Lex Fridman Podcast - #148 - Charles Isbell and Michael Littman: Machine Learning and Education

The following is a conversation with Charles Isbell

and Michael Whitman.

Charles is the Dean of the College of Computing

at Georgia Tech and Michael is a computer science professor

at Brown University.

I’ve spoken with each of them individually on this podcast

and since they are good friends in real life,

we all thought it would be fun

to have a conversation together.

Quick mention of each sponsor,

followed by some thoughts related to the episode.

Thank you to Athletic Greens,

the all in one drink that I start every day with

to cover all my nutritional bases.

Eight Sleep, a mattress that cools itself

and gives me yet another reason to enjoy sleep.

Masterclass, online courses

from some of the most amazing humans in history

and Cash App, the app I use to send money to friends.

Please check out the sponsors in the description

to get a discount and to support this podcast.

As a side note, let me say that having two guests

on the podcast is an experiment

that I’ve been meaning to do for a while.

In particular, because down the road,

I would like to occasionally be a kind of moderator

for debates between people that may disagree

in some interesting ways.

If you have suggestions for who you would like to see debate

on this podcast, let me know.

As with all experiments of this kind,

it is a learning process.

Both the video and the audio might need improvement.

I realized I think I should probably do

three or more cameras next time

as opposed to just two.

And also try different ways to mount the microphone

for the third person.

Also, after recording this intro,

I’m going to have to go figure out the thumbnail

for the video version of the podcast

since I usually put the guest’s head on the thumbnail.

And now there’s two heads and two names

to try to fit into the thumbnail.

It’s a kind of a bin packing problem

which in theoretical computer science

happens to be an NP hard problem.

Whatever I come up with, if you have better ideas

for the thumbnail, let me know as well.

And in general, I always welcome ideas

how this thing can be improved.

If you enjoy it, subscribe on YouTube,

review it with Five Stars and Apple Podcast,

follow on Spotify, support on Patreon,

or connect with me on Twitter at Lex Friedman.

And now here’s my conversation with Charles Isbell

and Michael Littman.

You’ll probably disagree about this question,

but what is your biggest, would you say, disagreement

about either something profound and very important

or something completely not important at all?

I don’t think we have any disagreements at all.

I’m not sure that’s true.

We walked into that one, didn’t we?

So one thing that you sometimes mention is that,

and we did this one on air too, as it were,

whether or not machine learning

is computational statistics.

It’s not.

But it is.

Well, it’s not.

And in particular, and more importantly,

it is not just computational statistics.

So what’s missing in the picture?

All the rest of it.

What’s missing?

That which is missing.

Oh, yes, well, you can’t be wrong now.

Well, it’s not just the statistics.

He doesn’t even believe this.

We’ve had this conversation before.

If it were just the statistics,

then we would be happy with where we are.

But it’s not just the statistics.

That’s why it’s computational statistics.

Or if it were just the computational.

I agree that machine learning is not just statistics.

It is not just statistics.

We can agree on that.

Nor is it just computational statistics.

It’s computational statistics.

It is computational.

What is the computational and computational statistics?

Does this take us into the realm of computing?

It does, but I think perhaps the way I can get him

to admit that he’s wrong is that it’s about rules.

It’s about rules.

It’s about symbols.

It’s about all these other things.

But statistics is not about rules?

I’m gonna say statistics is about rules.

But it’s not just the statistics, right?

It’s not just a random variable that you choose

and you have a probability.

I think you have a narrow view of statistics.

Okay, well then what would be the broad view of statistics

that would still allow it to be statistics

and not say history that would make

computational statistics okay?

Well, okay, so I had my first sort of research mentor,

a guy named Tom Landauer,

taught me to do some statistics, right?

And I was annoyed all the time

because the statistics would say

that what I was doing was not statistically significant.

And I was like, but, but, and basically what he said to me

is statistics is how you’re gonna keep

from lying to yourself, which I thought was really deep.

It is a way to keep yourself honest in a particular way.

I agree with that.

Yeah, and so you’re trying to find rules.

I’m just gonna bring it back to rules.

Wait, wait, wait, could you possibly try to define rules?

Even regular statisticians, noncomputational statisticians,

do spend some of their time evaluating rules, right?

Applying statistics to try to understand

does this rule capture this?

Does this not capture that?

You mean like hypothesis testing kind of thing?

Or like confidence intervals?

I think more like hypothesis.

Like I feel like the word statistic

literally means like a summary,

like a number that summarizes other numbers.

But I think the field of statistics

actually applies that idea to things like rules,

to understand whether or not a rule is valid.

Does software engineering statistics?

No.

Programming languages statistics?

No.

Because I think there’s a very,

it’s useful to think about a lot of what AI

and machine learning is or certainly should be

as software engineering, as programming languages.

Just to put it in language that you might understand,

the hyperparameters beyond the problem itself.

The hyperparameters is too many syllables

for me to understand.

The hyperparameters.

That’s better.

That goes around it, right?

It’s the decisions you choose to make.

It’s the metrics you choose to use.

It’s the loss function.

You wanna say the practice of machine learning

is different than the practice of statistics.

Like the things you have to worry about

and how you worry about them are different,

therefore they’re different.

Right.

At a very little, I mean, at the very least.

It’s that much is true.

It doesn’t mean that statistics,

computational or otherwise aren’t important.

I think they are.

I mean, I do a lot of that, for example.

But I think it goes beyond that.

I think that we could think about game theory

in terms of statistics,

but I don’t think it’s very as useful to do.

I mean, the way I would think about it

or a way I would think about it is this way.

Chemistry is just physics.

But I don’t think it’s as useful to think about chemistry

as being just physics.

It’s useful to think about it as chemistry.

The level of abstraction really matters here.

So I think it is,

there are contexts in which it is useful.

Yes.

I think of it that way, right?

So finding that connection is actually helpful.

And I think that’s when I emphasize

the computational statistics thing.

I think I want to befriend statistics and not absorb them.

Here’s the A way to think about it

beyond what I just said, right?

So what would you say,

and I want you to think back to a conversation

we had a very long time ago.

What would you say is the difference between,

say, the early 2000s, ICML

and what we used to call NIPS, NeurIPS?

Is there a difference?

A lot of, particularly on the machine learning

that was done there?

ICML was around that long.

Oh, yeah.

So iClear is the new conference, newish.

Yeah, I guess so.

And ICML was around the 2000.

So ICML predates that.

I think my most cited ICML paper is from 94.

Michael knows this better than me

because, of course, he’s significantly older than I am.

But the point is, what is the difference

between ICML and NeurIPS in the late 90s, early 2000s?

I don’t know what everyone else’s perspective would be,

but I had a particular perspective at that time,

which is I felt like ICML was more

of a computer science place

and that NIPS, NeurIPS was more of an engineering place,

like the kind of math that happened at the two places.

As a computer scientist,

I felt more comfortable with the ICML math.

And the NeurIPS people would say

that that’s because I’m dumb.

And that’s such an engineering thing to say, so.

I agree with that part of it,

but I do it a little differently.

