[MUSIC] I'd like to start with a
personal view on AI. Circa, late 1980s, I was a student at UCSD. My adviser was David Rumelhart, who's probably the
one person most responsible for this wave of AI. He reinvented back propagation, not really invented it,
but it's the chain rule, hard to say he invented it, but he was applying it to training of layered
neural networks. He spent about a
year doing that. He was next to my office and would come over
and show it to me. He wasn't an AI person and
really nor am I. I think, both of us were interested
in intelligence and understanding the science, but the Frankenstein
attitude of, let's build something like us. I don t think I have it and I don't think that
he had it either. He bequeathed me. He died early, sadly. He would be talked about
a lot more than some of the other names you're hearing
if he were still around, but he had Pick's disease, an affliction of the
frontal cortex. He bequeathed me the software that he wrote to do
back-propagation. The very first
software written to do it in the modern
era was called net. I worked on it for
several years as I was a young professor at MIT. I applied it to some problems
like kinematic chains, robots, [inaudible] and I
watched it solve the problem. I watched the mid score areas
go down and down and down. I wasn't stupified,
but I was impressed. It was very clear to
me there could be an era of brute force AI, and it would emerge
in inexorably. I had some data from a robot. It was enough to get
this thing to learn. It didn't know
anything about robots, so it was clear the writing was on the wall, it was
going to happen. The last couple of
years have been fast. The data got so large
and then [inaudible] worked even better than
we might have suspected. But the idea that
brute-force AI would be possible and it would
be taking over, and then everyone would
be trying to make money at it would cause even the legal profession to
get outraged and activated. That was clear, and so I
decided not to work on that. I wasn't interested in changing the world at
that level, of that kind. I didn't want to build
artificial humans and I didn't want to
make a lot of money. I wanted to make
human welfare better. I became an engineer
in a different sense, not of engineering
artificial human, but engineering
systems that work for all humans and are safe
and robust but adapt, even interesting and
exciting and all that. I worked a lot of
statistics like protection of these systems
though their error bars, [inaudible] talk
about the uncertainty out of distribution things. You collect some data over
here, but here's the reality, causal inference and
on and on and on. I've worked [inaudible]
around these things. I still think that's going on, it's going to be 20,30 years and involving lots
of you in the audience, all these other issues, just the predictive power of these neural nets
with large data. It's interesting, but it's not what everyone's
talking about. What is everyone talking about? Well, they're talking about the emergence of intelligence. We've figured out
intelligence to some level. We haven't. We have artifacts now that exhibit some intelligence, absolutely. They have some mental models, they do some things
that are beyond what we might have thought,
but they're not intelligent. We haven't discovered the
spark of intelligence yet, and we're not going to,
I think, very soon. We're not going to
discover it by looking at a trillion parameter of
trained data point objects, and look inside, just like
it's hard to figure it out by looking at our brain,
what is intelligence. If you look at what's
really happened in history, there have been
engineering disciplines emerging every 50 years or so. They changed human life more than just about anything
I can think of. Civil engineering,
mechanical engineering, chemical engineering, and so on. I did dig into the first one, chemical engineering,
part because my father was a
chemical engineer. In the 30s, there wasn't
chemical engineering yet, but there was an understanding of what happened when you
put molecules together. There was already quantum
mechanics, there was fluids, there was chemistry, and so on. It was clear that you could
start to build factories, but it weren't obvious
that you could do it in a field what you
did in the laboratory. In fact, when people tried, it didn't really work very well. You didn't make product. It was not economically
viable, it would explode. It was very hard, but people
did it for quite a while. Then over several decades, field of chemical
engineering emerged. It became a solid field of its own with its
own mathematics, it's own ideas, and
allowed us to start to envision just how
to do this in a controlled, socially useful way. It had its issues
and its problems. In the '60s, we
became more aware, but it changed life
for the better. All the things we're wearing,
all the things we do are chemical engineering
and medicines and so on. It's
all based on that. I know less about
electrical engineering, but obviously Maxwell's
equations already existed before there was
electrical engineering. We had a full understanding of the phenomenon at some level. We had to build modular
ways of thinking about it to bring electricity into
