speaker 1: the difficulty in the mystery of how zeros and ones 227 00:15:40,560 --> 00:15:44,920 Speaker 1: would ever come to equal our experience of being alive 228 00:15:45,040 --> 00:15:47,760 Speaker 1: in the world. Now, just to be very clear on 229 00:15:47,800 --> 00:15:51,880 Speaker 1: this point, we don't understand why we are conscious. There's 230 00:15:51,920 --> 00:15:54,040 Speaker 1: still a huge amount of work that has to be 231 00:15:54,080 --> 00:15:57,120 Speaker 1: done in biology to understand that. But this is just 232 00:15:57,160 --> 00:16:01,000 Speaker 1: to say that simply having zeros in one moving around 233 00:16:01,680 --> 00:16:06,560 Speaker 1: wouldn't by itself seem to be sufficient for conscious experience. 234 00:16:07,160 --> 00:16:10,520 Speaker 1: In other words, how do zeros and ones ever equal 235 00:16:10,640 --> 00:16:15,120 Speaker 1: the sting of a hot pepper, or the yellowness of 236 00:16:15,240 --> 00:16:19,720 Speaker 1: yellow or the beauty of a sunset. By the way, 237 00:16:19,760 --> 00:16:22,480 Speaker 1: I've covered the Chinese room argument in my TV show 238 00:16:22,600 --> 00:16:24,720 Speaker 1: The Brain, and if you're interested in that, I'll link 239 00:16:24,760 --> 00:16:28,960 Speaker 1: the video on Eagleman dot com slash podcast. Now, all 240 00:16:29,040 --> 00:16:31,840 Speaker 1: this is not a criticism of the approach of moving 241 00:16:31,960 --> 00:16:34,680 Speaker 1: zeros and ones around. But it is to point out 242 00:16:34,680 --> 00:16:39,000 Speaker 1: that we shouldn't confuse this type of Chinese room correlation 243 00:16:39,920 --> 00:16:45,040 Speaker 1: with real sentience or intelligence. And there's a deeper reason 244 00:16:45,120 --> 00:16:50,080 Speaker 1: to be suspicious too, because despite the incredible successes of 245 00:16:50,200 --> 00:16:54,480 Speaker 1: large language models, we also see that they sometimes make 246 00:16:54,880 --> 00:16:58,520 Speaker 1: decisions that expose the fact that they don't have any 247 00:16:58,600 --> 00:17:01,880 Speaker 1: meaningful model of the In other words, I think we 248 00:17:01,920 --> 00:17:05,480 Speaker 1: can gain some fast insight by paying attention to the 249 00:17:05,520 --> 00:17:08,840 Speaker 1: places where the AI is not working so well. So 250 00:17:08,920 --> 00:17:12,359 Speaker 1: I'll give three quick examples. The first has to do 251 00:17:12,440 --> 00:17:17,080 Speaker 1: with humor. AI has a very difficult time making an 252 00:17:17,119 --> 00:17:20,840 Speaker 1: original joke, and this is for a simple reason. To 253 00:17:21,000 --> 00:17:24,040 Speaker 1: make up a new joke, you need to know what 254 00:17:24,080 --> 00:17:27,760 Speaker 1: the ending is and then you work backwards to construct 255 00:17:27,880 --> 00:17:30,480 Speaker 1: the joke with red herrings so no one sees where 256 00:17:30,520 --> 00:17:33,399 Speaker 1: you're going and it happens at the way these large 257 00:17:33,480 --> 00:17:37,200 Speaker 1: language models work is all in the forward direction. They 258 00:17:37,240 --> 00:17:40,920 Speaker 1: decide what is the most probable word to come next, 259 00:17:41,160 --> 00:17:45,040 Speaker 1: So they're fine at parroting jokes back to us, but 260 00:17:45,119 --> 00:17:49,560 Speaker 1: they're total failures at building original jokes. And there's a 261 00:17:49,600 --> 00:17:52,240 Speaker 1: deeper point here as well. To build a joke, You 262 00:17:52,320 --> 00:17:56,440 Speaker 1: need to have some model, some idea of what will 263 00:17:56,440 --> 00:18:00,520 Speaker 1: be funny to a fellow human, what shared concept or 264 00:18:00,560 --> 00:18:04,200 Speaker 1: shared experience would make someone laugh. And for that, you 265 00:18:04,359 --> 00:18:07,959 Speaker 1: generally need to have the experience of a human life 266 00:18:08,000 --> 00:18:11,479 Speaker 1: with all of its joys and slings and arrows and 267 00:18:11,520 --> 00:18:14,199 Speaker 1: so on. And these large language models can do a 268 00:18:14,200 --> 00:18:18,120 Speaker 1: lot of things, but they don't have any model of 269 00:18:18,200 --> 00:18:22,680 Speaker 1: what it is to be a human. My second example 270 00:18:23,359 --> 00:18:25,920 Speaker 1: has to do with the flip side of making a joke, 271 00:18:25,960 --> 00:18:28,520 Speaker 1: which is getting a joke. And if you look carefully, 272 00:18:28,520 --> 00:18:31,639 Speaker 1: you will see how current AI always fails to catch 273 00:18:31,720 --> 00:18:34,359 Speaker 1: jokes that are thrown at it. It doesn't get jokes 274 00:18:34,400 --> 00:18:36,959 Speaker 1: because it doesn't have a model of what it is 275 00:18:37,000 --> 00:18:40,720 Speaker 1: to be a human. But this point goes beyond jokes. 276 00:18:41,119 --> 00:18:44,400 Speaker 1: One of the most remarkable feats of these large language 277 00:18:44,400 --> 00:18:49,440 Speaker 1: models is summarizing large texts, and in twenty twenty two, 278 00:18:49,520 --> 00:18:53,840 Speaker 1: open Ai announced how they could summarize entire books like 279 00:18:53,960 --> 00:18:57,000 Speaker 1: Alice in Wonderland. What it does is it generates a 280 00:18:57,040 --> 00:19:00,320 Speaker 1: summary of each chapter, and then it uses those after 281 00:19:00,359 --> 00:19:03,080 Speaker 1: summaries to make a summary of the whole book. So 282 00:19:03,200 --> 00:19:07,040 Speaker 1: for Alice in Wonderland, it generates the following. Alice falls 283 00:19:07,040 --> 00:19:09,399 Speaker 1: down a rabbit hole and grows to a giant size. 284 00:19:09,440 --> 00:19:12,919 Speaker 1: After drinking a mysterious bottle, she decides to focus on 285 00:19:13,119 --> 00:19:15,960 Speaker 1: growing back to her normal size and finding her way 286 00:19:16,000 --> 00:19:18,840 Speaker 1: into the garden. She meets the caterpillar, who tells