Melanie Mitchell: Concepts, Analogies, Common Sense & Future of AI

Lex Fridman (01:14.900)

but it helps if you listen to the ad

Lex Fridman (01:17.060)

and support this podcast by trying out the product

Lex Fridman (01:19.660)

or service being advertised.

Melanie Mitchell (01:22.660)

This show is presented by Cash App,

Lex Fridman (01:24.940)

the number one finance app in the App Store.

Melanie Mitchell (01:27.540)

I personally use Cash App to send money to friends,

Lex Fridman (01:30.260)

but you can also use it to buy, sell,

Lex Fridman (01:32.620)

and deposit Bitcoin in just seconds.

Lex Fridman (01:35.020)

Cash App also has a new investing feature.

Melanie Mitchell (01:38.180)

You can buy fractions of a stock, say $1 worth,

Lex Fridman (01:41.120)

no matter what the stock price is.

Melanie Mitchell (01:43.360)

Broker services are provided by Cash App Investing,

Lex Fridman (01:46.180)

a subsidiary of Square and member SIPC.

Melanie Mitchell (01:49.940)

I'm excited to be working with Cash App

Lex Fridman (01:51.860)

to support one of my favorite organizations called First.

Melanie Mitchell (01:55.260)

Best known for their first robotics and Lego competitions.

Lex Fridman (01:58.900)

They educate and inspire hundreds of thousands of students

Melanie Mitchell (02:02.620)

in over 110 countries

Lex Fridman (02:04.420)

and have a perfect rating on Charity Navigator,

Melanie Mitchell (02:06.940)

which means that donated money is used

Lex Fridman (02:09.020)

to maximum effectiveness.

Melanie Mitchell (02:11.660)

When you get Cash App from the App Store or Google Play

Lex Fridman (02:14.900)

and use code LexPodcast,

Melanie Mitchell (02:17.300)

you'll get $10 and Cash App will also donate $10 to First,

Lex Fridman (02:21.340)

which again is an organization

Melanie Mitchell (02:23.220)

that I've personally seen inspire girls and boys

Lex Fridman (02:26.100)

to dream of engineering a better world.

Lex Fridman (02:28.980)

And now here's my conversation with Melanie Mitchell.

Lex Fridman (02:33.720)

The name of your new book is Artificial Intelligence,

Melanie Mitchell (02:36.860)

subtitle, A Guide for Thinking Humans.

Lex Fridman (02:39.700)

The name of this podcast is Artificial Intelligence.

Lex Fridman (02:42.960)

So let me take a step back

Lex Fridman (02:44.100)

and ask the old Shakespeare question about roses.

Lex Fridman (02:46.980)

And what do you think of the term artificial intelligence

Lex Fridman (02:51.100)

for our big and complicated and interesting field?

Melanie Mitchell (02:55.520)

I'm not crazy about the term.

Lex Fridman (02:57.900)

I think it has a few problems

Melanie Mitchell (03:01.260)

because it means so many different things

Lex Fridman (03:04.380)

to different people.

Lex Fridman (03:05.640)

And intelligence is one of those words

Lex Fridman (03:07.480)

that isn't very clearly defined either.

Melanie Mitchell (03:10.060)

There's so many different kinds of intelligence,

Lex Fridman (03:14.420)

degrees of intelligence, approaches to intelligence.

Melanie Mitchell (03:18.900)

John McCarthy was the one who came up with the term

Lex Fridman (03:21.740)

artificial intelligence.

Lex Fridman (03:23.240)

And from what I read,

Lex Fridman (03:24.340)

he called it that to differentiate it from cybernetics,

Melanie Mitchell (03:28.800)

which was another related movement at the time.

Lex Fridman (03:33.720)

And he later regretted calling it artificial intelligence.

Melanie Mitchell (03:39.260)

Herbert Simon was pushing for calling it

Lex Fridman (03:41.920)

complex information processing,

Melanie Mitchell (03:45.420)

which got nixed,

Lex Fridman (03:47.100)

but probably is equally vague, I guess.

Melanie Mitchell (03:52.140)

Is it the intelligence or the artificial

Lex Fridman (03:55.360)

in terms of words that is most problematic, would you say?

Melanie Mitchell (03:58.720)

Yeah, I think it's a little of both.

Lex Fridman (04:01.060)

But it has some good sides

Melanie Mitchell (04:02.960)

because I personally was attracted to the field

Lex Fridman (04:07.060)

because I was interested in phenomenon of intelligence.

Lex Fridman (04:11.280)

And if it was called complex information processing,

Lex Fridman (04:13.620)

maybe I'd be doing something wholly different now.

Lex Fridman (04:16.220)

What do you think of, I've heard the term used,

Lex Fridman (04:18.700)

cognitive systems, for example, so using cognitive.

Melanie Mitchell (04:22.760)

Yeah, I mean, cognitive has certain associations with it.

Lex Fridman (04:27.840)

And people like to separate things like cognition

Lex Fridman (04:31.020)

and perception, which I don't actually think are separate.

Lex Fridman (04:33.940)

But often people talk about cognition as being different

Melanie Mitchell (04:37.740)

from sort of other aspects of intelligence.

Lex Fridman (04:41.420)

It's sort of higher level.

Lex Fridman (04:42.740)

So to you, cognition is this broad,

Lex Fridman (04:44.660)

beautiful mess of things that encompasses the whole thing.

Melanie Mitchell (04:47.900)

Memory, perception.

Lex Fridman (04:48.740)

Yeah, I think it's hard to draw lines like that.

Melanie Mitchell (04:53.040)

When I was coming out of grad school in 1990,

Lex Fridman (04:56.620)

which is when I graduated,

Melanie Mitchell (04:58.380)

that was during one of the AI winters.

Lex Fridman (05:01.560)

And I was advised to not put AI,

Melanie Mitchell (05:05.140)

artificial intelligence on my CV,

Lex Fridman (05:06.780)

but instead call it intelligence systems.

Lex Fridman (05:09.240)

So that was kind of a euphemism, I guess.

Lex Fridman (05:14.880)

What about to stick briefly on terms and words,

Melanie Mitchell (05:20.640)

the idea of artificial general intelligence,

Lex Fridman (05:24.100)

or like Yann LeCun prefers human level intelligence,

Melanie Mitchell (05:29.560)

sort of starting to talk about ideas

Lex Fridman (05:32.860)

that achieve higher and higher levels of intelligence

Lex Fridman (05:37.720)

and somehow artificial intelligence seems to be a term

Lex Fridman (05:41.320)

used more for the narrow, very specific applications of AI

Lex Fridman (05:45.320)

and sort of what set of terms appeal to you

Lex Fridman (05:51.320)

to describe the thing that perhaps we strive to create?

Melanie Mitchell (05:56.000)

People have been struggling with this

Lex Fridman (05:57.400)

for the whole history of the field

Lex Fridman (06:00.200)

and defining exactly what it is that we're talking about.

Lex Fridman (06:03.400)

You know, John Searle had this distinction

Melanie Mitchell (06:05.640)

between strong AI and weak AI.

Lex Fridman (06:08.520)

And weak AI could be general AI,

Lex Fridman (06:10.480)

but his idea was strong AI was the view

Lex Fridman (06:14.600)

that a machine is actually thinking,

Melanie Mitchell (06:18.460)

that as opposed to simulating thinking

Lex Fridman (06:22.600)

or carrying out processes that we would call intelligent.

Melanie Mitchell (06:30.940)

At a high level, if you look at the founding

Lex Fridman (06:34.480)

of the field of McCarthy and Searle and so on,

Melanie Mitchell (06:38.960)

are we closer to having a better sense of that line

Lex Fridman (06:44.520)

between narrow, weak AI and strong AI?

Melanie Mitchell (06:50.640)

Yes, I think we're closer to having a better idea

Lex Fridman (06:55.440)

of what that line is.

Melanie Mitchell (06:58.000)

Early on, for example, a lot of people thought

Lex Fridman (07:01.680)

that playing chess would be, you couldn't play chess

Melanie Mitchell (07:06.880)

if you didn't have sort of general human level intelligence.

Lex Fridman (07:11.160)

And of course, once computers were able to play chess

Melanie Mitchell (07:13.960)

better than humans, that revised that view.

Lex Fridman (07:18.400)

And people said, okay, well, maybe now we have to revise

Lex Fridman (07:22.080)

what we think of intelligence as.

Lex Fridman (07:25.320)

And so that's kind of been a theme

Melanie Mitchell (07:28.760)

throughout the history of the field

Lex Fridman (07:29.920)

is that once a machine can do some task,

Melanie Mitchell (07:34.360)

we then have to look back and say, oh, well,

Lex Fridman (07:37.280)

that changes my understanding of what intelligence is

Melanie Mitchell (07:39.680)

because I don't think that machine is intelligent,

Lex Fridman (07:43.040)

at least that's not what I wanna call intelligence.

Lex Fridman (07:45.600)

So do you think that line moves forever

Lex Fridman (07:47.640)

or will we eventually really feel as a civilization

Lex Fridman (07:51.280)

like we've crossed the line if it's possible?

Lex Fridman (07:54.060)

It's hard to predict, but I don't see any reason

Lex Fridman (07:56.520)

why we couldn't in principle create something

Lex Fridman (08:00.280)

that we would consider intelligent.

Melanie Mitchell (08:03.160)

I don't know how we will know for sure.

Lex Fridman (08:07.280)

Maybe our own view of what intelligence is

Melanie Mitchell (08:10.520)

will be refined more and more

Lex Fridman (08:12.480)

until we finally figure out what we mean

Melanie Mitchell (08:14.480)

when we talk about it.

Lex Fridman (08:17.160)

But I think eventually we will create machines

Melanie Mitchell (08:22.160)

in a sense that have intelligence.

Lex Fridman (08:24.440)

They may not be the kinds of machines we have now.

Lex Fridman (08:28.080)

And one of the things that that's going to produce

Lex Fridman (08:32.000)

is making us sort of understand

Melanie Mitchell (08:34.920)

our own machine like qualities

Lex Fridman (08:38.560)

that we in a sense are mechanical

Melanie Mitchell (08:43.080)

in the sense that like cells,

Lex Fridman (08:45.900)

cells are kind of mechanical.

Melanie Mitchell (08:47.700)

They have algorithms, they process information by

Lex Fridman (08:52.700)

and somehow out of this mass of cells,

Melanie Mitchell (08:57.060)

we get this emergent property that we call intelligence.

Lex Fridman (09:01.280)

But underlying it is really just cellular processing

Lex Fridman (09:07.360)

and lots and lots and lots of it.

Lex Fridman (09:10.460)

Do you think we'll be able to,

Lex Fridman (09:12.260)

do you think it's possible to create intelligence

Lex Fridman (09:14.440)

without understanding our own mind?

Melanie Mitchell (09:16.460)

You said sort of in that process

Lex Fridman (09:18.220)

we'll understand more and more,

Lex Fridman (09:19.500)

but do you think it's possible to sort of create

Lex Fridman (09:23.020)

without really fully understanding

Melanie Mitchell (09:26.140)

from a mechanistic perspective,

Lex Fridman (09:27.580)

sort of from a functional perspective

Lex Fridman (09:29.180)

how our mysterious mind works?

Lex Fridman (09:32.820)

If I had to bet on it, I would say,

Melanie Mitchell (09:36.280)

no, we do have to understand our own minds

Lex Fridman (09:39.460)

at least to some significant extent.

Lex Fridman (09:42.860)

But I think that's a really big open question.

Lex Fridman (09:47.140)

I've been very surprised at how far kind of

Melanie Mitchell (09:49.660)

brute force approaches based on say big data

Lex Fridman (09:53.820)

and huge networks can take us.

Melanie Mitchell (09:57.420)

I wouldn't have expected that.

Lex Fridman (09:59.120)

And they have nothing to do with the way our minds work.

Lex Fridman (10:03.100)

So that's been surprising to me, so it could be wrong.

Lex Fridman (10:06.820)

To explore the psychological and the philosophical,

Lex Fridman (10:09.580)

do you think we're okay as a species

Lex Fridman (10:11.800)

with something that's more intelligent than us?

Lex Fridman (10:16.020)

Do you think perhaps the reason

Lex Fridman (10:18.380)

we're pushing that line further and further

Melanie Mitchell (10:20.620)

is we're afraid of acknowledging

Lex Fridman (10:23.300)

that there's something stronger, better,

Lex Fridman (10:25.820)

smarter than us humans?

Lex Fridman (10:29.020)

Well, I'm not sure we can define intelligence that way

Melanie Mitchell (10:31.620)

because smarter than is with respect to what,

Lex Fridman (10:40.540)

computers are already smarter than us in some areas.

Melanie Mitchell (10:42.860)

They can multiply much better than we can.

Lex Fridman (10:45.580)

They can figure out driving routes to take

Melanie Mitchell (10:50.220)

much faster and better than we can.

Lex Fridman (10:51.820)

They have a lot more information to draw on.

Melanie Mitchell (10:54.420)

They know about traffic conditions and all that stuff.

Lex Fridman (10:57.400)

So for any given particular task,

Melanie Mitchell (11:02.220)

sometimes computers are much better than we are

Lex Fridman (11:04.660)

and we're totally happy with that, right?

Melanie Mitchell (11:07.100)

I'm totally happy with that.

Lex Fridman (11:08.540)

It doesn't bother me at all.

Melanie Mitchell (11:10.540)

I guess the question is which things about our intelligence

Lex Fridman (11:15.460)

would we feel very sad or upset

Lex Fridman (11:20.660)

that machines had been able to recreate?

Lex Fridman (11:24.460)

So in the book, I talk about my former PhD advisor,

Melanie Mitchell (11:27.460)

Douglas Hofstadter,

Lex Fridman (11:29.020)

who encountered a music generation program.

Lex Fridman (11:32.960)

And that was really the line for him,

Lex Fridman (11:36.820)

that if a machine could create beautiful music,

Melanie Mitchell (11:40.120)

that would be terrifying for him

Lex Fridman (11:44.100)

because that is something he feels

Melanie Mitchell (11:46.340)

is really at the core of what it is to be human,

Lex Fridman (11:50.180)

creating beautiful music, art, literature.

Melanie Mitchell (11:56.260)

He doesn't like the fact that machines

Lex Fridman (11:59.740)

can recognize spoken language really well.

Melanie Mitchell (12:05.400)

He personally doesn't like using speech recognition,

Lex Fridman (12:09.560)

but I don't think it bothers him to his core

Melanie Mitchell (12:11.620)

because it's like, okay, that's not at the core of humanity.

Lex Fridman (12:15.780)

But it may be different for every person

Lex Fridman (12:17.940)

what really they feel would usurp their humanity.

Lex Fridman (12:25.180)

And I think maybe it's a generational thing also.

Melanie Mitchell (12:27.380)

Maybe our children or our children's children

Lex Fridman (12:30.700)

will be adapted, they'll adapt to these new devices

Melanie Mitchell (12:35.900)

that can do all these tasks and say,

Lex Fridman (12:38.640)

yes, this thing is smarter than me in all these areas,

Lex Fridman (12:41.500)

but that's great because it helps me.

Lex Fridman (12:46.980)

Looking at the broad history of our species,

Lex Fridman (12:50.500)

why do you think so many humans have dreamed

Lex Fridman (12:52.700)

of creating artificial life and artificial intelligence

Lex Fridman (12:55.340)

throughout the history of our civilization?

Lex Fridman (12:57.340)

So not just this century or the 20th century,

Lex Fridman (13:00.700)

but really throughout many centuries that preceded it?

Lex Fridman (13:06.420)

That's a really good question,

Lex Fridman (13:07.820)

and I have wondered about that.

Lex Fridman (13:09.380)

Because I myself was driven by curiosity

Melanie Mitchell (13:16.660)

about my own thought processes

Lex Fridman (13:18.740)

and thought it would be fantastic

Melanie Mitchell (13:20.820)

to be able to get a computer

Lex Fridman (13:22.100)

to mimic some of my thought processes.

Melanie Mitchell (13:26.100)

I'm not sure why we're so driven.

Lex Fridman (13:28.940)

I think we want to understand ourselves better

Lex Fridman (13:33.940)

and we also want machines to do things for us.

Lex Fridman (13:40.240)

But I don't know, there's something more to it

Melanie Mitchell (13:42.160)

because it's so deep in the kind of mythology

Lex Fridman (13:45.480)

or the ethos of our species.

Lex Fridman (13:49.120)

And I don't think other species have this drive.

Lex Fridman (13:52.320)

So I don't know.

Melanie Mitchell (13:53.560)

If you were to sort of psychoanalyze yourself

Lex Fridman (13:55.960)

in your own interest in AI, are you,

Lex Fridman (13:59.960)

what excites you about creating intelligence?

Lex Fridman (14:07.480)

You said understanding our own selves?

Melanie Mitchell (14:09.760)

Yeah, I think that's what drives me particularly.

Lex Fridman (14:13.800)

I'm really interested in human intelligence,

Lex Fridman (14:22.480)

but I'm also interested in the sort of the phenomenon

Lex Fridman (14:25.800)

of intelligence more generally.

Lex Fridman (14:28.320)

And I don't think humans are the only thing

Lex Fridman (14:29.760)

with intelligence, or even animals.

Lex Fridman (14:34.240)

But I think intelligence is a concept

Lex Fridman (14:39.660)

that encompasses a lot of complex systems.

Lex Fridman (14:43.760)

And if you think of things like insect colonies

Lex Fridman (14:47.720)

or cellular processes or the immune system

Melanie Mitchell (14:52.000)

or all kinds of different biological

Lex Fridman (14:54.200)

or even societal processes have as an emergent property

Melanie Mitchell (14:59.200)

some aspects of what we would call intelligence.

Lex Fridman (15:02.660)

They have memory, they process information,

Melanie Mitchell (15:05.140)

they have goals, they accomplish their goals, et cetera.

Lex Fridman (15:08.500)

And to me, the question of what is this thing

Melanie Mitchell (15:12.700)

we're talking about here was really fascinating to me.

Lex Fridman (15:17.980)

And exploring it using computers seem to be a good way

Melanie Mitchell (15:22.300)

to approach the question.

Lex Fridman (15:23.980)

So do you think kind of of intelligence,

Lex Fridman (15:26.100)

do you think of our universe as a kind of hierarchy

Lex Fridman (15:28.580)

of complex systems?

Lex Fridman (15:30.140)

And then intelligence is just the property of any,

Lex Fridman (15:33.480)

you can look at any level and every level

Melanie Mitchell (15:36.820)

has some aspect of intelligence.

