Michael I. Jordan: Machine Learning, Recommender Systems, and the Future of AI
Guests: Michael I. Jordan
Full transcript · Read the full conversation in order, with its original context.
Contents10 chapters
§01Transcript · 00:00–13:01§02Transcript · 13:01–23:21§03Transcript · 23:21–33:54§04Transcript · 33:54–43:41§05Transcript · 43:41–54:00§06Transcript · 54:00–1:04:11§07Transcript · 1:04:11–1:15:41§08Transcript · 1:15:41–1:26:55§09Transcript · 1:26:55–1:38:18§10Transcript · 1:38:18–1:45:50§01Transcript · 00:00–13:01
The following is a conversation with Michael I. Jordan, a professor at Berkeley and one
of the most influential people in the history of machine learning, statistics, and artificial
intelligence.
He has been cited over 170,000 times and he has mentored many of the world class researchers
defining the field of AI today, including Andrew Ng, Zubin Garamani, Ben Taskar, and
Yoshua Bengio.
All this, to me, is as impressive as the over 32,000 points in the six NBA championships
of the Michael J. Jordan of basketball fame.
There's a nonzero probability that I talked to the other Michael Jordan given my connection
to and love of the Chicago Bulls of the 90s, but if I had to pick one, I'm going with
the Michael Jordan of statistics and computer science, or as Yann LeCun calls him, the Miles
Davis of machine learning.
In his blog post titled Artificial Intelligence, the Revolution Hasn't Happened Yet, Michael
argues for broadening the scope of the artificial intelligence field.
In many ways, the underlying spirit of this podcast is the same, to see artificial intelligence
as a deeply human endeavor, to not only engineer algorithms and robots, but to understand and
empower human beings at all levels of abstraction, from the individual to our civilization as
a whole.
This is the Artificial Intelligence Podcast.
If you enjoy it, subscribe and YouTube, give it five stars at Apple Podcast, support it
on Patreon, or simply connect with me on Twitter at Lex Friedman spelled F R I D M A N.
As usual, I'll do one or two minutes of ads now and never any ads in the middle that
can break the flow of the conversation.
I hope that works for you and doesn't hurt the listening experience.
This show is presented by Cash App, the number one finance app in the App Store.
When you get it, use code LEX PODCAST.
Cash App lets you send money to friends, buy Bitcoin, and invest in the stock market with
as little as $1.
Since Cash App does fractional share trading, let me mention that the order execution algorithm
that worked behind the scenes to create the abstraction of the fractional orders is to
me an algorithmic marvel.
Great props for the Cash App engineers for solving a hard problem that in the end provides
an easy interface that takes a step up to the next layer of abstraction over the stock
market, making trading more accessible for new investors and diversification much easier.
So once again, if you get Cash App from the App Store or Google Play and use the code
LEX PODCAST, you'll get $10 and Cash App will also donate $10 to First, one of my favorite
organizations that is helping to advance robotics and STEM education for young people around
the world.
And now, here's my conversation with Michael I. Jordan.
Given that you're one of the greats in the field of AI, machine learning, computer science,
and so on, you're trivially called the Michael Jordan of machine learning, although as you
know, you were born first, so technically MJ is the Michael I. Jordan of basketball.
But anyway, my favorite is Yann LeCun calling you the Miles Davis of machine learning because
as he says, you reinvent yourself periodically and sometimes leave fans scratching their
heads after you change direction.
So can you put at first your historian hat on and give a history of computer science
and AI as you saw it, as you experienced it, including the four generations of AI successes
that I've seen you talk about?
Sure.
Yeah, first of all, I much prefer Yann's metaphor.
Miles Davis was a real explorer in jazz and he had a coherent story.
So I think I have one, but it's not just the one you lived, it's the one you think about
later.
What the historian does is they look back and they revisit.
I think what's happening right now is not AI, that was an intellectual aspiration that's
still alive today as an aspiration.
But I think this is akin to the development of chemical engineering from chemistry or
electrical engineering from electromagnetism.
So if you go back to the 30s or 40s, there wasn't yet chemical engineering.
There was chemistry, there was fluid flow, there was mechanics and so on.
But people pretty clearly viewed interesting goals to try to build factories that make
chemicals products and do it viably, safely, make good ones, do it at scale.
So people started to try to do that, of course, and some factories worked, some didn't, some
were not viable, some exploded, but in parallel, developed a whole field called chemical engineering.
Electrical engineering is a field, it's no bones about it, it has theoretical aspects
to it, it has practical aspects.
It's not just engineering, quote unquote, it's the real thing, real concepts are needed.
Same thing with electrical engineering.
There was Maxwell's equations, which in some sense were everything you know about electromagnetism,
but you needed to figure out how to build circuits, how to build modules, how to put
them together, how to bring electricity from one point to another safely and so on and
so forth.
So a whole field that developed called electrical engineering.
I think that's what's happening right now, is that we have a proto field, which is statistics,
more of the theoretical side of it, algorithmic side of computer science, that was enough
to start to build things, but what things?
Systems that bring value to human beings and use human data and mix in human decisions.
The engineering side of that is all ad hoc.
That's what's emerging.
In fact, if you wanna call machine learning a field, I think that's what it is, that it's
a proto form of engineering based on statistical and computational ideas of previous generations.
But do you think there's something deeper about AI in his dreams and aspirations as
compared to chemical engineering and electrical engineering?
Well the dreams and aspirations maybe, but those are 500 years from now.
I think that that's like the Greeks sitting there and saying, it would be neat to get
to the moon someday.
I think we have no clue how the brain does computation.
We're just a clueless.
We're even worse than the Greeks on most anything interesting scientifically of our era.
Can you linger on that just for a moment because you stand not completely unique, but a little
bit unique in the clarity of that.
Can you elaborate your intuition of why we're, like where we stand in our understanding of
the human brain?
And a lot of people say, you know, scientists say we're not very far in understanding human
brain, but you're like, you're saying we're in the dark here.
Well, I know I'm not unique.
I don't even think in the clarity, but if you talk to real neuroscientists that really
study real synapses or real neurons, they agree, they agree.
It's a hundreds of year task and they're building it up slowly and surely.
What the signal is there is not clear.
We think we have all of our metaphors.
We think it's electrical, maybe it's chemical, it's a whole soup, it's ions and proteins
and it's a cell.
And that's even around like a single synapse.
If you look at a electron micrograph of a single synapse, it's a city of its own.
And that's one little thing on a dendritic tree, which is extremely complicated electrochemical
thing.
And it's doing these spikes and voltages are flying around and then proteins are taking
that and taking it down into the DNA and who knows what.
So it is the problem of the next few centuries.
It is fantastic.
But we have our metaphors about it.
Is it an economic device?
Is it like the immune system or is it like a layered set of, you know, arithmetic computations?
We have all these metaphors and they're fun.
But that's not real science per se.
There is neuroscience.
That's not neuroscience.
All right.
That's like the Greek speculating about how to get to the moon, fun, right?
And I think that I like to say this fairly strongly because I think a lot of young people
think we're on the verge because a lot of people who don't talk about it clearly let
it be understood that, yes, we kind of, this is a brain inspired, we're kind of close,
you know, breakthroughs are on the horizon.
And that's scrupulous people sometimes who need money for their labs.
That's what I'm saying, scrupulous, but people will oversell, I need money for my lab, I'm
studying computational neuroscience, I'm going to oversell it.
And so there's been too much of that.
So I'll step into the gray area between metaphor and engineering with, I'm not sure if you're
familiar with brain computer interfaces.
So a company like Elon Musk has Neuralink that's working on putting electrodes into
the brain and trying to be able to read, both read and send electrical signals.
Just as you said, even the basic mechanism of communication in the brain is not something
we understand.
But do you hope without understanding the fundamental principles of how the brain works,
we'll be able to do something interesting at that gray area of metaphor?
It's not my area.
So I hope in the sense, like anybody else hopes for some interesting things to happen
from research, I would expect more something like Alzheimer's will get figured out from
modern neuroscience.
There's a lot of human suffering based on brain disease and we throw things like lithium
at the brain, it kind of works, no one has a clue why.
That's not quite true, but mostly we don't know.
And that's even just about the biochemistry of the brain and how it leads to mood swings
and so on.
How thought emerges from that, we were really, really completely dim.
So that you might want to hook up electrodes and try to do some signal processing on that
and try to find patterns, fine, by all means, go for it.
It's just not scientific at this point.
So it's like kind of sitting in a satellite and watching the emissions from a city and
trying to infer things about the microeconomy, even though you don't have microeconomic concepts.
It's really that kind of thing.
And so yes, can you find some signals that do something interesting or useful?
Can you control a cursor or mouse with your brain?
Yeah, absolutely, and then I can imagine business models based on that and even medical applications
of that.
But from there to understanding algorithms that allow us to really tie in deeply from
the brain to computer, I just, no, I don't agree with Elon Musk.
I don't think that's even, that's not for our generations, not even for the century.
So just in hopes of getting you to dream, you've mentioned Kolmogorov and Turing might
pop up, do you think that there might be breakthroughs that will get you to sit back in five, 10
years and say, wow?
Oh, I'm sure there will be, but I don't think that there'll be demos that impress me.
I don't think that having a computer call a restaurant and pretend to be a human is
a breakthrough.
Right.
And people, you know, some people present it as such.
It's imitating human intelligence.
It's even putting coughs in the thing to make a bit of a PR stunt.
And so fine that the world runs on those things too.
And I don't want to diminish all the hard work and engineering that goes behind things
like that and the ultimate value to the human race.
But that's not scientific understanding.
And I know the people that work on these things, they are after scientific understanding.
In the meantime, they've got to kind of, you know, the trains got to run and they got mouths
to feed and they got things to do and there's nothing wrong with all that.
I would call that though, just engineering.
And I want to distinguish that between an engineering field, like electrical engineering
and chemical engineering that originally emerged, that had real principles and you really know
what you're doing and you had a little scientific understanding, maybe not even complete.
So it became more predictable and it really gave value to human life because it was understood.
And so we don't want to muddle too much these waters of, you know, what we're able to do
versus what we really can't do in a way that's going to impress the next.
So I don't need to be wowed, but I think that someone comes along in 20 years, a younger
person who's absorbed all the technology and for them to be wowed, I think they have to
be more deeply impressed.
A young Kolmogorov would not be wowed by some of the stunts that you see right now coming
from the big companies.
The demos, but do you think the breakthroughs from Kolmogorov would be, and give this question
a chance, do you think there'll be in the scientific fundamental principles arena or
do you think it's possible to have fundamental breakthroughs in engineering?
Meaning, you know, I would say some of the things that Elon Musk is working with SpaceX
and then others sort of trying to revolutionize the fundamentals of engineering, of manufacturing,
of saying, here's a problem we know how to do a demo of and actually taking it to scale.
Yeah.
So there's going to be all kinds of breakthroughs.
I just don't like that terminology.
I'm a scientist and I work on things day in and day out and things move along and eventually
you say, wow, something happened, but I don't like that language very much.
Also I don't like to prize theoretical breakthroughs over practical ones.
§02Transcript · 13:01–23:21
I tend to be more of a theoretician and I think there's lots to do in that arena right
now.
