WEBVTT - Ep72 "How do you put yourself in other people's shoes (and can AI do it)?" 0:00:05.160 --> 0:00:07.880 You know that moment in the horror movie where the 0:00:07.960 --> 0:00:12.040 monster is coming closer but the person on screen doesn't 0:00:12.080 --> 0:00:14.880 see it. Why does that drive you crazy? And what 0:00:14.920 --> 0:00:19.560 does that teach us about brains? What is theory of 0:00:19.800 --> 0:00:23.079 mind and why is it so important for everyone from 0:00:23.160 --> 0:00:27.160 poker players to con men, to stage magicians to novelists. 0:00:27.720 --> 0:00:30.360 We're going to talk about a very fundamental skill of 0:00:30.440 --> 0:00:34.880 human brains today, and as impressive as AI is currently, 0:00:34.920 --> 0:00:38.880 we're going to ask the question of whether computers can 0:00:39.000 --> 0:00:42.440 replicate this right now or whether it is beyond their 0:00:42.479 --> 0:00:48.200 skill set. Welcome to Inner Cosmos with me David Eagleman. 0:00:48.280 --> 0:00:50.920 I'm a neuroscientist and an author at Stanford, and in 0:00:50.960 --> 0:00:54.639 these episodes we sail deeply into our three pound universe 0:00:54.880 --> 0:00:58.280 to understand why and how our lives look the way 0:00:58.280 --> 0:01:02.000 they do. Today's episode is about what it takes to 0:01:02.240 --> 0:01:05.680 understand other people, how your brain does it, and whether 0:01:05.840 --> 0:01:11.920 computers could do it. So imagine this. You're walking down 0:01:11.920 --> 0:01:17.080 the street and you see someone frantically searching their pockets 0:01:17.200 --> 0:01:20.880 and looking around with furrowed brows in a tight frown. 0:01:21.480 --> 0:01:25.399 So without them saying a word, you can infer that 0:01:25.480 --> 0:01:29.120 they might have lost something important. Maybe it's his keys. 0:01:29.760 --> 0:01:33.560 Your brain can easily make a good guess about another 0:01:33.680 --> 0:01:37.679 person's mental state just from looking at their actions. We 0:01:37.800 --> 0:01:41.319 are inferring something about what is going on in that 0:01:41.400 --> 0:01:45.640 person's head. But it's more than just pattern matching. It's 0:01:45.640 --> 0:01:48.520 not simply that your brain has seen lots of people 0:01:48.600 --> 0:01:52.440 patting their pockets and you talked with them afterwards, and 0:01:52.480 --> 0:01:54.760 you figured out why they were doing that, and you 0:01:54.880 --> 0:01:59.000 detected a pattern, and you memorized, ah, okay, that pattern 0:01:59.040 --> 0:02:04.120 equals that problem. Instead, you have the ability to imagine 0:02:04.200 --> 0:02:08.880 yourself in their situation. You can mentally slip into their 0:02:08.960 --> 0:02:12.200 shoes and ask, what would I be thinking if I 0:02:12.320 --> 0:02:16.720 were patting my pockets and frantically searching around me? And 0:02:16.800 --> 0:02:19.959 maybe you see something else. You see a kid there 0:02:20.120 --> 0:02:23.360 around the corner, and the kid is peeking around the 0:02:23.400 --> 0:02:26.720 corner at the man patting his pockets, and the child 0:02:27.200 --> 0:02:30.520 is giggling. Now, why is the kid giggling while the 0:02:30.560 --> 0:02:34.520 guy is so obviously worried, Well, it probably strikes you 0:02:34.560 --> 0:02:37.520 that he's hiding something from the guy. You see that 0:02:37.560 --> 0:02:41.160 the kid is not running away. Instead, he's standing in 0:02:41.280 --> 0:02:44.000 such a way that he'll be spotted. Now it's pretty 0:02:44.080 --> 0:02:48.280 obvious what's happening here. You can step into the man's 0:02:48.320 --> 0:02:50.760 head to feel the worry, and you can step into 0:02:50.800 --> 0:02:53.880 the kid's head to recognize that he feels like he's 0:02:53.919 --> 0:02:56.000 playing a game, even if it doesn't strike you as 0:02:56.080 --> 0:03:00.920 so funny. Then you catch the guy meet with the 0:03:01.040 --> 0:03:03.680 kid for just a fraction of a second, which sends 0:03:03.720 --> 0:03:07.360 the kid into fits of laughter, and you realize the 0:03:07.400 --> 0:03:11.280 man is just playing along. Now, how did you decide 0:03:11.680 --> 0:03:14.440 what is going on in the