Ayanna Howard: Human-Robot Interaction & Ethics of Safety-Critical Systems

00:00:00.000 | The following is a conversation with Ayana Howard.

00:00:03.400 | She's a roboticist, professor at Georgia Tech

00:00:06.200 | and director of the Human Automation Systems Lab

00:00:09.840 | with research interests in human-robot interaction,

00:00:12.800 | assistive robots in the home, therapy gaming apps,

00:00:16.000 | and remote robotic exploration of extreme environments.

00:00:20.280 | Like me, in her work, she cares a lot

00:00:23.440 | about both robots and human beings.

00:00:26.360 | And so I really enjoyed this conversation.

00:00:29.560 | This is the Artificial Intelligence Podcast.

00:00:32.600 | If you enjoy it, subscribe on YouTube,

00:00:34.960 | give it five stars on Apple Podcast,

00:00:36.960 | follow on Spotify, support it on Patreon,

00:00:39.600 | or simply connect with me on Twitter,

00:00:41.720 | Alex Friedman, spelled F-R-I-D-M-A-N.

00:00:45.640 | I recently started doing ads at the end of the introduction.

00:00:48.680 | I'll do one or two minutes after introducing the episode

00:00:51.640 | and never any ads in the middle

00:00:53.200 | that can break the flow of the conversation.

00:00:55.480 | I hope that works for you

00:00:56.880 | and doesn't hurt the listening experience.

00:01:00.200 | This show is presented by Cash App,

00:01:02.280 | the number one finance app in the App Store.

00:01:04.800 | I personally use Cash App to send money to friends,

00:01:07.560 | but you can also use it to buy, sell,

00:01:09.320 | and deposit Bitcoin in just seconds.

00:01:11.720 | Cash App also has a new investing feature.

00:01:14.600 | You can buy fractions of a stock, say $1 worth,

00:01:17.560 | no matter what the stock price is.

00:01:19.680 | Brokerage services are provided by Cash App Investing,

00:01:22.600 | a subsidiary of Square and member SIPC.

00:01:25.880 | I'm excited to be working with Cash App

00:01:28.160 | to support one of my favorite organizations called FIRST,

00:01:31.600 | best known for their FIRST robotics and Lego competitions.

00:01:35.120 | They educate and inspire hundreds of thousands of students

00:01:38.400 | in over 110 countries

00:01:40.240 | and have a perfect rating at Charity Navigator,

00:01:42.840 | which means that donated money

00:01:44.160 | is used to maximum effectiveness.

00:01:46.880 | When you get Cash App from the App Store or Google Play

00:01:49.520 | and use code LEXPODCAST, you'll get $10

00:01:53.520 | and Cash App will also donate $10 to FIRST,

00:01:56.440 | which again is an organization

00:01:58.280 | that I've personally seen inspire girls and boys

00:02:01.080 | to dream of engineering a better world.

00:02:04.280 | And now here's my conversation with Ayana Howard.

00:02:08.320 | What or who is the most amazing robot you've ever met

00:02:13.640 | or perhaps had the biggest impact on your career?

00:02:16.760 | - I haven't met her, but I grew up with her,

00:02:21.120 | but of course Rosie.

00:02:22.760 | So, and I think it's because also--

00:02:25.240 | - Who's Rosie?

00:02:26.160 | - Rosie from the Jetsons.

00:02:27.800 | She is all things to all people, right?

00:02:31.000 | Think about it, like anything you wanted,

00:02:32.920 | it was like magic, it happened.

00:02:35.120 | - So people not only anthropomorphize,

00:02:37.920 | but project whatever they wish for the robot to be onto.

00:02:41.960 | - Onto Rosie, but also, I mean, think about it.

00:02:44.600 | She was socially engaging.

00:02:46.840 | She every so often had an attitude, right?

00:02:50.040 | She kept us honest.

00:02:51.960 | She would push back sometimes

00:02:53.800 | when George was doing some weird stuff,

00:02:57.080 | but she cared about people, especially the kids.

00:02:59.880 | She was like the perfect robot.

00:03:04.040 | - And you've said that people don't want

00:03:06.560 | their robots to be perfect.

00:03:08.360 | Can you elaborate that?

00:03:11.200 | What do you think that is?

00:03:12.080 | Just like you said, Rosie pushed back a little bit

00:03:14.840 | every once in a while.

00:03:15.800 | - Yeah, so I think it's that,

00:03:18.320 | so if you think about robotics in general,

00:03:19.920 | we want them because they enhance our quality of life.

00:03:24.000 | And usually that's linked to something that's functional.

00:03:27.080 | Even if you think of self-driving cars,

00:03:28.720 | why is there a fascination?

00:03:30.080 | Because people really do hate to drive.

00:03:31.600 | Like there's the Saturday driving where I can just speed,

00:03:35.400 | but then there's the, I have to go to work every day

00:03:37.600 | and I'm in traffic for an hour.

00:03:39.080 | I mean, people really hate that.

00:03:40.480 | And so robots are designed to basically enhance our ability

00:03:47.480 | to increase our quality of life.

00:03:49.800 | And so the perfection comes from this aspect of interaction.

00:03:54.400 | If I think about how we drive, if we drove perfectly,

00:04:00.120 | we would never get anywhere, right?

00:04:02.240 | So think about how many times you had to run past the light

00:04:07.200 | because you see the car behind you

00:04:09.080 | is about to crash into you,

00:04:10.440 | or that little kid kind of runs into the street

00:04:15.400 | and so you have to cross on the other side.

00:04:17.360 | See, there's no cars, right?

00:04:18.440 | Like if you think about it, we are not perfect drivers.

00:04:21.200 | Some of it is because it's our world.

00:04:23.560 | And so if you have a robot that is perfect

00:04:26.760 | in that sense of the word,

00:04:28.720 | they wouldn't really be able to function with us.

00:04:31.200 | - Can you linger a little bit on the word perfection?

00:04:34.520 | So from the robotics perspective, what does that word mean?

00:04:39.440 | And how is sort of the optimal behavior

00:04:42.880 | as you're describing different

00:04:44.480 | than what we think is perfection?

00:04:46.640 | - Yeah, so perfection, if you think about it

00:04:49.480 | in the more theoretical point of view,

00:04:52.000 | it's really tied to accuracy, right?

00:04:54.080 | So if I have a function,

00:04:55.640 | can I complete it at 100% accuracy with zero errors?

00:04:59.520 | And so that's kind of, if you think about perfection

00:05:04.200 | in the sense of the word.

00:05:05.040 | - And in the self-driving car realm,

00:05:07.560 | do you think from a robotics perspective,

00:05:10.480 | we kind of think that perfection means

00:05:13.960 | following the rules perfectly,

00:05:15.600 | sort of defining, staying in the lane, changing lanes.

00:05:19.640 | When there's a green light, you go,

00:05:20.920 | when there's a red light, you stop.

00:05:22.320 | And that's the, and be able to perfectly see

00:05:26.680 | all the entities in the scene.

00:05:29.160 | That's the limit of what we think of as perfection.

00:05:32.000 | - And I think that's where the problem comes

00:05:33.760 | is that when people think about perfection for robotics,

00:05:38.360 | the ones that are the most successful

00:05:40.840 | are the ones that are quote unquote perfect.

00:05:43.280 | Like I said, Rosie is perfect,

00:05:44.680 | but she actually wasn't perfect in terms of accuracy,

00:05:47.440 | but she was perfect in terms of how she interacted

00:05:50.400 | and how she adapted.

00:05:51.520 | And I think that's some of the disconnect

00:05:53.320 | is that we really want perfection

00:05:56.480 | with respect to its ability to adapt to us.

00:06:00.040 | We don't really want perfection

00:06:01.720 | with respect to 100% accuracy,

00:06:03.560 | with respect to the rules that we just made up anyway.

00:06:06.480 | Right, and so I think there's this disconnect sometimes

00:06:09.560 | between what we really want and what happens.

00:06:13.320 | And we see this all the time, like in my research, right?

00:06:16.000 | Like the optimal quote unquote optimal interactions

00:06:20.440 | are when the robot is adapting based on the person,

00:06:24.400 | not 100% following what's optimal based on the rules.

00:06:29.400 | - Just to linger on autonomous vehicles for a second,

00:06:32.720 | just your thoughts, maybe off the top of your head,

00:06:35.280 | how hard is that problem do you think

00:06:38.040 | based on what we just talked about?

00:06:39.920 | You know, there's a lot of folks

00:06:41.680 | in the automotive industry that are very confident

00:06:43.880 | from Elon Musk to Waymo to all these companies.

00:06:47.640 | How hard is it to solve that last piece?

00:06:50.480 | - The last mile.

00:06:51.440 | - The gap between the perfection and the human definition

00:06:56.440 | of how you actually function in this world?

00:06:59.480 | - Yeah, so this is a moving target.

00:07:00.600 | So I remember when all the big companies

00:07:04.480 | started to heavily invest in this.

00:07:06.800 | And there was a number of, even roboticists,

00:07:09.880 | as well as folks who were putting in the VCs

00:07:13.240 | and corporations, Elon Musk being one of them,

00:07:15.400 | that said, you know, self-driving cars on the road

00:07:18.360 | with people within five years.

00:07:20.920 | That was a little while ago.

00:07:24.600 | And now people are saying five years, 10 years, 20 years.

00:07:29.600 | Some are saying never, right?

00:07:31.560 | I think if you look at some of the things

00:07:33.760 | that are being successful is these,

00:07:39.480 | basically fixed environments

00:07:41.240 | where you still have some anomalies, right?

00:07:44.080 | You still have people walking, you still have stores,

00:07:46.560 | but you don't have other drivers, right?

00:07:50.160 | Like other human drivers are,

00:07:51.760 | is a dedicated space for the cars.

00:07:55.680 | Because if you think about robotics in general,

00:07:57.240 | where it's always been successful is in,

00:07:59.080 | I mean, you can say manufacturing,

00:08:00.640 | like way back in the day, right?

00:08:02.320 | It was a fixed environment.

00:08:03.360 | Humans were not part of the equation.

00:08:05.280 | We're a lot better than that.

00:08:07.240 | But like when we can carve out scenarios

00:08:11.000 | that are closer to that space,

00:08:13.840 | then I think that it's where we are.

00:08:16.720 | So a closed campus where you don't have self-driving cars

00:08:20.600 | and maybe some protection so that the students

00:08:23.800 | don't jet in front just because they wanna see what happens.

00:08:27.280 | Like having a little bit,

00:08:29.000 | I think that's where we're gonna see the most success

00:08:31.440 | in the near future.

00:08:32.280 | - And be slow moving.

00:08:33.720 | - Right, not 55, 60, 70 miles an hour,

00:08:37.880 | but the speed of a golf cart, right?

00:08:42.080 | - So that said, the most successful

00:08:45.200 | in the automotive industry robots operating today

00:08:47.880 | in the hands of real people

00:08:50.280 | are ones that are traveling over 55 miles an hour

00:08:53.920 | and in unconstrained environments,

00:08:55.560 | which is Tesla vehicles, so Tesla autopilot.

00:08:58.880 | So I just, I would love to hear sort of your,

00:09:01.720 | just thoughts of two things.

00:09:04.240 | So one, I don't know if you've gotten to see,

00:09:06.960 | you've heard about something called Smart Summon,

00:09:10.120 | where Tesla system, autopilot system,

00:09:13.480 | where the car drives zero occupancy,

00:09:15.880 | no driver in the parking lot,

00:09:17.960 | slowly sort of tries to navigate the parking lot

00:09:20.480 | to find itself to you.

00:09:22.680 | And there's some incredible amounts of videos

00:09:25.840 | and just hilarity that happens

00:09:27.640 | as it awkwardly tries to navigate this environment,

00:09:30.880 | but it's a beautiful nonverbal communication

00:09:33.560 | between machine and human that I think is a,

00:09:37.400 | it's like, it's some of the work that you do

00:09:39.320 | in this kind of interesting human-robot interaction space.

00:09:42.040 | So what are your thoughts in general about it?

00:09:43.800 | - So I do have that feature.

00:09:46.100 | - Do you drive a Tesla?

00:09:47.840 | - I do, mainly because I'm a gadget freak, right?

00:09:52.120 | So I'd say it's a gadget that happens to have some wheels.

00:09:55.640 | And yeah, I've seen some of the videos.

00:09:58.200 | - But what's your experience like?

00:09:59.400 | I mean, you're a human-robot interaction roboticist.

00:10:02.680 | You're a legit sort of expert in the field.

00:10:05.560 | So what does it feel for a machine to come to you?

00:10:08.040 | - It's one of these very fascinating things,

00:10:11.880 | but also I am hyper, hyper alert, right?

00:10:16.080 | Like I'm hyper alert, like my thumb is like,

00:10:20.560 | oh, okay, I'm ready to take over.

00:10:23.200 | Even when I'm in my car,

00:10:25.440 | I'm doing things like automated backing into,

00:10:28.760 | so there's like a feature where you can do this

00:10:30.640 | automating backing into a parking space,

00:10:33.200 | or bring the car out of your garage,

00:10:35.720 | or even, you know, pseudo autopilot on the freeway, right?

00:10:40.320 | I am hypersensitive.

00:10:42.280 | I can feel like as I'm navigating,

00:10:44.720 | I'm like, yeah, that's an error right there.

00:10:46.960 | Like I am very aware of it, but I'm also fascinated by it.

00:10:51.960 | And it does get better.

00:10:54.360 | Like I look and see it's learning

00:10:57.480 | from all of these people who are cutting it on.

00:11:00.440 | Like every time I cut it on, it's getting better, right?

00:11:04.160 | And so I think that's what's amazing about it is that.

00:11:07.160 | - This nice dance of you're still hyper vigilant.

00:11:10.360 | So you're still not trusting it at all.

00:11:12.840 | - Yeah.

00:11:13.680 | - And yet you're using it.

00:11:14.600 | On the highway, if I were to, like what,

00:11:17.600 | as a roboticist, we'll talk about trust a little bit.

00:11:20.300 | How do you explain that?

00:11:23.680 | You still use it.

00:11:25.080 | Is it the gadget freak part?

00:11:26.520 | Like where you just enjoy exploring technology

00:11:30.760 | or is that the right actually balance

00:11:33.720 | between robotics and humans is where you use it,

00:11:36.920 | but don't trust it.

00:11:38.360 | And somehow there's this dance

00:11:40.120 | that ultimately is a positive.

00:11:42.160 | - Yeah, so I think I just don't necessarily trust technology,

00:11:47.160 | but I'm an early adopter, right?

00:11:50.180 | So when it first comes out, I will use everything,

00:11:54.280 | but I will be very, very cautious of how I use it.

00:11:57.400 | - Do you read about it or do you explore it by just try it?

