WEBVTT 1 00:00:01.199 --> 00:00:06.200 Welcome to the Sentient Code, where intelligence is engineered, autonomy 2 00:00:06.280 --> 00:00:10.439 is emerging, and a line between human and machine grows thinner. 3 00:00:10.800 --> 00:00:15.359 Each episode, we decode the algorithms, explore the robotics, and 4 00:00:15.439 --> 00:00:19.000 examine the ideas shaping the future of artificial minds. 5 00:00:23.839 --> 00:00:25.760 I want you to do something for me right now. 6 00:00:26.719 --> 00:00:29.719 It's going to sound a little ridiculous, maybe a bit 7 00:00:29.760 --> 00:00:31.800 too simple, but just humor me for a second. Okay, 8 00:00:32.479 --> 00:00:34.520 if you're sitting at a desk, or maybe you're holding 9 00:00:34.560 --> 00:00:37.359 your phone while you're walking, I want you to reach 10 00:00:37.399 --> 00:00:40.520 out and touch something. Just reach your glass of water 11 00:00:40.880 --> 00:00:43.640 or pick up a pin. Go ahead, do it. 12 00:00:43.280 --> 00:00:45.280 It feels instant, doesn't it. 13 00:00:45.280 --> 00:00:48.799 It feels like absolutely nothing. It's seamless. I think I 14 00:00:48.880 --> 00:00:51.119 want that pen and poof my hand. Is there, my 15 00:00:51.159 --> 00:00:54.119 fingers wrap around it, I lift it. It's the most 16 00:00:54.240 --> 00:00:56.600 mundane thing in the world. We do it thousands of 17 00:00:56.640 --> 00:00:59.079 times a day without even registering it. 18 00:00:59.159 --> 00:01:02.320 But if we actually slow that fraction of a second down, 19 00:01:02.399 --> 00:01:04.400 I mean, really, if we could just freeze time right 20 00:01:04.480 --> 00:01:07.959 between the thought and the action, what just happened is 21 00:01:08.159 --> 00:01:13.120 arguably the most complex engineering feet in the known universe. 22 00:01:13.239 --> 00:01:14.879 Okay, hang on that's a big claim. 23 00:01:14.959 --> 00:01:15.799 I don't say that lightly. 24 00:01:15.959 --> 00:01:18.439 Let's unpack that. Yeah, because to me it just felt 25 00:01:18.439 --> 00:01:22.000 like magic. I didn't feel any engineering. I didn't feel 26 00:01:22.000 --> 00:01:24.920 any gears turning or you know, wires firing. 27 00:01:25.239 --> 00:01:28.640 That's because the interface is completely invisible to you. But 28 00:01:28.719 --> 00:01:32.280 here is the reality. Before your muscle fiber even twitched, 29 00:01:32.480 --> 00:01:37.640 a thought existed, an intention right get pin exactly. And 30 00:01:37.680 --> 00:01:42.159 that intention started as a literal storm of electrical activity 31 00:01:42.159 --> 00:01:45.079 in your motor cortex. It's a specific part of your brain, 32 00:01:45.200 --> 00:01:48.319 the scorm I like that. From there, it cascaded down 33 00:01:48.359 --> 00:01:52.000 through your brainstem, It rocketed down your spinal cord, branched 34 00:01:52.040 --> 00:01:55.040 out through all these periperal nerves in your arm, and 35 00:01:55.079 --> 00:01:58.040 then finally it hit the neuromuscular junction to tell your 36 00:01:58.079 --> 00:02:00.439 hand to close. All of that happened and in the 37 00:02:00.439 --> 00:02:01.519 time it took you to just. 38 00:02:01.640 --> 00:02:03.519 Reach It's a biological cascade. 39 00:02:03.680 --> 00:02:07.280 It's a massive data pipeline, and the biology is so fast, 40 00:02:07.439 --> 00:02:10.599 so integrated that for a healthy person, the thought and 41 00:02:10.639 --> 00:02:12.080 the movement feel identical. 42 00:02:12.159 --> 00:02:12.840 There's no lag. 43 00:02:12.960 --> 00:02:15.439 There is no lag, there is no user innercise. You 44 00:02:15.479 --> 00:02:16.360 are the machine. 45 00:02:16.520 --> 00:02:18.400 But and this is the heavy part, the part that 46 00:02:18.439 --> 00:02:22.240 really grounds why we're here today. That circuit isn't guaranteed. No, 47 00:02:22.520 --> 00:02:25.879 it's incredibly fragileve for millions of people. That lightning fast 48 00:02:25.919 --> 00:02:29.120 highway is cut. We're talking about things like als, spinal 49 00:02:29.120 --> 00:02:30.360 cord injuries. 50 00:02:30.000 --> 00:02:32.000 Strokes, praumatic brain injuries. 