We actually had a nice conversation

with Tom Dietrich about this in public.

On Twitter just a couple of days ago.

I put it a little differently,

which is that ICML was machine learning done

by a computer scientist.

And NeurIPS was machine learning done

by a computer scientist trying to impress statisticians.

Which was weird because it was the same people,

at least by the time I started paying attention.

But it just felt very, very different.

And I think that that perspective

of whether you’re trying to impress the statisticians

or you’re trying to impress the programmers

is actually very different and has real impact

on what you choose to worry about

and what kind of outcomes you come to.

So I think it really matters.

I think computational statistics is a means to an end.

It is not an end in some sense.

And I think that really matters here

in the same way that I don’t think computer science

is just engineering or just science

or just math or whatever.

Okay, so I’d have to now agree

that now we agree on everything.

Yes, yes.

The important thing here is that

my opinions may have changed,

but not the fact that I’m right,

I think is what we just came to.

Right, and my opinions may have changed

and not the fact that I’m wrong.

That’s right.

You lost me.

I’m not even.

I think I lost myself there too.

But anyway, we’re back.

This happens to us sometimes.

We’re sorry.

How does neural networks change this,

just to even linger on this topic,

change this idea of statistics,

how big of a pie statistics is

within the machine learning thing?

Like, because it sounds like hyperparameters

and also just the role of data.

You know, people are starting to use

this terminology of software 2.0,

which is like the act of programming

as a, like you’re a designer

in the hyperparameter space of neural networks,

and you’re also the collector and the organizer

and the cleaner of the data,

and that’s part of the programming.

So how did, on the NeurIPS versus ICML topic,

what’s the role of neural networks

in redefining the size and the role of machine learning?

I can’t wait to hear what Michael thinks about this,

but I would add one.

But you will.

That’s true, I will, I’ll force myself to.

I think there’s one other thing

I would add to your description,

which is the kind of software engineering part

of what does it mean to debug, for example.

But this is a difference between

the kind of computational statistics view

of machine learning and the computational view

of machine learning, which is, I think,

one is worried about the equation, as it were.

And by the way, this is not a value judgment.

I just think it’s about perspective.

But the kind of questions you would ask

when you start asking yourself,

well, what does it mean to program

and develop and build the system,

is a very computer sciencey view of the problem.

I mean, if you get on data science Twitter

and econ Twitter, you actually hear this a lot

with the economist and the data scientist

complaining about the machine learning people.

Well, it’s just statistics,

and I don’t know why they don’t see this.

But they’re not even asking the same questions.

They’re not thinking about it

as a kind of programming problem.

And I think that that really matters,

just asking this question.

I actually think it’s a little different

from programming in hyperparameter space

and sort of collecting the data.

But I do think that that immersion really matters.

So I’ll give you a quick example

of the way I think about this.

So I teach machine learning.

Michael and I have co taught a machine learning class,

which has now reached, I don’t know, 10,000 people at least

over the last several years, or somewhere there’s abouts.

And my machine learning assignments are of this form.

So the first one is something like,

implement these five algorithms,

KNN and SVMs and boosting and decision trees

and neural networks, and maybe that’s it, I can’t remember.

And when I say implement, I mean steal the code.

I am completely uninterested.

You get zero points for getting the thing to work.

I don’t want you spending your time worrying about

getting the corner case right of what happens

when you are trying to normalize distances

and the points on the thing.

And so you divide by zero.

I’m not interested in that, right?

Steal the code.

However, you’re going to run those algorithms

on two data sets.

The data sets have to be interesting.

What does it mean to be interesting?

Well, data sets interesting if it reveals differences

between algorithms, which presumably are all the same

because they can represent whatever they can represent.

And two data sets are interesting together

if they show different differences, as it were.

And you have to analyze them.

You have to justify their interestingness

and you have to analyze them in a whole bunch of ways.

But all I care about is the data in your analysis,

not the programming.

And I occasionally end up in these long discussions

with students, well, I don’t really,

I copy and paste the things that I’ve said

the other 15,000 times it’s come up,

which is they go, but the only way to learn,

really understand is to code them up,

which is a very programmer,

software engineering view of the world.

If you don’t program it, you don’t understand it,

which is, by the way, I think is wrong

in a very specific way.

But it is a way that you come to understand

because then you have to wrestle with the algorithm.

But the thing about machine learning

is it’s not just sorting numbers

where in some sense the data doesn’t matter.

What matters is, well, does the algorithm work

on these abstract things, one less than the other?

In machine learning, the data matters.

It matters more than almost anything.

And not everything, but almost anything.

And so as a result, you have to live with the data

and don’t get distracted by the algorithm per se.

And I think that that focus on the data

and what it can tell you

and what question it’s actually answering for you

as opposed to the question you thought you were asking

is a key and important thing about machine learning

and is a way that computationalists

as opposed to statisticians bring a particular view

about how to think about the process.

The statisticians, by contrast, bring,

I think I’d be willing to say,

a better view about the kind of formal math that’s behind it

and what an actual number ultimately is saying

about the data.

And those are both important, but they’re also different.

I didn’t really think of it this way

is to build intuition about the role of data,

the different characteristics of data

by having two data sets that are different

and they reveal the differences in the differences.

That’s a really fascinating,

that’s a really interesting educational approach.

The students love it, but not right away.

No, they love it at the end.

They love it later.

They love it at the end.

Not at the beginning.

Not even immediately after.

I feel like there’s a deep profound lesson

about education there.

Yeah.

That you can’t listen to students

about whether what you’re doing is the right

or the wrong thing.

Yeah, well, as a wise, Michael Lippmann once said to me

about children, which I think applies to teaching,

is you have to give them what they need

without bending to their will.

And students are like that.

You have to figure out what they need.

You’re a curator.

Your whole job is to curate and to present

because on their own,

they’re not gonna necessarily know where to search.

So you’re providing pushes in some direction

and learn space and you have to give them what they need

in a way that keeps them engaged enough

so that they eventually discover what they want

and they get the tools they need to go

and learn other things off of.

What’s your view?

Let me put on my Russian hat,

which believes that life is suffering.

I like Russian hats, by the way.

If you have one, I would like this.

Those are ridiculous, yes.

But in a delightful way.

But sure.

What do you think is the role of,

we talked about balance a little bit.

What do you think is the role of hardship in education?

Like I think the biggest things I’ve learned,

like what made me fall in love with math, for example,

is by being bad at it until I got good at it.

So like struggling with a problem,

which increased the level of joy I felt

when I finally figured it out.

And it always felt with me, with teachers,

especially modern discussions of education,

how can we make education more fun,

more engaging, more all those things?

Or from my perspective, it’s like,

you’re maybe missing the point

that education, that life is suffering.

Education is supposed to be hard

and that actually what increases the joy you feel

when you actually learn something.

Is that ridiculous?

Do you like to see your students suffer?

Okay, so this may be a point where we differ.

I suspect not.

I’m gonna do go on.

Well, what would your answer be?

I wanna hear you first.

