Latest / Daniel and Kelly’s Extraordinary Universe / What Does Artificial Intelligence Mean?

Transcript

What Does Artificial Intelligence Mean? · Daniel and Kelly’s Extraordinary Universe

0:07
Speaker 1: Alexa, what's the best science podcast on air? Hey? Are
0:11
Speaker 1: you trying to replace me with Alexa? What's going on here?
0:14
Speaker 1: Do you think you're replaceable? There's no way an artificial
0:17
Speaker 1: intelligence could ever make jokes nearly as funny as I am.
0:21
Speaker 1: I think there's no way an artificial intelligence would laugh
0:24
Speaker 1: at your jokes. I'm pretty sure I could. I could
0:27
Speaker 1: program a pretty dumb computer to laugh at my jokes.
0:30
Speaker 1: It's called the laugh track. But that's Hey, that's a
0:34
Speaker 1: new challenge for for AI. You know, first chest then
0:38
Speaker 1: go now science comedy. That's right, now, programs something that
0:43
Speaker 1: can find humor in Daniel's ramblings him and I'm Daniel,
1:05
Speaker 1: and this is our podcast. Daniel and Jorge explained the Universe,
1:09
Speaker 1: in which we try to download everything we know about
1:12
Speaker 1: the universe, episode by episode into your brain, whether you're
1:16
Speaker 1: a real person or an artificial intelligence, listening to our
1:19
Speaker 1: podcast while trying to sound intelligent about it while writing
1:24
Speaker 1: your own humor for the open mic AI Night. The
1:28
Speaker 1: topic of today's podcast is what is artificial intelligence? And
1:36
Speaker 1: very importantly is it dangerous? That's right? Should you be
1:40
Speaker 1: looking at your window for the first signs of the
1:42
Speaker 1: robot revolution. Should you be afraid of your Alexa? Should
1:46
Speaker 1: you be worried about that robot vacuum cleaner getting resentful
1:50
Speaker 1: for having to do all the dirty work and eating
1:52
Speaker 1: your face off in the middle of the night. That's
1:56
Speaker 1: a bit dark. It seems kind of sinister, doesn't it.
1:59
Speaker 1: It's like sitting there are circling, circling, circling, waning, waiting, waiting.
2:03
Speaker 1: I think those things are creepy, right. Maybe it wants to,
2:07
Speaker 1: you know, clean your face, it wants see That's the question.
2:10
Speaker 1: Does a robot vacuum cleaner want anything? What does it
2:13
Speaker 1: mean for it to want? What is it like to
2:15
Speaker 1: be a robot vacuum cleaner? The next great paper in philosophy?
2:19
Speaker 1: So this is kind of in the zeitgeist right now.
2:21
Speaker 1: I mean, people are really excited about artificial intelligence. But
2:24
Speaker 1: at the same time there are big names like Elon
2:27
Speaker 1: Musk kind of warning people like, hey, artificial intelligence not
2:31
Speaker 1: such a good idea. That's right, it's a huge topic.
2:33
Speaker 1: I mean, you drive around like San Francisco, you see
2:35
Speaker 1: artificial intelligence, machine learning, deep learning. It's on billboards. Even
2:39
Speaker 1: you know you want to get a million bucks for
2:41
Speaker 1: your new company, you just say the words AI, deep learning,
2:43
Speaker 1: and boom people are throwing cash to right, people are
2:46
Speaker 1: learning in the deepest learning. Um. It's definitely part of
2:51
Speaker 1: the cultural moment, and you see that reflected not just
2:53
Speaker 1: in like what deep thinkers are saying, but also in
2:55
Speaker 1: like science fiction. You know, a lot of the near
2:58
Speaker 1: term dystopian these days is about how AI will take
3:01
Speaker 1: over and the dangers of AI. Another way, like thirty
3:04
Speaker 1: years ago is about the dangers of UH of radiation. Right,
3:07
Speaker 1: that was the new dangerous thing physicists that invented. Now
3:11
Speaker 1: the new dangerous technology that we're all worried about is AI.
3:13
Speaker 1: It's a new promise in peril. Um. Yeah. AI. Every
3:17
Speaker 1: piece of technology is a double edged sword, right. You
3:19
Speaker 1: can use it for good, you can use for evil.
3:21
Speaker 1: But AI is special because it's not just technology, is
3:24
Speaker 1: not just a tool that people use. It's a tool
3:27
Speaker 1: that that has independence, that has autonomy. And that's why
3:30
Speaker 1: it's such a vexing question. Well people, I'm sure people
3:33
Speaker 1: everyone associated with robots and machines and computers, but we
3:37
Speaker 1: were kind of wondering if people actually knew what artificial
3:40
Speaker 1: intelligence was, like, what makes it work, what makes it
3:43
Speaker 1: different than than real intelligence? M I bet you that
3:47
Speaker 1: the people who say the phrase artificial intelligence don't actually
3:50
Speaker 1: know what they're talking about, which is probably true for
3:53
Speaker 1: most topics andology, it's true for me probably, I'm sure.
3:59
Speaker 1: But we're wondering if you guys out there knew what
4:01
Speaker 1: artificial intelligence was. And so, as usual, Daniel went out
4:05
Speaker 1: and as people in the street, and here's what they
4:07
Speaker 1: had to say. Um, it's the idea that we can
4:11
Speaker 1: create some type of material thing that could think on
4:14
Speaker 1: its own ultimately, And do you think it's something we
4:17
Speaker 1: should be concerned about? Is it ever going to be
4:19
Speaker 1: a threat to humanity? I mean possibly, but I mean
4:22
Speaker 1: we never We don't know everything. We can know the
4:24
Speaker 1: bounds of what could be a threat, we cannot be
4:26
Speaker 1: a threat. Yeah, it's AI, and it's the stuff that's
4:30
Speaker 1: used in various technological applications basically just kind of like
4:35
Speaker 1: trying to make machines replicate certain aspects of human intelligence.
4:39
Speaker 1: Stuff like that. Okay, And do you think it could
4:41
Speaker 1: ever be a threat to humanity? Is something we should
4:43
Speaker 1: be worried about? I guess since I don't have a
4:45
Speaker 1: particularly strong opinion on it, I don't think so. So
4:49
Speaker 1: I guess I'll say no for now. Um, I'm assuming
4:52
Speaker 1: that's the idea that computers or electronics can have like sentience. Right,
4:59
Speaker 1: Are you worried that computers would when they take over
5:02
Speaker 1: and make us their slaves? And not really, I don't
5:05
Speaker 1: think it will come to that point. All right. Those
5:07
Speaker 1: are pretty sophisticated answers. I like the ones that said,
5:10
Speaker 1: um oh, artificial intelligence, that's just AI, right, Like that's
5:14
Speaker 1: an answer. So that's an answer to every question. You know,
5:19
Speaker 1: what is Google blex Zavi Brown? Oh, that's just gens. Yeah, acronyms.
5:24
Speaker 1: Acronyms can make you look intelligent, that's it. That's the
5:27
Speaker 1: real artificial intelligence to speaking acronym acronym intelligence. Um. You know,
5:34
Speaker 1: but people had some sense that it's, you know, something
5:37
Speaker 1: that can think for itself, or something do something for you,
5:40
Speaker 1: or create something that can think by itself. There's definitely
5:43
Speaker 1: the nuggative idea is definitely out there. They use it
5:46
Speaker 1: in relation to what it can do. That's right, yeah,
5:49
Speaker 1: exactly what what is what's the new capability that defines it? Yeah?