51 00:02:32.080 --> 00:02:34.840 Yep. The heartbreaking thing I read in the research for 52 00:02:34.879 --> 00:02:39.439 this exploration is the description of what that actually feels like. 53 00:02:40.039 --> 00:02:42.400 It's not that the mind stops working. It's not that 54 00:02:42.439 --> 00:02:43.439 the desire. 55 00:02:43.120 --> 00:02:46.599 Fades precisely, that is the absolute key in many of 56 00:02:46.639 --> 00:02:50.719 these conditions. The motor cortex is still firing perfectly. The 57 00:02:50.759 --> 00:02:54.479 intention to move is there, the plan is perfect, The 58 00:02:54.560 --> 00:02:56.039 storm is brewing. 59 00:02:55.800 --> 00:02:57.759 That the wire is cut. The wire is. 60 00:02:57.759 --> 00:03:00.960 Cut, the signal screams down on the spinal cord, and 61 00:03:01.000 --> 00:03:03.280 it just it hits a dead end. 62 00:03:03.400 --> 00:03:05.560 It's a form of imprisonment. I think that's the phrase 63 00:03:05.560 --> 00:03:08.520 that's stuck with me. A fully intact, active mind locked 64 00:03:08.560 --> 00:03:11.479 inside a body that has stopped responding. It's a terrifying thought, 65 00:03:11.599 --> 00:03:14.560 it is, But it sets the stage perfectly for what 66 00:03:14.639 --> 00:03:17.479 we are exploring today. We are looking at the technology 67 00:03:17.479 --> 00:03:19.439 that is trying to bridge that gap. We're talking about 68 00:03:19.439 --> 00:03:21.240 brain computer interfaces, or. 69 00:03:21.199 --> 00:03:25.520 BCIs this is the frontier. If the biological highway is broken, 70 00:03:25.599 --> 00:03:28.159 the question becomes can we build a digital one? 71 00:03:28.319 --> 00:03:30.280 So we're not trying to fix the old road, we're 72 00:03:30.319 --> 00:03:31.840 building a whole new interstate. 73 00:03:31.960 --> 00:03:35.199 Exactly can we bypass the spinal cord entirely and wire 74 00:03:35.199 --> 00:03:36.800 the brain directly to a machine. 75 00:03:36.840 --> 00:03:39.080 And we aren't just talking about moving a mouse cursor 76 00:03:39.120 --> 00:03:40.840 on a screen, right, I mean that's where it started. 77 00:03:41.280 --> 00:03:43.639 The scope now is huge. 78 00:03:43.719 --> 00:03:46.159 Oh, it's way beyond that. We're talking about controlling advanced 79 00:03:46.240 --> 00:03:49.400 robotic arms. We're talking about synthesizing speech directly from the. 80 00:03:49.360 --> 00:03:52.360 Brain and eventually, well, stuff that sounds like straight up 81 00:03:52.360 --> 00:03:54.080 sci fi human augmentation. 82 00:03:54.400 --> 00:03:55.840 It does sound like sci fi, and it's easy to 83 00:03:55.840 --> 00:03:58.120 get carried away, but I do want to be the 84 00:03:58.199 --> 00:04:01.599 voice of let's say, engineering caution here. 85 00:04:01.479 --> 00:04:02.240 Around us a little. 86 00:04:03.080 --> 00:04:05.960 This is happening now, but it is incredibly hard, I 87 00:04:05.960 --> 00:04:11.800 mean unbelievably difficult. It is a collision of neuroscience, machine learning, robotics, 88 00:04:12.120 --> 00:04:15.840 and crucially ethics. And while the future possibilities are wild, 89 00:04:16.399 --> 00:04:19.800 the current reality is actually in some ways even more 90 00:04:19.839 --> 00:04:22.360 fascinating because of the sheer difficulty of the problem. 91 00:04:22.560 --> 00:04:25.360 So our mission today is to understand how we hack 92 00:04:25.480 --> 00:04:28.680 the brain. How do we listen to that electrical storm 93 00:04:28.720 --> 00:04:30.079 and translate it into action. 94 00:04:30.519 --> 00:04:31.360 Let's get into it. 95 00:04:31.480 --> 00:04:34.399 Okay, So to understand how we plug a computer into 96 00:04:34.399 --> 00:04:37.959 the brain, we first have to understand how the brain talks. 97 00:04:38.800 --> 00:04:41.399 I think most people have this vague idea that it's electrical. 98 00:04:41.480 --> 00:04:44.079 But yeah, it's not like a wall socket, right, you know, 99 00:04:44.279 --> 00:04:46.519 just you know, find the right port and play in 100 00:04:46.560 --> 00:04:47.480 a USB cable. 