Okay, well, I was gonna not answer the question.

You don’t want the students to know you enjoy them suffering?

No, no, no, no, no, no.

I was gonna say that there’s,

I think there’s a distinction that you can make

in the kind of suffering, right?

So I think you can be in a mode

where you’re suffering in a hopeless way

versus you’re suffering in a hopeful way, right?

Where you’re like, you can see that if you,

that you still have,

you can still imagine getting to the end, right?

And as long as people are in that mindset

where they’re struggling,

but it’s not a hopeless kind of struggling,

that’s productive.

I think that’s really helpful.

But it’s struggling, like if you break their will,

if you leave them hopeless.

No, that don’t, sure, some people are gonna,

whatever, lift themselves up by their bootstraps,

but like mostly you give up

and certainly it takes the joy out of it.

And you’re not gonna spend a lot of time

on something that brings you no joy.

So it is a bit of a delicate balance, right?

You have to thwart people in a way

that they still believe that there’s a way through.

Right, so that’s a, we strongly agree actually.

So I think, well, first off,

struggling and suffering aren’t the same thing, right?

Yeah, just being poetic.

Oh, no, no, I actually appreciate the poetry.

And one of the reasons I appreciate it

is that they are often the same thing

and often quite different, right?

So you can struggle without suffering,

you can certainly suffer pretty easily.

You don’t necessarily have to struggle to suffer.

So I think that you want people to struggle,

but that hope matters.

You have to, they have to understand

that they’re gonna get through it on the other side.

And it’s very easy to confuse the two.

I actually think Brown University has a very,

just philosophically has a very different take

on the relationship with their students,

particularly undergrads from say a place like Georgia Tech,

which is.

Which university is better?

Well, I have my opinions on that.

I mean, remember, Charles said,

it doesn’t matter what the facts are, I’m always right.

The correct answer is that it doesn’t matter,

they’re different.

But clearly, clearly the answer is different.

He went to a school like the school

where he is as an undergrad.

I went to a school, specifically the same school,

though it was changed a bit in the intervening years.

Brown or Georgia Tech?

No, I was talking about Georgia Tech.

And I went to an undergrad place

that’s a lot like the place where I work now.

And so it does seem like we’re more familiar

with these models.

So there’s a similarity between Brown and Yale?

Yeah, I think they’re quite similar, yeah.

And Duke.

Duke has some similarities too,

but it’s got a little Southern draw.

You’ve kind of worked your,

you’ve sort of worked at universities

that are like the places where you learned.

And the same would be true for me.

Are you uncomfortable venturing outside the box?

Is that what you’re saying?

Journeying out?

That’s not what I’m saying.

Yeah, Charles is definitely.

He only goes to places

that have institute in the name, right?

It has worked out that way.

Well, academic places anyway.

Well, no, I was a visiting scientist at UPenn

or visiting something at UPenn.

Oh, wow, I just understood your joke.

Which one?

Five minutes later.

I like to set the sort of time bomb.

The institute is in the,

that Charles only goes to places

that have institute in the name.

So I guess Georgia,

I forget that Georgia Tech

is Georgia Institute of Technology.

The number of people who refer to it

as Georgia Tech University is large

and incredibly irritating.

It’s one of the few things

that genuinely gets under my skin.

But like schools like Georgia Tech and MIT

have as part of the ethos,

like there is,

I wanna say there’s an abbreviation

that someone taught me,

like IHTFP, something like that.

Like there’s an expression

which is basically I hate being here,

which they say so proudly.

And that is definitely not the ethos at Brown.

Like Brown is,

there’s a little more pampering

and empowerment and stuff.

And it’s not like we’re gonna crush you

and you’re gonna love it.

So yeah, I think there’s a,

I think the ethoses are different.

That’s interesting, yeah.

We had Drown Proofing.

What’s that?

In order to graduate from Georgia Tech,

this is a true thing.

Feel free to look it up.

If you.

A lot of schools have this by the way.

No, actually Georgia Tech was barely the first.

Brandeis has it.

Had it.

I feel like Georgia Tech was the first

in a lot of things.

It was the first in a lot of things.

Had the first master’s degree.

First Bumblebee mascot.

Stop that.

First master’s in computer science actually.

Right, online master’s.

Well that too, but way back in the 60s.

NSF grant.

Yeah, yeah.

You’re the first information

and computer science master’s degree in the country.

But the Georgia Tech,

it used to be the case

that in order to graduate from Georgia Tech,

you had to take a Drown Proofing class.

Where effectively, they threw you in the water

and tied you up.

If you didn’t drown, you got to graduate.

Tied you up?

I believe so.

No.

There were certainly versions of it,

but I mean luckily they ended it

just before I had to graduate

because otherwise I would have never graduated.

It wasn’t going to happen.

I want to say 84, 83,

somewhere around then they ended it.

But yeah, you used to have to prove

you could tread water for some ridiculous amount of time

or you couldn’t graduate.

Two minutes.

No, it was more than two minutes.

I bet it was two minutes.

Okay, well we’ll look at it.

And it was in a bathtub.

Yeah, right.

You could just stare.

It was in a pool.

But it was a real thing.

But that idea that, you know, push you.

Fully clothed.

Yeah, fully clothed.

I bet it was that and not tied up.

Because who needs to learn how to swim when you’re tied?

Nobody.

But who needs to learn to swim

when you’re actually falling into the water dressed?

That’s a real thing.

I think your facts are getting in the way

with a good story.

Oh, that’s fair.

That’s fair.

I didn’t mean to.

All right, so they tie you up.

Sometimes the narrative matters.

But whatever it was, you had to,

it was called drown proofing for a reason.

The point of the story, Michael, is that it’s,

well, no, but that’s good.

It doesn’t bring it back to struggle.

That’s a part of what Georgia Tech has always been.

And we struggle with that, by the way,

about what we want to be, particularly as things go.

But you sort of,

how much can you be pushed without breaking?

And you come out of the other end stronger, right?

There’s a saying we used to have

when I was an undergrad there.

It was just Georgia Tech,

building tomorrow the night before.

Right?

And it was just kind of idea that,

give me something impossible to do

and I’ll do it in a couple of days

because that’s what I just spent

the last four or five or six years with.

That ethos definitely stuck to you.

Having now done a number of projects with you,

you definitely will do it the night before.

That’s not entirely true.

There’s nothing wrong with waiting until the last minute.

The secret is knowing when the last minute is.

Right, that’s brilliantly put.

Yeah, that is a definite Charles statement

that I am trying not to embrace.

And I appreciate that

because you helped move my last minute up.

That’s the social construct

the way you converge together

what the definition of last minute is.

We figure that all out together.

In fact, MIT, I’m sure a lot of universities have this,

but MIT has like MIT time

that everyone has always agreed together

that there is such a concept

and everyone just keeps showing up like 10 to 15 to 20,

depending on the department, late to everything.

So there’s like a weird drift that happens.

It’s kind of fascinating.

Yeah, we’re five minutes.

We’re five minutes.

