1 00:00:02,200 --> 00:00:04,900 In today's world, AI is everywhere. 2 00:00:04,900 --> 00:00:06,160 - What appointments do I have today? 3 00:00:06,160 --> 00:00:08,440 - Today is the Royal Institution Christmas Lectures. 4 00:00:08,440 --> 00:00:11,480 - Play me a "wake me up" morning mix. 5 00:00:11,480 --> 00:00:14,360 - # Another day's passing in your life... # 6 00:00:14,360 --> 00:00:17,280 - It's popping up in places we could never have imagined 7 00:00:17,280 --> 00:00:18,680 a few years ago. 8 00:00:18,680 --> 00:00:21,000 - # I'm on my way... # 9 00:00:21,000 --> 00:00:22,640 - Mike, you all set for tonight? 10 00:00:22,640 --> 00:00:24,120 I love your background! 11 00:00:24,120 --> 00:00:26,000 - Yeah, I'm just getting in the Christmas mood. 12 00:00:26,000 --> 00:00:28,200 - Computer, will I need a coat tonight? 13 00:00:28,200 --> 00:00:29,600 - Yes, take a coat. 14 00:00:29,600 --> 00:00:32,800 - How aware are we of the AI in our everyday lives? 15 00:00:32,800 --> 00:00:34,080 - # We're on a ride... # 16 00:00:34,080 --> 00:00:35,400 - Where can we find it? 17 00:00:35,400 --> 00:00:37,000 - # Ain't got no money... # 18 00:00:37,000 --> 00:00:38,320 - What does it do? 19 00:00:38,320 --> 00:00:40,080 - # You can't ignore it... # 20 00:00:40,080 --> 00:00:42,680 - That's what we'll be asking in tonight's lectures. 21 00:01:04,000 --> 00:01:07,040 APPLAUSE 22 00:01:23,920 --> 00:01:25,800 Hello, I'm Mike Wooldridge, 23 00:01:25,800 --> 00:01:28,440 and welcome back to the second of this year's 24 00:01:28,440 --> 00:01:30,640 Christmas lectures from the Royal Institution, 25 00:01:30,640 --> 00:01:33,000 an event supported by CGI. 26 00:01:33,000 --> 00:01:35,720 In the last lecture, we saw how today's 27 00:01:35,720 --> 00:01:39,640 artificial intelligence was inspired by the human brain. 28 00:01:39,640 --> 00:01:42,160 In this lecture, we're going to investigate 29 00:01:42,160 --> 00:01:44,800 how AI is shaping our lives today, 30 00:01:44,800 --> 00:01:47,400 and we're going to see how games have played 31 00:01:47,400 --> 00:01:51,120 an incredibly important role in advancing AI. 32 00:01:51,120 --> 00:01:53,440 We're going to come back to the video that we just saw 33 00:01:53,440 --> 00:01:55,000 a little bit later on in the lecture, 34 00:01:55,000 --> 00:01:58,840 and we're going to discuss exactly where AI features in that video. 35 00:01:58,840 --> 00:02:02,200 But first, we're going to start by exploring how AI 36 00:02:02,200 --> 00:02:05,320 features in one of our most popular forms of entertainment - 37 00:02:05,320 --> 00:02:06,920 video games. 38 00:02:06,920 --> 00:02:10,160 Please welcome top GT driver Martin Grady, 39 00:02:10,160 --> 00:02:13,040 and from Sony AI, Kaushik Subramanian. 40 00:02:13,040 --> 00:02:16,160 APPLAUSE 41 00:02:16,160 --> 00:02:17,440 Welcome, Martin. - Thank you. 42 00:02:21,000 --> 00:02:22,360 - Take a seat... 43 00:02:24,920 --> 00:02:26,400 ..and make yourself comfortable. 44 00:02:26,400 --> 00:02:29,120 Now, Martin, how long have you been racing? 45 00:02:29,120 --> 00:02:31,120 - I've been racing for 25 years. 46 00:02:31,120 --> 00:02:33,680 And I've been a pro-racing driver for 12 of those. 47 00:02:33,680 --> 00:02:36,000 - OK, so what we're going to get you to do today 48 00:02:36,000 --> 00:02:38,720 is we're going to get you to play against AI. 49 00:02:38,720 --> 00:02:41,000 We're going to get you to play against a racing AI 50 00:02:41,000 --> 00:02:42,520 that was developed by Sony AI 51 00:02:42,520 --> 00:02:44,840 for the game of Gran Turismo. 52 00:02:44,840 --> 00:02:47,560 By the way, who in the audience has got one of these things at home? 53 00:02:47,560 --> 00:02:48,760 Anybody? 54 00:02:48,760 --> 00:02:50,440 A couple of people have got one of those. 55 00:02:50,440 --> 00:02:52,280 So, Martin, you get yourself going. 56 00:02:52,280 --> 00:02:56,000 Kaushik, tell us, what did you want to do when you built this AI? 57 00:02:56,000 --> 00:02:58,280 - Right, the first challenge what we asked ourselves is, 58 00:02:58,280 --> 00:03:02,480 can we develop an AI that can race to beat the world's best 59 00:03:02,480 --> 00:03:05,520 Gran Turismo racers in a competitive race? 60 00:03:05,520 --> 00:03:09,680 And can the AI learn the required skills to perform at that level? 61 00:03:09,680 --> 00:03:11,880 It turns out we can. 62 00:03:11,880 --> 00:03:15,600 - Kaushik, where exactly is Sophy appearing in this, the AI? 63 00:03:15,600 --> 00:03:18,480 - So, Martin's raced with 19 other AI cars, 64 00:03:18,480 --> 00:03:20,800 and they're all controlled by Gran Turismo Sophy here. 65 00:03:20,800 --> 00:03:22,040 And the way you can see that is, 66 00:03:22,040 --> 00:03:23,440 if you look on the left-hand side, 67 00:03:23,440 --> 00:03:25,720 you can see all the names of the player cars. 68 00:03:25,720 --> 00:03:27,760 And there's a logo which looks like a heart, 69 00:03:27,760 --> 00:03:30,560 and all of those logos are the Gran Turismo Sophy cars. 70 00:03:30,560 --> 00:03:33,000 And all of those cars, they know what car they're driving, 71 00:03:33,000 --> 00:03:34,520 they're aware of each other 72 00:03:34,520 --> 00:03:36,600 and they're of course aware of Martin as well. 73 00:03:36,600 --> 00:03:39,040 - OK, but they're not all trying to gang up on Martin, are they? 74 00:03:39,040 --> 00:03:40,160 They're all independent? 75 00:03:40,160 --> 00:03:42,480 - They're all trying to get to first place independently. 76 00:03:42,480 --> 00:03:46,240 - So how did Sophy, the AI, get good enough in this race? 77 00:03:46,240 --> 00:03:47,920 - This AI, Gran Turismo Sophy, 78 00:03:47,920 --> 00:03:50,480 is trained using an approach called reinforcement learning, 79 00:03:50,480 --> 00:03:53,440 an approach where the AI learns by trial and error. 80 00:03:53,440 --> 00:03:56,000 So, at the start, it really doesn't know much about driving, 81 00:03:56,000 --> 00:03:58,000 much less the rules of racing. 82 00:03:58,000 --> 00:04:00,440 - OK, I think we've got the video showing here of the training. 83 00:04:00,440 --> 00:04:01,760 - Right, and when you watch this, 84 00:04:01,760 --> 00:04:03,560 you're going to see the cars colliding, 85 00:04:03,560 --> 00:04:05,880 you're going to see the cars completely losing control, 86 00:04:05,880 --> 00:04:07,000 going off the track. 87 00:04:07,000 --> 00:04:09,040 And this is just within a few hours of training. 88 00:04:09,040 --> 00:04:10,680 This is what it looks like. 89 00:04:10,680 --> 00:04:12,840 And after a while, with reinforcement learning, 90 00:04:12,840 --> 00:04:14,480 it tries many different things 91 00:04:14,480 --> 00:04:16,520 and then it learns, finally, the right sequence 92 00:04:16,520 --> 00:04:18,160 of actions to be able to race, 93 00:04:18,160 --> 00:04:20,400 like the way Martin's looking at right now. 94 00:04:20,400 --> 00:04:24,000 - LAUGHTER - Or crashing, apparently. 95 00:04:24,000 --> 00:04:26,480 OK, now, we're going to come back later on in the lecture 96 00:04:26,480 --> 00:04:29,440 and we're going to learn a lot more about reinforcement learning. 97 00:04:29,440 --> 00:04:33,320 But, Kaushik, why is a video game like this so difficult 98 00:04:33,320 --> 00:04:34,560 for reinforcement learning, 99 00:04:34,560 --> 00:04:36,320 and just for machine learning generally? 100 00:04:36,320 --> 00:04:38,200 - Racing has a lot of variability. 101 00:04:38,200 --> 00:04:40,000 There are everchanging situations. 102 00:04:40,000 --> 00:04:42,000 When you're driving at the edge of control 103 00:04:42,000 --> 00:04:44,320 in those situations, that can be hard. 104 00:04:44,320 --> 00:04:45,560 Now, in Gran Turismo, 105 00:04:45,560 --> 00:04:49,000 players have access to nearly 500 different cars, 106 00:04:49,000 --> 00:04:51,240 and when you take those cars and you put them in a race, 107 00:04:51,240 --> 00:04:53,440 those cars can interact in many different ways 108 00:04:53,440 --> 00:04:55,000 over multiple laps. 109 00:04:55,000 --> 00:04:56,480 And in all of those situations, 110 00:04:56,480 --> 00:04:59,440 the AI would need to be skilful and reliable, 111 00:04:59,440 --> 00:05:01,520 and that can be challenging. 112 00:05:01,520 --> 00:05:06,160 - OK. And how long did it take to train the AI that we're seeing? 113 00:05:06,160 --> 00:05:09,640 - A single experiment takes about 25 PlayStations 114 00:05:09,640 --> 00:05:12,360 and trains for nearly two weeks before it can race 115 00:05:12,360 --> 00:05:14,000 with other human players. 116 00:05:14,000 --> 00:05:16,480 - OK, so, Martin, tell us, how are you finding this? 117 00:05:16,480 --> 00:05:18,760 Is it different to what you've used previously? 118 00:05:18,760 --> 00:05:20,680 - Yeah, it's completely different. 119 00:05:20,680 --> 00:05:23,720 In the past, you would just race a standard car, 120 00:05:23,720 --> 00:05:25,800 just going on the same line over and over. 121 00:05:25,800 --> 00:05:27,600 Here, you can see they're going defensive, 122 00:05:27,600 --> 00:05:29,320 they're reacting to what I do. 123 00:05:29,320 --> 00:05:32,200 So, if I go for a move, they're going to move to defend that. 124 00:05:32,200 --> 00:05:34,240 - They can try and block you as you try to overtake? 125 00:05:34,240 --> 00:05:36,760 - Yeah, they react to my reactions, 126 00:05:36,760 --> 00:05:38,000 so then I have to react again. 127 00:05:38,000 --> 00:05:40,520 And it just makes it so much more fun to be able to play against that 128 00:05:40,520 --> 00:05:42,400 because there's endless possibilities then. 129 00:05:42,400 --> 00:05:44,520 As you can see, they're going to the right-hand side. 130 00:05:44,520 --> 00:05:46,360 It's just really good fun now, cos now I've got 131 00:05:46,360 --> 00:05:48,000 nowhere to go, essentially. - OK! 132 00:05:48,000 --> 00:05:52,240 Now, Martin, I can't help but notice you appear to be playing this game 133 00:05:52,240 --> 00:05:54,560 wearing a pair of socks. 134 00:05:54,560 --> 00:05:56,800 Is that a pro tip for playing Gran Turismo? 135 00:05:56,800 --> 00:05:58,000 - It is a pro tip. 136 00:05:58,000 --> 00:06:00,000 So take your shoes off, play in your socks. 137 00:06:00,000 --> 00:06:01,680 You get a more sensitive feel on the pedal 138 00:06:01,680 --> 00:06:04,000 and it allows you to just race that a little bit quicker. 139 00:06:04,000 --> 00:06:06,400 - OK, listen to the experts, kids. 140 00:06:06,400 --> 00:06:08,280 OK. Thank you, Martin. 141 00:06:08,280 --> 00:06:10,640 Thank you, Kaushik, for coming and showing us this game. 142 00:06:10,640 --> 00:06:12,080 APPLAUSE Thank you. 143 00:06:16,400 --> 00:06:19,360 How did the AI get so good? 144 00:06:19,360 --> 00:06:23,000 Well, we heard Kaushik mention reinforcement learning. 145 00:06:23,000 --> 00:06:27,320 That's what Kaushik and his team used to train GT Sophy. 146 00:06:27,320 --> 00:06:31,160 So what is reinforcement learning and how does it work? 147 00:06:31,160 --> 00:06:33,760 Now, to explain, I need some help, 148 00:06:33,760 --> 00:06:37,000 but my next guest is a little bit nervous. 149 00:06:37,000 --> 00:06:40,280 So, when she comes on, please, let's just have silent applause. 150 00:06:40,280 --> 00:06:42,440 Please keep it quiet. Don't alarm her. 151 00:06:42,440 --> 00:06:46,400 So, please, let's welcome Caitlin and her dog, Freya. 152 00:06:49,800 --> 00:06:52,000 FREYA BARKS 153 00:06:52,000 --> 00:06:53,640 - Hello! 154 00:06:53,640 --> 00:06:55,240 - Hello, Caitlin. 155 00:06:55,240 --> 00:06:57,000 Hello, Freya. - Hi. 156 00:06:57,000 --> 00:06:58,600 - Welcome to the Christmas lectures. 157 00:06:58,600 --> 00:07:01,760 - Thank you. - So, Caitlin, how long have you been working with Freya? 158 00:07:01,760 --> 00:07:04,560 - So, Freya and I have been working together for about three years. 159 00:07:04,560 --> 00:07:07,000 Her mum kindly let us work together when I was training 160 00:07:07,000 --> 00:07:08,640 to be a puppy school tutor, and now... 161 00:07:08,640 --> 00:07:10,000 - FREYA BARKS 162 00:07:10,000 --> 00:07:12,000 - ..just started some trick training together. 