101 00:04:48.319 --> 00:04:52.079 Not exactly. The brain is an electrochemical organ. It's wet, 102 00:04:52.240 --> 00:04:55.120 it's salty, right, not great for electronics, not at all. 103 00:04:55.160 --> 00:04:59.800 And you have roughly one hundred billion neurons. That number 104 00:04:59.839 --> 00:05:03.399 is staggering, one hundred billion. And when they communicate, they 105 00:05:03.439 --> 00:05:06.600 fire what's called an action potential, and all that is 106 00:05:06.680 --> 00:05:10.680 really is a tiny sharp pulse of voltage. It's caused 107 00:05:10.680 --> 00:05:13.600 by ions, things like sodium and potassium rushing in and 108 00:05:13.639 --> 00:05:14.560 out of a cell membrane. 109 00:05:14.600 --> 00:05:16.759 So it's a little chemical battery firing off a spark. 110 00:05:17.040 --> 00:05:19.120 That's a great way to think about it, a tiny 111 00:05:19.160 --> 00:05:21.600 little spark. And this isn't a sporadic thing. This is 112 00:05:21.639 --> 00:05:23.800 happening constantly all over your. 113 00:05:23.639 --> 00:05:27.000 Brain, even when you're just sitting here doing nothing constantly. 114 00:05:27.199 --> 00:05:29.800 It's a symphony of noise. Even when you are sleeping, 115 00:05:29.879 --> 00:05:34.680 it's humming with activity. But for our purposes today, for BCI, 116 00:05:34.879 --> 00:05:37.079 we are interested in a very specific part of the 117 00:05:37.079 --> 00:05:39.199 brain called the primary motor cortex. 118 00:05:39.319 --> 00:05:42.319 Okay, walk us through the geography. Here. Where is this 119 00:05:42.480 --> 00:05:44.000 command center for movement? 120 00:05:44.399 --> 00:05:47.800 Imagine a headband, a strip of tissue running roughly from 121 00:05:47.839 --> 00:05:49.560 ear to ear right across the top of your head. 122 00:05:50.079 --> 00:05:52.720 That is your command center for voluntary movement. 123 00:05:53.040 --> 00:05:56.639 And everything I consciously decide to move command starts there. 124 00:05:56.839 --> 00:06:01.079 Every single voluntary movement, Yes, from wiggling your to speaking 125 00:06:01.399 --> 00:06:01.800 a word. 126 00:06:01.920 --> 00:06:04.759 I've seen those diagrams of the homunculus. Yeah, it's that 127 00:06:04.839 --> 00:06:07.360 weird distorted little man drawn over the brain. It's a 128 00:06:07.399 --> 00:06:08.879 map right of the body. 129 00:06:08.959 --> 00:06:10.519 It is a map, a very famous one. If you 130 00:06:10.519 --> 00:06:11.600 were to open up the skull and. 131 00:06:11.600 --> 00:06:14.000 Please don't do this at humeh right, disclaimer. 132 00:06:13.759 --> 00:06:17.199 And you were to electrically stimulate specific spots on that strip, 133 00:06:17.639 --> 00:06:21.560 you would see specific body parts twitch. Stimulate. Here, the 134 00:06:21.560 --> 00:06:24.839 thumb moves move a centimeter over, the lip twitches. 135 00:06:24.959 --> 00:06:26.360 So it's one to one map it is. 136 00:06:26.560 --> 00:06:29.680 But here is the catch. The map isn't drawn to 137 00:06:29.720 --> 00:06:33.040 scale based on physical size. It's drawn to scale based 138 00:06:33.079 --> 00:06:34.319 on complexity of movement. 139 00:06:34.399 --> 00:06:38.319 Ah okay, so my back is physically huge, but on 140 00:06:38.360 --> 00:06:39.720 the brain map it's tiny. 141 00:06:40.040 --> 00:06:42.560 Your back takes up a minuscule amount of space on 142 00:06:42.560 --> 00:06:44.839 that map. You don't need fine motor control for your 143 00:06:44.879 --> 00:06:45.480 lower back. 144 00:06:45.639 --> 00:06:47.920 I just needed to, you know, hold me up right. 145 00:06:48.160 --> 00:06:51.720 But your hands, your face, your tongue, they take up 146 00:06:51.839 --> 00:06:54.279 massive amounts of real estate on that strip. Why the 147 00:06:54.319 --> 00:06:57.439 amount of neural processing power required to move your thumb 148 00:06:57.439 --> 00:07:00.720 and forefinger with precision to pick up a dime is well, 149 00:07:00.759 --> 00:07:04.040 it's astronomical compared to the processing power it takes to 150 00:07:04.079 --> 00:07:05.160 move your leg to walk. 151 00:07:05.319 --> 00:07:09.120 Because of the decks terity involved. We have incredible fine 152 00:07:09.319 --> 00:07:11.920 motor skills in our hands. Speech is another one, I 153 00:07:11.920 --> 00:07:13.079 imagine exactly. 