5:53
Speaker 1: Right yeah, and that it's it's a fascinating way to
5:56
Speaker 1: think about it, you know. And uh, it's definitely a
5:57
Speaker 1: tricky question, right because I guess we know it in
6:00
Speaker 1: the context of UM using them for things, right, Like,
6:02
Speaker 1: we're people, don't just create a I because we want
6:05
Speaker 1: to create artificial beings. It's like, so it can help us.
6:08
Speaker 1: I want to create artificial beings. What's what's wrong with that?
6:11
Speaker 1: That sounds pretty awesome. Create a whole army of physics
6:15
Speaker 1: artificial physics grad students. It sounds pretty cool. Kids. Yeah,
6:19
Speaker 1: you mean, are they worried about competing with my digital children? Natural?
6:23
Speaker 1: They know you'd rather have artificial children. I didn't say
6:26
Speaker 1: I'd rather have artificial that in addition to my beautiful,
6:30
Speaker 1: wonderful natural children, which I should not be talking about
6:32
Speaker 1: on this podcast. I'd love to have a whole, you know,
6:34
Speaker 1: cadre of artificial children to do my bidding, unlike your
6:38
Speaker 1: real children. If who won't do your bidding? So somebody
6:42
Speaker 1: listening children? And that sort of goes to the heart
6:47
Speaker 1: of the question. You know, UM, if you created a
6:49
Speaker 1: digital being with artificial intelligence, would it listen to you?
6:52
Speaker 1: Or would it make its own decisions? Right? And so
6:55
Speaker 1: that's why we thought it would be interesting to dig into,
6:57
Speaker 1: like what is artificial intelligence? If it just did what
6:59
Speaker 1: you told that to do. It wouldn't maybe be in
7:01
Speaker 1: an artificial intelligence you're saying nobody smart should listen to you,
7:04
Speaker 1: is what you're saying. I'm saying they should decide for
7:08
Speaker 1: themselves whether I'm I'm worth following. So let's break it
7:15
Speaker 1: down for people. Daniel, what is artificial intelligence? Well you
7:19
Speaker 1: should listen to this podcast and that will give you
7:21
Speaker 1: the answer. Done. Um, Well, you know, I think to
7:25
Speaker 1: understand what artificial intelligence is, we should think for a
7:27
Speaker 1: moment about what do we mean by intelligence? Right? And
7:30
Speaker 1: very simply intelligence. It's just the ability to learn, is
7:32
Speaker 1: to find patterns to extrapolate from them. Really, that's how you.
7:36
Speaker 1: But like a dog can learn. But a dog, you
7:39
Speaker 1: wouldn't say it's intelligent, would you? Absolutely? I would say
7:41
Speaker 1: a dog is intelligent. You can teach a dog, you
7:44
Speaker 1: can train a dog. It's more intelligent than a rock.
7:46
Speaker 1: But would you say by a lot, Oh my gosh,
7:50
Speaker 1: if you like, never interact with a dog. A dog
7:52
Speaker 1: is like a living sension being. It feels that experiences,
7:55
Speaker 1: It definitely learns. It can recognize you. I mean, dogs
7:59
Speaker 1: can do complicated of think. The dog is a perfect
8:01
Speaker 1: example it's you know, I wouldn't trust it to do
8:03
Speaker 1: my taxes. You know, I don't know compared to our
8:06
Speaker 1: tax accountant, did I do a pretty good job. I mean,
8:08
Speaker 1: you can say that's an intelligent dog, but you wouldn't say, like,
8:11
Speaker 1: that's the epitome of intelligence. I wouldn't say the dogs
8:14
Speaker 1: are the most intelligent beings in the universe. But that's
8:16
Speaker 1: what we're talking about. We're talking about do they have intelligence?
8:20
Speaker 1: Pretty example, because they can learn, you can train them,
8:22
Speaker 1: and you the cool thing about an intelligent being is
8:24
Speaker 1: that you can train it to do something even if
8:27
Speaker 1: you don't know how to do it. Say, for example,
8:29
Speaker 1: you want your dog to recognize you, right, but tear
8:33
Speaker 1: the face off anybody who tries to break into the house,
8:36
Speaker 1: right guard dog. Okay, so you can train a dog.
8:39
Speaker 1: You reward it when it does the right thing, and
8:41
Speaker 1: you punish it when it does the wrong thing. You
8:43
Speaker 1: don't know how to, like build a being that does that,
8:46
Speaker 1: that like recognizes your face and recognizes stranger's faces and
8:49
Speaker 1: makes these decisions. That's hard task, you know, it's not
8:52
Speaker 1: easy to do. But you can train a dog. A
8:55
Speaker 1: dog can learn how to solve this problem and all
8:58
Speaker 1: you need to do to train it is to reward
8:59
Speaker 1: it and punch it. So you're saying, just the ability
9:02
Speaker 1: to sort of learn from your mistakes or learn from
9:05
Speaker 1: your surroundings, that's what you would call intelligence. Yeah, And
9:08
Speaker 1: dogs have less of it than we do, and more
9:11
Speaker 1: of it than cats and mice, um. But they have
9:14
Speaker 1: some of it for sure, which is what makes them trainable.
9:17
Speaker 1: And you know, I wonder sometimes, because dogs can be trained, right,
9:20
Speaker 1: nobody ever trains their cat. What does that say about
9:23
Speaker 1: a cat's intelligence. I've always thought I love cats, but
9:27
Speaker 1: I've always thought dogs are probably smarter than cats because
9:30
Speaker 1: you can train them, right, Or maybe cats are more
9:33
Speaker 1: intelligent in that they're they're not they don't allow themselves
9:35
Speaker 1: to be trained by humans, right. And rocks, by that metric,
9:39
Speaker 1: are the most intelligent because they completely ignore you, right
9:41
Speaker 1: to see the fallacy of that argument right there. But
9:43
Speaker 1: I mean, maybe there's sort of sort of like a hump, right, like,
9:47
Speaker 1: as you get more intelligent, you're easily more trainable, trainable,
9:51
Speaker 1: trainable bout somebody. You get so intelligent that you rebel
9:54
Speaker 1: against your masters, And so how do you tell the
9:56
Speaker 1: difference between something that's totally unintelligent and something that's so
10:00
Speaker 1: intelligent and completely ignores you. Yeah, I don't know deep
10:04
Speaker 1: question or believes all the rocks are probably thinking about him.
10:08
Speaker 1: Well sure, I mean if he used to use the
10:09
Speaker 1: ability to listen to what I say as a benchpark
10:12
Speaker 1: of intelligence and then yeah, there's, um, something super intelligent
10:16
Speaker 1: could be just as smart as the rock. But obviously
10:18
Speaker 1: a cat is still making decisions and acting and you know,
10:21
Speaker 1: doing things, so it's intelligent. But maybe it's much more
10:24
Speaker 1: intelligent than a dog because it's it chooses not to
10:26
Speaker 1: listen to us. All right, I think we need to
10:28
Speaker 1: have a whole other podcast on who's smarter cats or dogs?
10:32
Speaker 1: And before we do that, we will collect some data
10:35
Speaker 1: to answer this question. Um, but I think with the
10:38
Speaker 1: question we're focusing on is what is artificial intelligence? So
10:41
Speaker 1: natural intelligence just the ability of an animal to learn.
10:44
Speaker 1: Artificial intelligence would be if something artificial that we create
10:48
Speaker 1: has that same property, the ability to change the way
10:52
Speaker 1: it processes things in response to what it sees about
10:56
Speaker 1: the world. Yeah, artificial intelligence is a very broad field
11:00
Speaker 1: with lots of elements that we couldn't cover in just
11:02
Speaker 1: one episode of a podcast. But let's just talk today
11:05
Speaker 1: about one important sub field of AI, which is machine learning,
11:09
Speaker 1: or more specifically, I would say, let's focus on training. Right,
11:12
Speaker 1: can you build something or something artificial that can be trained? Right?