163 00:07:12,000 --> 00:07:15,040 So she very much likes to use her voice, 164 00:07:15,040 --> 00:07:16,760 so I apologise for the ears. 165 00:07:16,760 --> 00:07:18,280 - OK. 166 00:07:18,280 --> 00:07:20,200 So what are you going to show us today, Caitlin? 167 00:07:20,200 --> 00:07:22,360 How are you going to show us how you train Freya? - Yes! 168 00:07:22,360 --> 00:07:28,000 So, today, Freya is going to show us how to pickpocket someone 169 00:07:28,000 --> 00:07:29,520 with a handkerchief. 170 00:07:29,520 --> 00:07:31,720 What we're going to do is use positive reinforcement 171 00:07:31,720 --> 00:07:35,800 to tell her that when she pulls on the handkerchief, 172 00:07:35,800 --> 00:07:37,280 then she gets a reward for that. 173 00:07:37,280 --> 00:07:39,880 So I use something called a marker. 174 00:07:39,880 --> 00:07:42,600 So, when I tell her to get it, I'll tell her yes 175 00:07:42,600 --> 00:07:45,000 and reward her for what she's done. 176 00:07:45,000 --> 00:07:46,560 That's the correct thing to do. 177 00:07:46,560 --> 00:07:49,000 So I'm going to tell her to get it. 178 00:07:49,000 --> 00:07:50,800 Yes! Good girl! - OK. 179 00:07:50,800 --> 00:07:52,000 - Very nice. 180 00:07:52,000 --> 00:07:55,200 So first I want to get her interested in this. 181 00:07:55,200 --> 00:07:56,800 Yes! - OK, I think she's interested. 182 00:07:56,800 --> 00:07:58,160 - Nice! 183 00:07:58,160 --> 00:08:01,720 So the way I do that is just make it interesting like this. 184 00:08:01,720 --> 00:08:03,240 If she pulls on it... Yes! 185 00:08:03,240 --> 00:08:04,960 And then give her that reward. 186 00:08:04,960 --> 00:08:07,160 And then I want to add a cue to it. 187 00:08:07,160 --> 00:08:09,520 So a little... Freya, get it! 188 00:08:09,520 --> 00:08:11,320 Yes! Good girl. 189 00:08:11,320 --> 00:08:12,640 And then reward. 190 00:08:12,640 --> 00:08:15,000 So it's just following that pattern. - OK. 191 00:08:15,000 --> 00:08:17,960 Now, rewards seem to be really important to Freya. 192 00:08:17,960 --> 00:08:19,520 Is this a big part of how she learns? 193 00:08:19,520 --> 00:08:22,880 - Absolutely. So what gets rewarded gets repeated. 194 00:08:22,880 --> 00:08:26,000 So the more I reward her for the thing I want her to do, 195 00:08:26,000 --> 00:08:27,520 the more it's going to happen. 196 00:08:27,520 --> 00:08:28,880 But what I want to make sure of 197 00:08:28,880 --> 00:08:30,760 is that she's also responding to the cue. 198 00:08:30,760 --> 00:08:33,080 So, when she pulled it there, I didn't give her the reward 199 00:08:33,080 --> 00:08:37,080 because I want her to respond to that to get it. - Because you hadn't given her the cue. - Exactly, yeah. 200 00:08:37,080 --> 00:08:40,240 And now the last part of this is the pickpocket part, 201 00:08:40,240 --> 00:08:42,720 which I will need your assistance. - OK. 202 00:08:42,720 --> 00:08:45,000 - If that's OK? - Yes. 203 00:08:45,000 --> 00:08:47,000 - So I'm going to give you that. 204 00:08:47,000 --> 00:08:49,320 - CAITLIN LAUGHS 205 00:08:49,320 --> 00:08:51,400 - And if you just pop that in your back pocket for me. 206 00:08:51,400 --> 00:08:52,560 Freya, come, come! 207 00:08:52,560 --> 00:08:54,400 Good girl! Sit! - OK. 208 00:08:54,400 --> 00:08:58,000 - Nice. Stay. - The things I do for the Christmas lectures. 209 00:08:58,000 --> 00:08:59,880 - Freya, get it! 210 00:08:59,880 --> 00:09:01,440 - FREYA BARKS 211 00:09:01,440 --> 00:09:03,360 - Yes! Good job! 212 00:09:03,360 --> 00:09:04,880 Very nice! 213 00:09:04,880 --> 00:09:07,760 - CAITLIN LAUGHS FREYA BARKS 214 00:09:07,760 --> 00:09:09,000 MIKE LAUGHS 215 00:09:09,000 --> 00:09:11,440 Freya, you're a lot smarter than my dog, 216 00:09:11,440 --> 00:09:12,720 I have to say. 217 00:09:12,720 --> 00:09:13,800 Well done. 218 00:09:13,800 --> 00:09:16,640 OK, thank you so much, Caitlin. Thank you, Freya. 219 00:09:16,640 --> 00:09:19,520 Remember, everybody, polite quiet applause, please. 220 00:09:19,520 --> 00:09:20,720 Thank you so much. - Thank you! 221 00:09:20,720 --> 00:09:23,000 - Come on, Freya! Let's go! - FREYA BARKS 222 00:09:23,000 --> 00:09:24,640 - She doesn't want to leave! 223 00:09:26,480 --> 00:09:29,640 So how is Freya learning? 224 00:09:29,640 --> 00:09:33,640 She's given a treat every time she does a task well. 225 00:09:33,640 --> 00:09:37,320 And in AI, we call that reinforcement learning. 226 00:09:37,320 --> 00:09:41,520 In the last lecture, we saw how AI learns from data, 227 00:09:41,520 --> 00:09:42,840 from training data. 228 00:09:42,840 --> 00:09:45,080 Now, reinforcement learning is similar, 229 00:09:45,080 --> 00:09:49,320 but the training data in this case comes in the form of rewards. 230 00:09:49,320 --> 00:09:52,000 And reinforcement learning is really good for games 231 00:09:52,000 --> 00:09:53,640 like Gran Turismo 232 00:09:53,640 --> 00:09:58,480 because most games have a score and we can use scores as the reward. 233 00:09:58,480 --> 00:10:02,280 When the AI scores a point, that's a reward. 234 00:10:02,280 --> 00:10:06,000 And what the AI is designed to do is to learn 235 00:10:06,000 --> 00:10:08,400 how to maximise its reward, 236 00:10:08,400 --> 00:10:10,000 to maximise its score, 237 00:10:10,000 --> 00:10:13,760 to get as many rewards as possible as quickly as possible, 238 00:10:13,760 --> 00:10:16,480 just like Freya getting her treats. 239 00:10:16,480 --> 00:10:20,280 And we can use neural networks to learn all that. 240 00:10:20,280 --> 00:10:23,120 That's called deep reinforcement learning. 241 00:10:23,120 --> 00:10:26,480 But sometimes reinforcement learning might not give us the outcomes 242 00:10:26,480 --> 00:10:27,880 that we were hoping for, 243 00:10:27,880 --> 00:10:30,160 and we're going to have a game to illustrate this idea. 244 00:10:30,160 --> 00:10:31,880 Can I have a volunteer, please? 245 00:10:31,880 --> 00:10:33,640 In the green there? Yeah, you come down. 246 00:10:33,640 --> 00:10:35,160 Yes, you've been selected! 247 00:10:35,160 --> 00:10:37,080 APPLAUSE 248 00:10:37,080 --> 00:10:39,360 Come on down. Just stand here. What's your name? - Nivea. 249 00:10:39,360 --> 00:10:41,440 - OK, well, welcome to the Christmas lectures. 250 00:10:41,440 --> 00:10:42,800 Now, I'm going to give you a task. 251 00:10:42,800 --> 00:10:44,480 I'm going to give you some instructions. 252 00:10:44,480 --> 00:10:46,880 And I just want you to follow those instructions, OK? 253 00:10:46,880 --> 00:10:49,160 Are you nimble on your feet? Are you a good runner? - Yes. 254 00:10:49,160 --> 00:10:51,640 - OK, excellent. Then you're going to be perfect for this. 255 00:10:51,640 --> 00:10:55,200 Now, what we've done is we've placed five bells 256 00:10:55,200 --> 00:10:56,840 around the lecture theatre. 257 00:10:56,840 --> 00:10:58,720 One there. One there. 258 00:10:58,720 --> 00:11:01,760 One there. Very nimble on my feet, as you can see. 259 00:11:01,760 --> 00:11:03,160 One there. 260 00:11:03,160 --> 00:11:05,480 And finally, one there. 261 00:11:05,480 --> 00:11:07,560 And what we're going to do is we're going to give you 262 00:11:07,560 --> 00:11:08,760 ten seconds, 263 00:11:08,760 --> 00:11:11,800 and I want to hear as many bell rings as possible. 264 00:11:11,800 --> 00:11:13,600 - Mm-hm. - Got it? - Yes. - Yeah. 265 00:11:13,600 --> 00:11:15,280 OK, all right, what we're going to do is 266 00:11:15,280 --> 00:11:17,000 we're going to count you down from three, 267 00:11:17,000 --> 00:11:18,800 and then we're going to say go, all right? 268 00:11:18,800 --> 00:11:20,080 Got it? 269 00:11:20,080 --> 00:11:23,360 Three, two, one. 270 00:11:23,360 --> 00:11:24,760 Go! AUDIENCE: - Go! 271 00:11:24,760 --> 00:11:26,600 - BELL RINGS 272 00:11:26,600 --> 00:11:29,080 - BELL RINGS Go, go, go! - Go, Nivea! Run! 273 00:11:29,080 --> 00:11:30,560 BELL RINGS Let's hear those bells! 274 00:11:30,560 --> 00:11:33,080 - Come on! Pick up the pace! 275 00:11:33,080 --> 00:11:35,360 - One, stop! 276 00:11:35,360 --> 00:11:36,520 OK, how many bell rings? 277 00:11:36,520 --> 00:11:38,760 Come back to the middle. We heard five, I think. 278 00:11:38,760 --> 00:11:40,160 Was that about right, everybody? 279 00:11:40,160 --> 00:11:42,000 Yeah? We heard five bell rings. 280 00:11:42,000 --> 00:11:44,920 That was really, really good, OK? 281 00:11:44,920 --> 00:11:47,120 Thank you so much for that. You can go back to your seat. 282 00:11:47,120 --> 00:11:49,360 APPLAUSE AND CHEERING 283 00:11:50,520 --> 00:11:54,000 Now we're going to carry out the same challenge, 284 00:11:54,000 --> 00:11:55,720 but with artificial intelligence. 285 00:11:55,720 --> 00:11:57,760 So please welcome the Royal Institution's 286 00:11:57,760 --> 00:12:00,000 special bell-ringing robot! 287 00:12:00,000 --> 00:12:01,480 Come on, bell-ringing robot. 288 00:12:03,240 --> 00:12:05,240 APPLAUSE 289 00:12:05,240 --> 00:12:07,360 DRILL WHIRS REPEATEDLY 290 00:12:10,480 --> 00:12:11,920 Give me that. Honestly. 291 00:12:13,280 --> 00:12:14,600 OK! 292 00:12:14,600 --> 00:12:17,000 Now, I'm going to give you ten seconds. 293 00:12:17,000 --> 00:12:20,760 I want to hear as many bell rings as possible. 294 00:12:20,760 --> 00:12:22,720 So let's count down the robot. 295 00:12:22,720 --> 00:12:27,560 Three, two, one, go! - Go! 296 00:12:27,560 --> 00:12:30,720 - BELL RINGS 297 00:12:34,160 --> 00:12:36,000 Stop! 298 00:12:36,000 --> 00:12:39,160 Did anybody manage to keep track of how many that was? 299 00:12:39,160 --> 00:12:42,200 OK, robot, thank you very much. 300 00:12:42,200 --> 00:12:43,400 There you go. 301 00:12:43,400 --> 00:12:46,640 DRILL WHIRS REPEATEDLY 302 00:12:48,080 --> 00:12:50,480 APPLAUSE 303 00:12:52,600 --> 00:12:55,400 So, our volunteer, 304 00:12:55,400 --> 00:12:58,000 you didn't do anything wrong at all. 305 00:12:58,000 --> 00:12:59,280 Because you're a human being. 306 00:12:59,280 --> 00:13:00,840 I gave you some instructions 307 00:13:00,840 --> 00:13:04,000 and you tried to interpret what I wanted you to do. 308 00:13:04,000 --> 00:13:06,400 And what I wanted in my instructions 309 00:13:06,400 --> 00:13:09,000 is to hear somebody running around the lecture theatre 310 00:13:09,000 --> 00:13:11,480 pressing each of those bells. 311 00:13:11,480 --> 00:13:14,560 But actually, those weren't exactly the instructions I gave you. 312 00:13:14,560 --> 00:13:16,680 What I said is, I just wanted to hear 313 00:13:16,680 --> 00:13:19,000 as many bell rings as possible. 314 00:13:19,000 --> 00:13:22,200 And our robot took the instructions literally. 315 00:13:22,200 --> 00:13:25,400 And the quickest way the robot could get a reward 316 00:13:25,400 --> 00:13:28,400 was just to stand there and bang that bell. 317 00:13:28,400 --> 00:13:32,600 The AI found a way to maximise its rewards 318 00:13:32,600 --> 00:13:36,000 without doing what I wanted it to do. 319 00:13:36,000 --> 00:13:40,480 The way we set up rewards when we use reinforcement learning 320 00:13:40,480 --> 00:13:42,360 is really important 321 00:13:42,360 --> 00:13:45,920 because sometimes we can set up rewards 322 00:13:45,920 --> 00:13:49,760 so that the AI discovers a way to maximise its rewards 323 00:13:49,760 --> 00:13:52,760 without doing what we wanted it to do. 324 00:13:52,760 --> 00:13:54,800 And that's what the robot did there. 325 00:13:54,800 --> 00:13:57,520 And now what we're going to see is a real example. 326 00:13:57,520 --> 00:14:00,000 So let's have a look at this video on the screen. 327 00:14:00,000 --> 00:14:03,920 In 2014, the AI company DeepMind 328 00:14:03,920 --> 00:14:06,440 trained an AI program to play this game. 329 00:14:06,440 --> 00:14:09,200 It's a 1970s video game called Breakout, 330 00:14:09,200 --> 00:14:11,200 and it uses reinforcement learning. 