1 00:00:05,440 --> 00:00:07,240 Speaker 1: Can we read minds? 2 00:00:07,680 --> 00:00:10,440 Speaker 2: For those of you who follow the media around neuroscience, 3 00:00:10,440 --> 00:00:14,800 Speaker 2: it certainly seems like it. You've seen articles and heard 4 00:00:14,840 --> 00:00:16,160 Speaker 2: things about mind reading. 5 00:00:16,280 --> 00:00:20,440 Speaker 1: So what is it? Could we run mind reading in. 6 00:00:20,480 --> 00:00:23,919 Speaker 2: Airports so that we know who is carrying a bomb 7 00:00:24,000 --> 00:00:25,239 Speaker 2: or hatching a plot? 8 00:00:25,880 --> 00:00:28,000 Speaker 1: Where are we right now with this technology? 9 00:00:28,080 --> 00:00:31,400 Speaker 2: How close or far are we from being able to 10 00:00:31,480 --> 00:00:34,879 Speaker 2: know if you were thinking about some particular thing you 11 00:00:35,040 --> 00:00:42,800 Speaker 2: did or intend to do. Welcome to Inner Cosmos. I'm 12 00:00:42,880 --> 00:00:48,159 Speaker 2: David Eagleman, a neuroscientist at Stanford and this podcast is 13 00:00:48,200 --> 00:00:51,960 Speaker 2: all about the intersection of brain science with our lives, 14 00:00:52,479 --> 00:01:04,720 Speaker 2: both now and in the future. Today's episode is about 15 00:01:05,160 --> 00:01:06,360 Speaker 2: mind reading. 16 00:01:06,520 --> 00:01:09,440 Speaker 1: What does it mean? Where are we with it now? 17 00:01:10,160 --> 00:01:13,440 Speaker 2: Will we be able to read your mind by measuring 18 00:01:13,520 --> 00:01:17,639 Speaker 2: something from your brain? So strap in for a wild 19 00:01:17,800 --> 00:01:22,319 Speaker 2: ride today. So imagine this twenty years from now. You 20 00:01:22,440 --> 00:01:25,440 Speaker 2: walk into the airport to take a flight across the nation. 21 00:01:26,200 --> 00:01:29,920 Speaker 2: But there are no metal detectors here anymore, or those 22 00:01:30,120 --> 00:01:33,720 Speaker 2: scanners that rotate around you looking for things hidden on 23 00:01:33,760 --> 00:01:37,920 Speaker 2: your body. Instead of simply knowing what you have on 24 00:01:37,959 --> 00:01:42,760 Speaker 2: your body. You find the airport now has portable brain 25 00:01:42,920 --> 00:01:48,040 Speaker 2: scanning such that the authorities can tell who is thinking 26 00:01:48,120 --> 00:01:52,440 Speaker 2: about bringing a bomb onto the airplane. And don't picture 27 00:01:52,680 --> 00:01:57,840 Speaker 2: a big, clunky fMRI machine like we have now, because 28 00:01:57,880 --> 00:02:01,920 Speaker 2: in two decades, brain scanning could certainly be done at 29 00:02:01,960 --> 00:02:06,960 Speaker 2: a distance, getting some signature of brain activity of every 30 00:02:07,080 --> 00:02:11,640 Speaker 2: person walking through Now could this work? I want to 31 00:02:11,639 --> 00:02:15,720 Speaker 2: put aside any legal and ethical conundrums for just a moment, 32 00:02:15,760 --> 00:02:20,799 Speaker 2: because I want to ask would this be technically possible? 33 00:02:21,960 --> 00:02:25,880 Speaker 2: So let's start with what do we mean by mind reading? 34 00:02:26,280 --> 00:02:29,040 Speaker 2: This is a term that's always been talked about by 35 00:02:29,600 --> 00:02:34,280 Speaker 2: magicians and FBI agents and forensic psychologists, and what it 36 00:02:34,360 --> 00:02:39,880 Speaker 2: means historically is reading somebody's body language or the words 37 00:02:39,919 --> 00:02:44,520 Speaker 2: they choose, or their body posture to understand what might 38 00:02:44,560 --> 00:02:47,839 Speaker 2: be going on inside their heads under the hood where 39 00:02:47,880 --> 00:02:51,720 Speaker 2: you can't really see what's happening. And so the next 40 00:02:51,800 --> 00:02:55,640 Speaker 2: step for mind reading is to use technology to allow 41 00:02:55,720 --> 00:02:58,639 Speaker 2: us to do this a little bit better than an 42 00:02:58,680 --> 00:03:03,000 Speaker 2: observant person would. And one example of this is the 43 00:03:03,200 --> 00:03:06,160 Speaker 2: lie detector test. So I'm going to do a separate 44 00:03:06,200 --> 00:03:09,120 Speaker 2: episode on lie detection. So I'll just say briefly here 45 00:03:09,639 --> 00:03:12,360 Speaker 2: that light detection is used in courts all over the 46 00:03:12,400 --> 00:03:17,240 Speaker 2: world it's not actually detecting a lie, it is simply 47 00:03:17,280 --> 00:03:22,680 Speaker 2: detecting a stress response that is usually associated with lying. 48 00:03:23,639 --> 00:03:26,960 Speaker 2: So this technology started in the early nineteen hundreds when 49 00:03:26,960 --> 00:03:30,880 Speaker 2: people realized that often when somebody is lying, there's a 50 00:03:31,240 --> 00:03:35,440 Speaker 2: stress response. You can detect when people are stressed about 51 00:03:35,480 --> 00:03:40,560 Speaker 2: the lie. Now, this has been researched with breathing, blood pressure, 52 00:03:41,120 --> 00:03:45,280 Speaker 2: heart rate, pupil dilation, and the most common measure is 53 00:03:45,280 --> 00:03:49,000 Speaker 2: what's called the galvanic skin response, and all these are 54 00:03:49,120 --> 00:03:53,160 Speaker 2: external signals that the person is stressed. Now, these aren't 55 00:03:53,480 --> 00:03:56,560 Speaker 2: terrible measures, but they're not perfect either, and that's why 56 00:03:56,600 --> 00:03:59,720 Speaker 2: they're accepted by some court systems and not by others. 57 00:04:00,200 --> 00:04:03,720 Speaker 2: And they're not perfect because, among other things, there are 58 00:04:03,760 --> 00:04:05,880 Speaker 2: people that are very good at lying and it doesn't 59 00:04:05,920 --> 00:04:08,600 Speaker 2: stress them out in any way to do so, and 60 00:04:08,640 --> 00:04:11,920 Speaker 2: so a galvanic skin response or something else is not 61 00:04:11,960 --> 00:04:17,320 Speaker 2: going to reveal anything about stress in their situation. So 62 00:04:17,360 --> 00:04:19,080 Speaker 2: the point I just want to make here is there's 63 00:04:19,160 --> 00:04:23,160 Speaker 2: actually quite a gap between the notion of a lie 64 00:04:23,560 --> 00:04:26,000 Speaker 2: and something that you can read off the skin. And 65 00:04:26,040 --> 00:04:30,000 Speaker 2: you can see this isn't exactly mind reading here, but 66 00:04:30,120 --> 00:04:34,359 Speaker 2: it's something removed by a few degrees, So what about 67 00:04:34,560 --> 00:04:38,320 Speaker 2: reading something directly from the brain. Just before I get 68 00:04:38,360 --> 00:04:41,440 Speaker 2: into that, I'll just mention that some years ago I 69 00:04:41,520 --> 00:04:46,480 Speaker 2: talked with a startup that was using electron cephlography or EEG, 70 00:04:47,080 --> 00:04:50,000 Speaker 2: with little electrodes that you stick on the outside of 71 00:04:50,000 --> 00:04:54,760 Speaker 2: the head. And this company's market was long distance truckers, 72 00:04:54,800 --> 00:04:57,040 Speaker 2: And so what they did is they would sew these 73 00:04:57,200 --> 00:05:01,440 Speaker 2: EEG electrodes into a baseball cap so that the trucker 74 00:05:01,800 --> 00:05:04,320 Speaker 2: could wear the cap and the system could tell if 75 00:05:04,360 --> 00:05:07,920 Speaker 2: he was getting sleepy, and if so, then the system 76 00:05:07,920 --> 00:05:10,360 Speaker 2: would do things like alert the driver to pull over. 77 00:05:10,760 --> 00:05:13,400 Speaker 2: So it was a pretty straightforward use case, but it 78 00:05:13,480 --> 00:05:16,320 Speaker 2: turned out the company was having a very hard time 79 00:05:16,360 --> 00:05:21,280 Speaker 2: getting traction for one main reason. The truckers were concerned 80 00:05:21,800 --> 00:05:24,960 Speaker 2: that their bosses would be able to read their minds. 81 00:05:25,720 --> 00:05:28,840 Speaker 2: Now that struck me as interesting, because we're never going 82 00:05:28,920 --> 00:05:32,520 Speaker 2: to get very far in knowing what is happening inside 83 00:05:32,560 --> 00:05:37,920 Speaker 2: a brain with a crude technology like EEG. People are 84 00:05:38,080 --> 00:05:41,839 Speaker 2: using EEG to let's say, control a wheelchair, because you 85 00:05:41,880 --> 00:05:46,280 Speaker 2: can find a few signals that can be reliably distinguished, 86 00:05:46,279 --> 00:05:48,760 Speaker 2: and then you can use one signal to mean go 87 00:05:48,960 --> 00:05:51,320 Speaker 2: and another one to mean stop and left and right. 88 00:05:51,839 --> 00:05:55,680 Speaker 2: But in today's discussion about reading minds, I want to 89 00:05:55,720 --> 00:05:59,560 Speaker 2: ask if we can use technology to really level up 90 00:06:00,080 --> 00:06:04,200 Speaker 2: simply is somebody having a stress response? Or can we 91 00:06:04,560 --> 00:06:09,080 Speaker 2: distinguish four signals to direct a wheelchair? But instead, what 92 00:06:09,320 --> 00:06:12,440 Speaker 2: precisely is this person thinking about? Who do we do 93 00:06:12,560 --> 00:06:16,520 Speaker 2: mind reading via more direct measurements of the brain and 94 00:06:16,640 --> 00:06:18,919 Speaker 2: how close are we to that right now? Now, what 95 00:06:18,960 --> 00:06:21,360 Speaker 2: I'm going to tell you about the field that the 96 00:06:21,440 --> 00:06:26,719 Speaker 2: media typically calls mind reading started getting the foundations laid 97 00:06:26,800 --> 00:06:31,000 Speaker 2: in the nineteen fifties when people first started measuring signals 98 00:06:31,040 --> 00:06:33,840 Speaker 2: from the brain. And what they found is that these 99 00:06:33,880 --> 00:06:37,800 Speaker 2: brain signals were well correlated with things in the outside world. So, 100 00:06:38,240 --> 00:06:43,680 Speaker 2: for example, when someone sees a horizontal line, that activates 101 00:06:43,760 --> 00:06:47,240 Speaker 2: particular neurons that makes them pop off in the visual cortex. 