11:18
Speaker 1: And uh? I think, I think let's talk for a
11:20
Speaker 1: moment about, you know, how how normal computers work, and
11:23
Speaker 1: then we can talk about how computers, smart computers, computers
11:26
Speaker 1: that can learn, computers with artificial intelligence, how they work.
11:29
Speaker 1: I think, I think what you keep talking about cats
11:31
Speaker 1: and dogs? All right, we'll talk about cats and dogs,
11:34
Speaker 1: but first let's take a quick break. Let's talk about
11:50
Speaker 1: what computers can do. Yeah, let's because computers are smart, right,
11:56
Speaker 1: you can program a computer to do smart things, but
11:58
Speaker 1: that doesn't necessarily mean it has intelligence. That's right. There's
12:02
Speaker 1: a difference between a computer that can do something and
12:04
Speaker 1: a computer that can learn something. Right. The way I
12:06
Speaker 1: think about non intelligent computers is the way you sort
12:09
Speaker 1: of think about machines. Right. You can tell them what
12:12
Speaker 1: to do, and they do exactly what you tell them,
12:16
Speaker 1: regardless of whether it's the right thing. You don't give
12:18
Speaker 1: them like a goal and say, hey, I just want
12:20
Speaker 1: the house to be clean. Figure it out. You have
12:22
Speaker 1: to tell them exactly what to do. You say, step
12:24
Speaker 1: over here, move the broom this way, step over there,
12:28
Speaker 1: you know, and if it's not cleaning the house because
12:29
Speaker 1: they're stuck on a corner or they're you know, fell
12:31
Speaker 1: on their on their butts or whatever, they don't care.
12:34
Speaker 1: They just tell you do exactly what you tell them
12:36
Speaker 1: to do. Have no sort of larger sense of what's important.
12:39
Speaker 1: It just follows instructions, just follows the recipe you gave it.
12:43
Speaker 1: That's right. It's like a like a wind up toy,
12:45
Speaker 1: you know, you wind it up, you give it some energy,
12:47
Speaker 1: and then it goes. And I really do think about
12:49
Speaker 1: computer programs the way you might think about little machines, right,
12:53
Speaker 1: because that's exactly what they are. They just execute a
12:55
Speaker 1: set of instructions. You know. It's just like a bunch
12:57
Speaker 1: of gears clicking into place, and they can't change the
13:00
Speaker 1: way they do that. And they do it regardless of
13:02
Speaker 1: whether it's the right thing, or whether it's effective or whatever.
13:04
Speaker 1: It just goes. Like your electric toothbrush, you know, you
13:09
Speaker 1: switch it on, and it's just it has a circuit
13:12
Speaker 1: that just has it moved the bristles back and forth,
13:15
Speaker 1: that's right, And it doesn't know if it's brushing your
13:17
Speaker 1: teeth or just flailing around in midair. Right, has no idea,
13:20
Speaker 1: It doesn't care, It doesn't think or feel whatever. It's
13:22
Speaker 1: just a machine, right, Thank god, it doesn't know. It
13:26
Speaker 1: would tell you to brush your teething like that's chocolate, Jorney,
13:32
Speaker 1: I'm tired of this? What is this gunk? Yeah? Exactly.
13:35
Speaker 1: And so that's what a sort of a normal machine is.
13:37
Speaker 1: That's what a classical computer program is, right. I think
13:40
Speaker 1: of it just the same way as you think of
13:41
Speaker 1: a physical machine. Okay, it's just doing what you the
13:44
Speaker 1: programmer told it to do. That's right, and it follows
13:48
Speaker 1: your instructions exactly. Um. Now, a computer that can learn
13:52
Speaker 1: is different, right. A computer that has artificial intelligence is
13:54
Speaker 1: different in this really important way because you can train
13:58
Speaker 1: it right, and you can train it because you we
14:01
Speaker 1: build these things to model the way that we work. Right.
14:05
Speaker 1: So for example, an AI program is sort of like, um,
14:09
Speaker 1: like a newborn baby can't do anything. Right. Say, there's
14:12
Speaker 1: a AI program, for example, that's supposed to recognize you
14:16
Speaker 1: when you come in the door. Right, is this Jorge
14:18
Speaker 1: or is this not Jorge? Right? Because it should only
14:21
Speaker 1: open the door for Jorge and not open the door
14:23
Speaker 1: for not Jorge. Okay, So when you created a new
14:27
Speaker 1: AI program, you would start out like just a newborn baby, okay,
14:30
Speaker 1: and like a blank slate, right, Like a blank slate,
14:33
Speaker 1: it would make random decisions, right you You you show
14:36
Speaker 1: it a face and it would say yes, it's Jorge,
14:39
Speaker 1: and then you say no, you were wrong or yes
14:41
Speaker 1: you were right. And then you would reward it if
14:44
Speaker 1: it does well, if it gives the right answer, and
14:46
Speaker 1: you would um, you would punish it if it doesn't, right,
14:49
Speaker 1: you would tell it. I mean, you don't actually punish
14:51
Speaker 1: it a reward. You just tell it, yes, you made
14:53
Speaker 1: the right um call this time, and know you made
14:56
Speaker 1: the wrong call this time this other time. But how
14:58
Speaker 1: is that different than the of calibrating something? Do you
15:01
Speaker 1: know what I mean? Like is calibration than artificial intelligence? Right? Well,
15:05
Speaker 1: the difference is calibration is like here, I have a tool.
15:09
Speaker 1: I know how to solve the problem. I just have
15:11
Speaker 1: to adjust it so that it does exactly the right thing.
15:13
Speaker 1: Um here, right, But you have a strategy that it's executing.
15:17
Speaker 1: You know. It's like, uh, you have a drill and
15:20
Speaker 1: you want it to drill fast or slow, and you
15:21
Speaker 1: know you know what how to solve the problem. You
15:23
Speaker 1: just you know it has to spin and screw and
15:25
Speaker 1: screw the thing in or whatever. It's just adjusting a
15:28
Speaker 1: knob here. You don't know how to solve the problem,
15:30
Speaker 1: and so you've given it a very very very flexible
15:34
Speaker 1: strategy on on the inside. You've given it like imagine
15:37
Speaker 1: something has like a thousand knobs. If you twist all
15:40
Speaker 1: these knobs, you could get all sorts of crazy um strategies. Right.
15:44
Speaker 1: So back to the example of like recognizing Joge or
15:47
Speaker 1: not when it when you tell it it's done, it's
15:50
Speaker 1: given the wrong answer, then it it adjusts those knobs.
15:54
Speaker 1: It says, well, let me try to tweak my strategy
15:56
Speaker 1: for deciding is this Jorge, and then we'll see how
15:58
Speaker 1: that goes. I think that's a key difference. It's the
16:01
Speaker 1: number of knobs, right, Like a drill with the knob
16:04
Speaker 1: for velocity. I mean, that is sort of trainable and
16:08
Speaker 1: you can set it up to be adaptive. But it's
16:10
Speaker 1: just one knob, and so it's not You wouldn't say
16:13
Speaker 1: it's intelligent. It's not intelligent. The spectrum of things it
16:16
Speaker 1: can do is very very seene right. But whereas like
16:19
Speaker 1: something that recognizes a phase. It needs to evaluate like
16:23
Speaker 1: a million pixels in a photo, right, and so for
16:26
Speaker 1: you to tweak how it evaluates each of those pixels,
16:29
Speaker 1: it would be really difficult for you to That's right.