331 00:14:11,200 --> 00:14:13,320 The more it plays, the better it gets. 332 00:14:13,320 --> 00:14:15,720 Now, at the beginning, most of the time, as you'll see, 333 00:14:15,720 --> 00:14:17,000 it's just missing. 334 00:14:17,000 --> 00:14:19,880 And if it manages to hit the ball, it's just really pure chance. 335 00:14:19,880 --> 00:14:22,000 But every time it knocks a brick out, 336 00:14:22,000 --> 00:14:23,080 it gets a point. 337 00:14:23,080 --> 00:14:26,160 Now, after a bit more training, watch, it never misses. 338 00:14:26,160 --> 00:14:30,680 It's reliably hitting that ball back every single time. 339 00:14:30,680 --> 00:14:34,320 And you might think that's actually about as good as it's going to get. 340 00:14:34,320 --> 00:14:36,920 It's hitting the ball every single time. 341 00:14:36,920 --> 00:14:39,400 But look what happened when they trained it 342 00:14:39,400 --> 00:14:40,760 just a little bit longer. 343 00:14:40,760 --> 00:14:42,720 Look at the left there. Look what goes on. 344 00:14:42,720 --> 00:14:45,800 It discovered completely on its own 345 00:14:45,800 --> 00:14:49,400 that the way to maximise its score is to drill a hole 346 00:14:49,400 --> 00:14:53,000 down the side of the wall and bounce the ball above. 347 00:14:53,000 --> 00:14:55,280 So games are fun for us to play, 348 00:14:55,280 --> 00:14:58,600 but they're also a great proving ground for AI 349 00:14:58,600 --> 00:15:01,240 because they can provide big challenges for AI 350 00:15:01,240 --> 00:15:05,000 without the possibility of hurting anybody or damaging anything. 351 00:15:05,000 --> 00:15:07,760 And, earlier, we spoke to the CEO of DeepMind, 352 00:15:07,760 --> 00:15:11,000 the company behind that video, Demis Hassabis. 353 00:15:11,000 --> 00:15:13,080 You can see at the beginning, the AI, by the way, 354 00:15:13,080 --> 00:15:14,320 is trying to focus on his image 355 00:15:14,320 --> 00:15:15,920 and not doing a terribly good job of it. 356 00:15:15,920 --> 00:15:18,600 I asked him about the importance of games 357 00:15:18,600 --> 00:15:20,200 for artificial intelligence. 358 00:15:20,200 --> 00:15:23,000 - Games and AI have always had a long history together. 359 00:15:23,000 --> 00:15:26,320 Actually, if you go all the way back to Turing and Shannon, 360 00:15:26,320 --> 00:15:28,240 who sort of invented the field of AI, 361 00:15:28,240 --> 00:15:31,000 they all started off with things like chess programs, 362 00:15:31,000 --> 00:15:33,280 and trying to figure out how could a machine, 363 00:15:33,280 --> 00:15:35,040 for example, play chess well? 364 00:15:35,040 --> 00:15:38,000 And we use games at DeepMind as a testing ground 365 00:15:38,000 --> 00:15:40,760 for our AI ideas and algorithmic ideas 366 00:15:40,760 --> 00:15:43,480 because you need a clear metric to measure them against. 367 00:15:43,480 --> 00:15:45,840 And, of course, games usually have scores 368 00:15:45,840 --> 00:15:48,480 that you can optimise or a win-loss condition, 369 00:15:48,480 --> 00:15:51,240 You know, so you can track very clearly if you're making progress. 370 00:15:51,240 --> 00:15:55,000 - So let's take a closer look at game-playing programs. 371 00:15:55,000 --> 00:15:58,000 OK. And one of the easiest games to play 372 00:15:58,000 --> 00:15:59,920 is noughts and crosses. 373 00:15:59,920 --> 00:16:01,240 Who plays noughts and crosses? 374 00:16:01,240 --> 00:16:02,920 Can you all play noughts and crosses? 375 00:16:02,920 --> 00:16:04,560 OK. In the United States, 376 00:16:04,560 --> 00:16:07,640 they don't call it noughts and crosses, they call it tic-tac-toe. 377 00:16:07,640 --> 00:16:09,480 And because the US is so big in AI, 378 00:16:09,480 --> 00:16:12,280 we have to call it tic-tac-toe in AI as well. 379 00:16:12,280 --> 00:16:13,880 I need a volunteer. 380 00:16:13,880 --> 00:16:17,120 Who's really good at playing tic-tac-toe? 381 00:16:17,120 --> 00:16:19,080 OK, you at the back. The second from the end. 382 00:16:19,080 --> 00:16:20,320 Come on down! 383 00:16:24,800 --> 00:16:26,280 Hello. Come here. 384 00:16:27,680 --> 00:16:29,400 What's your name? - Emmy. 385 00:16:29,400 --> 00:16:31,360 - Emmy? - Yes. - OK, thanks for joining us, Emmy. 386 00:16:31,360 --> 00:16:33,800 What I'd like you to do is come and stand over... 387 00:16:33,800 --> 00:16:35,320 Just stand about here. 388 00:16:35,320 --> 00:16:40,160 OK. Now, Emmy, we are going to get you to play tic-tac-toe 389 00:16:40,160 --> 00:16:43,200 live against the Royal Institution supercomputer. 390 00:16:43,200 --> 00:16:44,600 Just stand back a tiny little bit. 391 00:16:44,600 --> 00:16:48,000 And this tic-tac-toe supercomputer is called BrodeRick. 392 00:16:48,000 --> 00:16:49,480 What we're going to do is 393 00:16:49,480 --> 00:16:52,000 we're going to get you to play against 394 00:16:52,000 --> 00:16:53,680 a completely untrained BrodeRick. 395 00:16:53,680 --> 00:16:55,600 BrodeRick is going to be the crosses player. 396 00:16:55,600 --> 00:16:57,200 You are going to be the noughts player. 397 00:16:57,200 --> 00:16:58,880 And you're just going to move your piece 398 00:16:58,880 --> 00:17:00,720 just by pressing one of those buttons. 399 00:17:00,720 --> 00:17:01,960 So BrodeRick's moved first. 400 00:17:01,960 --> 00:17:03,360 You see where BrodeRick's moved? 401 00:17:03,360 --> 00:17:05,640 And now you can play one of your pieces. 402 00:17:07,000 --> 00:17:08,240 OK, moving to the middle. 403 00:17:08,240 --> 00:17:09,400 Smart move. 404 00:17:09,400 --> 00:17:10,720 BrodeRick moves. 405 00:17:10,720 --> 00:17:13,000 OK, you need to respond. 406 00:17:13,000 --> 00:17:15,000 OK, BrodeRick's going to respond again. 407 00:17:15,000 --> 00:17:16,640 You get to go again. 408 00:17:16,640 --> 00:17:17,960 And you've won! 409 00:17:17,960 --> 00:17:19,440 APPLAUSE 410 00:17:19,440 --> 00:17:20,680 Well done! 411 00:17:20,680 --> 00:17:22,320 Turn and face the audience! 412 00:17:23,600 --> 00:17:24,760 Take the applause. 413 00:17:24,760 --> 00:17:27,000 You never know when it's going to come in life. 414 00:17:27,000 --> 00:17:28,160 OK, so, well done. 415 00:17:28,160 --> 00:17:29,720 You beat BrodeRick. 416 00:17:29,720 --> 00:17:32,400 But BrodeRick hasn't got a clue how to play this game. 417 00:17:32,400 --> 00:17:36,160 So what we're now going to do is we're going to train BrodeRick 418 00:17:36,160 --> 00:17:37,880 to play tic-tac-toe, 419 00:17:37,880 --> 00:17:39,800 and we're going to train it using the technique 420 00:17:39,800 --> 00:17:42,280 that we've been talking about, the same technique that we saw 421 00:17:42,280 --> 00:17:44,800 with Caitlynn and Freya - reinforcement learning. 422 00:17:44,800 --> 00:17:47,000 So, this time, what we're going to do is tell BrodeRick 423 00:17:47,000 --> 00:17:48,760 to train itself, and we're going to do that 424 00:17:48,760 --> 00:17:50,720 by selecting number of players - zero. 425 00:17:50,720 --> 00:17:52,840 So do you want to select number of players - zero? 426 00:17:52,840 --> 00:17:54,720 And this is number of human players - zero. 427 00:17:54,720 --> 00:17:57,600 BrodeRick is just going to play against itself. 428 00:17:57,600 --> 00:17:59,840 Now, stay on, Emmy, because you're going to be playing 429 00:17:59,840 --> 00:18:01,360 the trained model in a moment. 430 00:18:01,360 --> 00:18:03,880 And what we're seeing now is BrodeRick 431 00:18:03,880 --> 00:18:06,760 starting to play itself. 432 00:18:06,760 --> 00:18:10,360 And it's cycling through more and more games. 433 00:18:10,360 --> 00:18:14,840 And every time it wins a game, it gets a reward. 434 00:18:14,840 --> 00:18:18,120 Every time it loses, it gets a punishment. 435 00:18:18,120 --> 00:18:19,800 When it gets a reward, 436 00:18:19,800 --> 00:18:23,640 that makes it more likely to play the moves it played again. 437 00:18:23,640 --> 00:18:25,840 And we're seeing lots of games coming up on the screen, 438 00:18:25,840 --> 00:18:29,320 but we're only seeing a tiny fraction of the total number 439 00:18:29,320 --> 00:18:31,600 of games that BrodeRick is playing. 440 00:18:32,720 --> 00:18:36,240 BrodeRick, training complete! 441 00:18:36,240 --> 00:18:40,320 It's played 20,000 games of tic-tac-toe 442 00:18:40,320 --> 00:18:42,160 while we've been watching it. 443 00:18:42,160 --> 00:18:43,560 Well, BrodeRick's trained. 444 00:18:43,560 --> 00:18:45,880 But, BrodeRick, how confident are you feeling? 445 00:18:45,880 --> 00:18:46,920 - Ah! BANG! 446 00:18:48,160 --> 00:18:52,520 - BrodeRick, it seems, is feeling very confident indeed. 447 00:18:52,520 --> 00:18:54,240 Emmy, how confident are you feeling? 448 00:18:54,240 --> 00:18:56,600 - I'm quite scared now. - You're quite scared now? 449 00:18:56,600 --> 00:18:58,840 There's nothing to be scared of, Emmy. 450 00:18:58,840 --> 00:19:00,360 Exactly the same procedure. 451 00:19:00,360 --> 00:19:02,240 Emmy, over to you. 452 00:19:02,240 --> 00:19:03,600 So we've trained BrodeRick. 453 00:19:03,600 --> 00:19:05,280 BrodeRick should be better. 454 00:19:05,280 --> 00:19:07,000 Let's see if that's the case. 455 00:19:08,200 --> 00:19:09,600 OK. 456 00:19:09,600 --> 00:19:11,000 Well spotted. 457 00:19:11,000 --> 00:19:12,240 OK. 458 00:19:18,040 --> 00:19:19,320 So it's a draw. 459 00:19:19,320 --> 00:19:21,360 How do you feel about that, Emmy? - Not that bad. 460 00:19:21,360 --> 00:19:22,640 - It's not bad at all, 461 00:19:22,640 --> 00:19:26,520 because BrodeRick is basically a perfect tic-tac-toe player. 462 00:19:26,520 --> 00:19:29,560 And if you have two perfect tic-tac-toe players 463 00:19:29,560 --> 00:19:32,360 like BrodeRick and Emmy playing against each other, 464 00:19:32,360 --> 00:19:34,840 that's what they're going to get - they're going to get a draw. 465 00:19:34,840 --> 00:19:39,000 Now, BrodeRick can't guarantee to win every time it plays, 466 00:19:39,000 --> 00:19:41,000 but it's never going to lose any more. 467 00:19:41,000 --> 00:19:43,800 It's completely different to the untrained model 468 00:19:43,800 --> 00:19:46,000 that we saw earlier on. 469 00:19:46,000 --> 00:19:50,560 OK. Now, in fact, under the hood, I have to tell you, 470 00:19:50,560 --> 00:19:53,000 it is not a supercomputer at all. 471 00:19:53,000 --> 00:19:55,760 Dan from the demo team is going to show us the truth 472 00:19:55,760 --> 00:19:57,000 about BrodeRick. 473 00:19:57,000 --> 00:19:59,280 And the truth about BrodeRick is that the computer 474 00:19:59,280 --> 00:20:01,000 is, in fact, this. 475 00:20:01,000 --> 00:20:02,640 It's a very basic computer. 476 00:20:02,640 --> 00:20:05,000 It's a £30 computer. 477 00:20:05,000 --> 00:20:09,000 That's all you need to learn how to play tic-tac-toe. 478 00:20:09,000 --> 00:20:11,720 So, Emmy, thank you for coming down and playing against BrodeRick. 479 00:20:11,720 --> 00:20:14,240 APPLAUSE And thank you, BrodeRick. 480 00:20:20,160 --> 00:20:24,200 Tic-tac-toe is a simple game. 481 00:20:24,200 --> 00:20:27,000 And one of the most important reasons it's simple 482 00:20:27,000 --> 00:20:28,240 is the following. 483 00:20:28,240 --> 00:20:31,000 The average number of moves that you can make 484 00:20:31,000 --> 00:20:35,240 at any point in tic-tac-toe is around about four. 485 00:20:35,240 --> 00:20:38,680 You start with nine possible moves for the crosses player, 486 00:20:38,680 --> 00:20:42,000 then the next player has eight possible moves and so on. 487 00:20:42,000 --> 00:20:45,400 On average, you've got about four possible moves 488 00:20:45,400 --> 00:20:46,680 available to you. 489 00:20:46,680 --> 00:20:50,760 And we call that the branching factor of the game. 490 00:20:50,760 --> 00:20:53,400 But even a branching factor of four 491 00:20:53,400 --> 00:20:56,600 means that there are close to 20,000 different ways 492 00:20:56,600 --> 00:21:00,000 that we can fill in a tic-tac-toe grid. 493 00:21:00,000 --> 00:21:03,880 Let's compare that to the game of chess. 