And so I wouldn't point to the Kolmogorovs, I might point to the Edisons of the era and
maybe Musk is a bit more like that.
But you know, Musk, God bless him, also will say things about AI that he knows very little
about and he leads people astray when he talks about things he doesn't know anything about.
Trying to program a computer to understand natural language, to be involved in a dialogue
we're having right now, that ain't going to happen in our lifetime.
You could fake it, you can mimic, sort of take old sentences that humans use and retread
them, but the deep understanding of language, no, it's not going to happen.
And so from that, I hope you can perceive that the deeper, yet deeper kind of aspects
and intelligence are not going to happen.
Now will there be breakthroughs?
No, I think that Google was a breakthrough, I think Amazon is a breakthrough, you know,
I think Uber is a breakthrough, you know, that bring value to human beings at scale
in new, brand new ways based on data flows and so on.
A lot of these things are slightly broken because there's not kind of an engineering
field that takes economic value in context of data and, you know, planetary scale and
worries about all the externalities, the privacy, you know, we don't have that field so we don't
think these things through very well.
I see that as emerging and that will be, you know, looking back from 100 years, that will
be a constituted breakthrough in this era, just like electrical engineering was a breakthrough
in the early part of the last century and chemical engineering was a breakthrough.
So the scale, the markets that you talk about and we'll get to will be seen as sort of breakthrough
and we're in the very early days of really doing interesting stuff there and we'll get
to that, but just taking a quick step back, can you give, kind of throw off the historian
hat.
I mean, you briefly said that the history of AI kind of mimics the history of chemical
engineering, but...
I keep saying machine learning.
You keep wanting to say AI, just to let you know, I don't, you know, I resist that.
I don't think this is about AI really was John McCarthy as almost a philosopher saying,
wouldn't it be cool if we could put thought in a computer?
If we could mimic the human capability to think or put intelligence in, in some sense
into a computer.
That's an interesting philosophical question and he wanted to make it more than philosophy.
He wanted to actually write down a logical formula and algorithms that would do that.
And that is a perfectly valid, reasonable thing to do.
That's not what's happening in this era.
So the reason I keep saying AI actually, and I'd love to hear what you think about it.
Machine learning has a very particular set of methods and tools.
Maybe your version of it is that mine doesn't, it's very, very open.
It does optimization, it does sampling, it does...
So systems that learn is what machine learning is.
Systems that learn and make decisions.
And make decisions.
So it's not just pattern recognition and, you know, finding patterns, it's all about
making decisions in real worlds and having close feedback loops.
So something like symbolic AI, expert systems, reasoning systems, knowledge based representation,
all of those kinds of things, search, does that neighbor fit into what you think of as
machine learning?
So I don't even like the word machine learning, I think that what the field you're talking
about is all about making large collections of decisions under uncertainty by large collections
of entities.
Right?
And there are principles for that, at that scale.
You don't have to say the principles are for a single entity that's making decisions, single
agent or single human.
It really immediately goes to the network of decisions.
Is a good word for that or no?
No, there's no good words for any of this.
That's kind of part of the problem.
So we can continue the conversation to use AI for all that.
I just want to kind of raise the flag here that this is not about, we don't know what
intelligence is and real intelligence.
We don't know much about abstraction and reasoning at the level of humans.
We don't have a clue.
We're not trying to build that because we don't have a clue.
Eventually it may emerge.
They'll make, I don't know if there'll be breakthroughs, but eventually we'll start
to get glimmers of that.
It's not what's happening right now.
Okay.
We're taking data.
We're trying to make good decisions based on that.
We're trying to scale.
We're trying to economically viably, we're trying to build markets.
We're trying to keep value at that scale and aspects of this will look intelligent.
Computers were so dumb before, they will seem more intelligent.
We will use that buzzword of intelligence so we can use it in that sense.
So machine learning, you can scope it narrowly as just learning from data and pattern recognition.
But when I talk about these topics, maybe data science is another word you could throw
in the mix, it really is important that the decisions are as part of it.
It's consequential decisions in the real world.
Am I going to have a medical operation?
Am I going to drive down the street?
Things where there's scarcity, things that impact other human beings or other environments
and so on.
How do I do that based on data?
How do I do that adaptively?
How do I use computers to help those kinds of things go forward?
Whatever you want to call that.
So let's call it AI.
Let's agree to call it AI, but let's not say that the goal of that is intelligence.
The goal of that is really good working systems at planetary scale that we've never seen before.
So reclaim the word AI from the Dartmouth conference from many decades ago of the dream
of humans.
I don't want to reclaim it.
I want a new word.
I think it was a bad choice.
I mean, if you read one of my little things, the history was basically that McCarthy needed
a new name because cybernetics already existed and he didn't like, no one really liked Norbert
Wiener.
Norbert Wiener was kind of an island to himself and he felt that he had encompassed all this
and in some sense he did.
You look at the language of cybernetics, it was everything we're talking about.
It was control theory and signal processing and some notions of intelligence and closed
feedback loops and data.
It was all there.
It's just not a word that lived on partly because of the maybe the personalities.
But McCarthy needed a new word to say, I'm different from you.
I'm not part of your show.
I got my own.
Invented this word and again, thinking forward about the movies that would be made about
it, it was a great choice.
But thinking forward about creating a sober academic and real world discipline, it was
a terrible choice because it led to promises that are not true that we understand.
We understand artificial perhaps, but we don't understand intelligence.
It's a small tangent because you're one of the great personalities of machine learning,
whatever the heck you call the field.
Do you think science progresses by personalities or by the fundamental principles and theories
and research that's outside of personalities?
Both.
And I wouldn't say there should be one kind of personality.
I have mine and I have my preferences and I have a kind of network around me that feeds
me and some of them agree with me and some of them disagree, but all kinds of personalities
are needed.
Right now, I think the personality that it's a little too exuberant, a little bit too ready
to promise the moon is a little bit too much in ascendance.
And I do think that there's some good to that.
It certainly attracts lots of young people to our field, but a lot of those people come
in with strong misconceptions and they have to then unlearn those and then find something
to do.
And so I think there's just got to be some multiple voices and I wasn't hearing enough
of the more sober voice.
So as a continuation of a fun tangent and speaking of vibrant personalities, what would
you say is the most interesting disagreement you have with Jan Lacune?
So Jan's an old friend and I just say that I don't think we disagree about very much
really.
He and I both kind of have a let's build it kind of mentality and does it work kind of
mentality and kind of concrete.
We both speak French and we speak French more together and we have a lot in common.
And so if one wanted to highlight a disagreement, it's not really a fundamental one.
I think it's just kind of what we're emphasizing.
Jan has emphasized pattern recognition and has emphasized prediction.
And it's interesting to try to take that as far as you can.
If you could do perfect prediction, what would that give you kind of as a thought experiment?
And I think that's way too limited.
We cannot do perfect prediction.
We will never have the data sets that allow me to figure out what you're about ready to
do, what question you're going to ask next.
I have no clue.
I will never know such things.
Moreover, most of us find ourselves during the day in all kinds of situations we had
no anticipation of that are kind of very, very novel in various ways.
And in that moment, we want to think through what we want.
And also there's going to be market forces acting on us.
I'd like to go down that street, but now it's full because there's a crane in the street.
I got it.
I got to think about that.
I got to think about what I might really want here.
And I got to sort of think about how much it costs me to do this action versus this
action.
I got to think about the risks involved.
A lot of our current pattern recognition and prediction systems don't do any risk evaluations.
They have no error bars, right?
I got to think about other people's decisions around me.
I got to think about a collection of my decisions, even just thinking about like a medical treatment,
you know, I'm not going to take a, the prediction of a neural net about my health, about something
consequential.
I'm not about ready to have a heart attack because some number is over 0.7.
Even if you had all the data in the world that ever been collected about heart attacks
better than any doctor ever had, I'm not going to trust the output of that neural net to
predict my heart attack.
I'm going to want to ask what if questions around that.
I'm going to want to look at some us or other possible data I didn't have, causal things.
I'm going to want to have a dialogue with a doctor about things we didn't think about
when he gathered the data.
You know, I could go on and on.
I hope you can see.
And I don't, I think that if you say predictions, everything that, that, that you're missing
all of this stuff.
And so prediction plus decision making is everything, but both of them are equally important.
And so the field has emphasized prediction, Jan rightly so has seen how powerful that
is.
But at the cost of people not being aware that decision making is where the rubber really
hits the road, where human lives are at stake, where risks are being taken, where you got
to gather more data.
You got to think about the error bars.
You got to think about the consequences of your decisions on others.
You got to think about the economy around your decisions, blah, blah, blah, blah.
I'm not the only one working on those, but we're a smaller tribe.
And right now we're not the one that people talk about the most.
But you know, if you go out in the real world and industry, you know, at Amazon, I'd say
half the people there are working on decision making and the other half are doing, you know,
the pattern recognition.
It's important.
And the words of pattern recognition and prediction, I think the distinction there, not to linger
on words, but the distinction there is more a constrained sort of in the lab data set
versus decision making is talking about consequential decisions in the real world, under the messiness
§03Transcript · 23:21–33:54
and the uncertainty of the real world.
And just the whole of it, the whole mess of it that actually touches human beings and
scale.
And the forces, that's the distinction.
It helps add those, that perspective, that broader perspective.
You're right.
I totally agree.
On the other hand, if you're a real prediction person, of course, you want it to be in the
real world.
You want to predict real world events.
I'm just saying that's not possible with just data sets.
That it has to be in the context of, you know, strategic things that someone's doing, data
they might gather, things they could have gathered, the reasoning process around data.
It's not just taking data and making predictions based on the data.
So one of the things that you're working on, I'm sure there's others working on it, but
I don't hear often it talked about, especially in the clarity that you talk about it, and
I think it's both the most exciting and the most concerning area of AI in terms of decision
making.
So you've talked about AI systems that help make decisions that scale in a distributed
way, millions, billions decisions, sort of markets of decisions.
Can you, as a starting point, sort of give an example of a system that you think about
when you're thinking about these kinds of systems?
Yeah, so first of all, you're absolutely getting into some territory, which I will be beyond
my expertise.
And there are lots of things that are going to be very not obvious to think about.
Just like, again, I like to think about history a little bit, but think about put yourself
back in the sixties.
There was kind of a banking system that wasn't computerized really.
There was database theory emerging and database people had to think about how do I actually
not just move data around, but actual money and have it be, you know, valid and have transactions
that ATMs happen that are actually, you know, all valid and so on and so forth.
So that's the kind of issues you get into when you start to get serious about sorts
of things like this.
I like to think about as kind of almost a thought experiment to help me think something
simpler, which is the music market.
And because there is, to first order, there is no music market in the world right now
and in our country, for sure.
There are something called things called record companies and they make money and they prop
up a few really good musicians and make them superstars and they all make huge amounts
of money.
But there's a long tail of huge numbers of people that make lots and lots of really good
music that is actually listened to by more people than the famous people.
They are not in a market.
They cannot have a career.
They do not make money.
The creators, the creators, the creators, the so called influencers or whatever that
diminishes who they are.
So there are people who make extremely good music, especially in the hip hop or Latin
world these days.
They do it on their laptop.
That's what they do on the weekend and they have another job during the week and they
put it up on SoundCloud or other sites.