her 287 00:19:18,880 --> 00:19:21,080 Speaker 1: that one side of a mushroom will make her grow taller, 288 00:19:21,359 --> 00:19:24,480 Speaker 1: the other side shorter. She eats the mushroom and returns 289 00:19:24,520 --> 00:19:27,240 Speaker 1: to her normal size. Alice attends a party with the 290 00:19:27,280 --> 00:19:30,800 Speaker 1: Mad Hatter and the march Hare. The Queen arrives and 291 00:19:30,920 --> 00:19:33,720 Speaker 1: orders the execution of the gardeners for making a mistake 292 00:19:33,800 --> 00:19:37,040 Speaker 1: with the roses. Alice saves them by putting them in 293 00:19:37,080 --> 00:19:39,760 Speaker 1: a flower pot. The King and Queen of Hearts preside 294 00:19:39,800 --> 00:19:42,760 Speaker 1: over a trial. The Queen gets angry and orders Alice 295 00:19:42,800 --> 00:19:45,680 Speaker 1: to be sentenced to death. Alice wakes up to find 296 00:19:45,680 --> 00:19:50,280 Speaker 1: her sister by her side. So that's pretty remarkable. It 297 00:19:50,320 --> 00:19:53,200 Speaker 1: took a whole book, and it was able to summarize 298 00:19:53,200 --> 00:19:56,520 Speaker 1: it down to a paragraph. But I kept reading these 299 00:19:56,560 --> 00:20:00,359 Speaker 1: text summaries carefully, and I got to the summary of 300 00:20:00,720 --> 00:20:04,040 Speaker 1: Act one of Romeo and Juliet, and here's what it says. 301 00:20:04,760 --> 00:20:08,440 Speaker 1: Romeo locks himself in his room, no longer in love 302 00:20:08,520 --> 00:20:11,840 Speaker 1: with rosalind Now, I think the engineers at open Ai 303 00:20:12,000 --> 00:20:14,879 Speaker 1: felt really satisfied with this summary. They thought it was 304 00:20:14,960 --> 00:20:17,280 Speaker 1: quite good, and my proof for this is that they 305 00:20:17,680 --> 00:20:21,800 Speaker 1: still display it proudly on their website. But I majored 306 00:20:21,880 --> 00:20:24,400 Speaker 1: in literature as an undergraduate, and I spend a lot 307 00:20:24,440 --> 00:20:27,560 Speaker 1: of time with shakespeare plays, and I immediately knew that 308 00:20:27,640 --> 00:20:32,240 Speaker 1: this summary was exactly wrong. The actual scene from Shakespeare 309 00:20:32,240 --> 00:20:38,000 Speaker 1: goes like this. His friend ben Voglio finds Romeo catatonically depressed, 310 00:20:38,440 --> 00:20:43,560 Speaker 1: and ben Volio says, what sadness lengthens Romeo's hours? And 311 00:20:43,640 --> 00:20:48,480 Speaker 1: Romeo says, not having that which having makes them short? 312 00:20:48,600 --> 00:20:52,560 Speaker 1: And ben Volio says in love, and Romeo says out 313 00:20:53,080 --> 00:20:56,399 Speaker 1: ben Reli says of love, and Romeo says out of 314 00:20:56,480 --> 00:21:00,199 Speaker 1: her favor, where I am in love? This this is 315 00:21:00,240 --> 00:21:05,720 Speaker 1: typical Shakespearean wordplay, where Romeo is expressing his grief of 316 00:21:05,760 --> 00:21:09,199 Speaker 1: being out of favor with Roslin, with whom he is 317 00:21:09,280 --> 00:21:12,120 Speaker 1: deeply in love. And when you read the play, it's 318 00:21:12,160 --> 00:21:16,560 Speaker 1: obvious that Romeo is not over Roslin. He's suffering over her. 319 00:21:16,600 --> 00:21:19,879 Speaker 1: He's almost suicidal. And this is an important piece of 320 00:21:19,920 --> 00:21:22,680 Speaker 1: the play, because the play is really about a young 321 00:21:22,720 --> 00:21:26,080 Speaker 1: man in love with the idea of being in love, 322 00:21:26,280 --> 00:21:29,639 Speaker 1: and that's why he later in the same act, falls 323 00:21:29,680 --> 00:21:33,600 Speaker 1: so hard into his relationship with Juliet, a relationship which 324 00:21:33,720 --> 00:21:36,840 Speaker 1: ends in their mutual suicide. By the way, as Friar 325 00:21:36,920 --> 00:21:41,760 Speaker 1: Lauren says of their relationship, these violent delights have violent ends. 326 00:21:42,240 --> 00:21:43,760 Speaker 1: And you get a bonus if you can tell me 327 00:21:43,800 --> 00:21:46,920 Speaker 1: where else you've heard that line more recently. Okay, anyway 328 00:21:46,960 --> 00:21:51,960 Speaker 1: back to the AI summary, The AI misses this wordplay entirely, 329 00:21:52,600 --> 00:21:57,960 Speaker 1: and it concludes that Romeo is out of love with Roslin. Again, 330 00:21:58,080 --> 00:22:01,480 Speaker 1: a human watching the play or reading the play immediately 331 00:22:01,520 --> 00:22:06,400 Speaker 1: gets that Romeo is making wordplay and his heartbroken over Roslin, 332 00:22:06,440 --> 00:22:10,000 Speaker 1: but the AI doesn't get that because it's reading words 333 00:22:10,119 --> 00:22:13,840 Speaker 1: only at a statistical level, not at a level of 334 00:22:13,920 --> 00:22:18,000 Speaker 1: understanding of what it is to be a human saying 335 00:22:18,240 --> 00:22:21,880 Speaker 1: those words. And that leads me to the third example, 336 00:22:22,320 --> 00:22:26,439 Speaker 1: which is the difficulty in understanding the physical world. So 337 00:22:26,560 --> 00:22:30,480 Speaker 1: consider a question like this, When President Biden walks into 338 00:22:30,520 --> 00:22:34,560 Speaker 1: a room, does his head come with him? So this 339 00:22:34,680 --> 00:22:38,119 Speaker 1: is famously difficult for AI to answer a question like this, 340 00:22:38,240 --> 00:22:42,200 Speaker 1: even though it's trivial for you because the AI doesn't 341 00:22:42,240 --> 00:22:46,639 Speaker 1: have an internal model of how everything physically hangs together 342 00:22:46,720 --> 00:22:49,320 Speaker 1: in the world. Last week, I was at the TED 343 00:22:49,400 --> 00:22:52,480 Speaker 1: conference and I heard a great talk by Yegin Choi, 344 00:22:52,880 --> 00:22:56,280 Speaker 1: and she was phrasing this problem as AI not having 345 00:22:56,760 --> 00:23:01,199 Speaker 1: common sense. She asked chat GPT the following question, it 346 00:23:01,280 --> 00:23:04,200 Speaker 1: takes six hours to dry six shirts in the sun, 347 00:23:04,640 --> 00:23:07,560 Speaker 1: how long does it take to dry thirty shirts? And 348 00:23:07,640 --> 00:23:11,399 Speaker 1: it answers thirty hours. Now you and I see that 349 00:23:11,440 --> 00:23:14,320 Speaker 1: the answer should be six hours, because we know the 350 00:23:14,359 --> 00:23:17,439 Speaker 1: sun doesn't care how many shirts are out there. But 351 00:23:17,560 --> 00:23:21,919 Speaker 1: chat GPT just doesn't get it because despite appearances, it 352 00:23:21,960 --> 00:23:25,840 Speaker 1: doesn't have a model of the world. And we've seen 353 00:23:25,880 --> 00:23:27,920 Speaker 1: this sort of thing for years. By the way, even 354 00:23:27,920 --> 00:23:32,879 Speaker 1: in mind blowingly impressive AI models that do image recognition, 355 00:23:32,920 --> 00:23:36,680 Speaker 1: they're so impressive in what they recognize, but then they'll 356 00:23:36,760 --> 00:23:40,679 Speaker 1: fail catastrophically. It's some easy picture making mistakes that a 357 00:23:40,760 --> 00:23:43,680 Speaker 1: human just wouldn't make. For example, there's one picture where 358 00:23:43,720 --> 00:23:46,280 Speaker 1: there's a boy holding a toothbrush and the AI says 359 00:23:46,720 --> 00:23:49,640 Speaker 1: it's a boy with a baseball bat. Okay, so there 360 00:23:49,640 --> 00:23:54,240 Speaker 1: are things that AI doesn't do that well. But that said, 361 00:23:54,280 --> 00:23:57,960 Speaker 1: there are other things that are mind blowing, things that 362 00:23:58,600 --> 00:24:01,360 Speaker 1: no one expected it to do. And this is why 363 00:24:01,400 --> 00:24:04,560 Speaker 1: I mentioned in my previous episode that we are in 364 00:24:04,640 --> 00:24:10,120 Speaker 1: an era of discovery more than just invention. Everyone's searching 365 00:24:10,200 --> 00:24:13,560 Speaker 1: and finding things that the AI can do that nobody 366 00:24:13,600 --> 00:24:17,160 Speaker 1: really expected or foresaw, including all the stuff that we're 367 00:24:17,160 --> 00:24:20,639 Speaker 1: now taking for granted, like oh, it can summarize books 368 00:24:20,720 --> 00:24:23,800 Speaker 1: or it can make art from text. And I want 369 00:24:23,840 --> 00:24:26,080 Speaker 1: to point out that a lot of the arguments that 370 00:24:26,119 --> 00:24:30,320 Speaker 1: people have been making about AI not being good at something, 371 00:24:30,520 --> 00:24:34,879 Speaker 1: these arguments have been changing rapidly. For example, just a 372 00:24:34,920 --> 00:24:38,000 Speaker 1: few months ago, people were arguing that AI would make 373 00:24:38,119 --> 00:24:41,080 Speaker 1: silly mistakes about things, and it couldn't really understand math 374 00:24:41,160 --> 00:24:45,119 Speaker 1: and would get math wrong and word problems. But in 375 00:24:45,160 --> 00:24:49,200 Speaker 1: a shockingly brief time, a lot of these shortcomings have 376 00:24:49,280 --> 00:24:53,000 Speaker 1: been mastered. So it's yet to be seen what challenges 377 00:24:53,119 --> 00:25:14,480 Speaker 1: will remain and for how long. So the evidence I've 378 00:25:14,520 --> 00:25:17,720 Speaker 1: presented so far is that AI doesn't have a great 379 00:25:17,800 --> 00:25:20,239 Speaker 1: model of what it's like to be human, but that 380 00:25:20,280 --> 00:25:25,600 Speaker 1: doesn't necessarily rule out that it has sentience or awareness, 381 00:25:25,760 --> 00:25:30,040 Speaker 1: even if it's of another flavor. It doesn't think like 382 00:25:30,080 --> 00:25:35,040 Speaker 1: a human, but maybe it stif thinks so is chat 383 00:25:35,080 --> 00:25:40,359 Speaker 1: GPT having some sort of experience? And how would we know? 384 00:25:42,119 --> 00:25:46,560 Speaker 1: In nineteen fifty, the brilliant mathematician and computer scientist Alan 385 00:25:46,680 --> 00:25:51,480 Speaker 1: Turing was asking this question, how could you determine whether 386 00:25:51,560 --> 00:25:56,600 Speaker 1: a machine exhibits human like intelligence? So he proposed an 387 00:25:56,640 --> 00:26:00,679 Speaker 1: experiment that he called the imitation game. You've got a 388 00:26:00,720 --> 00:26:05,840 Speaker 1: machine AI that's programmed to simulate human speech or conversation, 389 00:26:06,200 --> 00:26:08,800 Speaker 1: and you place it in a closed room, and in 390 00:26:08,840 --> 00:26:12,240 Speaker 1: a second room you have a real human, but the 391 00:26:12,280 --> 00:26:15,440 Speaker 1: doors are closed, so you don't know which room has 392 00:26:15,560 --> 00:26:19,360 Speaker 1: which machine or human. And now you are a person, 393 00:26:19,440 --> 00:26:24,359 Speaker 1: the evaluator, who communicates with both of them via a 394 00:26:24,560 --> 00:26:27,080 Speaker 1: computer terminal or I think of a nowadays like text 395 00:26:27,119 --> 00:26:31,840 Speaker 1: messaging with both of them. So you, the evaluator, engage 396 00:26:31,920 --> 00:26:35,600 Speaker 1: in a conversation with both closed rooms, one of which 397 00:26:35,640 --> 00:26:37,840 Speaker 1: has the machine and one the human, and your job 398 00:26:37,920 --> 00:26:40,879 Speaker 1: is simply to figure out which is which, which is 399 00:26:40,920 --> 00:26:43,160 Speaker 1: the machine and which is the human. And the only 400 00:26:43,280 --> 00:26:46,000 Speaker 1: thing that you have to work with are the texts 401 00:26:46,000 --> 00:26:49,160 Speaker 1: that are going back and forth. And if you, the evaluator, 402 00:26:49,359 --> 00:26:53,720 Speaker 1: cannot tell, that is the moment when machine intelligence has 403 00:26:53,840 --> 00:26:57,920 Speaker 1: finally arrived at the level of human intelligence. It has 404 00:26:58,160 --> 00:27:03,200 Speaker 1: passed the imitation or what we now call the Touring test. 405 00:27:04,240 --> 00:27:06,719 Speaker 1: And