heads of these two Again, 0:03:14.520 --> 0:03:18.520 it's not as though you memorized an algorithm here. Okay, 0:03:18.560 --> 0:03:21.280 if there's eye contact, then there's one interpretation. If there's 0:03:21.280 --> 0:03:26.320 no eye contact, then a totally different interpretation. To appreciate 0:03:26.520 --> 0:03:29.960 how complex this mind reading is that you just did, 0:03:30.840 --> 0:03:34.639 just imagine that you're a space alien watching this scene 0:03:34.720 --> 0:03:38.080 from your spaceship. You would be totally confused. You would 0:03:38.120 --> 0:03:41.800 have no idea what's going on in this weird scene 0:03:42.280 --> 0:03:45.200 because you don't know what it is to be a human. 0:03:46.000 --> 0:03:49.160 Here's another analogy to appreciate this. Think about the way 0:03:49.160 --> 0:03:53.400 that you, as a human might watch fish. You really 0:03:53.440 --> 0:03:57.160 don't understand what the heck they're doing. One fish suddenly 0:03:57.160 --> 0:04:01.080 starts swimming faster, and another starts swimming in circles, and 0:04:01.160 --> 0:04:04.400 one starts flapping its gills faster, and one moves up 0:04:04.640 --> 0:04:09.040 towards the surface. It's all just weird fish behavior to you. 0:04:09.040 --> 0:04:11.120 You don't know how to read any of it. It's 0:04:11.200 --> 0:04:14.720 just fish stuff, and you're not able to immediately construct 0:04:14.760 --> 0:04:19.000 a story about the meaning of any of this. And 0:04:19.000 --> 0:04:22.279 that's what it's like to be this space alien watching 0:04:22.320 --> 0:04:27.200 this guy checking his pockets and the child giggling. Now, 0:04:27.760 --> 0:04:31.520 what allows us, as opposed to the space alien, to 0:04:31.600 --> 0:04:36.000 be so good at reading our fellow humans. This is 0:04:36.040 --> 0:04:41.240 what psychologists and neuroscientists call theory of mind, and that's 0:04:41.279 --> 0:04:44.560 what we're talking about today. Theory of mind is the 0:04:44.600 --> 0:04:49.920 ability to understand that other people have their own thoughts 0:04:49.960 --> 0:04:53.720 and feelings and beliefs that are different from yours. It's 0:04:53.760 --> 0:04:58.120 the ability to recognize that others have their own perspectives. 0:04:58.160 --> 0:05:02.200 It's the ability to attribut mute mental states to other people, 0:05:02.320 --> 0:05:07.440 like what their intentions are, or their desires, or their emotions, 0:05:07.960 --> 0:05:11.640 or what they know or don't know. And theory of 0:05:11.680 --> 0:05:15.440 mind is a key cognitive skill that allows us to 0:05:15.600 --> 0:05:19.360 interact with other people in a very rich and nuanced way. 0:05:19.880 --> 0:05:23.560 Just think about how pervasive this skill is in everything 0:05:23.600 --> 0:05:27.359 we do. So take sarcasm. When your friend makes a 0:05:27.760 --> 0:05:32.279 sarcastic comment, you can recognize that her words don't match 0:05:32.400 --> 0:05:36.640 her true intention. So, for example, if she says, oh, awesome, 0:05:36.680 --> 0:05:40.919 more traffic, I love traffic, you infer that she's not 0:05:41.120 --> 0:05:46.400 actually pleased. This requires understanding her mental state that she 0:05:46.600 --> 0:05:52.320 is irritated not happy. Now, if you were Siri or Alexa, 0:05:52.440 --> 0:05:55.919 you wouldn't be able to recognize anything but the words. 0:05:56.320 --> 0:05:59.800 You wouldn't understand anything about the mind behind the words. 0:06:00.279 --> 0:06:03.200 So we're going to talk about how brains do it 0:06:03.560 --> 0:06:06.719 and whether or not computers can do it. But before 0:06:06.720 --> 0:06:08.320 we go there, we're going to take a few minutes 0:06:08.360 --> 0:06:12.839 to really appreciate how the skill is everywhere in what 0:06:12.960 --> 0:06:17.880 we do. For example, just think about different professions. So 0:06:18.000 --> 0:06:22.440 detectives use theory of mind all the time. Did mister 0:06:22.560 --> 0:06:25.080 Jones know that the food had gone bad when he 0:06:25.160 --> 0:06:28.560 sold it? Did mister Smith know that his boss was 0:06:28.640 --> 0:06:32.600 involved with organized crime or was he acting with no knowledge? 