homes, to make it be safe, to make it be useful if you learnt anything about circuits, think about communications
on top of waves. That became electrical
engineering. Those all took decades and they started with something that was a deep understanding. Well, I think you
can see if we've got this thing and this engineering
discipline is emerging. It's not building
factories and fields or bringing
electricity in homes, but it's building some of the notional factories like
transport and healthcare, and commerce and all
these systems that are computer networks
with dataflows and humans all involved in the
loop and various places. Those are objects over there somehow that are delivering
some product to us, we're in them, but those are the factories
of the modern era. That is what we're
talking about when we're talking about AI for me. We're not talking about how
we got this superintelligence over here that's going
to solve our problems. We got these massive systems that are putting us
all in the mix of it. It's just like
building a factory that may or may not work. Some of these are exploding. Some of them are hurting people. Some of them are doing
great things, absolutely, and some of them will, but that's really what's happening. Now, I think that
it's emergence is being worse by this
AI perspective. I don t think AI is
being very helpful here. Let's talk about
that for a moment. This was the 1950s perspective. John McCarthy and others said, rightly, we have this new
thing called a computer. It has hardware and software. It looks like a man and brain. Let's think about
what it means to put thought into computer. That's an exciting thought. That's a philosophical
aspiration, exciting, absolutely. The slightly naive thought
from a business point of view, if you will, or a
technological view, let's understand
intelligence and then great things will happen. It looks like a
bit of a cartoon, but if you go today to DeepMind and look at
their front page, it's solving intelligence and then great things will
happen, that's basically it. The naivety is breathtaking. We were going to
solve intelligence, whatever that might mean. We also have built
autonomous systems. Why autonomy? Where
did that come from? Well, not so clear, but if your agent is
tethered to a human, it's hard to brag that you've created an artificial
intelligent agent. That's really [inaudible]. Now, I agree there
are some cases like going into a
burning building or going up to the
Mars where it'd be nice to have them be
completely autonomous, but most intelligence
should be linked. Airplanes should
not be autonomous. They should be linked and
federated so they're safe. Horace should not be autonomous. They should be
linked and federated and communicated even, all part of a system that
was designed and so on. Very few things should be autonomous, they
should be linked. I think this is actually
a very unhelpful thing that was put in there
without much thought. Let me go back to this slide. Here's my main message here, it's an obvious one. But intelligence is
as much about the collective as it is
about the individual. Now there's two points
being made here, two allied but different ones. One of them, which I think
is really interesting, is that intelligence, we don't really know what
it is, but it doesn't just have to do with the
human brain and mind. Markets are intelligent. They bring food into cities
around the world every day, 365 days a year with no rain or shine for decades,
for centuries. That's an intelligent
thing to be doing. That market itself is intelligent and ways that
we aren't individually. Then ant colonies
are intelligent, so this is not a
new thought that the collectives
can be intelligent ways that you don't see
an each individual. I don't think we've actually
thought about that enough, we study the ant colonies and we think about the
markets and all that, but we don't realize
that we could create a brand new collectives, so it can be really amazing. We should think about the
intelligence of that level, instead of trying to replace a single human being
with this computer. The other part is even if we're putting into computers
into our midst, it's not about making
the computer happy or making the person who
built the computer rich, it's about making the
collective happening. Our goals and our aspiration should be at the level
of collectivity. For setting goals for
whatever you want to call this emerging engineering field, if you want to call it AI, fine, I don't like
that but whatever, you should be thinking
in terms of the collective as you're
designing your system. Now thinking about,
did my agents speak better language than a human being or didn't
beat people in chess? That was maybe the 1950s okay, but now it's
breathtakingly naive. Imagery of human beings
is a poor way to think about the implications
for collectives. Autonomy to me is a look, ma, no hands inspiration. Again, these are slightly
overrode strong, but there are many
attendant dangerous, so I'm going to get into that a little bit more
on the next slide, and then I'm also taking this other point
about there should be new forms of collectives
if we think about this. There's a lot of
further reading if you want to do on some of my perspectives that evolved