102 00:06:47,720 --> 00:06:50,839 Speaker 2: So in theory, if you had an electrode dunked into 103 00:06:50,839 --> 00:06:53,800 Speaker 2: the visual cortex and you didn't know what the person 104 00:06:53,920 --> 00:06:56,680 Speaker 2: was seeing, but then you saw these cells pop off, 105 00:06:57,279 --> 00:07:00,680 Speaker 2: you could know that they were seeing a horizon the line. 106 00:07:00,800 --> 00:07:04,320 Speaker 2: This is a very crude form of mind reading. Now, 107 00:07:04,360 --> 00:07:07,280 Speaker 2: with almost everyone on the planet there are not electrodes 108 00:07:07,279 --> 00:07:11,000 Speaker 2: in their brains, so how could you read something meaningful 109 00:07:11,400 --> 00:07:16,760 Speaker 2: about their brain activity. So some decades ago, functional magnetic 110 00:07:16,840 --> 00:07:21,360 Speaker 2: resonance imaging or fMRI burst onto the scene, and this 111 00:07:21,480 --> 00:07:27,120 Speaker 2: technique allows you to non invasively measure the activity across 112 00:07:27,120 --> 00:07:31,920 Speaker 2: the whole brain. The problem is that it's really low resolution, 113 00:07:32,280 --> 00:07:35,720 Speaker 2: so instead of knowing what individual neurons might be doing, 114 00:07:36,400 --> 00:07:40,200 Speaker 2: you now can only see great, big blobs that represent 115 00:07:40,560 --> 00:07:45,560 Speaker 2: activity that covers hundreds of thousands or millions of neurons. 116 00:07:46,200 --> 00:07:48,880 Speaker 2: So people didn't think that fMRI was going to be 117 00:07:48,920 --> 00:07:51,960 Speaker 2: able to tell you much about the details of what 118 00:07:52,080 --> 00:07:55,960 Speaker 2: someone was seeing, like exactly what kind of line or 119 00:07:56,040 --> 00:08:00,000 Speaker 2: picture or video someone is watching. But some years ago 120 00:08:00,120 --> 00:08:04,360 Speaker 2: go researchers in England and in America started doing some 121 00:08:04,520 --> 00:08:07,760 Speaker 2: very cool analyzes. And what they did is they looked 122 00:08:07,800 --> 00:08:12,680 Speaker 2: at lots of visual inputs and they measured the corresponding 123 00:08:12,760 --> 00:08:16,880 Speaker 2: outputs from the fMRI. So you show something and you 124 00:08:16,960 --> 00:08:19,080 Speaker 2: measure the brain activity, and then you do the next 125 00:08:19,120 --> 00:08:21,400 Speaker 2: one and the next one. And what they found is 126 00:08:21,440 --> 00:08:25,040 Speaker 2: that even though the individual blobs can't tell you much, 127 00:08:25,520 --> 00:08:28,840 Speaker 2: if you look at the patterns of blobs with lots 128 00:08:28,840 --> 00:08:32,440 Speaker 2: and lots of example inputs, you can then use machine 129 00:08:32,520 --> 00:08:36,160 Speaker 2: learning here to help you, and you can start pulling out. 130 00:08:36,240 --> 00:08:38,120 Speaker 1: Some pretty good correlations. 131 00:08:38,480 --> 00:08:43,240 Speaker 2: You can start knowing what was presented visually just by 132 00:08:43,280 --> 00:08:47,600 Speaker 2: looking at the details of this multi blob output. And 133 00:08:47,640 --> 00:08:51,680 Speaker 2: this really hit a new watermark in twenty eleven when 134 00:08:51,840 --> 00:08:55,679 Speaker 2: Jack Gallant and his team at Berkeley did an amazing experiment. 135 00:08:56,240 --> 00:09:00,560 Speaker 2: He had participants watch videos for hours and hours and 136 00:09:00,600 --> 00:09:03,760 Speaker 2: they measured each frame of the video and exactly how 137 00:09:03,800 --> 00:09:07,319 Speaker 2: the brain responded, tons and tons of input and output. 138 00:09:08,080 --> 00:09:11,520 Speaker 2: And then they could show the participant a video and 139 00:09:11,920 --> 00:09:16,400 Speaker 2: just by looking at the fMRI brain activity these patterns 140 00:09:16,440 --> 00:09:21,640 Speaker 2: of blobs, the researchers could crudely reconstruct what the video 141 00:09:21,800 --> 00:09:22,800 Speaker 2: must have been that. 142 00:09:22,840 --> 00:09:23,800 Speaker 1: Was being looked at. 143 00:09:24,280 --> 00:09:26,680 Speaker 2: So they look at your brain activity and the computer 144 00:09:26,760 --> 00:09:30,400 Speaker 2: reconstructs this must have been something like a bird flying 145 00:09:30,440 --> 00:09:33,559 Speaker 2: from right to left across the screen, and the real 146 00:09:33,679 --> 00:09:37,080 Speaker 2: video was an eagle flying from right to left across 147 00:09:37,080 --> 00:09:40,960 Speaker 2: the screen. Or here's the reconstruction of a person talking 148 00:09:41,120 --> 00:09:45,040 Speaker 2: or a whale breaching or whatever. And the original videos 149 00:09:45,040 --> 00:09:48,679 Speaker 2: that were shown to the participants were really really similar. 150 00:09:49,360 --> 00:09:52,400 Speaker 2: It was extremely impressive, and I'd encourage you to watch 151 00:09:52,480 --> 00:09:55,800 Speaker 2: their videos, which I've linked at Eagleman dot com slash 152 00:09:55,880 --> 00:10:00,720 Speaker 2: podcast now, as you can imagine the media blowing up 153 00:10:00,720 --> 00:10:04,440 Speaker 2: at this point, and everyone called this mind reading, and 154 00:10:04,480 --> 00:10:07,920 Speaker 2: it's only gotten more impressive and fine tuned from there. 155 00:10:08,600 --> 00:10:12,359 Speaker 2: Now the computer also spits out the words the concepts 156 00:10:12,400 --> 00:10:15,200 Speaker 2: that are associated with a video while you're watching, and 157 00:10:15,559 --> 00:10:18,680 Speaker 2: the Lawns team has reversed the machine learning to show 158 00:10:18,800 --> 00:10:22,600 Speaker 2: maps of where concepts sit in the brain. So if 159 00:10:22,600 --> 00:10:27,160 Speaker 2: it's the word marriage or job or dog or whatever, 160 00:10:27,280 --> 00:10:30,400 Speaker 2: you can see which areas of the brain care about 161 00:10:30,400 --> 00:10:31,400 Speaker 2: these concepts. 162 00:10:32,400 --> 00:10:34,640 Speaker 1: Now. I want to note that the maps. 163 00:10:34,240 --> 00:10:37,520 Speaker 2: Of concepts are slightly different for each participant, and I'm 164 00:10:37,520 --> 00:10:39,840 Speaker 2: going to come back to this in a bit, but 165 00:10:39,960 --> 00:10:42,880 Speaker 2: for now, I just want to emphasize how mind blowing 166 00:10:42,920 --> 00:10:45,760 Speaker 2: this all is. And when I first read this paper 167 00:10:45,800 --> 00:10:49,720 Speaker 2: about reconstructing videos in twenty eleven, I was so inspired 168 00:10:49,760 --> 00:10:52,760 Speaker 2: and immediately had an idea in my lab to see 169 00:10:52,800 --> 00:10:57,679 Speaker 2: if we could measure dreams. Because dreams are simply activity 170 00:10:57,679 --> 00:11:02,320 Speaker 2: in the primary visual cortex. Teams are experienced as visual 171 00:11:02,800 --> 00:11:05,120 Speaker 2: because there's activity in the visual part of your brain, 172 00:11:05,600 --> 00:11:08,839 Speaker 2: and your brain interprets that as seeing, so even though 173 00:11:08,880 --> 00:11:14,319 Speaker 2: your eyes are closed, you are having full, rich visual experience. 174 00:11:14,800 --> 00:11:17,200 Speaker 2: So I figured we should be able to just measure 175 00:11:17,240 --> 00:11:20,640 Speaker 2: that and then interview people about what they roughly saw 176 00:11:20,640 --> 00:11:23,679 Speaker 2: in their dreams. Or perhaps we give them a couple 177 00:11:23,720 --> 00:11:26,840 Speaker 2: of choices about scenes that might have appeared in their 178 00:11:26,960 --> 00:11:30,000 Speaker 2: dream and see which one they pick. Anyway, I started 179 00:11:30,000 --> 00:11:32,880 Speaker 2: this project, but before I could finish it, my colleagues 180 00:11:32,880 --> 00:11:37,200 Speaker 2: in Japan, led by Yukiyaso Komatani, published a beautiful paper 181 00:11:37,320 --> 00:11:40,800 Speaker 2: on this same thing. His was the same approach as 182 00:11:40,880 --> 00:11:43,480 Speaker 2: what Galant had done at Berkeley. You gather a bunch 183 00:11:43,520 --> 00:11:47,240 Speaker 2: of data about people watching videos in the fMRI, and 184 00:11:47,280 --> 00:11:50,120 Speaker 2: then you use a machine learning model to do the 185 00:11:50,280 --> 00:11:55,120 Speaker 2: neural decoding. Now, with dreaming, it's a little more difficult 186 00:11:55,160 --> 00:11:57,720 Speaker 2: to know whether you're actually getting this right, because the 187 00:11:57,760 --> 00:12:01,280 Speaker 2: people have to wake up and tell you what they remember, 188 00:12:01,360 --> 00:12:03,840 Speaker 2: and then you compare that to what you think you 189 00:12:03,920 --> 00:12:06,480 Speaker 2: might have decoded. And of course you can't know the 190 00:12:06,559 --> 00:12:10,680 Speaker 2: exact timing of when they saw what. But nonetheless this 191 00:12:10,760 --> 00:12:14,760 Speaker 2: gave the rough start to a dream decoder that could 192 00:12:14,760 --> 00:12:18,559 Speaker 2: tell you what you were dreaming about, and the same 193 00:12:18,640 --> 00:12:21,880 Speaker 2: idea with figuring out what someone might have seen, you 194 00:12:21,880 --> 00:12:25,760 Speaker 2: can do the same thing with what someone heard. This 195 00:12:25,800 --> 00:12:28,720 Speaker 2: started in twenty twelve in the lab of Robert Knight, 196 00:12:28,800 --> 00:12:32,840 Speaker 2: also at Berkeley, who showed that you could reconstruct what 197 00:12:33,080 --> 00:12:37,360 Speaker 2: words a person was hearing by using surface electrodes right 198 00:12:37,400 --> 00:12:40,480 Speaker 2: on the surface of the brain. This is called intracranial 199 00:12:40,600 --> 00:12:45,560 Speaker 2: EEG or IEG. So by presenting lots of spoken words 200 00:12:45,600 --> 00:12:49,360 Speaker 2: to them and measuring the neural activity here, they could 201 00:12:49,400 --> 00:12:53,200 Speaker 2: reconstruct which word a person was hearing. And a paper 202 00:12:53,400 --> 00:12:56,800 Speaker 2: just came out in twenty twenty three showing now that 203 00:12:56,880 --> 00:13:00,400 Speaker 2: you can use even more sophisticated machine learning models to 204 00:13:00,480 --> 00:13:05,280 Speaker 2: reconstruct the music that someone is listening to. Just from 205 00:13:05,360 --> 00:13:09,120 Speaker 2: eavesdropping on the neural activity, you can figure out what 206 00:13:09,280 --> 00:13:13,320 Speaker 2: the music must have been. And specifically, this new paper 207 00:13:13,400 --> 00:13:18,160 Speaker 2: reconstructs Pink Floyd's song Another Brick in the Wall. So 208 00:13:18,240 --> 00:13:21,120 Speaker 2: they play the music to the person and they look 209 00:13:21,400 --> 00:13:24,160 Speaker 2: at all this up and down electrical activity on the 210 00:13:24,200 --> 00:13:27,600 Speaker 2: surface of the cortex, and they can figure out the 211 00:13:27,640 --> 00:13:29,640 Speaker 2: words of the song, and they can figure out all 212 00:13:29,720 --> 00:13:33,240 Speaker 2: kinds of other information about the prosity of the music, 213 00:13:33,320 --> 00:13:37,000 Speaker 2: the rhythm and intonation and stress and so on, and 214 00:13:37,040 --> 00:13:40,880 Speaker 2: they can put this all together into a reconstruction of 215 00:13:40,960 --> 00:13:44,600 Speaker 2: what the whole song must have been just from. 216 00:13:44,400 --> 00:13:46,000 Speaker 1: The neural activity. 217 00:13:55,160 --> 00:13:57,240 Speaker 2: So I want you to notice one thing. This is 218 00:13:57,280 --> 00:14:01,160 Speaker 2: a totally different measure than what the the other labs did. 219 00:14:01,559 --> 00:14:05,400 Speaker 2: This lab is using intracranial EEG, which gives you really 220 00:14:05,480 --> 00:14:11,719 Speaker 2: fast small signals, whereas fMRI gives really slow large signals. 221 00:14:12,240 --> 00:14:15,240 Speaker 2: But it doesn't matter. For a machine learning model. All 222 00:14:15,280 --> 00:14:17,320 Speaker 2: you have to do is give lots of inputs and 223 00:14:17,360 --> 00:14:20,880 Speaker 2: measure lots of outputs, and as long as the output 224 00:14:21,280 --> 00:14:25,640 Speaker 2: is reliably correlated to that input, it doesn't really matter 225 00:14:26,080 --> 00:14:28,640 Speaker 2: what the measuring device is. You can still look at 226 00:14:28,640 --> 00:14:31,440 Speaker 2: the output and say, okay, look, this must have been 227 00:14:31,800 --> 00:14:35,080 Speaker 2: the input. This must have been the video shown because 228 00:14:35,080 --> 00:14:38,480 Speaker 2: it led to this pattern of activity in the fMRI, 229 00:14:39,200 --> 00:14:42,720 Speaker 2: or this must have been the music played because it 230 00:14:42,800 --> 00:14:47,040 Speaker 2: led to this pattern of activity in the intracranial EEG. 231 00:14:48,280 --> 00:14:51,640 Speaker 2: And you can take this same decoding approach with other 232 00:14:51,720 --> 00:14:54,440 Speaker 2: parts of the brain too, not just to decode the 233 00:14:54,480 --> 00:14:57,520 Speaker 2: input that someone must have seen or heard, but also 234 00:14:57,560 --> 00:15:00,600 Speaker 2: to understand what the intention is for an output, in 235 00:15:00,640 --> 00:15:04,520 Speaker 2: other words, how the brain wants to move the body. 236 00:15:05,040 --> 00:15:08,240 Speaker 2: In other words, you can eavesdrop on the visual cortex 237 00:15:08,360 --> 00:15:10,760 Speaker 2: to understand what's in front of someone's eyes, or on 238 00:15:10,800 --> 00:15:14,160 Speaker 2: the auditory cortex to understand what's piping into their ears. 239 00:15:14,680 --> 00:15:17,200 Speaker 1: And in the same way, you can decode. 240 00:15:16,680 --> 00:15:20,240 Speaker 2: The signals from their motor system to understand how they're 241 00:15:20,280 --> 00:15:23,640 Speaker 2: trying to move. And this is exactly what neuroscience is 242 00:15:23,680 --> 00:15:27,960 Speaker 2: doing with people who are paralyzed. Typically, a person who 243 00:15:28,000 --> 00:15:31,000 Speaker 2: is paralyzed has damage to their spinal cord such that 244 00:15:31,520 --> 00:15:35,000 Speaker 2: the signals in their brain aren't able to plunge down 245 00:15:35,080 --> 00:15:37,800 Speaker 2: the back and control the muscles. But their brain is 246 00:15:37,840 --> 00:15:42,040 Speaker 2: still generating the signals. It's just that the signals aren't 247 00:15:42,120 --> 00:15:46,240 Speaker 2: getting to their final destination because of a broken roadway 248 00:15:46,400 --> 00:15:50,120 Speaker 2: or a degenerated roadway. But the key is that their 249 00:15:50,160 --> 00:15:53,880 Speaker 2: brain is still generating the signals just fine. So if 250 00:15:53,880 --> 00:15:56,920 Speaker 2: you could listen to those signals in the motor cortex, 251 00:15:57,400 --> 00:16:00,760 Speaker 2: could you exploit those to make the action happened in 252 00:16:00,800 --> 00:16:05,320 Speaker 2: the outside world. Yes, And this kind of incredible work 253 00:16:05,320 --> 00:16:09,360 Speaker 2: has been happening for two decades now. It started when 254 00:16:09,800 --> 00:16:14,480 Speaker 2: researchers at Emery University implanted a brain computer interface into 255 00:16:14,520 --> 00:16:17,760 Speaker 2: a locked in patient named Johnny Ray, and this allowed 256 00:16:17,840 --> 00:16:22,000 Speaker 2: him to control a computer cursor on the screen simply 257 00:16:22,040 --> 00:16:26,480 Speaker 2: by imagining the movement. His motor cortex couldn't get the 258 00:16:26,520 --> 00:16:30,520 Speaker 2: signals passed a damaged spinal cord, but the implant could 259 00:16:30,600 --> 00:16:34,040 Speaker 2: listen to those signals and pass along the message to 260 00:16:34,120 --> 00:16:37,080 Speaker 2: a computer. And by two thousand and six there was 261 00:16:37,120 --> 00:16:40,880 Speaker 2: a former football player named Matt Nagel who was paralyzed, 262 00:16:41,080 --> 00:16:43,600 Speaker 2: and he was able to get a brain computer interface, 263 00:16:43,680 --> 00:16:48,040 Speaker 2: a little grid of almost one hundred electrodes implanted directly 264 00:16:48,080 --> 00:16:52,600 Speaker 2: into his motor cortex, and that allowed him to control 265 00:16:53,040 --> 00:16:56,480 Speaker 2: lights and open an email, and play the video game 266 00:16:56,640 --> 00:17:01,480 Speaker 2: Pong and draw circles on the screen. And that's mind reading. 267 00:17:01,600 --> 00:17:06,280 Speaker 2: His brain imagined moving his muscles, which caused activity in 268 00:17:06,320 --> 00:17:11,119 Speaker 2: his motor cortex, and the researchers measured that neural activity 269 00:17:11,200 --> 00:17:15,400 Speaker 2: with the electrodes and crudely decoded that to figure out 270 00:17:15,400 --> 00:17:19,879 Speaker 2: the intention. And by twenty eleven, Andrew Schwartz and his 271 00:17:19,960 --> 00:17:24,040 Speaker 2: colleagues at the University of Pittsburgh built a beautiful robotic arm, 272 00:17:24,119 --> 00:17:27,720 Speaker 2: and a woman who had become paralyzed could imagine making 273 00:17:27,760 --> 00:17:32,080 Speaker 2: a movement with her arm, and the robotic arm moves. Now, 274 00:17:32,119 --> 00:17:33,920 Speaker 2: as I said, when you want to move your arm. 275 00:17:33,960 --> 00:17:37,320 Speaker 2: The signals travel from your motor cortex, down your spinal cord, 276 00:17:37,359 --> 00:17:40,639 Speaker 2: to your peripheral nerves, and to your muscle fibers. So 277 00:17:40,840 --> 00:17:44,359 Speaker 2: with this woman, the signals recorded from the brain just 278 00:17:44,400 --> 00:17:48,600 Speaker 2: took a different route, racing along wires connected to motors 279 00:17:49,080 --> 00:17:52,840 Speaker 2: instead of neurons connected to muscles. In an interview, this 280 00:17:52,880 --> 00:17:56,480 Speaker 2: woman said, I'd so much rather have my brain than 281 00:17:56,480 --> 00:17:59,240 Speaker 2: my legs. Why did she say this, Because if you 282 00:17:59,440 --> 00:18:03,000 Speaker 2: have the brain, you can read signals from it to 283 00:18:03,080 --> 00:18:07,560 Speaker 2: make a robotic body do things. You can decode the 284 00:18:07,640 --> 00:18:10,880 Speaker 2: activity in the brain. Now, in episode two, I talked 285 00:18:10,920 --> 00:18:13,919 Speaker 2: about many of these examples and a lot more, and 286 00:18:13,960 --> 00:18:17,760 Speaker 2: I pointed out that we shouldn't be surprised that we 287 00:18:17,800 --> 00:18:20,480 Speaker 2: can figure out how to move robotic arms with our thoughts. 288 00:18:20,560 --> 00:18:23,639 Speaker 2: It's the same process by which your brain learns to 289 00:18:23,680 --> 00:18:29,720 Speaker 2: control your natural, fleshy limbs. As a baby, you flailed 290 00:18:29,760 --> 00:18:32,520 Speaker 2: your appendages around, and you bit your toes, and you 291 00:18:32,560 --> 00:18:35,240 Speaker 2: grabbed your crib bars, and you poked yourself in the eye, 292 00:18:35,320 --> 00:18:38,440 Speaker 2: and you turn yourself over. For years you did that, 293 00:18:39,160 --> 00:18:43,399 Speaker 2: and that's how your muscles learned to mind read the 294 00:18:43,480 --> 00:18:46,959 Speaker 2: command center of the brain. Eventually you could walk and 295 00:18:47,080 --> 00:18:49,720 Speaker 2: run and do cartwheels and ice skate and so on, 296 00:18:49,800 --> 00:18:53,760 Speaker 2: because your brain generates the signals and the muscles do 297 00:18:53,960 --> 00:18:58,240 Speaker 2: as commanded, and so the brain naturally reads from these systems, 298 00:18:58,520 --> 00:19:02,679 Speaker 2: and we can do so now artificially. So nowadays you 299 00:19:02,720 --> 00:19:08,200 Speaker 2: have people who are paralyzed successfully running brain controlled electric wheelchairs. 