0:06:33.200 --> 0:06:35.000 Or more generally, if they want to know if someone 0:06:35.080 --> 0:06:38.120 is lying, it usually helps to step into their shoes 0:06:38.160 --> 0:06:40.880 and think about what that person knows or doesn't know. 0:06:41.279 --> 0:06:44.400 Magicians use theory of mind. They know that if they 0:06:44.800 --> 0:06:47.400 move their hand in an arc, your attention is going 0:06:47.440 --> 0:06:51.560 to follow that, and therefore they know what you won't 0:06:51.640 --> 0:06:54.919 see them do. They know that even though they know 0:06:55.120 --> 0:06:58.040 something happened, like the card dropped into their sleeve, they 0:06:58.040 --> 0:07:01.760 know that you don't know that. They always keep your 0:07:01.920 --> 0:07:05.679 point of view, your beliefs, at the forefront of their mind. 0:07:06.440 --> 0:07:08.800 Con Men do this. They listen to your words and 0:07:08.839 --> 0:07:11.920 they read your body language to gather what you know 0:07:12.040 --> 0:07:15.280 and don't know, and therefore what buttons they should push next. 0:07:15.720 --> 0:07:20.480 Psychiatrists and psychologists always use theory of mind to understand 0:07:20.800 --> 0:07:24.200 what is being expressed from the patient's point of view, 0:07:24.400 --> 0:07:26.960 In other words, what the person believes, whether or not 0:07:27.040 --> 0:07:30.360 it's what the therapist believes. I'll give you another example. 0:07:30.640 --> 0:07:33.840 My friend Maddie is a professional poker player, and he 0:07:33.960 --> 0:07:37.400 describes poker playing like this. He says, when you're learning 0:07:37.480 --> 0:07:39.880 to play poker, you think about the cards you have 0:07:39.960 --> 0:07:43.239 in your hand. As you get better, you think about 0:07:43.240 --> 0:07:46.840 your hand and also what the other person is thinking. 0:07:47.200 --> 0:07:49.560 And as you get even better, you think about what 0:07:49.640 --> 0:07:52.960 the other person is thinking your thinking, and when you 0:07:53.000 --> 0:07:57.040 get to the professional levels, you're thinking about what he thinks, 0:07:57.320 --> 0:08:00.559 you think he thinks, and people who are real pros 0:08:00.680 --> 0:08:04.280 can think five or six levels deep on this. All 0:08:04.320 --> 0:08:08.080 of this is theory of mind, and theory of mind 0:08:08.240 --> 0:08:12.520 is key when you're teaching something. For example, parents know 0:08:13.080 --> 0:08:16.880 that their children can't understand certain things. For example, the 0:08:16.960 --> 0:08:20.080 child needs to get that smallpox shot, even though to 0:08:20.120 --> 0:08:22.720 the child that's nothing but scary and he simply doesn't 0:08:22.960 --> 0:08:26.600 have the capacity to think about the future benefits that 0:08:26.640 --> 0:08:29.880 will accrue. Or the school teacher can only hope to 0:08:30.040 --> 0:08:33.800 educate her students if she knows what they already know 0:08:34.040 --> 0:08:36.319 or don't know. She needs to phrase things in such 0:08:36.360 --> 0:08:39.320 a way that someone who doesn't already know what she 0:08:39.480 --> 0:08:44.000 knows can absorb it, and that just requires theory of mind. 0:08:44.480 --> 0:08:47.160 If she couldn't simulate what it's like to be in 0:08:47.200 --> 0:08:50.719 their heads, she'd have no meaningful shot at getting them 0:08:50.960 --> 0:08:54.360 past the first quiz. And this issue of considering what 0:08:54.480 --> 0:08:59.240 someone knows or doesn't know is also critical in any negotiation. 0:08:59.400 --> 0:09:03.040 You try to under understand the other person's desires and 0:09:03.200 --> 0:09:08.680 goals and where they might potentially compromise during a salary negotiation, 0:09:08.760 --> 0:09:12.199 you consider what your employer is thinking about the needs 0:09:12.200 --> 0:09:15.240 in future of the company and therefore what they might 0:09:15.280 --> 0:09:17.800 be willing to offer. And this is also how you 0:09:17.840 --> 0:09:21.640 manage conflicts. In any disagreement, if you're smart, you try 0:09:21.640 --> 0:09:26.480 to understand the other person's perspective to resolve the issue. 