over several decades, but I eventually
wrote a couple of papers at this level and
weren't just still improving, and I'll say a little
bit more at the first two on the next slide, and then the third
one is a collective, here's some of my colleagues,
we'll social scientists, and we wrote a paper
two years ago, this the title of my talk. One of the main points in there is about this autonomy issue, why autonomy and the social
science side of that, which I thought was
really fantastic, is that autonomy
has a real danger, which is that if
it's going to be autonomous like ChatGPT, it's got to be built by a
small number of people. Because if it's
built by everybody, somehow it's not really
that inteligent. With Wikipedia is not
intelligent, everybody built it. It has a tendency to concentrate the development of
this technology in the hands of small people, it just a natural small
numbers of people. In fact, OpenAI is supposed
to be open and distributed, it's not open now, it's
closed and that's continuing, so don't believe
it, when you talk about we're going to solve
the world's problems, we know everything,
it's dangerous, and these people will tell you that even
better than I will. I do want to dig into a
little bit about the points I was making in these
two articles. Just very briefly, I don't want to dig into this too much, but there's John
McCarthy who 1950s quite reasonably had this
exciting aspiration of thought in a computer. We don't have thought
in a computer, we have gradient ascent that mimics things in amazing ways, but it's not thought yet, probably won't be for awhile. I think what really happened in last 40 years is really what Doug Englebart was talking about intelligence
augmentation, like search engines and recommendation systems
and all of that. They are objects that are not intelligent
enough themselves, they can index
websites and all that, but they made all of
us more intelligent, they help to collect it, and they weren't called AI
and quite as much hype, not even nearly as it is now, there was all theoretics. But I think they had more impact than our current wave
of AI is going to have, frankly, trillion dollar
boost to economy and so on. I think in the meantime,
however, this is emerging. For a computer scientists
think the Internet of Things. But for an economist,
think markets and think new dataflows and new ways
of linking human beings. I'm going to talk
mostly about this, and really my talk is
for I think is that, what's exciting about the
current era, not that. Now I'll move a little away from provocative opinions and more
towards actual research. As a machine learning
person, if you go into machine-learning
conferences, you say, you folks are
thinking about this, about collectives and
not just centralizing everything and creating a
super human intelligence and all that stuff. They say, yeah, well, things
like federated learning. Here's federated learning, you've got a federated
server and it's trying to collect data from a
bunch of edge devices, and it takes in all
that data and it builds a better model than
you could build with each individual, and then that's
wonderful, isn't it? No. What's wrong
with this picture? I'm going to dig into it, but this came from Google, and that's Google
sitting up there. Google is happily collecting
everybody's data, telling everybody that we're going to make great
use of this data, we're going to give you a
model that will be great with speech recognition
or vision, whatever. There's some truth to that, it's not completely ridiculous, but it's missing the
fact that these are real agents who have their own interests
and their own data, and they have their
own desire to have an economically valued
lifestyle based on that partly, and they're not being
included in this vision. The ML part of this
is just about, hey, can we get gradients
cheaply up to here and compress and makes
sure all this thing works as an ML thing, it's not nearly enough to
be thinking about society. In fact, really the nodes in these graphs
are often people, their data is not just
something to be streamed, it might be music or
a novel or something, and I know Pam next week
we'll talk more about that people's creative acts for city, they aren't being ingested,
something missing there. They may want to get
benefits out of this, anyone want to opt in. If they get benefits, not just
opt-in because you will be protected or opt-in because you like it or opt out
because you don't like it, but because there's some
benefit to be had for you, so this is the field
that I work in now. Mechanism where learning
where mechanisms, and that's what I want to talk about the rest of the talk. Just to say as an academic, I try to think about organizations
and things like CDSS, very proud of that entity, it tries to bring some
of these thoughts into a collective entity
for all of campus. But just to say that I think these are three
disciplines that are particularly important to emphasize in this discussion. Certainly computer science,
as we've alluded to, but also definitely statistics, this is about the algorithms, this is about the uncertainty
and the decisions, and the two of them
have a bipartite relationship that's
called machine learning. It's all machine learning.