154 00:07:13.160 --> 00:07:15.959 And this is where the engineering challenge really begins. When 155 00:07:16.000 --> 00:07:18.000 you decide to move that hand. It's not just one 156 00:07:18.079 --> 00:07:20.519 neuron raising a little flag saying move left. 157 00:07:20.600 --> 00:07:22.720 That would be too easy. If it were one neuron, 158 00:07:22.800 --> 00:07:24.800 we could just find that one guy and listen to them. 159 00:07:25.000 --> 00:07:27.920 It would be vastly easier. We'd have solved this decades ago. 160 00:07:28.480 --> 00:07:33.000 But the brain uses something called population coding. A population, 161 00:07:33.399 --> 00:07:37.560 so a crowd, it's a distributed chorus. It's thousands of 162 00:07:37.680 --> 00:07:41.839 neurons firing in a complex, coordinated pattern. And they aren't 163 00:07:41.879 --> 00:07:47.160 just saying left. They are simultaneously encoding direction and speed 164 00:07:47.399 --> 00:07:50.519 and force and the kinematics of the whole trajectory, all 165 00:07:50.519 --> 00:07:53.639 at once, all at once. It is a massive, noisy, 166 00:07:53.720 --> 00:07:56.759 high dimensional data stream. So if I'm a BCI engineer, 167 00:07:56.800 --> 00:07:59.000 I'm not looking for a single switch I can flip. 168 00:07:59.439 --> 00:08:02.040 I'm trying to to a choir of thousands of people 169 00:08:02.079 --> 00:08:04.519 screaming slightly different instructions at once, and I have to 170 00:08:04.519 --> 00:08:05.800 figure out what song they're singing. 171 00:08:05.879 --> 00:08:08.199 That is a very very good analogy, and the problem 172 00:08:08.279 --> 00:08:12.319 gets even worse. Oh great, because usually the brain is cheating, cheating. 173 00:08:12.680 --> 00:08:13.639 How does the brain cheat? 174 00:08:13.759 --> 00:08:17.079 Feedback a constant stream of feedback. When you move your 175 00:08:17.079 --> 00:08:18.839 hand to pick up that water glass, you weren't just 176 00:08:18.839 --> 00:08:21.680 sending a command down. You are receiving data up. 177 00:08:22.160 --> 00:08:24.879 Your eyes see the hand moving, Your skin feels the 178 00:08:24.879 --> 00:08:28.439 air moving past it. Your muscles and joints have sensors 179 00:08:28.480 --> 00:08:31.600 proprioception that tell you exactly where your arm is in 180 00:08:31.600 --> 00:08:33.519 space without you even looking right. 181 00:08:33.679 --> 00:08:35.360 I can touch my nose with my eyes closed. 182 00:08:35.559 --> 00:08:40.039 That's proprioception, and your brain is constantly, millisecond by millisecond 183 00:08:40.159 --> 00:08:43.039 adjusting the plan based on that rich stream of feedback. 184 00:08:43.120 --> 00:08:44.840 Oh, I am a little too far to the left. 185 00:08:44.919 --> 00:08:48.200 Correct course, Wait, the glass is heavier than I thought. 186 00:08:48.240 --> 00:08:49.799 Apply more force constantly. 187 00:08:49.799 --> 00:08:52.639 It's a closed loop system. When we build a BCI, 188 00:08:52.799 --> 00:08:55.960 we are often flying blind. We are taking the output command, 189 00:08:56.039 --> 00:08:59.000 the shout from the choir, but we usually can't give 190 00:08:59.039 --> 00:09:02.000 the brain the same rich sensory feedback it's used to. 191 00:09:02.200 --> 00:09:05.279 We're intercepting the shout, but the brain can't hear the echo. 192 00:09:05.159 --> 00:09:08.159 Exactly, and that makes it the control problem so much harder. 193 00:09:08.200 --> 00:09:10.519 So we know the brain is shouting these electrical commands, 194 00:09:11.080 --> 00:09:14.879 the next big question is how do we listen? And 195 00:09:15.279 --> 00:09:17.519 looking at the research, this seems to be the biggest 196 00:09:17.519 --> 00:09:20.360 debate in the field right now. It's this fundamental trade off. 197 00:09:20.440 --> 00:09:23.600 It is the trade off resolution versus invasiveness. 198 00:09:24.159 --> 00:09:25.799 To cut or not to cut. 199 00:09:25.679 --> 00:09:27.240 That's a good way to put it. We can really 200 00:09:27.240 --> 00:09:30.120 think of this as three different scales of listening, three 201 00:09:30.159 --> 00:09:31.759 ways to eavesdrop on the brain. 202 00:09:32.120 --> 00:09:34.559 Okay, let's start with the most intense one, the one 203 00:09:34.559 --> 00:09:37.679 that requires a sterile operating room and a neurosurgeon. 