In fact, the classes will say,

well, this is no longer true actually,

but it used to be a class that started at eight,

but actually it started at eight oh five,

it ends at nine, actually it ends at eight fifty five.

Everything’s five minutes off

and nobody expects anything to start until five minutes

after the half hour, whatever it is.

It still exists.

It hurts my head.

Well, let’s rewind the clock back to the fifties and sixties

when you guys met, how did you,

I’m just kidding, I don’t know.

But what, can you tell the story of how you met?

So you’ve, like the internet and the world

kind of knows you as connected in some ways

in terms of education of teaching the world.

That’s like the public facing thing,

but how did you as human beings

and as collaborators meet?

I think there’s two stories.

One is how we met and the other is how we

got to know each other.

I’m not gonna say fell in love.

I’m gonna say that we came to understand that we

Had some common something.

Yeah, it’s funny.

Cause on the surface, I think we’re different

in a lot of ways, but there’s something

Yeah, I mean, now we complete each other’s

There you go.

Afternoon.

So I will tell the story of how we met

and I’ll let Michael tell the story of how we met.

Okay, all right.

Okay, so here’s how we met.

I was already at that point, it was AT&T labs.

There’s a long, interesting story there.

But anyway, I was there and Michael was coming to interview.

He was a professor at Duke at the time,

but decided for reasons that he wanted to be in New Jersey.

And so that would mean Bell Labs slash AT&T labs.

And we were doing the interview.

Interviews are very much like academic interviews.

And so I had to be there.

We all had to meet with him afterwards

and so on, one on one.

But it was obvious to me that he was going to be hired.

Like no matter what, because everyone loved him.

They were just talking about all the great stuff he did.

Oh, he did this great thing.

And you had just won something at AAAI, I think.

Or maybe you got 18 papers in AAAI that year.

I got the best paper award at AAAI

for the crossword stuff.

Right, exactly.

So that had all happened and everyone was going on

and on and on about it.

Actually, so Tinder was saying incredibly nice things

about you.

Really?

Yes.

He can be very grumpy.

Yes.

That’s nice to hear.

He was grumpily saying very nice things.

Oh, that makes sense.

And that does make sense.

So, you know, it was going to come.

So why was I meeting him?

I had something else I had to do.

I can’t remember what it was.

It probably involved comedy.

So he remembers meeting me

as inconveniencing his afternoon.

So he came.

So I eventually came to my office.

I was in the middle of trying to do something.

I can’t remember what.

And he came and he sat down.

And for reasons that are purely accidental,

despite what Michael thinks,

my desk at the time was set up

in such a way that it had sort of an L shape.

And the chair on the outside was always lower

than the chair that I was in.

And, you know, the kind of point was to…

The only reason I think that it was on purpose

is because you told me it was on purpose.

I don’t remember that.

Anyway, the thing is, is that, you know, it kind of gives…

His guest chair was really low

so that he could look down at everybody.

The idea was just to simply create a nice environment

that you were asking for a mortgage

and I was going to say no.

That was the point.

It was a very simple idea here.

Anyway, so we sat there

and we just talked for a little while.

And I think he got the impression that I didn’t like him.

Which wasn’t true.

I strongly got that impression.

The talk was really good.

The talk, by the way, was terrible.

And right after the talk,

I said to my host, Michael Kearns,

who ultimately was my boss.

I’m a huge fan.

I’m a friend and a huge fan of Michael, yeah.

Yeah, he is a remarkable person.

After my talk, I went into the…

He went into basketball.

I went…

Racquetball, he’s good at everything.

No, basketball.

No, but basketball and racquetball too.

Squash.

Squash, squash, squash, not racquetball.

Yes, squash, which is not…

Racquetball, yes.

Squash, no.

And I hope you hear that, Michael.

Oh, Michael Kearns.

As a game, not his skill level,

because I’m pretty sure he’s…

All right, there’s some competitiveness there,

but the point is that it was like the middle of the day,

I had full day of interviews.

I got met with people,

but then in the middle of the day, I gave a job talk.

And then there was gonna be more interviews,

but I pulled Michael aside and I said,

I think it’s in both of our best interest

if I just leave now, because that was so bad

that it’d just be embarrassing

if I have to talk to any more people.

You look bad for having invited me.

It’s just, let’s just forget this ever happened.

So I don’t think the talk went well.

That’s one of the most Michael Lipman set of sentences

I think I’ve ever heard.

He did great, or at least everyone knew he was great,

so maybe it didn’t matter.

I was there, I remember the talk,

and I remember him being very much the way

I remember him now, on any given week.

So it was good.

And we met and we talked about stuff.

He thinks I didn’t like him, but…

Because he was so grumpy.

Must’ve been the chair thing.

The chair thing and the low voice, I think.

But like, he obviously…

And that slight skeptical look.

Yes.

I have no idea what you’re talking about.

Well, I probably didn’t have any idea

what you were talking about.

Anyway, I liked him.

He asked me questions, I answered questions.

I felt bad about myself.

It was a normal day.

And then he left.

And then he left, and that’s how you met.

Can we take a…

And then I got hired and I was in the group.

Can we take a slight tangent on this topic of,

it sounds like, maybe you could speak

to the bigger picture.

It sounds like you’re quite self critical.

Who, Charles?

No, you.

Oh.

I can do better.

Try me again.

I’ll do better.

I’ll be so self critical.

I won’t.

Yeah, that was like a three out of 10 response.

So let’s try to work it up to five and six.

Yeah, I remember Marvin Minsky said on a video interview,

something that the key to success in academic research

is to hate everything you do.

For some reason…

I think I followed that because I hate everything he’s done.

That’s a good line.

That’s a six out of 10.

Maybe that’s a keeper.

But do you find that resonates with you at all

in how you think about talks and so on?

I would say it differently.

It’s not that.

No, not really.

That’s such an MIT view of the world though.

So I remember talking about this when, as a student,

you were basically told I will clean it up

for the purpose of the podcast.

My work is crap.

Then you go to a conference or something.

You’re like, everybody else’s work is crap.

Everybody else’s work is crap.

And you feel better and better about it, relatively speaking.

And then you sort of keep working on it.

I don’t hate my work.

That resonates with me.

Yes, I’ve never hated my work,

but I have been dissatisfied with it.

And I think being dissatisfied,

being okay with the fact that you’ve taken a positive step,

the derivative’s positive,

maybe even the second derivative’s positive,

that’s important because that’s a part of the hope, right?

But you have to, but I haven’t gotten there yet.

If that’s not there, that I haven’t gotten there yet,

then it’s hard to move forward, I think.

So I buy that, which is a little different

from hating everything that you do.

Yeah, I mean, there’s things that I’ve done

that I like better than I like myself.

So it’s separating me from the work, essentially.

So I think I am very critical of myself,

but sometimes the work I’m really excited about.

And sometimes I think it’s kind of good.

Does that happen right away?

So I found the work that I’ve liked, that I’ve done,

most of it, I liked it in retrospect

more when I was far away from it in time.

I have to be fairly excited about it to get done.