Lex Fridman (15:39.260)

So we're just like one little speck

Melanie Mitchell (15:40.920)

in that giant hierarchy of complex systems.

Lex Fridman (15:44.420)

I don't know if I would say any system

Melanie Mitchell (15:47.580)

like that has intelligence, but I guess what I wanna,

Lex Fridman (15:52.200)

I don't have a good enough definition of intelligence

Melanie Mitchell (15:55.260)

to say that.

Lex Fridman (15:56.740)

So let me do sort of a multiple choice, I guess.

Lex Fridman (15:59.220)

So you said ant colonies.

Lex Fridman (16:02.500)

So are ant colonies intelligent?

Lex Fridman (16:04.500)

Are the bacteria in our body intelligent?

Lex Fridman (16:09.420)

And then going to the physics world molecules

Lex Fridman (16:13.820)

and the behavior at the quantum level of electrons

Lex Fridman (16:18.580)

and so on, are those kinds of systems,

Lex Fridman (16:21.580)

do they possess intelligence?

Lex Fridman (16:22.900)

Like where's the line that feels compelling to you?

Melanie Mitchell (16:27.660)

I don't know.

Lex Fridman (16:28.500)

I mean, I think intelligence is a continuum.

Lex Fridman (16:30.520)

And I think that the ability to, in some sense,

Lex Fridman (16:35.160)

have intention, have a goal,

Melanie Mitchell (16:37.780)

have some kind of self awareness is part of it.

Lex Fridman (16:45.260)

So I'm not sure if, you know,

Melanie Mitchell (16:47.820)

it's hard to know where to draw that line.

Lex Fridman (16:50.340)

I think that's kind of a mystery.

Lex Fridman (16:52.380)

But I wouldn't say that the planets orbiting the sun

Lex Fridman (16:59.220)

is an intelligent system.

Melanie Mitchell (17:01.800)

I mean, I would find that maybe not the right term

Lex Fridman (17:05.060)

to describe that.

Lex Fridman (17:06.180)

And there's all this debate in the field

Lex Fridman (17:09.140)

of like what's the right way to define intelligence?

Lex Fridman (17:12.560)

What's the right way to model intelligence?

Lex Fridman (17:15.300)

Should we think about computation?

Lex Fridman (17:16.760)

Should we think about dynamics?

Lex Fridman (17:18.140)

And should we think about free energy

Lex Fridman (17:21.700)

and all of that stuff?

Lex Fridman (17:23.520)

And I think that it's a fantastic time to be in the field

Melanie Mitchell (17:28.300)

because there's so many questions

Lex Fridman (17:30.340)

and so much we don't understand.

Melanie Mitchell (17:32.020)

There's so much work to do.

Lex Fridman (17:33.840)

So are we the most special kind of intelligence

Melanie Mitchell (17:38.340)

in this kind of, you said there's a bunch

Lex Fridman (17:41.580)

of different elements and characteristics

Melanie Mitchell (17:43.880)

of intelligence systems and colonies.

Lex Fridman (17:47.160)

Is human intelligence the thing in our brain?

Melanie Mitchell (17:53.080)

Is that the most interesting kind of intelligence

Lex Fridman (17:55.360)

in this continuum?

Melanie Mitchell (17:57.060)

Well, it's interesting to us because it is us.

Lex Fridman (18:01.440)

I mean, interesting to me, yes.

Lex Fridman (18:03.360)

And because I'm part of, you know, human.

Lex Fridman (18:06.680)

But to understanding the fundamentals of intelligence,

Lex Fridman (18:08.760)

what I'm getting at, is studying the human,

Lex Fridman (18:11.000)

is sort of, if everything we've talked about,

Lex Fridman (18:13.160)

what you talk about in your book,

Lex Fridman (18:14.360)

what just the AI field, this notion,

Melanie Mitchell (18:18.600)

yes, it's hard to define,

Lex Fridman (18:19.800)

but it's usually talking about something

Melanie Mitchell (18:22.440)

that's very akin to human intelligence.

Lex Fridman (18:24.480)

Yeah, to me it is the most interesting

Melanie Mitchell (18:26.840)

because it's the most complex, I think.

Lex Fridman (18:29.960)

It's the most self aware.

Melanie Mitchell (18:32.120)

It's the only system, at least that I know of,

Lex Fridman (18:34.960)

that reflects on its own intelligence.

Lex Fridman (18:38.600)

And you talk about the history of AI

Lex Fridman (18:41.040)

and us, in terms of creating artificial intelligence,

Melanie Mitchell (18:45.000)

being terrible at predicting the future

Lex Fridman (18:48.480)

with AI, with tech in general.

Lex Fridman (18:50.880)

So why do you think we're so bad at predicting the future?

Lex Fridman (18:56.400)

Are we hopelessly bad?

Lex Fridman (18:59.080)

So no matter what, whether it's this decade

Lex Fridman (19:01.960)

or the next few decades, every time we make a prediction,

Melanie Mitchell (19:04.880)

there's just no way of doing it well,

Lex Fridman (19:06.920)

or as the field matures, we'll be better and better at it.

Melanie Mitchell (19:10.880)

I believe as the field matures, we will be better.

Lex Fridman (19:13.760)

And I think the reason that we've had so much trouble

Melanie Mitchell (19:16.040)

is that we have so little understanding

Lex Fridman (19:18.400)

of our own intelligence.

Lex Fridman (19:20.320)

So there's the famous story about Marvin Minsky

Lex Fridman (19:29.120)

assigning computer vision as a summer project

Melanie Mitchell (19:32.600)

to his undergrad students.

Lex Fridman (19:34.640)

And I believe that's actually a true story.

Melanie Mitchell (19:36.660)

Yeah, no, there's a write up on it.

Lex Fridman (19:39.320)

Everyone should read.

Melanie Mitchell (19:40.300)

It's like a, I think it's like a proposal

Lex Fridman (19:43.520)

that describes everything that should be done

Melanie Mitchell (19:46.000)

in that project.

Lex Fridman (19:46.840)

It's hilarious because it, I mean, you could explain it,

Lex Fridman (19:49.920)

but from my recollection, it describes basically

Lex Fridman (19:52.600)

all the fundamental problems of computer vision,

Melanie Mitchell (19:55.000)

many of which still haven't been solved.

Lex Fridman (19:57.680)

Yeah, and I don't know how far

Melanie Mitchell (19:59.560)

they really expect it to get.

Lex Fridman (1:00:00.560)

to me at least.

Melanie Mitchell (1:00:01.660)

It's easy to criticize them.

Lex Fridman (1:00:05.200)

Well, look, like exactly what you're saying,

Melanie Mitchell (1:00:07.200)

mental models sort of almost put a psychology hat on,

Lex Fridman (1:00:11.680)

say, look, these networks are clearly not able

Melanie Mitchell (1:00:15.280)

to achieve what we humans do with forming mental models,

Lex Fridman (1:00:18.360)

analogy making and so on.

Lex Fridman (1:00:20.060)

But that doesn't mean that they fundamentally cannot do that.

Lex Fridman (1:00:23.780)

Like it's very difficult to say that.

Melanie Mitchell (1:00:25.680)

I mean, at least to me,

Lex Fridman (1:00:26.600)

do you have a notion that the learning approach is really,

Melanie Mitchell (1:00:29.840)

I mean, they're going to not only are they limited today,

Lex Fridman (1:00:34.000)

but they will forever be limited

Melanie Mitchell (1:00:37.360)

in being able to construct such mental models.

Lex Fridman (1:00:42.400)

I think the idea of the dynamic perception is key here.

Melanie Mitchell (1:00:47.400)

The idea that moving your eyes around and getting feedback.

Lex Fridman (1:00:54.040)

And that's something that, you know,

Melanie Mitchell (1:00:56.920)

there's been some models like that.

Lex Fridman (1:00:58.320)

There's certainly recurrent neural networks

Melanie Mitchell (1:01:00.640)

that operate over several time steps.

Lex Fridman (1:01:03.800)

But the problem is that the actual, the recurrence is,

Melanie Mitchell (1:01:07.800)

you know, basically the feedback is at the next time step

Lex Fridman (1:01:13.760)

is the entire hidden state of the network,

Melanie Mitchell (1:01:18.760)

which is, it turns out that that doesn't work very well.

Lex Fridman (1:01:25.480)

But see, the thing I'm saying is mathematically speaking,

Melanie Mitchell (1:01:29.480)

it has the information in that recurrence

Lex Fridman (1:01:33.560)

to capture everything, it just doesn't seem to work.

Lex Fridman (1:01:38.560)

So like, you know, it's like,

Lex Fridman (1:01:40.560)

it's the same Turing machine question, right?

Melanie Mitchell (1:01:44.560)

Yeah, maybe theoretically, computers,

Lex Fridman (1:01:49.560)

anything that's Turing, a universal Turing machine

Melanie Mitchell (1:01:53.560)

can be intelligent, but practically,

Lex Fridman (1:01:56.560)

the architecture might be very specific.

Melanie Mitchell (1:01:59.560)

Kind of architecture to be able to create it.

Lex Fridman (1:02:04.560)

So just, I guess it sort of asks almost the same question

Melanie Mitchell (1:02:09.560)

again is how big of a role do you think deep learning needs,

Lex Fridman (1:02:14.560)

will play or needs to play in this, in perception?

Melanie Mitchell (1:02:20.560)

I think that deep learning as it's currently,

Lex Fridman (1:02:24.560)

as it currently exists, you know, will place,

Melanie Mitchell (1:02:27.560)

that kind of thing will play some role.

Lex Fridman (1:02:31.560)

But I think that there's a lot more going on in perception.

Lex Fridman (1:02:36.560)

But who knows, you know, the definition of deep learning,

Lex Fridman (1:02:39.560)

I mean, it's pretty broad.

Melanie Mitchell (1:02:41.560)

It's kind of an umbrella for a lot of different things.

Lex Fridman (1:02:43.560)

So what I mean is purely sort of neural networks.

Melanie Mitchell (1:02:45.560)

Yeah, and a feed forward neural networks.

Lex Fridman (1:02:48.560)

Essentially, or there could be recurrence,

Lex Fridman (1:02:50.560)

but sometimes it feels like,

Lex Fridman (1:02:53.560)

for instance, I talked to Gary Marcus,

Melanie Mitchell (1:02:55.560)

it feels like the criticism of deep learning

Lex Fridman (1:02:58.560)

is kind of like us birds criticizing airplanes

Melanie Mitchell (1:03:02.560)

for not flying well, or that they're not really flying.

Lex Fridman (1:03:07.560)

Do you think deep learning,

Lex Fridman (1:03:10.560)

do you think it could go all the way?

Lex Fridman (1:03:12.560)

Like Yann LeCun thinks.

Lex Fridman (1:03:14.560)

Do you think that, yeah,

Lex Fridman (1:03:17.560)

the brute force learning approach can go all the way?

Melanie Mitchell (1:03:21.560)

I don't think so, no.

Lex Fridman (1:03:23.560)

I mean, I think it's an open question,

Lex Fridman (1:03:25.560)

but I tend to be on the innateness side

Lex Fridman (1:03:29.560)

that there's some things that we've been evolved

Melanie Mitchell (1:03:35.560)

to be able to learn,

Lex Fridman (1:03:39.560)

and that learning just can't happen without them.

Lex Fridman (1:03:44.560)

So one example, here's an example I had in the book

Lex Fridman (1:03:47.560)

that I think is useful to me, at least, in thinking about this.

Lex Fridman (1:03:51.560)

So this has to do with

Lex Fridman (1:03:54.560)

the Deep Minds Atari game playing program, okay?

Lex Fridman (1:03:59.560)

And it learned to play these Atari video games

Lex Fridman (1:04:02.560)

just by getting input from the pixels of the screen,

Lex Fridman (1:04:08.560)

and it learned to play the game Breakout

Lex Fridman (1:04:15.560)

1,000% better than humans, okay?

Melanie Mitchell (1:04:18.560)

That was one of their results, and it was great.

Lex Fridman (1:04:20.560)

And it learned this thing where it tunneled through the side

Melanie Mitchell (1:04:23.560)

of the bricks in the breakout game,

Lex Fridman (1:04:26.560)

and the ball could bounce off the ceiling

Lex Fridman (1:04:28.560)

and then just wipe out bricks.

Lex Fridman (1:04:30.560)

Okay, so there was a group who did an experiment

Melanie Mitchell (1:04:36.560)

where they took the paddle that you move with the joystick

Lex Fridman (1:04:41.560)

and moved it up two pixels or something like that.

Lex Fridman (1:04:45.560)

And then they looked at a deep Q learning system

Lex Fridman (1:04:49.560)

that had been trained on Breakout and said,

Melanie Mitchell (1:04:51.560)

could it now transfer its learning

Lex Fridman (1:04:53.560)

to this new version of the game?

Melanie Mitchell (1:04:55.560)

Of course, a human could, and it couldn't.

Lex Fridman (1:04:58.560)

Maybe that's not surprising, but I guess the point is

Melanie Mitchell (1:05:00.560)

it hadn't learned the concept of a paddle.

Lex Fridman (1:05:04.560)

It hadn't learned the concept of a ball

Melanie Mitchell (1:05:07.560)

or the concept of tunneling.

Lex Fridman (1:05:09.560)

It was learning something, you know, we looking at it

Melanie Mitchell (1:05:12.560)

kind of anthropomorphized it and said,

Lex Fridman (1:05:16.560)

oh, here's what it's doing in the way we describe it.

Lex Fridman (1:05:18.560)

But it actually didn't learn those concepts.

Lex Fridman (1:05:21.560)

And so because it didn't learn those concepts,

Melanie Mitchell (1:05:23.560)

it couldn't make this transfer.

Lex Fridman (1:05:26.560)

Yes, so that's a beautiful statement,

Lex Fridman (1:05:28.560)

but at the same time, by moving the paddle,

Lex Fridman (1:05:31.560)

we also anthropomorphize flaws to inject into the system

Melanie Mitchell (1:05:36.560)

that will then flip how impressed we are by it.

Lex Fridman (1:05:39.560)

What I mean by that is, to me, the Atari games were,

Melanie Mitchell (1:05:43.560)

to me, deeply impressive that that was possible at all.

Lex Fridman (1:05:48.560)

So like I have to first pause on that,

Lex Fridman (1:05:50.560)

and people should look at that, just like the game of Go,

Lex Fridman (1:05:53.560)

which is fundamentally different to me

Melanie Mitchell (1:05:55.560)

than what Deep Blue did.

Lex Fridman (1:05:59.560)

Even though there's still a tree search,

Melanie Mitchell (1:06:03.560)

it's just everything DeepMind has done in terms of learning,

Lex Fridman (1:06:08.560)

however limited it is, is still deeply surprising to me.

Melanie Mitchell (1:06:11.560)

Yeah, I'm not trying to say that what they did wasn't impressive.

Lex Fridman (1:06:15.560)

I think it was incredibly impressive.

Melanie Mitchell (1:06:17.560)

To me, it's interesting.

Lex Fridman (1:06:19.560)

Is moving the board just another thing that needs to be learned?

Lex Fridman (1:06:24.560)

So like we've been able to, maybe, maybe,

Lex Fridman (1:06:27.560)

been able to, through the current neural networks,

Melanie Mitchell (1:06:29.560)

learn very basic concepts

Lex Fridman (1:06:31.560)

that are not enough to do this general reasoning,

Lex Fridman (1:06:34.560)

and maybe with more data.

Lex Fridman (1:06:37.560)

I mean, the interesting thing about the examples

Melanie Mitchell (1:06:41.560)

that you talk about beautifully

Lex Fridman (1:06:44.560)

is it's often flaws of the data.

Melanie Mitchell (1:06:48.560)

Well, that's the question.

Lex Fridman (1:06:49.560)

I mean, I think that is the key question,

Melanie Mitchell (1:06:51.560)

whether it's a flaw of the data or not.

Lex Fridman (1:06:53.560)

Because the reason I brought up this example

Melanie Mitchell (1:06:56.560)

was because you were asking,

Lex Fridman (1:06:57.560)

do I think that learning from data could go all the way?

Lex Fridman (1:07:01.560)

And this was why I brought up the example,

Lex Fridman (1:07:04.560)

because I think, and this is not at all to take away

Melanie Mitchell (1:07:09.560)

from the impressive work that they did,

Lex Fridman (1:07:11.560)

but it's to say that when we look at what these systems learn,

Melanie Mitchell (1:07:18.560)

do they learn the things

Lex Fridman (1:07:21.560)

that we humans consider to be the relevant concepts?

Lex Fridman (1:07:25.560)

And in that example, it didn't.

Lex Fridman (1:07:29.560)

Sure, if you train it on moving, you know, the paddle being

Melanie Mitchell (1:07:34.560)

in different places, maybe it could deal with,

Lex Fridman (1:07:38.560)

maybe it would learn that concept.

Melanie Mitchell (1:07:40.560)

I'm not totally sure.

Lex Fridman (1:07:42.560)

But the question is, you know, scaling that up

Melanie Mitchell (1:07:44.560)

to more complicated worlds,

Lex Fridman (1:07:48.560)

to what extent could a machine

Melanie Mitchell (1:07:51.560)

that only gets this very raw data

Lex Fridman (1:07:54.560)

learn to divide up the world into relevant concepts?

Lex Fridman (1:07:58.560)

And I don't know the answer,

Lex Fridman (1:08:01.560)

but I would bet that without some innate notion

Melanie Mitchell (1:08:08.560)

that it can't do it.

Lex Fridman (1:08:10.560)

Yeah, 10 years ago, I 100% agree with you

Melanie Mitchell (1:08:12.560)

as the most experts in AI system,

Lex Fridman (1:08:15.560)

but now I have a glimmer of hope.

Melanie Mitchell (1:08:19.560)

Okay, I mean, that's fair enough.

Lex Fridman (1:08:21.560)

And I think that's what deep learning did in the community is,

Melanie Mitchell (1:08:24.560)

no, no, if I had to bet all my money,

Lex Fridman (1:08:26.560)

it's 100% deep learning will not take us all the way.

Lex Fridman (1:08:29.560)

But there's still other, it's still,

Lex Fridman (1:08:31.560)

I was so personally sort of surprised by the Atari games,

Melanie Mitchell (1:08:36.560)

by Go, by the power of self play of just game playing

Lex Fridman (1:08:40.560)

against each other that I was like many other times

Melanie Mitchell (1:08:44.560)

just humbled of how little I know about what's possible

Lex Fridman (1:08:48.560)

in this approach.