494 00:21:03,880 --> 00:21:07,400 This is a much more interesting and much more difficult game. 495 00:21:07,400 --> 00:21:10,000 And one of the reasons that it's much more interesting 496 00:21:10,000 --> 00:21:13,440 and difficult is that the branching factor is larger. 497 00:21:13,440 --> 00:21:15,520 It's around about 35. 498 00:21:15,520 --> 00:21:17,240 And, remember, what that means 499 00:21:17,240 --> 00:21:19,000 is that from any position on the board, 500 00:21:19,000 --> 00:21:20,560 on average - 501 00:21:20,560 --> 00:21:22,480 not for every possible move, but on average - 502 00:21:22,480 --> 00:21:26,000 you've got about 35 possible moves available. 503 00:21:26,000 --> 00:21:28,000 Now let's have a look at another game. 504 00:21:28,000 --> 00:21:30,200 This is the game of Go. 505 00:21:30,200 --> 00:21:32,400 It originated in ancient China, 506 00:21:32,400 --> 00:21:35,760 but it's still hugely popular today in Asia. 507 00:21:35,760 --> 00:21:37,280 The aim in the game of Go 508 00:21:37,280 --> 00:21:40,000 is you're playing your black stones or white stones. 509 00:21:40,000 --> 00:21:43,360 And what you want to try and do is to cover as much 510 00:21:43,360 --> 00:21:46,000 of the board as possible with your stones 511 00:21:46,000 --> 00:21:49,000 and to surround your opponent's stones on the board. 512 00:21:49,000 --> 00:21:51,760 And you just take it in turns to place your stones. 513 00:21:51,760 --> 00:21:53,440 So it's a very simple game. 514 00:21:53,440 --> 00:21:56,000 It only really has three rules. 515 00:21:56,000 --> 00:22:00,000 But the branching factor of Go makes it phenomenally hard 516 00:22:00,000 --> 00:22:01,720 for human beings to play, 517 00:22:01,720 --> 00:22:05,000 and it makes it very, very hard for artificial intelligence. 518 00:22:05,000 --> 00:22:09,000 And that's what made it an important target for AI. 519 00:22:09,000 --> 00:22:14,000 Go has a branching factor of around about 250. 520 00:22:14,000 --> 00:22:16,000 Remember, that branching factor, 521 00:22:16,000 --> 00:22:18,280 that's the average number of moves 522 00:22:18,280 --> 00:22:21,000 that a player can make at any given time. 523 00:22:21,000 --> 00:22:22,480 Let's illustrate those numbers. 524 00:22:22,480 --> 00:22:26,800 If you take a look under your seat, you should find a card. 525 00:22:26,800 --> 00:22:29,240 OK? So take a look at the card. 526 00:22:29,240 --> 00:22:32,200 You don't need to do anything yet. Just take a look at the card. 527 00:22:32,200 --> 00:22:36,000 Now, if your card has a tic-tac-toe grid on it, 528 00:22:36,000 --> 00:22:38,000 stand up now. 529 00:22:38,000 --> 00:22:39,560 OK, so here we are. 530 00:22:39,560 --> 00:22:41,480 We've got four people in the audience 531 00:22:41,480 --> 00:22:43,680 holding a tic-tac-toe grid. 532 00:22:43,680 --> 00:22:47,400 That's the branching factor of tic-tac-toe. 533 00:22:47,400 --> 00:22:50,400 If your card has a chess board on it, 534 00:22:50,400 --> 00:22:52,320 which looks like this, 535 00:22:52,320 --> 00:22:54,640 then stand up now. 536 00:22:54,640 --> 00:22:58,880 That's 35 of our audience members that are standing up now. 537 00:22:58,880 --> 00:23:02,400 That's the branching factor of the game of chess. 538 00:23:02,400 --> 00:23:06,720 And, finally, if you have a Go board underneath your seat, 539 00:23:06,720 --> 00:23:08,240 something that looks like this, 540 00:23:08,240 --> 00:23:10,000 stand up now. 541 00:23:10,000 --> 00:23:14,440 That's pretty much the entire audience. 542 00:23:14,440 --> 00:23:18,240 That's the branching factor of the game of Go, 543 00:23:18,240 --> 00:23:21,320 and compare that to the branching factor of the game of chess 544 00:23:21,320 --> 00:23:25,280 and those four people that were stood up there at the beginning. 545 00:23:25,280 --> 00:23:28,240 OK, everybody, you can sit down now. Thank you. 546 00:23:28,240 --> 00:23:32,440 Let's think about the number of game states there are in the game of Go. 547 00:23:32,440 --> 00:23:33,840 And it's this number - 548 00:23:33,840 --> 00:23:37,360 ten to the power of 170. 549 00:23:37,360 --> 00:23:40,600 We can hardly fit that number on the screen. 550 00:23:40,600 --> 00:23:43,520 We never encounter numbers that large 551 00:23:43,520 --> 00:23:44,760 in our everyday life. 552 00:23:44,760 --> 00:23:47,200 They are literally astronomical. 553 00:23:48,240 --> 00:23:54,520 And what we've learned today is a really important lesson about AI. 554 00:23:54,520 --> 00:23:59,720 Trying to solve a problem by simply looking at all the alternatives 555 00:23:59,720 --> 00:24:01,800 is called brute force. 556 00:24:01,800 --> 00:24:04,320 And the AI lesson is this. 557 00:24:04,320 --> 00:24:08,000 For problems like playing Go, brute force doesn't work. 558 00:24:08,000 --> 00:24:09,200 It will never work. 559 00:24:09,200 --> 00:24:12,680 Because the numbers are just too astronomically large. 560 00:24:12,680 --> 00:24:16,240 Even if we turned the entire universe into a computer, 561 00:24:16,240 --> 00:24:20,000 it wouldn't be able to play just using brute force. 562 00:24:20,000 --> 00:24:24,240 So we need some way of reducing the number of alternatives to check. 563 00:24:24,240 --> 00:24:27,000 And these are often called heuristics. 564 00:24:27,000 --> 00:24:31,280 And we can use machine learning to help us build heuristics. 565 00:24:31,280 --> 00:24:32,800 And I asked Demis Hassabis 566 00:24:32,800 --> 00:24:35,840 how DeepMind trained AI to play Go. 567 00:24:35,840 --> 00:24:37,360 - In the case of AlphaGo, 568 00:24:37,360 --> 00:24:40,240 AlphaGo learnt for itself how to play Go, 569 00:24:40,240 --> 00:24:42,160 and the strategies it would use and so on, 570 00:24:42,160 --> 00:24:44,560 by playing against itself many millions of times. 571 00:24:44,560 --> 00:24:46,000 And it used learning systems, 572 00:24:46,000 --> 00:24:48,760 so reinforcement learning and tree search techniques, 573 00:24:48,760 --> 00:24:51,200 in order to figure out for itself what the right strategies 574 00:24:51,200 --> 00:24:52,760 were to be good at Go. 575 00:24:52,760 --> 00:24:57,400 - In 2016, AlphaGo played against and beat Lee Sedol, 576 00:24:57,400 --> 00:24:59,000 one of the world's greatest players. 577 00:24:59,000 --> 00:25:00,800 And a lot of people didn't expect that. 578 00:25:00,800 --> 00:25:03,440 - And so, of course, the cool thing about that is that, actually, 579 00:25:03,440 --> 00:25:06,440 AlphaGo, not only did it beat the world champion in 2016, 580 00:25:06,440 --> 00:25:10,120 it also came up with completely new ideas about Go - 581 00:25:10,120 --> 00:25:13,160 new strategies that no human players had ever thought of, 582 00:25:13,160 --> 00:25:16,000 even though Go is several thousands of years old. 583 00:25:39,000 --> 00:25:42,440 - AlphaGo's achievement was a really big step forward 584 00:25:42,440 --> 00:25:43,840 in the world of AI. 585 00:25:43,840 --> 00:25:45,800 Because of its complexity, 586 00:25:45,800 --> 00:25:49,240 we thought that Go was a barrier that AI wouldn't overcome 587 00:25:49,240 --> 00:25:51,840 for another decade, or maybe even more. 588 00:25:51,840 --> 00:25:56,520 And games like Go have been a really great training ground for AI, 589 00:25:56,520 --> 00:26:00,480 and the knowledge that we've derived from building AI to play games 590 00:26:00,480 --> 00:26:02,880 has led to major breakthroughs in science, 591 00:26:02,880 --> 00:26:04,280 as we're going to see later on. 592 00:26:04,280 --> 00:26:08,000 And, increasingly, AI is embedded in our lives, 593 00:26:08,000 --> 00:26:10,280 everywhere that we look. 594 00:26:10,280 --> 00:26:12,320 Let's go back to that video that we showed you 595 00:26:12,320 --> 00:26:13,680 at the beginning of the lecture. 596 00:26:13,680 --> 00:26:18,000 And this time, when you watch it, I want you to try to keep count 597 00:26:18,000 --> 00:26:21,320 of how many times you think the AI appears. 598 00:26:26,240 --> 00:26:27,680 - What appointments do I have today? 599 00:26:27,680 --> 00:26:30,000 - Today is the Royal Institution Christmas lecture. 600 00:26:30,000 --> 00:26:33,000 - Play me a "wake me up" morning mix. 601 00:26:33,000 --> 00:26:35,160 - # Another day's passing in your life 602 00:26:35,160 --> 00:26:37,000 # I heard you saying 603 00:26:37,000 --> 00:26:38,600 # Now is our time 604 00:26:38,600 --> 00:26:42,480 # I'm on the highway I'm on my way... # 605 00:26:42,480 --> 00:26:44,080 - Mike, you all set for tonight? 606 00:26:44,080 --> 00:26:45,520 I love your background! 607 00:26:45,520 --> 00:26:47,400 - Yeah, I'm just getting in the Christmas mood. 608 00:26:47,400 --> 00:26:49,720 - Computer, will I need a coat tonight? 609 00:26:49,720 --> 00:26:51,000 - Yes, take a coat. 610 00:26:51,000 --> 00:26:53,360 - # This time I ain't Ain't gonna stop 611 00:26:53,360 --> 00:26:56,880 # We're on a ride to reach the top 612 00:26:56,880 --> 00:27:00,040 # Ain't got no money But I'm full of fun 613 00:27:00,040 --> 00:27:03,000 # You can't ignore it No, you won't run away... # 614 00:27:03,000 --> 00:27:05,760 - I'd like to introduce you to the person who made that video, 615 00:27:05,760 --> 00:27:08,760 science author and broadcaster Dr Emily Grossman. 616 00:27:08,760 --> 00:27:11,680 APPLAUSE 617 00:27:11,680 --> 00:27:13,440 - Hiya. - Hello. 618 00:27:13,440 --> 00:27:14,760 - Hi. 619 00:27:16,000 --> 00:27:17,640 - And welcome, Emily! 620 00:27:17,640 --> 00:27:19,000 - Thank you. 621 00:27:19,000 --> 00:27:21,600 - Was what we watched just a normal day for you? 622 00:27:21,600 --> 00:27:22,760 - Yeah, pretty much. 623 00:27:22,760 --> 00:27:24,320 I mean, it was a pretty typical day, 624 00:27:24,320 --> 00:27:27,120 although I don't usually end up at the Christmas lectures. 625 00:27:27,120 --> 00:27:30,760 - OK. Now, Emily, before you came on, I asked the audience to watch 626 00:27:30,760 --> 00:27:34,600 that video and to keep track of how many times they thought 627 00:27:34,600 --> 00:27:36,280 that AI appeared in it. 628 00:27:36,280 --> 00:27:37,560 Shall we ask the audience? 629 00:27:37,560 --> 00:27:39,080 - Yeah, let's do that. - OK, over to you. 630 00:27:39,080 --> 00:27:41,240 - OK, what do you think? How many times? 631 00:27:41,240 --> 00:27:43,600 - 14. - 14 times?! 632 00:27:43,600 --> 00:27:45,200 Any other...? Any other suggestions? 633 00:27:45,200 --> 00:27:46,600 Over here, yeah! 634 00:27:46,600 --> 00:27:48,280 - Er, ten. - Ten times. 635 00:27:48,280 --> 00:27:50,680 - Emily, how many times did you think 636 00:27:50,680 --> 00:27:52,000 that AI appeared in that video? 637 00:27:52,000 --> 00:27:54,680 - Well, actually, somewhere between the two. 638 00:27:54,680 --> 00:27:57,720 It was actually 12 times that AI was in the video. 639 00:27:57,720 --> 00:27:59,120 So very close, both of you. 640 00:27:59,120 --> 00:28:01,680 But you've got to remember that those were only the times 641 00:28:01,680 --> 00:28:04,520 that I was interacting with AI, like, directly. 642 00:28:04,520 --> 00:28:06,360 So, like, when I was walking down the street, 643 00:28:06,360 --> 00:28:08,960 there might've been loads of stuff around that was using AI 644 00:28:08,960 --> 00:28:11,280 that I was sort of indirectly interacting with. 645 00:28:11,280 --> 00:28:13,240 So it could've been actually loads more than 12. 646 00:28:13,240 --> 00:28:15,120 - I think there probably was loads more than 12. 647 00:28:15,120 --> 00:28:16,640 We just don't realise it. - Exactly. 648 00:28:16,640 --> 00:28:18,440 - OK, were you surprised how many times 649 00:28:18,440 --> 00:28:19,960 AI appears in your everyday life? 650 00:28:19,960 --> 00:28:21,320 - I was, like, totally surprised. 651 00:28:21,320 --> 00:28:24,000 So, before I made this video, I had no idea 652 00:28:24,000 --> 00:28:26,160 how many times I might be interacting with AI 653 00:28:26,160 --> 00:28:27,360 on a daily basis. 