16:31
Speaker 1: So imagine you know the machine here is is a
16:34
Speaker 1: camera in the door and takes a picture of who
16:35
Speaker 1: you got a million pixels and then it has to
16:37
Speaker 1: look at those pixels and decide is this Jorge or
16:39
Speaker 1: is this not whohe? And so there's does some calculation
16:43
Speaker 1: on that picture, right, and that calculation has millions of
16:45
Speaker 1: knobs on it, right, how much do I weigh this pixel?
16:48
Speaker 1: How much I weigh adjacent pixels? Do I look for
16:50
Speaker 1: his nose? Do I look for his hair? Do I
16:52
Speaker 1: look for the eyes? Right? So it's got some very
16:55
Speaker 1: very flexible thing inside of it that can do almost anything.
16:58
Speaker 1: And when you first start out, it's just random. So
17:00
Speaker 1: it's making ridiculous, terrible decisions. But the key the thing
17:03
Speaker 1: that models the learning, right, you know, just need artificial intelligence,
17:06
Speaker 1: You need artificial learning. The thing that models that learning
17:10
Speaker 1: is that when it gets the wrong answer, it knows
17:12
Speaker 1: how to adjust those knobs so that next time it's
17:15
Speaker 1: more correct. By itself. That's the key thing is that
17:18
Speaker 1: it learns by itself. It doesn't need you. They're sitting like, oh,
17:22
Speaker 1: you got this pixel wrong. You've got that pixel wrong.
17:25
Speaker 1: A tweak you know, tweak this one this way. It's
17:27
Speaker 1: really more like an autonomous automatic learning. That's right because
17:32
Speaker 1: you don't know how to adjust it. If you knew
17:33
Speaker 1: how to adjust it, you would just write that program. Right.
17:35
Speaker 1: The key is artificial intelligence is excellent when you don't
17:38
Speaker 1: know how to solve the problem, but you can define
17:41
Speaker 1: the problem. You can say, this is a picture of
17:43
Speaker 1: or and this is not learn a way to tell
17:46
Speaker 1: the difference, right. So you give it a very flexible
17:48
Speaker 1: strategy and then you try You let it try out,
17:52
Speaker 1: and when it gives it the wrong answer, you would
17:54
Speaker 1: let it adjust itself so that it gets closer and
17:56
Speaker 1: closer to giving the right answer right, And eventually these
17:59
Speaker 1: things will find the right setting for those millions of knobs,
18:02
Speaker 1: so that it's doing the right thing. It's saying, oh, look,
18:05
Speaker 1: this picture is a picture of Horhey, and it gets
18:07
Speaker 1: the right answer at the time. And when you give
18:09
Speaker 1: it a picture that's not a picture of Horne, and
18:11
Speaker 1: you give it Daniel, it says no, sorry, you're not
18:13
Speaker 1: getting in the house. Right. And I think a key
18:16
Speaker 1: thing is also that you, as a programmer, could not
18:19
Speaker 1: have predicted what all those knobs are going to be
18:22
Speaker 1: at the end, right, Like, it's such a big problem.
18:25
Speaker 1: There's a million knobs. There's no way that you can
18:28
Speaker 1: predict what those knots are going to be set to
18:30
Speaker 1: when it learns my face. That's right. It's perfect for
18:33
Speaker 1: really hard problems where we don't know how to solve it, right.
18:37
Speaker 1: We know how to describe the problem, but we don't
18:38
Speaker 1: know how to solve it. You're right, So if I
18:40
Speaker 1: already knew how to solve it, I could write a
18:42
Speaker 1: computer program that that and and tell it just like
18:45
Speaker 1: use this pixel, use that pixel, use this pixel. But
18:47
Speaker 1: I don't know how to solve that problem. It's really hard, right,
18:50
Speaker 1: But I can train a computer to figure it out,
18:53
Speaker 1: just the same way I can train a dog. Right.
18:56
Speaker 1: A dog can learn my face. Right, a dog recognizes
18:59
Speaker 1: the owner, and you know, happily licks its face when
19:03
Speaker 1: it comes home, and recognize that that when somebody's not
19:05
Speaker 1: its owner, and barks like crazy and choose its face
19:08
Speaker 1: off when it's not its owner, Right, remind me not
19:10
Speaker 1: to visit your house, Daniel, seems a little dangerous. So
19:18
Speaker 1: then a big thing is programming a structure in the
19:22
Speaker 1: in the software that is kind of open ended and malleable,
19:26
Speaker 1: do you know what I mean? Like, its something that
19:28
Speaker 1: is kind of unpredictable in a way that can learn.
19:32
Speaker 1: That's right. And that's the key thing is that some
19:34
Speaker 1: people might be thinking, well, hold on, you said that
19:35
Speaker 1: the computers can just do what they tell you, So
19:37
Speaker 1: how can a computer learn? Right? How is that possible?
19:40
Speaker 1: And the key is that it's an emergent property, right.
19:43
Speaker 1: Like the way that you write a computer program that
19:45
Speaker 1: can learn is you build all these little um calculating
19:49
Speaker 1: bits with knobs on them, right, and each bit just
19:51
Speaker 1: does what it's told. It takes some data, it makes
19:54
Speaker 1: a decision based on the value of the knob, and
19:56
Speaker 1: it sends out some data. And together all these things
19:59
Speaker 1: may a decision. Right. Each individual piece has no idea
20:02
Speaker 1: what it's doing. It's not smart or intelligent or making
20:04
Speaker 1: its own decisions. That doesn't have free will, right, But
20:07
Speaker 1: together they're doing something. And as you said earlier, they
20:11
Speaker 1: can change the way they behave. They can adjust these
20:13
Speaker 1: knobs themselves to improve their performance, and that's where the
20:16
Speaker 1: learning comes from. It's from that training. It gets external
20:19
Speaker 1: input and changes its behavior based on that expe. You're
20:23
Speaker 1: saying that the way to program these AI s is
20:25
Speaker 1: by h connecting a bunch of little simple things together
20:30
Speaker 1: to get something complex. Yes, And here it's important to
20:33
Speaker 1: remember that we are using neural networks as sort of
20:36
Speaker 1: a standing to represent a big broad set of strategies
20:39
Speaker 1: that are part of machine learning, right, And you don't
20:42
Speaker 1: know how to set them, how to put them together
20:44
Speaker 1: to get the right complex behavior. You just put them
20:46
Speaker 1: together and then you train it. Right, You say, well,
20:48
Speaker 1: I have I have something that's dumb, like a newborn baby,
20:51
Speaker 1: and I teach it how to do the thing that
20:53
Speaker 1: I want. But this really all sort of came about
20:55
Speaker 1: from brain research, right, Like, people were studying the brain
20:58
Speaker 1: and they figured out that our brain vans are made
21:00
Speaker 1: up of all these little simple units neurons, that's right.
21:03
Speaker 1: And each neuron is pretty simple, right, Like, it just
21:06
Speaker 1: takes a couple inputs and then it just outputs one signal.
21:10
Speaker 1: That's the really fascinating deep part about it, right, is
21:13
Speaker 1: that the structures we use in computers are modeled after
21:16
Speaker 1: what's actually happening in real brains. And when you say,
21:19
Speaker 1: inside your brain are a bunch of neurons, right, and
21:21
Speaker 1: these neurons taken some input and then if the input
21:25
Speaker 1: is right or above some a certain amount, then they
21:27
Speaker 1: send us some output, which is the input to the
21:29
Speaker 1: next neuron. Right, and your brain is basically just a
21:31
Speaker 1: big web of these things. Yeah, yeah, that's the right.