00:12:01.040 | Do you do like crudely, to put it crudely,

00:12:05.000 | do you read the manual or do you learn through exploration?

00:12:07.960 | - I'm an explorer.

00:12:08.800 | If I have to read the manual, then I do design.

00:12:12.320 | Then it's a bad user interface.

00:12:14.200 | It's a failure.

00:12:15.040 | - Elon Musk is very confident that you kind of take it

00:12:19.560 | from where it is now to full autonomy.

00:12:21.800 | So from this human robot interaction,

00:12:24.480 | where you don't really trust and then you try

00:12:26.680 | and then you catch it when it fails to,

00:12:29.160 | it's going to incrementally improve itself into full,

00:12:33.920 | where you don't need to participate.

00:12:36.520 | What's your sense of that trajectory?

00:12:39.860 | Is it feasible?

00:12:41.040 | So the promise there is by the end of next year,

00:12:44.560 | by the end of 2020, is the current promise.

00:12:47.200 | What's your sense about that journey that Tesla's on?

00:12:52.200 | - So there's kind of three things going on now.

00:12:56.760 | I think in terms of will people go,

00:13:01.760 | like as a user, as a adopter,

00:13:04.800 | will you trust going to that point?

00:13:08.440 | I think so, right?

00:13:10.080 | Like there are some users and it's because what happens is

00:13:13.000 | when you're hypersensitive at the beginning

00:13:16.720 | and then the technology tends to work,

00:13:19.360 | your apprehension slowly goes away.

00:13:23.880 | And as people, we tend to swing to the other extreme, right?

00:13:28.280 | Because like, oh, I was like hyper, hyper fearful

00:13:30.960 | or hypersensitive and it was awesome.

00:13:33.980 | And we just tend to swing, that's just human nature.

00:13:37.400 | And so you will have, I mean, and I--

00:13:38.880 | - It's a scary notion because most people

00:13:41.560 | are now extremely untrusting of autopilot.

00:13:45.040 | They use it, but they don't trust it.

00:13:46.500 | And it's a scary notion that there's a certain point

00:13:48.920 | where you allow yourself to look at the smartphone

00:13:51.400 | for like 20 seconds.

00:13:53.160 | And then there'll be this phase shift

00:13:55.440 | where it'll be like 20 seconds, 30 seconds,

00:13:57.600 | one minute, two minutes.

00:13:58.980 | It's a scary proposition.

00:14:02.040 | - But that's people, right?

00:14:03.520 | That's just, that's humans.

00:14:05.600 | I mean, I think of even our use of,

00:14:10.000 | I mean, just everything on the internet, right?

00:14:12.780 | Think about how reliant we are on certain apps

00:14:16.900 | and certain engines, right?

00:14:19.420 | 20 years ago, people have been like, oh yeah, that's stupid.

00:14:22.700 | Like that makes no sense.

00:14:23.980 | Like, of course that's false.

00:14:25.900 | Like now it's just like, oh, of course I've been using it.

00:14:29.100 | It's been correct all this time.

00:14:30.780 | Of course, aliens, I didn't think they existed,

00:14:34.380 | but now it says they do, obviously.

00:14:37.620 | - 100%, earth is flat.

00:14:41.920 | Okay, but you said three things.

00:14:43.840 | So one is the human kind of structure.

00:14:44.680 | - Okay, so one is the human.

00:14:45.800 | And I think there will be a group of individuals

00:14:47.820 | that will swing, right?

00:14:50.160 | - Teenagers.

00:14:51.720 | - I mean, it'll be adults.

00:14:54.360 | There's actually an age demographic

00:14:56.400 | that's optimal for technology adoption.

00:15:00.160 | And you can actually find them.

00:15:02.280 | And they're actually pretty easy to find

00:15:03.960 | just based on their habits, based on...

00:15:07.000 | So someone like me who wasn't a roboticist

00:15:10.400 | would probably be the optimal kind of person, right?

00:15:13.560 | Early adopter, okay with technology,

00:15:15.640 | very comfortable and not hypersensitive, right?

00:15:20.000 | I'm just hypersensitive 'cause I designed this stuff.

00:15:23.560 | So there is a target demographic that will swing.

00:15:25.920 | The other one though is you still have these humans

00:15:29.800 | that are on the road.

00:15:31.360 | That one is a harder thing to do.

00:15:34.680 | And as long as we have people that are on the same streets,

00:15:40.360 | that's gonna be the big issue.

00:15:42.560 | And it's just because you can't possibly,

00:15:45.320 | I won't say, you can't possibly map

00:15:47.000 | some of the silliness of human drivers, right?

00:15:51.480 | Like as an example, when you're next to that car

00:15:56.320 | that has that big sticker called student driver, right?

00:15:59.880 | Like you are like, oh, either I am going to like go around.

00:16:04.680 | Like we know that that person is just gonna make mistakes

00:16:07.960 | that make no sense, right?

00:16:09.320 | How do you map that information?

00:16:11.000 | Or if I am in a car and I look over

00:16:14.360 | and I see two fairly young looking individuals

00:16:19.280 | and there's no student driver bumper

00:16:21.160 | and I see them chit chatting to each other,

00:16:22.880 | I'm like, oh, that's an issue, right?

00:16:26.200 | So how do you get that kind of information

00:16:28.560 | and that experience into basically an autopilot?

00:16:33.560 | - Yeah, and there's millions of cases like that

00:16:37.320 | where we take little hints to establish context.

00:16:41.240 | I mean, you said kind of beautifully poetic human things,

00:16:44.400 | but there's probably subtle things about the environment,

00:16:47.160 | about it being maybe time for commuters

00:16:52.160 | to start going home from work

00:16:55.360 | and therefore you can make some kind of judgment

00:16:57.200 | about the group behavior of pedestrians, blah, blah, blah.

00:17:00.520 | - Yes, yes, or even cities, right?

00:17:02.720 | Like if you're in Boston, how people cross the street,

00:17:07.160 | like lights are not an issue versus other places

00:17:10.680 | where people will actually wait for the crosswalk.

00:17:15.680 | - Seattle or somewhere peaceful.

00:17:18.240 | What I've also seen, so just even in Boston,

00:17:22.560 | that intersection to intersection is different.

00:17:25.520 | So every intersection has a personality of its own.

00:17:28.960 | So certain neighborhoods of Boston are different.

00:17:30.880 | So we're kind of, based on different timing of day,

00:17:35.320 | at night, it's all, there's a dynamic to human behavior

00:17:40.320 | that we kind of figure out ourselves.

00:17:42.480 | We're not able to introspect and figure it out,

00:17:46.120 | but somehow our brain learns it.

00:17:49.360 | - We do.

00:17:50.360 | - And so you're saying, is there a shortcut?

00:17:54.760 | Is there a shortcut though for,

00:17:56.440 | is there something that could be done, you think,

00:17:59.080 | that, you know, that's what we humans do.

00:18:02.680 | It's just like bird flight, right?

00:18:04.680 | That's the example they give for flight.

00:18:06.520 | Do you necessarily need to build a bird that flies

00:18:09.280 | or can you do an airplane?

00:18:10.740 | Is there a shortcut to it?

00:18:13.040 | - So I think that the shortcut is,

00:18:15.400 | and I kind of, I talk about it as a fixed space,

00:18:19.360 | where, so imagine that there's a neighborhood

00:18:23.320 | that's a new smart city or a new neighborhood that says,

00:18:26.840 | you know what, we are going to design this new city

00:18:31.520 | based on supporting self-driving cars.

00:18:33.840 | And then doing things, knowing that there's anomalies,

00:18:37.680 | knowing that people are like this, right?

00:18:39.640 | And designing it based on that assumption

00:18:42.120 | that like, we're going to have this,

00:18:43.960 | that would be an example of a shortcut.

00:18:45.560 | So you still have people, but you do very specific things

00:18:49.320 | to try to minimize the noise a little bit, as an example.

00:18:53.840 | - And the people themselves become accepting of the notion

00:18:56.200 | that there's autonomous cars, right?

00:18:57.760 | - Right, like they move into,

00:18:59.720 | so right now you have like a,

00:19:01.480 | you will have a self-selection bias, right?

00:19:03.600 | Like individuals will move into this neighborhood

00:19:06.280 | knowing like this is part of like the real estate pitch.

00:19:09.520 | Right?

00:19:10.680 | And so I think that's a way to do a shortcut.

00:19:14.200 | One, it allows you to deploy.

00:19:17.640 | It allows you to collect then data

00:19:19.640 | with these variances and anomalies,

00:19:22.640 | 'cause people are still people,

00:19:24.080 | but it's a safer space and is more of an accepting space.

00:19:28.880 | I.e. when something in that space might happen,

00:19:32.000 | because things do,

00:19:34.200 | because you already have the self-selection,

00:19:36.120 | like people would be, I think,

00:19:37.560 | a little more forgiving than other places.

00:19:40.800 | - And you said three things, did we cover all of them?

00:19:43.160 | - The third is legal law, liability,

00:19:46.440 | which I don't really want to touch,

00:19:47.920 | but it's still of concern.

00:19:51.000 | - And the mishmash with like, with policy as well,

00:19:53.360 | sort of government, all that whole--

00:19:55.840 | - That big ball of mess.

00:19:57.880 | - Yeah, gotcha.

00:19:59.200 | - So that's, so we're out of time now.

00:20:01.800 | Do you think from a robotics perspective,

00:20:06.040 | you know, if you're kind of honest of what cars do,

00:20:09.840 | they kind of threaten each other's life all the time.

00:20:14.840 | So cars are very, I mean,

00:20:17.680 | in order to navigate intersections, there's an assertiveness,

00:20:20.560 | there's a risk taking,

00:20:22.360 | and if you were to reduce it to an objective function,

00:20:25.320 | there's a probability of murder in that function,

00:20:28.800 | meaning you killing another human being,

00:20:31.960 | and you're using that, first of all,

00:20:34.440 | it has to be low enough to be acceptable to you

00:20:38.480 | on an ethical level as an individual human being,

00:20:41.360 | but it has to be high enough for people to respect you,

00:20:45.360 | to not sort of take advantage of you completely

00:20:47.600 | and jaywalk and funny and so on.

00:20:49.680 | So, I mean, I don't think there's a right answer here,

00:20:53.180 | but what's, how do we solve that?

00:20:56.160 | How do we solve that from a robotics perspective

00:20:58.040 | when danger and human life is at stake?

00:21:00.240 | - Yeah, as they say, cars don't kill people,

00:21:02.080 | people kill people.

00:21:02.920 | - People kill people.

00:21:05.160 | - Right, so I think--

00:21:08.720 | - Now robotic algorithms would be killing people.

00:21:10.840 | - Right, so it will be robotics algorithms that are,

00:21:14.480 | no, it will be robotic algorithms don't kill people,

00:21:17.080 | developers of robotic algorithms kill people, right?

00:21:19.840 | I mean, one of the things is people are still in the loop,

00:21:23.020 | and at least in the near and midterm,

00:21:26.640 | I think people will still be in the loop at some point,

00:21:29.480 | even if it's a developer.

00:21:30.360 | Like we're not necessarily at the stage where robots

00:21:33.600 | are programming autonomous robots

00:21:36.840 | with different behaviors quite yet.

00:21:39.400 | - It's a scary notion, sorry to interrupt,

00:21:42.360 | that a developer has some responsibility

00:21:47.360 | in the death of a human being.

00:21:49.720 | That's a heavy burden.

00:21:50.640 | - I mean, I think that's why the whole aspect of ethics

00:21:55.520 | in our community is so, so important, right?

00:21:59.040 | Because it's true.

00:22:00.240 | If you think about it, you can basically say,

00:22:04.880 | I'm not going to work on weaponized AI, right?

00:22:07.520 | Like people can say, that's not what I'm gonna do.

00:22:09.900 | But yet you are programming algorithms

00:22:12.760 | that might be used in healthcare algorithms

00:22:15.660 | that might decide whether this person

00:22:17.280 | should get this medication or not,

00:22:19.040 | and they don't, and they die.

00:22:20.660 | Okay, so that is your responsibility, right?

00:22:25.120 | And if you're not conscious and aware

00:22:27.360 | that you do have that power when you're coding

00:22:30.040 | and things like that, I think that's just not a good thing.

00:22:35.040 | Like we need to think about this responsibility

00:22:38.080 | as we program robots and computing devices

00:22:41.880 | much more than we are.

00:22:43.520 | - Yeah, so it's not an option to not think about ethics.

00:22:47.000 | I think it's a majority, I would say, of computer science.

00:22:50.940 | It's kind of a hot topic now, I think,

00:22:54.080 | about bias and so on, and we'll talk about it,

00:22:57.760 | but usually it's kind of,

00:22:59.160 | it's like a very particular group of people

00:23:02.680 | that work on that.

00:23:04.280 | And then people who do robotics are like,

00:23:06.960 | well, I don't have to think about that.

00:23:09.360 | There's other smart people thinking about it.

00:23:11.160 | It seems that everybody has to think about it.

00:23:14.560 | It's not, you can't escape the ethics,

00:23:17.040 | whether it's bias or just every aspect of ethics

00:23:21.120 | that has to do with human beings.

00:23:22.680 | - Everyone.

00:23:23.520 | - So think about, I'm gonna age myself,

00:23:25.720 | but I remember when we didn't have testers, right?

00:23:30.120 | And so what did you do?

00:23:31.080 | As a developer, you had to test your own code, right?

00:23:33.600 | Like you had to go through all the cases and figure it out,

00:23:36.120 | and then they realized that,

00:23:37.840 | we probably need to have testing

00:23:40.640 | because we're not getting all the things.

00:23:42.440 | And so from there, what happens is most developers,

00:23:45.560 | they do a little bit of testing,

00:23:47.320 | but it's usually like, okay, did my compiler bug out?

00:23:49.800 | Let me look at the warnings.

00:23:51.140 | Okay, is that acceptable or not?

00:23:52.920 | Right?

00:23:53.760 | Like that's how you typically think about it as a developer

00:23:55.880 | and you're just assume that it's going to go

00:23:58.280 | to another process and they're gonna test it out.

00:24:01.160 | But I think we need to go back to those early days

00:24:04.400 | when you're a developer, you're developing,

00:24:07.640 | there should be like this, say,

00:24:09.560 | okay, let me look at the ethical outcomes of this

00:24:12.240 | because there isn't a second,

00:24:13.840 | like testing ethical testers, right?

00:24:16.440 | It's you.

00:24:17.280 | We did it back in the early coding days.

00:24:21.240 | I think that's where we are with respect to ethics.

00:24:23.320 | Like, let's go back to what was good practices

00:24:26.360 | only because we were just developing the field.

00:24:30.080 | - Yeah, and it's a really heavy burden.