No, excited at the time, but then happy with the result.

But years later, or even I might go back,

you know what, that actually turned out to matter.

That turned out to matter.

Or, oh gosh, it turns out I’ve been thinking about that.

It’s actually influenced all the work that I’ve done since

without realizing it.

Boy, that guy was smart.

Yeah, that guy had a future.

Yeah, I think there’s something to it.

I think there’s something to the idea

you’ve got to hate what you do, but it’s not quite hate.

It’s just being unsatisfied.

And different people motivate themselves differently.

I don’t happen to motivate myself with self loathing.

I happen to motivate myself with something else.

So you’re able to sit back and be proud of,

in retrospect, of the work you’ve done.

Well, and it’s easier when you can connect it

with other people, because then you can be proud of them.

Proud of the people, yeah.

And then the question is.

You can still safely hate yourself.

Yeah, that’s right.

It’s win, win, Michael.

Or at least win, lose, which is what you’re looking for.

Oh, wow, there’s so many brilliant minds in this.

There’s levels.

So how did you actually meet me?

Yeah, Michael.

So the way I think about it is,

because we didn’t do much research together at AT&T,

but then we all got laid off.

So that sucked.

By the way, sorry to interrupt,

but that was one of the most magical places

historically speaking.

They did not appreciate what they had.

And how do we,

I feel like there’s a profound lesson in there too.

How do we get it, like what was, why was it so magical?

Is it just a coincidence of history?

Or is there something special about?

There were some really good managers

and people who really believed in machine learning

as this is gonna be important.

Let’s get the people who are thinking about this

in creative and insightful ways

and put them in one place and stir.

Yeah, but even beyond that, right?

It was Bell Labs at its heyday.

And even when we were there, which I think was past its heyday.

And to be clear, he’s gotten to be at Bell Labs.

I never got to be at Bell Labs.

I joined after that.

Yeah, I showed up in 91 as a grad student.

So I was there for a long time, every summer, except for two.

So twice I worked for companies

that had just stopped being Bell Labs.

Bellcore and then AT&T Labs.

So Bell Labs was several locations or for the research

or is it one?

I don’t know if Jersey’s involved somehow.

They’re all in Jersey.

Yeah, they’re all over the place.

But they were in a couple of places in Jersey.

Murray Hill was the Bell Labs place.

So you had an office in Murray Hill

at one point in your career.

Yeah, and I played Ultimate Frisbee

on the cricket pitch at Bell Labs at Murray Hill.

And then it became AT&T Labs when it split off

with loose during what we called Trivestiture.

Are you better than Michael Korn’s at Ultimate Frisbee?

Yeah. Oh, yeah.

Okay.

But I think that one’s not boasting.

I think Charles plays a lot of Ultimate

and I don’t think Michael does.

Yes, but that wasn’t the point.

The point is yes.

I’m finally better.

Sorry.

Okay, I have played on a championship winning

Ultimate Frisbee team or whatever,

Ultimate team with Charles.

So I know how good he is.

He’s really good.

How good I was anyway, when I was younger.

But the thing is.

I know how young he was when he was younger.

That’s true.

So much younger than now.

He’s older now.

Yeah, I’m older.

Michael was a much better basketball player than I was.

Michael Kearns.

Yes, no, not Michael.

Let’s be very clear about that.

To be clear, I’ve not played basketball with you.

So you don’t know how terrible I am,

but you have a probably pretty good guess.

And that you’re not as good as Michael Kearns.

He’s tall and athletic.

And he cared about it.

He’s very athletic.

He’s very good.

And probably competitive.

I love hanging out with Michael.

Anyway, but we were talking about something else,

although I no longer remember what it was.

What were we talking about?

Oh, Bell Labs.

But also Labs.

So this was kind of cool about what was magical about it.

The first thing you have to know

is that Bell Labs was an arm of the government, right?

Because AT&T was an arm of the government.

It was a monopoly.

And every month you paid a little thing on your phone bill,

which turned out was a tax

for all the research that Bell Labs was doing.

And they invented transistors and the laser

and whatever else is that they did.

The Big Bang or whatever, the cosmic background radiation.

Yeah, they did all that stuff.

They had some amazing stuff with directional microphones,

by the way.

I got to go in this room

where they had all these panels and everything.

And we would talk and one another,

and he’d move some panels around.

And then he would have me step two steps to the left.

And I couldn’t hear a thing he was saying

because nothing was bouncing off the walls.

And then he would shut it all down

and you could hear your heartbeat,

which is deeply disturbing to hear your heartbeat.

You can feel it.

I mean, you can feel it now.

There’s just so much all this sort of noise around.

Anyway, Bell Labs was about pure research.

It was a university, in some sense,

the purest sense of a university, but without students.

So it was all the faculty working with one another

and students would come in to learn.

They would come in for three or four months

during the summer and they would go away.

But it was just this kind of wonderful experience.

I could walk out my door.

In fact, I would often have to walk out my door

and deal with Rich Sutton and Michael Kearns

yelling at each other about whatever it is

they were yelling about the proper way

to prove something or another.

And I could just do that.

And Dave McAllister and Peter Stone

and all of these other people,

including, it’s a tender and then eventually Michael.

And it was just a place where you could think thoughts.

And it was okay because so long as once every 25 years or so

somebody invented a transistor, it paid for everything else.

You could afford to take the risk.

And then when that all went away,

it became harder and harder and harder to justify it

as far as the folks who were very far away were concerned.

And there was such a fast turnaround

among mental management on the AT&T side

that you never had a chance to really build a relationship.

At least people like us didn’t have a chance

to build a relationship.

So when the diaspora happened, it was amazing, right?

Everybody left and I think everybody ended up

at a great place and made a huge,

continued to do really good work with machine learning.

But it was a wonderful place.

And people will ask me, what’s the best job you’ve ever had?

And as a professor, anyway, the answer that I would give is

well, probably Bell Labs in some very real sense.

And I will never have a job like that again

because Bell Labs doesn’t exist anymore.

And Microsoft research is great and Google does good stuff.

And you can pick IBM, you can tell if you want to,

but Bell Labs was magical.

It was around for, it was an important time

and it represents a high watermark

in basic research in the US.

Is there something you could say about the physical proximity

and the chance collisions?

Like we live in this time of the pandemic

where everyone is maybe trying to see the silver lining

and accepting the remote nature of things.

Is there one of the things that people like faculty

that I talk to miss is the procrastination.

Like the chance to make everything is about meetings

that are supposed to be,

there’s not a chance to just talk about comic book

or whatever, like go into discussion that’s totally pointless.

So it’s funny you say this

because that’s how we met, met, it was exactly that.

So I’ll let Michael say that, but I’ll just add one thing

which is just that research is a social process

and it helps to have random social interactions

even if they don’t feel social at the time,

that’s how you get things done.

One of the great things about the AI Lab when I was there,

I don’t quite know what it looks like now

once they moved buildings,

but we had entire walls that were whiteboards

and people would just get up there

and they were just right and people would walk up

and you’d have arguments

and you’d explain things to one another

and you got so much out of the freedom to do that.