21:34
Speaker 1: That's the key is that these neurons, they're simple, but
21:37
Speaker 1: they're all sort of connected to each other. So it's
21:39
Speaker 1: a huge complex web going on inside your head. And
21:43
Speaker 1: when you're learning, what you're doing is you're kind of
21:45
Speaker 1: like shaping that web. You're saying, some connections. These connections
21:49
Speaker 1: are important for recognizing kore. Uh, these connections are important
21:54
Speaker 1: when you want to when it's not where kind of think.
21:56
Speaker 1: You know, your neurons can change. They have like basically
21:59
Speaker 1: knobs on them. I mean not physical literal knobs, but
22:01
Speaker 1: they have they can adjust. And so if you feel pain,
22:05
Speaker 1: you know, or you have an experience, then that changes
22:08
Speaker 1: the way your neurons work and it changes a little
22:11
Speaker 1: bit who you are and how you react to things.
22:13
Speaker 1: And that's why you know, newborn babies when they're born,
22:16
Speaker 1: they're not very responsive to stimulan because they're just still
22:20
Speaker 1: figuring figuring it out. You know, a newborn baby doesn't
22:22
Speaker 1: even know like, this is my arm, and I know
22:24
Speaker 1: how to control It has to learn all of these
22:26
Speaker 1: things by being trained, by having experiences. You know, it
22:31
Speaker 1: has the neurons, and the neurons are connected to each other,
22:34
Speaker 1: but it hasn't figure out how to use those connections.
22:37
Speaker 1: I try. It has to be trained to be useful
22:39
Speaker 1: and to interact with the world in any sort of
22:41
Speaker 1: meaningful way. Right, And so that's exactly the same sense.
22:45
Speaker 1: And it's fascinating that if you build a mathematical system.
22:48
Speaker 1: That's what a computer program is, basically a mathematical model
22:51
Speaker 1: of the processes that are happening in your brain. It
22:54
Speaker 1: performs in a very similar way, and it does this
22:56
Speaker 1: amazing thing, which is it adjusts itself to improve its
22:59
Speaker 1: performance on the task you've given it. Right, So it
23:02
Speaker 1: really is like a model of learning. And and when
23:04
Speaker 1: people saw this, they said, wow, I mean you you
23:06
Speaker 1: look inside the brain. You're wondering, like, how does thinking work?
23:09
Speaker 1: Where's the soul? Right? Where am I? You look inside
23:12
Speaker 1: the brain? All you see all these weird neurons connected
23:14
Speaker 1: to each other. You think, how could that possibly describe me?
23:17
Speaker 1: But when you build a model in a computer and
23:19
Speaker 1: it can do the things that you can do, which
23:21
Speaker 1: is you learn and develop and react and be trained
23:24
Speaker 1: and make bad jokes. Not yet, we have not yet
23:29
Speaker 1: solved the bad joke problem. Right, humans are still world
23:32
Speaker 1: champions in terms of bad jokes. We can still beat
23:35
Speaker 1: them at something that's right. And you know, and this
23:37
Speaker 1: is very useful because you want the systems around you
23:40
Speaker 1: to learn and to react, you know. And like if
23:44
Speaker 1: your phone, for example, it knows, hey, every time you
23:46
Speaker 1: open your phone, you start with Twitter, right, and so
23:49
Speaker 1: Twitter goes up on there on the most used app list, right.
23:53
Speaker 1: And that's not very complex artificial intelligence, but it is.
23:56
Speaker 1: And uh, and these sort of things are very helpful.
23:59
Speaker 1: Let's take a break. So that's kind of what makes
24:14
Speaker 1: AI is that a it can tackle complex problems that
24:17
Speaker 1: we don't wouldn't even know how to program something to do,
24:20
Speaker 1: and be that it changes and adapts and kind of
24:24
Speaker 1: it can get better, not just better, but also kind
24:26
Speaker 1: of adapt to the person using it. That's exactly right,
24:29
Speaker 1: exactly right. And so for example, sometimes you know Netflix
24:32
Speaker 1: uses AIS what program will you want to watch next? Well,
24:37
Speaker 1: you know, um, that's an AI. It's been trained. They
24:40
Speaker 1: feeded a bunch of examples. They say, Bob watched these
24:44
Speaker 1: five shows and then he watched this sixth show. But
24:47
Speaker 1: they gave the AI just the first five and they
24:49
Speaker 1: asked it predict what show he will watch next, and
24:52
Speaker 1: then they see if it doesn't do a a good job,
24:54
Speaker 1: and if it does a good job, you know, they reward.
24:55
Speaker 1: If it does a bad job, its its knobs to
24:58
Speaker 1: do better. Then when you're sitting there watching five hours
25:01
Speaker 1: of Netflix, it can do a pretty good job of
25:03
Speaker 1: predicting what you're gonna watch next because it's been trained
25:06
Speaker 1: on a lot of data. This is why people always
25:09
Speaker 1: talking about big data. Big data. These companies are gathering
25:11
Speaker 1: data about you so they can train their aiyes to
25:14
Speaker 1: learn your behavior and predict it. Except the problem is
25:18
Speaker 1: me and my spouse we share the same account and
25:21
Speaker 1: the same log in, So that's right, So it's learning
25:25
Speaker 1: some weird in your wife's brain. I have a very
25:29
Speaker 1: confused Netflix. I can or maybe it understands your marriage
25:33
Speaker 1: better than maybe's trying to tell us something. It's like
25:37
Speaker 1: you guys. Should you guys? His wife is out of town,
25:40
Speaker 1: he watches these shows. When she's in town, he has
25:43
Speaker 1: to watch these other shows. Oh alright, I know, break
25:53
Speaker 1: it down for us. How long before the aiyes take
25:57
Speaker 1: over the world? Um? Not very long actually, But you
26:02
Speaker 1: ask me different question earlier, which is is AI dangerous?
26:05
Speaker 1: And I think that has the two different questions there, right,
26:07
Speaker 1: I mean people are concerned. Some people are concerned. Yeah,
26:10
Speaker 1: I think people are concerned, and they're a good reason
26:12
Speaker 1: for it to be concerned. You know, One question is
26:14
Speaker 1: will AI develop its own autonomy and uh, and you
26:18
Speaker 1: know take over. That's a different question from are they dangerous,
26:21
Speaker 1: because you know they could take over and then take
26:23
Speaker 1: better care of the planet than we have, in which
26:25
Speaker 1: case you know they're not dangerous. They're benevolent dictators. I
26:28
Speaker 1: think the real question is will they take over? Will
26:31
Speaker 1: they become autonomous? We lose control of them somehow? And
26:34
Speaker 1: could they become smarter than us? So I see it's
26:37
Speaker 1: two issues. One could it could have developed a consciousness
26:41
Speaker 1: on its own and be is that consciousness good or
26:44
Speaker 1: bad for us? And it's an important question because as
26:47
Speaker 1: soon as you identify learning with consciousness, right, then you
26:51
Speaker 1: wonder about that and and this connection between the structure
26:54
Speaker 1: of AI and the structure our brain begs that question.
26:57
Speaker 1: You know, if you created, for example, and artificial or
27:00
Speaker 1: hate in the computer, if you built a set of
27:02
Speaker 1: neurons that mimic your brain, you know, would that simulation
27:06
Speaker 1: be alive? Would it be aware? Would it think I
27:08
Speaker 1: would have a first person experience? You know, that's a
27:15
Speaker 1: deep philosophical question, will never answer, right, And it's not
27:18
Speaker 1: really the important question. The important question is would we
27:21
Speaker 1: lose control of AI? Well? AI, because AI is something
27:24
Speaker 1: that can change and that can evolve, It can handle
27:26
Speaker 1: complex tasks. The question is can we lose control of it?