00:24:34.400 | I've had to feel it recently in the last few months,

00:24:37.560 | but I think it's a good one to feel like.

00:24:39.480 | I've gotten a message, more than one from people.

00:24:42.660 | I've unfortunately gotten some attention recently

00:24:47.480 | and I've gotten messages that say that

00:24:50.480 | I have blood in my hands

00:24:52.360 | because of working on semi-autonomous vehicles.

00:24:56.320 | So the idea that you have semi-autonomy

00:24:58.960 | means people would lose vigilance and so on.

00:25:02.040 | That's actually, be humans, as we describe.

00:25:05.200 | And because of that, because of this idea

00:25:08.160 | that we're creating automation,

00:25:10.080 | there'll be people be hurt because of it.

00:25:12.840 | And I think that's a beautiful thing.

00:25:14.560 | I mean, it's many nights where I wasn't able to sleep

00:25:17.800 | because of this notion.

00:25:19.080 | You really do think about people that might die

00:25:22.400 | because of this technology.

00:25:23.840 | Of course, you can then start rationalizing and saying,

00:25:26.560 | well, you know what?

00:25:27.400 | 40,000 people die in the United States every year

00:25:29.640 | and we're trying to ultimately try to save lives.

00:25:32.400 | But the reality is your code you've written

00:25:35.800 | might kill somebody.

00:25:36.680 | And that's an important burden to carry with you

00:25:38.920 | as you design the code.

00:25:40.200 | - I don't even think of it as a burden

00:25:43.800 | if we train this concept correctly from the beginning.

00:25:47.520 | And I use, and not to say that coding

00:25:49.640 | is like being a medical doctor, but think about it.

00:25:52.400 | Medical doctors, if they've been in situations

00:25:56.080 | where their patient didn't survive, right?

00:25:58.280 | Do they give up and go away?

00:26:00.800 | No, every time they come in,

00:26:02.480 | they know that there might be a possibility

00:26:05.440 | that this patient might not survive.

00:26:07.240 | And so when they approach every decision,

00:26:10.080 | like that's in their back of their head.

00:26:11.920 | And so why isn't that we aren't teaching?

00:26:15.840 | And those are tools though, right?

00:26:17.200 | They are given some of the tools to address that

00:26:19.720 | so that they don't go crazy.

00:26:21.480 | But we don't give those tools

00:26:24.240 | so that it does feel like a burden

00:26:26.160 | versus something of I have a great gift

00:26:28.720 | and I can do great, awesome good,

00:26:31.120 | but with it comes great responsibility.

00:26:33.360 | I mean, that's what we teach in terms of,

00:26:35.840 | if you think about the medical schools, right?

00:26:37.440 | Great gift, great responsibility.

00:26:39.560 | I think if we just change the messaging a little,

00:26:42.160 | great gift, being a developer, great responsibility.

00:26:45.600 | And this is how you combine those.

00:26:48.400 | - But do you think, I mean, this is really interesting.

00:26:51.200 | It's outside, I actually have no friends

00:26:54.320 | who are sort of surgeons or doctors.

00:26:57.300 | I mean, what does it feel like to make a mistake

00:27:01.320 | in a surgery and somebody to die because of that?

00:27:04.800 | Like, is that something you could be taught

00:27:07.040 | in medical school, sort of how to be accepting of that risk?

00:27:10.560 | - So, because I do a lot of work with healthcare robotics,

00:27:14.960 | I have not lost a patient, for example.

00:27:18.480 | The first one's always the hardest, right?

00:27:20.920 | But they really teach the value, right?

00:27:25.920 | So they teach responsibility, but they also teach the value.

00:27:30.840 | Like you're saving 40,000,

00:27:34.920 | but in order to really feel good about that,

00:27:38.320 | when you come to a decision,

00:27:40.160 | you have to be able to say at the end,

00:27:42.280 | I did all that I could possibly do, right?

00:27:45.360 | Versus a, well, I just picked the first widget and did,

00:27:48.840 | right, like, so every decision is actually thought through.

00:27:52.280 | It's not a habit, it's not a,

00:27:53.840 | let me just take the best algorithm

00:27:55.400 | that my friend gave me, right?

00:27:57.160 | It's a, is this it, is this the best?

00:27:59.600 | Have I done my best to do good, right?

00:28:03.200 | And so-- - You're right.

00:28:04.040 | And I think burden is the wrong word.

00:28:06.360 | It's a gift, but you have to treat it extremely seriously.

00:28:10.800 | - Correct.

00:28:12.640 | So on a slightly related note,

00:28:15.520 | in a recent paper,

00:28:16.360 | "The Ugly Truth About Ourselves and Our Robot Creations,"

00:28:20.160 | you discuss, you highlight some biases

00:28:24.320 | that may affect the function of various robotic systems.

00:28:27.120 | Can you talk through, if you remember examples of some?

00:28:30.120 | - There's a lot of examples.

00:28:31.320 | I usually-- - What is bias, first of all?

00:28:33.080 | - Yeah, so bias is this,

00:28:37.080 | and so bias, which is different than prejudice.

00:28:38.840 | So bias is that we all have these preconceived notions

00:28:41.880 | about particular, everything from particular groups

00:28:45.960 | to habits to identity, right?

00:28:49.720 | So we have these predispositions,

00:28:51.400 | and so when we address a problem,

00:28:54.080 | we look at a problem, we make a decision,

00:28:56.040 | those preconceived notions might affect our outputs,

00:29:01.040 | our outcomes.

00:29:02.240 | - So there, the bias could be positive and negative,

00:29:04.680 | and then is prejudice the negative kind of bias?

00:29:07.640 | - Prejudice is the negative, right?

00:29:09.200 | So prejudice is that not only are you aware of your bias,

00:29:13.560 | but you are then take it and have a negative outcome,

00:29:18.560 | even though you are aware.

00:29:20.680 | - And there could be gray areas too.

00:29:23.000 | - There's always gray areas.

00:29:24.680 | - That's the challenging aspect of all ethical questions.

00:29:27.600 | - So I always like, so there's a funny one,

00:29:30.040 | and in fact, I think it might be in the paper,

00:29:31.760 | 'cause I think I talk about self-driving cars,

00:29:34.200 | but think about this.

00:29:37.160 | For teenagers, right, typically,

00:29:40.600 | insurance companies charge quite a bit of money

00:29:44.560 | if you have a teenage driver.

00:29:46.760 | So you could say that's an age bias, right?

00:29:50.880 | But no one will, I mean, parents will be grumpy,

00:29:54.080 | but no one really says that that's not fair.

00:29:58.640 | - That's interesting.

00:29:59.480 | We don't, that's right, that's right.

00:30:01.720 | Everybody in human factors and safety research

00:30:07.120 | almost, I mean, is quite ruthlessly critical of teenagers.

00:30:12.120 | And we don't question, is that okay?

00:30:15.040 | Is that okay to be ageist in this kind of way?

00:30:17.160 | - It is, and it is ageist, right?

00:30:18.600 | It's definitely age, there's no question about it.

00:30:20.800 | And so this is the gray area, right?

00:30:24.960 | 'Cause you know that teenagers are more likely

00:30:29.840 | to be in accidents, and so there's actually some data to it.

00:30:33.080 | But then if you take that same example,

00:30:35.000 | and you say, well, I'm going to make the insurance higher

00:30:39.400 | for an area of Boston, because there's a lot of accidents.

00:30:44.400 | And then they find out that that's correlated

00:30:48.280 | with socioeconomics.

00:30:50.240 | Well, then it becomes a problem, right?

00:30:52.440 | Like that is not acceptable, but yet the teenager,

00:30:56.720 | which is age, it's against age, is, right?

00:31:01.720 | - And the way we figure that out as a society

00:31:03.960 | by having conversations, by having discourse,

00:31:06.200 | I mean, throughout history, the definition

00:31:08.360 | of what is ethical and not has changed,

00:31:11.360 | and hopefully always for the better.

00:31:14.360 | - Correct, correct.

00:31:15.480 | - So in terms of bias or prejudice in algorithms,

00:31:20.480 | what examples do you sometimes think about?

00:31:25.560 | - So I think about quite a bit the medical domain,

00:31:28.960 | just because historically, right?

00:31:31.320 | The healthcare domain has had these biases,

00:31:34.520 | typically based on gender and ethnicity, primarily,

00:31:39.520 | a little in age, but not so much.

00:31:42.320 | Historically, if you think about FDA and drug trials,

00:31:48.680 | it's harder to find women that aren't childbearing,

00:31:54.240 | and so you may not test on drugs at the same level, right?

00:31:57.160 | So there's these things.

00:31:59.320 | And so if you think about robotics, right?

00:32:02.920 | Something as simple as,

00:32:04.880 | I'd like to design an exoskeleton, right?

00:32:07.760 | What should the material be?

00:32:09.200 | What should the weight be?

00:32:10.160 | What should the form factor be?

00:32:12.560 | Who are you gonna design it around?

00:32:16.960 | I will say that in the US, women average height

00:32:21.120 | and weight is slightly different than guys.

00:32:23.440 | So who are you gonna choose?

00:32:25.880 | Like, if you're not thinking about it from the beginning,

00:32:29.080 | as, you know, okay, when I design this,

00:32:32.080 | and I look at the algorithms,

00:32:33.640 | and I design the control system,

00:32:34.960 | and the forces, and the torques,

00:32:37.080 | if you're not thinking about,

00:32:38.280 | well, you have different types of body structure,

00:32:41.720 | you're gonna design to what you're used to.

00:32:44.600 | Oh, this fits all the folks in my lab, right?

00:32:48.240 | - So think about it from the very beginning is important.

00:32:51.520 | What about sort of algorithms

00:32:52.920 | that train on data kind of thing?

00:32:56.120 | Sadly, our society already has a lot of negative bias.

00:33:01.120 | And so if we collect a lot of data,

00:33:03.360 | even if it's a balanced way,

00:33:06.320 | it's going to contain the same bias that a society contains.

00:33:09.040 | And so, yeah, is there things there that bother you?

00:33:13.760 | - Yeah, so you actually said something.

00:33:15.640 | You had said how we have biases,

00:33:19.960 | but hopefully we learn from them,

00:33:21.520 | and we become better, right?

00:33:23.120 | And so that's where we are now, right?

00:33:25.160 | So the data that we're collecting is historic.

00:33:28.640 | So it's based on these things

00:33:30.160 | when we knew it was bad to discriminate,

00:33:32.640 | but that's the data we have,

00:33:33.960 | and we're trying to fix it now,

00:33:36.120 | but we're fixing it based on the data

00:33:37.880 | that was used in the first place to discriminate.

00:33:39.760 | - Fix it in post.

00:33:40.680 | - Right, and so the decisions,

00:33:43.800 | and you can look at everything

00:33:45.000 | from the whole aspect of predictive policing,

00:33:49.320 | criminal recidivism.

00:33:51.440 | There was a recent paper that had the healthcare algorithms,

00:33:54.360 | which had a kind of a sensational titles.

00:33:58.240 | I'm not pro sensationalism in titles,

00:34:01.240 | but again, you read it, right?

00:34:03.760 | So it makes you read it,

00:34:05.800 | but I'm like, really?

00:34:07.040 | Like, ah, you could have--

00:34:08.960 | - What's the topic of the sensationalism?

00:34:10.960 | I mean, what's underneath it?

00:34:12.680 | If you could sort of educate me

00:34:16.320 | on what kind of bias creeps into the healthcare space.

00:34:19.200 | - Yeah, so--

00:34:20.040 | - I mean, you already kind of mentioned.

00:34:21.440 | - Yeah, so this one was,

00:34:23.120 | the headline was racist AI algorithms.

00:34:27.560 | Okay, like, okay, that's totally a clickbait title.

00:34:30.920 | And so you looked at it,

00:34:32.200 | and so there was data that these researchers had collected.

00:34:36.720 | I believe, I wanna say it was either science or nature.

00:34:39.480 | It was just published,

00:34:40.720 | but they didn't have the sensational title.

00:34:42.680 | It was like the media.

00:34:44.960 | And so they had looked at demographics,

00:34:47.560 | I believe, between black and white women, right?

00:34:52.240 | And they showed that there was a discrepancy

00:34:56.920 | in the outcomes, right?

00:34:59.280 | And so, and it was tied to ethnicity, tied to race.

00:35:02.480 | The piece that the researchers did

00:35:04.880 | actually went through the whole analysis,

00:35:07.480 | but of course--

00:35:08.880 | - I mean, the journalists with AI

00:35:11.120 | are problematic across the board, let's say.

00:35:14.320 | - And so this is a problem, right?

00:35:16.240 | And so there's this thing about,

00:35:18.280 | oh, AI, it has all these problems.

00:35:20.600 | We're doing it on historical data,

00:35:22.920 | and the outcomes aren't even based on gender

00:35:26.040 | or ethnicity or age.

00:35:28.120 | But I'm always saying is like,

00:35:30.280 | yes, we need to do better, right?

00:35:32.520 | We need to do better.

00:35:33.640 | It is our duty to do better,

00:35:35.400 | but the worst AI is still better than us.

00:35:39.880 | Like, you take the best of us,

00:35:41.960 | and we're still worse than the worst AI,

00:35:44.200 | at least in terms of these things.

00:35:45.680 | And that's actually not discussed, right?

00:35:48.040 | And so I think, and that's why the sensational title,

00:35:51.680 | right, and so it's like,

00:35:53.080 | so then you can have individuals go like,

00:35:54.400 | oh, we don't need to use this AI.

00:35:55.600 | I'm like, oh, no, no, no, no.

00:35:56.840 | I want the AI instead of the doctors

00:36:01.040 | that provided that data,

00:36:02.080 | 'cause it's still better than that, right?

00:36:04.280 | - I think that's really important to linger on.

00:36:06.840 | Is the idea that this AI is racist,

00:36:09.640 | it's like, well, compared to what?

00:36:14.320 | (laughs)

00:36:16.400 | I think we set, unfortunately,

00:36:20.160 | way too high of a bar for AI algorithms.

00:36:23.320 | And in the ethical space where perfect is,

00:36:26.040 | I would argue, probably impossible,

00:36:28.160 | then if we set the bar of perfection, essentially,

00:36:33.080 | of it has to be perfectly fair, whatever that means,

00:36:36.280 | it means we're setting it up for failure.

00:36:39.640 | But that's really important to say what you just said,

00:36:42.040 | which is, well, it's still better than some things.

00:36:45.000 | - And one of the things I think

00:36:46.960 | that we don't get enough credit for,

00:36:49.440 | just in terms of as developers,

00:36:52.240 | is that you can now poke at it, right?