Melanie Mitchell (1:08:49.560)

Yeah, I think fair enough.

Lex Fridman (1:08:51.560)

Self play is amazingly powerful.

Lex Fridman (1:08:53.560)

And that goes way back to Arthur Samuel, right,

Lex Fridman (1:08:58.560)

with his checker plane program,

Melanie Mitchell (1:09:01.560)

which was brilliant and surprising that it did so well.

Lex Fridman (1:09:06.560)

So just for fun, let me ask you on the topic of autonomous vehicles.

Melanie Mitchell (1:09:10.560)

It's the area that I work at least these days most closely on,

Lex Fridman (1:09:15.560)

and it's also area that I think is a good example that you use

Melanie Mitchell (1:09:20.560)

as sort of an example of things we as humans

Lex Fridman (1:09:25.560)

don't always realize how hard it is to do.

Melanie Mitchell (1:09:28.560)

It's like the constant trend in AI,

Lex Fridman (1:09:30.560)

but the different problems that we think are easy

Melanie Mitchell (1:09:32.560)

when we first try them and then realize how hard it is.

Lex Fridman (1:09:36.560)

Okay, so you've talked about autonomous driving

Melanie Mitchell (1:09:41.560)

being a difficult problem, more difficult than we realize.

Lex Fridman (1:09:44.560)

Humans give it credit for it.

Lex Fridman (1:09:46.560)

Why is it so difficult?

Lex Fridman (1:09:47.560)

What are the most difficult parts in your view?

Melanie Mitchell (1:09:51.560)

I think it's difficult because of the world is so open ended

Lex Fridman (1:09:56.560)

as to what kinds of things can happen.

Lex Fridman (1:09:59.560)

So you have sort of what normally happens,

Lex Fridman (1:10:05.560)

which is just you drive along and nothing surprising happens,

Lex Fridman (1:10:09.560)

and autonomous vehicles can do,

Lex Fridman (1:10:12.560)

the ones we have now evidently can do really well

Melanie Mitchell (1:10:17.560)

on most normal situations as long as the weather

Lex Fridman (1:10:21.560)

is reasonably good and everything.

Lex Fridman (1:10:24.560)

But if some, we have this notion of edge cases

Lex Fridman (1:10:28.560)

or things in the tail of the distribution,

Melanie Mitchell (1:10:32.560)

we call it the long tail problem,

Lex Fridman (1:10:34.560)

which says that there's so many possible things

Melanie Mitchell (1:10:37.560)

that can happen that was not in the training data

Lex Fridman (1:10:41.560)

of the machine that it won't be able to handle it

Melanie Mitchell (1:10:47.560)

because it doesn't have common sense.

Lex Fridman (1:10:50.560)

Right, it's the old, the paddle moved problem.

Melanie Mitchell (1:10:54.560)

Yeah, it's the paddle moved problem, right.

Lex Fridman (1:10:57.560)

And so my understanding, and you probably are more

Melanie Mitchell (1:10:59.560)

of an expert than I am on this,

Lex Fridman (1:11:01.560)

is that current self driving car vision systems

Melanie Mitchell (1:11:07.560)

have problems with obstacles, meaning that they don't know

Lex Fridman (1:11:12.560)

which obstacles, which quote unquote obstacles

Melanie Mitchell (1:11:15.560)

they should stop for and which ones they shouldn't stop for.

Lex Fridman (1:11:18.560)

And so a lot of times I read that they tend to slam

Melanie Mitchell (1:11:21.560)

on the brakes quite a bit.

Lex Fridman (1:11:23.560)

And the most common accidents with self driving cars

Melanie Mitchell (1:11:27.560)

are people rear ending them because they were surprised.

Lex Fridman (1:11:31.560)

They weren't expecting the machine, the car to stop.

Melanie Mitchell (1:11:35.560)

Yeah, so there's a lot of interesting questions there.

Lex Fridman (1:11:38.560)

Whether, because you mentioned kind of two things.

Lex Fridman (1:11:42.560)

So one is the problem of perception, of understanding,

Lex Fridman (1:11:46.560)

of interpreting the objects that are detected correctly.

Lex Fridman (1:11:51.560)

And the other one is more like the policy,

Lex Fridman (1:11:54.560)

the action that you take, how you respond to it.

Lex Fridman (1:11:57.560)

So a lot of the car's braking is a kind of notion of,

Lex Fridman (1:12:02.560)

to clarify, there's a lot of different kind of things

Melanie Mitchell (1:12:05.560)

that are people calling autonomous vehicles.

Lex Fridman (1:12:07.560)

But the L4 vehicles with a safety driver are the ones

Melanie Mitchell (1:12:12.560)

like Waymo and Cruise and those companies,

Lex Fridman (1:12:15.560)

they tend to be very conservative and cautious.

Lex Fridman (1:12:18.560)

So they tend to be very, very afraid of hurting anything

Lex Fridman (1:12:21.560)

or anyone and getting in any kind of accidents.

Lex Fridman (1:12:24.560)

So their policy is very kind of, that results

Lex Fridman (1:12:28.560)

in being exceptionally responsive to anything

Lex Fridman (1:12:31.560)

that could possibly be an obstacle, right?

Lex Fridman (1:12:33.560)

Right, which the human drivers around it,

Melanie Mitchell (1:12:38.560)

it behaves unpredictably.

Lex Fridman (1:12:41.560)

Yeah, that's not a very human thing to do, caution.

Melanie Mitchell (1:12:43.560)

That's not the thing we're good at, especially in driving.

Lex Fridman (1:12:46.560)

We're in a hurry, often angry and et cetera,

Melanie Mitchell (1:12:49.560)

especially in Boston.

Lex Fridman (1:12:51.560)

And then there's sort of another, and a lot of times,

Melanie Mitchell (1:12:55.560)

machine learning is not a huge part of that.

Lex Fridman (1:12:57.560)

It's becoming more and more unclear to me

Lex Fridman (1:13:00.560)

how much sort of speaking to public information

Lex Fridman (1:13:05.560)

because a lot of companies say they're doing deep learning

Lex Fridman (1:13:08.560)

and machine learning just to attract good candidates.

Lex Fridman (1:13:12.560)

The reality is in many cases,

Melanie Mitchell (1:13:14.560)

it's still not a huge part of the perception.

Lex Fridman (1:13:18.560)

There's LiDAR and there's other sensors

Melanie Mitchell (1:13:20.560)

that are much more reliable for obstacle detection.

Lex Fridman (1:13:23.560)

And then there's Tesla approach, which is vision only.

Lex Fridman (1:13:27.560)

And there's, I think a few companies doing that,

Lex Fridman (1:13:30.560)

but Tesla most sort of famously pushing that forward.

Lex Fridman (1:13:32.560)

And that's because the LiDAR is too expensive, right?

Lex Fridman (1:13:35.560)

Well, I mean, yes, but I would say

Melanie Mitchell (1:13:40.560)

if you were to for free give to every Tesla vehicle,

Lex Fridman (1:13:44.560)

I mean, Elon Musk fundamentally believes

Melanie Mitchell (1:13:47.560)

that LiDAR is a crutch, right, famously said that.

Lex Fridman (1:13:50.560)

That if you want to solve the problem of machine learning,

Melanie Mitchell (1:13:55.560)

LiDAR should not be the primary sensor is the belief.

Lex Fridman (1:14:00.560)

The camera contains a lot more information.

Lex Fridman (1:14:04.560)

So if you want to learn, you want that information.

Lex Fridman (1:14:08.560)

But if you want to not to hit obstacles, you want LiDAR, right?

Melanie Mitchell (1:14:13.560)

Sort of it's this weird trade off because yeah,

Lex Fridman (1:14:18.560)

sort of what Tesla vehicles have a lot of,

Melanie Mitchell (1:14:21.560)

which is really the thing, the fallback,

Lex Fridman (1:14:26.560)

the primary fallback sensor is radar,

Melanie Mitchell (1:14:29.560)

which is a very crude version of LiDAR.

Lex Fridman (1:14:32.560)

It's a good detector of obstacles

Lex Fridman (1:14:34.560)

except when those things are standing, right?

Lex Fridman (1:14:37.560)

The stopped vehicle.

Melanie Mitchell (1:14:39.560)

Right, that's why it had problems

Lex Fridman (1:14:41.560)

with crashing into stop fire trucks.

Melanie Mitchell (1:14:43.560)

Stop fire trucks, right.

Lex Fridman (1:14:44.560)

So the hope there is that the vision sensor

Melanie Mitchell (1:14:47.560)

would somehow catch that.

Lex Fridman (1:14:49.560)

And for, there's a lot of problems with perception.

Melanie Mitchell (1:14:54.560)

They are doing actually some incredible stuff in the,

Lex Fridman (1:15:00.560)

almost like an active learning space

Melanie Mitchell (1:15:02.560)

where it's constantly taking edge cases and pulling back in.

Lex Fridman (1:15:06.560)

There's this data pipeline.

Melanie Mitchell (1:15:08.560)

Another aspect that is really important

Lex Fridman (1:15:12.560)

that people are studying now is called multitask learning,

Melanie Mitchell (1:15:15.560)

which is sort of breaking apart this problem,

Lex Fridman (1:15:18.560)

whatever the problem is, in this case driving,

Melanie Mitchell (1:15:20.560)

into dozens or hundreds of little problems

Lex Fridman (1:15:24.560)

that you can turn into learning problems.

Lex Fridman (1:15:26.560)

So this giant pipeline, it's kind of interesting.

Lex Fridman (1:15:30.560)

I've been skeptical from the very beginning,

Lex Fridman (1:15:33.560)

but become less and less skeptical over time

Lex Fridman (1:15:35.560)

how much of driving can be learned.

Melanie Mitchell (1:15:37.560)

I still think it's much farther

Lex Fridman (1:15:39.560)

than the CEO of that particular company thinks it will be,

Lex Fridman (1:15:44.560)

but it's constantly surprising that

Lex Fridman (1:15:48.560)

through good engineering and data collection

Lex Fridman (1:15:51.560)

and active selection of data,

Lex Fridman (1:15:53.560)

how you can attack that long tail.

Lex Fridman (1:15:56.560)

And it's an interesting open question

Lex Fridman (1:15:58.560)

that you're absolutely right.

Melanie Mitchell (1:15:59.560)

There's a much longer tail

Lex Fridman (1:16:01.560)

and all these edge cases that we don't think about,

Lex Fridman (1:16:04.560)

but it's a fascinating question

Lex Fridman (1:16:06.560)

that applies to natural language and all spaces.

Lex Fridman (1:16:09.560)

How big is that long tail?

Lex Fridman (1:16:12.560)

And I mean, not to linger on the point,

Lex Fridman (1:16:16.560)

but what's your sense in driving

Lex Fridman (1:16:19.560)

in these practical problems of the human experience?

Lex Fridman (1:16:24.560)

Can it be learned?

Lex Fridman (1:16:26.560)

So the current, what are your thoughts of sort of

Melanie Mitchell (1:16:28.560)

Elon Musk thought, let's forget the thing that he says

Lex Fridman (1:16:31.560)

it'd be solved in a year,

Lex Fridman (1:16:33.560)

but can it be solved in a reasonable timeline

Lex Fridman (1:16:38.560)

or do fundamentally other methods need to be invented?

Lex Fridman (1:16:41.560)

So I don't, I think that ultimately driving,

Lex Fridman (1:16:47.560)

so it's a trade off in a way,

Melanie Mitchell (1:16:50.560)

being able to drive and deal with any situation that comes up

Lex Fridman (1:16:56.560)

does require kind of full human intelligence.

Lex Fridman (1:16:59.560)

And even in humans aren't intelligent enough to do it

Lex Fridman (1:17:02.560)

because humans, I mean, most human accidents

Melanie Mitchell (1:17:06.560)

are because the human wasn't paying attention

Lex Fridman (1:17:09.560)

or the humans drunk or whatever.

Lex Fridman (1:17:11.560)

And not because they weren't intelligent enough.

Lex Fridman (1:17:13.560)

And not because they weren't intelligent enough, right.

Melanie Mitchell (1:17:17.560)

Whereas the accidents with autonomous vehicles

Lex Fridman (1:17:23.560)

is because they weren't intelligent enough.

Melanie Mitchell (1:17:25.560)

They're always paying attention.

Lex Fridman (1:17:26.560)

Yeah, they're always paying attention.

Lex Fridman (1:17:27.560)

So it's a trade off, you know,

Lex Fridman (1:17:29.560)

and I think that it's a very fair thing to say

Melanie Mitchell (1:17:32.560)

that autonomous vehicles will be ultimately safer than humans

Lex Fridman (1:17:37.560)

because humans are very unsafe.

Melanie Mitchell (1:17:39.560)

It's kind of a low bar.

Lex Fridman (1:17:42.560)

But just like you said, I think humans got a better rap, right.

Melanie Mitchell (1:17:48.560)

Because we're really good at the common sense thing.

Lex Fridman (1:17:50.560)

Yeah, we're great at the common sense thing.

Melanie Mitchell (1:17:52.560)

We're bad at the paying attention thing.

Lex Fridman (1:17:53.560)

Paying attention thing, right.

Melanie Mitchell (1:17:54.560)

Especially when we're, you know, driving is kind of boring

Lex Fridman (1:17:56.560)

and we have these phones to play with and everything.

Lex Fridman (1:17:59.560)

But I think what's going to happen is that

Lex Fridman (1:18:06.560)

for many reasons, not just AI reasons,

Lex Fridman (1:18:09.560)

but also like legal and other reasons,

Lex Fridman (1:18:12.560)

that the definition of self driving is going to change

Melanie Mitchell (1:18:17.560)

or autonomous is going to change.

Lex Fridman (1:18:19.560)

It's not going to be just,

Melanie Mitchell (1:18:23.560)

I'm going to go to sleep in the back

Lex Fridman (1:18:24.560)

and you just drive me anywhere.

Melanie Mitchell (1:18:27.560)

It's going to be more certain areas are going to be instrumented

Lex Fridman (1:18:34.560)

to have the sensors and the mapping

Lex Fridman (1:18:37.560)

and all of the stuff you need for,

Lex Fridman (1:18:39.560)

that the autonomous cars won't have to have full common sense

Lex Fridman (1:18:43.560)

and they'll do just fine in those areas

Lex Fridman (1:18:46.560)

as long as pedestrians don't mess with them too much.

Melanie Mitchell (1:18:49.560)

That's another question.

Lex Fridman (1:18:51.560)

That's right.

Lex Fridman (1:18:52.560)

But I don't think we will have fully autonomous self driving

Lex Fridman (1:18:59.560)

in the way that like most,

Melanie Mitchell (1:19:01.560)

the average person thinks of it for a very long time.

Lex Fridman (1:19:04.560)

And just to reiterate, this is the interesting open question

Melanie Mitchell (1:19:09.560)

that I think I agree with you on,

Lex Fridman (1:19:11.560)

is to solve fully autonomous driving,

Melanie Mitchell (1:19:14.560)

you have to be able to engineer in common sense.

Lex Fridman (1:19:17.560)

Yes.

Melanie Mitchell (1:19:19.560)

I think it's an important thing to hear and think about.

Lex Fridman (1:19:23.560)

I hope that's wrong, but I currently agree with you

Melanie Mitchell (1:19:27.560)

that unfortunately you do have to have, to be more specific,

Lex Fridman (1:19:32.560)

sort of these deep understandings of physics

Lex Fridman (1:19:35.560)

and of the way this world works and also the human dynamics.

Lex Fridman (1:19:39.560)

Like you mentioned, pedestrians and cyclists,

Melanie Mitchell (1:19:41.560)

actually that's whatever that nonverbal communication

Lex Fridman (1:19:45.560)

as some people call it,

Melanie Mitchell (1:19:46.560)

there's that dynamic that is also part of this common sense.

Lex Fridman (1:19:50.560)

Right.

Lex Fridman (1:19:51.560)

And we humans are pretty good at predicting

Lex Fridman (1:19:55.560)

what other humans are going to do.

Lex Fridman (1:19:57.560)

And how our actions impact the behaviors

Lex Fridman (1:20:00.560)

of this weird game theoretic dance that we're good at somehow.

Lex Fridman (1:20:05.560)

And the funny thing is,

Lex Fridman (1:20:07.560)

because I've watched countless hours of pedestrian video

Lex Fridman (1:20:11.560)

and talked to people,

Lex Fridman (1:20:12.560)

we humans are also really bad at articulating

Melanie Mitchell (1:20:15.560)

the knowledge we have.

Lex Fridman (1:20:16.560)

Right.

Melanie Mitchell (1:20:17.560)

Which has been a huge challenge.

Lex Fridman (1:20:19.560)

Yes.

Lex Fridman (1:20:20.560)

So you've mentioned embodied intelligence.

Lex Fridman (1:20:23.560)

What do you think it takes to build a system

Lex Fridman (1:20:25.560)

of human level intelligence?

Lex Fridman (1:20:27.560)

Does it need to have a body?

Melanie Mitchell (1:20:29.560)

I'm not sure, but I'm coming around to that more and more.

Lex Fridman (1:20:34.560)

And what does it mean to be,

Melanie Mitchell (1:20:36.560)

I don't mean to keep bringing up Yann LeCun.

Lex Fridman (1:20:40.560)

He looms very large.

Melanie Mitchell (1:20:42.560)

Well, he certainly has a large personality.

Lex Fridman (1:20:45.560)

Yes.

Melanie Mitchell (1:20:46.560)

He thinks that the system needs to be grounded,

Lex Fridman (1:20:49.560)

meaning he needs to sort of be able to interact with reality,

Lex Fridman (1:20:53.560)

but doesn't think it necessarily needs to have a body.

Lex Fridman (1:20:56.560)

So when you think of...

Lex Fridman (1:20:57.560)

So what's the difference?

Lex Fridman (1:20:58.560)

I guess I want to ask,

Melanie Mitchell (1:21:00.560)

when you mean body,

Lex Fridman (1:21:01.560)

do you mean you have to be able to play with the world?

Melanie Mitchell (1:21:04.560)

Or do you also mean like there's a body

Lex Fridman (1:21:06.560)

that you have to preserve?

Melanie Mitchell (1:21:10.560)

Oh, that's a good question.

Lex Fridman (1:21:11.560)

I haven't really thought about that,

Lex Fridman (1:21:13.560)

but I think both, I would guess.