this reminds me of this great line in the 406 00:27:06,760 --> 00:27:11,879 Speaker 1: first episode of Westworld, where the protagonist William is talking 407 00:27:11,920 --> 00:27:14,560 Speaker 1: to the woman who's outfitting him for his adventure in 408 00:27:14,600 --> 00:27:17,080 Speaker 1: Westworld and giving him a hat and a gun and 409 00:27:17,119 --> 00:27:20,520 Speaker 1: so on, and he hesitantly asks, I hope you don't 410 00:27:20,520 --> 00:27:23,320 Speaker 1: mind if I ask you this question, but are you real? 411 00:27:23,880 --> 00:27:27,600 Speaker 1: And she says to him, if you can't tell, does 412 00:27:27,640 --> 00:27:30,760 Speaker 1: it matter? So I brought this up last episode in 413 00:27:30,800 --> 00:27:34,199 Speaker 1: the context of art, where we asked whether it matters 414 00:27:34,320 --> 00:27:36,760 Speaker 1: if the art is generated by an AI or a human, 415 00:27:37,200 --> 00:27:40,000 Speaker 1: But now this question comes up in the context of 416 00:27:40,480 --> 00:27:45,960 Speaker 1: intelligence and sentence. Does it matter whether we can tell 417 00:27:46,080 --> 00:27:49,199 Speaker 1: or not? Well, I think we're way beyond the Turing 418 00:27:49,280 --> 00:27:52,399 Speaker 1: test nowadays, but I don't feel like it gives us 419 00:27:52,400 --> 00:27:55,479 Speaker 1: a good answer to the question of whether the AI 420 00:27:55,600 --> 00:27:59,480 Speaker 1: is intelligent and is experiencing an inner life. I mean, 421 00:27:59,720 --> 00:28:02,479 Speaker 1: the Sturing test has been the test in the AI 422 00:28:02,600 --> 00:28:05,919 Speaker 1: world since the beginning. Why is it the perfect test? No, 423 00:28:06,160 --> 00:28:09,160 Speaker 1: but it's really hard to figure out how to test 424 00:28:09,200 --> 00:28:13,840 Speaker 1: for intelligence. But we have to be cautious about equating 425 00:28:14,240 --> 00:28:19,840 Speaker 1: conversational ability with sentience. Why well, for starters, let's just 426 00:28:19,920 --> 00:28:24,520 Speaker 1: acknowledge how easy it is for us to anthropomorphize. That 427 00:28:24,600 --> 00:28:28,960 Speaker 1: means to assign human qualities to everything around us. Like 428 00:28:29,040 --> 00:28:32,520 Speaker 1: we give animals human names and talk to them as 429 00:28:32,560 --> 00:28:36,520 Speaker 1: though they are people, and we project our emotions onto animals. 430 00:28:36,600 --> 00:28:40,720 Speaker 1: We make stories about animals that have human like qualities, 431 00:28:41,240 --> 00:28:43,960 Speaker 1: and we have animals that talk and wear clothes and 432 00:28:44,000 --> 00:28:48,040 Speaker 1: go on adventures in these stories. Every Pixar film that 433 00:28:48,120 --> 00:28:53,000 Speaker 1: you watch is about cars or toys or airplanes talking 434 00:28:53,040 --> 00:28:56,200 Speaker 1: and having emotions, and we don't even bad an eye 435 00:28:56,240 --> 00:28:59,920 Speaker 1: at that stuff. We can, in fact, just watch random 436 00:29:00,200 --> 00:29:04,240 Speaker 1: shapes moving around a computer screen and we will assign 437 00:29:04,800 --> 00:29:09,440 Speaker 1: intention and feel emotion depending on exactly how they're moving. 438 00:29:09,760 --> 00:29:11,840 Speaker 1: If you're interested in this, see the link on the 439 00:29:11,880 --> 00:29:15,640 Speaker 1: podcast page to the study by Heighter and Simil in 440 00:29:15,640 --> 00:29:19,880 Speaker 1: the nineteen forties where they move shapes around on a screen. Okay, 441 00:29:20,000 --> 00:29:22,960 Speaker 1: now this is all related to a point that I 442 00:29:23,040 --> 00:29:25,760 Speaker 1: brought up in the last episode, which is how easy 443 00:29:25,840 --> 00:29:28,880 Speaker 1: it is to pluck the strings on a human, or, 444 00:29:28,920 --> 00:29:33,600 Speaker 1: as the West World writers put it, how hackable humans are. 445 00:29:34,160 --> 00:29:36,080 Speaker 1: So I bring all this up to say that just 446 00:29:36,120 --> 00:29:40,440 Speaker 1: because you think that an answer sounds very clever or 447 00:29:40,480 --> 00:29:43,320 Speaker 1: it sounds like a human really tells us very little 448 00:29:43,360 --> 00:29:48,920 Speaker 1: about whether the AI is actually intelligent or sentient. It 449 00:29:48,960 --> 00:29:52,760 Speaker 1: only tells us something about the willingness of us as 450 00:29:52,840 --> 00:29:58,640 Speaker 1: observers to anthropomorphize, to assign intention where there is none, 451 00:29:58,920 --> 00:30:02,680 Speaker 1: Because what chat GPT does is take the structure of 452 00:30:02,800 --> 00:30:06,560 Speaker 1: language very impressively and spoon it back to us, and 453 00:30:06,640 --> 00:30:10,600 Speaker 1: we hear these well formed sentences, and we can hardly 454 00:30:11,080 --> 00:30:15,480 Speaker 1: help but impose sentience on the AI. And part of 455 00:30:15,480 --> 00:30:20,200 Speaker 1: the reason is that language is a super compressed package 456 00:30:20,240 --> 00:30:24,200 Speaker 1: that needs to be unpacked by the listener's brain for 457 00:30:24,320 --> 00:30:27,680 Speaker 1: its meaning. So we generally assume that when we send 458 00:30:27,720 --> 00:30:31,520 Speaker 1: our little package of sounds across the air, that it 459 00:30:31,920 --> 00:30:35,160 Speaker 1: unpacks and the other person understands exactly what we meant. 