00:36:55.920 | So it's harder to say, you know, is this hospital,

00:36:58.920 | is this city doing something, right,

00:37:01.120 | until someone brings in a civil case, right?

00:37:04.480 | Well, with AI, it can process through all this data

00:37:07.160 | and say, hey, yes, there was an issue here,

00:37:12.160 | but here it is, we've identified it,

00:37:14.560 | and then the next step is to fix it.

00:37:16.240 | I mean, that's a nice feedback loop

00:37:18.160 | versus like waiting for someone to sue someone else

00:37:21.400 | before it's fixed, right?

00:37:22.840 | And so I think that power,

00:37:25.160 | we need to capitalize on a little bit more, right?

00:37:27.680 | Instead of having the sensational titles,

00:37:29.760 | have the, okay, this is a problem,

00:37:33.400 | and this is how we're fixing it,

00:37:34.640 | and people are putting money to fix it

00:37:36.600 | because we can make it better.

00:37:38.680 | I look at like facial recognition, how Joy,

00:37:43.040 | she basically called out a couple of companies and said,

00:37:45.880 | hey, and most of them were like, oh, embarrassment,

00:37:50.560 | and the next time it had been fixed, right?

00:37:53.400 | It had been fixed better, right?

00:37:54.960 | And then it was like, oh, here's some more issues.

00:37:56.880 | And I think that conversation then moves that needle

00:38:01.880 | to having much more fair and unbiased and ethical aspects,

00:38:06.880 | as long as both sides, the developers are willing to say,

00:38:10.640 | okay, I hear you, yes, we are going to improve,

00:38:14.080 | and you have other developers who are like,

00:38:16.160 | hey, AI, it's wrong, but I love it, right?

00:38:19.720 | - Yes, so speaking of this really nice notion

00:38:23.080 | that AI is maybe flawed but better than humans,

00:38:27.080 | so just made me think of it,

00:38:29.240 | one example of flawed humans is our political system.

00:38:34.160 | Do you think, or you said judicial as well,

00:38:38.760 | do you have a hope for AI sort of being elected

00:38:43.760 | for president or running our Congress

00:38:49.840 | or being able to be a powerful representative of the people?

00:38:54.040 | - So I mentioned, and I truly believe

00:38:57.160 | that this whole world of AI is in partnerships with people.

00:39:01.480 | And so what does that mean?

00:39:02.680 | I don't believe, or maybe I just don't,

00:39:07.680 | I don't believe that we should have an AI for president,

00:39:11.640 | but I do believe that a president

00:39:13.760 | should use AI as an advisor, right?

00:39:16.120 | Like if you think about it,

00:39:17.640 | every president has a cabinet of individuals

00:39:22.120 | that have different expertise

00:39:23.880 | that they should listen to, right?

00:39:26.280 | That's kind of what we do.

00:39:28.040 | And you put smart people with smart expertise

00:39:31.160 | around certain issues and you listen.

00:39:33.520 | I don't see why AI can't function

00:39:35.760 | as one of those smart individuals giving input.

00:39:39.320 | So maybe there's an AI on healthcare,

00:39:41.080 | maybe there's an AI on education and right?

00:39:43.880 | Like all of these things that a human is processing, right?

00:39:48.800 | Because at the end of the day,

00:39:50.480 | there's people that are human

00:39:53.600 | that are going to be at the end of the decision.

00:39:55.560 | And I don't think as a world, as a culture, as a society,

00:39:59.360 | that we would totally believe, and this is us,

00:40:03.080 | like this is some fallacy about us,

00:40:05.360 | but we need to see that leader, that person as human.

00:40:10.360 | And most people don't realize that like leaders

00:40:15.480 | have a whole lot of advice, right?

00:40:17.040 | Like when they say something,

00:40:18.240 | it's not that they woke up, well, usually,

00:40:20.480 | they don't wake up in the morning and be like,

00:40:22.960 | I have a brilliant idea, right?

00:40:24.440 | It's usually a, okay, let me listen.

00:40:26.840 | I have a brilliant idea,

00:40:27.680 | but let me get a little bit of feedback on this.

00:40:30.000 | Like, okay.

00:40:31.120 | And then it's a, yeah, that was an awesome idea.

00:40:33.240 | Or it's like, yeah, let me go back.

00:40:36.000 | - We already talked through a bunch of them,

00:40:37.520 | but are there some possible solutions

00:40:41.600 | to the bias that's present in our algorithms

00:40:45.280 | beyond what we just talked about?

00:40:46.760 | - So I think there's two paths.

00:40:49.400 | One is to figure out how to systematically

00:40:53.800 | do the feedback and corrections.

00:40:56.600 | So right now it's ad hoc, right?

00:40:58.160 | It's a researcher identifies some outcomes

00:41:02.480 | that are not, don't seem to be fair, right?

00:41:05.440 | They publish it, they write about it,

00:41:07.960 | and either the developer or the companies

00:41:11.480 | that have adopted the algorithms may try to fix it, right?

00:41:14.280 | And so it's really ad hoc and it's not systematic.

00:41:18.880 | There's, it's just, it's kind of like, I'm a researcher.

00:41:22.480 | That seems like an interesting problem,

00:41:24.760 | which means that there's a whole lot out there

00:41:26.600 | that's not being looked at, right?

00:41:29.160 | 'Cause it's kind of researcher driven.

00:41:31.080 | And I don't necessarily have a solution,

00:41:35.720 | but that process I think could be done a little bit better.

00:41:40.720 | One way is I'm going to poke a little bit

00:41:45.080 | at some of the corporations, right?

00:41:48.320 | Like maybe the corporations, when they think about a product

00:41:51.760 | they should, instead of, in addition to hiring these bug,

00:41:56.760 | they give these--

00:41:59.120 | - Oh yeah, yeah, yeah.

00:42:01.280 | Like awards when you find a bug.

00:42:03.000 | - Yeah, security bug.

00:42:05.840 | Let's put it like, we will give the,

00:42:08.600 | whatever the award is that we give

00:42:10.320 | for the people who find these security holes,

00:42:12.680 | find an ethics hole, right?

00:42:14.040 | Like find an unfairness hole

00:42:15.440 | and we will pay you X for each one you find.

00:42:17.920 | I mean, why can't they do that?

00:42:19.680 | One is a win-win.

00:42:21.120 | They show that they're concerned about it,

00:42:23.160 | that this is important,

00:42:24.240 | and they don't have to necessarily dedicate

00:42:26.440 | their own internal resources.

00:42:28.880 | And it also means that everyone who has

00:42:31.240 | their own bias lens, like I'm interested in age

00:42:34.680 | and so I'll find the ones based on age

00:42:36.640 | and I'm interested in gender, right?

00:42:38.480 | Which means that you get

00:42:39.840 | all of these different perspectives.

00:42:41.640 | - But you think of it in a data driven way.

00:42:43.440 | So like, sort of, if we look at a company like Twitter,

00:42:48.800 | it gets, it's under a lot of fire

00:42:51.680 | for discriminating against certain political beliefs.

00:42:54.800 | - Correct.

00:42:55.880 | - And sort of, there's a lot of people,

00:42:58.080 | this is the sad thing,

00:42:59.240 | 'cause I know how hard the problem is

00:43:00.720 | and I know the Twitter folks are working really hard at it.

00:43:03.080 | Even Facebook, that everyone seems to hate,

00:43:04.960 | are working really hard at this.

00:43:06.880 | You know, the kind of evidence that people bring

00:43:09.320 | is basically anecdotal evidence.

00:43:11.240 | Well, me or my friend, all we said is X,

00:43:15.000 | and for that we got banned.

00:43:17.120 | And that's kind of a discussion of saying,

00:43:20.960 | well, look, that's usually, first of all,

00:43:23.280 | the whole thing is taken out of context.

00:43:25.480 | So they present sort of anecdotal evidence.

00:43:28.640 | And how are you supposed to, as a company,

00:43:31.120 | in a healthy way, have a discourse

00:43:33.080 | about what is and isn't ethical?

00:43:36.000 | How do we make algorithms ethical

00:43:38.040 | when people are just blowing everything?

00:43:40.760 | Like, they're outraged about a particular

00:43:45.120 | anecdotal piece of evidence

00:43:47.200 | that's very difficult to sort of contextualize

00:43:49.440 | in a big data-driven way.

00:43:51.620 | Do you have a hope for companies like Twitter and Facebook?

00:43:55.760 | - Yeah, so I think there's a couple of things going on.

00:43:59.760 | First off, remember this whole aspect

00:44:04.760 | of we are becoming reliant on technology.

00:44:09.360 | We're also becoming reliant on a lot of these,

00:44:14.360 | the apps and the resources that are provided.

00:44:18.000 | So some of it is kind of anger, like, I need you, right?

00:44:21.640 | And you're not working for me, right?

00:44:23.320 | - Yeah, not working for me, right.

00:44:24.680 | - But I think, and so some of it,

00:44:27.280 | and I wish that there was a little bit

00:44:31.400 | of change of rethinking.

00:44:32.840 | So some of it is like, oh, we'll fix it in-house.

00:44:35.560 | No, that's like, okay, I'm a fox,

00:44:39.000 | and I'm going to watch these hens

00:44:40.920 | because I think it's a problem that foxes eat hens.

00:44:44.080 | No, right?

00:44:45.200 | Like, use, like, be good citizens and say,

00:44:48.640 | look, we have a problem,

00:44:49.920 | and we are willing to open ourselves up

00:44:54.840 | for others to come in and look at it

00:44:57.080 | and not try to fix it in-house.

00:44:58.760 | Because if you fix it in-house,

00:45:00.500 | there's conflict of interest.

00:45:02.000 | If I find something, I'm probably going to want to fix it,

00:45:04.480 | and hopefully the media won't pick it up, right?

00:45:07.360 | And that then causes distrust

00:45:09.360 | because someone inside is going to be mad at you

00:45:11.920 | and go out and talk about how,

00:45:13.640 | yeah, they can the resume survey because, right?

00:45:17.800 | Like, be best people.

00:45:19.360 | Like, just say, look, we have this issue.

00:45:22.800 | Community, help us fix it,

00:45:24.480 | and we will give you, like, you know,

00:45:25.840 | the bug finder fee if you do.

00:45:28.160 | - Did you ever hope that the community,

00:45:31.320 | us as a human civilization on the whole is good

00:45:35.360 | and can be trusted to guide the future of our civilization

00:45:39.560 | into a positive direction?

00:45:41.000 | - I think so.

00:45:41.920 | So I'm an optimist, right?

00:45:44.160 | And, you know, there were some dark times in history,

00:45:48.520 | always.

00:45:50.040 | I think now we're in one of those dark times.

00:45:52.960 | I truly do.

00:45:53.800 | - In which aspect?

00:45:54.680 | - The polarization.

00:45:56.320 | And it's not just US, right?

00:45:57.600 | So if it was just US, I'd be like, yeah, it's a US thing,

00:46:00.080 | but we're seeing it, like, worldwide, this polarization.

00:46:03.520 | And so I worry about that.

00:46:06.580 | But I do fundamentally believe

00:46:10.240 | that at the end of the day, people are good, right?

00:46:13.480 | And why do I say that?

00:46:14.840 | Because anytime there's a scenario

00:46:17.760 | where people are in danger, and I will use,

00:46:20.880 | so Atlanta, we had a snowmageddon,

00:46:24.320 | and people can laugh about that.

00:46:26.680 | People at the time, so the city closed for, you know,

00:46:30.520 | little snow, but it was ice, and the city closed down.

00:46:33.480 | But you had people opening up their homes and saying,

00:46:35.800 | hey, you have nowhere to go, come to my house, right?

00:46:39.120 | Hotels were just saying, like, sleep on the floor.

00:46:41.840 | Like, places like, you know, the grocery stores were like,

00:46:44.440 | hey, here's food.

00:46:45.960 | There was no like, oh, how much are you gonna pay me?

00:46:47.960 | It was like this, such a community.

00:46:50.520 | And like, people who didn't know each other,

00:46:52.160 | strangers were just like, can I give you a ride home?

00:46:55.560 | And that was a point I was like, you know what?

00:46:57.920 | - That reveals that the deeper thing is,

00:47:03.120 | there's a compassionate love that we all have within us.

00:47:06.960 | It's just that when all of that is taken care of

00:47:09.480 | and get bored, we love drama.

00:47:11.280 | And that's, I think almost like the division

00:47:14.800 | is a sign of the times being good,

00:47:17.040 | is that it's just entertaining

00:47:19.000 | on some unpleasant mammalian level

00:47:22.840 | to watch, to disagree with others.

00:47:26.120 | And Twitter and Facebook are actually

00:47:28.120 | taking advantage of that in a sense

00:47:31.360 | because it brings you back to the platform,

00:47:34.040 | and they're advertiser-driven,

00:47:36.120 | so they make a lot of money.

00:47:37.680 | - So you go back and you click.

00:47:39.240 | - Love doesn't sell quite as well

00:47:41.240 | in terms of advertisement.

00:47:42.680 | - It doesn't.

00:47:44.920 | - So you've started your career

00:47:46.960 | at NASA Jet Propulsion Laboratory,

00:47:49.080 | but before I ask a few questions there,

00:47:51.960 | have you happened to have ever seen Space Odyssey,

00:47:54.440 | 2001 Space Odyssey?

00:47:55.840 | - Yes.

00:47:58.040 | - Okay, do you think Hal 9000,

00:48:01.440 | so we're talking about ethics,

00:48:03.440 | do you think Hal did the right thing

00:48:06.720 | by taking the priority of the mission

00:48:08.600 | over the lives of the astronauts?

00:48:10.280 | Do you think Hal is good or evil?

00:48:12.420 | Easy questions.

00:48:16.960 | - Yeah.

00:48:17.800 | Hal was misguided.

00:48:21.440 | - You're one of the people that would be

00:48:23.440 | in charge of an algorithm like Hal.

00:48:26.160 | - Yeah.

00:48:27.000 | - So how would you do better?

00:48:28.360 | - If you think about what happened was

00:48:32.280 | there was no fail safe, right?

00:48:35.400 | So perfection, right?

00:48:37.800 | Like what is that?

00:48:38.640 | I'm gonna make something that I think is perfect,

00:48:40.880 | but if my assumptions are wrong,

00:48:44.640 | it'll be perfect based on the wrong assumptions, right?

00:48:47.600 | That's something that you don't know until you deploy

00:48:51.760 | and then you're like, oh yeah, messed up.

00:48:53.880 | But what that means is that when we design software,

00:48:58.400 | such as in Space Odyssey,

00:49:00.720 | when we put things out,

00:49:02.200 | that there has to be a fail safe.

00:49:04.120 | There has to be the ability that once it's out there,

00:49:07.800 | we can grade it as an F and it fails

00:49:11.440 | and it doesn't continue, right?