204 00:09:38.440 --> 00:09:41.879 The finest scale is single unit recording. This means we 205 00:09:41.960 --> 00:09:46.080 are putting electrodes directly inside the brain tissue. We are 206 00:09:46.120 --> 00:09:49.320 physically penetrating the cortex and getting right up next to 207 00:09:49.360 --> 00:09:51.120 the cell bodies of the neurons. 208 00:09:51.440 --> 00:09:54.120 This is open brain sugery. We are drilling a hole 209 00:09:54.159 --> 00:09:54.679 in the skull. 210 00:09:54.840 --> 00:09:57.480 We are There's no way around it. But the advantage 211 00:09:57.519 --> 00:10:00.840 is clarity. The signal to noise ratio is incredible. 212 00:10:01.200 --> 00:10:04.200 Going back to your acquire analogy, what does this get us? 213 00:10:04.519 --> 00:10:06.360 This is like putting a microphone right in front of 214 00:10:06.440 --> 00:10:10.480 the soloist's mouth. You hear every breath, every subtle change, 215 00:10:10.519 --> 00:10:14.360 and pitch every single note. The precision is extraordinary, and. 216 00:10:14.279 --> 00:10:18.159 With that precision you can decode much more complex movements precisely. 217 00:10:18.519 --> 00:10:21.720 You can start to pull out signals for individual finger movements. 218 00:10:21.720 --> 00:10:25.720 For example, you're stabbing the brain, I mean gently. I'm sure, 219 00:10:25.799 --> 00:10:28.200 but that can't be good for long term health. 220 00:10:28.399 --> 00:10:31.480 It's not ideal. The brain is a very hostile environment 221 00:10:31.480 --> 00:10:35.440 for electronics. It's salty, it pulsates with every heartbeat, and 222 00:10:35.480 --> 00:10:38.600 the immune system absolutely hates foreign objects. 223 00:10:39.240 --> 00:10:41.360 It treats the electrode like a splinter. 224 00:10:41.120 --> 00:10:44.240 Exactly like a splinter. Over time, you get scarring. It's 225 00:10:44.279 --> 00:10:47.799 called glial scarring, where the brain's immune cells build up 226 00:10:47.840 --> 00:10:50.559 a wall of tissue around the electrode. 227 00:10:50.000 --> 00:10:51.559 And that pushes the neurons away. 228 00:10:51.720 --> 00:10:54.240 It pushes the neurons away from the electrode, the microphone 229 00:10:54.240 --> 00:10:58.159 gets muffled, the signal degrades. This long term stability is 230 00:10:58.159 --> 00:10:59.919 one of the biggest challenges for ENVA. 231 00:11:00.159 --> 00:11:03.759 Of BCIs okay, So that's that high risk, high reward option, 232 00:11:04.399 --> 00:11:08.080 maximum data, but maximum danger, and it degrades over time. 233 00:11:08.279 --> 00:11:09.320 What's the middle ground? 234 00:11:09.600 --> 00:11:13.799 The middle ground is usually called local field potentials or LFPs. 235 00:11:14.120 --> 00:11:18.200 Another term you'll here is ECoG, which stands for electrocordicography. 236 00:11:18.440 --> 00:11:20.080 So what's the difference in approach here? 237 00:11:20.279 --> 00:11:23.000 Instead of poking the sharp electrodes inside the tissue, you 238 00:11:23.080 --> 00:11:26.120 lay a flexible grid or a mat of electrodes directly 239 00:11:26.159 --> 00:11:28.559 on the surface of the brain, under the skull, but 240 00:11:28.639 --> 00:11:31.519 on top of the gray matter. Ah, So you're still 241 00:11:31.559 --> 00:11:35.480 inside the skull, which is invasive, but you aren't penetrating 242 00:11:35.480 --> 00:11:36.720 the brain tissue itself. 243 00:11:36.799 --> 00:11:40.639 Correct, you aren't triggering that same aggressive immune response. 244 00:11:41.320 --> 00:11:43.200 So in our choir analogy, where are we now? 245 00:11:43.240 --> 00:11:46.279 You aren't hearing the soloist anymore. You're hearing the entire 246 00:11:46.320 --> 00:11:49.440 soprano section as a group. You know they're singing, you 247 00:11:49.480 --> 00:11:51.960 know the general melody, but you can't pick out an 248 00:11:51.960 --> 00:11:52.879 individual voice. 249 00:11:52.919 --> 00:11:56.120 So you're summing up the activity of thousands of neurons. 