654 00:28:27,360 --> 00:28:29,560 So I've got, like, a sleep app on my phone, 655 00:28:29,560 --> 00:28:31,600 and I know that I use that pretty much every day, 656 00:28:31,600 --> 00:28:34,000 but I had no idea that that might have AI in it. 657 00:28:34,000 --> 00:28:39,000 And also things like an app for maps and even traffic lights. 658 00:28:39,000 --> 00:28:42,000 I just didn't realise all of that could be using AI as well. 659 00:28:42,000 --> 00:28:44,000 - OK, Emily, thanks so much for joining us. 660 00:28:44,000 --> 00:28:46,000 We really appreciate it. - Thank you. Thank you. 661 00:28:46,000 --> 00:28:48,320 - CHEERING AND APPLAUSE 662 00:28:51,120 --> 00:28:54,840 So, from that video, it's clear that AI is embedded 663 00:28:54,840 --> 00:28:57,680 in our lives already everywhere. 664 00:28:57,680 --> 00:29:01,680 Something like talking to a voice-activated personal assistant 665 00:29:01,680 --> 00:29:04,520 or using an automated translation app. 666 00:29:04,520 --> 00:29:07,320 Now, those things can very quickly get mundane. 667 00:29:07,320 --> 00:29:09,000 They seem very ordinary. 668 00:29:09,000 --> 00:29:10,240 But I have to tell you, 669 00:29:10,240 --> 00:29:13,000 it's incredibly frustrating for AI researchers 670 00:29:13,000 --> 00:29:16,000 when they deliver something like automated translation 671 00:29:16,000 --> 00:29:18,440 and everybody starts taking it for granted. 672 00:29:18,440 --> 00:29:21,800 But there are some areas where AI 673 00:29:21,800 --> 00:29:23,640 should never be taken for granted, 674 00:29:23,640 --> 00:29:26,840 where AI is saving lives. 675 00:29:26,840 --> 00:29:30,360 Let's have a look at one area that AI has the potential 676 00:29:30,360 --> 00:29:33,680 to have a beneficial impact on a global scale, 677 00:29:33,680 --> 00:29:38,000 where literally it will save lives and make lives better. 678 00:29:38,000 --> 00:29:39,600 AI in health care. 679 00:29:41,280 --> 00:29:45,000 Let's start by talking about Parkinson's disease. 680 00:29:45,000 --> 00:29:46,760 And to help explain this disease, 681 00:29:46,760 --> 00:29:49,360 I'd like you to welcome Dr Rutger Zietsma 682 00:29:49,360 --> 00:29:51,560 and former BBC technology correspondent 683 00:29:51,560 --> 00:29:53,040 Rory Cellan-Jones. 684 00:29:53,040 --> 00:29:55,520 APPLAUSE 685 00:30:00,560 --> 00:30:02,800 Welcome, Rutger. Welcome, Rory. 686 00:30:02,800 --> 00:30:05,320 So, Rory, first, can you tell us a little bit about 687 00:30:05,320 --> 00:30:08,400 what Parkinson's is and how it manifests itself? 688 00:30:08,400 --> 00:30:12,000 - Well, Parkinson's disease is a degenerative brain disease. 689 00:30:12,000 --> 00:30:15,600 It's basically what happens when the dopamine cells 690 00:30:15,600 --> 00:30:17,200 in your brain begin to die. 691 00:30:17,200 --> 00:30:19,000 And that affects an awful lot of things. 692 00:30:19,000 --> 00:30:22,800 And the symptoms include a tremor, insomnia, 693 00:30:22,800 --> 00:30:25,000 stiffness in movement. 694 00:30:25,000 --> 00:30:27,640 Tiny writing is another symptom. 695 00:30:27,640 --> 00:30:29,160 There are a whole bunch of symptoms 696 00:30:29,160 --> 00:30:30,560 and it's very various. 697 00:30:30,560 --> 00:30:33,000 Lots of people have one symptom, but not another. 698 00:30:33,000 --> 00:30:35,600 - OK, now, you have a really amazing story 699 00:30:35,600 --> 00:30:38,800 about how you came to be diagnosed. Can you tell us about that? 700 00:30:38,800 --> 00:30:40,520 - Well, I first suspected something 701 00:30:40,520 --> 00:30:42,440 when I was walking with my wife on holiday 702 00:30:42,440 --> 00:30:44,000 and I kept dragging my right foot. 703 00:30:44,000 --> 00:30:45,920 But then, a month later, I was on the television 704 00:30:45,920 --> 00:30:48,000 doing a live broadcast from the island... 705 00:30:48,000 --> 00:30:49,280 - We've got the broadcast here. 706 00:30:49,280 --> 00:30:50,680 - Yeah, from the island of Jersey. 707 00:30:50,680 --> 00:30:53,320 And as you can see, I didn't notice this at the time at all, 708 00:30:53,320 --> 00:30:55,280 but my hand was shaking quite violently. 709 00:30:55,280 --> 00:30:59,000 Somebody saw it, a neuroscience expert, 710 00:30:59,000 --> 00:31:00,640 and wrote in to the BBC and said, 711 00:31:00,640 --> 00:31:02,080 "You ought to get that sorted out 712 00:31:02,080 --> 00:31:04,480 "because that could be a symptom of Parkinson's disease." 713 00:31:04,480 --> 00:31:07,000 - OK, it's really incredible that you actually... 714 00:31:07,000 --> 00:31:08,840 Actually, what drew your attention to it 715 00:31:08,840 --> 00:31:11,760 is somebody spotting a broadcast that you were doing. 716 00:31:11,760 --> 00:31:13,520 OK, Rory, you take a seat. 717 00:31:13,520 --> 00:31:16,920 Rutger, tell us, how is AI being used to help 718 00:31:16,920 --> 00:31:18,560 diagnose Parkinson's disease? 719 00:31:18,560 --> 00:31:21,400 - Yes, I would like to introduce to you the NeuroMotor Pen, 720 00:31:21,400 --> 00:31:23,600 ideally with a brief demonstration. 721 00:31:23,600 --> 00:31:25,080 Rory, would you mind giving it a go? 722 00:31:25,080 --> 00:31:26,520 - I will give it a go. 723 00:31:26,520 --> 00:31:29,920 - All you need to do is draw an EL figure 724 00:31:29,920 --> 00:31:31,760 and do it about five times. 725 00:31:31,760 --> 00:31:34,000 The pen measures movements, 726 00:31:34,000 --> 00:31:36,560 but there's a link between the movement in the brain 727 00:31:36,560 --> 00:31:37,800 and the remote function. 728 00:31:37,800 --> 00:31:40,400 So we're effectively looking at what's going on with the brain 729 00:31:40,400 --> 00:31:42,480 and if a part of the brain controls fine movement, 730 00:31:42,480 --> 00:31:45,440 it might perhaps be affected by, for example, 731 00:31:45,440 --> 00:31:46,640 Parkinson's disease. 732 00:31:46,640 --> 00:31:49,640 So it basically calculates digital biomarkers, 733 00:31:49,640 --> 00:31:53,280 the symptom that the patient may be suffering from - 734 00:31:53,280 --> 00:31:55,800 things like tremor, a rhythmical shake 735 00:31:55,800 --> 00:31:57,440 or slowness of movement. 736 00:31:57,440 --> 00:32:00,000 - So why is this better or more preferable 737 00:32:00,000 --> 00:32:01,200 to a human diagnosing? 738 00:32:01,200 --> 00:32:03,000 - So the human will still be involved. 739 00:32:03,000 --> 00:32:05,960 The clinician is still in charge of making the diagnosis, 740 00:32:05,960 --> 00:32:08,840 but this device provides you with accurate, 741 00:32:08,840 --> 00:32:12,280 objective information, and that's really useful, especially 742 00:32:12,280 --> 00:32:15,160 if it's very subtle and it's not very obvious, when it's not easy 743 00:32:15,160 --> 00:32:16,880 to spot or interpret. 744 00:32:16,880 --> 00:32:19,400 So it makes the process a little bit easier and you can make 745 00:32:19,400 --> 00:32:20,720 a faster diagnosis. 746 00:32:20,720 --> 00:32:23,120 - OK. And what does it mean for people like Rory who've actually 747 00:32:23,120 --> 00:32:26,280 been diagnosed with the disease? What does it mean for them? 748 00:32:26,280 --> 00:32:28,680 - Well, Parkinson's disease develops differently 749 00:32:28,680 --> 00:32:30,080 in different individuals. 750 00:32:30,080 --> 00:32:34,160 So a treatment is ideally individualised for the patient. 751 00:32:35,480 --> 00:32:38,320 Currently, it's a big challenge to find out how well the person 752 00:32:38,320 --> 00:32:41,320 is doing today compared to, for example, three months ago. 753 00:32:41,320 --> 00:32:44,520 By taking these accurate readings, you can find out how much 754 00:32:44,520 --> 00:32:46,840 of a drug they need in order to feel well. 755 00:32:46,840 --> 00:32:49,040 - Now, I think we're ready to take a look at Rory's results, 756 00:32:49,040 --> 00:32:51,280 and they should be on this monitor over here. 757 00:32:51,280 --> 00:32:53,760 So perhaps you'd like to both join me, and Rutger, 758 00:32:53,760 --> 00:32:56,600 if you could just talk us through what we're seeing here? 759 00:32:56,600 --> 00:33:00,560 - OK. So the first output of the system was this was a confirmation 760 00:33:00,560 --> 00:33:02,280 of Parkinson's disease. 761 00:33:02,280 --> 00:33:06,040 - I've got Parkinson's disease?! - Correct. - Thank you for that. 762 00:33:06,040 --> 00:33:08,520 - And looking at the symptoms, so these are Rory's symptoms. 763 00:33:08,520 --> 00:33:11,440 This is interesting, because they're all sort of moderate 764 00:33:11,440 --> 00:33:13,600 to high, except tremor is relatively low. 765 00:33:13,600 --> 00:33:16,120 The reason is that he's using medication, 766 00:33:16,120 --> 00:33:19,360 and using that medication, there's less of a tremor. 767 00:33:19,360 --> 00:33:22,080 - OK. So these would be more over, the more severe the symptoms, 768 00:33:22,080 --> 00:33:23,680 is that right? - That's correct. Yeah. 769 00:33:23,680 --> 00:33:26,240 The other thing that's interesting here is the micrographia. 770 00:33:26,240 --> 00:33:29,200 So, as Rory was drawing, you could actually see the size 771 00:33:29,200 --> 00:33:30,440 of the writing deteriorates. 772 00:33:30,440 --> 00:33:32,680 It became smaller as he continued to write. 773 00:33:32,680 --> 00:33:34,680 OK, let's now have a look at somebody 774 00:33:34,680 --> 00:33:37,480 with severe Parkinson's disease who's not on any drug. 775 00:33:37,480 --> 00:33:39,920 These numbers are obviously a lot higher, simply because 776 00:33:39,920 --> 00:33:42,600 there's no suppression of, for example, the tremor. 777 00:33:42,600 --> 00:33:45,800 So what we're seeing here is really complex assessment 778 00:33:45,800 --> 00:33:47,320 made very simple. - OK. 779 00:33:47,320 --> 00:33:50,240 9.9 looks pretty dramatic there, doesn't it? 780 00:33:50,240 --> 00:33:52,760 Rory, what do you think of this? What do you make of this? 781 00:33:52,760 --> 00:33:55,800 - Well, this is really important because the trouble with Parkinson's 782 00:33:55,800 --> 00:33:58,360 is actually measuring it, because mostly it's the doctor 783 00:33:58,360 --> 00:34:00,720 telling you to walk up and down outside his office, 784 00:34:00,720 --> 00:34:03,360 get up out of a chair and do various exercises. 785 00:34:03,360 --> 00:34:06,640 This is beginning to develop an objective, 786 00:34:06,640 --> 00:34:08,400 scientific way of measuring it. 787 00:34:08,400 --> 00:34:11,400 And if you can measure it, then that's really important 788 00:34:11,400 --> 00:34:13,640 because you know whether the patient's getting worse, 789 00:34:13,640 --> 00:34:16,000 you know, for instance, if medication is working. 790 00:34:16,000 --> 00:34:17,760 And if you're searching for a wonder drug, 791 00:34:17,760 --> 00:34:21,200 you need to be able to measure what people are like before 792 00:34:21,200 --> 00:34:24,080 and after much more accurately. - OK. 793 00:34:24,080 --> 00:34:26,400 Thank you, Rutger, for joining us. And thank you, Rory, 794 00:34:26,400 --> 00:34:29,080 for sharing your story with us. Thank you. - Thank you. 795 00:34:33,840 --> 00:34:40,080 - The pen uses AI to help spot Parkinson's by comparing 796 00:34:40,080 --> 00:34:42,000 against large data sets. 797 00:34:42,000 --> 00:34:46,480 But AI very soon might be able to provide far more precise 798 00:34:46,480 --> 00:34:50,440 and sophisticated support for people that are pregnant. 799 00:34:50,440 --> 00:34:51,760 To explain how, 800 00:34:51,760 --> 00:34:56,080 please welcome Professor of Computer Science, Ana Namburete. 801 00:35:00,840 --> 00:35:02,440 Welcome, Ana. 802 00:35:02,440 --> 00:35:04,160 OK. - Thank you. 803 00:35:04,160 --> 00:35:08,480 - So, Ana, you are working to improve the monitoring 804 00:35:08,480 --> 00:35:10,800 of babies during pregnancy. - Yes. 805 00:35:10,800 --> 00:35:14,920 So I'm developing algorithms to provide better health care 806 00:35:14,920 --> 00:35:16,960 in pregnancy. And, in particular, 807 00:35:16,960 --> 00:35:19,880 I'm interested in how we can monitor the growth 808 00:35:19,880 --> 00:35:22,800 of a baby's brain when it's inside the womb. 809 00:35:22,800 --> 00:35:27,120 Now, the best and most accurate way of doing this is by collecting 810 00:35:27,120 --> 00:35:30,920 a 3D scan, and at present the technology that we use for this 811 00:35:30,920 --> 00:35:33,320 is an MRI machine. But an MRI machine, 812 00:35:33,320 --> 00:35:36,080 it's very large, it's very expensive. - Very expensive. 