460 00:30:35,520 --> 00:30:41,400 Speaker 1: So when I say justice or love or suffering, we 461 00:30:41,520 --> 00:30:44,400 Speaker 1: all have a different sense in our heads about what 462 00:30:44,440 --> 00:30:48,400 Speaker 1: that means, because I'm just sending a few phonemes across 463 00:30:48,440 --> 00:30:51,160 Speaker 1: the air, and you have to unpack those words and 464 00:30:51,280 --> 00:30:55,240 Speaker 1: interpret them within your own model of the world. I'm 465 00:30:55,280 --> 00:30:57,840 Speaker 1: going to come back to this point in future episodes, 466 00:30:57,880 --> 00:31:00,720 Speaker 1: but for now, the point I want to make is 467 00:31:00,760 --> 00:31:06,040 Speaker 1: that a large language model can generate text statistically and 468 00:31:06,080 --> 00:31:09,160 Speaker 1: we can be gobsmacked by the apparent depth of it. 469 00:31:09,520 --> 00:31:12,160 Speaker 1: But in part this is because we cannot help but 470 00:31:12,280 --> 00:31:15,760 Speaker 1: impose meaning on the words that we receive. We hear 471 00:31:15,800 --> 00:31:18,760 Speaker 1: a particular string of sounds and we cannot help but 472 00:31:18,920 --> 00:31:24,520 Speaker 1: assume meaning behind it. Okay, so maybe the imitation game 473 00:31:24,680 --> 00:31:29,080 Speaker 1: is not really the best test for meaningful intelligence, but 474 00:31:29,120 --> 00:31:33,040 Speaker 1: there are other tests out there. Because while the Turing 475 00:31:33,120 --> 00:31:38,320 Speaker 1: test measures something about AI language processing, it doesn't necessarily 476 00:31:38,400 --> 00:31:44,080 Speaker 1: require the AI to demonstrate creative thinking or originality, and 477 00:31:44,120 --> 00:31:47,480 Speaker 1: so that leads us to the Loveless test, named after 478 00:31:47,920 --> 00:31:52,200 Speaker 1: Ada Loveless, who is the nineteenth century mathematician who's often 479 00:31:52,240 --> 00:31:55,000 Speaker 1: thought of as the world's first computer programmer. And she 480 00:31:55,200 --> 00:32:00,400 Speaker 1: once said quote, only when computers originate things should be 481 00:32:00,400 --> 00:32:05,280 Speaker 1: believed to have minds. So the Loveless test was proposed 482 00:32:05,280 --> 00:32:08,120 Speaker 1: in two thousand and one, and this test focuses on 483 00:32:08,200 --> 00:32:12,440 Speaker 1: the creative capabilities of AI systems. So to pass the 484 00:32:12,520 --> 00:32:17,360 Speaker 1: Loveless test, a machine has to create an original work, 485 00:32:17,680 --> 00:32:19,880 Speaker 1: such as a piece of art or a novel that 486 00:32:19,960 --> 00:32:24,680 Speaker 1: it was not explicitly designed to produce. This test aims 487 00:32:24,720 --> 00:32:29,160 Speaker 1: to assess whether AI systems can exhibit creativity and autonomy, 488 00:32:29,200 --> 00:32:32,680 Speaker 1: which are key aspects of what we think about with consciousness. 489 00:32:32,840 --> 00:32:36,719 Speaker 1: And the idea is that true sentience involves creative and 490 00:32:36,760 --> 00:32:41,280 Speaker 1: original thinking, not just the ability to follow pre programmed 491 00:32:41,400 --> 00:32:44,440 Speaker 1: rules or algorithms. And I'll just note that over a 492 00:32:44,560 --> 00:32:48,320 Speaker 1: decade ago, the scientist A. Mark Rydel proposed the loveless 493 00:32:48,320 --> 00:32:51,360 Speaker 1: two point zero test, which gets the human evaluator to 494 00:32:51,520 --> 00:32:56,480 Speaker 1: specify the constraints that will make the output novel and surprising. 495 00:32:56,760 --> 00:33:00,920 Speaker 1: So the example that l used in his paper is, quote, 496 00:33:01,280 --> 00:33:03,920 Speaker 1: create a story in which a boy falls in love 497 00:33:03,920 --> 00:33:07,120 Speaker 1: with a girl, Aliens abduct the boy, and the girl 498 00:33:07,240 --> 00:33:09,560 Speaker 1: saves the world with the help of a talking cat. 499 00:33:10,120 --> 00:33:13,640 Speaker 1: But we now know that this is totally trivial for chat, 500 00:33:13,680 --> 00:33:17,640 Speaker 1: GPTE or BARD or any large language model. And I 501 00:33:17,680 --> 00:33:20,400 Speaker 1: think this tells us that these sorts of games with 502 00:33:20,920 --> 00:33:25,520 Speaker 1: making conversation or making text or art are insufficient to 503 00:33:25,600 --> 00:33:29,600 Speaker 1: actually assess intelligence. Why because it's not so hard to 504 00:33:29,720 --> 00:33:33,640 Speaker 1: mix things up to make them seem original and intelligent 505 00:33:34,120 --> 00:33:37,960 Speaker 1: when it's really just doing a mashup. So I want 506 00:33:38,000 --> 00:33:40,320 Speaker 1: to turn to another test that I think is more 507 00:33:40,440 --> 00:33:43,800 Speaker 1: powerful than the Turing test of the Loveless test, and 508 00:33:43,960 --> 00:33:47,720 Speaker 1: probably easier to judge, and that is this, if a 509 00:33:47,880 --> 00:33:51,880 Speaker 1: system is truly intelligent, it should be able to do 510 00:33:52,680 --> 00:33:58,160 Speaker 1: scientific discovery. A version of the scientific discovery test was 511 00:33:58,360 --> 00:34:02,440 Speaker 1: first proposed by a scientist named Shao cheng Xiang a 512 00:34:02,480 --> 00:34:05,400 Speaker 1: few years ago, and he pointed out that the most 513 00:34:05,400 --> 00:34:10,239 Speaker 1: important thing that humans do is make scientific discoveries, and 514 00:34:10,280 --> 00:34:14,520 Speaker 1: the day our AI can make real discoveries is the 515 00:34:14,640 --> 00:34:17,920 Speaker 1: day they become as smart as we are. Now. I 516 00:34:18,000 --> 00:34:21,360 Speaker 1: want to propose an important change to this test, and 517 00:34:21,400 --> 00:34:38,319 Speaker 1: then I think we'll be getting somewhere. So here's the 518 00:34:38,360 --> 00:34:42,960 Speaker 1: scenario I'm envisioning. Let's say that I ask Ai some question, 519 00:34:43,080 --> 00:34:46,640 Speaker 1: a question in the biomedical space about what kind of 520 00:34:46,760 --> 00:34:49,600 Speaker 1: drug would be best suited to bind to this receptor 521 00:34:49,640 --> 00:34:52,880 Speaker 1: and trigger a cascade that causes a particular gene to 522 00:34:52,880 --> 00:34:56,040 Speaker 1: get suppressed. Okay, So imagine that I ask that to 523 00:34:56,160 --> 00:35:00,640 Speaker 1: chat GPT and it tells me some mind blowing, amazing 524 00:35:01,040 --> 00:35:05,360 