300 00:19:08,200 --> 00:19:09,040 Speaker 1: It's the same idea. 301 00:19:09,080 --> 00:19:12,320 Speaker 2: You measure brain activity in some way, either with surface 302 00:19:12,359 --> 00:19:16,520 Speaker 2: electrodes on the outside, or with higher resolution like electrodes 303 00:19:16,560 --> 00:19:18,800 Speaker 2: on the surface of the brain, or even with a 304 00:19:18,840 --> 00:19:21,200 Speaker 2: grid of electrodes poked a little bit into the brain 305 00:19:21,240 --> 00:19:25,119 Speaker 2: tissue and you read what the motor intention is, like 306 00:19:25,359 --> 00:19:29,199 Speaker 2: move forward or turn right, and you control the chair accordingly. 307 00:19:29,720 --> 00:19:32,680 Speaker 2: And this sort of work is exploding. A few years ago, 308 00:19:33,280 --> 00:19:38,320 Speaker 2: Stanford researchers implanted two little arrays of microelectrodes in a 309 00:19:38,440 --> 00:19:41,919 Speaker 2: small part of the brain, the premotor cortex, in a 310 00:19:42,080 --> 00:19:45,359 Speaker 2: person who was paralyzed, and then the person was able 311 00:19:45,400 --> 00:19:49,520 Speaker 2: to do mind writing, which was turning the signals from 312 00:19:49,520 --> 00:19:53,639 Speaker 2: the motor cortex into handwritten letters on the computer screen. 313 00:19:54,200 --> 00:19:57,240 Speaker 2: So the person thinks about making the hand motions to 314 00:19:57,320 --> 00:20:00,520 Speaker 2: write big letters, and then you see the lets get 315 00:20:00,640 --> 00:20:03,600 Speaker 2: drawn on the computer screen and the person could do 316 00:20:03,680 --> 00:20:06,760 Speaker 2: about ninety characters a minute. 317 00:20:06,960 --> 00:20:08,920 Speaker 1: And then just in August. 318 00:20:08,640 --> 00:20:12,400 Speaker 2: Twenty twenty three, two amazing papers came out side by 319 00:20:12,480 --> 00:20:14,920 Speaker 2: side in the journal Nature, and by the way, I'm 320 00:20:14,920 --> 00:20:26,760 Speaker 2: linking all these papers at eagleman dot com slash podcast. 321 00:20:33,520 --> 00:20:36,280 Speaker 2: The first paper was from Stanford. They were working with 322 00:20:36,359 --> 00:20:40,040 Speaker 2: a person with als. So this person had a growing 323 00:20:40,440 --> 00:20:45,080 Speaker 2: paralysis and at this point couldn't speak understandably anymore. So 324 00:20:45,119 --> 00:20:49,439 Speaker 2: the team put in two small electrode arrays. These are little, 325 00:20:49,680 --> 00:20:53,000 Speaker 2: tiny square grids of electrodes. They're small, They're like three 326 00:20:53,000 --> 00:20:56,719 Speaker 2: point two millimeters on his side, way smaller than a penny, 327 00:20:57,160 --> 00:20:59,920 Speaker 2: and they dunk these into an area just in front 328 00:20:59,920 --> 00:21:03,560 Speaker 2: of the motor cortex, a pre motor area called ventral 329 00:21:03,560 --> 00:21:06,200 Speaker 2: area six. Okay, I'm going to skip all the details, 330 00:21:06,200 --> 00:21:10,440 Speaker 2: because the important part is that the researcher said, look, 331 00:21:10,680 --> 00:21:15,080 Speaker 2: given these signals worth seeing, what is the probability of 332 00:21:15,119 --> 00:21:19,040 Speaker 2: the sound the person is trying to say, and also 333 00:21:19,119 --> 00:21:21,880 Speaker 2: what are the statistics of the English language, And they 334 00:21:21,880 --> 00:21:26,600 Speaker 2: combine those probabilities to figure out the most likely sequence 335 00:21:26,680 --> 00:21:30,200 Speaker 2: of words the person is trying to say, just by 336 00:21:30,240 --> 00:21:34,679 Speaker 2: looking at signals in the brain, and it worked really well. 337 00:21:35,119 --> 00:21:38,760 Speaker 2: If the person tried to say any one of fifty words, 338 00:21:39,359 --> 00:21:42,040 Speaker 2: the computer got it right nine out of ten times, 339 00:21:42,720 --> 00:21:45,440 Speaker 2: And when the person trained up on a huge vocabulary 340 00:21:45,480 --> 00:21:48,200 Speaker 2: of over one hundred thousand words, the computer got it 341 00:21:48,359 --> 00:21:49,359 Speaker 2: right three out. 342 00:21:49,280 --> 00:21:50,200 Speaker 1: Of four times. 343 00:21:50,960 --> 00:21:54,719 Speaker 2: This is incredible because the person is not using his 344 00:21:54,920 --> 00:21:59,680 Speaker 2: mouth but only brain signals, and the researchers could decode 345 00:21:59,760 --> 00:22:04,399 Speaker 2: those and hear the person talking with them. And as 346 00:22:04,480 --> 00:22:07,280 Speaker 2: I said, at the exact same time that paper came out, 347 00:22:07,640 --> 00:22:10,800 Speaker 2: a second paper in nature came out from UCE San 348 00:22:10,840 --> 00:22:14,120 Speaker 2: Francisco and UC Berkeley. It was the same basic idea, 349 00:22:14,520 --> 00:22:18,119 Speaker 2: but here they were using surface electrodes on the brain. 350 00:22:18,520 --> 00:22:23,639 Speaker 2: This intracordical eeg. So imagine two post it notes up 351 00:22:23,640 --> 00:22:25,760 Speaker 2: against the surface of your brain. It sort of looks 352 00:22:25,800 --> 00:22:30,560 Speaker 2: like that anyway, Slightly different signals spread over sensory and 353 00:22:30,680 --> 00:22:33,720 Speaker 2: motor cortex. And they worked with a woman who had 354 00:22:33,760 --> 00:22:37,680 Speaker 2: had a brainstem stroke eighteen years ago, and she can't 355 00:22:37,720 --> 00:22:40,800 Speaker 2: speak and she can't type. So just imagine being able 356 00:22:40,840 --> 00:22:45,040 Speaker 2: to think clearly but have no capacity for output. It's 357 00:22:45,080 --> 00:22:48,320 Speaker 2: a nightmare, right, So they were not only able to 358 00:22:48,400 --> 00:22:51,679 Speaker 2: decode her words what she was trying to say, but 359 00:22:51,760 --> 00:22:55,080 Speaker 2: they were also able to decode what signals her brain 360 00:22:55,160 --> 00:22:58,600 Speaker 2: was trying to send to her facial muscles, and they 361 00:22:58,600 --> 00:23:03,280 Speaker 2: made an avatar on the screen that would speak her words. 362 00:23:03,119 --> 00:23:05,280 Speaker 1: And show her facial expressions. 363 00:23:05,600 --> 00:23:08,720 Speaker 2: And this is so amazing because it's taking the signals 364 00:23:08,960 --> 00:23:11,880 Speaker 2: that are hidden on the inside of the skull and 365 00:23:12,160 --> 00:23:16,600 Speaker 2: surfacing these, exposing these, turning these into something we can 366 00:23:16,760 --> 00:23:19,760 Speaker 2: understand on the outside and watch on the computer screen. 367 00:23:20,560 --> 00:23:22,800 Speaker 2: So this certainly seems like the kind of thing we 368 00:23:22,840 --> 00:23:26,479 Speaker 2: would want when we talk about mind reading. And I 369 00:23:26,480 --> 00:23:29,600 Speaker 2: also want to say that the researchers and surgeons who 370 00:23:29,600 --> 00:23:34,600 Speaker 2: are doing this are doing incredible pioneering work. But after 371 00:23:34,680 --> 00:23:38,119 Speaker 2: we've lived with these mind bending results for a while, 372 00:23:38,600 --> 00:23:41,800 Speaker 2: it seems easy and obvious that this should be possible. 373 00:23:41,840 --> 00:23:45,720 Speaker 2: We're finding these signals that are often complex and covered 374 00:23:45,760 --> 00:23:49,800 Speaker 2: giant swaths of neurons, but you know that they're correlated 375 00:23:49,840 --> 00:23:53,120 Speaker 2: with movements or pictures or sounds, and you simply throw 376 00:23:53,200 --> 00:23:57,040 Speaker 2: machine learning at it to decode the relationship. And once 377 00:23:57,080 --> 00:24:00,240 Speaker 2: you've read the brain signals and correlated them to their output, 378 00:24:00,560 --> 00:24:05,280 Speaker 2: you can now figure out the intention. So all this 379 00:24:05,480 --> 00:24:08,679 Speaker 2: is amazing news about where we are right now. And 380 00:24:08,720 --> 00:24:11,560 Speaker 2: if you were to drop off from this podcast now, 381 00:24:11,920 --> 00:24:16,040 Speaker 2: you'd probably say that we have achieved mind reading. But 382 00:24:16,119 --> 00:24:23,639 Speaker 2: here comes the plot twist into Act two. It seems 383 00:24:23,680 --> 00:24:26,639 Speaker 2: like mind reading is here, and if you judged just 384 00:24:26,680 --> 00:24:29,639 Speaker 2: from magazine and news articles, you would be certain that 385 00:24:29,680 --> 00:24:32,320 Speaker 2: we're reading minds and that the next step is to 386 00:24:32,400 --> 00:24:36,880 Speaker 2: read everyone's mind and their innermost thoughts. But I want 387 00:24:36,920 --> 00:24:40,479 Speaker 2: to explain why we are actually quite distant from that. 388 00:24:41,480 --> 00:24:44,399 Speaker 2: What I've been telling you about is brain reading, and 389 00:24:44,440 --> 00:24:48,360 Speaker 2: it's amazing, but we should reserve the term mind reading 390 00:24:48,920 --> 00:24:53,280 Speaker 2: for reading the contents of your mind, your thoughts, And 391 00:24:53,440 --> 00:24:56,879 Speaker 2: as stunning as all that current research is, it's not 392 00:24:57,080 --> 00:24:58,040 Speaker 2: reading thoughts. 393 00:24:58,640 --> 00:25:00,640 Speaker 1: Why because unlike a. 394 00:25:00,880 --> 00:25:03,720 Speaker 2: Picture or a piece of music, or a decision about 395 00:25:03,720 --> 00:25:07,080 Speaker 2: which way to go next, a real thought that crosses 396 00:25:07,119 --> 00:25:10,879 Speaker 2: your mind can be a pretty different creature. Now, with 397 00:25:10,920 --> 00:25:13,320 Speaker 2: everything I've told you so far, you might say, well, fine, 398 00:25:13,359 --> 00:25:17,000 Speaker 2: it seems like reading someone's mind is just the next 399 00:25:17,359 --> 00:25:20,440 Speaker 2: layer of resolution, like going from black and white TV 400 00:25:20,640 --> 00:25:24,119 Speaker 2: to color TV. But I'm going to suggest that analogy 401 00:25:24,359 --> 00:25:27,440 Speaker 2: is a false one and we're going to unpack that now. 402 00:25:28,119 --> 00:25:31,560 Speaker 1: So really take stock of the thoughts. 