00:49:13.160 | There's some way that it can be brought in

00:49:16.120 | and removed and that's aspect.

00:49:19.720 | Because that's what happened with Hal.

00:49:21.160 | It was like assumptions were wrong.

00:49:23.840 | It was perfectly correct based on those assumptions.

00:49:27.920 | And there was no way to change it,

00:49:31.080 | change the assumptions at all.

00:49:34.080 | - And the change, the fallback would be to a human.

00:49:37.080 | So you ultimately think like human should be,

00:49:40.080 | it's not turtles or AI all the way down.

00:49:45.640 | It's at some point there's a human

00:49:47.280 | that actually makes a decision.

00:49:48.120 | - I still think that, and again,

00:49:49.840 | because I do human robot interaction,

00:49:51.440 | I still think the human needs to be part of the equation

00:49:55.040 | at some point.

00:49:56.480 | - So what, just looking back,

00:49:58.480 | what are some fascinating things in robotic space

00:50:01.960 | that NASA was working at the time?

00:50:03.520 | Or just in general, what have you gotten to play with

00:50:07.720 | and what are your memories from working at NASA?

00:50:10.080 | - Yeah, so one of my first memories

00:50:12.600 | was they were working on a surgical robot system

00:50:18.640 | that could do eye surgery, right?

00:50:21.880 | And this was back in, oh my gosh, it must've been,

00:50:25.680 | oh, maybe '92, '93, '94.

00:50:30.560 | - So it's like almost like a remote operation.

00:50:32.840 | - Yeah, it was remote operation.

00:50:34.480 | And in fact, you can even find some old tech reports on it.

00:50:38.360 | So think of it, like now we have Da Vinci, right?

00:50:41.600 | Like think of it, but these were like the late '90s, right?

00:50:45.840 | And I remember going into the lab one day

00:50:48.200 | and I was like, what's that, right?

00:50:50.960 | And of course it wasn't pretty, right?

00:50:53.880 | 'Cause the technology, but it was like functional

00:50:56.600 | and you had this individual that could use

00:50:59.200 | a version of haptics to actually do the surgery.

00:51:01.920 | And they had this mock-up of a human face

00:51:04.320 | and like the eyeballs and you can see this little drill.

00:51:08.400 | And I was like, oh, that is so cool.

00:51:11.640 | That one I vividly remember

00:51:13.680 | because it was so outside of my like possible thoughts

00:51:18.600 | of what could be done.

00:51:20.000 | - Just the kind of precision.

00:51:21.320 | And I mean, what's the most amazing of a thing like that?

00:51:26.080 | - I think it was the precision.

00:51:28.200 | It was the kind of first time that I had physically seen

00:51:33.200 | this robot machine human interface, right?

00:51:39.600 | Versus, 'cause manufacturing had been,

00:51:42.360 | you saw those kinds of big robots, right?

00:51:44.520 | But this was like, oh, this is in a person.

00:51:48.040 | There's a person and a robot like in the same space.

00:51:51.400 | - The meeting them in person.

00:51:53.040 | Like for me, it was a magical moment that I can't,

00:51:56.040 | as life transforming that I recently met Spot Mini

00:51:59.640 | from Boston Dynamics.

00:52:01.320 | I don't know why, but on the human robot interaction,

00:52:04.720 | for some reason I realized how easy it is

00:52:07.800 | to anthropomorphize.

00:52:09.800 | And it was, I don't know, it was almost like falling in love

00:52:13.520 | this feeling of meeting.

00:52:14.840 | And I've obviously seen these robots a lot

00:52:17.440 | in video and so on, but meeting in person,

00:52:19.280 | just having that one-on-one time.

00:52:21.000 | - It's different. - It's different.

00:52:22.440 | So have you had a robot like that in your life

00:52:25.160 | that made you maybe fall in love with robotics?

00:52:28.400 | Sort of like meeting in person?

00:52:30.600 | - I mean, I loved robotics.

00:52:35.040 | - From the beginning. - Yeah, so.

00:52:37.000 | I was a 12-year-old, like I'm gonna be a roboticist.

00:52:39.600 | Actually, I called it cybernetics.

00:52:41.280 | But so my motivation was bionic woman.

00:52:44.800 | I don't know if you know that.

00:52:46.480 | And so, I mean, that was like a seminal moment,

00:52:49.600 | but I didn't meet, like that was TV, right?

00:52:52.440 | Like it wasn't like I was in the same space and I met,

00:52:54.640 | I was like, oh my gosh, you're like real.

00:52:56.640 | - Just linking on bionic woman, which by the way,

00:52:58.920 | because I read that about you, I watched a bit of it

00:53:03.320 | and it's just so, no offense, terrible.

00:53:05.640 | (laughing)

00:53:06.480 | - It's cheesy. - It's cheesy.

00:53:07.880 | - If you look at it now. - It's cheesy.

00:53:08.960 | - I've seen a couple of reruns lately.

00:53:10.800 | (laughing)

00:53:11.640 | - But it's, but of course at the time,

00:53:14.280 | it was probably-- - But the sound effects.

00:53:16.120 | - It was probably the sound of my imagination.

00:53:16.960 | (laughing)

00:53:19.200 | Especially when you're younger, it just catches you.

00:53:23.160 | But which aspect, did you think of it,

00:53:24.760 | you mentioned cybernetics, did you think of it as robotics

00:53:27.760 | or did you think of it as almost

00:53:29.440 | constructing artificial beings?

00:53:31.680 | Like is it the intelligent part

00:53:34.240 | that captured your fascination

00:53:37.040 | or was it the whole thing, like even just the limbs

00:53:39.400 | and just the-- - So for me,

00:53:40.520 | it would have, in another world,

00:53:42.960 | I probably would have been more of a biomedical engineer

00:53:46.840 | because what fascinated me was the parts,

00:53:50.040 | like the bionic parts, the limbs, those aspects of it.

00:53:55.040 | - Are you especially drawn to humanoid or human-like robots?

00:53:59.640 | - I would say human-like, not humanoid, right?

00:54:03.040 | And when I say human-like,

00:54:04.200 | I think it's this aspect of that interaction,

00:54:07.800 | whether it's social and it's like a dog, right?

00:54:10.680 | Like that's human-like because it understands us,

00:54:14.120 | it interacts with us at that very social level.

00:54:17.680 | Humanoids are part of that,

00:54:21.860 | but only if they interact with us as if we are human.

00:54:26.860 | - But just to linger on NASA for a little bit,

00:54:30.920 | what do you think, maybe if you have other memories,

00:54:34.080 | but also what do you think is the future of robots in space?

00:54:38.560 | We mentioned how, but there's incredible robots

00:54:41.880 | that NASA's working on in general,

00:54:43.400 | thinking about in our, as we venture out,

00:54:48.160 | human civilization ventures out into space.

00:54:50.440 | What do you think the future of robots is there?

00:54:52.240 | - Yeah, so I mean, there's the near term.

00:54:53.680 | For example, they just announced the rover

00:54:57.280 | that's going to the moon, which, you know,

00:55:00.720 | that's kind of exciting, but that's like near term.

00:55:06.040 | You know, my favorite, favorite, favorite series

00:55:11.040 | is "Star Trek," right?

00:55:13.300 | You know, I really hope, and even "Star Trek,"

00:55:17.160 | like if I calculate the years, I wouldn't be alive,

00:55:20.060 | but I would really, really love to be in that world.

00:55:25.060 | Like even if it's just at the beginning,

00:55:28.440 | like, you know, like "Voyage," like "Adventure 1."

00:55:33.160 | - So basically living in space.

00:55:35.720 | - Yeah.

00:55:36.560 | - With what robots, what do robots--

00:55:39.760 | - With data.

00:55:40.600 | - What role--

00:55:41.420 | - The data would have to be, even though that wasn't,

00:55:42.840 | you know, that was like later, but--

00:55:44.760 | - So data is a robot that has human-like qualities.

00:55:49.160 | - Right, without the emotion ship, yeah.

00:55:51.080 | - You don't like emotion in your robots.

00:55:52.200 | - Well, so data with the emotion ship

00:55:54.240 | was kind of a mess, right?

00:55:58.440 | It took a while for that, him to adapt,

00:56:04.640 | but, and so why was that an issue?

00:56:08.600 | The issue is, is that emotions make us irrational agents.

00:56:13.600 | That's the problem.

00:56:15.200 | And yet he could think through things,

00:56:20.040 | even if it was based on an emotional scenario, right?

00:56:23.440 | Based on pros and cons.

00:56:25.100 | But as soon as you made him emotional,

00:56:28.520 | one of the metrics he used for evaluation

00:56:31.160 | was his own emotions.

00:56:33.280 | Not people around him, right?

00:56:35.480 | Like, and so--

00:56:37.280 | - We do that as children, right?

00:56:39.000 | So we're very egocentric when we're young.

00:56:40.920 | - We are very egocentric.

00:56:42.320 | - And so isn't that just an early version

00:56:44.900 | of the emotion ship then?

00:56:46.400 | I haven't watched much "Star Trek."

00:56:48.240 | - Except I have also met adults.

00:56:50.740 | Right, and so that is a developmental process,

00:56:54.600 | and I'm sure there's a bunch of psychologists

00:56:57.600 | that could go through, like you can have a 60-year-old adult

00:57:00.640 | who has the emotional maturity of a 10-year-old, right?

00:57:04.640 | And so there's various phases that people should go through

00:57:08.880 | in order to evolve, and sometimes you don't.

00:57:11.460 | - So how much psychology do you think,

00:57:14.840 | a topic that's rarely mentioned in robotics,

00:57:17.600 | but how much does psychology come to play

00:57:19.720 | when you're talking about HRI, human-robot interaction,

00:57:23.600 | when you have to have robots

00:57:25.000 | that actually interact with humans?

00:57:26.120 | - Tons.

00:57:27.800 | Like my group, as well as I,

00:57:30.400 | read a lot in the cognitive science literature

00:57:33.320 | as well as the psychology literature,

00:57:36.200 | because they understand a lot about human-human relations

00:57:41.200 | and developmental milestones and things like that.

00:57:45.960 | And so we tend to look to see what's been done out there.

00:57:50.960 | Sometimes what we'll do is we'll try to match that

00:57:55.640 | to see is that human-human relationship

00:57:58.440 | the same as human-robot?

00:58:01.020 | Sometimes it is, and sometimes it's different.

00:58:03.100 | And then when it's different, we try to figure out,

00:58:05.880 | okay, why is it different in this scenario,

00:58:09.060 | but it's the same in the other scenario, right?

00:58:11.920 | And so we try to do that quite a bit.

00:58:15.360 | - Would you say that's, if we're looking at the future

00:58:17.840 | of human-robot interaction,

00:58:19.160 | would you say the psychology piece is the hardest?

00:58:23.120 | I mean, it's a funny notion for you as,

00:58:25.440 | I don't know if you consider, yeah.

00:58:27.400 | I mean, one way to ask it,

00:58:28.440 | do you consider yourself a roboticist or a psychologist?

00:58:32.040 | - Oh, I consider myself a roboticist

00:58:33.640 | that plays the act of a psychologist.

00:58:36.280 | - But if you were to look at yourself

00:58:38.160 | sort of 20, 30 years from now,

00:58:42.400 | do you see yourself more and more wearing the psychology hat?

00:58:45.400 | Another way to put it is,

00:58:49.040 | are the hard problems in human-robot interactions

00:58:51.640 | fundamentally psychology, or is it still robotics,

00:58:55.840 | the perception, manipulation, planning,

00:58:57.760 | all that kind of stuff?

00:58:59.500 | - It's actually neither.

00:59:01.720 | The hardest part is the adaptation and the interaction.

00:59:05.200 | So-- - The learning.

00:59:07.200 | - It's the interface, it's the learning.

00:59:08.880 | And so if I think of,

00:59:10.820 | I've become much more of a roboticist/AI person

00:59:16.880 | than when I, like originally, again,

00:59:19.080 | I was about the bionics.

00:59:20.200 | I was electrical engineer, I was control theory, right?

00:59:24.080 | And then I started realizing that my algorithms

00:59:28.840 | needed human data, right?

00:59:30.640 | And so then I was like, okay, what is this human thing?

00:59:32.800 | How do I incorporate human data?

00:59:34.400 | And then I realized that human perception had,

00:59:38.720 | there was a lot in terms of how we perceive the world,

00:59:41.080 | and so trying to figure out

00:59:41.960 | how do I model human perception for my,

00:59:44.440 | and so I became a HRI person,

00:59:47.600 | human-robot interaction person,

00:59:49.360 | from being a control theory

00:59:51.200 | and realizing that humans actually offered quite a bit.

00:59:54.340 | And then when you do that,

00:59:56.080 | you become more of an artificial intelligence, AI,

00:59:59.320 | and so I see myself evolving more in this AI world

01:00:04.320 | under the lens of robotics,

01:00:09.600 | having hardware, interacting with people.

01:00:12.140 | - So you're a world-class expert researcher in robotics,

01:00:17.880 | and yet others, there's a few,

01:00:21.160 | it's a small but fierce community of people,

01:00:24.200 | but most of them don't take the journey

01:00:26.640 | into the H of HRI, into the human.

01:00:29.440 | So why did you brave into the interaction with humans?

01:00:34.440 | It seems like a really hard problem.

01:00:36.880 | - It's a hard problem, and it's very risky as an academic.

01:00:39.880 | - Yes.

01:00:41.120 | - And I knew that when I started down that journey,

01:00:46.240 | that it was very risky as an academic

01:00:49.960 | in this world that was nuanced, it was just developing.

01:00:53.520 | We didn't even have a conference, right, at the time.

01:00:56.800 | Because it was the interesting problems.

01:01:00.160 | That was what drove me.

01:01:01.600 | It was the fact that I looked at what interests me

01:01:06.600 | in terms of the application space and the problems,

01:01:10.440 | and that pushed me into trying to figure out

01:01:14.960 | what people were and what humans were

01:01:16.840 | and how to adapt to them.

01:01:18.180 | If those problems weren't so interesting,

01:01:21.280 | I'd probably still be sending rovers to glaciers, right?

01:01:26.320 | But the problems were interesting.

01:01:28.120 | And the other thing was that they were hard, right?

01:01:30.640 | So I like having to go into a room and being like,

01:01:35.640 | "I don't know what to do."

01:01:37.040 | And then going back and saying,

01:01:38.020 | "Okay, I'm gonna figure this out."

01:01:39.820 | I'm not driven when I go in like,

01:01:42.280 | "Oh, there are no surprises."

01:01:44.080 | Like, I don't find that satisfying.

01:01:47.360 | If that was the case,

01:01:48.200 | I'd go someplace and make a lot more money, right?