Lex Fridman (1:21:15.560)

Because I think intelligence,

Melanie Mitchell (1:21:23.560)

it's so hard to separate it from our desire

Lex Fridman (1:21:29.560)

for self preservation,

Melanie Mitchell (1:21:31.560)

our emotions,

Lex Fridman (1:21:34.560)

all that non rational stuff

Melanie Mitchell (1:21:37.560)

that kind of gets in the way of logical thinking.

Lex Fridman (1:21:43.560)

Because the way,

Melanie Mitchell (1:21:46.560)

if we're talking about human intelligence

Lex Fridman (1:21:48.560)

or human level intelligence, whatever that means,

Melanie Mitchell (1:21:51.560)

a huge part of it is social.

Lex Fridman (1:21:55.560)

We were evolved to be social

Lex Fridman (1:21:58.560)

and to deal with other people.

Lex Fridman (1:22:01.560)

And that's just so ingrained in us

Melanie Mitchell (1:22:05.560)

that it's hard to separate intelligence from that.

Lex Fridman (1:22:09.560)

I think AI for the last 70 years

Melanie Mitchell (1:22:14.560)

or however long it's been around,

Lex Fridman (1:22:16.560)

it has largely been separated.

Melanie Mitchell (1:22:18.560)

There's this idea that there's like,

Lex Fridman (1:22:20.560)

it's kind of very Cartesian.

Melanie Mitchell (1:22:23.560)

There's this thinking thing that we're trying to create,

Lex Fridman (1:22:27.560)

but we don't care about all this other stuff.

Lex Fridman (1:22:30.560)

And I think the other stuff is very fundamental.

Lex Fridman (1:22:34.560)

So there's idea that things like emotion

Melanie Mitchell (1:22:37.560)

can get in the way of intelligence.

Lex Fridman (1:22:40.560)

As opposed to being an integral part of it.

Melanie Mitchell (1:22:42.560)

Integral part of it.

Lex Fridman (1:22:43.560)

So, I mean, I'm Russian,

Lex Fridman (1:22:45.560)

so romanticize the notions of emotion and suffering

Lex Fridman (1:22:48.560)

and all that kind of fear of mortality,

Melanie Mitchell (1:22:51.560)

those kinds of things.

Lex Fridman (1:22:52.560)

So in AI, especially.

Lex Fridman (1:22:56.560)

By the way, did you see that?

Lex Fridman (1:22:57.560)

There was this recent thing going around the internet.

Melanie Mitchell (1:23:00.560)

Some, I think he's a Russian or some Slavic

Lex Fridman (1:23:03.560)

had written this thing,

Melanie Mitchell (1:23:05.560)

anti the idea of super intelligence.

Lex Fridman (1:23:08.560)

I forgot, maybe he's Polish.

Melanie Mitchell (1:23:10.560)

Anyway, so it all these arguments

Lex Fridman (1:23:12.560)

and one was the argument from Slavic pessimism.

Melanie Mitchell (1:23:15.560)

My favorite.

Lex Fridman (1:23:19.560)

Do you remember what the argument is?

Melanie Mitchell (1:23:21.560)

It's like nothing ever works.

Lex Fridman (1:23:23.560)

Everything sucks.

Lex Fridman (1:23:27.560)

So what do you think is the role?

Lex Fridman (1:23:29.560)

Like that's such a fascinating idea

Melanie Mitchell (1:23:31.560)

that what we perceive as sort of the limits of the human mind,

Lex Fridman (1:23:38.560)

which is emotion and fear and all those kinds of things

Melanie Mitchell (1:23:42.560)

are integral to intelligence.

Lex Fridman (1:23:45.560)

Could you elaborate on that?

Lex Fridman (1:23:47.560)

Like why is that important, do you think?

Lex Fridman (1:23:54.560)

For human level intelligence.

Melanie Mitchell (1:23:58.560)

At least for the way the humans work,

Lex Fridman (1:24:00.560)

it's a big part of how it affects how we perceive the world.

Melanie Mitchell (1:24:04.560)

It affects how we make decisions about the world.

Lex Fridman (1:24:07.560)

It affects how we interact with other people.

Melanie Mitchell (1:24:10.560)

It affects our understanding of other people.

Lex Fridman (1:24:14.560)

For me to understand what you're likely to do,

Melanie Mitchell (1:24:21.560)

I need to have kind of a theory of mind

Lex Fridman (1:24:22.560)

and that's very much a theory of emotion

Lex Fridman (1:24:27.560)

and motivations and goals.

Lex Fridman (1:24:32.560)

And to understand that,

Melanie Mitchell (1:24:35.560)

we have this whole system of mirror neurons.

Lex Fridman (1:24:42.560)

I sort of understand your motivations

Melanie Mitchell (1:24:45.560)

through sort of simulating it myself.

Lex Fridman (1:24:49.560)

So it's not something that I can prove that's necessary,

Lex Fridman (1:24:55.560)

but it seems very likely.

Lex Fridman (1:24:58.560)

So, okay.

Melanie Mitchell (1:25:01.560)

You've written the op ed in the New York Times titled

Lex Fridman (1:25:04.560)

We Shouldn't Be Scared by Superintelligent AI

Lex Fridman (1:25:07.560)

and it criticized a little bit Stuart Russell and Nick Bostrom.

Lex Fridman (1:25:13.560)

Can you try to summarize that article's key ideas?

Lex Fridman (1:25:18.560)

So it was spurred by an earlier New York Times op ed

Lex Fridman (1:25:22.560)

by Stuart Russell, which was summarizing his book

Melanie Mitchell (1:25:26.560)

called Human Compatible.

Lex Fridman (1:25:28.560)

And the article was saying if we have superintelligent AI,

Melanie Mitchell (1:25:36.560)

we need to have its values aligned with our values

Lex Fridman (1:25:40.560)

and it has to learn about what we really want.

Lex Fridman (1:25:43.560)

And he gave this example.

Lex Fridman (1:25:45.560)

What if we have a superintelligent AI

Lex Fridman (1:25:48.560)

and we give it the problem of solving climate change

Lex Fridman (1:25:52.560)

and it decides that the best way to lower the carbon

Lex Fridman (1:25:56.560)

in the atmosphere is to kill all the humans?

Lex Fridman (1:25:59.560)

Okay.

Lex Fridman (1:26:00.560)

So to me, that just made no sense at all

Lex Fridman (1:26:02.560)

because a superintelligent AI,

Melanie Mitchell (1:26:08.560)

first of all, trying to figure out what a superintelligence means

Lex Fridman (1:26:13.560)

and it seems that something that's superintelligent

Melanie Mitchell (1:26:21.560)

can't just be intelligent along this one dimension of,

Lex Fridman (1:26:24.560)

okay, I'm going to figure out all the steps,

Melanie Mitchell (1:26:26.560)

the best optimal path to solving climate change

Lex Fridman (1:26:30.560)

and not be intelligent enough to figure out

Melanie Mitchell (1:26:32.560)

that humans don't want to be killed,

Lex Fridman (1:26:36.560)

that you could get to one without having the other.

Melanie Mitchell (1:26:39.560)

And, you know, Bostrom, in his book,

Lex Fridman (1:26:43.560)

talks about the orthogonality hypothesis

Melanie Mitchell (1:26:46.560)

where he says he thinks that a system's,

Lex Fridman (1:26:51.560)

I can't remember exactly what it is,

Lex Fridman (1:26:52.560)

but like a system's goals and its values

Lex Fridman (1:26:56.560)

don't have to be aligned.

Melanie Mitchell (1:26:58.560)

There's some orthogonality there,

Lex Fridman (1:27:00.560)

which didn't make any sense to me.

Lex Fridman (1:27:02.560)

So you're saying in any system that's sufficiently

Lex Fridman (1:27:06.560)

not even superintelligent,

Lex Fridman (1:27:07.560)

but as opposed to greater and greater intelligence,

Lex Fridman (1:27:09.560)

there's a holistic nature that will sort of,

Melanie Mitchell (1:27:11.560)

a tension that will naturally emerge

Lex Fridman (1:27:14.560)

that prevents it from sort of any one dimension running away.

Melanie Mitchell (1:27:17.560)

Yeah, yeah, exactly.

Lex Fridman (1:27:19.560)

So, you know, Bostrom had this example

Melanie Mitchell (1:27:23.560)

of the superintelligent AI that makes,

Lex Fridman (1:27:28.560)

that turns the world into paper clips

Melanie Mitchell (1:27:30.560)

because its job is to make paper clips or something.

Lex Fridman (1:27:33.560)

And that just, as a thought experiment,

Melanie Mitchell (1:27:35.560)

didn't make any sense to me.

Lex Fridman (1:27:37.560)

Well, as a thought experiment

Lex Fridman (1:27:39.560)

or as a thing that could possibly be realized?

Lex Fridman (1:27:42.560)

Either.

Lex Fridman (1:27:43.560)

So I think that, you know,

Lex Fridman (1:27:45.560)

what my op ed was trying to do was say

Melanie Mitchell (1:27:47.560)

that intelligence is more complex

Lex Fridman (1:27:50.560)

than these people are presenting it.

Melanie Mitchell (1:27:53.560)

That it's not like, it's not so separable.

Lex Fridman (1:27:58.560)

The rationality, the values, the emotions,

Melanie Mitchell (1:28:03.560)

the, all of that, that it's,

Lex Fridman (1:28:06.560)

the view that you could separate all these dimensions

Lex Fridman (1:28:09.560)

and build a machine that has one of these dimensions

Lex Fridman (1:28:12.560)

and it's superintelligent in one dimension,

Lex Fridman (1:28:14.560)

but it doesn't have any of the other dimensions.

Lex Fridman (1:28:17.560)

That's what I was trying to criticize

Melanie Mitchell (1:28:22.560)

that I don't believe that.

Lex Fridman (1:28:24.560)

So can I read a few sentences

Lex Fridman (1:28:28.560)

from Yoshua Bengio who is always super eloquent?

Lex Fridman (1:28:35.560)

So he writes,

Melanie Mitchell (1:28:38.560)

I have the same impression as Melanie

Lex Fridman (1:28:40.560)

that our cognitive biases are linked

Melanie Mitchell (1:28:42.560)

with our ability to learn to solve many problems.

Lex Fridman (1:28:45.560)

They may also be a limiting factor for AI.

Melanie Mitchell (1:28:49.560)

However, this is a may in quotes.

Lex Fridman (1:28:53.560)

Things may also turn out differently

Lex Fridman (1:28:55.560)

and there's a lot of uncertainty

Lex Fridman (1:28:56.560)

about the capabilities of future machines.

Lex Fridman (1:28:59.560)

But more importantly for me,

Lex Fridman (1:29:02.560)

the value alignment problem is a problem

Melanie Mitchell (1:29:04.560)

well before we reach some hypothetical superintelligence.

Lex Fridman (1:29:08.560)

It is already posing a problem

Melanie Mitchell (1:29:10.560)

in the form of super powerful companies

Lex Fridman (1:29:13.560)

whose objective function may not be sufficiently aligned

Melanie Mitchell (1:29:17.560)

with humanity's general wellbeing,

Lex Fridman (1:29:19.560)

creating all kinds of harmful side effects.

Lex Fridman (1:29:21.560)

So he goes on to argue that the orthogonality

Lex Fridman (1:29:28.560)

and those kinds of things,

Melanie Mitchell (1:29:29.560)

the concerns of just aligning values

Lex Fridman (1:29:32.560)

with the capabilities of the system

Melanie Mitchell (1:29:34.560)

is something that might come long

Lex Fridman (1:29:37.560)

before we reach anything like superintelligence.

Lex Fridman (1:29:40.560)

So your criticism is kind of really nice to saying

Lex Fridman (1:29:44.560)

this idea of superintelligent systems

Melanie Mitchell (1:29:46.560)

seem to be dismissing fundamental parts

Lex Fridman (1:29:48.560)

of what intelligence would take.

Lex Fridman (1:29:50.560)

And then Yoshua kind of says, yes,

Lex Fridman (1:29:53.560)

but if we look at systems that are much less intelligent,

Melanie Mitchell (1:29:57.560)

there might be these same kinds of problems that emerge.

Lex Fridman (1:30:02.560)

Sure, but I guess the example that he gives there

Lex Fridman (1:30:06.560)

of these corporations, that's people, right?

Lex Fridman (1:30:09.560)

Those are people's values.

Melanie Mitchell (1:30:11.560)

I mean, we're talking about people,

Lex Fridman (1:30:13.560)

the corporations are,

Melanie Mitchell (1:30:16.560)

their values are the values of the people

Lex Fridman (1:30:20.560)

who run those corporations.

Lex Fridman (1:30:21.560)

But the idea is the algorithm, that's right.

Lex Fridman (1:30:24.560)

So the fundamental person,

Melanie Mitchell (1:30:26.560)

the fundamental element of what does the bad thing

Lex Fridman (1:30:30.560)

is a human being.

Melanie Mitchell (1:30:31.560)

Yeah.

Lex Fridman (1:30:32.560)

But the algorithm kind of controls the behavior

Melanie Mitchell (1:30:36.560)

of this mass of human beings.

Lex Fridman (1:30:38.560)

Which algorithm?

Melanie Mitchell (1:30:40.560)

For a company that's the,

Lex Fridman (1:30:42.560)

so for example, if it's an advertisement driven company

Melanie Mitchell (1:30:44.560)

that recommends certain things

Lex Fridman (1:30:47.560)

and encourages engagement,

Lex Fridman (1:30:50.560)

so it gets money by encouraging engagement

Lex Fridman (1:30:53.560)

and therefore the company more and more,

Melanie Mitchell (1:30:57.560)

it's like the cycle that builds an algorithm

Lex Fridman (1:31:00.560)

that enforces more engagement

Lex Fridman (1:31:03.560)

and may perhaps more division in the culture

Lex Fridman (1:31:05.560)

and so on, so on.

Lex Fridman (1:31:07.560)

I guess the question here is sort of who has the agency?

Lex Fridman (1:31:12.560)

So you might say, for instance,

Melanie Mitchell (1:31:14.560)

we don't want our algorithms to be racist.

Lex Fridman (1:31:17.560)

Right.

Lex Fridman (1:31:18.560)

And facial recognition,

Lex Fridman (1:31:21.560)

some people have criticized some facial recognition systems

Melanie Mitchell (1:31:23.560)

as being racist because they're not as good

Lex Fridman (1:31:26.560)

on darker skin than lighter skin.

Melanie Mitchell (1:31:29.560)

That's right.

Lex Fridman (1:31:30.560)

Okay.

Lex Fridman (1:31:31.560)

But the agency there,

Lex Fridman (1:31:33.560)

the actual facial recognition algorithm

Melanie Mitchell (1:31:36.560)

isn't what has the agency.

Lex Fridman (1:31:38.560)

It's not the racist thing, right?

Melanie Mitchell (1:31:41.560)

It's the, I don't know,

Lex Fridman (1:31:44.560)

the combination of the training data,

Melanie Mitchell (1:31:48.560)

the cameras being used, whatever.

Lex Fridman (1:31:51.560)

But my understanding of,

Lex Fridman (1:31:53.560)

and I agree with Bengio there that he,

Lex Fridman (1:31:56.560)

I think there are these value issues

Melanie Mitchell (1:31:59.560)

with our use of algorithms.

Lex Fridman (1:32:02.560)

But my understanding of what Russell's argument was

Melanie Mitchell (1:32:09.560)

is more that the machine itself has the agency now.

Lex Fridman (1:32:14.560)

It's the thing that's making the decisions

Lex Fridman (1:32:17.560)

and it's the thing that has what we would call values.

Lex Fridman (1:32:21.560)

Yes.

Lex Fridman (1:32:22.560)

So whether that's just a matter of degree,

Lex Fridman (1:32:25.560)

it's hard to say, right?

Lex Fridman (1:32:27.560)

But I would say that's sort of qualitatively different

Lex Fridman (1:32:30.560)

than a face recognition neural network.

Lex Fridman (1:32:34.560)

And to broadly linger on that point,

Lex Fridman (1:32:38.560)

if you look at Elon Musk or Stuart Russell or Bostrom,

Melanie Mitchell (1:32:42.560)

people who are worried about existential risks of AI,

Lex Fridman (1:32:45.560)

however far into the future,

Melanie Mitchell (1:32:47.560)

the argument goes is it eventually happens.

Lex Fridman (1:32:50.560)

We don't know how far, but it eventually happens.

Lex Fridman (1:32:53.560)

Do you share any of those concerns

Lex Fridman (1:32:56.560)

and what kind of concerns in general do you have about AI

Lex Fridman (1:32:59.560)

that approach anything like existential threat to humanity?

Lex Fridman (1:33:06.560)

So I would say, yes, it's possible,

Lex Fridman (1:33:10.560)

but I think there's a lot more closer in existential threats to humanity.

Lex Fridman (1:33:15.560)

As you said, like a hundred years for your time.

Melanie Mitchell (1:33:18.560)

It's more than a hundred years.

Lex Fridman (1:33:20.560)

More than a hundred years.

Melanie Mitchell (1:33:21.560)

Maybe even more than 500 years.

Lex Fridman (1:33:23.560)

I don't know.

Lex Fridman (1:33:24.560)

So the existential threats are so far out that the future is,

Lex Fridman (1:33:29.560)

I mean, there'll be a million different technologies

Melanie Mitchell (1:33:32.560)

that we can't even predict now

Lex Fridman (1:33:34.560)

that will fundamentally change the nature of our behavior,

Melanie Mitchell (1:33:37.560)

reality, society, and so on before then.

Lex Fridman (1:33:39.560)

Yeah, I think so.

Melanie Mitchell (1:33:40.560)

I think so.

Lex Fridman (1:33:41.560)

And we have so many other pressing existential threats going on right now.

Melanie Mitchell (1:33:46.560)

Nuclear weapons even.

Lex Fridman (1:33:47.560)

Nuclear weapons, climate problems, poverty, possible pandemics.

Melanie Mitchell (1:33:57.560)

You can go on and on.

Lex Fridman (1:33:59.560)

And I think worrying about existential threat from AI

Melanie Mitchell (1:34:05.560)

is not the best priority for what we should be worrying about.

Lex Fridman (1:34:13.560)

That's kind of my view, because we're so far away.

Lex Fridman (1:34:15.560)

But I'm not necessarily criticizing Russell or Bostrom or whoever

Lex Fridman (1:34:24.560)

for worrying about that.

Lex Fridman (1:34:26.560)

And I think some people should be worried about it.

Lex Fridman (1:34:29.560)

It's certainly fine.