Speaker 1: clever answer, one that had previously not been known, something 525 00:35:05,400 --> 00:35:09,000 Speaker 1: that's never been known by scientists before. We would assume 526 00:35:09,239 --> 00:35:14,200 Speaker 1: naturally that it has done some extraordinary scientific reasoning, but 527 00:35:14,280 --> 00:35:19,200 Speaker 1: that won't necessarily be the reason that it passes. Instead, 528 00:35:19,840 --> 00:35:23,359 Speaker 1: it might pass simply because it's more well read than 529 00:35:23,400 --> 00:35:26,440 Speaker 1: I am, or than any other human on the planet 530 00:35:26,480 --> 00:35:29,520 Speaker 1: by literally millions of times. So the way to think 531 00:35:29,560 --> 00:35:35,000 Speaker 1: about this is to picture a typical giant biomedical library, 532 00:35:35,120 --> 00:35:37,960 Speaker 1: where there's some fact stored at a paper and a 533 00:35:38,040 --> 00:35:41,279 Speaker 1: journal over here on this shelf in this book, and 534 00:35:41,360 --> 00:35:46,080 Speaker 1: there's another seemingly dissociated fact over on this shelf seven 535 00:35:46,120 --> 00:35:49,719 Speaker 1: stacks away, and there's a third fact all the way 536 00:35:49,760 --> 00:35:52,480 Speaker 1: on the other side of the library, on the bottom shelf, 537 00:35:52,920 --> 00:35:56,040 Speaker 1: in a book from nineteen seventy nine. And it's almost 538 00:35:56,440 --> 00:36:01,000 Speaker 1: infinitesimally unlikely that any human could even hope to have 539 00:36:01,120 --> 00:36:05,200 Speaker 1: read one one millionth of the biomedical literature, and really 540 00:36:05,280 --> 00:36:08,400 Speaker 1: really unlikely that she would be able to catch those 541 00:36:08,480 --> 00:36:11,480 Speaker 1: three facts and hold them in mind at the same time. 542 00:36:12,440 --> 00:36:15,400 Speaker 1: But this is trivial, of course, for a large language 543 00:36:15,400 --> 00:36:19,200 Speaker 1: model with hundreds of billions of nodes. So I think 544 00:36:19,600 --> 00:36:23,160 Speaker 1: that we will see new science getting done by CHATGPT, 545 00:36:24,040 --> 00:36:28,320 Speaker 1: not because it is conceptualizing, not because it's doing human 546 00:36:28,480 --> 00:36:32,040 Speaker 1: like reasoning, but because it doesn't know that these are 547 00:36:32,120 --> 00:36:35,480 Speaker 1: disparate facts spread around the library. It simply knows these 548 00:36:35,520 --> 00:36:38,560 Speaker 1: as three facts that seem to fit together. And so 549 00:36:38,680 --> 00:36:41,920 Speaker 1: with the right sort of questions, we might find that 550 00:36:42,040 --> 00:36:46,720 Speaker 1: sometimes AI generates something amazing and it seems to pass 551 00:36:46,800 --> 00:36:50,239 Speaker 1: the scientific discovery test. So this is going to be 552 00:36:50,239 --> 00:36:53,799 Speaker 1: incredibly useful for science. And I've never been able to 553 00:36:53,920 --> 00:36:57,680 Speaker 1: escape the feeling as I sift through Google scholar and 554 00:36:57,760 --> 00:37:00,920 Speaker 1: the thousands of papers published each month that have something 555 00:37:00,960 --> 00:37:04,760 Speaker 1: could hold all the knowledge and mind at once, each 556 00:37:05,040 --> 00:37:08,680 Speaker 1: page in every journal, and every gene in the genome, 557 00:37:08,960 --> 00:37:11,920 Speaker 1: and all the pages about chemistry and physics and mathematical 558 00:37:11,960 --> 00:37:15,720 Speaker 1: techniques and astrophysics and so on. Then you'd have lots 559 00:37:15,719 --> 00:37:19,239 Speaker 1: of puzzle pieces that could potentially make lots of connections. 560 00:37:19,560 --> 00:37:21,759 Speaker 1: And you know this might lead to the retirement of 561 00:37:21,880 --> 00:37:25,840 Speaker 1: many scientists, or at minimum lead to a better use 562 00:37:25,960 --> 00:37:30,040 Speaker 1: of our time. There's a depressing sense in which each scientist, 563 00:37:30,080 --> 00:37:33,400 Speaker 1: each one of us, finds little pieces of the puzzle, 564 00:37:33,760 --> 00:37:37,000 Speaker 1: and in the twinkling of a single human lifetime, a 565 00:37:37,040 --> 00:37:41,160 Speaker 1: busy scientist might collect up a handful of different puzzle pieces. 566 00:37:41,800 --> 00:37:46,799 Speaker 1: The most voracious reader, the most assiduous worker, the most 567 00:37:46,800 --> 00:37:50,319 Speaker 1: creative synthesizer of ideas, can only hope to collect a 568 00:37:50,400 --> 00:37:53,840 Speaker 1: small number of puzzle pieces and pray that some of 569 00:37:53,880 --> 00:37:56,520 Speaker 1: them might fit together. So this is going to be 570 00:37:56,600 --> 00:38:02,960 Speaker 1: massively important. But I wanted to find two categories of 571 00:38:03,000 --> 00:38:06,080 Speaker 1: scientific discovery. The first is what I just described, which 572 00:38:06,120 --> 00:38:09,600 Speaker 1: is science where things that already exist in literature can 573 00:38:09,640 --> 00:38:13,680 Speaker 1: be pieced together. And let's call that level one discovery. 