You had to be okay with people challenging

every fricking word you said,

which I would sometimes find deeply irritating,

but most of the time it was quite useful.

But the sort of pointlessness and the interaction

was in some sense the point, at least for me.

Yeah, I think offline yesterday I mentioned

Josh Tenenbaum and he’s very much, he’s a man,

he’s such an inspiration in the childlike way

that he pulls you in on any topic.

It doesn’t even have to be about machine learning

or the brain, he’ll just pull you in

to a closest writable surface,

which is still, you can find whiteboards

at MIT everywhere, and just like basically cancel

all meetings and talk for a couple hours

about some aimless thing and it feels like

the whole world, the time space continuum kind of warps

and that becomes the most important thing.

And then it’s just, it’s definitely something

worth missing in this world where everything’s remote.

There’s some magic to the physical presence.

Whenever I wonder myself whether MIT really is

as great as I remember it, I just go talk to Josh.

Yeah, you know, that’s funny.

There’s a few people in this world that carry

the best of what particular institutions stand for, right?

And it’s.

It’s Josh.

I mean, I don’t, my guess is he’s unaware of this.

That’s the point.

Yeah.

That the masters are not aware of their mastery.

So.

How did we meet?

Yes, but first a tangent, no.

How did you meet me?

So I’m not sure what you were thinking,

but when it started to dawn on me

that maybe we had a longer term bond

was after we all got laid off.

And you had decided at that point

that we were still paid.

We were given an opportunity to like do a job search

and kind of make a transition,

but it was clear that we were done.

And I would go to my office to work

and you would go to my office to keep me from working.

That was my recollection of it.

You had decided that there was no,

really no point in working for the company

because our relationship with the company was done.

Yeah, but remember I felt that way beforehand.

It wasn’t about the company.

It was about the set of people there

doing really cool things.

And it always, always been that way.

But we were working on something together.

Oh yeah, yeah, yeah.

That’s right.

So at the very end, we all got laid off,

but then our boss came to, our boss’s boss came to us

because our boss was Michael Kearns

and he had jumped ship brilliantly, like perfect timing.

Like things like right before the ship was about to sink,

he was like, gotta go and landed perfectly

Because Michael Kearns.

And leaving the rest of us to go like, this is fine.

And then it was clear that it wasn’t fine

and we were all toast.

So we had this sort of long period of time.

But then our boss figured out, okay, wait,

maybe we can save a couple of these people

if we can have them do something really useful.

And the useful thing was we were gonna make

basically an automated assistant

that could help you with your calendar.

You could like tell it things

and it would respond appropriately.

It would just kind of integrate across

all sorts of your personal information.

And so me and Charles and Peter Stone

were set up as the crack team

to actually solve this problem.

Other people maybe were too theoretical that they thought,

but we could actually get something done.

So we sat down to get something done

and there wasn’t time and it wouldn’t have saved us anyway.

And so it all kind of went downhill.

But the interesting, I think, coda to that

is that our boss’s boss is a guy named Ron Brockman.

And when he left AT&T,

cause we were all laid off,

he went to DARPA, started up a program there

that became KALO,

which is the program from which Siri sprung,

which is a digital assistant

that helps you with your calendar

and a bunch of other things.

It really, in some ways got its start

with me and Charles and Peter trying to implement this vision

that Ron Brockman had,

that he ultimately got implemented

through his role at DARPA.

So when I’m trying to feel less bad

about having been laid off

from what is possibly the greatest job of all time,

I think about, well, we kind of helped birth Siri.

That’s something.

And then he did other things too.

But we got to spend a lot of time in his office

and talk about lots of things.

We got to spend a lot of time in my office, yeah.

Yeah, yeah.

And so then we went on our merry way.

Everyone went to different places.

Charles landed at Georgia Tech,

which was what he always dreamed he would do.

And so that worked out well.

I came up with a saying at the time,

which is luck favors the Charles.

It’s kind of like luck favors the prepared,

but Charles, like he wished something

and then it would basically happen just the way he wanted.

It was inspirational to see things go that way.

Things worked out.

And we stayed in touch.

And then I think it really helped

when you were working on,

I mean, you’d kept me in the loop for things like threads

and the work that you were doing at Georgia Tech.

But then when they were starting

their online master’s program,

he knew that I was really excited about MOOCs

and online teaching.

And he’s like, I have a plan.

And I’m like, tell me your plan.

He’s like, I can’t tell you the plan yet.

Cause they were deep in negotiations

between Georgia Tech and Udacity to make this happen.

And they didn’t want it to leak.

So Charles would kept teasing me about it,

but wouldn’t tell me what was actually going on.

And eventually it was announced and he said,

I would like you to teach the machine learning course

with me.

I’m like, that can’t possibly work.

But it was a great idea.

And it was super fun.

It was a lot of work to put together,

but it was really great.

Was that the first time you thought about,

first of all, was it the first time

you got seriously into teaching?

I mean, I was a professor.

This was already after you jumped to,

so like there’s a little bit of jumping around in time.

Yeah, sorry about that.

There’s a pretty big jump in time.

So like the MOOCs thing.

So Charles got to Georgia Tech and he,

I mean, maybe Charles, maybe this is a Charles story.

I got to Georgia Tech in 2002.

He got to Georgia Tech in 2002.

And worked on things like revamping the curriculum,

the undergraduate curriculum,

so that it had some kind of semblance of modular structure

because computer science was at the time

moving from a fairly narrow specific set of topics

to touching a lot of other parts of intellectual life.

And the curriculum was supposed to reflect that.

And so Charles played a big role in kind of redesigning that.

And then the.

And for my labors, I ended up as associate dean.

Right, he got to become associate dean

of charge of educational stuff.

Yeah, I was under.

This should be a valuable lesson.

If you’re good at something,

they will give you responsibility to do more of that thing.

Well.

Until you.

Don’t show competence.

Don’t show competence if you.

Don’t want responsibility.

Here’s what they say.

The reward for good work is more work.

The reward for bad work is less work.

Which, I don’t know.

Depending on what you’re trying to do that week,

one of those is better than the other.

Well, one of the problems with the word work,

sorry to interrupt, is that it seems to be an antonym

in this particular language.

We have the opposite of happiness.

But it seems like they’re.

That’s one of, you know, we talked about balance.

It’s always like work life balance.

It always rubbed me the wrong way as a terminology.

I know it’s just words.

Right, the opposite of work is play.

But ideally, work is play.

Oh, I can’t tell you how much time I’d spend.

Certainly, when I was at Bell Labs,

except for a few very key moments,

as a professor, I would do this too.

I would just say, I cannot believe

they’re paying me to do this.

Because it’s fun.

It’s something that I would do for a hobby

if I could anyway.

So that sort of worked out.

Are you sure you want to be saying that

when this is being recorded?

As a dean, that is not true at all.

I need a raise.

But I think here with this,

even though a lot of time passed,

Mike and I talked almost every, well, we texted,

almost every day during the period.