01:01:51.060 | I think I stay an academic and choose to do this

01:01:55.020 | because I can go into a room and like, "That's hard."

01:01:58.320 | - Yeah, I think just from my perspective,

01:02:01.760 | maybe you can correct me on it,

01:02:03.240 | but if I just look at the field of AI broadly,

01:02:06.760 | it seems that human-robot interaction

01:02:10.080 | has one of the most number of open problems.

01:02:15.080 | People, especially relative to how many people

01:02:20.320 | are willing to acknowledge that there are.

01:02:22.480 | Because most people are just afraid of the humans,

01:02:26.160 | so they don't even acknowledge

01:02:27.240 | how many open problems there are.

01:02:28.200 | But in terms of difficult problems to solve,

01:02:30.880 | exciting spaces, it seems to be incredible for that.

01:02:35.800 | - It is, and it's exciting.

01:02:38.720 | - You've mentioned trust before.

01:02:40.020 | What role does trust,

01:02:43.320 | from interacting with autopilot

01:02:46.840 | to in the medical context,

01:02:48.440 | what role does trust play

01:02:49.720 | in the human-robot interaction space?

01:02:51.320 | - So some of the things I study in this domain

01:02:53.920 | is not just trust, but it really is overtrust.

01:02:56.920 | - How do you think about overtrust?

01:02:58.120 | Like, what is, first of all,

01:02:59.520 | what is trust and what is overtrust?

01:03:03.360 | - Basically, the way I look at it is

01:03:05.800 | trust is not what you click on a survey.

01:03:08.040 | Trust is about your behavior.

01:03:09.560 | So if you interact with the technology

01:03:11.900 | based on the decision or the actions of the technology,

01:03:17.280 | as if you trust that decision,

01:03:19.680 | then you're trusting.

01:03:20.720 | Even in my group, we've done surveys that,

01:03:26.620 | on the thing, do you trust robots?

01:03:28.260 | Of course not.

01:03:29.100 | Would you follow this robot in an abandoned building?

01:03:31.660 | Of course not.

01:03:32.920 | And then you look at their actions,

01:03:34.440 | and you're like, clearly your behavior

01:03:37.220 | does not match what you think, right?

01:03:39.640 | Or what you think you would like to think, right?

01:03:41.980 | And so I'm really concerned about the behavior,

01:03:44.040 | 'cause that's really, at the end of the day,

01:03:45.800 | when you're in the world,

01:03:47.340 | that's what will impact others around you.

01:03:50.480 | It's not whether before you went onto the street,

01:03:52.920 | you clicked on, like, I don't trust self-driving cars.

01:03:55.680 | - Yeah, that, from an outsider perspective,

01:03:58.680 | it's always frustrating to me.

01:04:00.600 | Well, I read a lot,

01:04:01.480 | so I'm insider in a certain philosophical sense.

01:04:06.040 | It's frustrating to me how often trust is used in surveys,

01:04:10.700 | and how people say, make claims

01:04:14.420 | out of any kind of finding they make

01:04:16.220 | about somebody clicking on answer.

01:04:18.720 | Because trust is, yeah, behavior,

01:04:23.460 | just, you said it beautifully,

01:04:24.640 | I mean, the action, your own behavior is what trust is.

01:04:28.140 | I mean, everything else is not even close.

01:04:30.800 | It's almost like absurd, comedic poetry,

01:04:35.640 | poetry that you weave around your actual behavior.

01:04:38.560 | So some people can say they trust,

01:04:42.120 | I trust my wife, husband, or not, whatever,

01:04:46.100 | but the actions is what speaks volumes.

01:04:48.280 | - You bug their car.

01:04:49.900 | (laughing)

01:04:51.080 | You probably don't trust them.

01:04:52.320 | - I trust them, I'm just making sure.

01:04:53.840 | No, no, that's, yeah.

01:04:55.640 | - Like, even if you think about cars,

01:04:57.320 | I think it's a beautiful case.

01:04:58.600 | I came here at some point, I'm sure,

01:05:01.280 | on either Uber or Lyft, right?

01:05:03.640 | I remember when it first came out.

01:05:05.840 | I bet if they had had a survey,

01:05:08.080 | would you get in the car with a stranger and pay them?

01:05:11.480 | - Yes.

01:05:12.720 | - How many people do you think would have said,

01:05:15.360 | like, really?

01:05:16.720 | Wait, even worse, would you get in the car

01:05:18.760 | with a stranger at 1 a.m. in the morning

01:05:22.000 | to have them drop you home as a single female?

01:05:25.680 | Like, how many people would say, that's stupid?

01:05:29.360 | - Yeah.

01:05:30.200 | - And now look at where we are.

01:05:31.640 | I mean, people put kids, right?

01:05:34.040 | Like, oh yeah, my child has to go to school,

01:05:37.760 | and I, yeah, I'm gonna put my kid in this car

01:05:40.640 | with a stranger.

01:05:42.420 | I mean, it's just fascinating how,

01:05:45.320 | like, what we think we think

01:05:47.160 | is not necessarily matching our behavior.

01:05:49.760 | - Yeah, and certainly with robots, with autonomous vehicles,

01:05:52.400 | and all the kinds of robots you work with,

01:05:54.760 | that's, it's, yeah, it's, the way you answer it,

01:06:00.440 | especially if you've never interacted with that robot before.

01:06:03.440 | If you haven't had the experience,

01:06:05.720 | you being able to respond correctly on a survey is impossible.

01:06:09.640 | But what role does trust play in the interaction,

01:06:13.440 | do you think?

01:06:14.280 | Like, is it good to, is it good to trust a robot?

01:06:19.280 | What does overtrust mean?

01:06:21.720 | Or is it good to, kind of how you feel

01:06:24.080 | about autopilot currently, which is like,

01:06:26.560 | from a robotics perspective, is like--

01:06:29.440 | - So cautious.

01:06:30.280 | - So very cautious.

01:06:31.640 | - Yeah, so this is still an open area of research,

01:06:35.040 | but basically what I would like in a perfect world

01:06:40.040 | is that people trust the technology when it's working 100%,

01:06:45.000 | and people will be hypersensitive and identify when it's not.

01:06:49.160 | But of course we're not there.

01:06:51.080 | That's the ideal world.

01:06:52.800 | And, but we find is that people swing, right?

01:06:56.600 | They tend to swing,

01:06:58.400 | which means that if my first,

01:07:01.400 | and like, we have some papers,

01:07:03.000 | like first impressions is everything, right?

01:07:05.360 | If my first instance with technology,

01:07:07.760 | with robotics is positive, it mitigates any risk,

01:07:12.760 | it correlates with like best outcomes,

01:07:16.960 | it means that I'm more likely to either not see it

01:07:21.600 | when it makes some mistakes or faults,

01:07:24.320 | or I'm more likely to forgive it.

01:07:28.240 | And so this is a problem

01:07:30.400 | because technology is not 100% accurate, right?

01:07:32.680 | It's not 100% accurate, although it may be perfect.

01:07:35.120 | - How do you get that first moment right, do you think?

01:07:37.720 | There's also an education about the capabilities

01:07:40.720 | and limitations of the system.

01:07:42.520 | Do you have a sense of how you educate people correctly

01:07:45.760 | in that first interaction?

01:07:47.160 | - Again, this is an open-ended problem.

01:07:50.280 | So one of the study that actually has given me some hope

01:07:55.040 | that I was trying to figure out how to put in robotics.

01:07:57.680 | So there was a research study

01:08:01.320 | that has showed for medical AI systems,

01:08:03.480 | giving information to radiologists about,

01:08:07.880 | here, you need to look at these areas on the X-ray.

01:08:12.880 | What they found was that when the system provided one choice,

01:08:20.600 | there was this aspect of either no trust or overtrust, right?

01:08:25.600 | Like, I don't believe it at all, or a yes, yes, yes, yes.

01:08:32.000 | And they would miss things, right?

01:08:36.440 | Instead, when the system gave them multiple choices,

01:08:40.680 | like here are the three, even if it knew,

01:08:42.680 | like it had estimated that the top area

01:08:45.320 | that you need to look at was some place on the X-ray.

01:08:49.880 | If it gave like one plus others,

01:08:54.160 | the trust was maintained

01:08:57.600 | and the accuracy of the entire population increased, right?

01:09:02.600 | So basically it was a, you're still trusting the system,

01:09:07.600 | but you're also putting in a little bit

01:09:09.280 | of like your human expertise,

01:09:11.600 | like your human decision processing into the equation.

01:09:15.640 | So it helps to mitigate that overtrust risk.

01:09:18.640 | - Yeah, so there's a fascinating balance to have to strike.

01:09:21.640 | - I haven't figured out, again,

01:09:22.720 | in robotics, it's still an open research.

01:09:24.680 | - Open area research, exactly.

01:09:26.760 | So what are some exciting applications

01:09:29.000 | of human-robot interaction?

01:09:30.200 | You started a company, maybe you can talk

01:09:31.920 | about the exciting efforts there,

01:09:35.720 | but in general also, what other space can robots interact

01:09:39.600 | with humans and help?

01:09:41.080 | - Yeah, so besides healthcare,

01:09:42.400 | 'cause that's my bias lens,

01:09:44.560 | my other bias lens is education.

01:09:47.120 | I think that, well, one, we definitely,

01:09:52.040 | in the US, we're doing okay with teachers,

01:09:54.800 | but there's a lot of school districts

01:09:56.920 | that don't have enough teachers.

01:09:58.320 | If you think about the teacher-student ratio

01:10:01.960 | for at least public education,

01:10:04.840 | in some districts, it's crazy.

01:10:06.720 | It's like, how can you have learning in that classroom?

01:10:10.400 | Because you just don't have the human capital.

01:10:13.000 | And so if you think about robotics,

01:10:15.520 | bringing that in to classrooms,

01:10:18.480 | as well as the after-school space,

01:10:20.360 | where they offset some of this lack of resources

01:10:25.120 | in certain communities, I think that's a good place.

01:10:28.720 | And then turning, on the other end,

01:10:30.920 | is using these systems then for workforce retraining

01:10:35.280 | and dealing with some of the things

01:10:38.960 | that are going to come out later on of job loss,

01:10:43.040 | like thinking about robots and in AI systems

01:10:45.880 | for retraining and workforce development.

01:10:48.320 | I think that's exciting areas that can be pushed even more,

01:10:53.200 | and it would have a huge, huge impact.

01:10:56.760 | - What would you say are some of the open problems

01:10:59.640 | in education, sort of, it's exciting.

01:11:03.200 | So young kids and the older folks

01:11:08.720 | or just folks of all ages who need to be retrained,

01:11:12.560 | who need to sort of open themselves up

01:11:14.240 | to a whole 'nother area of work.

01:11:17.680 | What are the problems to be solved there?

01:11:20.040 | How do you think robots can help?

01:11:22.400 | - We have the engagement aspect, right?

01:11:24.800 | So we can figure out the engagement.

01:11:26.440 | That's not a-- - What do you mean

01:11:27.480 | by engagement?

01:11:28.880 | - So identifying whether a person is focused is,

01:11:33.880 | like, that we can figure out.

01:11:38.840 | What we can figure out,

01:11:40.800 | and there's some positive results in this,

01:11:44.640 | is that personalized adaptation based on any concepts.

01:11:49.640 | So imagine I think about, I have an agent,

01:11:54.680 | and I'm working with a kid learning, I don't know,

01:11:59.680 | algebra two.

01:12:01.600 | Can that same agent then switch and teach

01:12:05.880 | some type of new coding skill to a displaced mechanic?

01:12:10.880 | Like, what does that actually look like, right?

01:12:14.520 | Like, hardware might be the same,

01:12:17.680 | content is different, two different target demographics

01:12:21.360 | of engagement, like, how do you do that?

01:12:24.640 | - How important do you think personalization

01:12:26.880 | is in human-robot interaction?

01:12:28.640 | And not just mechanic or student,

01:12:32.040 | but like literally to the individual human being?

01:12:35.400 | - I think personalization is really important,

01:12:37.600 | but a caveat is that I think we'd be okay

01:12:42.160 | if we can personalize to the group, right?

01:12:44.720 | And so if I can label you as along some certain dimensions,

01:12:49.720 | then even though it may not be you specifically,

01:12:56.520 | I can put you in this group.

01:12:58.240 | So the sample size, this is how they best learn,

01:13:00.520 | this is how they best engage.

01:13:03.240 | Even at that level, it's really important.

01:13:06.840 | And it's because, I mean, it's one of the reasons

01:13:09.680 | why educating in large classrooms is so hard, right?

01:13:13.400 | You teach to the median, but there's these individuals

01:13:18.200 | that are struggling, and then you have

01:13:20.480 | highly intelligent individuals,

01:13:22.400 | and those are the ones that are usually kind of left out.

01:13:26.400 | So highly intelligent individuals may be disruptive,

01:13:28.960 | and those who are struggling might be disruptive

01:13:30.920 | because they're both bored.

01:13:33.040 | - Yeah, and if you narrow the definition of the group

01:13:35.560 | or in the size of the group enough,

01:13:37.560 | you'll be able to address their individual,

01:13:40.400 | it's not individual needs, but really the most--

01:13:42.680 | - Group needs. - Group,

01:13:43.960 | most important group needs.

01:13:45.720 | Right, and that's kind of what a lot

01:13:47.320 | of successful recommender systems do, Spotify and so on.

01:13:51.000 | So it's sad to believe, but I'm, as a music listener,

01:13:53.840 | probably in some sort of large group.

01:13:55.880 | (laughing)

01:13:56.920 | It's very sadly predictable.

01:13:58.120 | - You have been labeled.

01:13:59.280 | - Yeah, I've been labeled, and successfully so,

01:14:02.120 | because they're able to recommend stuff that I--

01:14:04.640 | - Yeah, but applying that to education, right?

01:14:07.760 | There's no reason why it can't be done.

01:14:09.800 | - Do you have a hope for our education system?

01:14:13.120 | - I have more hope for workforce development,

01:14:16.240 | and that's because I'm seeing investments.

01:14:19.720 | Even if you look at VC investments in education,

01:14:23.320 | the majority of it has lately been going

01:14:26.200 | to workforce retraining, right?

01:14:28.600 | And so I think that government investments is increasing.

01:14:32.960 | There's like a claim, and some of it's based on fear, right?

01:14:36.160 | Like AI's gonna come and take over all these jobs.

01:14:38.080 | What are we gonna do with all these non-paying taxes

01:14:41.560 | that aren't coming to us by our citizens?

01:14:44.400 | And so I think I'm more hopeful for that.

01:14:47.240 | Not so hopeful for early education,

01:14:51.840 | because it's this, it's still a who's gonna pay for it,

01:14:56.440 | and you won't see the results for like 16 to 18 years.