Lex Fridman (1:34:30.560)

But I was more getting at their view of what intelligence is.

Lex Fridman (1:34:38.560)

So I was more focusing on their view of superintelligence

Melanie Mitchell (1:34:42.560)

than just the fact of them worrying.

Lex Fridman (1:34:49.560)

And the title of the article was written by the New York Times editors.

Melanie Mitchell (1:34:54.560)

I wouldn't have called it that.

Lex Fridman (1:34:55.560)

We shouldn't be scared by superintelligence.

Melanie Mitchell (1:34:58.560)

No.

Melanie Mitchell (1:34:59.560)

If you wrote it, it'd be like we should redefine what you mean by superintelligence.

Melanie Mitchell (1:35:02.560)

I actually said something like superintelligence is not a sort of coherent idea.

Lex Fridman (1:35:13.560)

But that's not something the New York Times would put in.

Lex Fridman (1:35:18.560)

And the follow up argument that Yoshua makes also,

Lex Fridman (1:35:22.560)

not argument, but a statement, and I've heard him say it before.

Lex Fridman (1:35:25.560)

And I think I agree.

Lex Fridman (1:35:27.560)

He kind of has a very friendly way of phrasing it.

Melanie Mitchell (1:35:30.560)

It's good for a lot of people to believe different things.

Lex Fridman (1:35:34.560)

He's such a nice guy.

Melanie Mitchell (1:35:36.560)

Yeah.

Lex Fridman (1:35:37.560)

But it's also practically speaking like we shouldn't be like,

Melanie Mitchell (1:35:42.560)

while your article stands, like Stuart Russell does amazing work.

Lex Fridman (1:35:46.560)

Bostrom does amazing work.

Melanie Mitchell (1:35:48.560)

You do amazing work.

Lex Fridman (1:35:49.560)

And even when you disagree about the definition of superintelligence

Melanie Mitchell (1:35:53.560)

or the usefulness of even the term,

Lex Fridman (1:35:56.560)

it's still useful to have people that like use that term, right?

Lex Fridman (1:36:01.560)

And then argue.

Lex Fridman (1:36:02.560)

Sure.

Melanie Mitchell (1:36:03.560)

I absolutely agree with Benjo there.

Lex Fridman (1:36:05.560)

And I think it's great that, you know,

Lex Fridman (1:36:08.560)

and it's great that New York Times will publish all this stuff.

Lex Fridman (1:36:10.560)

That's right.

Melanie Mitchell (1:36:11.560)

It's an exciting time to be here.

Lex Fridman (1:36:13.560)

What do you think is a good test of intelligence?

Melanie Mitchell (1:36:16.560)

Is natural language ultimately a test that you find the most compelling,

Lex Fridman (1:36:21.560)

like the original or the higher levels of the Turing test kind of?

Melanie Mitchell (1:36:28.560)

Yeah, I still think the original idea of the Turing test

Lex Fridman (1:36:33.560)

is a good test for intelligence.

Melanie Mitchell (1:36:36.560)

I mean, I can't think of anything better.

Lex Fridman (1:36:38.560)

You know, the Turing test, the way that it's been carried out so far

Melanie Mitchell (1:36:42.560)

has been very impoverished, if you will.

Lex Fridman (1:36:47.560)

But I think a real Turing test that really goes into depth,

Melanie Mitchell (1:36:52.560)

like the one that I mentioned, I talk about in the book,

Lex Fridman (1:36:54.560)

I talk about Ray Kurzweil and Mitchell Kapoor have this bet, right?

Melanie Mitchell (1:36:59.560)

That in 2029, I think is the date there,

Lex Fridman (1:37:04.560)

a machine will pass the Turing test and they have a very specific,

Melanie Mitchell (1:37:09.560)

like how many hours, expert judges and all of that.

Lex Fridman (1:37:14.560)

And, you know, Kurzweil says yes, Kapoor says no.

Melanie Mitchell (1:37:17.560)

We only have like nine more years to go to see.

Lex Fridman (1:37:21.560)

But I, you know, if something, a machine could pass that,

Melanie Mitchell (1:37:27.560)

I would be willing to call it intelligent.

Lex Fridman (1:37:30.560)

Of course, nobody will.

Melanie Mitchell (1:37:33.560)

They will say that's just a language model, if it does.

Lex Fridman (1:37:37.560)

So you would be comfortable, so language, a long conversation that,

Melanie Mitchell (1:37:43.560)

well, yeah, you're, I mean, you're right,

Lex Fridman (1:37:45.560)

because I think probably to carry out that long conversation,

Melanie Mitchell (1:37:48.560)

you would literally need to have deep common sense understanding of the world.

Lex Fridman (1:37:52.560)

I think so.

Lex Fridman (1:37:54.560)

And the conversation is enough to reveal that.

Lex Fridman (1:37:57.560)

I think so.

Lex Fridman (1:37:59.560)

So another super fun topic of complexity that you have worked on, written about.

Lex Fridman (1:38:09.560)

Let me ask the basic question.

Lex Fridman (1:38:10.560)

What is complexity?

Lex Fridman (1:38:12.560)

So complexity is another one of those terms like intelligence.

Melanie Mitchell (1:38:17.560)

It's perhaps overused.

Lex Fridman (1:38:18.560)

But my book about complexity was about this wide area of complex systems,

Melanie Mitchell (1:38:29.560)

studying different systems in nature, in technology,

Melanie Mitchell (1:38:35.560)

in society in which you have emergence, kind of like I was talking about with intelligence.

Melanie Mitchell (1:38:41.560)

You know, we have the brain, which has billions of neurons.

Lex Fridman (1:38:45.560)

And each neuron individually could be said to be not very complex compared to the system as a whole.

Lex Fridman (1:38:53.560)

But the system, the interactions of those neurons and the dynamics,

Lex Fridman (1:38:58.560)

creates these phenomena that we call intelligence or consciousness,

Melanie Mitchell (1:39:04.560)

you know, that we consider to be very complex.

Lex Fridman (1:39:08.560)

So the field of complexity is trying to find general principles that underlie all these systems

Melanie Mitchell (1:39:16.560)

that have these kinds of emergent properties.

Lex Fridman (1:39:19.560)

And the emergence occurs from like underlying the complex system is usually simple, fundamental interactions.

Melanie Mitchell (1:39:27.560)

Yes.

Lex Fridman (1:39:28.560)

And the emergence happens when there's just a lot of these things interacting.

Melanie Mitchell (1:39:34.560)

Yes.

Lex Fridman (1:39:35.560)

Sort of what, and then most of science to date, can you talk about what is reductionism?

Melanie Mitchell (1:39:45.560)

Well, reductionism is when you try and take a system and divide it up into its elements,

Lex Fridman (1:39:54.560)

whether those be cells or atoms or subatomic particles, whatever your field is,

Lex Fridman (1:40:02.560)

and then try and understand those elements.

Lex Fridman (1:40:06.560)

And then try and build up an understanding of the whole system by looking at sort of the sum of all the elements.

Lex Fridman (1:40:13.560)

So what's your sense?

Melanie Mitchell (1:40:15.560)

Whether we're talking about intelligence or these kinds of interesting complex systems,

Melanie Mitchell (1:40:20.560)

is it possible to understand them in a reductionist way,

Lex Fridman (1:40:24.560)

which is probably the approach of most of science today, right?

Melanie Mitchell (1:40:29.560)

I don't think it's always possible to understand the things we want to understand the most.

Lex Fridman (1:40:35.560)

So I don't think it's possible to look at single neurons and understand what we call intelligence,

Melanie Mitchell (1:40:45.560)

to look at sort of summing up, and sort of the summing up is the issue here.

Melanie Mitchell (1:40:54.560)

One example is that the human genome, right, so there was a lot of work on excitement about sequencing the human genome

Melanie Mitchell (1:41:03.560)

because the idea would be that we'd be able to find genes that underlies diseases.

Lex Fridman (1:41:10.560)

But it turns out that, and it was a very reductionist idea, you know, we figure out what all the parts are,

Lex Fridman (1:41:18.560)

and then we would be able to figure out which parts cause which things.

Lex Fridman (1:41:22.560)

But it turns out that the parts don't cause the things that we're interested in.

Melanie Mitchell (1:41:25.560)

It's like the interactions, it's the networks of these parts.

Lex Fridman (1:41:30.560)

And so that kind of reductionist approach didn't yield the explanation that we wanted.

Lex Fridman (1:41:37.560)

What do you, what do you use the most beautiful complex system that you've encountered?

Lex Fridman (1:41:43.560)

The most beautiful.

Melanie Mitchell (1:41:45.560)

That you've been captivated by.

Lex Fridman (1:41:47.560)

Is it sort of, I mean, for me, is the simplest to be cellular automata.

Melanie Mitchell (1:41:54.560)

Oh, yeah. So I was very captivated by cellular automata and worked on cellular automata for several years.

Lex Fridman (1:42:01.560)

Do you find it amazing or is it surprising that such simple systems, such simple rules in cellular automata can create sort of seemingly unlimited complexity?

Melanie Mitchell (1:42:14.560)

Yeah, that was very surprising to me.

Lex Fridman (1:42:16.560)

How do you make sense of it? How does that make you feel?

Lex Fridman (1:42:18.560)

Is it just ultimately humbling or is there a hope to somehow leverage this into a deeper understanding and even able to engineer things like intelligence?

Lex Fridman (1:42:29.560)

It's definitely humbling.

Lex Fridman (1:42:31.560)

How humbling in that also kind of awe inspiring that it's that awe inspiring like part of mathematics that these credibly simple rules can produce this very beautiful, complex, hard to understand behavior.

Lex Fridman (1:42:50.560)

And that's, it's mysterious, you know, and surprising still.

Lex Fridman (1:42:58.560)

But exciting because it does give you kind of the hope that you might be able to engineer complexity just from simple rules.

Lex Fridman (1:43:09.560)

Can you briefly say what is the Santa Fe Institute, its history, its culture, its ideas, its future?

Lex Fridman (1:43:14.560)

So I've never, as I mentioned to you, I've never been, but it's always been this, in my mind, this mystical place where brilliant people study the edge of chaos.

Lex Fridman (1:43:24.560)

Yeah, exactly.

Lex Fridman (1:43:26.560)

So the Santa Fe Institute was started in 1984 and it was created by a group of scientists, a lot of them from Los Alamos National Lab, which is about a 40 minute drive from Santa Fe Institute.

Melanie Mitchell (1:43:45.560)

They were mostly physicists and chemists, but they were frustrated in their field because they felt so that their field wasn't approaching kind of big interdisciplinary questions like the kinds we've been talking about.

Lex Fridman (1:44:03.560)

And they wanted to have a place where people from different disciplines could work on these big questions without sort of being siloed into physics, chemistry, biology, whatever.

Lex Fridman (1:44:17.560)

So they started this institute and this was people like George Cowen, who was a chemist in the Manhattan Project, and Nicholas Metropolis, a mathematician, physicist, Marie Gail Mann, physicist.

Lex Fridman (1:44:37.560)

So some really big names here.

Melanie Mitchell (1:44:39.560)

Ken Arrow, Nobel Prize winning economist, and they started having these workshops.

Lex Fridman (1:44:47.560)

And this whole enterprise kind of grew into this research institute that itself has been kind of on the edge of chaos its whole life because it doesn't have a significant endowment.

Lex Fridman (1:45:03.560)

And it's just been kind of living on whatever funding it can raise through donations and grants and however it can, you know, business associates and so on.

Lex Fridman (1:45:21.560)

But it's a great place. It's a really fun place to go think about ideas that you wouldn't normally encounter.

Lex Fridman (1:45:28.560)

I saw Sean Carroll, a physicist. Yeah, he's on the external faculty.

Lex Fridman (1:45:34.560)

And you mentioned that there's, so there's some external faculty and there's people that are...

Melanie Mitchell (1:45:37.560)

A very small group of resident faculty, maybe about 10 who are there for five year terms that can sometimes get renewed.

Lex Fridman (1:45:48.560)

And then they have some postdocs and then they have this much larger on the order of 100 external faculty or people like me who come and visit for various periods of time.

Lex Fridman (1:45:59.560)

So what do you think is the future of the Santa Fe Institute?

Lex Fridman (1:46:02.560)

And if people are interested, like what's there in terms of the public interaction or students or so on that could be a possible interaction with the Santa Fe Institute or its ideas?

Lex Fridman (1:46:15.560)

Yeah, so there's a few different things they do.

Melanie Mitchell (1:46:18.560)

They have a complex system summer school for graduate students and postdocs and sometimes faculty attend too.

Lex Fridman (1:46:25.560)

And that's a four week, very intensive residential program where you go and you listen to lectures and you do projects and people really like that.

Melanie Mitchell (1:46:35.560)

I mean, it's a lot of fun.

Lex Fridman (1:46:37.560)

They also have some specialty summer schools.

Melanie Mitchell (1:46:41.560)

There's one on computational social science.

Lex Fridman (1:46:45.560)

There's one on climate and sustainability, I think it's called.

Melanie Mitchell (1:46:52.560)

There's a few and then they have short courses where just a few days on different topics.

Melanie Mitchell (1:46:59.560)

They also have an online education platform that offers a lot of different courses and tutorials from SFI faculty.

Melanie Mitchell (1:47:09.560)

Including an introduction to complexity course that I taught.

Melanie Mitchell (1:47:13.560)

Awesome. And there's a bunch of talks too online from the guest speakers and so on.

Melanie Mitchell (1:47:19.560)

They host a lot of...

Melanie Mitchell (1:47:20.560)

Yeah, they have sort of technical seminars and colloquia and they have a community lecture series like public lectures and they put everything on their YouTube channel so you can see it all.

Melanie Mitchell (1:47:33.560)

Watch it.

Lex Fridman (1:47:34.560)

Douglas Hofstadter, author of Ghetto Escherbach, was your PhD advisor.

Melanie Mitchell (1:47:40.560)

He mentioned a couple of times in collaborator.

Lex Fridman (1:47:43.560)

Do you have any favorite lessons or memories from your time working with him that continues to this day?

Melanie Mitchell (1:47:50.560)

Just even looking back throughout your time working with him.

Melanie Mitchell (1:47:55.560)

One of the things he taught me was that when you're looking at a complex problem, to idealize it as much as possible to try and figure out what is the essence of this problem.

Lex Fridman (1:48:11.560)

And this is how the copycat program came into being was by taking analogy making and saying, how can we make this as idealized as possible but still retain really the important things we want to study?

Lex Fridman (1:48:25.560)

And that's really been a core theme of my research, I think.

Lex Fridman (1:48:33.560)

And I continue to try and do that.

Lex Fridman (1:48:36.560)

And it's really very much kind of physics inspired. Hofstadter was a PhD in physics.

Melanie Mitchell (1:48:42.560)

That was his background.

Lex Fridman (1:48:44.560)

It's like first principles kind of thing.

Melanie Mitchell (1:48:46.560)

You're reduced to the most fundamental aspect of the problem so that you can focus on solving that fundamental aspect.

Lex Fridman (1:48:52.560)

Yeah.

Lex Fridman (1:48:53.560)

And in AI, people used to work in these micro worlds, right?

Lex Fridman (1:48:57.560)

Like the blocks world was very early important area in AI.

Lex Fridman (1:49:02.560)

And then that got criticized because they said, oh, you can't scale that to the real world.

Lex Fridman (1:49:09.560)

And so people started working on much more real world like problems.

Lex Fridman (1:49:14.560)

But now there's been kind of a return even to the blocks world itself.

Melanie Mitchell (1:49:19.560)

We've seen a lot of people who are trying to work on more of these very idealized problems for things like natural language and common sense.

Lex Fridman (1:49:28.560)

So that's an interesting evolution of those ideas.

Lex Fridman (1:49:31.560)

So perhaps the blocks world represents the fundamental challenges of the problem of intelligence more than people realize.

Melanie Mitchell (1:49:38.560)

It might. Yeah.

Melanie Mitchell (1:49:41.560)

When you look back at your body of work and your life, you've worked in so many different fields.

Lex Fridman (1:49:46.560)

Is there something that you're just really proud of in terms of ideas that you've gotten a chance to explore, create yourself?

Lex Fridman (1:49:54.560)

So I am really proud of my work on the copycat project.

Melanie Mitchell (1:49:59.560)

I think it's really different from what almost everyone has done in AI.

Lex Fridman (1:50:04.560)

I think there's a lot of ideas there to be explored.

Lex Fridman (1:50:08.560)

And I guess one of the happiest days of my life.

Melanie Mitchell (1:50:14.560)

You know, aside from like the births of my children was the birth of copycat when it actually started to be able to make really interesting analogies.

Lex Fridman (1:50:24.560)

And I remember that very clearly.

Lex Fridman (1:50:27.560)

It was a very exciting time.

Melanie Mitchell (1:50:30.560)

Well, you kind of gave life to an artificial system.

Lex Fridman (1:50:34.560)

That's right.

Melanie Mitchell (1:50:35.560)

In terms of what people can interact, I saw there's like a, I think it's called MetaCat.

Lex Fridman (1:50:40.560)

MetaCat.

Melanie Mitchell (1:50:41.560)

MetaCat.

Lex Fridman (1:50:42.560)

And there's a Python 3 implementation.

Melanie Mitchell (1:50:45.560)

If people actually wanted to play around with it and actually get into it and study it and maybe integrate into whether it's with deep learning or any other kind of work they're doing.

Lex Fridman (1:50:54.560)

What would you suggest they do to learn more about it and to take it forward in different kinds of directions?

Melanie Mitchell (1:51:00.560)

Yeah, so that there's Douglas Hofstadter's book called Fluid Concepts and Creative Analogies talks in great detail about copycat.

Melanie Mitchell (1:51:09.560)

I have a book called Analogy Making as Perception, which is a version of my PhD thesis on it.

Melanie Mitchell (1:51:16.560)

There's also code that's available that you can get it to run.

Lex Fridman (1:51:20.560)

I have some links on my webpage to where people can get the code for it.

Lex Fridman (1:51:25.560)

And I think that that would really be the best way to get into it.

Lex Fridman (1:51:28.560)

Just dive in and play with it.

Melanie Mitchell (1:51:30.560)

Well, Melanie, it was an honor talking to you.

Lex Fridman (1:51:33.560)

I really enjoyed it.

Melanie Mitchell (1:51:34.560)

Thank you so much for your time today.

Lex Fridman (1:51:35.560)

Thanks.

Melanie Mitchell (1:51:36.560)

It's been really great.

