WEBVTT 00:00:01.040 --> 00:00:04.938 I'm here to offer you a new way to think about my field, 00:00:04.938 --> 00:00:06.729 artificial intelligence. 00:00:06.729 --> 00:00:08.751 I think the purpose of AI 00:00:08.751 --> 00:00:13.286 is to empower humans with machine intelligence, 00:00:13.286 --> 00:00:15.760 and as machines get smarter, 00:00:15.760 --> 00:00:18.487 we get smarter. 00:00:18.487 --> 00:00:20.658 I call this humanistic AI: 00:00:20.658 --> 00:00:24.352 artificial intelligence designed to meet human needs 00:00:24.352 --> 00:00:28.419 by collaborating and augmenting people. 00:00:28.419 --> 00:00:30.084 Now today I'm happy to see 00:00:30.084 --> 00:00:32.163 that the idea that intelligent assistant 00:00:32.163 --> 00:00:33.921 is mainstream. 00:00:33.921 --> 00:00:39.700 It's the well-accepted metaphor for the interface between humans and AI. 00:00:39.700 --> 00:00:43.252 And the one I helped create is called Siri. NOTE Paragraph 00:00:43.252 --> 00:00:44.447 Now you know Siri. 00:00:44.447 --> 00:00:47.209 Siri is the thing that knows your intent 00:00:47.209 --> 00:00:49.274 and helps you do it for you, 00:00:49.274 --> 00:00:50.590 helps you get things done, 00:00:50.590 --> 00:00:53.282 but what you might not know is that we designed Siri 00:00:53.282 --> 00:00:55.865 as humanistic AI, 00:00:55.865 --> 00:00:58.616 to augment people with a conversational interface 00:00:58.616 --> 00:01:01.915 that made it possible for them to use mobile computing, 00:01:01.915 --> 00:01:06.981 regardless of who they were and their abilities. NOTE Paragraph 00:01:06.981 --> 00:01:08.581 Now for most of us, 00:01:08.581 --> 00:01:11.385 the impact of this technology 00:01:11.385 --> 00:01:14.080 is to make things a little bit easier to use, 00:01:14.080 --> 00:01:17.463 but for my friend Daniel, 00:01:17.463 --> 00:01:21.217 the impact of the AI in these systems is a life-changer. 00:01:21.217 --> 00:01:26.165 You see, Daniel is a really social guy, 00:01:26.165 --> 00:01:28.846 and he's blind and quadriplegic 00:01:28.846 --> 00:01:31.148 which makes it hard to use those devices 00:01:31.148 --> 00:01:32.680 that we all take for granted. 00:01:32.680 --> 00:01:34.178 The last time I was at his house, his brother said, 00:01:34.178 --> 00:01:36.211 "Hang on a second, Daniel's not ready. 00:01:36.211 --> 00:01:39.845 He's on the phone with a woman he met online." 00:01:39.845 --> 00:01:42.158 I'm like, "That's cool, how'd he do it?" 00:01:42.158 --> 00:01:46.577 Well, Daniel uses Siri to manage his own social life, 00:01:46.577 --> 00:01:49.002 his email, text, and phone, 00:01:49.002 --> 00:01:52.644 without depending on his caregivers. 00:01:52.644 --> 00:01:55.125 This is kind of interesting, right? 00:01:55.125 --> 00:01:55.830 The irony here is great. 00:01:55.830 --> 00:01:58.741 Here's the man whose relationship with AI 00:01:58.741 --> 00:02:02.964 helps him have relationships with genuine human beings. 00:02:02.964 --> 00:02:07.274 And this is humanistic AI. NOTE Paragraph 00:02:07.274 --> 00:02:09.908 Another example with life-changing consequences 00:02:09.908 --> 00:02:13.002 is diagnosing cancer. 00:02:13.002 --> 00:02:15.573 When a doctor suspects cancer, 00:02:15.573 --> 00:02:17.820 they take a sample, send it to a pathologist 00:02:17.820 --> 00:02:19.662 who looks at it under a microscope. 