Machine learning, it hasn't taught much about the social and about the
incentives and all that, that's was a field it does,
that's called economics. Economics has long
had an alliance to statistics called
the econometrics, but it's mostly about
measuring the economy, and not so much about
building algorithms and mechanisms that do
things like over here. There certainly is a part of it, but that's really more
mechanism design, different part of the
company economics. But economics and computer
times have had an alliance, that's called
Algorithmic Game Theory, where you're talking about the mechanisms and the
operatives and all that. Three bipartite
relationships in academia, this thing has almost
no statistics, this thing has almost none
of this, we've missed it, there's a triad that all
has to come together, and this is not a
provocative thing to say. If you go into any industry
now that's working sufficient scale
with real impact in the real world taking
Amazon or whatever, all three of those
disciplines around the table every real-world
problem they tried to solve. It's often operations
to such people who have that triad already
in their brains, and then there's of course
around this applied math and sociology, all the fields are
represented here, this was just pick
out three that I think are
particularly important. Now, that's the academic side. I tend to be driven more
about what's happened in the real-world and what are we trying to do in the real-world. This is Steve Stoute,
he's a friend of mine, he's somebody I admired deeply, he's a legendary hip hop
producer, entrepreneur. He and I talked several
years ago about this idea of information
technology shouldn't be about empowering
collectives and it should be about building
multi-way markets, and that is flowered into accompany called
United Masters, I'm on the board there,
scientific advisor on the board, and it's thinking about
music in a different way. Music nowadays, more people
are listening to it than ever before in history by
a factor of 1,000, more people are making it more than ever in history
by a factor of 1,000. Here's the amazing thing,
if you look at the data, 95 percent of the songs listened to today
around the world, written by people
you've never heard of and written in
the last six months. Wow, something amazing is
happening. That's great. It's not just everyone's
listening to the Beatles or Beyonce, not at all. But the Beatles and
Beyonce are getting still paid huge
amounts of money, and all the people that
are writing and doing actual music are not getting
paid anything roughly. Occasionally, they
get enough streams. The payment on
streams on Spotify, it's like 0.0002 or something
ridiculous like that. There's no market, that
means there's no jobs, and actually most of these
people doing the songs are 16-19 year old
living in the city. Should be a job, this is what
they're really good habits when people are listening
to. Why isn't there? We'll, United Masters thought about that and thought, okay, this is a minimally
start to set up an actual to a market, that's
what data should be doing. We just not taking your song and streaming it with the Spotify, Spotify streams it to people's, Spotify creates a subscription, advertising model makes money, and then maybe
throws a little bit back at the producers, no. Let's think, producer relationship directly to
who's listening to me. At the beginning of the
week, United Masters artists gets to see a dashboard, here's a map of
the United States, they see 10,000 people
listening to my songs and you do say
every vendor there, but I can imagine it's a place, and they tell the people of owners there and look on
popular, they see that, yeah, I want you to come here to
show you can make some money, then you can be even
more connected, you can go play at
people's weddings , it's a two-way market. A lot of people
bought into this, a lot of young musicians
did not sign with record companies and a lot of the actually well-known ones, there is now over two
million artists who signed with United Masters,
it's really working. Then Steve had the
brilliant idea, let's make it a
three-way market. He went to the NBA, National
Basketball Association and said we got this
two-way market, you are in. All the songs you're
streaming on the NBA website, usually is Beyonce and
Kanye or whatever, you're paying them
vast amounts of money. Why don't you have these songs that
people like them more? They're more fresh and all that. NBA signed a contract, and now all the music
on the NBA website, is coming from United
Masters artists. When you listen to one of them, the artist gets the money, not Spotify or
somebody in between. This is cool, this is changing the world, this is changing music and
this is not just a US, this can be done
obviously Brazil, Africa, China, you name it, and
it can create jobs. This can create a million jobs, I think that feels
quite reasonable, we've had lots of discussions
about this in each country. Thinking about AI
in this way, yes, some jobs will be lost, but hey, we can also create new jobs. That was the first
part of my talk. Second part, that's
the motivation. That's why I do what
I do. Second part is, what are you going
to do with that? I'm an academic
and a researcher, I want to do actual mathematics. I don't want to write
algorithms and I want to make students get
excited about this. I wanted to have them just try things out and be empirical. I want them to actually
think about foundations, do actual theorems and so on. Here's some of the things
I've been working on for the last 10,15 years. One of those words look a little familiar, but most of them not. This is not the standard
AI list of things. I'm going to emphasize three of them in this talk quickly. We're just going to say
some of the first one. A lot of the work in machine
learning really in fact, when we say AI,
just to be clear, almost all the actual progress has been in