250 00:11:57.159 --> 00:12:01.039 Less detail, but maybe safer and more stable over time. 251 00:12:01.200 --> 00:12:05.080 Exactly. It's a compromise. You lose that fine grain specific 252 00:12:05.120 --> 00:12:08.159 data you might need for say, playing a piano with 253 00:12:08.200 --> 00:12:11.840 a robotic hand, but it's much more stable long term. 254 00:12:11.960 --> 00:12:15.320 Which brings us to option three, the one everyone wants 255 00:12:15.320 --> 00:12:16.840 to work because it doesn't involve. 256 00:12:16.559 --> 00:12:18.480 A drill, the Holy Grail, non. 257 00:12:18.360 --> 00:12:21.679 Invasive, the headset, the EEG cap that you just put on. 258 00:12:21.679 --> 00:12:25.440 Right EEG or electro and cephlography. You're just placing electrodes 259 00:12:25.480 --> 00:12:28.440 on the scalp with some conductive gel, no surgery, just 260 00:12:28.480 --> 00:12:30.200 a funny looking hat with a lot of wires. 261 00:12:30.279 --> 00:12:33.600 It sounds perfect, that's the catch. The catch is the skull. 262 00:12:34.360 --> 00:12:38.120 The skull is a massive electrical insulator. It smears and 263 00:12:38.120 --> 00:12:40.519 distorts the tiny electrical signals from the. 264 00:12:40.440 --> 00:12:42.799 Brain and the sources used a great analogy for this. 265 00:12:43.159 --> 00:12:45.080 They said, it's like standing in the parking lot of 266 00:12:45.080 --> 00:12:48.200 a football stadium and trying to listen to the conversation 267 00:12:48.399 --> 00:12:50.440 of two people on the fifty yard line. 268 00:12:50.480 --> 00:12:53.000 It's a perfect analogy. You can hear the roar of 269 00:12:53.039 --> 00:12:55.679 the crowd, you know when a touchdown happens, because the 270 00:12:55.720 --> 00:12:59.320 whole stadium erupts. You can tell if the crowd is excited. 271 00:12:59.000 --> 00:13:01.720 Or bored, big large scale signals. 272 00:13:01.840 --> 00:13:05.919 Exactly, you can detect large scale brain states, but you 273 00:13:06.039 --> 00:13:10.240 cannot hear the quarterback calling the play. The specific detailed 274 00:13:10.279 --> 00:13:12.080 commands are lost in the noise. 275 00:13:12.360 --> 00:13:14.000 So if I want to control a cursor on a 276 00:13:14.039 --> 00:13:17.720 screen to just lect yes or no, the stadium roar 277 00:13:17.799 --> 00:13:18.399 is enough for that. 278 00:13:18.679 --> 00:13:21.759 Yes, that's a big signal. We can detect that. But 279 00:13:21.799 --> 00:13:23.919 if you want to control a robotic hand to gently 280 00:13:23.919 --> 00:13:25.960 pick up a grape without squashing. 281 00:13:25.480 --> 00:13:26.960 It, you need to hear the quarterback. 282 00:13:27.039 --> 00:13:28.679 You need to hear the quarterback. You need to be 283 00:13:28.799 --> 00:13:29.840 inside the stadium. 284 00:13:30.159 --> 00:13:32.639 That is the cool irony of this whole field, isn't it. 285 00:13:32.639 --> 00:13:36.480 It feels like the central conflict. To get the magical 286 00:13:36.559 --> 00:13:39.840 sci fi dexterity that could change lives, you currently have 287 00:13:39.919 --> 00:13:42.679 to drill a hole in someone's skull, that's right. But 288 00:13:42.799 --> 00:13:46.360 to make the technology accessible and safe for everyone, you 289 00:13:46.440 --> 00:13:49.919 have to use the non invasive methods, which for now 290 00:13:50.320 --> 00:13:53.159 lose too much fidelity for those complex tasks. 291 00:13:53.519 --> 00:13:56.759 That is the central tension, The engineering dilemma that every 292 00:13:56.799 --> 00:13:59.639 single lab in company in this space is wrestling with 293 00:14:00.200 --> 00:14:03.440 is essentially how can we get the highest resolution signal 294 00:14:03.639 --> 00:14:06.399 with the absolute minimum amount of harm to the patient. 295 00:14:06.600 --> 00:14:09.000 So let's assume we've made that choice. We've gone for 296 00:14:09.000 --> 00:14:11.200 the high fidelity option. We've drilled the whole, we've got 297 00:14:11.240 --> 00:14:14.519 the wires in, we're listening to the soloist. Now comes 298 00:14:14.519 --> 00:14:16.919 the part that honestly breaks my brain a little bit. 299 00:14:16.840 --> 00:14:20.879 The decoding the machine learning problem. This is where neuroscience 300 00:14:20.919 --> 00:14:23.080 stops and computer science really begins. 