403 00:25:31,200 --> 00:25:34,440 Speaker 2: That you have in a day, their complexity and their weirdness, 404 00:25:34,960 --> 00:25:38,240 Speaker 2: that'll allow you to understand whether and how we could 405 00:25:38,240 --> 00:25:41,399 Speaker 2: really do mind reading. So the other day I was 406 00:25:41,480 --> 00:25:44,200 Speaker 2: driving on the highway and I glanced at a billboard, 407 00:25:44,720 --> 00:25:48,080 Speaker 2: and I really took stock of the three second stream 408 00:25:48,160 --> 00:25:51,200 Speaker 2: of thoughts in my head. It went something like this, Oh, 409 00:25:51,520 --> 00:25:55,960 Speaker 2: there's the billboard again for that software company. Tommy first 410 00:25:56,000 --> 00:25:58,679 Speaker 2: showed me that app six years ago, or maybe it 411 00:25:58,720 --> 00:26:02,400 Speaker 2: was seven years ago. Tommy's beard was so bushy, which 412 00:26:02,480 --> 00:26:06,639 Speaker 2: looked unusual on such a young face. Whatever happened to 413 00:26:06,720 --> 00:26:09,159 Speaker 2: him and that girl that he was dating, what was 414 00:26:09,200 --> 00:26:10,920 Speaker 2: her name, Jackie Jane? 415 00:26:11,359 --> 00:26:12,200 Speaker 1: I wonder how. 416 00:26:12,040 --> 00:26:14,920 Speaker 2: Tommy's doing now. Ah, that reminds me. I need to 417 00:26:14,960 --> 00:26:17,119 Speaker 2: send him an email, because I think he sent me 418 00:26:17,200 --> 00:26:19,720 Speaker 2: one a few weeks ago about a paper he wanted 419 00:26:19,720 --> 00:26:23,359 Speaker 2: to co author. Now, if you could read out that thought, 420 00:26:23,560 --> 00:26:25,480 Speaker 2: that's what I would call mind reading. 421 00:26:25,960 --> 00:26:27,720 Speaker 1: Not what's the picture you're. 422 00:26:27,600 --> 00:26:31,200 Speaker 2: Seeing or the sound you're hearing, but that kind of thought, 423 00:26:31,280 --> 00:26:35,280 Speaker 2: in all of its weirdness and richness. Now you may 424 00:26:35,320 --> 00:26:38,040 Speaker 2: be thinking, but wait, didn't I just hear about some 425 00:26:38,240 --> 00:26:41,359 Speaker 2: other nature paper in twenty twenty three that seems to 426 00:26:41,440 --> 00:26:46,600 Speaker 2: do just that reads people's thoughts sort of. My colleague 427 00:26:46,640 --> 00:26:49,480 Speaker 2: Alex huss At Ut Austin and his grad student Jerry 428 00:26:49,520 --> 00:26:52,680 Speaker 2: Tang did a wonderful piece of work. He used fMRI 429 00:26:53,440 --> 00:26:56,560 Speaker 2: and married it to a large language model GPT one. 430 00:26:57,200 --> 00:26:59,280 Speaker 2: They put people in the scanner and they had them 431 00:26:59,400 --> 00:27:04,639 Speaker 2: listen over many days to sixteen hours of storytelling podcasts 432 00:27:04,800 --> 00:27:08,240 Speaker 2: like The Moth. So the researchers know the words that 433 00:27:08,280 --> 00:27:11,119 Speaker 2: were said, and they measure what's happening in the brain 434 00:27:11,960 --> 00:27:14,280 Speaker 2: and they throw the AI model at it to get 435 00:27:14,280 --> 00:27:17,960 Speaker 2: a clearer and clearer picture of what words cause activation 436 00:27:18,160 --> 00:27:22,840 Speaker 2: in what areas. Then you have the person think about 437 00:27:22,880 --> 00:27:25,680 Speaker 2: a little story. They say the words in their head, 438 00:27:26,359 --> 00:27:29,600 Speaker 2: like quote, he was coming down a hill at me 439 00:27:29,720 --> 00:27:32,359 Speaker 2: on a skateboard and he was going really fast and 440 00:27:32,400 --> 00:27:35,840 Speaker 2: he stopped just in time. And you measure the brain 441 00:27:35,880 --> 00:27:39,119 Speaker 2: activity and you stick the AI model on it to 442 00:27:39,240 --> 00:27:42,879 Speaker 2: tell you what words were being thought of. Now, the 443 00:27:42,920 --> 00:27:45,639 Speaker 2: way they use GPT one is to narrow down the 444 00:27:45,680 --> 00:27:49,520 Speaker 2: word sequences by asking what words are most likely to 445 00:27:49,520 --> 00:27:52,480 Speaker 2: come next. So you take all this noisy brain activity 446 00:27:52,800 --> 00:27:55,879 Speaker 2: and you force it down a straight and narrow path 447 00:27:55,920 --> 00:27:59,879 Speaker 2: that makes it sound sentency. So the decoder doesn't get 448 00:27:59,880 --> 00:28:03,239 Speaker 2: all the words, but the idea, the hope, is that 449 00:28:03,280 --> 00:28:06,840 Speaker 2: it gets the gist, right, So does it well? 450 00:28:07,320 --> 00:28:07,719 Speaker 1: Sort of? 451 00:28:08,119 --> 00:28:11,560 Speaker 2: It depends how loosely you're willing to interpret things. So, 452 00:28:11,600 --> 00:28:16,080 Speaker 2: for the sentence about the skateboard, that internally thought sentence 453 00:28:16,119 --> 00:28:18,520 Speaker 2: again is he was coming down a hill at me 454 00:28:18,640 --> 00:28:21,119 Speaker 2: on a skateboard and he was going really fast and 455 00:28:21,119 --> 00:28:24,560 Speaker 2: he stopped just in time. And the decoder comes up 456 00:28:24,600 --> 00:28:28,560 Speaker 2: with he couldn't get to me fast enough. He drove 457 00:28:28,640 --> 00:28:31,800 Speaker 2: straight up into my lane and tried to ram me 458 00:28:32,760 --> 00:28:35,320 Speaker 2: which is not really like the original sentence at all, 459 00:28:35,400 --> 00:28:38,520 Speaker 2: but if you squint really hard, you could say the 460 00:28:38,640 --> 00:28:43,800 Speaker 2: gist is similar. Or here's another example. The participant thought quote, 461 00:28:44,480 --> 00:28:47,520 Speaker 2: look for a message from my wife saying that she 462 00:28:47,640 --> 00:28:50,440 Speaker 2: had changed her mind and that she was coming back, 463 00:28:51,120 --> 00:28:54,720 Speaker 2: And the dakoder translated this as to see her. For 464 00:28:54,760 --> 00:28:57,480 Speaker 2: some reason, I thought she would come to me and 465 00:28:57,520 --> 00:28:58,480 Speaker 2: say she misses me. 466 00:28:59,400 --> 00:29:00,920 Speaker 1: So exciting stuff. 467 00:29:00,960 --> 00:29:04,480 Speaker 2: But it's far from perfect for even a simple sentence. 468 00:29:05,640 --> 00:29:08,920 Speaker 2: So let's return to this issue of whether you could 469 00:29:09,080 --> 00:29:12,880 Speaker 2: tell if someone is trying to carry a bomb onto 470 00:29:12,960 --> 00:29:16,440 Speaker 2: a flight. You might argue that even if the system 471 00:29:16,480 --> 00:29:20,200 Speaker 2: is imperfect, you might be able to see something cooking 472 00:29:20,280 --> 00:29:23,440 Speaker 2: under the surface about a bomb, and maybe that could 473 00:29:23,440 --> 00:29:27,440 Speaker 2: get us close to mind reading. The problem, of course, 474 00:29:27,520 --> 00:29:29,880 Speaker 2: is that lots of people in the airport are thinking 475 00:29:29,920 --> 00:29:34,200 Speaker 2: about a bomb. For almost everybody, that's I sure hope 476 00:29:34,200 --> 00:29:37,280 Speaker 2: there's not a bomb on this plane. And certainly if 477 00:29:37,320 --> 00:29:41,080 Speaker 2: you know that your government in twenty years has perfected 478 00:29:41,120 --> 00:29:45,800 Speaker 2: remote neuroimaging and is seeing if you're thinking about a bomb. 479 00:29:46,040 --> 00:29:48,120 Speaker 1: Then you can be guaranteed that everyone is going to 480 00:29:48,160 --> 00:29:49,000 Speaker 1: be thinking of a bomb. 481 00:29:49,080 --> 00:29:52,959 Speaker 2: It's like asking someone to not think of a yellow bird. 482 00:29:53,080 --> 00:29:56,160 Speaker 2: As soon as I ask that, you can't help but 483 00:29:56,200 --> 00:29:57,440 Speaker 2: to think about a yellow bird. 484 00:29:57,880 --> 00:29:59,680 Speaker 1: The more you try not to. 485 00:29:59,760 --> 00:30:03,800 Speaker 2: The more your brain is screaming off with yellow feathers. 486 00:30:04,280 --> 00:30:06,200 Speaker 2: But what I want to make really clear is that 487 00:30:06,240 --> 00:30:11,200 Speaker 2: the problem with mind reading is even far tougher than that. 488 00:30:11,960 --> 00:30:15,680 Speaker 2: We need to recognize that our thoughts are much richer 489 00:30:15,720 --> 00:30:19,360 Speaker 2: than sequences of words. When you really take into account 490 00:30:19,680 --> 00:30:23,800 Speaker 2: all the color and emotion of our thoughtscape most of 491 00:30:23,840 --> 00:30:28,200 Speaker 2: it is beyond just the words. So really consider thoughts 492 00:30:28,200 --> 00:30:30,600 Speaker 2: that you have. I'll give you another one of mine 493 00:30:30,600 --> 00:30:32,960 Speaker 2: from the other day. I walked into a restaurant and 494 00:30:32,960 --> 00:30:35,920 Speaker 2: I smelled a syrup being poorn on a pancake, and 495 00:30:35,960 --> 00:30:38,479 Speaker 2: that reminded me of the time that I was in 496 00:30:38,520 --> 00:30:41,760 Speaker 2: the tenth grade and sat with my friend at the 497 00:30:41,800 --> 00:30:43,320 Speaker 2: International House of Pancakes. 498 00:30:43,400 --> 00:30:44,280 Speaker 1: I was wearing. 