Eventually it gets streamed.
It now gets turned into bits.
It's not economically valuable.
The information is lost.
It gets put up there.
People stream it.
You walk around in a big city, you see people with headphones, especially young kids listening
to music all the time.
If you look at the data, very little of the music they are listening to is the famous
people's music and none of it's old music.
It's all the latest stuff.
But the people who made that latest stuff are like some 16 year old somewhere who will
never make a career out of this, who will never make money.
Of course there will be a few counter examples.
The record companies incentivize to pick out a few and highlight them.
Long story short, there's a missing market there.
There is not a consumer producer relationship at the level of the actual creative acts.
The pipelines and Spotify's of the world that take this stuff and stream it along, they
make money off of subscriptions or advertising and those things.
They're making the money.
All right.
And then they will offer bits and pieces of it to a few people again to highlight that
they simulate a market.
Anyway, a real market would be if you're a creator of music that you actually are somebody
who's good enough that people want to listen to you, you should have the data available
to you.
There should be a dashboard showing a map of the United States.
So in last week, here's all the places your songs were listened to.
It should be transparent, vetable, so that if someone down in Providence sees that you're
being listened to 10,000 times in Providence, that they know that's real data.
You know it's real data.
They will have you come give a show down there.
They will broadcast to the people who've been listening to you that you're coming.
If you do this right, you could go down there and make $20,000.
You do that three times a year, you start to have a career.
So in this sense, AI creates jobs.
It's not about taking away human jobs.
It's creating new jobs because it creates a new market.
Once you've created a market, you've now connected up producers and consumers.
The person who's making the music can say to someone who comes to their shows a lot,
hey, I'll play at your daughter's wedding for $10,000.
You'll say 8,000.
They'll say 9,000.
Then again, you can now get an income up to $100,000.
You're not going to be a millionaire.
And now even think about really the value of music is in these personal connections,
even so much so that a young kid wants to wear a tshirt with their favorite musician's
signature on it.
So if they listen to the music on the internet, the internet should be able to provide them
with a button that they push and the merchandise arrives the next day.
We can do that.
And now why should we do that?
Well, because the kid who bought the shirt will be happy, but more the person who made
the music will get the money.
There's no advertising needed.
So you can create markets between producers and consumers, take 5% cut.
Your company will be perfectly sound.
It'll go forward into the future and it will create new markets and that raises human happiness.
Now this seems like, well, this is easy, just create this dashboard, kind of create some
connections and all that.
But if you think about Uber or whatever, you think about the challenges in the real world
of doing things like this, and there are actually new principles going to be needed.
You're trying to create a new kind of two way market at a different scale that's ever
been done before.
There's going to be unwanted aspects of the market.
There'll be bad people.
There'll be the data will get used in the wrong ways, it'll fail in some ways, it won't
deliver about.
You have to think that through.
Just like anyone who ran a big auction or ran a big matching service in economics will
think these things through.
And so that maybe doesn't get at all the huge issues that can arise when you start to create
markets, but it starts to, at least for me, solidify my thoughts and allow me to move
forward in my own thinking.
Yeah.
So I talked to the head of research at Spotify actually, and I think their longterm goal,
they've said, is to have at least one million creators make a comfortable living putting
on Spotify.
So I think you articulate a really nice vision of the world and the digital and the cyberspace
of markets.
What do you think companies like Spotify or YouTube or Netflix can do to create such markets?
Is it an AI problem?
Is it an interface problem for interface design?
Is it some other kind of, is it an economics problem?
Who should they hire to solve these problems?
Well, part of it's not just top down.
So the Silicon Valley has this attitude that they know how to do it.
They will create the system just like Google did with the search box that will be so good
that they'll just, everyone will adopt that.
It's everything you said, but really I think missing that kind of culture.
So it's literally that 16 year old who's able to create the songs.
You don't create that as a Silicon Valley entity.
You don't hire them per se.
You have to create an ecosystem in which they are wanted and that they belong.
And so you have to have some cultural credibility to do things like this.
Netflix, to their credit, wanted some of that credibility and they created shows, content.
They call it content.
It's such a terrible word, but it's culture.
And so with movies, you can kind of go give a large sum of money to somebody graduating
from the USC film school.
It's a whole thing of its own, but it's kind of like rich white people's thing to do.
And American culture has not been so much about rich white people.
It's been about all the immigrants, all the Africans who came and brought that culture
and those rhythms to this world and created this whole new thing.
American culture.
And so companies can't artificially create that.
They can't just say, hey, we're here.
We're going to buy it up.
You've got a partner.
And so anyway, not to denigrate, these companies are all trying and they should, and I'm sure
they're asking these questions and some of them are even making an effort.
But it is partly a respect the culture as a technology person.
You've got to blend your technology with cultural meaning.
How much of a role do you think the algorithm, so machine learning has in connecting the
consumer to the creator, sort of the recommender system aspect of this?
Yeah.
It's a great question.
I think pretty high.
There's no magic in the algorithms, but a good recommender system is way better than
a bad recommender system.
And recommender systems is a billion dollar industry back even 10, 20 years ago.
And it continues to be extremely important going forward.
What's your favorite recommender system, just so we can put something, well, just historically
I was one of the, when I first went to Amazon, I first didn't like Amazon because they put
the book people out of business or the library, the local booksellers went out of business.
I've come to accept that there probably are more books being sold now and poor people
reading them than ever before.
And then local book stores are coming back.
So that's how economics sometimes work.
You go up and you go down.
But anyway, when I finally started going there and I bought a few books, I was really pleased
to see another few books being recommended to me that I never would have thought of.
And I bought a bunch of them.
So they obviously had a good business model.
But I learned things and I still to this day kind of browse using that service.
And I think lots of people get a lot, that is a good aspect of a recommendation system.
I'm learning from my peers in an indirect way.
And their algorithms are not meant to have them impose what we learn.
It really is trying to find out what's in the data.
It doesn't work so well for other kinds of entities, but that's just the complexity of
human life.
Like shirts, I'm not going to get recommendations on shirts, but that's interesting.
If you try to recommend restaurants, it's hard.
It's hard to do it at scale.
But a blend of recommendation systems with other economic ideas, matchings and so on
is really, really still very open research wise.
And there's new companies that are going to emerge that do that well.
What do you think is going to the messy, difficult land of say politics and things like that,
§04Transcript · 33:54–43:41
that YouTube and Twitter have to deal with in terms of recommendation systems?
Being able to suggest, I think Facebook just launched Facebook news.
So recommend the kind of news that are most likely for you to be interesting.
Do you think this is AI solvable, again, whatever term we want to use, do you think it's a solvable
problem for machines or is it a deeply human problem that's unsolvable?
So I don't even think about it at that level.
I think that what's broken with some of these companies, it's all monetization by advertising.
They're not, at least Facebook, I want to critique them, but they didn't really try
to connect a producer and a consumer in an economic way, right?
No one wants to pay for anything.
And so they all, you know, starting with Google and Facebook, they went back to the playbook
of, you know, the television companies back in the day.
No one wanted to pay for this signal.
They will pay for the TV box, but not for the signal, at least back in the day.
And so advertising kind of filled that gap and advertising was new and interesting and
it somehow didn't take over our lives quite, right?
Fast forward, Google provides a service that people don't want to pay for.
And so somewhat surprisingly in the nineties, they made, they ended up making huge amounts
so they cornered the advertising market.
It didn't seem like that was going to happen, at least to me.
These little things on the right hand side of the screen just did not seem all that economically
interesting, but that companies had maybe no other choice.
The TV market was going away and billboards and so on.
So they've, they got it.
And I think that sadly that Google just has, it was doing so well with that at making such
money.
They didn't think much more about how, wait a minute, is there a producer consumer relationship
to be set up here?
Not just between us and the advertisers market to be created.
Is there an actual market between the producer consumer?
They're the producers, the person who created that video clip, the person that made that
website, the person who could make more such things, the person who could adjust it as
a function of demand, the person on the other side who's asking for different kinds of things,
you know?
So you see glimmers of that now there's influencers and there's kind of a little glimmering of
a market, but it should have been done 20 years ago.
It should have been thought about.
It should have been created in parallel with the advertising ecosystem.
And then Facebook inherited that.
And I think they also didn't think very much about that.
So fast forward and now they are making huge amounts of money off of advertising.
And the news thing and all these clicks is just feeding the advertising.
It's all connected up to the advertiser.
So you want more people to click on certain things because that money flows to you, Facebook.
You're very much incentivized to do that.
And when you start to find it's breaking, people are telling you, well, we're getting
into some troubles.
You try to adjust it with your smart AI algorithms, right?
And figure out what are bad clicks.
So maybe it shouldn't be click through rate, it should be something else.
I find that pretty much hopeless.
It does get into all the complexity of human life and you can try to fix it.
You should, but you could also fix the whole business model.
And the business model is that really, what are, are there some human producers and consumers
out there?
Is there some economic value to be liberated by connecting them directly?
Is it such that it's so valuable that people will be able to pay for it?
All right.
And micro payments, like small payments.
Micro, but even have to be micro.
So I like the example, suppose I'm going, next week I'm going to India.
Never been to India before.
Right?
I have a couple of days in Mumbai, I have no idea what to do there.
Right?
And I could go on the web right now and search.
It's going to be kind of hopeless.
I'm not going to find, you know, I have lots of advertisers in my face.
Right?
What I really want to do is broadcast to the world that I am going to Mumbai and have someone
on the other side of a market look at me and, and there's a recommendation system there.
So I'm not looking at all possible people coming to Mumbai.
They're looking at the people who are relevant to them.
So someone in my age group, someone who kind of knows me in some level, I give up a little
privacy by that, but I'm happy because what I'm going to get back is this person can make
a little video for me, or they're going to write a little two page paper on here's the
cool things that you want to do and move by this week, especially, right?
I'm going to look at that.
I'm not going to pay a micro payment.
I'm going to pay, you know, a hundred dollars or whatever for that.
It's real value.
It's like journalism.
Um, and as an honest subscription, it's that I'm going to pay that person in that moment.
Company's going to take 5% of that.
And that person has now got it.
It's a gig economy, if you will, but you know, done for it, you know, thinking about a little
bit behind YouTube, there was actually people who could make more of those things.
If they were connected to a market, they would make more of those things independently.
You don't have to tell them what to do.
You don't have to incentivize them any other way.
Um, and so, yeah, these companies, I don't think have thought long and hard about that.
So I do distinguish on Facebook on the one side, who just not thought about these things
at all.
I think, uh, thinking that AI will fix everything, uh, and Amazon thinks about them all the time
because they were already out in the real world.
They were delivering packages, people's doors.
They were, they were worried about a market.
They were worried about sellers and, you know, they worry and some things they do are great.
Some things maybe not so great, but you know, they're in that business model.
And then I'd say Google sort of hovers somewhere in between.
I don't, I don't think for a long, long time they got it.
I think they probably see that YouTube is more pregnant with possibility than, than,
than they might've thought and that they're probably heading that direction.
Um, but uh, you know, Silicon Valley has been dominated by the Google Facebook kind of mentality
and the subscription and advertising and that is, that's the core problem, right?