00:02:19.662 --> 00:02:22.629 Now, pathologists look at hundreds of slides 00:02:22.629 --> 00:02:25.694 and millions of cells every day, 00:02:25.694 --> 00:02:27.104 so to support this task, 00:02:27.104 --> 00:02:31.691 some researchers made an AI classifier. 00:02:31.691 --> 00:02:35.809 Now the classifier says, "Is this cancer or is this not cancer?" 00:02:35.809 --> 00:02:38.063 looking at the picture. 00:02:38.063 --> 00:02:40.408 The classifier was pretty good, 00:02:40.408 --> 00:02:43.779 but not as good as the person, 00:02:43.779 --> 00:02:46.282 who got it right most of the time. 00:02:46.282 --> 00:02:49.070 But when they combine the ability 00:02:49.070 --> 00:02:50.986 of the machine and the human together, 00:02:50.986 --> 00:02:57.014 accuracy went to 99.5 percent. 00:02:57.014 --> 00:03:00.386 Adding that AI to a partnership 00:03:00.386 --> 00:03:03.908 eliminated 85 percent of the errors that the human pathologist 00:03:03.908 --> 00:03:06.900 would have made working alone. 00:03:06.900 --> 00:03:08.065 That's a lot of cancer 00:03:08.065 --> 00:03:10.828 that would have otherwise gone untreated. 00:03:10.828 --> 00:03:12.396 Now, for the curious, it turns out that the human is better at 00:03:12.396 --> 00:03:16.627 rejecting false positives 00:03:16.627 --> 00:03:20.339 and the machine was better at recognizing those hard-to-spot cases, 00:03:20.339 --> 00:03:23.112 but the lesson here isn't about which agent is better 00:03:23.112 --> 00:03:25.689 at this image classification task. 00:03:25.689 --> 00:03:27.844 Those things are changing every day. 00:03:27.844 --> 00:03:29.343 The lesson here 00:03:29.343 --> 00:03:32.727 is that by combining the abilities of the human and machine, 00:03:32.727 --> 00:03:34.811 it created a partnership 00:03:34.811 --> 00:03:38.395 that had superhuman performance. 00:03:38.395 --> 00:03:42.827 And that is humanistic AI. NOTE Paragraph 00:03:42.827 --> 00:03:44.836 Now let's look at another example 00:03:44.836 --> 00:03:47.241 with turbocharging performance. 00:03:47.241 --> 00:03:48.883 This is design. 00:03:48.883 --> 00:03:51.291 Now, let's say you're an engineer. 00:03:51.291 --> 00:03:53.009 You want to design a new frame for a drum. 00:03:53.009 --> 00:03:54.921 You get out your favorite software tools, 00:03:54.921 --> 00:03:56.107 cad tools, 00:03:56.107 --> 00:04:00.345 and you enter the form and the materials and then you analyze performance. 00:04:00.345 --> 00:04:02.408 That gives you one design. 00:04:02.408 --> 00:04:04.406 If you give those same tools to an AI, 00:04:04.406 --> 00:04:06.322 it can generate thousands of designs. 00:04:06.322 --> 00:04:12.307 This video by Autodesk is amazing. 00:04:12.307 --> 00:04:14.331 This is real stuff. 00:04:14.331 --> 00:04:17.187 So now this transforms how we do design. 00:04:17.187 --> 00:04:18.768 The human engineer now 00:04:18.768 --> 00:04:21.978 says what the machine, what the design should achieve, 00:04:21.978 --> 00:04:25.234 and the machine says, here's the possibilities. 00:04:25.234 --> 00:04:29.014 Now in her job now, the engineer's job 00:04:29.014 --> 00:04:31.846 is to pick the one that best meets the goals of the design, 00:04:31.846 --> 00:04:34.230 which she knows as a human 00:04:34.230 --> 00:04:35.592 better than anyone else, 00:04:35.592 --> 00:04:37.565 using human judgment and expertise. 