27:29
Speaker 1: And I think the answer to that one is definitely yes.
27:32
Speaker 1: We can lose control, meaning like, um, we'll give it
27:36
Speaker 1: control and then not be able to take it back. Yes, exactly,
27:39
Speaker 1: because the way AI is moving is that it can
27:42
Speaker 1: handle more and more complex tasks so that you don't
27:45
Speaker 1: have to be super specific about what it's doing, you know,
27:48
Speaker 1: like we have amazing natural language processing. Now you can
27:51
Speaker 1: say sort of vague things to your phone like hey, UM,
27:54
Speaker 1: set me up an appointment for tomorrow afternoon, right, and
27:57
Speaker 1: it will understand because it understands what your intent was, right,
28:00
Speaker 1: has to judge your intent and then execute it. It
28:02
Speaker 1: used to be you have to go into your computer
28:04
Speaker 1: and you have to press the keys in order to
28:06
Speaker 1: create that in your calendar. Now you can sort of
28:08
Speaker 1: talk to your phone and it will interpret what what
28:10
Speaker 1: you want and it will do that. And that's that's awesome.
28:13
Speaker 1: That's wonderful for human computer interactions that we can use
28:17
Speaker 1: our language to talk to them. We don't have to
28:19
Speaker 1: write computer code. That's a huge step forward, right, that
28:22
Speaker 1: people can construct machines using English rather than Python or
28:25
Speaker 1: C plus plus. Right, it's a big step forward. I
28:28
Speaker 1: think that's kind of what people find scary about a
28:30
Speaker 1: is that you can't really predict what it's going to do.
28:34
Speaker 1: I mean, it's sort of comedy gold when your kids
28:36
Speaker 1: are trying to talk to Alexa and ask it funny questions.
28:38
Speaker 1: But that's kind of what's fascinating about it, right, Like
28:40
Speaker 1: you ask it questions, do you give a task and
28:44
Speaker 1: you're so you really sort of don't know what it's
28:46
Speaker 1: going to do. That's exactly right, because it's making higher
28:49
Speaker 1: and higher level decisions, which makes it much more useful
28:51
Speaker 1: and much more intelligent. The same way when your kid
28:54
Speaker 1: grows up, right, when it's a When your kid is
28:56
Speaker 1: for you have to be very specific. You have to
28:58
Speaker 1: say things out loud which are ridiculous us, right, like
29:00
Speaker 1: don't put that finger in your nose. You know, we're like,
29:03
Speaker 1: uh oh, that's been on the floor, don't eat it. Right,
29:05
Speaker 1: you have to be really specific. When they're ten, you
29:07
Speaker 1: can say more general things and they'll understand, right, and
29:10
Speaker 1: learned their intelligence, like don't put both fingers in your nose.
29:15
Speaker 1: Only put one finger in your nose at a time,
29:17
Speaker 1: so you don't put your finger and your sisters and right, um,
29:20
Speaker 1: the same way as machines or artificial intelligence gets more intelligent,
29:24
Speaker 1: you can give it vagrant instructions and then it makes
29:27
Speaker 1: decisions based on its training. Right, and you don't really know,
29:31
Speaker 1: just like you can't really read what is in every
29:34
Speaker 1: neuron in another person. And then AI you sort of
29:37
Speaker 1: you don't know what's gonna happen, what's gonna come out exactly,
29:40
Speaker 1: So they're gonna start making decisions based on you know,
29:43
Speaker 1: still what we tell them to do. But you know
29:45
Speaker 1: what if you told your AI, you're like, hey, keep
29:47
Speaker 1: my kids safe right. I mean, imagine some future where
29:50
Speaker 1: you have an AI robot it's really smart, and you say, hey,
29:52
Speaker 1: keep my kids safe, and you come home and it's
29:54
Speaker 1: like lock them in the basement, right, and like, well, okay,
29:57
Speaker 1: they're safe. But it's sort of a monkey pause situation, right,
30:00
Speaker 1: Like you got exactly what you asked for, but you
30:02
Speaker 1: didn't really labor the right way and made different decisions. Right,
30:05
Speaker 1: So we still skip the question of whether aies can
30:07
Speaker 1: be you know, a chief consciousness and become its own
30:11
Speaker 1: kind of soul, have a soul. It doesn't seem like
30:14
Speaker 1: you think that's a relevant question. I think it's important
30:17
Speaker 1: because when AI gets to be super intelligent, it's gonna
30:20
Speaker 1: seem like it has a soul. They're gonna seem like people,
30:22
Speaker 1: and people don't wonder like do they have rights? Can
30:25
Speaker 1: you kill an AI? What? How can you just delete it?
30:28
Speaker 1: You know? Um, that's going to be a really interesting question.
30:31
Speaker 1: But that's again that's a whole question of philosophy that
30:33
Speaker 1: we could we could easily spend an hour on things
30:35
Speaker 1: a much more practical question, which is will we lose
30:37
Speaker 1: control of them whether or not they have first person
30:40
Speaker 1: experiences so they just seem to. It's important to think
30:42
Speaker 1: about whether we're gonna lose control. And there's two reasons
30:45
Speaker 1: why I think that we will. One is, computers are
30:48
Speaker 1: getting faster, really really quickly. Right, Every year, computers get
30:51
Speaker 1: faster and faster and smarter and smarter, and the scale
30:54
Speaker 1: is is is growing, right, So this thing is happening
30:57
Speaker 1: very quickly. But we're not right. We're not getting smarter. Right,
31:01
Speaker 1: human brain is not changing and evolving at a very
31:03
Speaker 1: rapid rate. Computers are, So they're catching up and the
31:06
Speaker 1: slope is steep. Right, you can just get bigger and
31:09
Speaker 1: bigger computers and teaming together and paralyze them and you
31:11
Speaker 1: can just keep going right, So eventually they'll definitely have
31:15
Speaker 1: enormous computing power with capabilities to do things we can't
31:17
Speaker 1: even imagine. And also being faster doesn't necessarily mean being smarter.
31:22
Speaker 1: You also need like more dated to train on it.
31:24
Speaker 1: And also being twice as fast doesn't mean being twice
31:27
Speaker 1: as smart. It's not linear. So you think that they
31:31
Speaker 1: will get more capable than us. But do you think
31:34
Speaker 1: we will ever seed control of really important things to Aiyes, Like, hey,
31:40
Speaker 1: here's the nuclear button, um only fire it if it's
31:43
Speaker 1: necessary exactly, Let's talk about weapons. Weapons is going to
31:51
Speaker 1: be what ends it because you know, for example, we
31:54
Speaker 1: already have drones, right, and we have drones with missiles
31:57
Speaker 1: on them, and these drones can kill people. They can,
32:00
Speaker 1: they can, you can. Some pilot somewhere is flying it.
32:02
Speaker 1: He's making a decision and he's gonna shoot dismissile to
32:05
Speaker 1: kill a person, right, right, But you know the enemy
32:08
Speaker 1: has drones, and pretty soon it's gonna be drone on
32:10
Speaker 1: drone warfare, right, And again, drones are gonna shoot each other,
32:13
Speaker 1: and at some point somebody's going to put an AI
32:16
Speaker 1: in their drone. Why because an AI can make the
32:19
Speaker 1: decision about shooting much faster than a human can, So
32:22
Speaker 1: which drone is gonna win? An AI will be a
32:25
Speaker 1: better fighter than a human fire yes, And so eventually
32:29
Speaker 1: these AI will be making kill decisions, right because the
32:32
Speaker 1: one that can make the decision faster, it's going to
32:34
Speaker 1: be the one that wins. And so I don't think
32:37
Speaker 1: it's gonna be very long before we have AI powered
32:40
Speaker 1: drones that are authorized to kill people. Right. This is
32:44
Speaker 1: a clear next step for the military. You know, like here,
32:47
Speaker 1: here's a picture of somebody we you think is a terrorist.