The fake news actually rides on top of that because it means that you're monetizing with
clip through rate and that is the core problem.
You got to remove that.
So advertisement, if we're going to linger on that, I mean, that's an interesting thesis.
I don't know if everyone really deeply thinks about that.
So you're right.
The thought is the advertising model is the only thing we have, the only thing we'll ever
have.
We have to fix, we have to build algorithms that despite that business model, you know,
find the better angels of our nature and do good by society and by the individual.
But you think we can slowly, you think, first of all, there's a difference between should
and could.
So you're saying we should slowly move away from the advertising model and have a direct
connection between the consumer and the creator.
The question I also have is, can we, because the advertising model is so successful now
in terms of just making a huge amount of money and therefore being able to build a big company
that provides, has really smart people working that create a good service.
Do you think it's possible?
And just to clarify, you think we should move away?
Well, I think we should.
Yeah.
But we is the, you know, me.
So society.
Yeah.
Well, the companies, I mean, so first of all, full disclosure, I'm doing a day a week at
Amazon because I kind of want to learn more about how they do things.
So, you know, I'm not speaking for Amazon in any way, but, you know, I did go there
because I actually believe they get a little bit of this or trying to create these markets.
And they don't really use, advertising is not a crucial part of it.
Well, that's a good question.
So it has become not crucial, but it's become more and more present if you go to Amazon
website.
And, you know, without revealing too many deep secrets about Amazon, I can tell you
that, you know, a lot of people in the company question this and there's a huge questioning
going on.
You do not want a world where there's zero advertising.
That actually is a bad world.
Okay.
So here's a way to think about it.
You're a company that like Amazon is trying to bring products to customers, right?
And the customer, at any given moment, you want to buy a vacuum cleaner, say, you want
to know what's available for me.
And, you know, it's not going to be that obvious.
You have to do a little bit of work at it.
The recommendation system will sort of help, right?
But now suppose this other person over here has just made the world, you know, they spent
a huge amount of energy.
They had a great idea.
They made a great vacuum cleaner.
They know they really did it.
They nailed it.
It's an MIT, you know, whiz kid that made a great new vacuum cleaner, right?
It's not going to be in the recommendation system.
No one will know about it.
The algorithms will not find it and AI will not fix that.
Okay.
At all.
Right.
How do you allow that vacuum cleaner to start to get in front of people, be sold well advertising.
And here, what advertising is, it's a signal that you're, you believe in your product enough
that you're willing to pay some real money for it.
And to me as a consumer, I look at that signal.
I say, well, first of all, I know these are not just cheap little ads cause we have now
right now there.
I know that, you know, these are super cheap, you know, pennies.
If I see an ad where it's actually, I know the company is only doing a few of these and
they're making, you know, real money is kind of flowing and I see an ad, I may pay more
attention to it.
And I actually might want that because I see, Hey, that guy spent money on his vacuum cleaner.
Maybe there's something good there.
So I will look at it.
And so that's part of the overall information flow in a good market.
So advertising has a role, but the problem is of course that that signal is now completely
gone because it just, you know, dominant by these tiny little things that add up to big
money for the company, you know?
So I think it will just, I think it will change because the societies just don't, you know,
stick with things that annoy a lot of people and advertising currently annoys people more
than it provides information.
And I think that a Google probably is smart enough to figure out that this is a dead,
this is a bad model, even though it's a hard, huge amount of money and they'll have to figure
out how to pull it away from it slowly.
And I'm sure the CEO there will figure it out, but they need to do it.
And they needed it to, so if you reduce advertising, not to zero, but you reduce it at the same
time you bring up producer, consumer, actual real value being delivered.
So real money is being paid and they take a 5% cut that 5% could start to get big enough
to cancel out the lost revenue from the kind of the poor kind of advertising.
And I think that a good company will do that, will realize that.
And Facebook, you know, again, God bless them.
They bring, you know, grandmothers, they bring children's pictures into grandmothers lives.
It's fantastic.
But they need to think of a new business model and that's the core problem there.
Until they start to connect producer consumer, I think they will just continue to make money
and then buy the next social network company and then buy the next one and the innovation
level will not be high and the health issues will not go away.
So I apologize that we kind of returned to words, I don't think the exact terms matter,
§05Transcript · 43:41–54:00
but in sort of defense of advertisement, don't you think the kind of direct connection between
consumer and creator producer is what advertisement strives to do, right?
So that is best advertisement is literally now Facebook is listening to our conversation
and heard that you're going to India and will be able to actually start automatically for
you making these connections and start giving this offer.
So like, I apologize if it's just a matter of terms, but just to draw a distinction,
is it possible to make advertisements just better and better and better algorithmically
to where it actually becomes a connection, almost a direct connection?
That's a good question.
So let's component on that.
First of all, what we just talked about, I was defending advertising.
Okay.
So I was defending it as a way to get signals into a market that don't come any other way,
especially algorithmically.
It's a sign that someone spent money on it, it's a sign they think it's valuable.
And if I think that if other things, someone else thinks it's valuable, and if I trust
other people, I might be willing to listen.
I don't trust that Facebook though, who's an intermediary between this.
I don't think they care about me.
Okay.
I don't think they do.
And I find it creepy that they know I'm going to India next week because of our conversation.
Why do you think that is?
So what, could you just put your PR hat on?
Why do you think you find Facebook creepy and not trust them as do majority of the population?
So they're out of the Silicon Valley companies, I saw like not approval rate, but there's
ranking of how much people trust companies and Facebook is in the gutter.
In the gutter, including people inside of Facebook.
So what do you attribute that to?
Because when I...
Come on, you don't find it creepy that right now we're talking that I might walk out on
the street right now that some unknown person who I don't know kind of comes up to me and
says, I hear you're going to India.
I mean, that's not even Facebook.
That's just, I want transparency in human society.
I want to have, if you know something about me, there's actually some reason you know
something about me.
That's something that if I look at it later and audit it kind of, I approve.
You know something about me because you care in some way.
There's a caring relationship even, or an economic one or something.
Not just that you're someone who could exploit it in ways I don't know about or care about
or I'm troubled by or whatever.
We're in a world right now where that happens way too much and that Facebook knows things
about a lot of people and could exploit it and does exploit it at times.
I think most people do find that creepy.
It's not for them.
It's not that Facebook is not doing it because they care about them in a real sense.
And they shouldn't.
They should not be a big brother caring about us.
That is not the role of a company like that.
Why not?
Wait, not the big brother part, but the caring, the trusting.
I mean, don't those companies, just to link on it because a lot of companies have a lot
of information about us.
I would argue that there's companies like Microsoft that has more information about
us than Facebook does and yet we trust Microsoft more.
Well, Microsoft is pivoting.
Microsoft, you know, under Satya Nadella has decided this is really important.
We don't want to do creepy things.
Really want people to trust us to actually only use information in ways that they really
would approve of, that we don't decide, right?
And I'm just kind of adding that the health of a market is that when I connect to someone
who produces a consumer, it's not just a random producer or consumer, it's people who see
each other.
They don't like each other, but they sense that if they transact, some happiness will
go up on both sides.
If a company helps me to do that in moments that I choose of my choosing, then fine.
So, and also think about the difference between, you know, browsing versus buying, right?
There are moments in my life I just want to buy, you know, a gadget or something.
I need something for that moment.
I need some ammonia for my house or something because I got a problem with a spill.
I want to just go in.
I don't want to be advertised at that moment.
I don't want to be led down various, you know, that's annoying.
I want to just go and have it be extremely easy to do what I want.
Other moments I might say, no, it's like today I'm going to the shopping mall.
I want to walk around and see things and see people and be exposed to stuff.
So I want control over that though.
I don't want the company's algorithms to decide for me, right?
I think that's the thing.
There's a total loss of control if Facebook thinks they should take the control from us
of deciding when we want to have certain kinds of information, when we don't, what information
that is, how much it relates to what they know about us that we didn't really want them
to know about us.
I don't want them to be helping me in that way.
I don't want them to be helping them by they decide they have control over what I want
and when.
I totally agree.
Facebook, by the way, I have this optimistic thing where I think Facebook has the kind
of personal information about us that could create a beautiful thing.
So I'm really optimistic of what Facebook could do.
It's not what it's doing, but what it could do.
So I don't see that.
I think that optimism is misplaced because there's not a bit, you have to have a business
model behind these things.
Create a beautiful thing is really, let's be, let's be clear.
It's about something that people would value.
And I don't think they have that business model and I don't think they will suddenly
discover it by what, you know, a long hot shower.
I disagree.
I disagree in terms of, you can discover a lot of amazing things in a shower.
So I didn't say that.
I said, they won't come, they won't do it, but in the shower, I think a lot of other
people will discover it.
I think that this guy, so I should also, full disclosure, there's a company called United
Masters, which I'm on their board and they've created this music market and I have a hundred
thousand artists now signed on and they've done things like gone to the NBA and the NBA,
the music you find behind NBA clips right now is their music, right?
That's a company that had the right business model in mind from the get go, right?
Executed on that.
And from day one, there was value brought to, so here you have a kid who made some songs
who suddenly their songs are on the NBA website, right?
That's real economic value to people.
And so, you know, so you and I differ on the optimism of being able to sort of change the
direction of the Titanic, right?
So I, yeah, I'm older than you, so I've seen some Titanic's crash, got it.
But and just to elaborate, cause I totally agree with you and I just want to know how
difficult you think this problem is of, so for example, I want to read some news and
I would, there's a lot of times in the day where something makes me either smile or think
in a way where I like consciously think this really gave me value.
Like I sometimes listen to the daily podcasts in the New York times, way better than the
New York times themselves, by the way, for people listening.
That's like real journalism is happening for some reason in the podcast space.
It doesn't make sense to me, but often I listen to it 20 minutes and I would be willing to
pay for that, like $5, $10 for that experience.
And how difficult, that's kind of what you're getting at is that little transaction.
How difficult is it to create a frictionless system like Uber has, for example, for other
things?
What's your intuition there?
So I, first of all, I pay little bits of money to, you know, to send, there's something
called courts that does financial things.
I like medium as a site, I don't pay there, but I would.
You had a great post on medium.
I would have loved to pay you a dollar and not others.
I wouldn't have wanted it per se because there should be also sites where that's not actually
the goal.
The goal is to actually have a broadcast channel that I monetize in some other way if I chose
to.
I mean, I could now people know about it.
I could, I'm not doing it, but that's fine with me.
Also the musicians who are making all this music, I don't think the right model is that
you pay a little subscription fee to them, right?
Because people can copy the bits too easily and it's just not that somewhere the value
is.
The value is that a connection was made between real human beings, then you can follow up
on that.
All right.
And create yet more value.
So no, I think there's a lot of open questions here, hot open questions, but also, yeah,
I do want good recommendation systems that recommend cool stuff to me.
But it's pretty hard, right?
I don't like them to recommend stuff just based on my browsing history.
I don't like the based on stuff they know about me, quote unquote.
What's unknown about me is the most interesting.
So this is the, this is the really interesting question.
We may disagree, maybe not.
I think that I love recommender systems and I want to give them everything about me in
a way that I trust.
Yeah.
But you, but you don't, because, so for example, this morning I clicked on a, you know, I was
pretty sleepy this morning.