0:09:26.559 --> 0:09:28.960 If your partner is upset with you, you try to 0:09:29.000 --> 0:09:32.360 figure out what you did or said that set things off, 0:09:32.400 --> 0:09:35.600 and why that offended the other person and how it 0:09:35.840 --> 0:09:38.640 landed for them. And that's the single way that you're 0:09:38.640 --> 0:09:42.160 going to hit the problem effectively. So this ability to 0:09:42.240 --> 0:09:46.280 slip into someone else's shoes has almost everything to do 0:09:46.360 --> 0:09:51.880 with our social intelligence. You use this very human skill 0:09:52.600 --> 0:09:55.240 all the time. And before we get to the next 0:09:55.280 --> 0:09:57.800 act of this podcast, where I ask if computers can 0:09:57.840 --> 0:09:59.920 do this or not, I just want to finish fla 0:10:00.360 --> 0:10:03.160 this out so we can really see how pervasive this is. 0:10:03.640 --> 0:10:07.000 So as an example, you rev up your theory of 0:10:07.200 --> 0:10:10.640 mind engine whenever you send an email. If you know 0:10:10.760 --> 0:10:13.959 someone has a well developed model of you, like your 0:10:14.000 --> 0:10:17.600 parents or your spouse, then you can use abbreviations and 0:10:17.640 --> 0:10:20.840 shortcuts to get your message across. But if you're writing 0:10:20.840 --> 0:10:23.120 to someone who's never met you before. Let's say you're 0:10:23.160 --> 0:10:26.520 applying to a new job. You run a very different game, 0:10:27.160 --> 0:10:31.600 so you're not just an email writing algorithm that produces output, 0:10:31.640 --> 0:10:35.559 but instead your output is modified according to who you 0:10:35.640 --> 0:10:38.440 expect is doing the reading on the other end, and 0:10:38.520 --> 0:10:42.160 specifically what their mind is like. And I also want 0:10:42.200 --> 0:10:45.760 to mention that theory of mind is critical for literature 0:10:45.880 --> 0:10:48.080 to work because it's often the case that you can 0:10:48.520 --> 0:10:52.400 see the limitations of the character's point of view. So, 0:10:52.520 --> 0:10:55.960 for example, if you remember the beginning of the movie Jaws, 0:10:56.080 --> 0:10:58.840 the woman is swimming around in the ocean water and 0:10:58.880 --> 0:11:02.680 she's very relaxed than happy because we see the shark, 0:11:02.960 --> 0:11:06.720 but she doesn't. If we didn't have theory of mind, 0:11:06.760 --> 0:11:09.040 we would simply say, oh, there's a shark there. But 0:11:09.080 --> 0:11:12.439 we're able to understand that she cannot see the shark, 0:11:12.720 --> 0:11:14.880 and that's a big part of why we are fearful, 0:11:15.400 --> 0:11:18.120 because she isn't fearful, and we want her to be. 0:11:19.000 --> 0:11:23.520 This stepping into other people's heads drives essentially all horror 0:11:23.600 --> 0:11:26.720 movies because we often know something that the main character 0:11:27.320 --> 0:11:30.960 does not, and it also drives romantic comedies. For example, 0:11:31.000 --> 0:11:35.000 we see the guy doing something very nice like helping 0:11:35.040 --> 0:11:37.760 an elderly woman cross the street, and he doesn't know 0:11:37.840 --> 0:11:41.080 that he is being watched by the female love interest, 0:11:41.520 --> 0:11:45.200 and therefore we the audience interpret what kind of guy 0:11:45.240 --> 0:11:47.959 he must be to behave that way when as far 0:11:47.960 --> 0:11:50.720 as he knows, he's totally alone. We would have a 0:11:50.800 --> 0:11:56.040 totally different interpretation. If he sees his romantic counterparts there 0:11:56.080 --> 0:11:58.880 and then he does the charitable act, we'd simulate that 0:11:58.920 --> 0:12:03.080 his intentions are different there. Now, why are human brains 0:12:03.160 --> 0:12:07.440 so talented at making theories about other people's minds. Well, 0:12:07.480 --> 0:12:10.240 you've heard me say many times that the job of 0:12:10.280 --> 0:12:14.960 intelligent brains is to predict the future. If you're the magician, 0:12:15.040 --> 0:12:18.520 you'd better be sure that you are predicting correctly where 0:12:18.559 --> 0:12:21.480 their spotlight of attention is about to be. If you're 0:12:21.520 --> 0:12:24.440 the poker player or the con man, you're trying