Charles, at one point, took me,

the ICML conference, the machine learning conference

was in Atlanta.

I was the chair, the general chair of the conference.

Charles was my publicity chair or something like that,

or fundraising chair.

Yeah, but he decided it’d be really funny

if he didn’t actually show up for the conference

in his own home city.

So he didn’t, but he did at one point

pick me up at the conference in his Tesla

and drove me to the Atlanta mall

and forced me to buy an iPhone

because he didn’t like how it was to text with me

and thought it would be better for him

if I had an iPhone, the text would be somehow smoother.

And it was.

And it is, and his life is better.

And my life is better.

And so, yeah, but it was, yeah,

Charles forced me to get an iPhone

so that he could text me more efficiently.

I thought that was an interesting moment.

It works for me.

Anyway, so we kept talking the whole time

and then eventually we did the teaching thing

and it was great.

And there’s a couple of reasons for that, by the way.

One is I really wanted to do something different.

Like you’ve got this medium here,

people claim it can change things.

What’s a thing that you could do in this medium

that you could not do otherwise besides edit, right?

I mean, what could you do?

And being able to do something with another person

was that kind of thing.

It’s very hard.

I mean, you can take turns,

but teaching together, having conversations is very hard.

So that was a cool thing.

The second thing, give me an excuse

to do more stuff with him.

Yeah, I always thought, he makes it sound brilliant.

And it is, I guess.

But at the time it really felt like

I’ve got a lot to do, Charles is saying,

and it would be great if Michael could teach the course

and I could just hang out.

Yeah, just kind of coast on that.

Well, that’s what the second class was more like that.

Because the second class was explicitly like that.

But the first class, it was at least half.

Yeah, but I do all the stuff.

So the structure that we came up with.

I think you’re once again letting the facts

get in the way of a good story.

I should just let Charles talk to us.

But that’s the facts that he saw.

So that was kind of true for 7642.

Yeah, that was sort of true for 7642,

which is the reinforcement learning class,

because that was really his class.

You started with reinforcement learning or machine learning?

Intro machine learning, 7641,

which is supervised learning, unsupervised learning,

and reinforcement learning and decision making,

cram all that in there,

the kind of assignments that we talked about earlier.

And then eventually, about a year later,

we did a follow on 7642,

which is reinforcement learning and decision making.

The first class was based on something

I’d been teaching at that point for well over a decade.

And the second class was based on something

Michael had been teaching.

Actually, I learned quite a bit

teaching that class with him, but he drove most of that.

But the first one I drove most, it was all my material.

Although I had stolen that material originally

from slides I found online from Michael,

who had originally stolen that material

from, I guess, slides he found online,

probably from Andrew Moore,

because the jokes were the same anyway.

At least some of the, at least when I found the slides,

some of the stuff with it.

Is that true?

Yes, every machine learning class taught in the early 2000s

stole from Andrew Moore.

A particular joke or two?

At least the structure.

Now, I did, and he did, actually,

a lot more with reinforcement learning and such,

and game theory and those kinds of things.

But, you know, we all sort of built in.

You mean in the research world?

No, no, no, in that class.

No, I mean in teaching that class.

The coverage was different than what we started.

Most people were just doing supervised learning

and maybe a little bit of clustering and whatnot,

but we took it all the way to machine learning.

A lot of it just comes from Tom Mitchell’s book.

Oh, no, yeah, except, well,

half of it comes from Tom Mitchell’s book, right?

I mean, the other half doesn’t.

This is why it’s all readings, right?

Because certain things weren’t invented

when Tom wrote that stuff.

Yeah, okay, that’s true.

All right, but it was quite good.

But there’s a reason for that besides, you know,

just, I wanted to do it.

I wanted to do something new,

and I wanted to do something with him,

which is a realization,

which is despite what you might believe,

he’s an introvert and I’m an introvert,

or I’m on the edge of being an introvert anyway.

But both of us, I think, enjoy the energy of the crowd,

right?

There’s something about talking to people

and bringing them into whatever we find interesting

that is empowering, energizing, or whatever.

And I found the idea of staring alone at a computer screen

and then talking off of materials

less inspiring than I wanted it to be.

And I had in fact done a MOOC for Udacity on algorithms.

And it was a week in a dark room talking at the screen,

writing on the little pad.

And I didn’t know this was happening,

but they had watched,

the crew had watched some of the videos

while, you know, like in the middle of this,

and they’re like, something’s wrong.

You’re sort of shutting down.

And I think a lot of it was I’ll make jokes

and no one would laugh.

And I felt like the crowd hated me.

Now, of course, there was no crowd.

So like, it wasn’t rational.

But each time I tried it and I got no reaction,

it just was taking the energy out of my performance,

out of my presentation.

Such a fantastic metaphor for grad school.

Anyway, by working together,

we could play off each other and have a good time.

And keep the energy up,

because you can’t let your guard down for a moment

with Charles, he’ll just overpower you.

I have no idea what you’re talking about.

But we would work really well together, I thought,

and we knew each other,

so I knew that we could sort of make it work.

Plus, I was the associate dean,

so they had to do what I told them to do.

We had to make it work.

And so it worked out very well, I thought,

well enough that we.

With great power comes great power.

That’s right.

And we became smooth and curly.

And that’s when we did the overfitting thriller video.

Yeah, that’s a thing.

So can we just, like, smooth and curly,

where did that come from?

Okay, so it happened.

It was completely spontaneous.

These are nicknames you go by.

Yeah, so it’s what the students call us.

He was lecturing.

So the way that we structured the lectures

is one of us is the lecturer

and one of us is basically the student.

And so he was lecturing on.

The lecturer prepares all the materials,

comes up with the quizzes,

and then the student comes in not knowing anything.

So it was just like being on campus.

And I was doing game theory in particular,

the Prisoner’s Dilemma.

Prisoner’s Dilemma.

And so he needed to set up a little Prisoner’s Dilemma grid.

So he drew it and I could see what he was drawing.

And the Prisoner’s Dilemma consists of two players,

two parties.

So he decided he would make little cartoons

of the two of us.

And so there was two criminals, right,

that were deciding whether or not to rat each other out.

One of them he drew as a circle with a smiley face

and a kind of goatee thing, smooth head.

And the other one with all sorts of curly hair.

And he said, this is smooth and curly.

I said, smooth and curly?

He said, no, no, smooth with a V.

It’s very important that it have a V.

And then the students really took to that.

Like they found that relatable.

He started singing Smooth Criminal by Michael Jackson.

Yeah, yeah, yeah.

And those names stuck.

So we now have a video series,

an episode, our kind of first actual episode

should be coming out today,

Smooth and Curly on video,

where the two of us discuss episodes of Westworld.

We watch Westworld and we’re like, huh,

what does this say about computer science and AI?

And we’ve never, we did not watch it.

I mean, no, it’s on season three or whatever we have.

As of this recording, it’s on season three.

We’ve watched now two episodes total.

Yeah, I think I watched three.

What do you think about Westworld?

Two episodes in.