Lex Fridman (1:51:38.560)

Thanks for listening to this conversation with Melanie Mitchell.

Lex Fridman (1:51:41.560)

And thank you to our presenting sponsor, Cash App.

Lex Fridman (1:51:44.560)

Download it.

Melanie Mitchell (1:51:45.560)

Use code LexPodcast.

Melanie Mitchell (1:51:47.560)

You will get $10 and $10 will go to FIRST, a STEM education nonprofit that inspires hundreds of thousands of young minds to learn and to dream of engineering our future.

Melanie Mitchell (1:51:58.560)

If you enjoy this podcast, subscribe on YouTube, give it five stars on Apple Podcast, support it on Patreon or connect with me on Twitter.

Lex Fridman (1:52:06.560)

And now let me leave you with some words of wisdom from Douglas Hofstadter and Melanie Mitchell.

Melanie Mitchell (1:52:12.560)

Without concepts, there can be no thought.

Lex Fridman (1:52:15.560)

Without analogies, there can be no concepts.

Lex Fridman (1:52:18.560)

And Melanie adds, how to form and fluidly use concepts is the most important open problem in AI.

Lex Fridman (1:52:27.560)

Thank you for listening and hope to see you next time.

Lex Fridman (20:01.400)

But I think that, and they're really,

Lex Fridman (20:04.300)

Marvin Minsky is a super smart guy

Lex Fridman (20:06.120)

and very sophisticated thinker.

Lex Fridman (20:08.400)

But I think that no one really understands

Melanie Mitchell (20:12.960)

or understood, still doesn't understand

Lex Fridman (20:16.240)

how complicated, how complex the things that we do are

Melanie Mitchell (20:22.160)

because they're so invisible to us.

Lex Fridman (20:24.640)

To us, vision, being able to look out at the world

Lex Fridman (20:27.640)

and describe what we see, that's just immediate.

Lex Fridman (20:31.660)

It feels like it's no work at all.

Lex Fridman (20:33.360)

So it didn't seem like it would be that hard,

Lex Fridman (20:35.920)

but there's so much going on unconsciously,

Melanie Mitchell (20:39.320)

sort of invisible to us that I think we overestimate

Lex Fridman (20:44.440)

how easy it will be to get computers to do it.

Lex Fridman (20:50.020)

And sort of for me to ask an unfair question,

Lex Fridman (20:53.880)

you've done research, you've thought about

Melanie Mitchell (20:56.520)

many different branches of AI through this book,

Lex Fridman (20:59.880)

widespread looking at where AI has been, where it is today.

Melanie Mitchell (21:06.360)

If you were to make a prediction,

Lex Fridman (21:08.840)

how many years from now would we as a society

Melanie Mitchell (21:12.120)

create something that you would say

Lex Fridman (21:15.760)

achieved human level intelligence

Lex Fridman (21:19.720)

or superhuman level intelligence?

Lex Fridman (21:23.140)

That is an unfair question.

Melanie Mitchell (21:25.120)

A prediction that will most likely be wrong.

Lex Fridman (21:28.520)

But it's just your notion because.

Melanie Mitchell (21:30.000)

Okay, I'll say more than 100 years.

Lex Fridman (21:34.300)

More than 100 years.

Lex Fridman (21:35.320)

And I quoted somebody in my book who said that

Lex Fridman (21:38.520)

human level intelligence is 100 Nobel Prizes away,

Melanie Mitchell (21:44.660)

which I like because it's a nice way to sort of,

Lex Fridman (21:48.040)

it's a nice unit for prediction.

Lex Fridman (21:51.800)

And it's like that many fantastic discoveries

Lex Fridman (21:55.680)

have to be made.

Lex Fridman (21:56.600)

And of course there's no Nobel Prize in AI, not yet at least.

Lex Fridman (22:03.120)

If we look at that 100 years,

Melanie Mitchell (22:05.300)

your sense is really the journey to intelligence

Lex Fridman (22:10.240)

has to go through something more complicated

Melanie Mitchell (22:15.680)

that's akin to our own cognitive systems,

Lex Fridman (22:19.400)

understanding them, being able to create them

Melanie Mitchell (22:21.640)

in the artificial systems,

Lex Fridman (22:24.480)

as opposed to sort of taking the machine learning

Melanie Mitchell (22:26.880)

approaches of today and really scaling them

Lex Fridman (22:30.280)

and scaling them and scaling them exponentially

Melanie Mitchell (22:33.560)

with both compute and hardware and data.

Lex Fridman (22:37.920)

That would be my guess.

Melanie Mitchell (22:42.200)

I think that in the sort of going along in the narrow AI

Lex Fridman (22:47.200)

that the current approaches will get better.

Melanie Mitchell (22:54.840)

I think there's some fundamental limits

Lex Fridman (22:56.840)

to how far they're gonna get.

Melanie Mitchell (22:59.360)

I might be wrong, but that's what I think.

Lex Fridman (23:01.800)

And there's some fundamental weaknesses that they have

Melanie Mitchell (23:06.680)

that I talk about in the book that just comes

Lex Fridman (23:10.920)

from this approach of supervised learning requiring

Melanie Mitchell (23:20.760)

sort of feed forward networks and so on.

Lex Fridman (23:27.120)

It's just, I don't think it's a sustainable approach

Melanie Mitchell (23:31.240)

to understanding the world.

Lex Fridman (23:34.200)

Yeah, I'm personally torn on it.

Melanie Mitchell (23:36.460)

Sort of everything you read about in the book

Lex Fridman (23:39.480)

and sort of what we're talking about now,

Melanie Mitchell (23:41.160)

I agree with you, but I'm more and more,

Lex Fridman (23:45.800)

depending on the day, first of all,

Melanie Mitchell (23:48.040)

I'm deeply surprised by the success

Lex Fridman (23:50.080)

of machine learning and deep learning in general.

Melanie Mitchell (23:52.760)

From the very beginning, when I was,

Lex Fridman (23:54.920)

it's really been my main focus of work.

Melanie Mitchell (23:57.280)

I'm just surprised how far it gets.

Lex Fridman (23:59.380)

And I'm also think we're really early on

Melanie Mitchell (24:03.560)

in these efforts of these narrow AI.

Lex Fridman (24:07.080)

So I think there'll be a lot of surprise

Melanie Mitchell (24:09.360)

of how far it gets.

Lex Fridman (24:11.880)

I think we'll be extremely impressed.

Melanie Mitchell (24:14.360)

Like my sense is everything I've seen so far,

Lex Fridman (24:17.120)

and we'll talk about autonomous driving and so on,

Melanie Mitchell (24:19.480)

I think we can get really far.

Lex Fridman (24:21.760)

But I also have a sense that we will discover,

Melanie Mitchell (24:24.720)

just like you said, is that even though we'll get

Lex Fridman (24:27.560)

really far in order to create something

Melanie Mitchell (24:30.680)

like our own intelligence, it's actually much farther

Lex Fridman (24:32.880)

than we realize.

Melanie Mitchell (24:34.680)

I think these methods are a lot more powerful

Lex Fridman (24:37.160)

than people give them credit for actually.

Lex Fridman (24:39.120)

So that of course there's the media hype,

Lex Fridman (24:41.160)

but I think there's a lot of researchers in the community,

Lex Fridman (24:43.700)

especially like not undergrads, right?

Lex Fridman (24:46.680)

But like people who've been in AI,

Melanie Mitchell (24:48.820)

they're skeptical about how far deep learning can get.

Lex Fridman (24:50.940)

And I'm more and more thinking that it can actually

Melanie Mitchell (24:54.640)

get farther than they'll realize.

Lex Fridman (24:56.960)

It's certainly possible.

Melanie Mitchell (24:58.440)

One thing that surprised me when I was writing the book

Lex Fridman (25:00.840)

is how far apart different people in the field are

Melanie Mitchell (25:03.800)

on their opinion of how far the field has come

Lex Fridman (25:08.400)

and what is accomplished and what's gonna happen next.

Melanie Mitchell (25:11.520)

What's your sense of the different,

Lex Fridman (25:13.760)

who are the different people, groups, mindsets,

Lex Fridman (25:17.520)

thoughts in the community about where AI is today?

Lex Fridman (25:22.760)

Yeah, they're all over the place.

Lex Fridman (25:24.080)

So there's kind of the singularity transhumanism group.

Lex Fridman (25:30.760)

I don't know exactly how to characterize that approach,

Melanie Mitchell (25:33.200)

which is sort of the sort of exponential,

Lex Fridman (25:36.560)

exponential progress where we're on the sort of

Melanie Mitchell (25:41.320)

almost at the hugely accelerating part of the exponential.

Lex Fridman (25:45.720)

And in the next 30 years,

Melanie Mitchell (25:49.680)

we're going to see super intelligent AI and all that,

Lex Fridman (25:54.080)

and we'll be able to upload our brains and that.

Lex Fridman (25:57.360)

So there's that kind of extreme view that most,

Lex Fridman (26:00.480)

I think most people who work in AI don't have.

Melanie Mitchell (26:04.600)

They disagree with that.

Lex Fridman (26:06.040)

But there are people who are,

Melanie Mitchell (26:09.280)

maybe aren't singularity people,

Lex Fridman (26:12.880)

but they do think that the current approach

Melanie Mitchell (26:16.840)

of deep learning is going to scale

Lex Fridman (26:20.000)

and is going to kind of go all the way basically

Lex Fridman (26:23.800)

and take us to true AI or human level AI

Lex Fridman (26:26.680)

or whatever you wanna call it.

Lex Fridman (26:29.100)

And there's quite a few of them.

Lex Fridman (26:30.840)

And a lot of them, like a lot of the people I've met

Melanie Mitchell (26:34.760)

who work at big tech companies in AI groups

Lex Fridman (26:40.160)

kind of have this view that we're really not that far.

Melanie Mitchell (26:46.160)

Just to linger on that point,

Lex Fridman (26:47.360)

sort of if I can take as an example, like Yann LeCun,

Melanie Mitchell (26:50.920)

I don't know if you know about his work

Lex Fridman (26:52.600)

and so his viewpoints on this.

Melanie Mitchell (26:54.400)

I do.

Lex Fridman (26:55.240)

He believes that there's a bunch of breakthroughs,

Melanie Mitchell (26:57.760)

like fundamental, like Nobel prizes that are needed still.

Lex Fridman (27:01.040)

But I think he thinks those breakthroughs

Melanie Mitchell (27:03.540)

will be built on top of deep learning.

Lex Fridman (27:06.540)

And then there's some people who think

Melanie Mitchell (27:08.520)

we need to kind of put deep learning

Lex Fridman (27:11.280)

to the side a little bit as just one module

Melanie Mitchell (27:14.440)

that's helpful in the bigger cognitive framework.

Lex Fridman (27:17.760)

Right, so I think somewhat I understand Yann LeCun

Melanie Mitchell (27:22.000)

is rightly saying supervised learning is not sustainable.

Lex Fridman (27:27.960)

We have to figure out how to do unsupervised learning,

Melanie Mitchell (27:31.080)

that that's gonna be the key.

Lex Fridman (27:34.000)

And I think that's probably true.

Melanie Mitchell (27:39.360)

I think unsupervised learning

Lex Fridman (27:40.720)

is gonna be harder than people think.

Melanie Mitchell (27:43.280)

I mean, the way that we humans do it.

Lex Fridman (27:47.040)

Then there's the opposing view,

Melanie Mitchell (27:50.920)

there's the Gary Marcus kind of hybrid view

Lex Fridman (27:55.840)

where deep learning is one part,

Lex Fridman (27:58.120)

but we need to bring back kind of these symbolic approaches

Lex Fridman (28:02.200)

and combine them.

Melanie Mitchell (28:03.400)

Of course, no one knows how to do that very well.

Lex Fridman (28:06.640)

Which is the more important part to emphasize

Lex Fridman (28:10.360)

and how do they fit together?

Lex Fridman (28:12.040)

What's the foundation?

Lex Fridman (28:13.760)

What's the thing that's on top?

Lex Fridman (28:15.400)

What's the cake?

Lex Fridman (28:16.220)

What's the icing?

Lex Fridman (28:17.060)

Right.

Melanie Mitchell (28:18.600)

Then there's people pushing different things.

Lex Fridman (28:22.680)

There's the people, the causality people who say,

Melanie Mitchell (28:26.640)

deep learning as it's formulated today

Lex Fridman (28:28.680)

completely lacks any notion of causality.

Lex Fridman (28:32.040)

And that's, dooms it.

Lex Fridman (28:35.120)

And therefore we have to somehow give it

Melanie Mitchell (28:37.680)

some kind of notion of causality.

Lex Fridman (28:41.300)

There's a lot of push

Melanie Mitchell (28:45.080)

from the more cognitive science crowd saying,

Lex Fridman (28:51.400)

we have to look at developmental learning.

Melanie Mitchell (28:54.120)

We have to look at how babies learn.

Lex Fridman (28:56.720)

We have to look at intuitive physics,

Melanie Mitchell (29:00.960)

all these things we know about physics.

Lex Fridman (29:03.000)

And as somebody kind of quipped,

Melanie Mitchell (29:05.280)

we also have to teach machines intuitive metaphysics,

Lex Fridman (29:08.800)

which means like objects exist.

Melanie Mitchell (29:14.540)

Causality exists.

Lex Fridman (29:17.480)

These things that maybe we're born with.

Melanie Mitchell (29:19.260)

I don't know that they don't have the,

Lex Fridman (29:21.800)

machines don't have any of that.

Melanie Mitchell (29:23.800)

They look at a group of pixels

Lex Fridman (29:26.600)

and maybe they get 10 million examples,

Lex Fridman (29:31.380)

but they can't necessarily learn

Lex Fridman (29:34.360)

that there are objects in the world.

Lex Fridman (29:38.160)

So there's just a lot of pieces of the puzzle

Lex Fridman (29:41.160)

that people are promoting

Lex Fridman (29:44.040)

and with different opinions of like how important they are

Lex Fridman (29:47.640)

and how close we are to being able to put them all together

Melanie Mitchell (29:52.000)

to create general intelligence.

Lex Fridman (29:54.080)

Looking at this broad field,

Lex Fridman (29:56.580)

what do you take away from it?

Lex Fridman (29:57.800)

Who is the most impressive?

Melanie Mitchell (29:59.580)

Is it the cognitive folks,

Lex Fridman (30:01.720)

the Gary Marcus camp, the on camp,

Melanie Mitchell (30:05.120)

unsupervised and their self supervised.

Lex Fridman (30:07.000)

There's the supervisors and then there's the engineers

Melanie Mitchell (30:09.640)

who are actually building systems.

Lex Fridman (30:11.560)

You have sort of the Andrej Karpathy at Tesla

Melanie Mitchell (30:14.720)

building actual, it's not philosophy,

Lex Fridman (30:17.960)

it's real like systems that operate in the real world.

Lex Fridman (30:21.040)

What do you take away from all this beautiful variety?

Lex Fridman (30:23.880)

I don't know if,

Melanie Mitchell (30:25.600)

these different views are not necessarily

Lex Fridman (30:27.520)

mutually exclusive.

Lex Fridman (30:29.640)

And I think people like Yann LeCun

Lex Fridman (30:34.640)

agrees with the developmental psychology of causality,

Melanie Mitchell (30:39.600)

intuitive physics, et cetera.

Lex Fridman (30:43.160)

But he still thinks that it's learning,

Melanie Mitchell (30:45.960)

like end to end learning is the way to go.

Lex Fridman (30:48.280)

Will take us perhaps all the way.

Melanie Mitchell (30:50.080)

Yeah, and that we don't need,

Lex Fridman (30:51.080)

there's no sort of innate stuff that has to get built in.

Melanie Mitchell (30:56.880)

This is, it's because it's a hard problem.

Lex Fridman (31:02.240)

I personally, I'm very sympathetic

Melanie Mitchell (31:05.280)

to the cognitive science side,

Lex Fridman (31:07.200)

cause that's kind of where I came in to the field.

Melanie Mitchell (31:10.460)

I've become more and more sort of an embodiment adherent

Lex Fridman (31:15.460)

saying that without having a body,

Melanie Mitchell (31:18.540)

it's gonna be very hard to learn

Lex Fridman (31:20.840)

what we need to learn about the world.

Melanie Mitchell (31:24.420)

That's definitely something I'd love to talk about

Lex Fridman (31:26.840)

in a little bit.

Melanie Mitchell (31:28.760)

To step into the cognitive world,

Lex Fridman (31:31.520)

then if you don't mind,

Melanie Mitchell (31:32.760)

cause you've done so many interesting things.

Lex Fridman (31:34.240)

If you look to copycat,

Melanie Mitchell (31:36.920)

taking a couple of decades step back,

Lex Fridman (31:40.240)

you, Douglas Hofstadter and others

Melanie Mitchell (31:43.320)

have created and developed copycat

Lex Fridman (31:45.040)

more than 30 years ago.

Melanie Mitchell (31:48.680)

That's painful to hear.

Lex Fridman (31:50.880)

So what is it?

Lex Fridman (31:51.920)

What is copycat?

Lex Fridman (31:54.280)

It's a program that makes analogies

Melanie Mitchell (31:57.800)

in an idealized domain,

Lex Fridman (32:00.680)

idealized world of letter strings.

Lex Fridman (32:03.580)

So as you say, 30 years ago, wow.

Lex Fridman (32:06.520)

So I started working on it

Melanie Mitchell (32:07.880)

when I started grad school in 1984.

Lex Fridman (32:12.600)

Wow, dates me.

Lex Fridman (32:17.960)

And it's based on Doug Hofstadter's ideas

Lex Fridman (32:21.680)

about that analogy is really a core aspect of thinking.

Melanie Mitchell (32:30.240)

I remember he has a really nice quote

Lex Fridman (32:32.360)

in the book by himself and Emmanuel Sandor

Melanie Mitchell (32:36.900)

called Surfaces and Essences.

Lex Fridman (32:38.760)

I don't know if you've seen that book,

Lex Fridman (32:39.760)

but it's about analogy and he says,

Lex Fridman (32:43.880)

without concepts, there can be no thought

Lex Fridman (32:46.800)

and without analogies, there can be no concepts.

Lex Fridman (32:51.120)

So the view is that analogy

Melanie Mitchell (32:52.560)

is not just this kind of reasoning technique

Lex Fridman (32:55.040)

where we go, shoe is to foot as glove is to what,

Melanie Mitchell (33:01.880)

these kinds of things that we have on IQ tests or whatever,

Lex Fridman (33:05.440)

but that it's much deeper,

Melanie Mitchell (33:06.540)

it's much more pervasive in every thing we do,

Lex Fridman (33:10.960)

in our language, our thinking, our perception.