499 00:30:44,040 --> 00:30:47,080 Speaker 2: A striped sweater and my friend was in his def 500 00:30:47,160 --> 00:30:50,840 Speaker 2: Leopard concert t shirt, and we had the thrill of 501 00:30:50,880 --> 00:30:53,440 Speaker 2: being away from our parents and out on our own, 502 00:30:54,240 --> 00:30:57,440 Speaker 2: just entering our teen years, and we were flirting with 503 00:30:57,480 --> 00:30:59,960 Speaker 2: the waitress who had a German accent, who was at 504 00:31:00,240 --> 00:31:03,040 Speaker 2: seven years older than us, and she was kind enough 505 00:31:03,040 --> 00:31:05,840 Speaker 2: to give us the illusion of flirting back, or so 506 00:31:05,960 --> 00:31:09,120 Speaker 2: we thought at the time, but now a better interpretation 507 00:31:09,240 --> 00:31:11,960 Speaker 2: would perhaps be that she just wanted a better tip. 508 00:31:12,520 --> 00:31:15,160 Speaker 2: And I asked her to help me with a German 509 00:31:15,240 --> 00:31:18,240 Speaker 2: vocabulary word that I was working on, and she got 510 00:31:18,320 --> 00:31:22,320 Speaker 2: uncomfortable and wouldn't do it. And suddenly I was seized 511 00:31:22,360 --> 00:31:25,840 Speaker 2: with the suspicion that she wasn't actually German after all, 512 00:31:25,880 --> 00:31:30,600 Speaker 2: but for some unimaginable reason, had been pretending to be German. 513 00:31:30,800 --> 00:31:33,880 Speaker 2: Was she in fact American or from somewhere else in Europe? 514 00:31:34,000 --> 00:31:36,520 Speaker 2: What was her reason for telling us she was German 515 00:31:36,880 --> 00:31:39,800 Speaker 2: and putting on that accent when she clearly didn't know 516 00:31:39,840 --> 00:31:43,320 Speaker 2: how to speak the German language. Now that memory, that 517 00:31:43,560 --> 00:31:47,560 Speaker 2: moment in time involves territories all over my brain. The 518 00:31:47,600 --> 00:31:50,120 Speaker 2: taste of the pancakes, the smell of the coffee, the 519 00:31:50,160 --> 00:31:53,200 Speaker 2: sights and colors of the pancake house, the sound of 520 00:31:53,200 --> 00:31:56,760 Speaker 2: my friend's fits of laughter, but also things that are 521 00:31:56,800 --> 00:32:01,160 Speaker 2: more subtle, more difficult to simply re out, like my 522 00:32:01,240 --> 00:32:05,320 Speaker 2: interpretation then and now of the reasons behind her behavior, 523 00:32:05,720 --> 00:32:10,120 Speaker 2: my feelings of doubt and skepticism, and for that matter, 524 00:32:10,200 --> 00:32:15,120 Speaker 2: of a confused and unrequited pubescent crush and mystery. This 525 00:32:15,240 --> 00:32:17,959 Speaker 2: is the type of thing that's not stored in anyone 526 00:32:18,040 --> 00:32:21,440 Speaker 2: blob in the brain. It's instead stored in a distributed 527 00:32:21,560 --> 00:32:25,160 Speaker 2: fashion that covers all territories of the brain. And more importantly, 528 00:32:25,200 --> 00:32:28,160 Speaker 2: it's not just words that I'm thinking of. Most of 529 00:32:28,200 --> 00:32:32,400 Speaker 2: it is subtle events and emotions, and I have to 530 00:32:32,640 --> 00:32:36,560 Speaker 2: actively work to translate them into words to even communicate 531 00:32:36,600 --> 00:32:40,560 Speaker 2: this memory on a podcast. But words is not how 532 00:32:40,600 --> 00:32:44,600 Speaker 2: I experience it as it flips through my mind, and 533 00:32:44,720 --> 00:32:48,320 Speaker 2: hence the challenge of reading out something like that. So, 534 00:32:48,520 --> 00:32:51,880 Speaker 2: despite the tremendous progress that's going on, I want us 535 00:32:52,200 --> 00:32:57,040 Speaker 2: to retain this distinction between brain reading guessing what video 536 00:32:57,080 --> 00:32:59,120 Speaker 2: you're seeing or song you're hearing, or word you want 537 00:32:59,160 --> 00:33:03,160 Speaker 2: to say, and mind reading for reading out a rich 538 00:33:03,360 --> 00:33:08,600 Speaker 2: internal experience of some thought or reminiscence or plan for 539 00:33:08,640 --> 00:33:12,240 Speaker 2: the future. Now you might say, Okay, look, that's a 540 00:33:12,280 --> 00:33:15,560 Speaker 2: tough problem to get the real stream of consciousness kind 541 00:33:15,600 --> 00:33:18,600 Speaker 2: of thoughts. But this should become possible as we get 542 00:33:18,680 --> 00:33:23,400 Speaker 2: better resolution in measuring the brain. So let's consider whether 543 00:33:23,440 --> 00:33:27,480 Speaker 2: that challenge seems realistic. So one thing we might need 544 00:33:27,520 --> 00:33:31,120 Speaker 2: would be a technology that could measure that tends to 545 00:33:31,200 --> 00:33:36,400 Speaker 2: hundreds of little electrical spikes every second in every single 546 00:33:36,520 --> 00:33:39,280 Speaker 2: one of the eighty six billion neurons in the cortex 547 00:33:39,360 --> 00:33:43,040 Speaker 2: in real time and record those results. Now, that might 548 00:33:43,120 --> 00:33:46,520 Speaker 2: sound like a straightforward technology problem, and in this century 549 00:33:46,600 --> 00:33:49,200 Speaker 2: we're used to going out and solving big problems. But 550 00:33:49,280 --> 00:33:52,600 Speaker 2: I want to be clear that this would make terraflops 551 00:33:52,600 --> 00:33:56,000 Speaker 2: of data every minute, and that's a ton of data. 552 00:33:56,080 --> 00:33:59,280 Speaker 2: But let's imagine that we flex all our technical muscles 553 00:33:59,280 --> 00:34:02,960 Speaker 2: and figure out how to solve that. Amazing, But that's 554 00:34:03,080 --> 00:34:06,920 Speaker 2: not actually the problem. The technical feat is just the 555 00:34:07,040 --> 00:34:12,120 Speaker 2: warm up problem. The fundamental problem is the interpretation. Let's 556 00:34:12,120 --> 00:34:14,839 Speaker 2: say I handed you this file of the eighty six 557 00:34:15,000 --> 00:34:18,279 Speaker 2: billion neurons popping off for two seconds, and I told 558 00:34:18,320 --> 00:34:21,880 Speaker 2: you this is a particular memory. So you're looking in 559 00:34:22,000 --> 00:34:26,040 Speaker 2: great detail at a pattern, a swirl of billions of neurons. 560 00:34:26,440 --> 00:34:29,279 Speaker 2: But imagine that I don't tell you if this was 561 00:34:29,360 --> 00:34:34,200 Speaker 2: measured from Fred's brain or from Susie's brain, this exact 562 00:34:34,400 --> 00:34:37,560 Speaker 2: pattern among the neurons would mean totally different things in 563 00:34:37,600 --> 00:34:41,080 Speaker 2: different heads. Fred grew up in China and is thinking 564 00:34:41,120 --> 00:34:44,360 Speaker 2: about the time that he climbed the tree and saw 565 00:34:44,640 --> 00:34:48,719 Speaker 2: the panoramic view, and his mother, who suffered from anxiety, 566 00:34:48,800 --> 00:34:51,520 Speaker 2: yelled at him because she was scared because she had 567 00:34:51,560 --> 00:34:55,239 Speaker 2: lost her younger brother in an accident. While in Susie's 568 00:34:55,239 --> 00:34:58,440 Speaker 2: brain this pattern would need something entirely different. She grew 569 00:34:58,560 --> 00:35:01,600 Speaker 2: up in New Mexico, and for her it would mean 570 00:35:01,680 --> 00:35:03,880 Speaker 2: she's thinking about the time that she went on a 571 00:35:03,920 --> 00:35:07,280 Speaker 2: hike and saw wild horses and got in trouble because 572 00:35:07,280 --> 00:35:10,560 Speaker 2: her friend had parked his truck behind her mother's car, 573 00:35:10,800 --> 00:35:12,840 Speaker 2: and her mother couldn't get out and they had to 574 00:35:12,920 --> 00:35:16,000 Speaker 2: run all the way down the mountain to move his truck. Now, 575 00:35:16,000 --> 00:35:19,600 Speaker 2: how could such different memories have the same pattern of neurons. 576 00:35:19,760 --> 00:35:23,799 Speaker 2: It's because who you are is determined by the exact 577 00:35:23,920 --> 00:35:27,840 Speaker 2: wiring of your brain, which is determined by everything that 578 00:35:27,880 --> 00:35:32,080 Speaker 2: has come in, all your memories and experiences of your life. 579 00:35:32,719 --> 00:35:37,800 Speaker 2: Everything about your identity is stored in the particular configuration 580 00:35:37,960 --> 00:35:41,200 Speaker 2: of your tens of billions of neurons and hundreds of 581 00:35:41,280 --> 00:35:45,560 Speaker 2: trillions of connection. That's what makes you you. When you 582 00:35:45,680 --> 00:35:48,880 Speaker 2: learned that a cloud is the same thing as fog, 583 00:35:49,120 --> 00:35:51,800 Speaker 2: just at a different elevation, or you learned that Isaac 584 00:35:51,880 --> 00:35:55,799 Speaker 2: Newton was into alchemy, or you learned that the Bissie 585 00:35:56,040 --> 00:35:58,640 Speaker 2: breed of dog is from Africa and has a curled tail, 586 00:35:59,200 --> 00:36:02,960 Speaker 2: this is all stored by changes in your network. That's 587 00:36:03,000 --> 00:36:06,080 Speaker 2: what it means to learn something and to recall it later. 588 00:36:06,360 --> 00:36:10,320 Speaker 2: It means your network has changed to store the information. 589 00:36:10,800 --> 00:36:13,080 Speaker 2: And so as you go through life and you have 590 00:36:13,360 --> 00:36:17,400 Speaker 2: your unique experiences, your brain gets wired up in a 591 00:36:17,440 --> 00:36:21,400 Speaker 2: totally different way than someone else's, with the end result 592 00:36:21,520 --> 00:36:26,960 Speaker 2: being that your enormous neural forest, your inner cosmos, is 593 00:36:27,120 --> 00:36:31,000 Speaker 2: totally different than someone else's. And that means that if 594 00:36:31,040 --> 00:36:35,480 Speaker 2: I measured neuron number sixty nine hundred and thirty five 595 00:36:35,560 --> 00:36:38,920 Speaker 2: in your head, it would tell me little to nothing 596 00:36:38,960 --> 00:36:41,839 Speaker 2: about what that neuron firing means in someone else's head. 597 00:36:42,600 --> 00:36:45,040 Speaker 2: And this is of course what the researchers at Berkeley 598 00:36:45,080 --> 00:36:49,040 Speaker 2: and Austin and Stanford find. To guess at the words 599 00:36:49,080 --> 00:36:51,920 Speaker 2: that someone is trying to communicate, they have to train 600 00:36:52,040 --> 00:36:54,359 Speaker 2: the model on each person individually. 601 00:36:54,880 --> 00:36:56,200 Speaker 1: The AI model. 