301 00:14:23.159 --> 00:14:27.159 Because the brain isn't sending computer code, it's not sending English, 302 00:14:27.639 --> 00:14:30.840 it's sending what sounds like static. How on earth do 303 00:14:30.879 --> 00:14:33.759 we turn a storm of electrical noise into the command 304 00:14:34.039 --> 00:14:35.639 move robot arm left. 305 00:14:35.759 --> 00:14:39.200 This is where the progress in AI and machine learning 306 00:14:39.240 --> 00:14:42.039 over the last decade has completely changed the game. In 307 00:14:42.080 --> 00:14:44.519 the old days. By old days, I mean, you know, 308 00:14:45.039 --> 00:14:50.919 ten fifteen years ago, we used relatively simple linear models, simple. 309 00:14:50.639 --> 00:14:53.279 Math, like more firing in this area equals more speed 310 00:14:53.320 --> 00:14:53.919 in that direction. 311 00:14:54.159 --> 00:14:57.480 Relatively Yes, it was essentially drawing a straight line. If 312 00:14:57.519 --> 00:15:01.000 this neuron fires fast, move hand right, that one fires fast, 313 00:15:01.080 --> 00:15:04.519 move hand up, and surprisingly that works for basic stuff. 314 00:15:04.639 --> 00:15:07.480 Really yeah, the brain is robust enough that even a 315 00:15:07.519 --> 00:15:10.480 simple MAC like that captures something meaningful. But it hit 316 00:15:10.519 --> 00:15:14.360 a ceiling. You couldn't do complex, fluid, coordinated movements. It 317 00:15:14.399 --> 00:15:15.799 was always jerky and slow. 318 00:15:16.120 --> 00:15:17.799 Enter deep learning exactly. 319 00:15:18.080 --> 00:15:21.039 Now we use much more sophisticated tools. We use recurrent 320 00:15:21.080 --> 00:15:24.159 neural networks RNNs and transformers, the same kind of AI 321 00:15:24.240 --> 00:15:27.919 architecture behind things like chat GPT, but applied to neural 322 00:15:27.960 --> 00:15:29.279 signals instead of words. 323 00:15:29.679 --> 00:15:31.879 How does that work? How do you apply a language 324 00:15:31.879 --> 00:15:32.879 model to brainwaves? 325 00:15:33.360 --> 00:15:36.320 We stop thinking about individual neurons and start treating the 326 00:15:36.320 --> 00:15:38.840 neural activity as a trajectory through what we call high 327 00:15:38.879 --> 00:15:39.720 dimensional space. 328 00:15:39.960 --> 00:15:43.759 Okay, hold on, you're losing me. High dimensional space. Explain 329 00:15:43.840 --> 00:15:44.919 that to me like I'm five. 330 00:15:45.200 --> 00:15:47.679 Okay, Think of it this way. Instead of looking at 331 00:15:47.679 --> 00:15:50.159 one neuron at a time, the AI looks at the 332 00:15:50.200 --> 00:15:53.159 state of the entire population of neurons. We're recording from 333 00:15:53.679 --> 00:15:56.759 say one hundred of them, all at once, at a 334 00:15:56.799 --> 00:15:58.120 single snapshot in time. 335 00:15:58.519 --> 00:16:01.480 So it's looking at the whole choir, not just one singer. 336 00:16:01.559 --> 00:16:04.639 The whole choir, and it does this many times a second. 337 00:16:05.519 --> 00:16:08.879 Imagine a flock of birds turning in the sky. If 338 00:16:08.919 --> 00:16:13.240 you only watch one bird, its movement seems chaotic and unpredictable, right, 339 00:16:13.480 --> 00:16:15.720 But if you watch the shape of the entire flock, 340 00:16:15.840 --> 00:16:18.080 you can see clear patterns. You can predict where the 341 00:16:18.080 --> 00:16:21.360 group is going. The AI learns the shape of the intention. 342 00:16:21.960 --> 00:16:25.519 It recognizes that this specific swirl of activity, this shape 343 00:16:25.519 --> 00:16:29.080 in high dimensional space means prepared to grasp, and that 344 00:16:29.240 --> 00:16:31.080 swirl means accelery forward. 345 00:16:31.320 --> 00:16:34.159 So it's pattern recognition on a massive, massive scale. 346 00:16:34.240 --> 00:16:36.720 That's all it is. But it's incredibly powerful. 347 00:16:36.759 --> 00:16:39.759 But there's a catch. The sources all mentioned something they 348 00:16:39.759 --> 00:16:42.399 call the drift, which sounds like a horror movie title, 349 00:16:42.440 --> 00:16:44.240 by the way, or maybe a racing movie. 