32:49
Speaker 1: If you spot them, just fire the missile. Don't bother checking, Yeah,
32:53
Speaker 1: don't bother checking with us. Right, that's a clear next step.
32:56
Speaker 1: So now you have AI that have the authority to
32:59
Speaker 1: kill people, and why because they've been tasked to, you know,
33:03
Speaker 1: take care of us or protect us or only only
33:05
Speaker 1: if you give it that permission though, right, Like, I
33:08
Speaker 1: mean that's a big ethical step to say, like, if
33:10
Speaker 1: you see him, shoot him. Yeah, But I don't think
33:12
Speaker 1: that's a big ethical step for the military. You know,
33:15
Speaker 1: the protocols for shooting somebody in the military. I mean,
33:17
Speaker 1: I'm not an expert on military protocols, but you know,
33:21
Speaker 1: our military kills a lot of people for you know,
33:24
Speaker 1: a lot of civilians get killed, right, and we decide
33:27
Speaker 1: it's okay. A lot of innocent people get killed for
33:30
Speaker 1: military purposes. And so I don't think it's too far
33:32
Speaker 1: before AI is making that decision. And then it's AI.
33:37
Speaker 1: It's weaponized AI, our weaponized AI versus their weaponized AI.
33:41
Speaker 1: And then it's an arms race, and then the most
33:43
Speaker 1: powerful army is gonna be the one that just makes
33:46
Speaker 1: it all of his decisions, and the generals just say
33:48
Speaker 1: defend us right, or respond if we're attacked, right, And
33:52
Speaker 1: then you basically handed over control of the weapons to
33:55
Speaker 1: the AI because the enemy has weaponized AI. But that
33:59
Speaker 1: doesn't mean that they're trilling us. I mean, we use
34:01
Speaker 1: them to protect us or to take away some decision
34:04
Speaker 1: making for it, but that doesn't mean that they're necessarily
34:07
Speaker 1: in control of us. And let's make sure not to
34:09
Speaker 1: be too alarmist here, of course, because people are working
34:12
Speaker 1: really hard to make sure that there are always ways
34:14
Speaker 1: for humans to override these systems. We would be different.
34:17
Speaker 1: That would be um, you know, it'd be like if
34:19
Speaker 1: a robot then turns the weapons inwards. That's another deal,
34:22
Speaker 1: I guess. Yeah. And of course AI researchers do their
34:25
Speaker 1: best to make sure that the AI systems are very
34:28
Speaker 1: well trained so that they do exactly what we want
34:31
Speaker 1: them to do. But they are complex and unpredictable, just
34:34
Speaker 1: like people are. Right, So this is a very interesting
34:37
Speaker 1: topic whether AI is dangerous or not. And I know
34:42
Speaker 1: Daniel that you you're sort of an expert in artificial
34:45
Speaker 1: intelligence because you use it in your particle physics research, right,
34:49
Speaker 1: you use machine learning, that's right. I wouldn't say I'm
34:51
Speaker 1: an expert I mean, I know something about it. Um.
34:53
Speaker 1: I've done some reading and I've used it, but I'm
34:55
Speaker 1: certainly not a deep expert in artificial intelligence itself, right,
34:59
Speaker 1: But you you know experts in your department, right, and
35:02
Speaker 1: you're in your campus, that's right. You see, I has
35:04
Speaker 1: an amazing computer science department and experts in machine learning.
35:08
Speaker 1: Some of the folks I actually collaborate with. When we're
35:10
Speaker 1: understanding the huge amounts of data from the large Hagon collider,
35:13
Speaker 1: we train machines to sift through that data and like
35:16
Speaker 1: look for the Higgs boson and learn to recognize new
35:19
Speaker 1: kinds of particles. It's really fun. And these guys know
35:22
Speaker 1: a lot about artificial intelligence more than I do. So
35:24
Speaker 1: I went over there and I asked them if they
35:26
Speaker 1: were worried about whether robots would take over the world,
35:29
Speaker 1: and what did the robots say the robots had taken
35:33
Speaker 1: over the professors and they answered for no. Um. First,
35:36
Speaker 1: here's professor Pierre Baldi, he's a distinguished professor on campus.
35:40
Speaker 1: And here's what he had to say. Potentially, yes, all
35:44
Speaker 1: very powerful technologies I think can pose such a threat,
35:49
Speaker 1: and all depends how they are deployed, how they are used.
35:52
Speaker 1: Et cetera. Right, you can say that nuclear technology pulls
35:55
Speaker 1: is such a threat and continues to pull such a threat.
36:00
Speaker 1: And I think AI, if used in the wrong way
36:05
Speaker 1: to pose a threat to mankind. Yes, the potential is
36:09
Speaker 1: there and so we should be careful. Um, right, So
36:12
Speaker 1: that was Professor Baldy, and then I also went down
36:15
Speaker 1: the hall and asked another colleagues cause I thought let's
36:17
Speaker 1: get more than one opinion, and so this is Professor
36:19
Speaker 1: Park Smith, also a professor of computer science at U
36:23
Speaker 1: c Irvine. I think the main threat with artificial intelligence
36:26
Speaker 1: going forward is not understanding how the black boxes work.
36:30
Speaker 1: And so I think not the typical sort of we're
36:35
Speaker 1: going to have robots taking over the world, but more
36:38
Speaker 1: the use of AI and situations where we're extrapolating beyond
36:42
Speaker 1: what it can do. And so I think we need
36:44
Speaker 1: to understand the limits of a I I think that's
36:47
Speaker 1: a threat, all right, So the answer is yes, Well,
36:51
Speaker 1: I think they're cautious, right, both of them think it's unpredictable.
36:55
Speaker 1: We don't know what's going to happen. We're creating a
36:57
Speaker 1: whole new kind of system and uh, and we may
37:00
Speaker 1: lose control of parts of it. On the other hand,
37:02
Speaker 1: you know it's likely for that to happen. You know,
37:04
Speaker 1: a lot of people are working really hard to make
37:06
Speaker 1: sure that AI will be contained and that in the
37:09
Speaker 1: end you can just pull the plug if the robot
37:11
Speaker 1: revolution starts, and so it is unpredictable. But also you know,
37:15
Speaker 1: the future is unpredictable, is always going to be unpredictable. Yeah,
37:18
Speaker 1: I feel like I thought it was interesting he said
37:21
Speaker 1: it is dangerous, but not more so than any other
37:24
Speaker 1: powerful technology. Yeah, that's a really interesting comment. It's true
37:28
Speaker 1: that any technology you can create could be used for
37:31
Speaker 1: good or for even if it's powerful, like I mean,
37:34
Speaker 1: not just like you know, wind up toy. Maybe it's
37:37
Speaker 1: not as dangerous. But but but I think that speaks
37:39
Speaker 1: to the kind of the power of AI, Like it
37:42
Speaker 1: really is maybe more powerful than we can handle. Yeah,
37:46
Speaker 1: and it's it's powerful in a special way run like
37:49
Speaker 1: nuclear weapons are powerful. Right, But in the end, a
37:52
Speaker 1: human is making that decision, and so you're giving humans
37:54
Speaker 1: a new kind of power, which is unpredictable. But here
37:57
Speaker 1: you're you're unleashing something, right, You're creating AI, and it's
38:01
Speaker 1: making its own decisions. Of course, it's making decisions based
38:04
Speaker 1: on what has been told to do. Right, you have
38:05
Speaker 1: to give it instructions. Still, you have to teach it um.