01:15:01.440 | It's hard for people to wrap their heads around that.

01:15:06.040 | - But on the retraining part, what are your thoughts?

01:15:10.680 | There's a candidate, Andrew Yang, running for president,

01:15:14.880 | saying that sort of AI automation robots--

01:15:19.040 | - Universal basic income.

01:15:21.080 | - Universal basic income in order to support us

01:15:24.000 | as we kind of automation takes people's jobs

01:15:26.760 | and allows you to explore and find other means.

01:15:30.200 | Like, do you have a concern of society transforming effects

01:15:35.200 | of automation and robots and so on?

01:15:40.520 | - I do.

01:15:41.360 | I do know that AI robotics will displace workers.

01:15:46.200 | Like, we do know that.

01:15:48.000 | But there'll be other workers

01:15:49.520 | that will be defined new jobs.

01:15:54.520 | What I worry about is, that's not what I worry about,

01:15:57.480 | like, will all the jobs go away?

01:15:59.520 | What I worry about is the type of jobs that will come out.

01:16:02.240 | Right, like people who graduate from Georgia Tech

01:16:05.040 | will be okay, right?

01:16:06.400 | We give them the skills, they will adapt

01:16:08.560 | even if their current job goes away.

01:16:10.720 | I do worry about those that don't have

01:16:13.640 | that quality of an education, right?

01:16:15.480 | Will they have the ability, the background

01:16:19.560 | to adapt to those new jobs?

01:16:21.760 | That, I don't know.

01:16:23.080 | That I worry about, which will create

01:16:25.720 | even more polarization in our society,

01:16:29.600 | internationally, and everywhere.

01:16:31.320 | I worry about that.

01:16:33.000 | I also worry about not having equal access

01:16:36.920 | to all these wonderful things that AI can do

01:16:39.640 | and robotics can do.

01:16:41.160 | I worry about that.

01:16:43.160 | You know, people like me from Georgia Tech,

01:16:47.320 | from say, MIT, will be okay, right?

01:16:50.400 | But that's such a small part of the population

01:16:53.400 | that we need to think much more globally

01:16:56.000 | of having access to the beautiful things,

01:16:58.560 | whether it's AI in healthcare, AI in education,

01:17:01.600 | AI in politics, right?

01:17:05.200 | I worry about that.

01:17:06.040 | - And that's part of the thing that you were talking about

01:17:08.200 | is people that build the technology

01:17:09.680 | have to be thinking about ethics,

01:17:12.480 | have to be thinking about access and all those things,

01:17:15.240 | and not just a small subset.

01:17:17.920 | Let me ask some philosophical, slightly romantic questions.

01:17:22.160 | - All right. - People that listen to this

01:17:24.560 | will be like, here he goes again.

01:17:26.320 | Okay, do you think one day we'll build an AI system

01:17:31.320 | that a person can fall in love with

01:17:35.640 | and it would love them back?

01:17:38.040 | Like in the movie "Her," for example.

01:17:39.960 | - Yeah, although she kind of didn't fall in love with him,

01:17:43.400 | or she fell in love with like a million other people,

01:17:45.680 | something like that.

01:17:46.840 | - You're the jealous type, I see.

01:17:48.600 | (laughing)

01:17:49.880 | We humans are the jealous type.

01:17:50.960 | - Yes, so I do believe that we can design systems

01:17:55.120 | where people would fall in love with their robot,

01:17:59.480 | with their AI partner.

01:18:03.280 | That I do believe.

01:18:05.160 | Because it's actually,

01:18:06.360 | and I don't like to use the word manipulate,

01:18:09.000 | but as we see, there are certain individuals

01:18:12.360 | that can be manipulated

01:18:13.400 | if you understand the cognitive science about it, right?

01:18:16.360 | - Right, so I mean, if you could think

01:18:18.560 | of all close relationship and love in general

01:18:21.440 | as a kind of mutual manipulation,

01:18:24.800 | that dance, the human dance.

01:18:27.200 | I mean, manipulation is a negative connotation.

01:18:29.560 | - And that's why I don't like to use that word particularly.

01:18:32.880 | - I guess another way to phrase it is,

01:18:34.320 | you're getting at it as it could be

01:18:36.000 | algorithmatized or something.

01:18:37.440 | It could be--

01:18:38.480 | - The relationship building part can be.

01:18:40.720 | I mean, just think about it.

01:18:41.920 | We have, and I don't use dating sites,

01:18:44.920 | but from what I heard,

01:18:46.400 | there are some individuals that have been dating

01:18:50.480 | that have never saw each other, right?

01:18:52.920 | In fact, there's a show I think

01:18:54.200 | that tries to weed out fake people.

01:18:57.640 | Like there's a show that comes out, right?

01:18:59.560 | Because people start faking.

01:19:02.080 | Like, what's the difference of that person

01:19:05.240 | on the other end being an AI agent, right?

01:19:08.160 | And having a communication,

01:19:09.440 | are you building a relationship remotely?

01:19:12.280 | Like there's no reason why that can't happen.

01:19:14.920 | - In terms of human-robot interaction,

01:19:17.680 | so what role, you've kind of mentioned with data,

01:19:20.640 | emotion being, can be problematic if not implemented well,

01:19:25.440 | I suppose.

01:19:26.320 | What role does emotion and some other human-like things,

01:19:30.600 | the imperfect things come into play here

01:19:32.880 | for good human-robot interaction and something like love?

01:19:37.400 | - Yeah, so in this case, and you had asked,

01:19:39.880 | can an AI agent love a human back?

01:19:43.800 | I think they can emulate love back, right?

01:19:47.440 | And so what does that actually mean?

01:19:49.040 | It just means that if you think about their programming,

01:19:52.320 | they might put the other person's needs in front of theirs

01:19:56.280 | in certain situations, right?

01:19:58.040 | You look at, think about it as return on investment.

01:20:00.440 | Like, what's my return on investment?

01:20:01.800 | As part of that equation, that person's happiness,

01:20:04.600 | has some type of algorithm waiting to it.

01:20:08.000 | And the reason why is because I care about them, right?

01:20:11.440 | That's the only reason, right?

01:20:13.760 | But if I care about them and I show that,

01:20:15.600 | then my final objective function

01:20:18.320 | is length of time of the engagement, right?

01:20:20.600 | So you can think of how to do this actually quite easily.

01:20:24.040 | And so--

01:20:24.880 | - But that's not love?

01:20:26.520 | - Well, so that's the thing.

01:20:30.000 | I think it emulates love

01:20:32.600 | because we don't have a classical definition of love.

01:20:37.600 | - Right, but, and we don't have the ability

01:20:41.640 | to look into each other's minds to see the algorithm.

01:20:45.480 | And I mean, I guess what I'm getting at is,

01:20:48.760 | is it possible that, especially if that's learned,

01:20:51.040 | especially if there's some mystery

01:20:52.600 | and black box nature to the system,

01:20:55.240 | how is that, you know--

01:20:57.680 | - How is it any different?

01:20:58.600 | How is it any different in terms of sort of,

01:21:00.680 | if the system says, "I'm conscious, I'm afraid of death,"

01:21:04.160 | and it does indicate that it loves you,

01:21:10.160 | another way to sort of phrase it,

01:21:12.520 | I'd be curious to see what you think.

01:21:14.200 | Do you think there'll be a time

01:21:15.720 | when robots should have rights?

01:21:20.160 | You've kind of phrased the robot in a very roboticist way,

01:21:23.440 | and just a really good way, but saying,

01:21:25.680 | "Okay, well, there's an objective function,

01:21:27.920 | "and I could see how you can create

01:21:30.600 | "a compelling human-robot interaction experience

01:21:33.360 | "that makes you believe that the robot cares for your needs,

01:21:36.280 | "and even something like loves you."

01:21:39.040 | But what if the robot says, "Please don't turn me off"?

01:21:43.840 | What if the robot starts making you feel

01:21:46.560 | like there's an entity, a being, a soul there, right?

01:21:50.160 | Do you think there'll be a future,

01:21:52.200 | hopefully you won't laugh too much at this,

01:21:55.800 | but where they do ask for rights?

01:22:00.120 | - So I can see a future if we don't address it

01:22:05.120 | in the near term, where these agents,

01:22:10.400 | as they adapt and learn, could say,

01:22:12.360 | "Hey, this should be something that's fundamental."

01:22:15.920 | I hopefully think that we would address it

01:22:18.960 | before it gets to that point.

01:22:20.200 | - You think that's a bad future?

01:22:22.280 | Is that a negative thing, where they ask,

01:22:25.440 | "We're being discriminated against"?

01:22:27.840 | - I guess it depends on what role

01:22:31.200 | have they attained at that point, right?

01:22:34.440 | And so if I think about now--

01:22:35.960 | - Careful what you say, because the robots 50 years from now

01:22:39.320 | will be listening to this, and you'll be on TV saying,

01:22:42.200 | "This is what roboticists used to believe."

01:22:44.520 | - Well, right?

01:22:45.360 | And so this is my, and as I said, I have a biased lens,

01:22:48.800 | and my robot friends will understand that.

01:22:50.920 | But so if you think about it,

01:22:54.000 | and I actually put this in kind of the, as a roboticist,

01:22:59.000 | you don't necessarily think of robots as human,

01:23:02.600 | with human rights, but you could think of them

01:23:05.120 | either in the category of property,

01:23:09.280 | or you could think of them in the category of animals.

01:23:12.960 | And so both of those have different types of rights.

01:23:18.440 | So animals have their own rights as a living being,

01:23:22.840 | but they can't vote, they can't write,

01:23:25.160 | they can be euthanized.

01:23:28.240 | But as humans, if we abuse them, we go to jail.

01:23:32.080 | So they do have some rights that protect them,

01:23:36.120 | but don't give them the rights of citizenship.

01:23:39.320 | And then if you think about property,

01:23:42.400 | property, the rights are associated with the person.

01:23:45.840 | So if someone vandalizes your property,

01:23:49.640 | or steals your property, there are some rights,

01:23:53.960 | but it's associated with the person who owns that.

01:23:57.820 | If you think about it, back in the day,

01:24:01.640 | and if you remember, we talked about

01:24:03.480 | how society has changed, women were property, right?

01:24:08.280 | They were not thought of as having rights.

01:24:11.960 | They were thought of as property of, like their--

01:24:15.880 | - Yeah, assaulting a woman meant assaulting the property

01:24:18.880 | of somebody else's person.

01:24:20.160 | - Exactly, and so what I envision is,

01:24:22.960 | is that we will establish some type of norm at some point,

01:24:27.880 | but that it might evolve, right?

01:24:29.640 | Like if you look at women's rights now,

01:24:32.000 | like there are still some countries that don't have,

01:24:35.840 | and the rest of the world is like,

01:24:37.160 | why that makes no sense, right?

01:24:39.760 | And so I do see a world where we do establish

01:24:42.560 | some type of grounding.

01:24:44.600 | It might be based on property rights,

01:24:46.200 | it might be based on animal rights.

01:24:48.080 | And if it evolves that way,

01:24:51.200 | I think we will have this conversation at that time,

01:24:54.960 | because that's the way our society

01:24:56.560 | traditionally has evolved.

01:24:59.040 | - Beautifully put.

01:25:01.960 | Just out of curiosity, Anki, Jibo, Mayfield Robotics,

01:25:06.960 | with their robot Curie, Sci-Fi Works, Rethink Robotics,

01:25:10.240 | were all these amazing robotics companies

01:25:12.480 | led, created by incredible roboticists,

01:25:16.240 | and they've all went out of business recently.

01:25:21.240 | Why do you think they didn't last longer?

01:25:23.520 | Why is it so hard to run a robotics company,

01:25:26.960 | especially one like these,

01:25:29.800 | which are fundamentally HRI,

01:25:33.440 | Human Robot Interaction Robots?

01:25:36.000 | - Yeah. - Or personal robots.

01:25:37.360 | - Each one has a story.

01:25:38.960 | Only one of them I don't understand, and that was Anki.

01:25:43.040 | That's actually the only one I don't understand.

01:25:45.240 | - I don't understand it either.

01:25:46.320 | It's-- - No, no, I mean,

01:25:47.160 | I look at it from the outside.

01:25:49.000 | I've looked at their sheets.

01:25:50.920 | I've looked at the data that's--

01:25:52.280 | - Oh, you mean business-wise,

01:25:53.560 | you don't understand, gotcha. - Yeah, yeah.

01:25:55.240 | And I look at that data,

01:25:59.200 | and I'm like, they seem to have product-market fit.

01:26:02.720 | So that's the only one I don't understand.

01:26:05.720 | The rest of it was product-market fit.

01:26:08.280 | - What's product-market fit,

01:26:09.880 | just out of, how do you think about it?

01:26:11.960 | - Yeah, so although we think robotics was getting there,

01:26:15.680 | but I think it's just the timing,

01:26:17.640 | their clock just timed out.

01:26:20.400 | I think if they'd been given a couple of more years,

01:26:23.160 | they would've been okay.

01:26:25.120 | But the other ones were still fairly early

01:26:28.680 | by the time they got into the market.

01:26:30.160 | And so product-market fit is,

01:26:32.760 | I have a product that I wanna sell at a certain price.

01:26:37.200 | Are there enough people out there, the market,

01:26:40.080 | that are willing to buy the product at that market price

01:26:42.800 | for me to be a functional, viable,

01:26:46.000 | profit-bearing company, right?

01:26:48.920 | So product-market fit.

01:26:50.440 | If it costs you $1,000, and everyone wants it,

01:26:55.360 | and only is willing to pay a dollar,

01:26:57.400 | you have no product-market fit,

01:26:59.320 | even if you could sell it for, you know,

01:27:01.960 | it's enough for a dollar, 'cause you can't--

01:27:03.720 | - So how hard is it for robots?

01:27:05.440 | So if maybe, if you look at iRobot,

01:27:07.640 | the company that makes Roombas, vacuum cleaners,

01:27:10.840 | can you comment on, did they find the right product,

01:27:14.200 | market-product fit?

01:27:15.200 | Like, are people willing to pay for robots

01:27:18.680 | is also another kind of question underlying all this.

01:27:20.360 | - So if you think about iRobot and their story, right?

01:27:23.800 | Like when they first, they had enough of a runway, right?

01:27:28.760 | When they first started,

01:27:29.880 | they weren't doing vacuum cleaners, right?

01:27:31.440 | They were a military, they were contracts, primarily,

01:27:35.320 | government contracts, designing robots.

01:27:37.840 | - Military robots.

01:27:38.680 | - Yeah, I mean, that's what they were.

01:27:39.520 | That's how they started, right?

01:27:40.960 | And then-- - They still do

01:27:41.800 | a lot of incredible work there.