813 00:35:36,080 --> 00:35:39,680 - And it's unsuitable for many people and it's unavailable 814 00:35:39,680 --> 00:35:41,800 in most settings. - Whereas this? 815 00:35:41,800 --> 00:35:44,800 - Well, the advantage of using AI in health care 816 00:35:44,800 --> 00:35:47,720 is that it can revolutionise the way in which we use 817 00:35:47,720 --> 00:35:49,880 existing medical technology. 818 00:35:49,880 --> 00:35:53,360 So in pregnancy monitoring, we have been using ultrasound 819 00:35:53,360 --> 00:35:56,200 machines for over half a century. 820 00:35:56,200 --> 00:36:00,400 And while we can't have a pregnant woman or scan a pregnant woman 821 00:36:00,400 --> 00:36:03,560 in a lecture theatre, we can scan a phantom, 822 00:36:03,560 --> 00:36:06,920 which is the 3D model of a mother's abdomen. 823 00:36:06,920 --> 00:36:09,600 So can I have a volunteer, please, to help me? 824 00:36:09,600 --> 00:36:13,280 - OK, we need a volunteer for this. Hard to see with all the lights. 825 00:36:13,280 --> 00:36:15,760 - Is there a keen person here? Yes? - Come on. 826 00:36:21,880 --> 00:36:23,280 What's your name? 827 00:36:23,280 --> 00:36:26,200 - Bea. - Bea. OK, Bea, Ana's going to show you what to do. 828 00:36:26,200 --> 00:36:29,680 - OK, so what we're going to do is try to find the baby 829 00:36:29,680 --> 00:36:31,520 inside the mother's belly. 830 00:36:31,520 --> 00:36:34,000 So what you have here is an ultrasound probe. 831 00:36:34,000 --> 00:36:36,200 You're going to place this in between your thumb 832 00:36:36,200 --> 00:36:37,880 and your forefinger. How does that feel? 833 00:36:37,880 --> 00:36:39,400 Comfortable? - OK. - OK. 834 00:36:39,400 --> 00:36:41,920 So you're going to place it on the mother's belly, 835 00:36:41,920 --> 00:36:45,880 and what you'll see is the image of the foetus emerging. 836 00:36:45,880 --> 00:36:47,320 Do you see that? - Yeah. 837 00:36:47,320 --> 00:36:50,440 - Yeah. What we're going to try to do is to find the profile view. 838 00:36:50,440 --> 00:36:53,560 And what that is, is the view where you can see the forehead, 839 00:36:53,560 --> 00:36:55,920 the nose, the lips and the neck. Yeah? 840 00:36:55,920 --> 00:36:59,120 So we're going to just move the probe. 841 00:36:59,120 --> 00:37:02,360 You can sort of see it there, right? 842 00:37:02,360 --> 00:37:07,160 So what you see is this bright kind of oval, that's the skull. 843 00:37:07,160 --> 00:37:09,560 And inside it, those are the brain structures. 844 00:37:09,560 --> 00:37:12,640 And it's really important to study the brain because it's the most 845 00:37:12,640 --> 00:37:14,800 complex organ in our bodies, 846 00:37:14,800 --> 00:37:18,000 and it grows most rapidly when we're in the womb. 847 00:37:18,000 --> 00:37:22,080 OK. So I think we can take that off. - OK. Thanks so much, Bea. 848 00:37:24,720 --> 00:37:26,280 OK. 849 00:37:26,280 --> 00:37:28,440 So, Ana, that that was an ordinary ultrasound. 850 00:37:28,440 --> 00:37:29,880 But you're using AI, right? 851 00:37:29,880 --> 00:37:33,840 You're going to show us how you use AI to help you with this. 852 00:37:33,840 --> 00:37:38,240 - This is a typical ultrasound image of the baby's brain inside the womb. 853 00:37:38,240 --> 00:37:41,480 And what you see, that's the bright oval again, right? 854 00:37:41,480 --> 00:37:43,040 So that's the skull. 855 00:37:43,040 --> 00:37:45,360 And what you notice is that we're looking at a video, 856 00:37:45,360 --> 00:37:49,440 but it doesn't allow us to see or measure the brain in 3D. 857 00:37:49,440 --> 00:37:55,200 Right? So what AI allows us to do is to turn this fairly grainy video 858 00:37:55,200 --> 00:37:58,680 into a 3D reconstruction of the brain. 859 00:37:58,680 --> 00:38:01,920 And the way it does that is by taking the video, 860 00:38:01,920 --> 00:38:04,920 which is essentially image slices through the brain. 861 00:38:04,920 --> 00:38:07,760 And that's what you see here on the right-hand side 862 00:38:07,760 --> 00:38:09,120 as this red sheet. 863 00:38:09,120 --> 00:38:13,080 And what the AI has done is that it's predicted the position 864 00:38:13,080 --> 00:38:16,160 and orientation of the slice of the brain that we see 865 00:38:16,160 --> 00:38:18,960 in the video, but in 3D space. 866 00:38:18,960 --> 00:38:23,520 And ultimately it cleverly combines them to reconstruct or produce 867 00:38:23,520 --> 00:38:25,440 a single 3D image. 868 00:38:25,440 --> 00:38:30,600 And from this 3D image, it allows us to measure the growth of individual 869 00:38:30,600 --> 00:38:34,080 brain regions and identify babies that may be in need 870 00:38:34,080 --> 00:38:36,240 of further support. - Wow. 871 00:38:36,240 --> 00:38:38,280 Ana, that's absolutely incredible work. 872 00:38:38,280 --> 00:38:42,000 Thank you so much for coming and telling us about it. Thank you. 873 00:38:43,720 --> 00:38:47,320 So AI is already pushing back the boundaries 874 00:38:47,320 --> 00:38:51,400 across medical science, but it's also revolutionising 875 00:38:51,400 --> 00:38:54,160 our understanding of biology. 876 00:38:54,160 --> 00:38:56,640 To help us explore this, please welcome 877 00:38:56,640 --> 00:39:01,560 Professor of Molecular Biophysics from King's College, Rivka Isaacson. 878 00:39:11,120 --> 00:39:16,960 So welcome, Rivka. Now, Professor of Molecular Biophysics. 879 00:39:16,960 --> 00:39:18,920 What do you do exactly? 880 00:39:18,920 --> 00:39:24,320 - Well, in my lab, we use physics technology to work out the shapes 881 00:39:24,320 --> 00:39:28,320 of proteins, which are the molecular machines that do all kinds 882 00:39:28,320 --> 00:39:31,000 of important jobs inside our bodies. 883 00:39:31,000 --> 00:39:33,400 - So why are proteins so complex? 884 00:39:33,400 --> 00:39:37,120 - Well, they're made as a sequence of amino acids, 885 00:39:37,120 --> 00:39:41,000 so it's almost like threading beads with all different shapes 886 00:39:41,000 --> 00:39:43,320 and properties onto a string. 887 00:39:43,320 --> 00:39:46,680 And then that straight line has to fold up into a really 888 00:39:46,680 --> 00:39:50,880 complicated three dimensional structure like this, 889 00:39:50,880 --> 00:39:54,640 which is lysozyme, a protein from our tears 890 00:39:54,640 --> 00:39:57,360 that kills germs that fall into our eyes. 891 00:39:57,360 --> 00:39:59,680 - OK, so why is knowing the structure 892 00:39:59,680 --> 00:40:02,760 like this of a protein so important? 893 00:40:02,760 --> 00:40:06,720 - Well, kind of like taking a machine apart to understand 894 00:40:06,720 --> 00:40:08,240 the way it works. 895 00:40:08,240 --> 00:40:11,120 We can't really do that with proteins because they're so small 896 00:40:11,120 --> 00:40:15,040 and they're on the nano scale and they're in our tiny, tiny cells. 897 00:40:15,040 --> 00:40:18,960 So we don't have an easy way to find out their shape. 898 00:40:18,960 --> 00:40:22,280 But if you know their shape, then you can work out what they do 899 00:40:22,280 --> 00:40:25,280 and how they stick to each other and how they move. 900 00:40:25,280 --> 00:40:27,080 It's like clockwork, kind of. 901 00:40:27,080 --> 00:40:29,560 - OK, so if you know their shape, you can understand how they can 902 00:40:29,560 --> 00:40:31,880 help you or hurt you. Is that right? - Absolutely. 903 00:40:31,880 --> 00:40:33,720 - OK. Now while we've been talking, 904 00:40:33,720 --> 00:40:37,600 it has not escaped our audience's notice that hanging above us 905 00:40:37,600 --> 00:40:41,360 is a rather large, threatening, and strange contraption. 906 00:40:41,360 --> 00:40:42,920 So what is it? 907 00:40:42,920 --> 00:40:46,600 - So apparently this is a protein, 908 00:40:46,600 --> 00:40:50,560 and it's a lot bigger, 50 million times bigger 909 00:40:50,560 --> 00:40:52,600 than the ones in my lab. 910 00:40:52,600 --> 00:40:55,800 - So this is a sequence of amino acids, is that right? 911 00:40:55,800 --> 00:40:58,320 The linear sequence of amino acids. 912 00:40:58,320 --> 00:41:01,160 - I'm sure you can tell that's exactly what it is! 913 00:41:01,160 --> 00:41:03,480 - We all knew immediately that this was a sequence 914 00:41:03,480 --> 00:41:05,360 of amino acids! OK. 915 00:41:05,360 --> 00:41:09,080 Well, OK, we're going to fold this sequence in a moment. 916 00:41:09,080 --> 00:41:11,640 But before I do that, I want you to see if you can guess 917 00:41:11,640 --> 00:41:13,600 what shape it's going to be. 918 00:41:13,600 --> 00:41:15,440 Who wants to have a guess on the front? 919 00:41:15,440 --> 00:41:17,320 What shape do you think it's going to be? 920 00:41:17,320 --> 00:41:20,200 - I think it might be a sequence of rectangles. 921 00:41:20,200 --> 00:41:24,200 - A sequence of rectangles. Anybody else? You on this side? 922 00:41:24,200 --> 00:41:27,800 - I think it's going to be pentagons. - A pentagon? OK. 923 00:41:27,800 --> 00:41:30,080 Both really good guesses. 924 00:41:30,080 --> 00:41:32,280 But now we're going to find out. 925 00:41:32,280 --> 00:41:35,000 OK, so, team, let's fold our protein! 926 00:41:41,680 --> 00:41:43,040 OK! 927 00:41:54,920 --> 00:41:57,520 It spells Royal Institution, and... 928 00:41:57,520 --> 00:42:01,000 - My initials! - ..Rivka Isaacson! 929 00:42:01,000 --> 00:42:02,880 We planned that all along! 930 00:42:04,080 --> 00:42:08,120 OK. So, Rivka, what causes these amino acids to fold 931 00:42:08,120 --> 00:42:09,920 in the way that they do? 932 00:42:09,920 --> 00:42:13,680 - Well, so, for example, if you have a positive charge 933 00:42:13,680 --> 00:42:17,080 over here and a negative charge over here, 934 00:42:17,080 --> 00:42:20,680 then they want to be near each other in the same way that a magnet 935 00:42:20,680 --> 00:42:23,080 would attract another magnet. 936 00:42:23,080 --> 00:42:25,640 And there are lots of other different properties 937 00:42:25,640 --> 00:42:28,760 of the amino acids that want to be in particular places 938 00:42:28,760 --> 00:42:31,320 within the three dimensional structure. 939 00:42:31,320 --> 00:42:35,280 And that's what helps it find its most energetically 940 00:42:35,280 --> 00:42:37,080 favourable position. 941 00:42:37,080 --> 00:42:41,400 - OK. So, finding out the structure of proteins is a really important part 942 00:42:41,400 --> 00:42:43,760 of your job. How did you do it before AI? 943 00:42:43,760 --> 00:42:48,200 - Well, there are lots of ways to do it, but they're all very expensive 944 00:42:48,200 --> 00:42:49,920 and take years and years. 945 00:42:49,920 --> 00:42:53,840 - And, so, getting to grips with protein structure, protein folding, 946 00:42:53,840 --> 00:42:55,360 is a really important problem. 947 00:42:55,360 --> 00:42:57,840 Rivka, thank you so much for coming and telling us about 948 00:42:57,840 --> 00:43:00,800 such a difficult problem. Thank you very much. - Thank you. 949 00:43:07,280 --> 00:43:09,600 - So how can AI help? 950 00:43:09,600 --> 00:43:13,160 Let's welcome a scientist from the team responsible for one 951 00:43:13,160 --> 00:43:17,760 of the most important scientific advances made possible by AI, 952 00:43:17,760 --> 00:43:21,280 DeepMind's Kathryn Tunyasuvunakool! 953 00:43:29,920 --> 00:43:32,280 Kathryn, welcome to the Christmas lectures. 954 00:43:32,280 --> 00:43:34,760 So tell us, what is AlphaFold? 955 00:43:34,760 --> 00:43:38,520 - So AlphaFold is an AI system that we developed for predicting 956 00:43:38,520 --> 00:43:40,040 the structure of proteins. 957 00:43:40,040 --> 00:43:42,840 And it takes as input that amino acid sequence and then... 958 00:43:42,840 --> 00:43:47,520 - Like that big thing that we saw above us. - Exactly like that big sausage of drainpipes. - Yeah. 959 00:43:47,520 --> 00:43:51,200 - And then it tries to output the folded up protein. 