00:04:37.565 --> 00:04:38.937 In this case, the winning form 00:04:38.937 --> 00:04:41.753 looks kind of like something nature would have designed, 00:04:41.753 --> 00:04:44.107 minus a few million years of evolution 00:04:44.107 --> 00:04:48.054 and all that unnecessary fur. NOTE Paragraph 00:04:48.054 --> 00:04:51.987 Now let's see where this idea 00:04:51.987 --> 00:04:53.369 of humanistic AI might lead us 00:04:53.369 --> 00:04:56.611 if we follow it into the speculative beyond. 00:04:56.611 --> 00:04:59.704 What's a kind of augmentation that we would all like to have? 00:04:59.704 --> 00:05:03.084 Well, how about cognitive enhancement? 00:05:03.084 --> 00:05:05.334 Instead of asking 00:05:05.334 --> 00:05:07.706 how smart can we make our machines, 00:05:07.706 --> 00:05:09.304 let's ask how smart 00:05:09.304 --> 00:05:12.885 can our machines make us? 00:05:12.885 --> 00:05:15.833 I mean, take memory for example. 00:05:15.833 --> 00:05:19.200 Memory is the foundation of human intelligence, 00:05:19.200 --> 00:05:23.733 but human memory is famously flawed. 00:05:23.733 --> 00:05:25.829 We're great at telling stories, 00:05:25.829 --> 00:05:28.086 but not getting the details right, 00:05:28.086 --> 00:05:31.082 and our memories, they decay over time. 00:05:31.082 --> 00:05:35.568 I mean, like, where did the '60s go, and can I go there too? 00:05:35.568 --> 00:05:39.133 But what if you could have a memory 00:05:39.133 --> 00:05:41.672 that was as good as computer memory, 00:05:41.672 --> 00:05:45.960 and was about your life? 00:05:45.960 --> 00:05:48.772 What if you could remember every person you ever met, 00:05:48.772 --> 00:05:50.497 how to pronounce their name, 00:05:50.497 --> 00:05:52.613 their family details, their favorite sports, 00:05:52.613 --> 00:05:54.635 the last conversation you had with them? 00:05:54.635 --> 00:05:57.313 If you had this memory all your life, 00:05:57.313 --> 00:06:00.634 you could have the AI look at all the interactions 00:06:00.634 --> 00:06:02.113 you had with people over time 00:06:02.113 --> 00:06:03.728 and help you reflect 00:06:03.728 --> 00:06:06.816 on the long arc of your relationships. 00:06:06.816 --> 00:06:09.699 What if you could have the AI read 00:06:09.699 --> 00:06:11.795 everything you've ever read 00:06:11.795 --> 00:06:15.344 and listen to every song you've ever heard? 00:06:15.344 --> 00:06:17.944 From the tiniest clue, it could help you retrieve 00:06:17.944 --> 00:06:22.161 anything you've ever seen or heard before. 00:06:22.161 --> 00:06:24.567 Imagine what that would do for the ability to make new connections 00:06:24.567 --> 00:06:27.059 and form new ideas. 00:06:27.059 --> 00:06:30.010 And what about our bodies? 00:06:30.010 --> 00:06:34.225 What if we could remember the consequences of every food we eat, 00:06:34.225 --> 00:06:36.676 every pill we take, 00:06:36.676 --> 00:06:39.159 every all-nighter we pull? 00:06:39.159 --> 00:06:41.892 We could do our own science on our own data 00:06:41.892 --> 00:06:44.052 about what makes us feel good 00:06:44.052 --> 00:06:45.774 and stay healthy, 00:06:45.774 --> 00:06:48.166 and imagine how this could revolutionize 00:06:48.166 --> 00:06:53.084 the way we manage allergies and chronic disease. NOTE Paragraph 00:06:53.084 --> 00:06:55.908 I believe that AI will make 00:06:55.908 --> 00:06:59.281 personal memory enhancement a reality. 