Lex Fridman (33:16.080)

So he had a view that was a very active perception idea.

Lex Fridman (33:20.920)

So the idea was that instead of having kind of

Melanie Mitchell (33:26.680)

a passive network in which you have input

Lex Fridman (33:31.680)

that's being processed through these feed forward layers

Lex Fridman (33:35.480)

and then there's an output at the end,

Lex Fridman (33:37.080)

that perception is really a dynamic process

Melanie Mitchell (33:41.440)

where like our eyes are moving around

Lex Fridman (33:43.360)

and they're getting information

Lex Fridman (33:44.760)

and that information is feeding back

Lex Fridman (33:47.040)

to what we look at next, influences,

Lex Fridman (33:50.640)

what we look at next and how we look at it.

Lex Fridman (33:53.200)

And so copycat was trying to do that,

Melanie Mitchell (33:56.080)

kind of simulate that kind of idea

Lex Fridman (33:57.720)

where you have these agents,

Melanie Mitchell (34:02.640)

it's kind of an agent based system

Lex Fridman (34:04.120)

and you have these agents that are picking things

Melanie Mitchell (34:07.160)

to look at and deciding whether they were interesting

Lex Fridman (34:10.680)

or not and whether they should be looked at more

Lex Fridman (34:13.580)

and that would influence other agents.

Lex Fridman (34:15.880)

Now, how do they interact?

Lex Fridman (34:17.560)

So they interacted through this global kind of

Lex Fridman (34:20.040)

what we call the workspace.

Lex Fridman (34:22.160)

So it's actually inspired by the old blackboard systems

Lex Fridman (34:25.480)

where you would have agents that post information

Melanie Mitchell (34:28.920)

on a blackboard, a common blackboard.

Lex Fridman (34:30.840)

This is like very old fashioned AI.

Lex Fridman (34:33.560)

Is that, are we talking about like in physical space?

Lex Fridman (34:36.280)

Is this a computer program?

Melanie Mitchell (34:37.120)

It's a computer program.

Lex Fridman (34:38.320)

So agents posting concepts on a blackboard kind of thing?

Melanie Mitchell (34:41.960)

Yeah, we called it a workspace.

Lex Fridman (34:43.920)

And the workspace is a data structure.

Melanie Mitchell (34:48.440)

The agents are little pieces of code

Lex Fridman (34:50.720)

that you could think of them as little detectors

Melanie Mitchell (34:54.080)

or little filters that say,

Lex Fridman (34:55.960)

I'm gonna pick this place to look

Lex Fridman (34:57.480)

and I'm gonna look for a certain thing

Lex Fridman (34:59.080)

and is this the thing I think is important, is it there?

Lex Fridman (35:03.040)

So it's almost like, you know, a convolution in a way,

Lex Fridman (35:06.960)

except a little bit more general and saying,

Lex Fridman (35:10.800)

and then highlighting it in the workspace.

Lex Fridman (35:14.680)

Once it's in the workspace,

Lex Fridman (35:16.320)

how do the things that are highlighted

Lex Fridman (35:18.000)

relate to each other?

Lex Fridman (35:18.880)

Like what's, is this?

Lex Fridman (35:19.720)

So there's different kinds of agents

Melanie Mitchell (35:21.560)

that can build connections between different things.

Lex Fridman (35:23.640)

So just to give you a concrete example,

Lex Fridman (35:25.600)

what CopyCat did was it made analogies

Lex Fridman (35:28.400)

between strings of letters.

Lex Fridman (35:30.360)

So here's an example.

Lex Fridman (35:31.960)

ABC changes to ABD.

Lex Fridman (35:35.360)

What does IJK change to?

Lex Fridman (35:39.200)

And the program had some prior knowledge

Melanie Mitchell (35:41.200)

about the alphabet, knew the sequence of the alphabet.

Lex Fridman (35:45.160)

It had a concept of letter, successor of letter.

Melanie Mitchell (35:49.320)

It had concepts of sameness.

Lex Fridman (35:50.960)

So it has some innate things programmed in.

Lex Fridman (35:55.120)

But then it could do things like say,

Lex Fridman (35:58.360)

discover that ABC is a group of letters in succession.

Lex Fridman (36:06.400)

And then an agent can mark that.

Lex Fridman (36:11.000)

So the idea that there could be a sequence of letters,

Melanie Mitchell (36:16.200)

is that a new concept that's formed

Lex Fridman (36:18.160)

or that's a concept that's innate?

Melanie Mitchell (36:19.400)

That's a concept that's innate.

Lex Fridman (36:21.480)

Sort of, can you form new concepts

Melanie Mitchell (36:23.680)

or are all concepts innate? No.

Lex Fridman (36:25.040)

So in this program, all the concepts

Melanie Mitchell (36:28.520)

of the program were innate.

Lex Fridman (36:30.240)

So, cause we weren't, I mean,

Melanie Mitchell (36:32.240)

obviously that limits it quite a bit.

Lex Fridman (36:35.600)

But what we were trying to do is say,

Melanie Mitchell (36:37.200)

suppose you have some innate concepts,

Lex Fridman (36:40.400)

how do you flexibly apply them to new situations?

Lex Fridman (36:45.160)

And how do you make analogies?

Lex Fridman (36:47.800)

Let's step back for a second.

Lex Fridman (36:49.040)

So I really liked that quote that you say,

Lex Fridman (36:51.760)

without concepts, there could be no thought

Lex Fridman (36:53.760)

and without analogies, there can be no concepts.

Lex Fridman (36:56.600)

In a Santa Fe presentation,

Melanie Mitchell (36:58.480)

you said that it should be one of the mantras of AI.

Lex Fridman (37:00.880)

Yes.

Lex Fridman (37:01.880)

And that you also yourself said,

Lex Fridman (37:04.320)

how to form and fluidly use concept

Melanie Mitchell (37:06.640)

is the most important open problem in AI.

Lex Fridman (37:09.880)

Yes.

Lex Fridman (37:11.240)

How to form and fluidly use concepts

Lex Fridman (37:14.500)

is the most important open problem in AI.

Lex Fridman (37:16.980)

So let's, what is a concept and what is an analogy?

Lex Fridman (37:21.880)

A concept is in some sense a fundamental unit of thought.

Lex Fridman (37:28.200)

So say we have a concept of a dog, okay?

Lex Fridman (37:38.560)

And a concept is embedded in a whole space of concepts

Lex Fridman (37:45.120)

so that there's certain concepts that are closer to it

Lex Fridman (37:48.720)

or farther away from it.

Melanie Mitchell (37:50.240)

Are these concepts, are they really like fundamental,

Lex Fridman (37:53.120)

like we mentioned innate, almost like axiomatic,

Melanie Mitchell (37:55.600)

like very basic and then there's other stuff

Lex Fridman (37:57.960)

built on top of it?

Lex Fridman (37:58.880)

Or does this include everything?

Lex Fridman (38:01.080)

Are they complicated?

Melanie Mitchell (38:04.360)

You can certainly form new concepts.

Lex Fridman (38:06.980)

Right, I guess that's the question I'm asking.

Lex Fridman (38:08.720)

Can you form new concepts

Lex Fridman (38:10.080)

that are complex combinations of other concepts?

Melanie Mitchell (38:14.360)

Yes, absolutely.

Lex Fridman (38:15.960)

And that's kind of what we do in learning.

Lex Fridman (38:20.000)

And then what's the role of analogies in that?

Lex Fridman (38:22.960)

So analogy is when you recognize

Melanie Mitchell (38:27.200)

that one situation is essentially the same

Lex Fridman (38:33.320)

as another situation.

Lex Fridman (38:35.560)

And essentially is kind of the key word there

Lex Fridman (38:38.760)

because it's not the same.

Lex Fridman (38:39.980)

So if I say, last week I did a podcast interview

Lex Fridman (38:44.980)

actually like three days ago in Washington, DC.

Lex Fridman (38:52.980)

And that situation was very similar to this situation,

Lex Fridman (38:56.580)

although it wasn't exactly the same.

Melanie Mitchell (38:58.380)

It was a different person sitting across from me.

Lex Fridman (39:00.780)

We had different kinds of microphones.

Melanie Mitchell (39:03.380)

The questions were different.

Lex Fridman (39:04.740)

The building was different.

Melanie Mitchell (39:06.140)

There's all kinds of different things,

Lex Fridman (39:07.140)

but really it was analogous.

Melanie Mitchell (39:10.220)

Or I can say, so doing a podcast interview,

Lex Fridman (39:14.700)

that's kind of a concept, it's a new concept.

Melanie Mitchell (39:17.540)

I never had that concept before this year essentially.

Lex Fridman (39:23.020)

I mean, and I can make an analogy with it

Melanie Mitchell (39:27.220)

like being interviewed for a news article in a newspaper.

Lex Fridman (39:31.380)

And I can say, well, you kind of play the same role

Melanie Mitchell (39:35.660)

that the newspaper reporter played.

Lex Fridman (39:40.100)

It's not exactly the same

Melanie Mitchell (39:42.060)

because maybe they actually emailed me some written questions

Lex Fridman (39:45.020)

rather than talking and the writing,

Melanie Mitchell (39:48.300)

the written questions are analogous

Lex Fridman (39:52.060)

to your spoken questions.

Lex Fridman (39:53.260)

And there's just all kinds of similarities.

Lex Fridman (39:55.100)

And this somehow probably connects to conversations

Melanie Mitchell (39:57.420)

you have over Thanksgiving dinner,

Lex Fridman (39:58.820)

just general conversations.

Melanie Mitchell (40:01.060)

There's like a thread you can probably take

Lex Fridman (40:03.520)

that just stretches out in all aspects of life

Melanie Mitchell (40:06.700)

that connect to this podcast.

Lex Fridman (40:08.440)

I mean, conversations between humans.

Melanie Mitchell (40:11.460)

Sure, and if I go and tell a friend of mine

Lex Fridman (40:16.920)

about this podcast interview, my friend might say,

Melanie Mitchell (40:20.740)

oh, the same thing happened to me.

Lex Fridman (40:22.900)

Let's say, you ask me some really hard question

Lex Fridman (40:27.020)

and I have trouble answering it.

Lex Fridman (40:29.260)

My friend could say, the same thing happened to me,

Lex Fridman (40:31.640)

but it was like, it wasn't a podcast interview.

Lex Fridman (40:34.100)

It wasn't, it was a completely different situation.

Lex Fridman (40:39.100)

And yet my friend is seeing essentially the same thing.

Lex Fridman (40:43.340)

We say that very fluidly, the same thing happened to me.

Melanie Mitchell (40:46.540)

Essentially the same thing.

Lex Fridman (40:48.940)

But we don't even say that, right?

Melanie Mitchell (40:50.180)

We just say the same thing.

Lex Fridman (40:51.020)

You imply it, yes.

Melanie Mitchell (40:51.860)

Yeah, and the view that kind of went into say copycat,

Lex Fridman (40:56.860)

that whole thing is that that act of saying

Melanie Mitchell (41:00.860)

the same thing happened to me is making an analogy.

Lex Fridman (41:04.540)

And in some sense, that's what's underlies

Melanie Mitchell (41:07.820)

all of our concepts.

Lex Fridman (41:10.660)

Why do you think analogy making that you're describing

Lex Fridman (41:14.020)

is so fundamental to cognition?

Lex Fridman (41:17.020)

Like it seems like it's the main element action

Melanie Mitchell (41:20.020)

of what we think of as cognition.

Lex Fridman (41:23.820)

Yeah, so it can be argued that all of this

Melanie Mitchell (41:28.260)

generalization we do of concepts

Lex Fridman (41:31.500)

and recognizing concepts in different situations

Melanie Mitchell (41:39.580)

is done by analogy.

Lex Fridman (41:42.620)

That that's, every time I'm recognizing

Melanie Mitchell (41:48.220)

that say you're a person, that's by analogy

Lex Fridman (41:53.740)

because I have this concept of what person is

Lex Fridman (41:55.660)

and I'm applying it to you.

Lex Fridman (41:57.360)

And every time I recognize a new situation,

Melanie Mitchell (42:02.580)

like one of the things I talked about in the book

Lex Fridman (42:06.540)

was the concept of walking a dog,

Melanie Mitchell (42:09.700)

that that's actually making an analogy

Lex Fridman (42:11.780)

because all of the details are very different.

Lex Fridman (42:15.420)

So reasoning could be reduced down

Lex Fridman (42:19.420)

to essentially analogy making.

Lex Fridman (42:21.780)

So all the things we think of as like,

Lex Fridman (42:25.220)

yeah, like you said, perception.

Lex Fridman (42:26.820)

So what's perception is taking raw sensory input

Lex Fridman (42:29.680)

and it's somehow integrating into our understanding

Melanie Mitchell (42:33.020)

of the world, updating the understanding.

Lex Fridman (42:34.740)

And all of that has just this giant mess of analogies

Melanie Mitchell (42:39.180)

that are being made.

Lex Fridman (42:40.180)

I think so, yeah.

Melanie Mitchell (42:42.540)

If you just linger on it a little bit,

Lex Fridman (42:44.280)

like what do you think it takes to engineer

Lex Fridman (42:47.260)

a process like that for us in our artificial systems?

Lex Fridman (42:52.140)

We need to understand better, I think,

Lex Fridman (42:56.900)

how we do it, how humans do it.

Lex Fridman (43:02.700)

And it comes down to internal models, I think.

Melanie Mitchell (43:07.840)

People talk a lot about mental models,

Lex Fridman (43:11.140)

that concepts are mental models,

Melanie Mitchell (43:13.300)

that I can, in my head, I can do a simulation

Lex Fridman (43:18.300)

of a situation like walking a dog.

Lex Fridman (43:22.500)

And there's some work in psychology

Lex Fridman (43:25.580)

that promotes this idea that all of concepts

Melanie Mitchell (43:29.420)

are really mental simulations,

Lex Fridman (43:31.800)

that whenever you encounter a concept

Melanie Mitchell (43:35.100)

or situation in the world or you read about it or whatever,

Lex Fridman (43:38.100)

you do some kind of mental simulation

Melanie Mitchell (43:40.680)

that allows you to predict what's gonna happen,

Lex Fridman (43:44.020)

to develop expectations of what's gonna happen.

Lex Fridman (43:47.980)

So that's the kind of structure I think we need,

Lex Fridman (43:51.580)

is that kind of mental model that,

Lex Fridman (43:55.580)

and in our brains, somehow these mental models

Lex Fridman (43:58.060)

are very much interconnected.

Melanie Mitchell (44:01.300)

Again, so a lot of stuff we're talking about

Lex Fridman (44:03.700)

are essentially open problems, right?

Lex Fridman (44:05.960)

So if I ask a question, I don't mean

Lex Fridman (44:08.700)

that you would know the answer, only just hypothesizing.

Lex Fridman (44:11.340)

But how big do you think is the network graph,

Lex Fridman (44:19.300)

data structure of concepts that's in our head?

Melanie Mitchell (44:23.300)

Like if we're trying to build that ourselves,

Lex Fridman (44:26.500)

like it's, we take it,

Melanie Mitchell (44:28.140)

that's one of the things we take for granted.

Lex Fridman (44:29.580)

We think, I mean, that's why we take common sense

Melanie Mitchell (44:32.060)

for granted, we think common sense is trivial.

Lex Fridman (44:34.720)

But how big of a thing of concepts

Melanie Mitchell (44:38.940)

is that underlies what we think of as common sense,

Lex Fridman (44:42.400)

for example?

Melanie Mitchell (44:44.580)

Yeah, I don't know.

Lex Fridman (44:45.460)

And I'm not, I don't even know what units to measure it in.

Lex Fridman (44:48.420)

Can you say how big is it?

Lex Fridman (44:50.260)

That's beautifully put, right?

Melanie Mitchell (44:51.980)

But, you know, we have, you know, it's really hard to know.

Lex Fridman (44:55.700)

We have, what, a hundred billion neurons or something.

Melanie Mitchell (45:00.900)

I don't know.

Lex Fridman (45:02.880)

And they're connected via trillions of synapses.

Lex Fridman (45:07.860)

And there's all this chemical processing going on.

Lex Fridman (45:10.540)

There's just a lot of capacity for stuff.

Lex Fridman (45:13.740)

And their information's encoded

Lex Fridman (45:15.860)

in different ways in the brain.

Melanie Mitchell (45:17.180)

It's encoded in chemical interactions.

Lex Fridman (45:19.900)

It's encoded in electric, like firing and firing rates.

Lex Fridman (45:24.220)

And nobody really knows how it's encoded,

Lex Fridman (45:25.780)

but it just seems like there's a huge amount of capacity.

Lex Fridman (45:29.020)

So I think it's huge.

Lex Fridman (45:30.860)

It's just enormous.

Lex Fridman (45:32.460)

And it's amazing how much stuff we know.

Lex Fridman (45:36.740)

Yeah.

Lex Fridman (45:38.140)

And for, but we know, and not just know like facts,

Lex Fridman (45:42.780)

but it's all integrated into this thing

Melanie Mitchell (45:44.860)

that we can make analogies with.

Lex Fridman (45:46.540)

Yes.

Melanie Mitchell (45:47.380)

There's a dream of Semantic Web,

Lex Fridman (45:49.300)

and there's a lot of dreams from expert systems

Melanie Mitchell (45:53.000)

of building giant knowledge bases.

Lex Fridman (45:56.300)

Do you see a hope for these kinds of approaches

Melanie Mitchell (45:58.980)

of building, of converting Wikipedia

Lex Fridman (46:01.180)

into something that could be used in analogy making?

Melanie Mitchell (46:05.160)

Sure.

Lex Fridman (46:07.280)

And I think people have made some progress

Melanie Mitchell (46:09.600)

along those lines.

Lex Fridman (46:10.540)

I mean, people have been working on this for a long time.

Lex Fridman (46:13.320)

But the problem is,

Lex Fridman (46:14.800)

and this I think is the problem of common sense.

Melanie Mitchell (46:17.760)

Like people have been trying to get

Lex Fridman (46:19.120)

these common sense networks.

Lex Fridman (46:21.000)

Here at MIT, there's this concept net project, right?

Lex Fridman (46:25.420)

But the problem is that, as I said,

Melanie Mitchell (46:27.480)

most of the knowledge that we have is invisible to us.

Lex Fridman (46:31.920)

It's not in Wikipedia.