So I can tell you so far,

I’m just guessing what’s gonna happen next.

It seems like bad things are gonna happen

with the robots uprising.

It’s a lot of.

Spoiler alert.

So I have not, I have not,

I mean, you know, I vaguely remember a movie existing.

So I assume it’s related to that, but.

That was more my time than your time, Charles.

That’s right, cause you’re much older than I am.

I think the important thing here is that

it’s narrative, right?

It’s all about telling a story.

That’s the whole driving thing.

But the idea that they would give these reveries,

that they would make people,

they would make them.

Let them remember.

Remember the awful things that happened.

The terrible things that happened.

Who could possibly think that was gonna,

I gotta, I mean, I don’t know.

I’ve only seen the first two episodes

or maybe the third one.

I think I’ve only seen the first one.

You know what it was?

You know what the problem is?

That the robots were actually designed by Hannibal Lecter.

That’s true.

They weren’t.

So like, what do you think is gonna happen?

Bad things.

It’s clear that things are happening

and characters are being introduced

and we don’t yet know anything,

but still I was just struck by how

it’s all driven by narrative and story.

And there’s all these implied things like programming,

the programming interface is talking to them

about what’s going on in their heads,

which is both, I mean, artistically,

it’s probably useful to film it that way.

But think about how it would work in real life.

That just seems very great.

But there was, we saw in the second episode,

there’s a screen.

You could see things.

They were wearing like Kubrick’s glasses.

In the world.

It was quite interesting to just kind of ask this question

so far.

I mean, I assume it veers off into Never Never Land

at some point.

So we don’t know.

We can’t answer that question.

I’m also a fan of a guy named Alex Garland.

He’s a director of Ex Machina.

Mm hmm.

And he is the first,

I wonder if Kubrick was like this actually,

is he like studies,

what would it take to program an AI systems?

Like he’s curious enough to go into that direction.

On the Westworld side,

I felt there was more emphasis on the narratives

than like actually asking like computer science questions.

Yeah.

Like, how would you build this?

How would you, and.

How would you debug it?

I still think, to me, that’s the key issue.

They were terrible debuggers.

Yeah.

Well, they said specifically,

so we make a change and we put it out in the world

and that’s bad because something terrible could happen.

Like if you’re putting things out in the world

and you’re not sure whether something terrible

is going to happen, your process is probably.

I just feel like there should have been someone

whose sole job it was to walk around and poke his head in

and say, what could possibly go wrong?

Just over and over again.

I would have loved if there was an,

and I did watch a lot more and I’m not giving anything away.

I would have loved it if there was like an episode

where like the new intern is like debugging

a new model or something and like it just keeps failing

and they’re like, all right.

And then it’s more turns into like a episode

of Silicon Valley or something like that.

Yes.

Versus like this ominous AI systems

that are constantly like threatening the fabric

of this world that’s been created.

Yeah.

Yeah, and you know the other,

this reminds me of something that,

so I agree that that should be very cool,

at least for the small percentage of people

who care about debugging systems.

But the other thing is.

Right, debugging, the series.

Yeah, it falls into, think of the sequels,

fear of the debugger.

Oh my gosh.

Anyway, so.

It’s a nightmare show, it’s a horror movie.

I think that’s where we lose people, by the way,

early on is the people who either decide,

either figure out debugging or think debugging is terrible.

This is where we lose people in computer science.

This is a part of the struggle versus suffering, right?

You get through it and you kind of get the skills of it,

or you’re just like, this is dumb,

and this is a dumb way to do anything.

And I think that’s when we lose people.

But, well, I’ll leave it at that.

But I think that there’s something really, really neat

about framing it that way.

But what I don’t like about all of these things,

and I love Tex Machina, by the way,

although the ending was very depressing.

One of the things I have to talk to Alex about,

he says that the thing that nobody noticed he put in

is at the end, spoiler alert,

the robot turns and looks at the camera and smiles, briefly.

And to him, he thought that his definition

of passing the general version of the Turing test,

or the consciousness test, is smiling for no one.

It’s like the Chinese room kind of experiment.

It’s not always trying to act for others,

but just on your own, being able to have a relationship

with the actual experience and just take it in.

I don’t know, he said nobody noticed the magic of it.

I have this vague feeling that I remember the smile,

but now you’ve just put the memory in my head,

so probably not.

But I do think that that’s interesting.

Although, by looking at the camera,

you are smiling for the audience, right?

You’re breaking the fourth wall.

It seems, I mean, well, that’s a limitation of the medium.

But I like that idea.

But here’s the problem I have with all of those movies,

all of them, is that, but I know why it’s this way,

and I enjoy those movies, and Westworld,

is it sets up the problem of AI as succeeding

and then having something we cannot control.

But it’s not the bad part of AI.

The bad part of AI is the stuff

we’re living through now, right?

It’s using the data to make decisions that are terrible.

It’s not the intelligence that’s gonna go out there

and surpass us and take over the world

or lock us into a room to starve to death slowly

over multiple days.

It’s instead the tools that we’re building

that are allowing us to make the terrible decisions

we would have less efficiently made before, right?

Computers are very good at making us more efficient,

including being more efficient at doing terrible things.

And that’s the part of the AI we have to worry about.

It’s not the true intelligence that we’re gonna build

sometime in the future, probably long after we’re around.

But I think that whole framing of it

sort of misses the point, even though it is inspiring.

And I was inspired by those ideas, right?

I got into this in part

because I wanted to build something like that.

Philosophical questions are interesting to me,

but that’s not where the terror comes from.

The terror comes from the everyday.

And you can construct situations

in the subtlety of the interaction between AI and the human,

like with social networks,

all the stuff you’re doing

with interactive artificial intelligence.

But I feel like Cal 9000 came a little bit closer to that

in 2001 Space Odyssey,

because it felt like a personal assistant.

It felt like closer to the AI systems we have today.

And the real things we might actually encounter,

which is over relying in some fundamental way

on our dumb assistants or on social networks,

like over offloading too much of us

onto things that require internet and power and so on

and thereby becoming powerless as a standalone entity.

And then when that thing starts to misbehave

in some subtle way, it creates a lot of problems.

And those problems are dramatized when you’re in space,

because you don’t have a way to walk away.

Well, as the man said,

once we started making the decisions for you,

it stopped being your world, right?

That’s the matrix, Michael, in case you don’t remember.

But on the other hand, I could say no,

because isn’t that what we do with people anyway?

You know, just kind of the shared intelligence

that is humanity is relying on other people constantly.

I mean, we hyper specialize, right?

As individuals, we’re still generally intelligent.

We make our own decisions in a lot of ways,

but we leave most of this up to other people.

And that’s perfectly fine.

And by the way, everyone doesn’t necessarily share our goals.

Sometimes they seem to be quite against us.

Sometimes we make decisions that others would see

as against our own interests.

And yet we somehow manage it, manage to survive.

I’m not entirely sure why an AI

would actually make that worse or even different, really.

You mentioned the matrix.

Do you think we’re living in a simulation?

It does feel like a thought game