574 00:38:14,000 --> 00:38:16,759 Speaker 1: And these large language models will be awesome at level 575 00:38:16,760 --> 00:38:18,640 Speaker 1: one because they've read every paper and they have a 576 00:38:18,640 --> 00:38:22,360 Speaker 1: perfect memory. But I want to distinguish a second level 577 00:38:22,440 --> 00:38:25,719 Speaker 1: of scientific discovery, and this is the one I'm interested in. 578 00:38:26,000 --> 00:38:29,160 Speaker 1: I'll call this level two, and that is science that 579 00:38:29,239 --> 00:38:34,279 Speaker 1: requires conceptualization to get to the next step, not just 580 00:38:34,440 --> 00:38:39,200 Speaker 1: remixing what's already there. Conceptualization like when the young Albert 581 00:38:39,239 --> 00:38:43,200 Speaker 1: Einstein imagined something that he had never seen before. He 582 00:38:43,239 --> 00:38:45,920 Speaker 1: asked himself, what would it be like if I could 583 00:38:46,000 --> 00:38:49,040 Speaker 1: catch up with a beam of light and write it 584 00:38:49,120 --> 00:38:51,960 Speaker 1: like a surfer riding a wave. And this is how 585 00:38:52,000 --> 00:38:56,720 Speaker 1: he derived this special theory of relativity. This isn't something 586 00:38:56,760 --> 00:38:59,919 Speaker 1: he looked up and found three facts that clicked. Again, 587 00:39:00,640 --> 00:39:05,160 Speaker 1: he imagined he asked new questions. He tried out a 588 00:39:05,239 --> 00:39:08,800 Speaker 1: new model of the world, one in which time runs 589 00:39:08,840 --> 00:39:11,640 Speaker 1: differently depending on how fast you're going, and then he 590 00:39:11,800 --> 00:39:15,400 Speaker 1: worked backwards to see if that model could work. Or 591 00:39:15,480 --> 00:39:19,560 Speaker 1: consider when Charles Darwin thought about the species that he 592 00:39:19,640 --> 00:39:22,560 Speaker 1: saw around him, and he imagined all the species that 593 00:39:22,600 --> 00:39:26,000 Speaker 1: he didn't see but who might have existed, and he 594 00:39:26,040 --> 00:39:29,279 Speaker 1: was able to put together a new mental model in 595 00:39:29,320 --> 00:39:33,240 Speaker 1: which most species don't make it and we only see 596 00:39:33,280 --> 00:39:39,319 Speaker 1: those whose mutations cause survival advantages or reproductive advantages. These 597 00:39:39,360 --> 00:39:42,680 Speaker 1: weren't facts that he just collected from some papers. He 598 00:39:42,800 --> 00:39:47,320 Speaker 1: was trying out a new model of the world. Now 599 00:39:47,600 --> 00:39:50,920 Speaker 1: this kind of science isn't just for the big giant stuff. 600 00:39:51,040 --> 00:39:54,319 Speaker 1: Most meaningful science is actually driven by this kind of 601 00:39:54,800 --> 00:39:59,840 Speaker 1: imagination of new models. Just as one example, I recently 602 00:40:00,040 --> 00:40:03,359 Speaker 1: in an episode about whether time runs in slow motion 603 00:40:03,480 --> 00:40:06,600 Speaker 1: when you're in fear for your life. And so when 604 00:40:06,640 --> 00:40:10,120 Speaker 1: I wondered about this question, I realized there were two 605 00:40:10,320 --> 00:40:13,560 Speaker 1: hypotheses that might explain it, and I thought up an 606 00:40:13,600 --> 00:40:17,160 Speaker 1: experiment to discriminate those two hypotheses. And then we built 607 00:40:17,160 --> 00:40:21,400 Speaker 1: a wristband that flashes information at a particular speed and 608 00:40:21,480 --> 00:40:23,760 Speaker 1: had people wear, and we dropped them from one hundred 609 00:40:23,760 --> 00:40:26,640 Speaker 1: and fifty foot tall tower into a net below. A 610 00:40:26,920 --> 00:40:31,400 Speaker 1: large language model presumably couldn't do that because it's just 611 00:40:31,520 --> 00:40:35,279 Speaker 1: playing statistical word games. And unless someone had thought of 612 00:40:35,320 --> 00:40:40,480 Speaker 1: that experiment and written it down, JATGPT would never say, Okay, 613 00:40:40,520 --> 00:40:43,080 Speaker 1: here's a new framework, and how we can design an 614 00:40:43,120 --> 00:40:46,000 Speaker 1: experiment to put this to the test. So this is 615 00:40:46,040 --> 00:40:49,120 Speaker 1: what I wanted to find as the most meaningful test 616 00:40:49,640 --> 00:40:53,680 Speaker 1: for a human level of intelligence. When AI can do 617 00:40:54,239 --> 00:40:59,319 Speaker 1: science in this way, generating new ideas and frameworks, not 618 00:40:59,400 --> 00:41:05,000 Speaker 1: just clicking act together, then we will have matched human intelligence. 619 00:41:08,040 --> 00:41:09,719 Speaker 1: And I just want to take one more angle on 620 00:41:09,760 --> 00:41:13,000 Speaker 1: this to make the picture clear. The way a scientist 621 00:41:13,160 --> 00:41:17,440 Speaker 1: reads a journal paper is not simply by correlating words 622 00:41:17,440 --> 00:41:20,520 Speaker 1: and extracting keywords, although that might be part of it, 623 00:41:20,640 --> 00:41:24,759 Speaker 1: but also by realizing what was not said. Why did 624 00:41:24,800 --> 00:41:28,320 Speaker 1: the authors cut off the x axis here at thirty 625 00:41:28,680 --> 00:41:31,480 Speaker 1: What if they had extended this graph, would the line 626 00:41:31,560 --> 00:41:34,840 Speaker 1: have reversed in its trend? And why didn't the authors 627 00:41:34,920 --> 00:41:38,560 Speaker 1: mention the hypothesis of Smith at all? And does that 628 00:41:38,640 --> 00:41:42,080 Speaker 1: graph look too perfect? You know? One of my mentors, 629 00:41:42,120 --> 00:41:46,239 Speaker 1: Francis Krik operated under the assumption that he should disbelieve 630 00:41:46,400 --> 00:41:48,880 Speaker 1: twenty five percent of what he read in the literature. 631 00:41:49,400 --> 00:41:53,520 Speaker 1: Is this because of fraud or error, or statistical fluctuations 632 00:41:53,600 --> 00:41:57,239 Speaker 1: or manipulation or the waste basket effect? Who cares? The 633 00:41:57,280 --> 00:42:01,080 Speaker 1: bottom line is that the literature is rife with errors, 634 00:42:01,480 --> 00:42:06,160 Speaker 1: and depending on the field, some estimates put the ireproducibility 635 00:42:06,640 --> 00:42:11,719 Speaker 1: at fifty percent. So when scientists read papers they know this, 636 00:42:11,920 --> 00:42:15,920 Speaker 1: just as Francis Crick did. They read in an entirely 637 00:42:15,920 --> 00:42:20,360 Speaker 1: different manner than Google Translate or Watson or chat GPT 638 00:42:20,640 --> 00:42:25,640 Speaker 1: or any of the correlational methods they extrapolate. They read 639 00:42:25,680 --> 00:42:28,719 Speaker 1: the paper and wonder about other possibilities. They chew on 640 00:42:28,800 --> 00:42:32,440 Speaker 1: what's missing. They envision the next step. They think of 641 00:42:32,480 --> 00:42:37,240 Speaker 1: the next experiment that could confirm or disconfirm the hypotheses 642 00:42:37,600 --> 00:42:40,600 Speaker 1: and the frameworks in the paper. To my mind, the 643 00:42:40,640 --> 00:42:43,440 Speaker 1: meaningful goal of AI is not going to be found 644 00:42:43,480 --> 00:42:47,000 Speaker 1: in number crunching and looking for facts that click together. 