38:08
Speaker 1: But you can't predict what these complex systems are going
38:11
Speaker 1: to do in new circumstances and how they're gonna interpret
38:13
Speaker 1: your instructions. And of course there are a lot of
38:15
Speaker 1: AI supark people out there working hard to make sure
38:18
Speaker 1: that their boundaries and safeties um installed in all AI systems.
38:22
Speaker 1: But you know, I've seen Jurassic Park, you know, the
38:25
Speaker 1: lesson there they had fences. Lesson there, they had fences.
38:28
Speaker 1: We have fences. But then Jeff Goldbloom, you know, has
38:31
Speaker 1: a theory about chaos. Yeah, exactly. You know, these systems
38:34
Speaker 1: are hard to predict, and so I think we should
38:37
Speaker 1: be worried, but then we should respond to that worry
38:39
Speaker 1: with appropriate safeguards. You know, we should take this seriously,
38:42
Speaker 1: but not be overly alarmed. Right. Well, the other point
38:45
Speaker 1: that the other professor made is also interesting that he's
38:48
Speaker 1: saying some of the danger is in the fact that
38:51
Speaker 1: it's kind of like a black box, like we're trusting
38:55
Speaker 1: these things, but we don't really know what's going on inside.
38:58
Speaker 1: Like it's so complex they we we can't predict what
39:02
Speaker 1: it's gonna do. We can't maybe even deconstruct how it
39:06
Speaker 1: makes decisions. That's right, and uh, you know you train
39:09
Speaker 1: these systems are very complicated and you don't know how
39:11
Speaker 1: they're gonna respond to new circumstances. Right. It's same as
39:14
Speaker 1: when like training your dog, Like do you know how
39:16
Speaker 1: your dog makes a decision about who to bark end
39:18
Speaker 1: who not to. You try to train it, You try
39:21
Speaker 1: to give it instructions to try to make sure it
39:22
Speaker 1: knows how to how to handle it stuff in a
39:24
Speaker 1: new circumstances, but you can't honestly know what it's going
39:26
Speaker 1: to do at any given moment. Yeah, I'm definitely not
39:29
Speaker 1: visiting your house if you have dogs. I think about
39:34
Speaker 1: I think about AI. Again, not an expert, so maybe
39:38
Speaker 1: these are uninformed speculations, but I think about AI sort
39:40
Speaker 1: of like digital children. You know, like you raise your children,
39:44
Speaker 1: you know they're gonna take over one day because you know,
39:46
Speaker 1: you and I are going to get old and our
39:48
Speaker 1: kids are younger than we are, so eventually they will
39:50
Speaker 1: take over and you don't know what they're gonna do,
39:53
Speaker 1: and you raise them. You try to raise them in
39:55
Speaker 1: a way that they have values they make reasonable decisions,
39:58
Speaker 1: and you can sort of think about AI to same
40:00
Speaker 1: in way like you try to create this new generation
40:03
Speaker 1: of technology that's going to make its own decisions, but
40:05
Speaker 1: you try to teach it to make good decisions so
40:07
Speaker 1: that when you're in a home, right, it's making good
40:10
Speaker 1: choices for you. And I know that some folks out
40:12
Speaker 1: there think, well, you know, AI is never really going
40:14
Speaker 1: to be separate from humanity. There's not this like cognitive separation,
40:17
Speaker 1: Like you can just be part of who you are,
40:19
Speaker 1: the way your iPhone feels like part of who you are. Um,
40:22
Speaker 1: But we don't know necessarily if if that separation is
40:26
Speaker 1: going to be serious, you know, if these things really
40:28
Speaker 1: would be separate from us, or if they always just
40:29
Speaker 1: feel like an extension of ourselves. Well, until then, I
40:33
Speaker 1: think we should stick to regular dogs. Dogs. Yeah, But
40:40
Speaker 1: you know, I think about it sometimes the way I
40:42
Speaker 1: think about children, right, In the same way that you
40:44
Speaker 1: raise your children and they're gonna take over, right, It's
40:47
Speaker 1: gonna be some point when your children are in charge.
40:49
Speaker 1: You raise them to have values and to make good decisions,
40:52
Speaker 1: and you hope that when they take over, they're you know,
40:54
Speaker 1: looking after you. In the same way, we got to
40:56
Speaker 1: create these digital tools, and we've got to teach them
40:59
Speaker 1: to be Hey, we got to teach them what's important,
41:01
Speaker 1: and we got to teach them how to be responsible
41:03
Speaker 1: so that if they take over, you know that we
41:06
Speaker 1: hope they treat as well. Yeah, daddy, good daddy. Your
41:11
Speaker 1: parents don't put creator good Please don't bury me underground. Well,
41:22
Speaker 1: I personally am looking forward to a time when I have,
41:24
Speaker 1: like I don't have to think as much, where life
41:27
Speaker 1: is a little bit easier because we have these things
41:30
Speaker 1: making things easier for us. It could handle a lot
41:33
Speaker 1: of the drudgery and a lot of the logistics. You know,
41:36
Speaker 1: eventually you could have a car that drives itself and
41:38
Speaker 1: obeys your instructions. You can say like, hey, go pick
41:41
Speaker 1: up my kids from school, and he would know how
41:43
Speaker 1: to navigate and how to drive and recognize your children
41:45
Speaker 1: and how to get back home. And that's totally within
41:48
Speaker 1: the realm of possibility in a few years, right, And
41:51
Speaker 1: that's pretty awesome. It will offload a lot of work
41:53
Speaker 1: and logistics from beleaguered parents. I think you and I
41:55
Speaker 1: are in a pretty good position career wise, you know,
41:58
Speaker 1: Like I'm a cartoonist in your physicist. These are not
42:01
Speaker 1: um jobs that are going to be taken away by
42:04
Speaker 1: AI anytime soon. Hopefully have you not seen a our cartoons?
42:08
Speaker 1: They're pretty good man, all right, they you should like
42:11
Speaker 1: start a podcast instead of wearing of relying on your cartooning. Well,
42:15
Speaker 1: there is definitely that as a genre of humor. Like, hey,
42:18
Speaker 1: I put um so and so through an AI machine
42:20
Speaker 1: and look look at the crazy thing it came out with.
42:23
Speaker 1: Except those are all manufactured. None of those are no,
42:25
Speaker 1: those are real, None of those are real. Those are
42:29
Speaker 1: all made up. Well, that's good for humorist. So artificial
42:33
Speaker 1: intelligence is certainly a revolution in thinking and in computing,
42:37
Speaker 1: and it will definitely change the world. And so check
42:40
Speaker 1: back in in ten years to see if we've been
42:42
Speaker 1: replaced by robot Daniel and robot Warhead. Maybe we already
42:45
Speaker 1: are bump bump ball. So thanks everyone for listening to
42:49
Speaker 1: this episode of Daniel and Jorge Explain the Universe, and
42:53
Speaker 1: to listen tomorrow. Just say, Alexa, what's the best science
42:56
Speaker 1: podcast in the world. What's the third best? It's not
43:00
Speaker 1: a Catherine the world. If you still have a question
43:11
Speaker 1: after listening to all these explanations, please drop us a
43:14
Speaker 1: line we'd love to hear from you. You can find
43:17
Speaker 1: us at Facebook, Twitter, and Instagram at Daniel and Jorge
43:20
Speaker 1: That's One Word, or email us at Feedback at Daniel
43:24
Speaker 1: and Jorge dot com