Melanie Mitchell (46:33.200)

It's very basic things about intuitive physics,

Lex Fridman (46:42.320)

intuitive psychology, intuitive metaphysics,

Melanie Mitchell (46:46.400)

all that stuff.

Lex Fridman (46:47.240)

If you were to create a website

Melanie Mitchell (46:49.200)

that described intuitive physics, intuitive psychology,

Lex Fridman (46:53.480)

would it be bigger or smaller than Wikipedia?

Lex Fridman (46:56.480)

What do you think?

Lex Fridman (46:58.940)

I guess described to whom?

Melanie Mitchell (47:00.680)

I'm sorry, but.

Lex Fridman (47:03.880)

No, that's really good.

Melanie Mitchell (47:05.360)

That's exactly right, yeah.

Lex Fridman (47:07.060)

That's a hard question,

Melanie Mitchell (47:07.900)

because how do you represent that knowledge

Lex Fridman (47:10.560)

is the question, right?

Melanie Mitchell (47:12.080)

I can certainly write down F equals MA

Lex Fridman (47:15.760)

and Newton's laws and a lot of physics

Melanie Mitchell (47:19.680)

can be deduced from that.

Lex Fridman (47:23.280)

But that's probably not the best representation

Melanie Mitchell (47:27.060)

of that knowledge for doing the kinds of reasoning

Lex Fridman (47:32.320)

we want a machine to do.

Melanie Mitchell (47:35.760)

So, I don't know, it's impossible to say now.

Lex Fridman (47:40.400)

And people, you know, the projects,

Melanie Mitchell (47:43.160)

like there's the famous psych project, right,

Lex Fridman (47:46.520)

that Douglas Linnaught did that was trying.

Lex Fridman (47:50.040)

That thing's still going?

Lex Fridman (47:51.080)

I think it's still going.

Lex Fridman (47:52.080)

And the idea was to try and encode

Lex Fridman (47:54.800)

all of common sense knowledge,

Melanie Mitchell (47:56.280)

including all this invisible knowledge

Lex Fridman (47:58.480)

in some kind of logical representation.

Lex Fridman (48:03.480)

And it just never, I think, could do any of the things

Lex Fridman (48:09.200)

that he was hoping it could do,

Melanie Mitchell (48:11.000)

because that's just the wrong approach.

Lex Fridman (48:13.920)

Of course, that's what they always say, you know.

Lex Fridman (48:16.760)

And then the history books will say,

Lex Fridman (48:18.880)

well, the psych project finally found a breakthrough

Melanie Mitchell (48:21.900)

in 2058 or something.

Lex Fridman (48:24.480)

So much progress has been made in just a few decades

Melanie Mitchell (48:28.500)

that who knows what the next breakthroughs will be.

Lex Fridman (48:31.980)

It could be.

Melanie Mitchell (48:32.820)

It's certainly a compelling notion

Lex Fridman (48:34.700)

what the psych project stands for.

Melanie Mitchell (48:37.540)

I think Linnaught was one of the earliest people

Lex Fridman (48:39.940)

to say common sense is what we need.

Melanie Mitchell (48:43.540)

That's what we need.

Lex Fridman (48:44.780)

All this like expert system stuff,

Melanie Mitchell (48:46.980)

that is not gonna get you to AI.

Lex Fridman (48:49.140)

You need common sense.

Lex Fridman (48:50.420)

And he basically gave up his whole academic career

Lex Fridman (48:56.180)

to go pursue that.

Lex Fridman (48:57.660)

And I totally admire that,

Lex Fridman (48:59.420)

but I think that the approach itself will not,

Melanie Mitchell (49:06.020)

in 2040 or wherever, be successful.

Lex Fridman (49:09.020)

What do you think is wrong with the approach?

Lex Fridman (49:10.300)

What kind of approach might be successful?

Lex Fridman (49:14.640)

Well, if I knew that.

Lex Fridman (49:15.480)

Again, nobody knows the answer, right?

Lex Fridman (49:16.940)

If I knew that, you know, one of my talks,

Melanie Mitchell (49:19.060)

one of the people in the audience,

Lex Fridman (49:21.080)

this is a public lecture,

Melanie Mitchell (49:22.200)

one of the people in the audience said,

Lex Fridman (49:24.220)

what AI companies are you investing in?

Melanie Mitchell (49:27.040)

I'm like, well, I'm a college professor for one thing,

Lex Fridman (49:31.840)

so I don't have a lot of extra funds to invest,

Lex Fridman (49:34.740)

but also like no one knows what's gonna work in AI, right?

Lex Fridman (49:39.320)

That's the problem.

Melanie Mitchell (49:41.520)

Let me ask another impossible question

Lex Fridman (49:43.120)

in case you have a sense.

Melanie Mitchell (49:44.760)

In terms of data structures

Lex Fridman (49:46.460)

that will store this kind of information,

Lex Fridman (49:49.520)

do you think they've been invented yet,

Lex Fridman (49:51.880)

both in hardware and software?

Lex Fridman (49:54.600)

Or is it something else needs to be, are we totally, you know?

Lex Fridman (49:58.280)

I think something else has to be invented.

Melanie Mitchell (50:01.920)

That's my guess.

Lex Fridman (50:03.560)

Is the breakthroughs that's most promising,

Lex Fridman (50:06.440)

would that be in hardware or in software?

Lex Fridman (50:09.720)

Do you think we can get far with the current computers?

Lex Fridman (50:12.680)

Or do we need to do something that you see?

Lex Fridman (50:14.800)

I see what you're saying.

Melanie Mitchell (50:16.400)

I don't know if Turing computation

Lex Fridman (50:18.560)

is gonna be sufficient.

Melanie Mitchell (50:19.880)

Probably, I would guess it will.

Lex Fridman (50:22.040)

I don't see any reason why we need anything else.

Lex Fridman (50:26.020)

So in that sense, we have invented the hardware we need,

Lex Fridman (50:29.000)

but we just need to make it faster and bigger,

Lex Fridman (50:31.900)

and we need to figure out the right algorithms

Lex Fridman (50:34.300)

and the right sort of architecture.

Melanie Mitchell (50:39.620)

Turing, that's a very mathematical notion.

Lex Fridman (50:43.080)

When we try to have to build intelligence,

Melanie Mitchell (50:44.920)

it's now an engineering notion

Lex Fridman (50:46.800)

where you throw all that stuff.

Melanie Mitchell (50:48.320)

Well, I guess it is a question.

Lex Fridman (50:53.440)

People have brought up this question,

Lex Fridman (50:56.200)

and when you asked about, like, is our current hardware,

Lex Fridman (51:00.680)

will our current hardware work?

Melanie Mitchell (51:02.240)

Well, Turing computation says that our current hardware

Lex Fridman (51:08.800)

is, in principle, a Turing machine, right?

Lex Fridman (51:13.300)

So all we have to do is make it faster and bigger.

Lex Fridman (51:16.480)

But there have been people like Roger Penrose,

Melanie Mitchell (51:20.200)

if you might remember, that he said,

Lex Fridman (51:22.560)

Turing machines cannot produce intelligence

Melanie Mitchell (51:26.440)

because intelligence requires continuous valued numbers.

Lex Fridman (51:30.480)

I mean, that was sort of my reading of his argument.

Lex Fridman (51:34.800)

And quantum mechanics and what else, whatever.

Lex Fridman (51:38.440)

But I don't see any evidence for that,

Melanie Mitchell (51:41.680)

that we need new computation paradigms.

Lex Fridman (51:48.060)

But I don't know if we're, you know,

Melanie Mitchell (51:50.440)

I don't think we're gonna be able to scale up

Lex Fridman (51:53.880)

our current approaches to programming these computers.

Lex Fridman (51:58.400)

What is your hope for approaches like CopyCat

Lex Fridman (52:00.760)

or other cognitive architectures?

Melanie Mitchell (52:02.680)

I've talked to the creator of SOAR, for example.

Lex Fridman (52:04.640)

I've used ActR myself.

Melanie Mitchell (52:06.000)

I don't know if you're familiar with it.

Lex Fridman (52:07.040)

Yeah, I am.

Lex Fridman (52:07.880)

What do you think is,

Lex Fridman (52:10.120)

what's your hope of approaches like that

Melanie Mitchell (52:12.040)

in helping develop systems of greater

Lex Fridman (52:15.840)

and greater intelligence in the coming decades?

Melanie Mitchell (52:19.960)

Well, that's what I'm working on now,

Lex Fridman (52:22.160)

is trying to take some of those ideas and extending it.

Lex Fridman (52:26.080)

So I think there are some really promising approaches

Lex Fridman (52:30.120)

that are going on now that have to do with

Melanie Mitchell (52:34.120)

more active generative models.

Lex Fridman (52:39.520)

So this is the idea of this simulation in your head,

Melanie Mitchell (52:42.760)

the concept, when you, if you wanna,

Lex Fridman (52:46.160)

when you're perceiving a new situation,

Melanie Mitchell (52:49.880)

you have some simulations in your head.

Lex Fridman (52:51.280)

Those are generative models.

Melanie Mitchell (52:52.560)

They're generating your expectations.

Lex Fridman (52:54.600)

They're generating predictions.

Lex Fridman (52:55.920)

So that's part of a perception.

Lex Fridman (52:57.240)

You have a metamodel that generates a prediction

Melanie Mitchell (53:00.680)

then you compare it with, and then the difference.

Lex Fridman (53:03.560)

And you also, that generative model is telling you

Melanie Mitchell (53:07.560)

where to look and what to look at

Lex Fridman (53:09.480)

and what to pay attention to.

Lex Fridman (53:11.640)

And it, I think it affects your perception.

Lex Fridman (53:14.080)

It's not that just you compare it with your perception.

Melanie Mitchell (53:16.680)

It becomes your perception in a way.

Lex Fridman (53:21.960)

It's kind of a mixture of the bottom up information

Melanie Mitchell (53:28.320)

coming from the world and your top down model

Lex Fridman (53:31.880)

being imposed on the world is what becomes your perception.

Lex Fridman (53:36.160)

So your hope is something like that

Lex Fridman (53:37.400)

can improve perception systems

Lex Fridman (53:39.600)

and that they can understand things better.

Lex Fridman (53:41.760)

Yes. To understand things.

Melanie Mitchell (53:42.920)

Yes.

Lex Fridman (53:44.160)

What's the, what's the step,

Lex Fridman (53:47.160)

what's the analogy making step there?

Lex Fridman (53:49.520)

Well, there, the idea is that you have this

Melanie Mitchell (53:54.040)

pretty complicated conceptual space.

Lex Fridman (53:57.120)

You can talk about a semantic network or something like that

Melanie Mitchell (54:00.420)

with these different kinds of concept models

Lex Fridman (54:04.240)

in your brain that are connected.

Melanie Mitchell (54:07.280)

So, so let's, let's take the example of walking a dog.

Lex Fridman (54:10.920)

So we were talking about that.

Melanie Mitchell (54:12.360)

Okay.

Lex Fridman (54:13.600)

Let's say I see someone out in the street walking a cat.

Melanie Mitchell (54:16.640)

Some people walk their cats, I guess.

Lex Fridman (54:18.600)

Seems like a bad idea, but.

Melanie Mitchell (54:19.880)

Yeah.

Lex Fridman (54:21.760)

So my model, my, you know,

Melanie Mitchell (54:23.480)

there's connections between my model of a dog

Lex Fridman (54:27.220)

and model of a cat.

Lex Fridman (54:28.920)

And I can immediately see the analogy

Lex Fridman (54:33.120)

of that those are analogous situations,

Lex Fridman (54:38.760)

but I can also see the differences

Lex Fridman (54:40.840)

and that tells me what to expect.

Lex Fridman (54:43.280)

So also, you know, I have a new situation.

Lex Fridman (54:48.640)

So another example with the walking the dog thing

Melanie Mitchell (54:51.280)

is sometimes people,

Lex Fridman (54:52.960)

I see people riding their bikes with a leash,

Melanie Mitchell (54:55.120)

holding a leash and the dogs running alongside.

Lex Fridman (54:57.640)

Okay, so I know that the,

Melanie Mitchell (55:00.200)

I recognize that as kind of a dog walking situation,

Lex Fridman (55:03.940)

even though the person's not walking, right?

Lex Fridman (55:06.800)

And the dog's not walking.

Lex Fridman (55:08.480)

Because I have these models that say, okay,

Melanie Mitchell (55:14.120)

riding a bike is sort of similar to walking

Lex Fridman (55:16.580)

or it's connected, it's a means of transportation,

Lex Fridman (55:20.180)

but I, because they have their dog there,

Lex Fridman (55:22.840)

I assume they're not going to work,

Lex Fridman (55:24.400)

but they're going out for exercise.

Lex Fridman (55:26.360)

You know, these analogies help me to figure out

Melanie Mitchell (55:30.240)

kind of what's going on, what's likely.

Lex Fridman (55:33.120)

But sort of these analogies are very human interpretable.

Lex Fridman (55:37.240)

So that's that kind of space.

Lex Fridman (55:38.980)

And then you look at something

Melanie Mitchell (55:40.480)

like the current deep learning approaches,

Lex Fridman (55:43.420)

they kind of help you to take raw sensory information

Lex Fridman (55:46.680)

and to sort of automatically build up hierarchies

Lex Fridman (55:49.440)

of what you can even call them concepts.

Melanie Mitchell (55:52.960)

They're just not human interpretable concepts.

Lex Fridman (55:55.600)

What's your, what's the link here?

Lex Fridman (55:58.640)

Do you hope, sort of the hybrid system question,

Lex Fridman (56:05.720)

how do you think the two can start to meet each other?

Melanie Mitchell (56:08.220)

What's the value of learning in this systems of forming,

Lex Fridman (56:14.040)

of analogy making?

Melanie Mitchell (56:16.040)

The goal of, you know, the original goal of deep learning

Lex Fridman (56:20.600)

in at least visual perception was that

Melanie Mitchell (56:24.260)

you would get the system to learn to extract features

Lex Fridman (56:27.320)

that at these different levels of complexity.

Lex Fridman (56:30.120)

So maybe edge detection and that would lead into learning,

Lex Fridman (56:34.000)

you know, simple combinations of edges

Lex Fridman (56:36.640)

and then more complex shapes

Lex Fridman (56:38.120)

and then whole objects or faces.

Lex Fridman (56:42.740)

And this was based on the ideas

Lex Fridman (56:47.960)

of the neuroscientists, Hubel and Wiesel,

Melanie Mitchell (56:51.480)

who had seen, laid out this kind of structure in brain.

Lex Fridman (56:58.740)

And I think that's right to some extent.

Melanie Mitchell (57:02.020)

Of course, people have found that the whole story

Lex Fridman (57:05.840)

is a little more complex than that.

Lex Fridman (57:07.320)

And the brain of course always is

Lex Fridman (57:09.120)

and there's a lot of feedback.

Lex Fridman (57:10.520)

So I see that as absolutely a good brain inspired approach

Lex Fridman (57:22.860)

to some aspects of perception.

Lex Fridman (57:25.680)

But one thing that it's lacking, for example,

Lex Fridman (57:29.460)

is all of that feedback, which is extremely important.

Melanie Mitchell (57:33.300)

The interactive element that you mentioned.

Lex Fridman (57:36.420)

The expectation, right, the conceptual level.

Melanie Mitchell (57:39.020)

Going back and forth with the expectation,

Lex Fridman (57:42.220)

the perception and just going back and forth.

Melanie Mitchell (57:44.180)

So, right, so that is extremely important.

Lex Fridman (57:47.940)

And, you know, one thing about deep neural networks

Melanie Mitchell (57:52.180)

is that in a given situation,

Lex Fridman (57:54.960)

like, you know, they're trained, right?

Melanie Mitchell (57:56.660)

They get these weights and everything,

Lex Fridman (57:58.340)

but then now I give them a new image, let's say.

Melanie Mitchell (58:02.400)

They treat every part of the image in the same way.

Lex Fridman (58:09.860)

You know, they apply the same filters at each layer

Melanie Mitchell (58:13.540)

to all parts of the image.

Lex Fridman (58:15.900)

There's no feedback to say like,

Melanie Mitchell (58:17.600)

oh, this part of the image is irrelevant.

Lex Fridman (58:20.860)

I shouldn't care about this part of the image.

Melanie Mitchell (58:23.060)

Or this part of the image is the most important part.

Lex Fridman (58:27.020)

And that's kind of what we humans are able to do

Melanie Mitchell (58:30.120)

because we have these conceptual expectations.

Lex Fridman (58:33.140)

So there's a, by the way, a little bit of work in that.

Melanie Mitchell (58:35.580)

There's certainly a lot more in what's under the,

Lex Fridman (58:38.900)

called attention in natural language processing knowledge.

Melanie Mitchell (58:42.480)

It's a, and that's exceptionally powerful.

Lex Fridman (58:46.820)

And it's a very, just as you say,

Melanie Mitchell (58:49.240)

it's a really powerful idea.

Lex Fridman (58:50.660)

But again, in sort of machine learning,

Melanie Mitchell (58:53.380)

it all kind of operates in an automated way.

Lex Fridman (58:55.740)

That's not human interpret.

Melanie Mitchell (58:56.940)

It's not also, okay, so that, right.

Lex Fridman (58:59.340)

It's not dynamic.

Melanie Mitchell (59:00.300)

I mean, in the sense that as a perception

Lex Fridman (59:03.420)

of a new example is being processed,

Melanie Mitchell (59:08.540)

those attention's weights don't change.

Lex Fridman (59:12.780)

Right, so I mean, there's a kind of notion

Melanie Mitchell (59:17.540)

that there's not a memory.

Lex Fridman (59:20.340)

So you're not aggregating the idea of like,

Melanie Mitchell (59:23.820)

this mental model.

Lex Fridman (59:25.040)

Yes.

Melanie Mitchell (59:26.540)

I mean, that seems to be a fundamental idea.

Lex Fridman (59:28.600)

There's not a really powerful,

Melanie Mitchell (59:30.940)

I mean, there's some stuff with memory,

Lex Fridman (59:32.380)

but there's not a powerful way to represent the world

Melanie Mitchell (59:37.820)

in some sort of way that's deeper than,

Lex Fridman (59:42.300)

I mean, it's so difficult because, you know,

Melanie Mitchell (59:45.300)

neural networks do represent the world.

Lex Fridman (59:47.580)

They do have a mental model, right?

Lex Fridman (59:50.860)

But it just seems to be shallow.

Lex Fridman (59:53.000)

It's hard to criticize them at the fundamental level,