645 00:42:47,600 --> 00:42:50,840 Speaker 1: It's going to often be something else. It's going to 646 00:42:50,880 --> 00:42:56,160 Speaker 1: require an AI that learns how humans think, how they behave, 647 00:42:56,360 --> 00:42:59,719 Speaker 1: what they don't say, what they didn't think of, what 648 00:42:59,760 --> 00:43:03,600 Speaker 1: they misthought about, what they should think about. And one 649 00:43:03,640 --> 00:43:06,640 Speaker 1: more thing, I should note that these different levels I've outlined, 650 00:43:07,040 --> 00:43:11,600 Speaker 1: from fitting facts together versus imagining new world models, they're 651 00:43:11,600 --> 00:43:15,959 Speaker 1: probably gonna end up with blurry boundaries. So maybe chat 652 00:43:16,000 --> 00:43:19,840 Speaker 1: GPT will come up with something, and you won't always 653 00:43:20,040 --> 00:43:24,799 Speaker 1: know whether it's piecing together a few disparate pieces in 654 00:43:24,880 --> 00:43:29,160 Speaker 1: the literature what I'm calling level one, or whether it's 655 00:43:29,800 --> 00:43:33,240 Speaker 1: come up with something that is truly a new world 656 00:43:33,360 --> 00:43:36,759 Speaker 1: model that's not a simple clicking together but a genuine 657 00:43:37,239 --> 00:43:40,800 Speaker 1: process of generating a new framework to explain the data. 658 00:43:40,920 --> 00:43:44,880 Speaker 1: So distinguishing the levels of discovery is probably not going 659 00:43:44,960 --> 00:43:47,600 Speaker 1: to be an easy task with a bright line between them, 660 00:43:48,080 --> 00:43:51,359 Speaker 1: but I think it will clarify some things to make 661 00:43:51,400 --> 00:43:55,799 Speaker 1: this distinction. And last thing, I don't necessarily know that 662 00:43:55,840 --> 00:43:59,520 Speaker 1: there's something magical and ineffable about the way that humans 663 00:43:59,560 --> 00:44:03,600 Speaker 1: do this. Presumably we're running algorithms too, it's just that 664 00:44:03,640 --> 00:44:07,800 Speaker 1: they're running on self configuring wetwear. I have seen tens 665 00:44:07,800 --> 00:44:10,880 Speaker 1: of thousands of science experiments in my career, so I 666 00:44:11,040 --> 00:44:14,279 Speaker 1: know the process of asking a question and figuring out 667 00:44:14,640 --> 00:44:17,000 Speaker 1: what we'll put it to the test. So we may 668 00:44:17,040 --> 00:44:19,120 Speaker 1: get to level two and it may be sooner than 669 00:44:19,160 --> 00:44:21,480 Speaker 1: we expect, but I just want to be clear that 670 00:44:21,719 --> 00:44:25,279 Speaker 1: right now we have not figured out the human algorithms. 671 00:44:25,680 --> 00:44:29,680 Speaker 1: So the current version of AI, as massively impressive as 672 00:44:29,760 --> 00:44:34,560 Speaker 1: it is, does not do level two scientific problem solving. 673 00:44:34,840 --> 00:44:37,720 Speaker 1: And that's when we're going to know that we've crossed 674 00:44:37,800 --> 00:44:41,319 Speaker 1: a new kind of line into a machine that is 675 00:44:41,520 --> 00:44:45,799 Speaker 1: truly intelligent. So let's wrap up. At least for now. 676 00:44:45,920 --> 00:44:48,480 Speaker 1: Humans still have to do the science, by which I 677 00:44:48,560 --> 00:44:52,400 Speaker 1: mean the conceptual work, wherein we take a framework for 678 00:44:52,560 --> 00:44:55,800 Speaker 1: understanding the world and we rethink it and we mentally 679 00:44:55,920 --> 00:44:59,320 Speaker 1: simulate whether a new model of the world could explain 680 00:44:59,400 --> 00:45:01,839 Speaker 1: the observed data, and we come up with a way 681 00:45:01,880 --> 00:45:05,200 Speaker 1: to test that new model. It's not just searching for facts. 682 00:45:05,640 --> 00:45:07,640 Speaker 1: So I'm definitely not saying we won't get to the 683 00:45:07,640 --> 00:45:11,640 Speaker 1: next level where AI can conceptualize things and predict forward 684 00:45:11,719 --> 00:45:14,600 Speaker 1: and build new knowledge. This might be a week from now, 685 00:45:14,719 --> 00:45:16,680 Speaker 1: or it might be a century from now. Who knows 686 00:45:17,040 --> 00:45:19,120 Speaker 1: how hard a problem that's going to turn out to be. 687 00:45:19,520 --> 00:45:21,640 Speaker 1: But I want us to be clear eyed on where 688 00:45:21,680 --> 00:45:26,400 Speaker 1: we are right now, because sometimes in the blindingly impressive 689 00:45:26,520 --> 00:45:29,640 Speaker 1: light of what current AI is doing, it can be 690 00:45:29,719 --> 00:45:33,960 Speaker 1: difficult to see, what's missing and where we might be heading. 691 00:45:38,360 --> 00:45:40,920 Speaker 1: That's all for this week. To find out more and 692 00:45:40,960 --> 00:45:43,680 Speaker 1: to share your thoughts, head over to eagleman dot com 693 00:45:43,719 --> 00:45:47,680 Speaker 1: slash Podcasts, and you can also watch full episodes of 694 00:45:47,760 --> 00:45:51,400 Speaker 1: Inner Cosmos on YouTube. Subscribe to my channel so you 695 00:45:51,440 --> 00:45:54,719 Speaker 1: can follow along each week for new updates. I'd love 696 00:45:54,760 --> 00:45:58,480 Speaker 1: to hear your questions, so please send those to podcast 697 00:45:58,640 --> 00:46:01,560 Speaker 1: at eagleman dot com and I will do a special 698 00:46:01,600 --> 00:46:05,480 Speaker 1: episode where I answer questions. Until next time. I'm David 699 00:46:05,480 --> 00:46:07,880 Speaker 1: Eagelman and this is Inner Cosmos