I clicked on a story about the queen of England.
Yes.
Right.
I do not give a damn about the queen of England.
I really do not.
But it was clickbait.
It kind of looked funny and I had to say, what the heck are they talking about?
I don't want to have my life, you know, heading that direction.
Now that's in my browsing history.
The system in any reasonable system will think that I care about the queen of England.
That's browsing history.
Right.
But, but you're saying all the trace, all the digital exhaust or whatever, that's been
kind of the models.
If you collect all this stuff, you're going to figure all of us out.
Well, if you're trying to figure out like kind of one person like Trump or something,
maybe you could figure him out.
But if you're trying to figure out, you know, 500 million people, you know, no way, no way.
You think so?
No, I do.
I think so.
I think we are, humans are just amazingly rich and complicated.
Every one of us has our little quirks, every one of us has our little things that could
intrigue us that we don't even know it will intrigue us.
And there's no sign of it in our past, but by God, there it comes and you know, you fall
in love with it.
And I don't want a company trying to figure that out for me and anticipate that I want
them to provide a forum, a market, a place that I kind of go and by hook or by crook,
this happens, you know, I I'm walking down the street and I hear some Chilean music being
played and I never knew I liked Chilean music, but wow.
So there is that side and I want them to provide a limited, but you know, interesting place
to go.
Right.
And so don't try to use your AI to kind of, you know, figure me out and then put me in
a world where you figured me out, you know, no, create huge spaces for human beings where
our creativity and our style will be enriched and come forward and it'll be a lot of more
transparency.
I won't have people randomly, anonymously putting comments up and I'll special based
on stuff they know about me, facts that, you know, we are so broken right now.
§06Transcript · 54:00–1:04:11
If you're, you know, especially if you're a celebrity, but you know, it's about anybody
that anonymous people are hurting lots and lots of people right now.
That's part of this thing that Silicon Valley is thinking that, you know, just collect all
this information and use it in a great way.
So no, I'm not, I'm not a pessimist, I'm very much an optimist by nature, but I think that's
just been the wrong path for the whole technology to take.
Be more limited, create, let humans rise up.
Don't try to replace them.
That's the AI mantra.
Don't try to anticipate them.
Don't try to predict them because you're, you're, you're not going to, you're not going
to be able to do those things.
You're going to make things worse.
Okay.
So right now, just give this a chance.
Right now, the recommender systems are the creepy people in the shadow watching your
every move.
So they're looking at traces of you.
They're not directly interacting with you, sort of the, your close friends and family,
the way they know you is by having conversation, by actually having interactions back and forth.
Do you think there's a place for recommender systems sort of to step, cause you, you just
emphasize the value of human to human connection, but yeah, just give it a chance, AI human
connection.
Is there a role for an AI system to have conversations with you in terms of, to try to figure out
what kind of music you like, not by just watching what you listening to, but actually having
a conversation, natural language or otherwise.
Yeah, no, I'm, I'm, so I'm not against it.
I just wanted to push back against the, maybe you're saying you have options for Facebook.
So there I think it's misplaced, but, but I think that distributing, yeah, no, so good
for you.
Go for it.
That's a hard spot to be in.
Yeah, no, good.
Human interaction, like on our daily, the context around me in my own home is something
that I don't want some big company to know about at all, but I would be more than happy
to have technology help me with it.
Which kind of technology?
Well, you know, just, Alexa, Amazon, well, a good, Alexa's done right.
And I think Alexa is a research platform right now more than anything else.
But Alexa done right, you know, could do things like I, I leave the water running in my garden
and I say, Hey, Alexa, the water's running in my garden.
And even have Alexa figure out that that means when my wife comes home, that she should be
told about that.
That's a little bit of a reasoning.
I would call that AI and by any kind of stretch, it's a little bit of reasoning and it actually
kind of would make my life a little easier and better.
And you know, I don't, I wouldn't call this a wow moment, but I kind of think that overall
rises human happiness up to have that kind of thing.
But not when you're lonely, Alexa, knowing loneliness.
No, no, I don't want Alexa to be, feel intrusive.
And I don't want just the designer of the system to kind of work all this out.
I really want to have a lot of control and I want transparency and control.
And if a company can stand up and give me that in the context of new technology, I think
they're good.
First of all, be way more successful than our current generation.
And like I said, I was mentioning Microsoft, I really think they're, they're pivoting to
kind of be the trusted old uncle, but you know, I think that they get that this is a
way to go, that if you let people find technology, empowers them to have more control and have
and have control, not just over privacy, but over this rich set of interactions, that that
people are going to like that a lot more.
And that's, that's the right business model going forward.
What does control over privacy look like?
Do you think you should be able to just view all the data that?
No, it's much more than that.
I mean, first of all, it should be an individual decision.
Some people don't want privacy.
They want their whole life out there.
Other people's want it.
Privacy is not a zero one.
It's not a legal thing.
It's not just about which data is available, which is not.
I like to recall to people that, you know, a couple hundred years ago, everyone, there
was not really big cities, everyone lived in on the countryside and villages and villages.
Everybody knew everything about you.
Very, you didn't have any privacy.
Is that bad?
Are we better off now?
Well, you know, arguably no, because what did you get for that loss of certain kinds
of privacy?
Well, people help each other if they, because they know everything about you.
They know something's bad's happening, they will help you with that.
Right.
And now you live in a big city, no one knows about that.
You get no help.
So it kind of depends the answer.
I want certain people who I trust and there should be relationships.
I should kind of manage all those, but who knows what about me?
I should have some agency there.
It shouldn't, I shouldn't be a drift in a sea of technology where I have no agency.
I don't want to go reading things and checking boxes.
So I don't know how to do that.
And I'm not a privacy researcher per se.
I just, I recognize the vast complexity of this.
It's not just technology.
It's not just legal scholars meeting technologists.
There's gotta be kind of a whole layers around it.
And so I, when I alluded to this emerging engineering field, this is a big part of it.
When electrical engineering came, I'm not one around at the time, but you just didn't
plug electricity into walls and all kinds of work.
You don't have to have like underwriters laboratory that reassured you that that plug's not going
to burn up your house and that that machine will do this and that and everything.
There'll be whole people who can install things.
There'll be people who can watch the installers.
There'll be a whole layers, you know, an onion of these kinds of things.
And for things as deep and interesting as privacy, which is as least as interesting
as electricity, that's going to take decades to kind of work out, but it's going to require
a lot of new structures that we don't have right now.
So it's kind of hard to talk about it.
And you're saying there's a lot of money to be made if you get it right.
So something you should look at.
A lot of money to be made in all these things that provide human services and people recognize
them as useful parts of their lives.
So yeah.
So yeah, the dialogue sometimes goes from the exuberant technologists to the no technology
is good, kind of.
And that's, you know, in our public discourse, you know, and as far as you see too much of
this kind of thing and the sober discussions in the middle, which are the challenge he
wants to have or where we need to be having our conversations.
And you know, there's just not actually, there's not many forum fora for those.
You know, there's, that's, that's kind of what I would look for.
Maybe I could go and I could read a comment section of something and it would actually
be this kind of dialogue going back and forth.
You don't see much of this, right?
Which is why actually there's a resurgence of podcasts out of all, because people are
really hungry for conversation, but there's technology is not helping much.
So comment sections of anything, including YouTube is not hurting and not helping.
Yeah.
And you think technically speaking, it's possible to help.
I don't know the answers, but it's a, it's a, it's a less anonymity, a little more locality,
you know, worlds that you kind of enter in and you trust the people there in those worlds
so that when you start having a discussion, you know, not only is that people are not
going to hurt you, but it's not going to be a total waste of your time because there's
a lot of wasting of time that, you know, a lot of us, I pulled out of Facebook early
on cause it was clearly going to waste a lot of my time even though there was some value.
And so, yeah, worlds that are somehow you enter in and you know what you're getting
and it's kind of appeals to you and you might, new things might happen, but you kind of have
some, some trust in that world.
And there's some deep, interesting, complex psychological aspects around anonymity, how
that changes human behavior that's quite dark.
Quite dark.
Yeah.
I think a lot of us are, especially those of us who really loved the advent of technology.
I love social networks when they came out.
I was just, I didn't see any negatives there at all.
But then I started seeing comment sections.
I think it was maybe, you know, with the CNN or something.
And I started to go, wow, this, this darkness I just did not know about and, and our technology
is now amplifying it.
So sorry for the big philosophical question, but on that topic, do you think human beings,
cause you've also, out of all things, had a foot in psychology too, the, do you think
human beings are fundamentally good?
Like all of us have good intent that could be mind or is it depending on context and
environment, everybody could be evil.
So my answer is fundamentally good.
But fundamentally limited.
All of us have very, you know, blinkers on.
We don't see the other person's pain that easily.
We don't see the other person's point of view that easily.
We're very much in our own head, in our own world.
And on my good days, I think the technology could open us up to, you know, more perspectives
and more less blinkered and more understanding, you know, a lot of wars in human history happened
because of just ignorance.
They didn't, they, they thought the other person was doing this while their person wasn't
doing this.
And we have a huge amounts of that.
But in my lifetime, I've not seen technology really help in that way yet.
And I do, I do, I do believe in that, but you know, no, I think fundamentally humans
are good.
The people suffer, people have grievances because you have grudges and those things
cause them to do things they probably wouldn't want.
They regret it often.
So no, I, I think it's a, you know, part of the progress of technology is to indeed allow
it to be a little easier to be the real good person you actually are.
Well, but do you think individual human life or society could be modeled as an optimization
problem?
Not the way I think typically, I mean, that's, you're talking about one of the most complex
phenomenon in the whole, you know, in all of which the individual human life or society
as a whole.
Both, both.
I mean, individual human life is amazingly complex.
And so you know, optimization is kind of just one branch of mathematics that talks about
certain kinds of things.
And it just feels way too limited for the complexity of such things.
What properties of optimization problems do you think, so do you think most interesting
problems that could be solved through optimization, what kind of properties does that surface
have non convexity, convexity, linearity, all those kinds of things, saddle points?
Well, so optimization is just one piece of mathematics.
You know, there's like, you just, even in our era, we're aware that say sampling is
coming up, examples of something coming up with a distribution.
What's optimization?
What's sampling?
Well, they, you can, if you're a kind of a certain kind of mathematician, you can try
to blend them and make them seem to be sort of the same thing.
But optimization is roughly speaking, trying to find a point that, a single point that
is the optimum of a criterion function of some kind.
And sampling is trying to, from that same surface, treat that as a distribution or density
and find points that have high density.
So I want the entire distribution in a sampling paradigm and I want the, you know, the single
point, that's the best point in the optimization paradigm.
Now if you were optimizing in the space of probability measures, the output of that could
§07Transcript · 1:04:11–1:15:41
be a whole probability distribution.
So you can start to make these things the same.
But in mathematics, if you go too high up that kind of abstraction hierarchy, you start
to lose the, you know, the ability to do the interesting theorems.
So you kind of don't try that.
You don't try to overly over abstract.
So as a small tangent, what kind of worldview do you find more appealing?
One that is deterministic or stochastic?
Well, that's easy.
I mean, I'm a statistician.
You know, the world is highly stochastic.
I don't know what's going to happen in the next five minutes, right?