01:27:42.880 | But yeah, that was the initial thing

01:27:44.800 | that gave them enough funding to--

01:27:46.760 | - To then try to, the vacuum cleaner is what I've been told

01:27:50.880 | was not like their first rendezvous

01:27:54.080 | in terms of designing a product, right?

01:27:56.640 | And so they were able to survive

01:27:59.440 | until they got to the point that they found

01:28:02.360 | a product price market, right?

01:28:05.640 | And even with, if you look at the Roomba,

01:28:09.200 | the price point now is different

01:28:10.640 | than when it was first released, right?

01:28:12.360 | It was an early adopter price,

01:28:13.560 | but they found enough people who were willing to fund it.

01:28:16.760 | And I mean, I forgot what their loss profile was

01:28:20.400 | for the first couple of years,

01:28:22.240 | but they became profitable in sufficient time

01:28:25.920 | that they didn't have to close their doors.

01:28:28.200 | - So they found the right,

01:28:29.240 | there's still people willing to pay a large amount of money,

01:28:32.760 | sort of over $1,000 for a vacuum cleaner.

01:28:36.040 | Unfortunately for them,

01:28:37.840 | now that they've proved everything out

01:28:39.240 | and figured it all out, now there's competitors.

01:28:40.920 | - Yeah, and so that's the next thing, right?

01:28:43.560 | The competition, and they have quite a number,

01:28:46.680 | even internationally, like there's some products out there

01:28:50.240 | you can go to Europe and be like,

01:28:52.480 | "Oh, I didn't even know this one existed."

01:28:55.120 | So this is the thing though,

01:28:56.880 | like with any market, I would,

01:29:00.400 | this is not a bad time,

01:29:03.680 | although as a roboticist, it's kind of depressing,

01:29:06.480 | but I actually think about things like with,

01:29:10.640 | I would say that all of the companies

01:29:13.160 | that are now in the top five or six,

01:29:15.880 | they weren't the first to the stage, right?

01:29:19.720 | Like Google was not the first search engine,

01:29:22.880 | sorry AltaVista, right?

01:29:24.880 | Facebook was not the first, sorry, MySpace, right?

01:29:28.440 | Like think about it, they were not the first players.

01:29:31.200 | Those first players, like they're not in the top five,

01:29:35.560 | 10 of Fortune 500 companies, right?

01:29:39.520 | They proved, they started to prove out the market,

01:29:44.080 | they started to get people interested,

01:29:46.480 | they started the buzz,

01:29:48.440 | but they didn't make it to that next level.

01:29:50.200 | But the second batch, right?

01:29:52.400 | The second batch, I think might make it to the next level.

01:29:57.400 | - When do you think the Facebook of, ugh.

01:30:02.480 | - The Facebook of robotics.

01:30:03.800 | - Sorry, I take that phrase back

01:30:07.360 | because people deeply, for some reason,

01:30:09.640 | well, I know why, but it's, I think,

01:30:11.760 | exaggerated distrust Facebook

01:30:13.840 | because of the privacy concerns and so on.

01:30:15.640 | And with robotics, one of the things you have to make sure

01:30:18.560 | is all the things we've talked about

01:30:20.240 | is to be transparent and have people deeply trust you

01:30:23.160 | to let a robot into their lives, into their home.

01:30:25.960 | When do you think the second batch of robots,

01:30:28.800 | is it five, 10 years, 20 years,

01:30:32.320 | that we'll have robots in our homes

01:30:34.880 | and robots in our hearts?

01:30:36.720 | - So if I think about, 'cause I try to follow the VC

01:30:40.200 | kind of space in terms of robotic investments.

01:30:43.360 | And right now, and I don't know

01:30:45.080 | if they're gonna be successful,

01:30:46.120 | I don't know if this is the second batch,

01:30:49.400 | but there's only one batch that's focused on

01:30:51.600 | the first batch, right?

01:30:53.040 | And then there's all these self-driving Xs, right?

01:30:56.440 | And so I don't know if they're a first batch of something

01:30:59.720 | or if, I don't know quite where they fit in,

01:31:03.240 | but there's a number of companies,

01:31:05.680 | the co-robot, I call them co-robots,

01:31:08.600 | that are still getting VC investments.

01:31:11.480 | Some of them have some of the flavor

01:31:14.640 | of like Rethink Robotics,

01:31:15.880 | some of them have some of the flavor of like Curie.

01:31:19.120 | - What's a co-robot?

01:31:20.840 | - So basically a robot and human working in the same space.

01:31:25.840 | So some of the companies are focused on manufacturing.

01:31:30.600 | So having a robot and human working together in a factory,

01:31:35.600 | some of these co-robots are robots and humans

01:31:40.000 | working in the home, working in clinics.

01:31:42.200 | Like there's different versions of these companies

01:31:44.080 | in terms of their products, but they're all,

01:31:47.040 | so Rethink Robotics would be like one of the first,

01:31:50.320 | at least well-known companies focused on this space.

01:31:54.720 | So I don't know if this is a second batch

01:31:56.840 | or if this is still part of the first batch,

01:32:01.080 | that I don't know.

01:32:02.120 | And then you have all these other companies

01:32:03.880 | in this self-driving space.

01:32:06.960 | And I don't know if that's a first batch

01:32:09.520 | or again, a second batch.

01:32:11.280 | - Yeah, so there's a lot of mystery about this now.

01:32:14.000 | Of course, it's hard to say that this is the second batch

01:32:16.520 | until it proves out, right?

01:32:18.520 | - Correct. - Yeah, exactly.

01:32:19.360 | - Yeah, we need a unicorn.

01:32:20.640 | - Yeah, exactly.

01:32:21.740 | Why do you think people are so afraid,

01:32:26.840 | at least in popular culture of legged robots

01:32:30.520 | like those worked in Boston Dynamics

01:32:32.440 | or just robotics in general?

01:32:34.200 | If you were to psychoanalyze that fear,

01:32:36.360 | what do you make of it?

01:32:38.040 | And should they be afraid, sorry?

01:32:39.840 | - So should people be afraid?

01:32:41.520 | I don't think people should be afraid, but with a caveat.

01:32:45.400 | I don't think people should be afraid

01:32:47.160 | given that most of us in this world

01:32:51.520 | understand that we need to change something, right?

01:32:55.760 | So given that.

01:32:58.200 | Now, if things don't change, be very afraid.

01:33:01.600 | - Which is the dimension of change that's needed?

01:33:04.520 | - So thinking about the ramifications,

01:33:07.880 | thinking about like the ethics,

01:33:09.520 | thinking about like the conversation is going on, right?

01:33:12.840 | It's no longer a, we're gonna deploy it and forget that,

01:33:17.760 | this is a car that can kill pedestrians

01:33:20.440 | that are walking across the street, right?

01:33:22.600 | We're not in that stage.

01:33:24.440 | We're putting these roads out.

01:33:25.920 | There are people out there.

01:33:27.600 | A car could be a weapon.

01:33:28.920 | People are now, solutions aren't there yet,

01:33:33.240 | but people are thinking about this

01:33:35.440 | as we need to be ethically responsible

01:33:38.560 | as we send these systems out,

01:33:40.920 | robotics, medical, self-driving.

01:33:43.160 | - And military too.

01:33:44.560 | - And military.

01:33:45.400 | - Which is not as often talked about,

01:33:47.120 | but it's really where probably these robots

01:33:50.360 | will have a significant impact as well.

01:33:52.040 | - Correct, correct, right?

01:33:53.240 | Making sure that they can think rationally,

01:33:57.480 | even having the conversations,

01:33:58.840 | who should pull the trigger, right?

01:34:01.360 | - But overall, you're saying if we start to think more

01:34:03.720 | and more as a community about these ethical issues,

01:34:05.840 | people should not be afraid.

01:34:07.080 | - Yeah, I don't think people should be afraid.

01:34:08.760 | I think that the return on investment,

01:34:10.640 | the positive impact will outweigh

01:34:14.160 | any of the potentially negative impacts.

01:34:17.600 | - Do you have worries of existential threats of robots

01:34:21.800 | or AI that some people kind of talk about

01:34:25.800 | and romanticize about in the next decade,

01:34:28.920 | next few decades?

01:34:30.240 | - No, I don't.

01:34:31.600 | Singularity would be an example.

01:34:33.960 | So my concept is that, so remember,

01:34:36.640 | robots, AI is designed by people.

01:34:39.840 | It has our values.

01:34:41.520 | And I always correlate this with a parent and a child, right?

01:34:45.360 | So think about it, as a parent, what do we want?

01:34:47.400 | We want our kids to have a better life than us.

01:34:50.120 | We want them to expand.

01:34:52.560 | We want them to experience the world.

01:34:56.080 | And then as we grow older, our kids think and know

01:35:00.000 | they're smarter and better and more intelligent

01:35:03.280 | and have better opportunities.

01:35:05.040 | And they may even stop listening to us.

01:35:08.480 | They don't go out and then kill us, right?

01:35:10.800 | Like think about it.

01:35:11.640 | It's because it's instilled in them values.

01:35:14.480 | We instilled in them this whole aspect of community.

01:35:17.720 | And yes, even though you're maybe smarter

01:35:20.120 | and have more money and da, da, da,

01:35:22.760 | it's still about this love, caring relationship.

01:35:27.080 | And so that's what I believe.

01:35:28.040 | So even if like, you know, we've created the singularity

01:35:30.840 | in some archaic system back in like 1980

01:35:34.120 | that suddenly evolves,

01:35:35.680 | the fact is it might say, I am smarter, I am sentient.

01:35:40.560 | These humans are really stupid,

01:35:43.600 | but I think it'll be like, yeah,

01:35:45.600 | but I just can't destroy them.

01:35:48.000 | - Yeah, for sentimental value.

01:35:49.600 | Still just to come back for Thanksgiving dinner

01:35:53.280 | every once in a while.

01:35:54.120 | - Exactly.

01:35:55.440 | - That's so beautifully put.

01:35:57.560 | You've also said that "The Matrix"

01:35:59.840 | may be one of your more favorite AI-related movies.

01:36:03.800 | Can you elaborate why?

01:36:05.680 | - Yeah, it is one of my favorite movies.

01:36:07.920 | And it's because it represents

01:36:11.280 | kind of all the things I think about.

01:36:14.160 | So there's a symbiotic relationship

01:36:16.160 | between robots and humans, right?

01:36:20.200 | That symbiotic relationship is that they don't destroy us,

01:36:22.560 | they enslave us, right?

01:36:24.680 | But think about it.

01:36:25.680 | Even though they enslaved us,

01:36:30.360 | they needed us to be happy, right?

01:36:32.920 | And in order to be happy,

01:36:33.960 | they had to create this crude world

01:36:35.520 | that they then had to live in, right?

01:36:37.080 | That's the whole premise.

01:36:38.720 | But then there were humans that had a choice, right?

01:36:44.480 | Like you had a choice to stay in this horrific,

01:36:47.760 | horrific world where it was your fantasy life

01:36:51.320 | with all of the anomalies, perfection, but not accurate.

01:36:54.840 | Or you can choose to be on your own

01:36:58.040 | and like have maybe no food for a couple of days,

01:37:02.600 | but you were totally autonomous.

01:37:05.280 | And so I think of that as, and that's why.

01:37:08.080 | So it's not necessarily us being enslaved,

01:37:09.800 | but I think about us having the symbiotic relationship.

01:37:13.160 | Robots and AI, even if they become sentient,

01:37:15.880 | they're still part of our society

01:37:17.200 | and they will suffer just as much as we.

01:37:20.840 | - And there will be some kind of equilibrium

01:37:23.920 | that we'll have to find some symbiotic relationship.

01:37:26.800 | - Right, and then you have the ethicists,

01:37:28.320 | the robotics folks that were like,

01:37:30.280 | no, this has got to stop.

01:37:32.400 | I will take the other pill in order to make a difference.

01:37:36.320 | - So if you could hang out for a day with a robot,

01:37:41.200 | real or from science fiction, movies, books, safely,

01:37:45.320 | and get to pick his or her, their brain, who would you pick?

01:37:50.320 | - Gotta say it's Data.

01:37:58.240 | - Data.

01:37:59.360 | - I was gonna say Rosie, but I don't,

01:38:01.600 | I'm not really interested in her brain.

01:38:04.200 | I'm interested in Data's brain.

01:38:06.360 | - Data pre or post-emotionship?

01:38:09.000 | - Pre.

01:38:09.840 | - But don't you think it'd be a more interesting

01:38:14.720 | conversation post-emotionship?

01:38:16.640 | - Yeah, it would be drama.

01:38:18.200 | And I, you know, I'm human.

01:38:20.440 | I deal with drama all the time.

01:38:23.320 | But the reason why I wanna pick Data's brain

01:38:25.320 | is because I could have a conversation with him

01:38:30.040 | and ask, for example,

01:38:32.880 | how can we fix this ethics problem, right?

01:38:35.680 | And he could go through like the rational thinking

01:38:39.280 | and through that,

01:38:40.280 | he could also help me think through it as well.

01:38:43.080 | And so that's, there's like these questions,

01:38:44.760 | fundamental questions I think I could ask him

01:38:47.280 | that he would help me also learn from.

01:38:50.800 | And that fascinates me.

01:38:52.480 | - I don't think there's a better place to end it.

01:38:55.880 | Aiyana, thank you so much for talking to me.

01:38:57.360 | It was an honor.

01:38:58.200 | - Thank you, thank you.

01:38:59.040 | This was fun.

01:38:59.880 | - Thanks for listening to this conversation

01:39:03.280 | and thank you to our presenting sponsor, Cash App.

01:39:06.480 | Download it, use code LexPodcast.

01:39:09.120 | You'll get $10 and $10 will go to FIRST,

01:39:11.920 | a STEM education nonprofit that inspires hundreds

01:39:14.720 | of thousands of young minds

01:39:16.400 | to become future leaders and innovators.

01:39:19.280 | If you enjoy this podcast, subscribe on YouTube,

01:39:22.120 | get five stars on Apple Podcast,

01:39:24.120 | follow on Spotify, support on Patreon,

01:39:26.760 | or simply connect with me on Twitter.

01:39:29.800 | And now let me leave you with some words of wisdom

01:39:32.400 | from Arthur C. Clarke.

01:39:33.960 | Whether we are based on carbon or on silicon

01:39:38.520 | makes no fundamental difference.

01:39:40.560 | We should each be treated with appropriate respect.

01:39:43.600 | Thank you for listening and hope to see you next time.

01:39:47.720 | (upbeat music)

01:39:50.320 | (upbeat music)

01:39:52.920 | [BLANK_AUDIO]

Ayanna Howard: Human-Robot Interaction & Ethics of Safety-Critical Systems | Lex Fridman Podcast #66

Chapters