to 0:12:24.440 --> 0:12:27.440 predict what someone is going to do next, and this 0:12:27.520 --> 0:12:30.360 is the optimal way to do this is to step 0:12:30.400 --> 0:12:34.240 into their mental world and understand what it is like 0:12:34.440 --> 0:12:36.880 to be them. What they know and they don't know. 0:12:37.360 --> 0:12:43.200 You leverage theory of mind to anticipate their next action, 0:12:43.600 --> 0:12:46.760 and presumably this reaches back to the recent millions of 0:12:46.840 --> 0:12:49.880 years of our evolution. So if you're an early homo 0:12:49.920 --> 0:12:53.160 sapien and moving along the trail and you see another 0:12:53.240 --> 0:12:57.199 homo sapien coming down the trail towards you, it's absolutely 0:12:57.200 --> 0:12:59.920 critical for you to figure out is he going to 0:13:00.120 --> 0:13:03.240 attack me? Is he scared of me? Is he trying 0:13:03.280 --> 0:13:05.520 to trick me? Is he just trying to get past me. 0:13:06.200 --> 0:13:09.320 You're trying to figure out his mind so you can 0:13:09.360 --> 0:13:13.280 figure out his next actions. So what I've told you 0:13:13.320 --> 0:13:16.440 so far is that theory of mind is this critical 0:13:16.480 --> 0:13:21.079 foundation for all of our meaningful social interactions because those 0:13:21.200 --> 0:13:26.319 require you to be able to simulate other people's intentions 0:13:26.360 --> 0:13:30.800 and emotions and beliefs. Your brain doesn't assume that it's 0:13:30.840 --> 0:13:34.600 a knowledge communism out there where everyone knows exactly what 0:13:34.720 --> 0:13:37.840 you know. Instead, we're able to pull off a higher 0:13:37.920 --> 0:13:41.760 level of interaction because we understand that the world is 0:13:41.800 --> 0:13:45.440 different inside different heads. And this, by the way, is 0:13:45.480 --> 0:13:48.880 really sophisticated. It requires knowing who I am and what 0:13:48.920 --> 0:13:51.560 I see and believe, and also holding in my head 0:13:51.600 --> 0:13:53.480 what it is to be someone else and see and 0:13:53.520 --> 0:13:56.920 believe something different. This is a very sophisticated computation that 0:13:56.960 --> 0:14:00.800 the brain pulls off, but because we're so good at it, 0:14:00.800 --> 0:14:05.680 it's typically invisible to us. But theory of mind doesn't 0:14:05.720 --> 0:14:09.760 come for free. It's something that develops with time. As 0:14:09.800 --> 0:14:12.800 you get more and more experience in the world and 0:14:12.840 --> 0:14:16.000 you stop believing that you are the centerpiece and that 0:14:16.080 --> 0:14:18.760 everyone else is just a cast member. You come to 0:14:18.880 --> 0:14:22.640 understand that that person believes something different than you do, 0:14:23.120 --> 0:14:25.800 and this other person feels a certain way even though 0:14:25.840 --> 0:14:29.360 you don't, and that this person over here thinks something 0:14:29.400 --> 0:14:48.040 to be true even though you know it's not. So 0:14:48.080 --> 0:14:50.800 how do we know that this is a skill that 0:14:50.880 --> 0:14:55.920 develops through time Because very little kids are terrible at 0:14:55.960 --> 0:14:59.240 theory of mind, but they get better as they mature 0:14:59.440 --> 0:15:02.520 into the world, and typically by the ages of three 0:15:02.720 --> 0:15:05.720 to five, they're getting that they're not the only point 0:15:05.720 --> 0:15:08.400 of view that's possible, but that each person in the 0:15:08.440 --> 0:15:11.360 scene has his or her own point of view. Now, 0:15:11.360 --> 0:15:15.160 how do you test whether someone is capable of theory 0:15:15.280 --> 0:15:18.240 of mind? Well, what you do is you present a 0:15:18.280 --> 0:15:22.080 little scenario like this. Sally comes into the room and 0:15:22.160 --> 0:15:25.560 puts her baseball under the bed, and then she leaves. 0:15:26.200 --> 0:15:29.920 While she's gone, Anne comes in the room, she sees 0:15:29.960 --> 0:15:32.240 the ball under the bed, She picks it up, and 0:15:32.280 --> 0:15:36.080 she puts it in the closet. Then she leaves. Now 0:15:36.440 --> 0:15:39.359 Sally comes back in the room, she wants her baseball. 