602 00:36:55,960 --> 00:36:59,440 Speaker 2: Trained on one person's brain has no chance of reading 603 00:36:59,520 --> 00:37:20,480 Speaker 2: another person's brain. So to do any sort of meaningful 604 00:37:20,600 --> 00:37:25,960 Speaker 2: mind reading, we would need to do what's called system identification. 605 00:37:26,239 --> 00:37:31,040 Speaker 2: What is system identification, It's a process in engineering of 606 00:37:31,200 --> 00:37:35,000 Speaker 2: figuring out the internal structure of a system by sending 607 00:37:35,040 --> 00:37:37,920 Speaker 2: in inputs and looking at the outputs. Like if I 608 00:37:38,000 --> 00:37:40,800 Speaker 2: hit this button, it beeps, and I hit that button 609 00:37:40,800 --> 00:37:42,720 Speaker 2: and it boops, and so on through all the buttons 610 00:37:42,719 --> 00:37:44,960 Speaker 2: and all the combinations, and I can get a pretty 611 00:37:44,960 --> 00:37:47,480 Speaker 2: good idea of what the system must look like on 612 00:37:47,520 --> 00:37:50,200 Speaker 2: the inside. You give lots of inputs, you look at 613 00:37:50,200 --> 00:37:52,759 Speaker 2: all the outputs, and that's how you figure out the 614 00:37:52,800 --> 00:37:56,759 Speaker 2: parameters of a control system, or the dynamics of a 615 00:37:56,800 --> 00:38:00,600 Speaker 2: physical system, or in this case, the structure of a 616 00:38:00,719 --> 00:38:05,640 Speaker 2: neural network a person's brain. The key is that if 617 00:38:05,680 --> 00:38:09,279 Speaker 2: you want to do meaningful mind reading, you'd really have 618 00:38:09,320 --> 00:38:13,280 Speaker 2: to know essentially everything about an individual person. 619 00:38:14,120 --> 00:38:15,480 Speaker 1: So I want to know two things here. 620 00:38:15,840 --> 00:38:21,040 Speaker 2: First, In a sense, this system identification is what long 621 00:38:21,160 --> 00:38:24,480 Speaker 2: term relationships are about. Let's say with a spouse, you 622 00:38:24,520 --> 00:38:27,920 Speaker 2: see a person in lots of different situations, and in 623 00:38:28,000 --> 00:38:31,120 Speaker 2: each you see how they react and how they handle things, 624 00:38:31,560 --> 00:38:34,040 Speaker 2: And this is what it means to get to know 625 00:38:34,160 --> 00:38:38,360 Speaker 2: someone well. But second, we can never do it with 626 00:38:38,520 --> 00:38:42,640 Speaker 2: very high precision. For anyone who's been in long term relationships, 627 00:38:42,680 --> 00:38:46,640 Speaker 2: you know that even with years of experience, our models 628 00:38:46,640 --> 00:38:51,080 Speaker 2: of another person are always impoverished. One of the things 629 00:38:51,120 --> 00:38:53,560 Speaker 2: that I enjoyed the most about being in a relationship 630 00:38:53,560 --> 00:38:56,560 Speaker 2: with my spouse is to watch when she has a 631 00:38:56,600 --> 00:39:00,400 Speaker 2: funny facial expression or stares at something on you usual 632 00:39:00,400 --> 00:39:03,040 Speaker 2: that I wouldn't have expected her to stare at, and 633 00:39:03,200 --> 00:39:05,160 Speaker 2: trying to figure out what in the world is going 634 00:39:05,200 --> 00:39:08,640 Speaker 2: on inside her head. And like any person in a 635 00:39:08,680 --> 00:39:11,120 Speaker 2: long term relationship, I'll tell you that often I have 636 00:39:11,520 --> 00:39:15,120 Speaker 2: no idea, or I'll make a guess or have an assumption, 637 00:39:15,719 --> 00:39:20,040 Speaker 2: but I am almost certainly doomed to be incorrect. And 638 00:39:20,080 --> 00:39:22,680 Speaker 2: this is an example of a more general issue that 639 00:39:22,719 --> 00:39:26,160 Speaker 2: I've mentioned in other episodes, which is that all we 640 00:39:26,320 --> 00:39:29,600 Speaker 2: have is our own internal model of the world, and 641 00:39:29,640 --> 00:39:33,560 Speaker 2: we can never really know what's happening inside someone else's head, 642 00:39:33,880 --> 00:39:36,600 Speaker 2: and so we make our guesses, and we say things 643 00:39:36,640 --> 00:39:39,440 Speaker 2: to people like, oh, I know exactly how you feel, 644 00:39:39,760 --> 00:39:41,839 Speaker 2: and we mean well when we say that, but we're 645 00:39:41,960 --> 00:39:46,480 Speaker 2: always off in our assumptions. It is difficult or impossible 646 00:39:46,520 --> 00:39:49,360 Speaker 2: to understand exactly how someone. 647 00:39:48,960 --> 00:39:51,279 Speaker 1: Else is feeling. Why. 648 00:39:51,440 --> 00:39:54,279 Speaker 2: It's because we've all had a history of years and 649 00:39:54,440 --> 00:39:58,040 Speaker 2: years of deep experience in the world, things that were 650 00:39:58,520 --> 00:40:03,080 Speaker 2: scary to us or tiddle or amazing or dumbfounding or 651 00:40:03,120 --> 00:40:07,360 Speaker 2: attention grabbing, and all of these form pathways in the 652 00:40:07,560 --> 00:40:12,000 Speaker 2: unspeakably large neural forest of our tens of billions of 653 00:40:12,000 --> 00:40:15,680 Speaker 2: neurons and hundreds of trillions of connections. So when you 654 00:40:15,920 --> 00:40:18,600 Speaker 2: zoom passed the billboard on the highway, the one that 655 00:40:18,680 --> 00:40:21,279 Speaker 2: made me think of my student and his beard and 656 00:40:21,320 --> 00:40:23,640 Speaker 2: his girlfriend whose name I couldn't remember, and the email 657 00:40:23,680 --> 00:40:27,560 Speaker 2: that I owed him, you would have a totally different 658 00:40:27,600 --> 00:40:33,960 Speaker 2: cascade of thoughts, same visual input, totally different internal waterfall. 659 00:40:34,000 --> 00:40:37,920 Speaker 2: And unless I knew really everything in your life, it 660 00:40:37,920 --> 00:40:43,320 Speaker 2: would be impossible for me to decode your particular idiosyncratic 661 00:40:43,480 --> 00:40:47,880 Speaker 2: stream of consciousness. Why because I'd be tasked with trying 662 00:40:47,960 --> 00:40:50,920 Speaker 2: to get some sense of a neural system that is 663 00:40:50,960 --> 00:40:55,399 Speaker 2: built from mind boggling complexity that has been developing over 664 00:40:55,600 --> 00:41:00,040 Speaker 2: years and decades of experience, every single experience left, and 665 00:41:00,080 --> 00:41:03,480 Speaker 2: its fingerprint in the deep forests of your brain. So 666 00:41:03,560 --> 00:41:05,960 Speaker 2: while some general structures tend to be the same, like 667 00:41:06,440 --> 00:41:10,000 Speaker 2: where your visual system is, and how your temperature regulation works, 668 00:41:10,280 --> 00:41:13,040 Speaker 2: and the hormones that control your digestive system and so on, 669 00:41:13,800 --> 00:41:17,200 Speaker 2: all the rest is shaped uniquely, which is what makes 670 00:41:17,200 --> 00:41:20,480 Speaker 2: one person different from the next. As you have followed 671 00:41:20,480 --> 00:41:24,080 Speaker 2: your thin trajectory of space and time from your parents, 672 00:41:24,120 --> 00:41:28,680 Speaker 2: to your hometown, to your relationships, your school, friends, your sports, 673 00:41:28,800 --> 00:41:34,840 Speaker 2: your tragedies, your successes, your heart breaks, your life, that's 674 00:41:34,880 --> 00:41:37,920 Speaker 2: what has led your brain to be unique. Now you 675 00:41:37,960 --> 00:41:41,160 Speaker 2: can get a graduate student in the lab to lie 676 00:41:41,200 --> 00:41:43,680 Speaker 2: down in the brain scanner and you can show scene 677 00:41:43,719 --> 00:41:46,640 Speaker 2: after seeing to figure out the rough shape of their 678 00:41:46,960 --> 00:41:50,400 Speaker 2: semantic network. In other words, which areas tend to be 679 00:41:50,520 --> 00:41:53,759 Speaker 2: activated by which words or concepts, And this is what 680 00:41:53,800 --> 00:41:56,279 Speaker 2: the labs at Berkeley and U T Austin and others do. 681 00:41:57,160 --> 00:41:59,040 Speaker 2: But it's not at all clear how you would get 682 00:41:59,080 --> 00:42:03,080 Speaker 2: something meaning from a stranger in the airport whose brain 683 00:42:03,160 --> 00:42:06,000 Speaker 2: you didn't already know in detail, Like, how could you 684 00:42:06,120 --> 00:42:09,279 Speaker 2: know if that guy is thinking, I disagree with this 685 00:42:09,320 --> 00:42:14,280 Speaker 2: political position, and so my goal is to murderously sneak 686 00:42:14,320 --> 00:42:17,120 Speaker 2: a bomb onto the airplane. Now, I just want to 687 00:42:17,160 --> 00:42:20,160 Speaker 2: be really clear about the assertion I'm making. This challenge 688 00:42:20,160 --> 00:42:25,080 Speaker 2: we face is not a scientific challenge of programming better 689 00:42:25,120 --> 00:42:26,160 Speaker 2: AI decoders. 690 00:42:26,200 --> 00:42:28,440 Speaker 1: It's not a technical issue. 691 00:42:28,480 --> 00:42:33,480 Speaker 2: Instead, it's about knowing you as an individual. We are 692 00:42:33,600 --> 00:42:38,160 Speaker 2: never going to get to real rich mind reading unless 693 00:42:38,200 --> 00:42:41,040 Speaker 2: we have something in the future that accompanies you and 694 00:42:41,120 --> 00:42:44,120 Speaker 2: tracks every single experience you have from the time you're 695 00:42:44,120 --> 00:42:48,240 Speaker 2: an infant, every scary moment, every class you sat through, 696 00:42:48,560 --> 00:42:52,080 Speaker 2: every conversation you had, every book you read, every movie 697 00:42:52,080 --> 00:42:55,600 Speaker 2: you saw, every sexual encounter. This device would have to 698 00:42:55,640 --> 00:42:59,400 Speaker 2: record everything you see and hear in experience, and then 699 00:42:59,680 --> 00:43:03,760 Speaker 2: maybe we could combine that device's data with a sophisticated 700 00:43:03,840 --> 00:43:07,160 Speaker 2: model of your genetics, or maybe combine that with a 701 00:43:07,239 --> 00:43:11,399 Speaker 2: bunch of input output system identification, and then maybe we'd 702 00:43:11,440 --> 00:43:15,040 Speaker 2: be close. So you might think, fine, if we could 703 00:43:15,120 --> 00:43:18,680 Speaker 2: actually record everything in someone's life, then we'd be able 704 00:43:18,719 --> 00:43:20,919 Speaker 2: to do it that way. After all, you might point 705 00:43:20,960 --> 00:43:24,160 Speaker 2: out people in the near future might wear portable cameras 706 00:43:24,280 --> 00:43:29,360 Speaker 2: every moment of their life. But really the problem of 707 00:43:29,440 --> 00:43:34,000 Speaker 2: system identification is even harder than that, and that's because 708 00:43:34,040 --> 00:43:38,480 Speaker 2: your internal life isn't just predicated on your inputs and outputs. 