00:06:59.281 --> 00:07:02.347 I can't say when or what form factors are involved, 00:07:02.347 --> 00:07:04.803 but I think it's inevitable 00:07:04.803 --> 00:07:09.314 because the very things 00:07:09.314 --> 00:07:11.099 that make AI successful today, 00:07:11.099 --> 00:07:13.478 the availability of comprehensive data 00:07:13.478 --> 00:07:17.049 and the ability for machines to make sense of that data 00:07:17.049 --> 00:07:20.664 can be applied to the data of our lives, 00:07:20.664 --> 00:07:22.794 and those data are here today, 00:07:22.794 --> 00:07:24.475 available for all of us because we lead 00:07:24.475 --> 00:07:28.816 digitally mediated lives, 00:07:28.816 --> 00:07:32.111 in mobile and online. NOTE Paragraph 00:07:32.111 --> 00:07:34.760 In my view, a personal memory is a private memory. 00:07:34.760 --> 00:07:40.305 We get to choose what is and is not recalled and retained. 00:07:40.305 --> 00:07:42.031 It's absolutely essential 00:07:42.031 --> 00:07:45.477 that this be kept very secure. NOTE Paragraph 00:07:45.477 --> 00:07:47.202 Now for most of us, 00:07:47.202 --> 00:07:49.848 the impact of augmented personal memory 00:07:49.848 --> 00:07:52.676 will be a more improved mental gain, 00:07:52.676 --> 00:07:56.658 maybe, hopefully, a bit more social grace. 00:07:56.658 --> 00:07:58.925 But for the millions who suffer 00:07:58.925 --> 00:08:02.191 from Alzheimer's and dementia, 00:08:02.191 --> 00:08:05.436 the difference that augmented memory could make 00:08:05.436 --> 00:08:09.405 is a difference between a life of isolation 00:08:09.405 --> 00:08:12.429 and a life of dignity and connection. 00:08:12.429 --> 00:08:14.161 We are in the middle 00:08:14.161 --> 00:08:16.005 of a renaissance 00:08:16.005 --> 00:08:18.819 in artificial intelligence right now. 00:08:18.819 --> 00:08:20.905 I mean, in just the past few years 00:08:20.905 --> 00:08:23.185 we're beginning to see solutions to AI problems 00:08:23.185 --> 00:08:26.403 that we have struggled with 00:08:26.403 --> 00:08:29.369 literally for decades: 00:08:29.369 --> 00:08:32.110 speech understanding, text understanding, 00:08:32.110 --> 00:08:34.329 image understanding. 00:08:34.329 --> 00:08:39.483 We have a choice in how we use this powerful technology. 00:08:39.483 --> 00:08:42.683 We can choose to use AI to automate 00:08:42.683 --> 00:08:44.517 and compete with us, 00:08:44.517 --> 00:08:48.192 or we can use AI to augment and collaborate with us, 00:08:48.192 --> 00:08:50.967 to overcome our cognitive limitations 00:08:50.967 --> 00:08:54.866 and to help us do what we want to do 00:08:54.866 --> 00:08:56.617 only better. 00:08:56.617 --> 00:08:59.732 And as we discover new ways 00:08:59.732 --> 00:09:03.094 to give machines intelligence, 00:09:03.094 --> 00:09:05.960 we can distribute that intelligence 00:09:05.960 --> 00:09:08.498 to all of the AI assistants in the world, 00:09:08.498 --> 00:09:13.340 and therefore to every person, regardless of circumstance. 00:09:13.340 --> 00:09:14.892 And that is why, 00:09:14.892 --> 00:09:18.811 every time a machine gets smarter, 00:09:18.811 --> 00:09:21.140 we get smarter. NOTE Paragraph 00:09:21.140 --> 00:09:24.996 That is an AI worth spreading. NOTE Paragraph 00:09:24.996 --> 00:09:26.865 Thank you. NOTE Paragraph 00:09:26.865 --> 00:09:33.122 (Applause)