Because what you're going to ask, what we're going to do, what I'll say.
Due to the uncertainty.
Due to the...
Massive uncertainty.
Yeah.
You know, massive uncertainty.
And so the best I can do is have come rough sense or probability distribution on things
and somehow use that in my reasoning about what to do now.
So how does the distributed at scale when you have multi agent systems look like?
So optimization can optimize sort of, it makes a lot more sense, sort of at least from my
from robotics perspective, for a single robot, for a single agent, trying to optimize some
objective function.
When you start to enter the real world, this game theoretic concept starts popping up.
That's how do you see optimization in this?
Because you've talked about markets in a scale.
What does that look like?
Do you see it as optimization?
Do you see it as sampling?
Do you see like, how should you mark?
These all blend together.
And a system designer thinking about how to build an incentivized system will have a blend
of all these things.
So, you know, a particle in a potential well is optimizing a functional called a Lagrangian,
right?
The particle doesn't know that.
There's no algorithm running that does that.
It just happens.
And so it's a description mathematically of something that helps us understand as analysts
what's happening, right?
And so the same thing will happen when we talk about, you know, mixtures of humans and
computers and markets and so on and so forth, there'll be certain principles that allow
us to understand what's happening, whether or not the actual algorithms are being used
by any sense is not clear.
Now at some point, I may have set up a multi agent or market kind of system.
And I'm now thinking about an individual agent in that system.
And they're asked to do some task and they're incentivized in some way, they get certain
signals and they have some utility.
What they will do at that point is they just won't know the answer, they may have to optimize
to find an answer.
Okay, so an artist could be embedded inside of an overall market.
You know, and game theory is very, very broad.
It is often studied very narrowly for certain kinds of problems.
But it's roughly speaking, this is just the, I don't know what you're going to do.
So I kind of anticipate that a little bit, and you anticipate what I'm anticipating.
And we kind of go back and forth in our own minds.
We run kind of thought experiments.
You've talked about this interesting point in terms of game theory, you know, most optimization
problems really hate saddle points, maybe you can describe what saddle points are.
But I've heard you kind of mentioned that there's a there's a branch of optimization
that you could try to explicitly look for saddle points as a good thing.
Oh, not optimization.
That's just game theory that that so there's all kinds of different equilibria in game
theory.
And some of them are highly explanatory behavior.
They're not attempting to be algorithmic.
They're just trying to say, if you happen to be at this equilibrium, you would see certain
kind of behavior.
And we see that in real life.
That's what an economist wants to do, especially behavioral economists in continuous differential
game theory, you're in continuous spaces, a some of the simplest equilibria are saddle
points and Nash equilibrium as a saddle point.
It's a special kind of saddle point.
So classically, in game theory, you were trying to find Nash equilibria and an algorithmic
game theory, you're trying to find algorithms that would find them.
And so you're trying to find saddle points.
I mean, so that's literally what you're trying to do.
But you know, any economist knows that Nash equilibria have their limitations.
They are definitely not that explanatory in many situations.
They're not what you really want.
There's other kind of equilibria.
And there's names associated with these because they came from history with certain people
working on them, but there will be new ones emerging.
So you know, one example is a Stackelberg equilibrium.
So you know, Nash, you and I are both playing this game against each other or for each other,
maybe it's cooperative, and we're both going to think it through and then we're going to
decide and we're going to do our thing simultaneously.
You know, in a Stackelberg, no, I'm going to be the first mover.
I'm going to make a move.
You're going to look at my move and then you're going to make yours.
Now since I know you're going to look at my move, I anticipate what you're going to do.
And so I don't do something stupid, but then I know that you are also anticipating me.
So we're kind of going back and forth on why, but there is then a first mover thing.
And so those are different equilibria, right?
And so just mathematically, yeah, these things have certain topologies and certain shapes
that are like, what's it, algorithmically or dynamically, how do you move towards them?
How do you move away from things?
You know, so some of these questions have answers, they've been studied, others do not.
And especially if it becomes stochastic, especially if there's large numbers of decentralized
things, there's just, you know, young people get in this field who kind of think it's all
done because we have, you know, TensorFlow.
Well, no, these are all open problems and they're really important and interesting.
And it's about strategic settings.
How do I collect data?
Suppose I don't know what you're going to do because I don't know you very well, right?
Well, I got to collect data about you.
So maybe I want to push you into a part of the space where I don't know much about you
so I can get data.
Cause, and then later I'll realize that you'll never, you'll never go there because of the
way the game is set up.
You know, that's part of the overall, you know, data analysis context is that.
Even the game of poker is fascinating space, whenever there's any uncertainty, a lack of
information, it's a super exciting space.
Just to linger on optimization for a second.
So when we look at deep learning, it's essentially minimization of a complicated loss function.
So is there something insightful or hopeful that you see in the kinds of function surface
that loss functions, the deep learning and in the real world is trying to optimize over?
Is there something interesting as it's just the usual kind of problems of optimization?
I think from an optimization point of view, that surface, first of all, it's pretty smooth.
And secondly, if there's over, if it's over parameterized, there's kind of lots of paths
down to reasonable Optima.
And so kind of the getting downhill to the, to an optimum is viewed as not as hard as
you might've expected in high dimensions.
The fact that some Optima tend to be really good ones and others not so good.
And you tend to, it's not, sometimes you find the good ones is sort of still needs explanation.
Yeah.
But, but the particular surface is coming from the particular generation of neural nets.
I kind of suspect those will, those will change in 10 years.
It will not be exactly those surfaces.
There'll be some others that are an optimization theory will help contribute to why other surfaces
or why other algorithms.
Years of arithmetic operations with a little bit of nonlinearity, that's not, that didn't
come from neuroscience per se.
I mean, maybe in the minds of some of the people working on it, they were thinking about
brains, but they were arithmetic circuits in all kinds of fields, computer science control
theory and so on.
And that layers of these could transform things in certain ways.
And that if it's smooth, maybe you could find parameter values is a sort of big discovery
that it's working, it's able to work at this scale.
But I don't think that we're stuck with that and we're, we're certainly not stuck with
that cause we're understanding the brain.
So in terms of on the algorithm side sort of gradient descent, do you think we're stuck
with gradient descent as a variance of it?
What variance do you find interesting or do you think there'll be something else invented
that is able to walk all over these optimization spaces in more interesting ways?
So there's a co design of the surface and the, or the architecture and the algorithm.
So if you just ask if we stay with the kind of architectures that we have now and not
just neural nets, but you know, phase retrieval architectures or matrix completion architectures
and so on.
You know, I think we've kind of come to a place where yeah, a stochastic gradient algorithms
are dominant and there are versions that are a little better than others.
They have more guarantees, they're more robust and so on.
And there's ongoing research to kind of figure out which is the best arm for which situation.
But I think that that'll start to co evolve, that that'll put pressure on the actual architecture.
And so we shouldn't do it in this particular way, we should do it in a different way because
this other algorithm is now available if you do it in a different way.
So that I can't really anticipate that co evolution process, but you know, gradients
are amazing mathematical objects.
They have a lot of people who start to study them more deeply mathematically are kind of
shocked about what they are and what they can do.
Think about it this way, suppose that I tell you if you move along the x axis, you go uphill
in some objective by three units, whereas if you move along the y axis, you go uphill
by seven units, right?
Now I'm going to only allow you to move a certain unit distance, right?
What are you going to do?
Well, most people will say that I'm going to go along the y axis, I'm getting the biggest
bang for my buck, you know, and my buck is only one unit, so I'm going to put all of
it in the y axis, right?
And why should I even take any of my strength, my step size and put any of it in the x axis
because I'm getting less bang for my buck.
That seems like a completely clear argument and it's wrong because the gradient direction
is not to go along the y axis, it's to take a little bit of the x axis.
And to understand that, you have to know some math and so even a trivial so called operator
like gradient is not trivial and so, you know, exploiting its properties is still very important.
Now we know that just pervading descent has got all kinds of problems, it gets stuck in
many ways and it had never, you know, good dimension dependence and so on.
So my own line of work recently has been about what kinds of stochasticity, how can we get
dimension dependence, how can we do the theory of that and we've come up pretty favorable
results with certain kinds of stochasticity.
We have sufficient conditions generally.
We know if you do this, we will give you a good guarantee.
We don't have necessary conditions that it must be done a certain way in general.
So stochasticity, how much randomness to inject into the walking along the gradient?
And what kind of randomness?
Why is randomness good in this process?
Why is stochasticity good?
Yeah, so I can give you simple answers but in some sense again, it's kind of amazing.
Stochasticity just, you know, particular features of a surface that could have hurt you if you
were doing one thing deterministically won't hurt you because by chance, there's very little
chance that you would get hurt.
So here stochasticity, it just kind of saves you from some of the particular features of
surfaces.
In fact, if you think about surfaces that are discontinuous in our first derivative,
like an absolute value function, you will go down and hit that point where there's nondifferentiability.
And if you're running a deterministic algorithm at that point, you can really do something
bad.
Whereas stochasticity just means it's pretty unlikely that's going to happen, that you're
going to hit that point.
So it's again, nontrivial to analyze but especially in higher dimensions, also stochasticity,
§08Transcript · 1:15:41–1:26:55
our intuition isn't very good about it but it has properties that kind of are very appealing
in high dimensions for a lot of large number of reasons.
So it's all part of the mathematics to kind of, that's what's fun to work in the field
is that you get to try to understand this mathematics.
But long story short, you know, partly empirically, it was discovered stochastic gradient is very
effective and theory kind of followed, I'd say, that but I don't see that we're getting
clearly out of that.
What's the most beautiful, mysterious, a profound idea to you in optimization?
I don't know the most.
But let me just say that Nesterov's work on Nesterov acceleration to me is pretty surprising
and pretty deep.
Can you elaborate?
Well Nesterov acceleration is just that, suppose that we are going to use gradients
to move around in a space.
For the reasons I've alluded to, they're nice directions to move.
And suppose that I tell you that you're only allowed to use gradients, you're not going
to be allowed to use this local person that can only sense kind of the change in the surface.
But I'm going to give you kind of a computer that's able to store all your previous gradients.
And so you start to learn some something about the surface.
And I'm going to restrict you to maybe move in the direction of like a linear span of
all the gradients.
So you can't kind of just move in some arbitrary direction, right?
So now we have a well defined mathematical complexity model.
There's certain classes of algorithms that can do that and others that can't.
And we can ask for certain kinds of surfaces, how fast can you get down to the optimum?
So there's answers to these.
So for a smooth convex function, there's an answer, which is one over the number of steps
squared.
You will be within a ball of that size after k steps.
Gradient descent in particular has a slower rate, it's one over k.
So you could ask, is gradient descent actually, even though we know it's a good algorithm,
is it the best algorithm?
And the answer is no.
Well, not clear yet, because one over k squared is a lower bound.
That's probably the best you can do.
Gradient is one over k, but is there something better?
And so I think as a surprise to most, Nesterov discovered a new algorithm that has got two
pieces to it.
It's two gradients and puts those together in a certain kind of obscure way.
And the thing doesn't even move downhill all the time.
It sometimes goes back uphill.
And if you're a physicist, that kind of makes some sense.