350 00:16:44.360 --> 00:16:46.279 It is a bit of a nightmare for engineers, so 351 00:16:46.360 --> 00:16:49.279 maybe horror movie is right. The problem is that the 352 00:16:49.320 --> 00:16:53.519 brain is not a static chip. It's plastic. It changes, 353 00:16:53.679 --> 00:16:57.919 it learns, it learns, Neurons die or they change their tuning. 354 00:16:57.960 --> 00:17:00.399 The way they fire in response to a certain intention, 355 00:17:01.240 --> 00:17:04.559 or on a purely mechanical level, the electro to array 356 00:17:04.680 --> 00:17:08.240 might shift by a few micrometers because the brain wobbles 357 00:17:08.240 --> 00:17:09.799 when you sneeze or turn your head quickly. 358 00:17:09.799 --> 00:17:11.960 Oh, I never even thought of that. So the algorithm 359 00:17:11.960 --> 00:17:14.400 I spent all morning training at nine zero zero. 360 00:17:14.160 --> 00:17:17.279 Am might be completely useless by two point zero pm. 361 00:17:17.440 --> 00:17:20.839 The map has changed the relationship between the neural signals 362 00:17:20.880 --> 00:17:23.200 and the intention has drifted. 363 00:17:23.359 --> 00:17:26.079 That sounds incredibly frustrating for the user. I mean, imagine 364 00:17:26.119 --> 00:17:29.039 your computer mounts working perfectly in the morning, but by lunchtime, 365 00:17:29.319 --> 00:17:32.240 moving your handwright makes the cursor go up. You'd go crazy. 366 00:17:32.359 --> 00:17:36.240 It happens, and traditionally the solution was clumsy. It was recalibration. 367 00:17:36.519 --> 00:17:37.160 What does that mean. 368 00:17:37.759 --> 00:17:39.759 It means you have to stop everything you're doing and 369 00:17:39.799 --> 00:17:42.119 you have to do a twenty minute training session. Okay, 370 00:17:42.160 --> 00:17:45.359 now think about moving left four minute. Now think right 371 00:17:45.400 --> 00:17:48.680 for a minute. Now think up. Every single day, sometimes 372 00:17:48.799 --> 00:17:49.680 multiple times a day. 373 00:17:49.880 --> 00:17:52.039 That's a job. That's not a tool. If I have 374 00:17:52.079 --> 00:17:54.039 to train my phone for twenty minutes before I can 375 00:17:54.079 --> 00:17:55.400 send a text message, I'm throwing the. 376 00:17:55.359 --> 00:17:58.640 Phone away exactly. It's a huge barrier to adoption. So 377 00:17:58.640 --> 00:18:02.160 the cutting edge right now is in adaptive algorithms. 378 00:18:01.720 --> 00:18:02.759 So they adapt on the fly. 379 00:18:03.200 --> 00:18:06.240 The AI continues to learn while you use it. It 380 00:18:06.279 --> 00:18:09.599 watches for errors. If it sees you correcting a mistake, 381 00:18:09.839 --> 00:18:12.240 like you clearly intended to go left, but the cursor 382 00:18:12.279 --> 00:18:14.640 went up and you immediately pulled it back down, the 383 00:18:14.680 --> 00:18:17.599 AI logs that and says, oops, my map was wrong 384 00:18:17.720 --> 00:18:20.240 for that movement. Let me update that in real time. 385 00:18:20.319 --> 00:18:22.440 So it learns from the user's frustration. 386 00:18:22.240 --> 00:18:25.440 In a sense. Yes, it's trying to dynamically realign the 387 00:18:25.480 --> 00:18:29.000 machines map with the brains map moment by moment so 388 00:18:29.160 --> 00:18:31.079 the user doesn't have to stop and recalibrate. 389 00:18:31.279 --> 00:18:33.039 Okay, there's one more piece to this puzzle that we 390 00:18:33.119 --> 00:18:35.079 touched on earlier, the missing. 391 00:18:34.759 --> 00:18:36.720 Feedback, the closed sleep problem. 392 00:18:36.839 --> 00:18:40.319 My biological hand feels the glass, it knows when it's 393 00:18:40.359 --> 00:18:45.240 made contact. A BCI controlled robot arm does not. So 394 00:18:45.279 --> 00:18:48.599 the user is relying entirely on vision right, just watching the. 395 00:18:48.720 --> 00:18:52.480 R move mostly Yes, and that's very unnatural and slow. 396 00:18:52.640 --> 00:18:55.279 You're constantly overcorrecting because you don't have that sense of 397 00:18:55.680 --> 00:18:56.720 touch or position. 398 00:18:57.279 --> 00:18:59.920 So what's the frontier there? How? Do you send signals 399 00:19:00.079 --> 00:19:00.799 back to the brain. 400 00:19:01.519 --> 00:19:05.119 This is truly bleeding edge research. The idea is to 401 00:19:05.160 --> 00:19:08.079