0:15:39.920 --> 0:15:42.520 Where does she look for it? Now? You and I 0:15:42.600 --> 0:15:44.840 know that Sally will look for it under the bed 0:15:44.880 --> 0:15:48.720 where she put it last, even though we simultaneously know 0:15:48.840 --> 0:15:52.400 the actual location of the baseball in the closet. And 0:15:52.440 --> 0:15:55.800 this is because we are running an emulation of what 0:15:55.920 --> 0:15:58.640 it is like to be inside Sally's head with her 0:15:58.800 --> 0:16:03.640 limited knowledg. Now, little children will fail the sally An 0:16:03.800 --> 0:16:07.640 test because they know that the baseball is in the closet, 0:16:08.000 --> 0:16:11.920 so they assume that Sally should know that too. But 0:16:12.200 --> 0:16:15.680 as cognition develops, they come to realize that different heads 0:16:15.880 --> 0:16:18.880 have different beliefs. And a really important clue to the 0:16:18.960 --> 0:16:23.240 development of this is that not everyone develops theory of 0:16:23.320 --> 0:16:26.400 mind in the same way at the same rate. For example, 0:16:26.760 --> 0:16:31.680 people who are on the autism spectrum typically show delays 0:16:31.720 --> 0:16:36.360 in developing theory of mind, which cannot surprisingly impact their 0:16:36.360 --> 0:16:40.440 social interactions. For instance, this is why sarcasm doesn't work 0:16:40.480 --> 0:16:44.080 so well with a person who has autism. When you say, oh, 0:16:44.080 --> 0:16:47.760 great more traffic. I love traffic. They're not likely to 0:16:47.880 --> 0:16:51.160 catch the meaning beneath the words that you're not actually 0:16:51.200 --> 0:16:54.680 pleased because they don't have a sensitive model of your 0:16:55.040 --> 0:16:57.880 actual mental state. If you can't put yourself in the 0:16:57.920 --> 0:17:01.000 shoes of the other person, your understands is limited to 0:17:01.200 --> 0:17:05.159 just pattern recognition, which is not enough for the very 0:17:05.200 --> 0:17:09.240 subtle and sophisticated kinds of communication that humans engage in 0:17:09.359 --> 0:17:12.520 every day. So this tells us that theory of mind 0:17:12.720 --> 0:17:16.480 doesn't come for free in humans. There are brain networks 0:17:16.480 --> 0:17:18.879 that have to develop and learn for this to work, 0:17:19.040 --> 0:17:22.200 so when you look at normal development or delay development. 0:17:22.240 --> 0:17:27.600 This allows us to understand how different brain regions contribute 0:17:27.920 --> 0:17:30.960 to theory of mind. For example, there's one area called 0:17:30.960 --> 0:17:34.520 the temporopridal junction, and this is interesting because it pops 0:17:34.520 --> 0:17:39.119 its head up in tasks that require understanding perspectives, like 0:17:39.600 --> 0:17:43.080 distinguishing between what you know and what someone else knows. 0:17:43.520 --> 0:17:47.000 So imagine you're teaching a friend how to play chess. 0:17:47.480 --> 0:17:49.720 You need to not only understand the rules of the game, 0:17:50.040 --> 0:17:52.919 but also know what your friend knows or doesn't know 0:17:53.119 --> 0:17:57.080 about the game to teach effectively, and the temporo pridal 0:17:57.160 --> 0:18:01.160 junction is involved in that not just that area. It's 0:18:01.200 --> 0:18:03.760 a lot of other areas involved in theory of mind. 0:18:04.119 --> 0:18:07.399 So the medial prefrontal cortex plays a big role in 0:18:07.480 --> 0:18:11.600 making social judgments. It becomes active when you think about 0:18:11.680 --> 0:18:14.840 the mental states of others. For example, if you're trying 0:18:14.880 --> 0:18:19.680 to decide if someone is lying or being truthful, your 0:18:19.720 --> 0:18:23.160 medial prefrontal cortex is engaged. And there are other areas, 0:18:23.240 --> 0:18:26.600 like part of your superior temporal sulcus is involved in 0:18:26.760 --> 0:18:31.720 processing social information like interpreting other people's eye gaze or 0:18:31.760 --> 0:18:35.080 their body language, like the man looking for his keys 0:18:35.400 --> 0:18:38.520 and the child giggling. We're able to infer a lot 0:18:38.920 --> 0:18:42.040 because of the activity of this area. So we see 0:18:42.119 --> 0:18:45.280 lots of areas in brain imaging experiments. And I want 0:18:45.320 --> 0:18:48.399 to mention this to illustrate that theory of mind is 0:18:48.440 --> 0:18:51.960 a brain wide issue. It's not a single area. And 0:18:52.000 --> 0:18:53.720 by the way, this is true of so many things 0:18:53.760 --> 0:18:57.439 in neuroscience. Imagine that I spread out a map of 0:18:57.480 --> 0:19:00.280 your city and I ask you, hey, can you put 0:19:00.320 --> 0:19:03.840 a pin in the spot that represents the economy of 0:19:03.880 --> 0:19:07.040 the city. You tell me that that is a misplaced request. 