You're building up some momentum and that is kind of the right intuition, but that intuition
is not enough to understand kind of how to do it and why it works.
But it does.
It achieves one over k squared and it has a mathematical structure and it's still kind
of to this day, a lot of us are writing papers and trying to explore that and understand
it.
So there are lots of cool ideas and optimization, but just kind of using gradients, I think
is number one that goes back, you know, 150 years.
And then Nesterov, I think has made a major contribution with this idea.
So like you said, gradients themselves are in some sense, mysterious.
They're not as trivial as...
Not as trivial.
Coordinate descent is more of a trivial one.
You just pick one of the coordinates.
That's how we think.
That's how our human mind thinks.
That's how our human minds think.
And gradients are not that easy for our human mind to grapple with.
An absurd question, but what is statistics?
So here it's a little bit, it's somewhere between math and science and technology.
It's somewhere in that convex hole.
So it's a set of principles that allow you to make inferences that have got some reason
to be believed and also principles that allow you to make decisions where you can have some
reason to believe you're not going to make errors.
So all of that requires some assumptions about what do you mean by an error?
What do you mean by the probabilities?
But after you start making some of those assumptions, you're led to conclusions that, yes, I can
guarantee that if you do this in this way, your probability of making an error will be
small.
Your probability of continuing to not make errors over time will be small.
And the probability that you found something that's real will be small, will be high.
So decision making is a big part of that.
Decision making is a big part.
Yeah.
So statistics, short history was that, it goes back as a formal discipline, 250 years
or so.
It was called inverse probability because around that era, probability was developed
sort of especially to explain gambling situations.
Of course, interesting.
So you would say, well, given the state of nature is this, there's a certain roulette
board that has a certain mechanism and what kind of outcomes do I expect to see?
And especially if I do things long amounts of time, what outcomes will I see?
And the physicists started to pay attention to this.
And then people said, well, let's turn the problem around.
What if I saw certain outcomes, could I infer what the underlying mechanism was?
That's an inverse problem.
And in fact, for quite a while, statistics was called inverse probability.
That was the name of the field.
And I believe that it was Laplace who was working in Napoleon's government who needed
to do a census of France, learn about the people there.
So he went and gathered data and he analyzed that data to determine policy and said, well,
let's call this field that does this kind of thing statistics because the word state
is in there.
In French, that's etat, but it's the study of data for the state.
So anyway, that caught on and it's been called statistics ever since.
But by the time it got formalized, it was sort of in the 30s.
And around that time, there was game theory and decision theory developed nearby.
People in that era didn't think of themselves as either computer science or statistics or
control or econ.
They were all the above.
And so Von Neumann is developing game theory, but also thinking of that as decision theory.
Wald is an econometrician developing decision theory and then turning that into statistics.
And so it's all about, here's not just data and you analyze it, here's a loss function.
Here's what you care about.
Here's the question you're trying to ask.
Here is a probability model and here's the risk you will face if you make certain decisions.
And to this day, in most advanced statistical curricula, you teach decision theory as the
starting point and then it branches out into the two branches of Bayesian and frequentist.
But that's all about decisions.
In statistics, what is the most beautiful, mysterious, maybe surprising idea that you've
come across?
Yeah, good question.
I mean, there's a bunch of surprising ones.
There's something that's way too technical for this thing, but something called James
Stein estimation, which is kind of surprising and really takes time to wrap your head around.
Can you try to maybe...
I think I don't want to even want to try.
Let me just say a colleague at Steven Stigler at University of Chicago wrote a really beautiful
paper on James Stein estimation, which helps to...
It's views a paradox.
It kind of defeats the mind's attempts to understand it, but you can and Steve has a
nice perspective on that.
So one of the troubles with statistics is that it's like in physics that are in quantum
physics, you have multiple interpretations.
There's a wave and particle duality in physics and you get used to that over time, but it
still kind of haunts you that you don't really quite understand the relationship.
The electron's a wave and electron's a particle.
Well the same thing happens here.
There's Bayesian ways of thinking and frequentist, and they are different.
They sometimes become sort of the same in practice, but they are physically different.
And then in some practice, they are not the same at all.
They give you rather different answers.
And so it is very much like wave and particle duality, and that is something that you have
to kind of get used to in the field.
Can you define Bayesian and frequentist?
Yeah in decision theory you can make, I have a video that people could see.
It's called are you a Bayesian or a frequentist and kind of help try to make it really clear.
It comes from decision theory.
So you know, decision theory, you're talking about loss functions, which are a function
of data X and parameter theta.
They're a function of two arguments.
Okay.
Neither one of those arguments is known.
You don't know the data a priori.
It's random and the parameters unknown.
All right.
So you have a function of two things you don't know, and you're trying to say, I want that
function to be small.
I want small loss, right?
Well what are you going to do?
So you sort of say, well, I'm going to average over these quantities or maximize over them
or something so that, you know, I turn that uncertainty into something certain.
So you could look at the first argument and average over it, or you could look at the
second argument and average over it.
That's Bayesian and frequentist.
So the frequentist says, I'm going to look at the X, the data, and I'm going to take
that as random and I'm going to average over the distribution.
So I take the expectation loss under X. Theta is held fixed, right?
That's called the risk.
And so it's looking at other, all the data sets you could get, right?
And say, how well will a certain procedure do under all those data sets?
That's called a frequentist guarantee, right?
So I think it is very appropriate when like you're building a piece of software and you're
shipping it out there and people are using it on all kinds of data sets.
You want to have a stamp, a guarantee on it that as people run it on many, many data sets
that you never even thought about that 95% of the time it will do the right thing.
Perfectly reasonable.
The Bayesian perspective says, well, no, I'm going to look at the other argument of the
loss function, the theta part, okay?
That's unknown and I'm uncertain about it.
So I could have my own personal probability for what it is, you know, how many tall people
are there out there?
I'm trying to infer the average height of the population while I have an idea roughly
what the height is.
So I'm going to average over the theta.
So now that loss function as only now, again, one argument's gone, now it's a function of
X and that's what a Bayesian does is they say, well, let's just focus on the particular
X we got, the data set we got, we condition on that.
Conditional on the X, I say something about my loss.
That's a Bayesian approach to things.
And the Bayesian will argue that it's not relevant to look at all the other data sets
you could have gotten and average over them, the frequentist approach.
It's really only the data sets you got, right?
And I do agree with that, especially in situations where you're working with a scientist, you
can learn a lot about the domain and you're really only focused on certain kinds of data
and you gathered your data and you make inferences.
I don't agree with it though, that, you know, in the sense that there are needs for frequentist
guarantees, you're writing software, people are using it out there, you want to say something.
So these two things have to got to fight each other a little bit, but they have to blend.
So long story short, there's a set of ideas that are right in the middle that are called
empirical Bayes.
And empirical Bayes sort of starts with the Bayesian framework.
It's kind of arguably philosophically more, you know, reasonable and kosher.
Write down a bunch of the math that kind of flows from that, and then realize there's
a bunch of things you don't know because it's the real world and you don't know everything.
So you're uncertain about certain quantities.
At that point, ask, is there a reasonable way to plug in an estimate for those things?
Okay.
§09Transcript · 1:26:55–1:38:18
And in some cases, there's quite a reasonable thing to do, to plug in, there's a natural
thing you can observe in the world that you can plug in and then do a little bit more
mathematics and assure yourself it's really good.
So based on math or based on human expertise, what's, what, what are good?
Oh, they're both going in.
The Bayesian framework allows you to put a lot of human expertise in, but the math kind
of guides you along that path and then kind of reassures you the end, you could put that
stamp of approval under certain assumptions, this thing will work.
So you asked the question, what's my favorite, you know, or what's the most surprising, nice
idea.
So one that is more accessible is something called false discovery rate, which is, you
know, you're making not just one hypothesis test or making one decision, you're making
a whole bag of them.
And in that bag of decisions, you look at the ones where you made a discovery, you announced
that something interesting had happened.
All right.
That's going to be some subset of your big bag.
In the ones you made a discovery, which subset of those are bad?
Or false, false discoveries.
You'd like the fraction of your false discoveries among your discoveries to be small.
That's a different criterion than accuracy or precision or recall or sensitivity and
specificity.
It's a different quantity.
Those latter ones are almost all of them have more of a frequentist flavor.
They say, given the truth is that the null hypothesis is true.
Here's what accuracy I would get, or given that the alternative is true, here's what
I would get.
So it's kind of going forward from the state of nature to the data.
The Bayesian goes the other direction from the data back to the state of nature.
And that's actually what false discovery rate is.
It says, given you made a discovery, okay, that's conditioned on your data.
What's the probability of the hypothesis?
It's going the other direction.
And so the classical frequency look at that, well, I can't know that there's some priors
needed in that.
And the empirical Bayesian goes ahead and plows forward and starts writing down these formulas
and realizes at some point, some of those things can actually be estimated in a reasonable
way.
And so it's kind of, it's a beautiful set of ideas.
So I, this kind of line of argument has come out.
It's not certainly mine, but it sort of came out from Robbins around 1960.
Brad Efron has written beautifully about this in various papers and books.
And the FDR is, you know, Benjamin in Israel, John Story did this Bayesian interpretation
and so on.
And he used to absorb these things over the years and find it a very healthy way to think
about statistics.
Let me ask you about intelligence to jump slightly back out into philosophy, perhaps.
You said that maybe you can elaborate, but you said that defining just even the question
of what is intelligence is a very difficult question.
Is it a useful question?
Do you think we'll one day understand the fundamentals of human intelligence and what
it means, you know, have good benchmarks for general intelligence that we put before our
machines?
So I don't work on these topics so much that you're really asking the question for a psychologist
really.
And I studied some, but I don't consider myself at least an expert at this point.
You know, a psychologist aims to understand human intelligence, right?
And I think many psychologists I know are fairly humble about this.
They might try to understand how a baby understands, you know, whether something's a solid or liquid
or whether something's hidden or not.
And maybe how a child starts to learn the meaning of certain words, what's a verb, what's
a noun and also, you know, slowly but surely trying to figure out things.
But humans ability to take a really complicated environment, reason about it, abstract about
it, find the right abstractions, communicate about it, interact and so on is just, you
know, really staggeringly rich and complicated.
And so, you know, I think in all humility, we don't think we're kind of aiming for that
in the near future.
A certain psychologist doing experiments with babies in the lab or with people talking has
a much more limited aspiration.
And you know, Kahneman and Tversky would look at our reasoning patterns and they're not
deeply understanding all the how we do our reasoning, but they're sort of saying, hey,
here's some oddities about the reasoning and some things you should think about it.
But also, as I emphasize in some things I've been writing about, you know, AI, the revolution
hasn't happened yet.
Yeah.
Great blog post.
I've been emphasizing that, you know, if you step back and look at intelligent systems
of any kind and whatever you mean by intelligence, it's not just the humans or the animals or,
you know, the plants or whatever, you know, so a market that brings goods into a city,
you know, food to restaurants or something every day is a system.
It's a decentralized set of decisions.
Looking at it from far enough away, it's just like a collection of neurons.
Every neuron is making its own little decisions, presumably in some way.
And if you step back enough, every little part of an economic system is making all of
its decisions.
And just like with the brain, who knows what an individual neuron does and what the overall
goal is, right?