960 00:43:51,200 --> 00:43:52,880 - OK. So how did you create it? 961 00:43:52,880 --> 00:43:55,640 - So AlphaFold is an example of a neural network, 962 00:43:55,640 --> 00:43:58,680 and we trained it using all of the experimentally determined 963 00:43:58,680 --> 00:44:00,600 structures that Rivka talked about. 964 00:44:00,600 --> 00:44:03,080 So there are thousands of those that have been painstakingly 965 00:44:03,080 --> 00:44:04,680 determined by scientists. 966 00:44:04,680 --> 00:44:08,440 We run AlphaFold on those training proteins and we compare its 967 00:44:08,440 --> 00:44:10,400 prediction against the true structure, 968 00:44:10,400 --> 00:44:13,320 and then we try and make it more accurate on those examples. 969 00:44:13,320 --> 00:44:16,080 - So I think you can show us how AlphaFold is actually used 970 00:44:16,080 --> 00:44:18,280 by scientists like Rivka? - Absolutely. 971 00:44:18,280 --> 00:44:20,920 So AlphaFold can produce a new prediction in a matter 972 00:44:20,920 --> 00:44:24,040 of minutes, or you can look up a prediction in the AlphaFold 973 00:44:24,040 --> 00:44:26,000 protein structure database. 974 00:44:26,000 --> 00:44:28,640 And once you have a prediction, it's useful for figuring out 975 00:44:28,640 --> 00:44:31,080 what the protein does, how it interacts, 976 00:44:31,080 --> 00:44:32,720 and for planning experiments. 977 00:44:32,720 --> 00:44:35,840 So you can see the structure of one of our predicted proteins here, 978 00:44:35,840 --> 00:44:39,040 and we actually provide confidence measures to help the scientists 979 00:44:39,040 --> 00:44:41,720 understand which parts of this prediction they should rely on. 980 00:44:41,720 --> 00:44:43,720 - So this is a real protein we're looking at now? 981 00:44:43,720 --> 00:44:46,400 - This is a real protein, yes. - That is complex. - It really is. 982 00:44:46,400 --> 00:44:48,960 It's much more complex than that RI example. 983 00:44:48,960 --> 00:44:51,360 And when you look at the blue regions, those are the parts 984 00:44:51,360 --> 00:44:54,280 where the model is confident, and the orange regions are parts 985 00:44:54,280 --> 00:44:55,760 where the model is less confident. 986 00:44:55,760 --> 00:44:57,960 So that helps scientists use this prediction. 987 00:44:57,960 --> 00:45:01,360 - So once scientists know this kind of structure, 988 00:45:01,360 --> 00:45:04,040 that's then going to allow them to focus on the stuff 989 00:45:04,040 --> 00:45:06,600 that they really need to focus on, right? - Yeah, exactly. 990 00:45:06,600 --> 00:45:09,360 Structural biology is all about interpreting these things. 991 00:45:09,360 --> 00:45:12,600 - OK, Kathryn, thank you so much for joining us and telling us 992 00:45:12,600 --> 00:45:15,080 about AlphaFold. - Thank you. 993 00:45:20,960 --> 00:45:24,680 - I asked Demis Hassabis to explain the benefits of AlphaFold 994 00:45:24,680 --> 00:45:26,120 for researchers. 995 00:45:26,120 --> 00:45:29,080 - The rule of thumb is it normally takes one PhD student 996 00:45:29,080 --> 00:45:31,600 their entire PhD, so up to five years, 997 00:45:31,600 --> 00:45:34,560 just to find the structure of one protein 998 00:45:34,560 --> 00:45:38,280 using complicated systems like X-ray crystallography. 999 00:45:38,280 --> 00:45:40,920 And so we managed over the last couple of years 1000 00:45:40,920 --> 00:45:44,960 to fold all 200 million proteins known to science. 1001 00:45:44,960 --> 00:45:47,120 So, out there in nature. 1002 00:45:47,120 --> 00:45:51,320 And so only a tiny fraction of those were known before experimentally. 1003 00:45:51,320 --> 00:45:55,440 And if you times the five years of a PhD by 200 million, 1004 00:45:55,440 --> 00:45:59,360 you get a billion years of PhD time it potentially saved. 1005 00:45:59,360 --> 00:46:02,480 - What's the impact been of AlphaFold on people like Rivka? 1006 00:46:02,480 --> 00:46:06,040 - It's been amazing, the impact, and even over the last two or three 1007 00:46:06,040 --> 00:46:09,280 years, over a million researchers around the world have used 1008 00:46:09,280 --> 00:46:11,480 AlphaFold in its structures. We think that's almost 1009 00:46:11,480 --> 00:46:13,800 every biologist in the world. - Thank you, Dennis. 1010 00:46:13,800 --> 00:46:18,200 So AI is already answering really important problems 1011 00:46:18,200 --> 00:46:20,920 in the world of science and medicine. 1012 00:46:20,920 --> 00:46:24,600 But what about creativity and the arts? 1013 00:46:24,600 --> 00:46:29,240 For a long time, artists assumed that they were going to be immune 1014 00:46:29,240 --> 00:46:31,680 from the impact of AI. 1015 00:46:31,680 --> 00:46:35,160 So are artists under threat from developments in AI? 1016 00:46:35,160 --> 00:46:37,880 Surely it can't do what a human artist does? 1017 00:46:37,880 --> 00:46:41,680 To help us understand, please welcome artist Eric Drass. 1018 00:46:46,520 --> 00:46:47,960 Welcome, Eric. 1019 00:46:50,400 --> 00:46:53,120 So, Eric, tell us what you do? 1020 00:46:53,120 --> 00:46:55,320 - Well, I'm a visual artist and I've been working with 1021 00:46:55,320 --> 00:46:58,040 and collaborating with AI for about five years now. 1022 00:46:58,040 --> 00:47:02,040 - OK. And we're going to use AI to create some pictures live 1023 00:47:02,040 --> 00:47:05,240 in front of this audience, is that right? - We're going to try. 1024 00:47:05,240 --> 00:47:08,120 - And you're going to steer the process through your laptop, right? 1025 00:47:08,120 --> 00:47:12,680 So, what I need, we're going to ask you for some suggestions 1026 00:47:12,680 --> 00:47:14,280 of pictures to create. 1027 00:47:14,280 --> 00:47:16,840 So put your hands up if you'd like to suggest something 1028 00:47:16,840 --> 00:47:20,040 that you want the AI to create. OK. 1029 00:47:20,040 --> 00:47:22,600 Tell us, what picture would you like to see? - Apples. 1030 00:47:22,600 --> 00:47:26,000 - An apple? Just an apple? OK, Eric, do we think we can do that? 1031 00:47:26,000 --> 00:47:28,440 - I think we can probably do that, yeah. - OK. I didn't even manage to 1032 00:47:28,440 --> 00:47:31,720 get back to you and you'd already come up with a picture of an apple! - It's already there! 1033 00:47:31,720 --> 00:47:34,400 We've already got an apple. - OK. That certainly looks like an apple. 1034 00:47:34,400 --> 00:47:36,440 Do you think we can make it a bit more interesting? 1035 00:47:36,440 --> 00:47:40,080 - Well, let's see if we can make it out of Lego, perhaps. 1036 00:47:44,120 --> 00:47:46,920 - I'm not sure I would have guessed that that was an apple! 1037 00:47:46,920 --> 00:47:48,880 But, yeah, that's what you can do with Lego. 1038 00:47:48,880 --> 00:47:51,720 Let's have another, let's have a look at somebody else. 1039 00:47:51,720 --> 00:47:55,200 We haven't been at the back here, so let's just come to the back. 1040 00:47:55,200 --> 00:47:57,240 OK. 1041 00:47:57,240 --> 00:48:00,040 And you? - A dog. - A dog. 1042 00:48:00,040 --> 00:48:02,680 Do you want it doing anything? Begging? 1043 00:48:02,680 --> 00:48:06,200 - Chasing after a ball. - A dog chasing after a ball. 1044 00:48:06,200 --> 00:48:09,280 Great prompt. Eric, off you go. See if you can do the same again. 1045 00:48:09,280 --> 00:48:12,960 See if you can do it before I manage to get back down. 1046 00:48:14,520 --> 00:48:17,800 - Ah-ha! - Chasing a ball, photorealistic! 1047 00:48:17,800 --> 00:48:19,680 OK. - There's a lot of tails there! 1048 00:48:21,600 --> 00:48:24,720 - Yeah, the dog does appear to have two tails! - It does! 1049 00:48:24,720 --> 00:48:26,880 - How many legs has it got, Eric? 1050 00:48:26,880 --> 00:48:28,160 - Many, many legs! 1051 00:48:28,160 --> 00:48:29,880 - So is it just going out on the internet, 1052 00:48:29,880 --> 00:48:32,200 searching for that picture and it's just pulling that in 1053 00:48:32,200 --> 00:48:33,840 and showing it to us? - Well, no. 1054 00:48:33,840 --> 00:48:36,600 Each of these images is created dynamically on the fly 1055 00:48:36,600 --> 00:48:39,720 by the neural network and has never existed before. 1056 00:48:39,720 --> 00:48:41,880 - So that image, these images never existed before? 1057 00:48:41,880 --> 00:48:45,360 - Not at all. This is what's known as text to image. 1058 00:48:45,360 --> 00:48:48,600 So to turn text into an image, you need two parts. 1059 00:48:48,600 --> 00:48:52,560 You need something that can understand text, and you need 1060 00:48:52,560 --> 00:48:54,600 something that can make pictures. 1061 00:48:54,600 --> 00:48:57,760 So if we look at the first part about understanding text, 1062 00:48:57,760 --> 00:49:02,280 part of the neural network is able to tell you how well the picture 1063 00:49:02,280 --> 00:49:05,000 it's trying to create matches the words that you've put in. 1064 00:49:05,000 --> 00:49:08,320 So in this case, this photograph, with the prompt 1065 00:49:08,320 --> 00:49:12,080 "a tabby cat in a basket" gives us like a 93% score. 1066 00:49:12,080 --> 00:49:14,800 It's pretty confident that that's what it looks like. 1067 00:49:14,800 --> 00:49:17,160 If we give it the same picture with a different prompt 1068 00:49:17,160 --> 00:49:20,400 a London bus, for example, it only says 5%. 1069 00:49:20,400 --> 00:49:24,400 So what we've got here is like a reinforcement learning we heard about earlier. 1070 00:49:24,400 --> 00:49:27,320 This is the signal that goes back into the system to tell the machine 1071 00:49:27,320 --> 00:49:30,120 how well it's doing when it's trying to make a picture. 1072 00:49:30,120 --> 00:49:32,880 Now, the second part is making the picture, 1073 00:49:32,880 --> 00:49:36,280 which is using diffusion training. Now, this is a bit of a head fry, 1074 00:49:36,280 --> 00:49:39,400 but we're going to try and explain it to you. 1075 00:49:39,400 --> 00:49:43,920 What you do is you take an image and you add some visual noise to it, 1076 00:49:43,920 --> 00:49:47,720 adding a few random pixels to it. And we're teaching the network 1077 00:49:47,720 --> 00:49:50,080 the relationship between these pictures. 1078 00:49:50,080 --> 00:49:52,840 So I know that if I add a bit more noise, it turns into this. 1079 00:49:52,840 --> 00:49:55,520 And a bit more and a bit more, and you keep adding noise. 1080 00:49:55,520 --> 00:49:58,000 - And the picture is disappearing. - And the picture disappears 1081 00:49:58,000 --> 00:50:00,240 until you've only got a field of noise. 1082 00:50:00,240 --> 00:50:03,520 And if you train that on lots and lots of images, 1083 00:50:03,520 --> 00:50:07,360 for example, lots and lots of cats with lots of different text, 1084 00:50:07,360 --> 00:50:11,440 the machine is then able to reconstruct images from the noise 1085 00:50:11,440 --> 00:50:14,200 by taking a bit of noise away each time. 1086 00:50:14,200 --> 00:50:17,720 So if I'm going to try the prompt "a tabby cat" 1087 00:50:17,720 --> 00:50:20,440 and I start with just pure noise, 1088 00:50:20,440 --> 00:50:24,160 at each step, the network tries to take away the dots 1089 00:50:24,160 --> 00:50:27,680 that will make it look a bit more like a tabby cat. 1090 00:50:27,680 --> 00:50:29,960 And you do that over and over and over again, 1091 00:50:29,960 --> 00:50:32,000 until the image emerges. 1092 00:50:32,000 --> 00:50:34,440 This model has not been trained only on cats, obviously. 1093 00:50:34,440 --> 00:50:37,080 It's been trained on millions and millions and millions 1094 00:50:37,080 --> 00:50:40,240 of images from the internet, along with their associated text. 1095 00:50:40,240 --> 00:50:44,280 So we've got a bus or a watercolour painting or a block of cheese. 1096 00:50:44,280 --> 00:50:48,080 People playing chess. And once it has all these concepts inside, 1097 00:50:48,080 --> 00:50:50,000 you can start blending them together. 