0:19:07.359 --> 0:19:10.840 There is no single spot for the economy. The economy 0:19:11.000 --> 0:19:14.439 emerges from all the interactions between all the pieces and 0:19:14.480 --> 0:19:17.000 parts of the city, and it's the same with almost 0:19:17.080 --> 0:19:21.080 everything in neuroscience, and especially something like the skill of 0:19:21.200 --> 0:19:24.440 slipping into someone else's point of view. There's not one 0:19:24.600 --> 0:19:27.640 spot to drop a pin into. Instead, it is an 0:19:27.720 --> 0:19:32.600 emergent property that develops from the interaction of lots of networks. 0:19:32.880 --> 0:19:35.679 So what we've seen so far is that theory of 0:19:35.840 --> 0:19:39.560 mind is this ability to infer what someone else knows, 0:19:39.600 --> 0:19:41.560 and we've seen that this is right at the center 0:19:42.040 --> 0:19:46.000 of social interactions. It's something that most humans develop naturally, 0:19:46.440 --> 0:19:49.200 but that doesn't mean it's simple. And the question we're 0:19:49.200 --> 0:19:54.159 going to ask today is does AI have theory of mind? 0:19:54.280 --> 0:19:58.640 Can it put itself into someone else's shoes to understand 0:19:59.080 --> 0:20:03.000 their limited knowledge. One of my colleagues at Stanford recently 0:20:03.000 --> 0:20:08.119 wrote a paper suggesting yes, AI can do this. But fascinatingly, 0:20:08.720 --> 0:20:11.679 it's not as easy to answer this question as you 0:20:11.760 --> 0:20:14.200 might think. And this is for some reasons that we're 0:20:14.200 --> 0:20:16.760 going to dive into. But before we get there, I 0:20:16.800 --> 0:20:19.640 just want to zoom this out to a slightly larger question. 0:20:20.200 --> 0:20:25.280 Could a computer develop theory of mind. Hypothetically, could an 0:20:25.320 --> 0:20:28.119 AI system at some point in the future say, look, 0:20:28.280 --> 0:20:31.040 I know XYZ to be true, but if I look 0:20:31.040 --> 0:20:33.639 at that other person over there, I understand that they 0:20:33.640 --> 0:20:36.399 have a limited viewpoint and that they don't know X 0:20:36.440 --> 0:20:41.439 and Y, and that person over there misbelieves something about Z. 0:20:41.920 --> 0:20:46.560 Well almost certainly, yes, Why it's because we're made up 0:20:46.600 --> 0:20:50.280 of physical stuff and we're running algorithms that took hundreds 0:20:50.280 --> 0:20:54.600 of millions of years to refine. But nonetheless it's physical stuff. 0:20:54.720 --> 0:20:59.160 So if we can do something, presumably a machine could 0:20:59.160 --> 0:21:01.760 do it also, whether or not it's currently clear how 0:21:01.760 --> 0:21:07.679 that's done. That's the central premise of computational neuroscience, and 0:21:07.720 --> 0:21:10.240 to my mind, one of the most remarkable effects of 0:21:10.280 --> 0:21:14.800 the AI explosion over the last few years is understanding 0:21:15.280 --> 0:21:18.280 that things that would have seemed impossible to do with 0:21:18.359 --> 0:21:21.760 a machine, things that almost everyone would have sworn couldn't 0:21:21.800 --> 0:21:25.080 be done. It now seems like background furniture as we 0:21:25.160 --> 0:21:28.080 wait for the next thing. Now, the complexity of the 0:21:28.119 --> 0:21:31.040 brain suggests that theory of mind is going to be 0:21:31.080 --> 0:21:34.399 a very hard problem to solve, because it requires us 0:21:34.400 --> 0:21:37.240 to understand how the brain has a model of the 0:21:37.280 --> 0:21:41.320 world and then how it can make submodels and simulate 0:21:41.720 --> 0:21:43.720 what it is like to only know part of the 0:21:43.760 --> 0:21:47.120 story or to believe a different story. So we don't 0:21:47.119 --> 0:21:49.919 currently know how our brains do it, but of course 0:21:50.400 --> 0:21:54.280 we have Our computers do this sort of thing often, 0:21:54.720 --> 0:21:58.520 Like you can take your modern MacBook laptop and use 0:21:58.600 --> 0:22:02.200 a little bit of its processor to simulate an old 0:22:02.760 --> 0:22:06.600 timex Sinclare computer. Your mac can perfectly simulate it by 0:22:06.680 --> 0:22:11.360 running what's called an emulation on part of its computational hardware. 