machine learning. The classical AI story is not
what's led to the progress. Those people call
themselves now, everybody is AI, aren't we? Yeah, machine learning
people who resisted that, but it's hard to resist. The PR stopped. Mostly companies like Google changed
from ML to AI. The ML people are really
great at finding optima. We can go downhill in billion dimensional space with saddle points that
we can avoid them. We can prove theorems and
looking at it actually works. We're really good at that. But
real economic systems with multiple compete agents and all that are not about optimization. That will be central
planning that didn't work. It's about equilibria. It's about not
static equilibrium, dynamic equilibrium
and it's about the algorithms that do that and making those good and real. Very little research on that. There is some stochastic
extra gradient methods so on. A couple of my students
in the audience are real world experts on that. But much less than
you would expect. It's partly because
of this perspective. It's all about optimizing from a single agent
point of view. This is a topic that's Berkeley. It's Berkeley
highlight right now, conformal prediction
and all that. My group and others are
really working on that. This is an attempt to really bring in the
economics folks. I partner with them
going forward more. But anyway, I've been talking
about these three others. These are just three choices I thought were fun
to talk about. Partly because I get to
show pictures of some of my great students and postdocs. This is Steven Bates,
who's a postdoc here. Michael is actually down in South Bay and then Jake
was a student here. This is going to give three vignettes,
the rest of the talk. All that blend machine-learning, which is already
a bit of a blend with economics with something. When we say economics really that mostly means incentives. Thinking about, I'm going to treat you
seriously you agent. You have to be talking about your language,
your utilities, and I want to incentivize notice that make you want
to be involved, not just telling
you to be involved. There's an area I'll talk
about of the economics of contract theory
that has not been playing together with
machine learning and it's a real
opportunity here. That's what we're going to
talk about just very briefly. The theory of incentives. There's books on this,
has several branches auctions are certainly a branch of the theory of incentives. You all know about auctions. You probably know less about another branch called
contract theory. It's a asymmetric situation which is a little bit
rare for economics. Most things are symmetric,
crossing curves, equilibrium, Nash
equilibrium, so on. But here is asymmetric. We have a principal who
wants to get something done, but they don't have the
knowledge or the willpower or the resources to do it themselves so they want to get some agents
to help them out. Now the agents know
more than they do. But now there's a
question of how much am I going to incentivize you? How much am I going to pay
you for the job you're doing? I could say, well,
how much do you know how skilled are you? Jennifer, I'm going
to pay you for being a dean. How
skilled are you? The best, exactly. I'm going to give you
a really high price. Now, she was
incentivized to have heard of person next
to us say that. I really want to know
how good you are. You're not going to
tell me, obviously. Pricing things when we have an asymmetry of information
is just very hard. You know what they worked out. This is in the '60s, they revolutionized
the aircraft industry because we all know the results. There's not one price for
every seat on the airplane. It's obvious why I'm taking an airplane from
here down to Los Angeles. There may be a few people who really need to get there today. They really want to get there. They'll pay $1,000 or their business class and
then someone else is paying their business
people, someone else's pay. They're happy to
pay 1,000, so on. What I might do
is set a price of 900 and coax them
to buy the ticket. They are happy they get the
extra $100 of surplus and then I get them on airplane
and anything above a grade. But now I have my airplane is empty and that's not going to actually be a
good business model. What I could do is say, fill
the rest of it with people. I offer them for $100. But now the first-class
people would get mad, that is the same deal. What they did is they created of course different fare classes. They create what's
called a contract or a menu of contracts. Service price and critically, they gave that same
menu to everybody. It's illegal. It's not price discrimination illegally and people then self-select. Here's the amazing thing for
the students in the room. You can't believe
this, but there are people for whom for
a little glass of red wine and being first in line will pay $1,000 to
go on an airplane. They'll feel good about it,
they feel so good about it. Those of you who are willing
to do that will be amazed. There are people who
are happy to always pay $100 and they don't get the glass of red wine and
amp to sit in the back. Everybody is actually
happy. They self-selected. Now I had to set up
a menu correctly. If I set the menu wrong, then these people are
going to pretend to be these people and so on. Anyway, the contract theory,
people work this out. It sounds like we should be
doing this semester learning. Well, the problem is
there's no data here. It's all smart people
writing down values and probability
distributions and curves across at certain places and they design the
thing that way. That work for the
airline industry. But it's not going to work
for us going forward. It's not the right model. We have been working on
this and here's our one of our killer apps for this,
the clinical trials. As you probably all know, tens of millions of
dollars are spent every year on clinical trials. For all diseases, literally tens of millions. What are these things?