1098 00:50:50,000 --> 00:50:53,280 So I can ask for a tabby cat playing chess, for example, 1099 00:50:53,280 --> 00:50:56,560 because it understands what chess is, it understands what cats are. 1100 00:50:56,560 --> 00:50:59,360 Or a tabby cat on a London bus. 1101 00:50:59,360 --> 00:51:01,960 - AUDIENCE: - Aw! 1102 00:51:01,960 --> 00:51:05,560 - Now the idea of randomising and then de-randomising, 1103 00:51:05,560 --> 00:51:08,360 just introducing noise until the image disappears 1104 00:51:08,360 --> 00:51:11,160 and then somehow being able to kind of un-randomise, 1105 00:51:11,160 --> 00:51:13,600 this, to be honest, is one of the weirdest 1106 00:51:13,600 --> 00:51:16,600 and most wonderful developments in AI over the last few years. 1107 00:51:16,600 --> 00:51:18,320 And there's quite a lot of mathematics 1108 00:51:18,320 --> 00:51:19,760 going on behind the scenes. 1109 00:51:19,760 --> 00:51:23,480 So even though the image that the AI creates is new, 1110 00:51:23,480 --> 00:51:27,400 it's worth thinking about how it knows what to create. 1111 00:51:27,400 --> 00:51:29,520 And it needs training data. 1112 00:51:29,520 --> 00:51:32,120 Remember, AI needs training data. 1113 00:51:32,120 --> 00:51:36,760 The more training data it has, the better it's going to be. 1114 00:51:36,760 --> 00:51:40,760 And so lots of artists are concerned that their images, 1115 00:51:40,760 --> 00:51:44,480 their creative output, is being used to train AI, 1116 00:51:44,480 --> 00:51:48,800 which then is used to create images like the one that we've seen. 1117 00:51:48,800 --> 00:51:53,120 So what is it like for you as an artist using this tool? 1118 00:51:53,120 --> 00:51:56,440 - What is fascinating, it's like having an infinite number 1119 00:51:56,440 --> 00:51:58,800 of studio assistants. I can ask them to make some ideas 1120 00:51:58,800 --> 00:52:01,280 for me and they come back and produce thousands of versions 1121 00:52:01,280 --> 00:52:04,320 and I can choose the ones that I like and manipulate them further. 1122 00:52:04,320 --> 00:52:08,360 So it's actually an incredible resource for creativity. 1123 00:52:08,360 --> 00:52:10,640 - OK. But you're the artist, you're steering it. 1124 00:52:10,640 --> 00:52:13,760 I mean, you're not just giving it a prompt. It's doing it, right? 1125 00:52:13,760 --> 00:52:15,600 - Well, to some degree, you give it a prompt 1126 00:52:15,600 --> 00:52:16,920 and it does something. 1127 00:52:16,920 --> 00:52:19,000 But it doesn't always do what you want it to. - OK. 1128 00:52:19,000 --> 00:52:21,440 - And it takes a lot of work to get the desired outputs. 1129 00:52:21,440 --> 00:52:25,760 - OK. Still images are great, but what about videos? 1130 00:52:25,760 --> 00:52:30,080 Now, before the lectures, we asked you to suggest some dreams 1131 00:52:30,080 --> 00:52:32,520 that you'd had. If you can dream it, 1132 00:52:32,520 --> 00:52:35,440 we think AI can produce a video of it. 1133 00:52:35,440 --> 00:52:37,200 And we picked out from your suggestions 1134 00:52:37,200 --> 00:52:39,560 some of the more interesting and original ones. 1135 00:52:39,560 --> 00:52:41,360 And here are a couple of the ones that we got. 1136 00:52:41,360 --> 00:52:45,520 "There was a cat which was eating a fish and then more fish came 1137 00:52:45,520 --> 00:52:48,000 "and slapped the cat with banana peels." 1138 00:52:48,000 --> 00:52:50,760 OK, hands up if you were responsible for this 1139 00:52:50,760 --> 00:52:53,720 rather remarkable... OK, this was you. 1140 00:52:53,720 --> 00:52:55,640 OK, I don't know what you've been eating! 1141 00:52:55,640 --> 00:52:59,680 "The sky was pink. I was walking and fell off a cliff. 1142 00:52:59,680 --> 00:53:02,600 "I landed on a huge marshmallow." Of course you did! 1143 00:53:02,600 --> 00:53:04,520 "With a packet of jelly beans on top." 1144 00:53:04,520 --> 00:53:06,600 Of course there was a packet of jelly beans on top! 1145 00:53:06,600 --> 00:53:08,840 "I ate the jelly beans and the marshmallow 1146 00:53:08,840 --> 00:53:12,680 "and carried on walking. It happened again and again." 1147 00:53:12,680 --> 00:53:14,320 What a remarkable life you lead! 1148 00:53:14,320 --> 00:53:18,040 But we picked one from Arabella. Arabella, where are you? 1149 00:53:18,040 --> 00:53:20,160 Arabella, you're in the middle. Right. 1150 00:53:20,160 --> 00:53:22,280 I'm going to have to pass the microphone. 1151 00:53:22,280 --> 00:53:24,760 OK, Arabella, what was your prompt? 1152 00:53:24,760 --> 00:53:28,240 - So I was baking, baking a mug cake, 1153 00:53:28,240 --> 00:53:33,480 and the mug cake started talking to me and telling me some jokes, 1154 00:53:33,480 --> 00:53:35,120 and, yeah. 1155 00:53:35,120 --> 00:53:38,360 - OK, so Eric, talk us through what you managed to do 1156 00:53:38,360 --> 00:53:40,480 with that prompt? - The first thing I discovered 1157 00:53:40,480 --> 00:53:42,920 is that AI is not very good at mug cakes! 1158 00:53:44,640 --> 00:53:47,400 It's blended cakes and mugs quite nicely, 1159 00:53:47,400 --> 00:53:50,960 but I don't think that's quite what you had in your mind. 1160 00:53:50,960 --> 00:53:54,600 And even if you take one of these cakes, these mug cakes, 1161 00:53:54,600 --> 00:53:57,600 and I ask it to tell me a joke, it's very difficult to predict 1162 00:53:57,600 --> 00:53:59,720 how the video model is going to work. 1163 00:53:59,720 --> 00:54:04,040 And as you can see, this is not quite a cake telling a joke. 1164 00:54:08,120 --> 00:54:10,040 Here's another one. I quite like this one. 1165 00:54:10,040 --> 00:54:13,080 It's got a lovely smiley face, but it decided just to rotate him 1166 00:54:13,080 --> 00:54:15,920 for some reason. I don't know why. 1167 00:54:15,920 --> 00:54:19,320 So it's early days with this technology, 1168 00:54:19,320 --> 00:54:21,440 and it's a bit of trial and error. 1169 00:54:21,440 --> 00:54:26,880 So I changed your prompt slightly and went for a cupcake telling jokes 1170 00:54:26,880 --> 00:54:29,240 because the model knows about cupcakes slightly better. 1171 00:54:29,240 --> 00:54:32,480 And I decided the cupcake wanted to start its career 1172 00:54:32,480 --> 00:54:36,840 in standup comedy at a comedy club called The Happy Cake. 1173 00:54:36,840 --> 00:54:38,720 Now, as you can see, text is one of the things 1174 00:54:38,720 --> 00:54:40,720 that AI is not very good at at the moment. 1175 00:54:40,720 --> 00:54:43,320 - It can't spell happy. - It can't spell happy. 1176 00:54:43,320 --> 00:54:47,360 And in the end I went for this one, the happy ca-e-ake, 1177 00:54:47,360 --> 00:54:50,520 just because I thought it was great. 1178 00:54:50,520 --> 00:54:53,840 And the character I chose was this guy, the cupcake, 1179 00:54:53,840 --> 00:54:56,640 which we'll see. It took quite a lot of time to get the cupcake 1180 00:54:56,640 --> 00:54:59,960 to tell a joke. As you can see, it grew a moustache in this one, 1181 00:54:59,960 --> 00:55:02,240 I don't really know why. 1182 00:55:02,240 --> 00:55:08,360 I was able to generate a video for you that contains the cupcake. 1183 00:55:08,360 --> 00:55:11,880 Everything you see has been generated by AI, all of the video, 1184 00:55:11,880 --> 00:55:14,720 all of the images, the voices, are AI generated. 1185 00:55:14,720 --> 00:55:18,240 The music at the end is generated by AI, including the singing, 1186 00:55:18,240 --> 00:55:22,640 and the terrible, terrible joke that he tells came from ChatGPT. 1187 00:55:22,640 --> 00:55:25,960 - OK, so let's see the video, Eric! 1188 00:55:25,960 --> 00:55:29,840 SLURPING SQUELCH 1189 00:55:29,840 --> 00:55:34,440 LIQUID POURING, BLENDER BUZZES 1190 00:55:34,440 --> 00:55:36,880 - I love telling jokes. Wow! Open Mic Night 1191 00:55:36,880 --> 00:55:39,720 at the Happy Cake Comedy Club! This is my chance! 1192 00:55:39,720 --> 00:55:44,400 - ENGINE ROARS, TYRES SQUEAL 1193 00:55:44,400 --> 00:55:47,920 - Why did the cupcake become a stand up comedian? 1194 00:55:47,920 --> 00:55:51,040 Because it had a talent for delivering punch lines 1195 00:55:51,040 --> 00:55:53,600 that were as sweet as its frosting! 1196 00:55:55,040 --> 00:55:56,600 Wow, I'm a hit! 1197 00:55:56,600 --> 00:55:59,360 - # Everybody loves Comedy Cake Mix! # 1198 00:55:59,360 --> 00:56:02,080 LAUGHTER AND APPLAUSE 1199 00:56:07,440 --> 00:56:11,760 Let's go back to Arabella and see what Arabella made of that. 1200 00:56:11,760 --> 00:56:16,120 Was that what you had in mind, Arabella? - Erm, I guess so! 1201 00:56:18,440 --> 00:56:21,480 - Were you surprised at how good it was or how bad it was? 1202 00:56:21,480 --> 00:56:23,520 - Both, really! - Both! 1203 00:56:23,520 --> 00:56:25,000 It kind of was a bit of both, 1204 00:56:25,000 --> 00:56:27,600 but it was kind of amazing, wasn't it? 1205 00:56:27,600 --> 00:56:29,360 - The singing was interesting! 1206 00:56:30,640 --> 00:56:32,880 - Arabella, thank you so much. 1207 00:56:36,960 --> 00:56:39,360 Arabella, Eric, you're going to have to fight over 1208 00:56:39,360 --> 00:56:42,000 who gets the Oscar for this one, I think! 1209 00:56:42,000 --> 00:56:44,800 OK. Now, remember, AI created that, 1210 00:56:44,800 --> 00:56:46,880 but it created it under your direction. 1211 00:56:46,880 --> 00:56:49,640 You just didn't give it Arabella's prompt and all of that appeared? 1212 00:56:49,640 --> 00:56:51,640 - If only. Not yet. We're not quite at that point. 1213 00:56:51,640 --> 00:56:55,360 Took me about a day or so of coercing the images out of it. 1214 00:56:55,360 --> 00:56:57,480 I had to put them together, add sound effects. 1215 00:56:57,480 --> 00:56:59,960 There's still quite a lot of human activity involved. 1216 00:56:59,960 --> 00:57:03,720 - And if that went to a team of human professionals, how long do you think it might have taken them? 1217 00:57:03,720 --> 00:57:06,400 - I imagine it would take them a week or two to generate all the assets 1218 00:57:06,400 --> 00:57:09,000 and build the whole thing. So it's definitely a time saver. 1219 00:57:09,000 --> 00:57:11,840 - OK. The technology is getting better and better all the time. 1220 00:57:11,840 --> 00:57:14,440 Eric, thank you so much for joining us. - Thank you. 1221 00:57:18,920 --> 00:57:22,720 - In this second lecture, we've looked at how AI is being used 1222 00:57:22,720 --> 00:57:25,680 all around us in our everyday life. 1223 00:57:25,680 --> 00:57:29,640 It's reshaping our health care systems and asking us to confront 1224 00:57:29,640 --> 00:57:33,160 what creativity means, what art means. 1225 00:57:33,160 --> 00:57:37,200 Does AI democratise the means of producing art? 1226 00:57:37,200 --> 00:57:42,080 Is art created with AI tools less emotional or engaging 1227 00:57:42,080 --> 00:57:43,880 than that created by humans? 1228 00:57:43,880 --> 00:57:45,720 Is it even art? 1229 00:57:45,720 --> 00:57:47,440 You have to decide that. 1230 00:57:47,440 --> 00:57:50,120 But, remember, AI is a tool. 1231 00:57:50,120 --> 00:57:52,640 It's not a mind like you or me. 1232 00:57:54,040 --> 00:57:57,040 In lecture three, we're going to look into the future. 1233 00:57:57,040 --> 00:57:59,840 We're going to consider the dreams that we have for AI, 1234 00:57:59,840 --> 00:58:01,960 but also the nightmares. 1235 00:58:01,960 --> 00:58:05,080 We're going to consider the ways in which this new technology 1236 00:58:05,080 --> 00:58:08,680 might develop in directions that we don't want it to develop. 1237 00:58:08,680 --> 00:58:10,360 And we're going to try and understand 1238 00:58:10,360 --> 00:58:13,560 how we can avoid those directions. Thank you. 1239 00:58:17,560 --> 00:58:19,240 AI is here now. 1240 00:58:19,240 --> 00:58:21,960 Discover the different ways it may already be revolutionising 1241 00:58:21,960 --> 00:58:24,720 your life in my exclusive interview. 1242 00:58:24,720 --> 00:58:28,320 To watch, head to bbc.co.uk/ri 1243 00:58:28,320 --> 00:58:30,720 and follow the links to the Open University.