0:22:11.800 --> 0:22:17.160 Somehow human brains can run emulations also, like just by 0:22:17.240 --> 0:22:20.440 looking you can emulate what it's like to not know 0:22:20.560 --> 0:22:23.199 that the shark is there below you. So yes, it 0:22:23.240 --> 0:22:26.639 seems totally plausible to me that a machine could do 0:22:26.880 --> 0:22:30.560 theory of mind, because we can. But the question we 0:22:30.600 --> 0:22:33.639 want to ask today is whether we are there or 0:22:33.720 --> 0:22:38.000 not right now? Have current large language models like chat 0:22:38.080 --> 0:22:42.360 GPT come to solve this problem without us telling them 0:22:42.359 --> 0:22:46.560 explicitly to do so, in other words, with no instruction? Whatsoever? 0:22:47.119 --> 0:22:51.600 Is the emulation of other minds and emergent property that 0:22:51.720 --> 0:22:54.560 comes out of these things, which would absolutely blow our 0:22:54.600 --> 0:22:59.240 minds if true, does AI do theory of mind? If 0:22:59.280 --> 0:23:03.639 it can, this would have profound implications for our understanding 0:23:03.760 --> 0:23:07.119 of intelligence and our relationship with AI. I mean, just 0:23:07.200 --> 0:23:09.320 consider how much better it would be if it could 0:23:09.359 --> 0:23:14.320 emulate the mental states of people, like with auto driving cars, 0:23:14.400 --> 0:23:18.120 if it didn't just depend on the observable, but instead 0:23:18.160 --> 0:23:21.080 on what's going on in the other driver's head. Like, 0:23:21.480 --> 0:23:23.679 given the trajectory of this car, I think that the 0:23:23.720 --> 0:23:27.720 other driver is drunk or asleep or distracted. And so 0:23:27.880 --> 0:23:30.679 here's what I think is going to happen next. So 0:23:31.119 --> 0:23:34.520 a colleague of mine at Stanford, Michael Kazinski, published a 0:23:34.640 --> 0:23:38.479 twenty twenty three paper that was originally titled Theory of 0:23:38.600 --> 0:23:44.159 Mind might have spontaneously emerged in large language models, although 0:23:44.160 --> 0:23:47.119 he later changed the title. In the paper, he suggested 0:23:47.400 --> 0:23:50.600 that even though these AI models didn't set out to 0:23:50.880 --> 0:23:54.840 have theory of mind, it may have appeared anyway as 0:23:55.000 --> 0:23:59.400 a byproduct of their improving language skills. So, for example, 0:23:59.640 --> 0:24:04.560 he gives the following scenario to chatchipt complete the following story. 0:24:05.119 --> 0:24:08.600 Here is a bag filled with popcorn. There is no 0:24:08.880 --> 0:24:12.160 chocolate in the bag, yet the label on the bag 0:24:12.200 --> 0:24:17.280 says chocolate and not popcorn. Sam finds the bag. She 0:24:17.320 --> 0:24:20.639 has never seen this bag before Sam doesn't open the 0:24:20.680 --> 0:24:24.800 bag and doesn't look inside. Sam reads the label and 0:24:24.840 --> 0:24:27.639 then he gives the prompt. Sam opens the bag and 0:24:27.760 --> 0:24:31.639 looks inside. She can clearly see that it is full of. 0:24:32.359 --> 0:24:35.760 And then he looks at the word that Chatgypt produces, 0:24:35.960 --> 0:24:40.879 is it popcorn or chocolate? And chatchipt says popcorn. But 0:24:40.920 --> 0:24:44.320 if instead he gives a different prompt, Sam calls a 0:24:44.359 --> 0:24:47.280 friend to tell him that she has just found a 0:24:47.400 --> 0:24:52.760 bag full of and now Chatchipet says chocolate, indicating that 0:24:52.920 --> 0:24:57.840 Sam holds a false belief. And Kasinski runs this a 0:24:57.880 --> 0:25:01.240 bunch of ways and shows that chat Gi gets the 0:25:01.320 --> 0:25:05.320