Well, these are statistical tests
at the FDA runs. The FDA is a statistical entity. It's trying to do good false
positive control type 1, type 2 error control to
make sure that most of the drugs on the market
are not false positives. They were actually good drug. That's why you have
to get 36,000 people to get a vaccine. You want to
make sure that it's awful. But that statistical perspective is not enough for this problem. This is really a contract
theory problem because the FDA is not deciding what are the
candidate drugs to test, some randomly picking
some candidates. Those candidate drugs are
coming from the drug companies. You got to think about
why are they going to send you certain
candidates and not others? What cancer they get to send in. They have private information. They're not willing just to tell the FDA how good their drug is. If I go to you and say
how good is your drug? Because I want to
price various things. I want to get a license. I want to say how many
people to test and so on. We've put money and all this
stuff, they're going to lie. Lying is not a bad
thing by the way, lying just means that there's
an information subsidy think you should be
able to exploit. Well, I take it out
right away from me. Here's a statistical protocol. If it's a bad drug, doesn't mean really it's
going to hurt people. The drug company does test
that it doesn't hurt people, but it may not do anything. Most drugs on the market
don't actually do anything. Not most, but many. Let's suppose it's
one of those drugs. Well, the FDA will ensure you that the false discovery rate, the probability of
approving given that it's a bad drug is only 0.05 and they will also show that if it
happens to be a good drug, they will discover that
fact would probably 0.8. These numbers aren't
exactly right, but this is standard numbers for industry for type 1,
type 2 error control. Is this a good protocol? Well, yeah it's optimal
in a statistical sense. It's the Neyman-Pearson test. But is it actually
a good protocol? No. Now let's bring
in economics. Suppose that up small profit
is to be made for this drug. Cost 20 million to
run the clinical trial and if you're approved, let's say you'll
make 200 million. It's not a very big market,
relatively small one. Now we can do a calculation, both the FDA and the CEO
of the drug company, they could do this calculation. If the drug happened to be bad. They don't know
if it is or not. Either side knows. But if
it were counterfactually, the expected profit would
be minus 10 million. The CEO could do
that calculation. They say, Oh boy,
don't send drugs up there and pay the 20
million to pay to play, unless you're really sure
it's pretty good drug. How can we really be sure? Well, you gather some
more data internally, you put your best
engineers on it and so on. Then you don't
tell the FDA that. You still send it up
hoping for a false positive if it is
not a good drug. But that's bad fit
and bad number. You don't want to hope
too much for false value, you're going to lose
a lot of money. That mostly you will send only the drugs that
look really good. If we were looking at regime,
thing would be great. But we're not, we're probably more
working in this regime, $20 million to run the trial. If you're approved to billion, that's more like
ibuprofen or something. If you do the same exact
calculation on both sides, it's not a hidden calculation, the expected profit if it were a bad drug, was 80 million. You want to send as
many candidates up to the FDA as possible that'll
go to test a lot of things. They'll do their type
1 error control, but there'll be some
false positives, your drug on the market, you'll make that amount of
money for a few years and people will say then something else will
come in dysplasia. You didn't hurt anybody
made a lot of money. That's what happens.