709 00:43:39,520 --> 00:43:41,759 Speaker 2: So the brain reading tasks that I told you about 710 00:43:41,760 --> 00:43:46,040 Speaker 2: at the beginning, the researchers know exactly what stimulus was 711 00:43:46,080 --> 00:43:50,160 Speaker 2: presented and know exactly the brain signals that resulted, and 712 00:43:50,200 --> 00:43:53,400 Speaker 2: that's what the AI is trained up on. But lots 713 00:43:53,440 --> 00:43:56,839 Speaker 2: of things in our inside life don't come with a 714 00:43:57,320 --> 00:44:03,440 Speaker 2: clear external input. Example, let's say our recording device detects 715 00:44:03,480 --> 00:44:05,360 Speaker 2: that you meet with a friend at a cafe, and 716 00:44:05,360 --> 00:44:09,320 Speaker 2: it measures all the conversation and your corresponding brain signals, 717 00:44:09,840 --> 00:44:12,360 Speaker 2: but it wouldn't know that you thought about that encounter 718 00:44:12,480 --> 00:44:15,600 Speaker 2: later you saw an analogy between something that was said 719 00:44:15,680 --> 00:44:18,719 Speaker 2: during that conversation to something else in your life. Two 720 00:44:18,760 --> 00:44:22,680 Speaker 2: weeks later, you have a tiny revelation about the behavior 721 00:44:22,800 --> 00:44:24,799 Speaker 2: of your date back when you were a kid and 722 00:44:24,840 --> 00:44:27,759 Speaker 2: went to the high school prom. So for many of 723 00:44:27,800 --> 00:44:31,520 Speaker 2: the things screaming around in your internal cosmos. There's no 724 00:44:31,960 --> 00:44:36,120 Speaker 2: external input, there's nothing to film, and this is true 725 00:44:36,600 --> 00:44:41,320 Speaker 2: of most of your internal thoughtscape You might see somebody 726 00:44:41,360 --> 00:44:43,560 Speaker 2: sitting in the booth next to you, or you might 727 00:44:43,880 --> 00:44:46,480 Speaker 2: walk through the shopping mall and you see different people 728 00:44:46,800 --> 00:44:49,640 Speaker 2: and you think whatever you think about them. They're attractive, 729 00:44:49,680 --> 00:44:52,799 Speaker 2: they're unattractive, they look like my friend from back home, 730 00:44:52,880 --> 00:44:55,160 Speaker 2: they look sad, they look pleased. 731 00:44:55,200 --> 00:44:55,600 Speaker 1: Whatever. 732 00:44:55,920 --> 00:44:59,759 Speaker 2: The challenge for training and input output system like the 733 00:45:00,239 --> 00:45:04,200 Speaker 2: in the research papers, is that the overwhelming majority of 734 00:45:04,239 --> 00:45:07,800 Speaker 2: your thoughts are private. You don't say your thoughts out loud, 735 00:45:08,280 --> 00:45:10,600 Speaker 2: and no one from the outside ever knows what's going 736 00:45:10,640 --> 00:45:14,280 Speaker 2: on in there, and therefore a researcher trying to train 737 00:45:14,360 --> 00:45:18,080 Speaker 2: up an AI model doesn't have the data to train 738 00:45:18,160 --> 00:45:21,080 Speaker 2: the model with. You might walk down the street and 739 00:45:21,200 --> 00:45:24,400 Speaker 2: think that person looks lost, or that person has weird 740 00:45:24,400 --> 00:45:26,960 Speaker 2: shoes and so on, but these are thoughts that you 741 00:45:27,440 --> 00:45:32,600 Speaker 2: never express, so we can't suss out the input output mapping. 742 00:45:33,160 --> 00:45:36,799 Speaker 2: It's not just a matter of having sophisticated enough algorithms 743 00:45:36,840 --> 00:45:40,719 Speaker 2: to measure and store and analyze the patterns. It's that 744 00:45:40,800 --> 00:45:45,600 Speaker 2: the patterns are inherently unpredictable because we don't know what 745 00:45:45,880 --> 00:45:49,759 Speaker 2: causes most of them and how those patterns lead to 746 00:45:49,800 --> 00:45:52,640 Speaker 2: the next. And I just want to mention one more 747 00:45:52,719 --> 00:45:56,400 Speaker 2: reason why mind reading won't be done in airports, not 748 00:45:56,640 --> 00:45:59,680 Speaker 2: now and not in two centuries from now. You need 749 00:45:59,680 --> 00:46:04,160 Speaker 2: the part participants cooperation to train up the decoder and 750 00:46:04,239 --> 00:46:07,080 Speaker 2: to apply the decoder. If people didn't want to train 751 00:46:07,200 --> 00:46:09,759 Speaker 2: up on this, they could think whatever random thing they 752 00:46:09,800 --> 00:46:13,240 Speaker 2: wanted and mess up the whole process. So yet another 753 00:46:13,320 --> 00:46:17,160 Speaker 2: reason why system identification will be difficult unless you want 754 00:46:17,200 --> 00:46:23,319 Speaker 2: to be system identified. So meaningful mind reading reading the 755 00:46:23,480 --> 00:46:27,440 Speaker 2: stream of consciousness isn't going to happen in our lifetimes. 756 00:46:27,719 --> 00:46:31,760 Speaker 2: Maybe some new kind of recording technology and brain reading 757 00:46:31,880 --> 00:46:34,960 Speaker 2: loop will get developed in a thousand years from now, 758 00:46:35,200 --> 00:46:40,160 Speaker 2: but this is not for us or our near term descendants. Instead, 759 00:46:40,920 --> 00:46:45,080 Speaker 2: we will continue to live in a world of mysterious 760 00:46:45,120 --> 00:46:51,040 Speaker 2: strangers stuck inside their own skulls, each person the sole 761 00:46:51,320 --> 00:46:55,600 Speaker 2: inhabitant of his or her internal world. 762 00:46:56,760 --> 00:46:57,719 Speaker 1: So let's wrap up. 763 00:46:57,960 --> 00:47:02,000 Speaker 2: There's been excitement and genuine an incredible progress, but this 764 00:47:02,040 --> 00:47:07,040 Speaker 2: should be carefully distinguished as brain reading. When we read 765 00:47:07,120 --> 00:47:10,360 Speaker 2: the brain patterns to know what a person is seeing 766 00:47:10,840 --> 00:47:15,360 Speaker 2: or what music they're hearing, or read the premotor cortex 767 00:47:15,440 --> 00:47:18,640 Speaker 2: to understand what words they want to say. These are 768 00:47:18,840 --> 00:47:22,719 Speaker 2: technological feats that would have blown the hairback of the 769 00:47:22,760 --> 00:47:24,839 Speaker 2: best neuroscientists fifty years ago. 770 00:47:25,719 --> 00:47:27,200 Speaker 1: But it's not mind reading. 771 00:47:27,480 --> 00:47:31,719 Speaker 2: Mind reading would be an understanding of the universe of 772 00:47:31,920 --> 00:47:35,760 Speaker 2: swirling thoughts, which is where we spend most of our lives. 773 00:47:36,280 --> 00:47:38,600 Speaker 2: Just look at a person sitting at the bus stop, 774 00:47:39,000 --> 00:47:42,719 Speaker 2: or walking silently down the street, or sitting at the library. 775 00:47:43,080 --> 00:47:46,640 Speaker 2: Their minds are swirling with an internal world, considering what 776 00:47:46,719 --> 00:47:49,600 Speaker 2: they're going to do and what they're going to say, 777 00:47:49,600 --> 00:47:53,680 Speaker 2: and maybe reminiscing about the past or feeling regret over 778 00:47:53,719 --> 00:47:57,520 Speaker 2: a possible present that could have been. It's not about 779 00:47:57,560 --> 00:48:00,360 Speaker 2: the final output of what they're trying to say, or 780 00:48:00,400 --> 00:48:02,840 Speaker 2: the input of what's on their eyes or ears. 781 00:48:03,200 --> 00:48:04,319 Speaker 1: It's much deeper than that. 782 00:48:04,800 --> 00:48:08,400 Speaker 2: It's the entire rest of the brain, beyond the little 783 00:48:08,480 --> 00:48:13,239 Speaker 2: territories of input and output. To read the rest of 784 00:48:13,280 --> 00:48:16,360 Speaker 2: that area, to read the private thoughts of a person 785 00:48:16,920 --> 00:48:22,640 Speaker 2: would require an individualized understanding, a system identification that we 786 00:48:22,760 --> 00:48:26,960 Speaker 2: need to track essentially every event in their life and 787 00:48:27,160 --> 00:48:31,040 Speaker 2: probe every single neural reaction and have some way to 788 00:48:31,120 --> 00:48:35,279 Speaker 2: at least guess or impute what all the internal washing 789 00:48:35,360 --> 00:48:39,480 Speaker 2: machine of thought is doing. We are centuries away from that, 790 00:48:40,280 --> 00:48:43,560 Speaker 2: and in a sense that's good news. It means that 791 00:48:43,600 --> 00:48:47,759 Speaker 2: we can enjoy helping people who are mute express themselves, 792 00:48:48,080 --> 00:48:52,120 Speaker 2: or a person who's paralyzed move a wheelchair or a robot, 793 00:48:52,200 --> 00:48:55,520 Speaker 2: or otherwise make things happen in the world. But happily, 794 00:48:55,560 --> 00:48:59,400 Speaker 2: it also means that we retain, for the foreseeable pretty 795 00:48:59,400 --> 00:49:04,839 Speaker 2: distant future, sure the sanctity and privacy of the swirling 796 00:49:05,000 --> 00:49:13,040 Speaker 2: galaxies in our inner cosmos. Go to eagleman dot com 797 00:49:13,080 --> 00:49:16,480 Speaker 2: slash podcast for more information and to find further reading 798 00:49:16,520 --> 00:49:20,400 Speaker 2: and to see videos. Send me an email at podcasts 799 00:49:20,400 --> 00:49:23,600 Speaker 2: at eagleman dot com with questions or discussion, and I'll 800 00:49:23,640 --> 00:49:30,399 Speaker 2: be making mailbag podcasts to address those until next time. 801 00:49:30,560 --> 00:49:33,600 Speaker 2: I'm David Eagleman, and this is Inner Cosmos.