But something happens at some aggregate level, same thing with the economy.
People eat in a city and it's robust.
It works at all scales, small villages to big cities.
It's been working for thousands of years.
It works rain or shine, so it's adaptive.
So all the kind of, you know, those are adjectives one tends to apply to intelligent systems.
Robust, adaptive, you know, you don't need to keep adjusting it, self healing, whatever.
Plus not perfect.
You know, intelligences are never perfect and markets are not perfect.
But I do not believe in this era that you cannot, that you can say, well, our computers
are, our humans are smart, but you know, no markets are not, more markets are.
So they are intelligent.
Now we humans didn't evolve to be markets.
We've been participating in them, right?
But we are not ourselves a market per se.
The neurons could be viewed as the market.
There's economic, you know, neuroscience kind of perspective.
That's interesting to pursue all that.
The point though is, is that if you were to study humans and really be the world's best
psychologist studied for thousands of years and come up with the theory of human intelligence,
you might have never discovered principles of markets, you know, supply demand curves
and you know, matching and auctions and all that.
Those are real principles and they lead to a form of intelligence that's not maybe human
intelligence.
It's arguably another kind of intelligence.
There probably are third kinds of intelligence or fourth that none of us are really thinking
too much about right now.
So if you really, and then all of those are relevant to computer systems in the future.
Certainly the market one is relevant right now.
Whereas the understanding of human intelligence is not so clear that it's relevant right now.
Probably not.
So if you want general intelligence, whatever one means by that, or, you know, understanding
intelligence in a deep sense and all that, it is definitely has to be not just human
intelligence.
It's gotta be this broader thing.
And that's not a mystery.
Markets are intelligent.
So, you know, it's definitely not just a philosophical stance to say we've got to move beyond intelligence.
That sounds ridiculous.
Yeah.
But it's not.
And in that blog post, you define different kinds of like intelligent infrastructure,
AI, which I really like is some of the concepts you've just been describing.
Do you see ourselves, if we see earth, human civilization as a single organism, do you
think the intelligence of that organism, when you think from the perspective of markets
and intelligence infrastructure is increasing, is it increasing linearly?
Is it increasing exponentially?
What do you think the future of that intelligence?
Yeah, I don't know.
I don't tend to think, I don't tend to answer questions like that because you know, that's
science fiction.
I'm hoping to catch you off guard.
Well again, because you said it's so far in the future, it's fun to ask and you'll probably,
you know, like you said, predicting the future is really nearly impossible.
But say as an axiom, one day we create a human level, a superhuman level intelligent, not
the scale of markets, but the scale of an individual.
What do you think it is, what do you think it would take to do that?
Or maybe to ask another question is how would that system be different than the biological
human beings that we see around us today?
Is it possible to say anything interesting to that question or is it just a stupid question?
It's not a stupid question, but it's science fiction.
Science fiction.
And so I'm totally happy to read science fiction and think about it from time in my own life.
I loved, there was this like brain in a vat kind of, you know, little thing that people
were talking about when I was a student, I remember, you know, imagine that, you know,
between your brain and your body, there's a, you know, there's a bunch of wires, right?
And suppose that every one of them was replaced with a literal wire.
And then suppose that wire was turned in actually a little wireless, you know, there's a receiver
and sender.
So the brain has got all the senders and receiver, you know, on all of its exiting, you know,
axons and all the dendrites down to the body have replaced with senders and receivers.
Now you could move the body off somewhere and put the brain in a vat, right?
And then you could do things like start killing off those senders and receivers one by one.
And after you've killed off all of them, where is that person?
You know, they thought they were out in the body walking around the world and they moved
on.
So those are science fiction things.
Those are fun to think about.
It's just intriguing about where is, what is thought, where is it and all that.
And I think every 18 year old should take philosophy classes and think about these things.
And I think that everyone should think about what could happen in society that's kind of
bad and all that.
But I really don't think that's the right thing for most of us that are my age group
to be doing and thinking about.
I really think that we have so many more present, you know, first challenges and dangers and
real things to build and all that such that, you know, spending too much time on science
fiction, at least in public for like this, I think is not what we should be doing.
Maybe over beers in private.
That's right.
Well, I'm not going to broadcast where I have beers because this is going to go on Facebook
and I don't want a lot of people showing up there.
But yeah, I'll, I love Facebook, Twitter, Amazon, YouTube.
I have I'm optimistic and hopeful, but maybe, maybe I don't have grounds for such optimism
and hope.
But let me ask, you've mentored some of the brightest sort of some of the seminal figures
in the field.
Can you give advice to people who are undergraduates today?
What does it take to take, you know, advice on their journey if they're interested in
machine learning and in the ideas of markets from economics and psychology and all the
kinds of things that you've exploring?
What steps should they take on that journey?
Well, yeah, first of all, the door is open and second, it's a journey.
I like your language there.
It is not that you're so brilliant and you have great, brilliant ideas and therefore
that's just, you know, that's how you have success or that's how you enter into the field.
It's that you apprentice yourself, you spend a lot of time, you work on hard things, you
try and pull back and you be as broad as you can, you talk to lots of people.
And it's like entering in any kind of a creative community.
There's years that are needed and human connections are critical to it.
So, you know, I think about, you know, being a musician or being an artist or something,
you don't just, you know, immediately from day one, you know, you're a genius and therefore
you do it.
No, you, you know, practice really, really hard on basics and you be humble about where
§10Transcript · 1:38:18–1:45:50
you are and then, and you realize you'll never be an expert on everything.
So you kind of pick and there's a lot of randomness and a lot of kind of luck, but luck just kind
of picks out which branch of the tree you go down, but you'll go down some branch.
So yeah, it's a community.
So the graduate school is, I still think is one of the wonderful phenomena that we have
in our, in our world.
It's very much about apprenticeship with an advisor.
It's very much about a group of people you belong to.
It's a four or five year process.
So it's plenty of time to start from kind of nothing to come up to something, you know,
more, more expertise, and then to start to have your own creativity start to flower,
even surprising your own self.
And it's a very cooperative endeavor.
I think a lot of people think of science as highly competitive and I think in some other
fields it might be more so.
Here it's way more cooperative than you might imagine.
And people are always teaching each other something and people are always more than
happy to be clear that, so I feel I'm an expert on certain kinds of things, but I'm very much
not expert on lots of other things and a lot of them are relevant and a lot of them are,
I should know, but should in some society, you know, you don't.
So I'm always willing to reveal my ignorance to people around me so they can teach me things.
And I think a lot of us feel that way about our field.
So it's very cooperative.
I might add it's also very international because it's so cooperative.
We see no barriers.
And so that the nationalism that you see, especially in the current era and everything
is just at odds with the way that most of us think about what we're doing here, where
this is a human endeavor and we cooperate and are very much trying to do it together
for the, you know, the benefit of everybody.
So last question, where and how and why did you learn French and which language is more
beautiful English or French?
Great question.
So first of all, I think Italian is actually more beautiful than French and English.
And I also speak that.
So I'm married to an Italian and I have kids and we speak Italian.
Anyway, all kidding aside, every language allows you to express things a bit differently.
And it is one of the great fun things to do in life is to explore those things.
So in fact, when I kids or teens or college students ask me what they study, I say, well,
do what your heart, where your heart is, certainly do a lot of math.
Math is good for everybody, but do some poetry and do some history and do some language too.
You know, throughout your life, you'll want to be a thinking person.
You'll want to have done that.
For me, French I learned when I was, I'd say a late teen, I was living in the middle of
the country in Kansas and not much was going on in Kansas with all due respect to Kansas.
And so my parents happened to have some French books on the shelf and just in my boredom,
I pulled them down and I found this is fun.
And I kind of learned the language by reading.
And when I first heard it spoken, I had no idea what was being spoken, but I realized
I had somehow knew it from some previous life and so I made the connection.
But then I traveled and just I love to go beyond my own barriers and my own comfort
or whatever.
And I found myself on trains in France next to say older people who had lived a whole
life of their own.
And the ability to communicate with them was special and the ability to also see myself
in other people's shoes and have empathy and kind of work on that language as part of that.
So after that kind of experience and also embedding myself in French culture, which
is quite amazing, languages are rich, not just because there's something inherently
beautiful about it, but it's all the creativity that went into it.
So I learned a lot of songs, read poems, read books.
And then I was here actually at MIT where we're doing the podcast today and a young
professor not yet married and not having a lot of friends in the area.
So I just didn't have, I was kind of a bored person.
I said, I heard a lot of Italians around.
There's happened to be a lot of Italians at MIT, an Italian professor for some reason.
And so I was kind of vaguely understanding what they were talking about.
I said, well, I should learn this language too.
So I did.
And then later met my spouse and Italian became a part of my life.
But I go to China a lot these days.
I go to Asia, I go to Europe and every time I go, I kind of am amazed by the richness
of human experience and the people don't have any idea if you haven't traveled, kind of
how amazingly rich and I love the diversity.
It's not just a buzzword to me.
It really means something.
I love to embed myself with other people's experiences.
And so yeah, learning language is a big part of that.
I think I've said in some interview at some point that if I had millions of dollars and
infinite time or whatever, what would you really work on if you really wanted to do
AI?
And for me, that is natural language and really done right.
Deep understanding of language.
That's to me, an amazingly interesting scientific challenge.
One we're very far away on.
One we're very far away, but good natural language.
People are kind of really invested then.
I think a lot of them see that's where the core of AI is that if you understand that
you really help human communication, you understand something about the human mind, the semantics
that come out of the human mind and I agree, I think that will be such a long time.
So I didn't do that in my career just cause I kind of, I was behind in the early days.
I didn't kind of know enough of that stuff.
I was at MIT, I didn't learn much language and it was too late at some point to kind
of spend a whole career doing that, but I admire that field and so in my little way
by learning language, you know, kind of that part of my brain has been trained up.
Jan was right.
You truly are the Miles Davis of machine learning.
I don't think there's a better place than it.
Mike it was a huge honor talking to you today.
Merci beaucoup.
All right.
It's been my pleasure.
Thanks for listening to this conversation with Michael I. Jordan and thank you to our
presenting sponsor, Cash App.
Download it, use code LEXPodcast, you'll get $10 and $10 will go to FIRST, an organization
that inspires and educates young minds to become science and technology innovators of
tomorrow.
If you enjoy this podcast, subscribe on YouTube, give it five stars on Apple Podcast, support
on Patreon, or simply connect with me on Twitter at Lex Friedman.
And now let me leave you with some words of wisdom from Michael I. Jordan from his blog
post titled Artificial Intelligence, the revolution hasn't happened yet, calling for broadening
the scope of the AI field.
We should embrace the fact that what we are witnessing is the creation of a new branch
of engineering.
The term engineering is often invoked in a narrow sense in academia and beyond with overtones
of cold, effectless machinery and negative connotations of loss of control by humans.
But an engineering discipline can be what we want it to be.
In the current era, we have a real opportunity to conceive of something historically new,
a human centric engineering discipline.
I will resist giving this emerging discipline a name, but if the acronym AI continues to
be used, let's be aware of the very real limitations of this placeholder.
Let's broaden our scope, tone down the hype, and recognize the serious challenges ahead.
Thank you for listening and hope to see you next time.