How do you fix this? Well, you just have to
realize that this is a contract problem and it's
got a statistical side to it. You blend the two fields. I'm not going to
get into details. We have a paper on
this, but here is our new approach to
statistical count. We call this physical
contract theory. There's a protocol which
you now as an agent, a drug company can
opt into or not. If you don't opt in, you just fine. We walk away. If you opt in, you pay
a reservation price r. Then I give you a
menu of functions. I'll say a little
bit more about that. They turn the random
clinical trial result into a utility for you. That's what they
are. There's a price to pay for each one of them. That's standard
crunch activity to do this and this,
but this is new. Now we have a statistical
trial and it yields a random variable
that's drawn from distribution that depends on the true parameter no
one actually knows. But we draw the random variable, we do the clinical trial. Now we get payoffs. The agent gets a
payoff which just depends on their choice of
function from the menu. The FDA gets utility
that depends on the choice function to
the menu plus the truth. Because the FDA, if they approve a lot of bad
drugs over time, people will realize this and
they'll be mad at the FDA. That is the right way
to design these things, and it's straightforward
at some level. We now prove some
theorems about this. First of all, this is too busy. Have a slide, I don't
get any details, but if you're going to
do any of this work, you have to talk about
incentive alignment. Are people wanting to play? There's basically
a little condition saying under the
null hypothesis, when you're a bad drug, this how much you would make minus your reservation price
is got to be negative. You don't want the FDA
just to be losing money. Anyway, you can set up that
very natural definition of incentive compatibility. Now here's an amazing fact. There is an object in statistics
is called an E-value. It's like a p-value, but
p-values have some problems. They're not terrible objects, but they don't
aggregate very well. It's a tail probability under the null hypothesis
That's a p-value. An E-value is a random variable whose expectation on the null hypothesis is less
than equal to one, so with an expectation
rather tail probability, it behaves better
under aggregation. It looks a bit like
a martingale. It is. Therefore you could do
it over time and stop when you want a lot
of nice properties. Statisticians know about this. It's not that common to know
about it, but it's known. It's thing we try to teach
at the undergraduate level, by the way, data
science classes. We have a theorem now
which says that a contract is incentive aligned
and economics contract theory
concept if and only if all pay off
functions are E-values. We have a characterization
of optimal contracts. These linked the two
fields that their foundations and allows you to start to design
optimum contracts. We've been doing this
in various domains. We went back to
federated learning. We said, what if
these are agents that need to be incentivized, whether some economic value
that passes back and forth? How do we structure
the contract in that situation and
we're now of a paper? I'm ready when she will go
and we wrote a paper on this, basically adapting the theory to this and it really solves
the free-riding problem. Which is that if I have some good data and I
could send it up there, but I have a little privacy loss and it cost me money to do it. But I know Eli sitting
next to me and he has some of the same
data that I have. I'm just going to watch Eli
send up the data and I'm not going to send that if he's
sending it to be writing. This solves that problem or it gives you leverage
on that problem. That was vignette number 1, vignette number 2 and 3
will be able to order. The main thing about
that vignette was just the economics is really
brought together with machine learning at their core and we're solving a real-world
problem by doing that. Otherwise would not be solved. We'd just be throwing stuff
out there hoping it works. This is a little bit more
of an academic exercise, but I really like it. I get to again show up two migrate students is
Lydia and Horia. This is competing bandits
and matching markets. There is the learning side,
there's the economic side. I just want to show you how these two ideas come together. In learning, one of
the key problems is exploration and exploitation. We're not seeing that in
the current generation of ChatGPT. It's
just exploiting. It takes all his data and it
just uses the training data. But in real life,
you don't know what the right answer is and you have to explore a little bit and give up a little utility
to try things out and share that
information with others. If you're talking about a
collective, there's lots of this sharing and
exploring together. Anyway, the bandit algorithms are a perfect model of this. Our agent is choosing one of the choices
and getting a reward. There is some unknown reward
distribution behind that. They maybe try another one. They get a reward. They're
trying to figure out which of the arms has
the highest mean reward. This AB testing industry, this is being done 10,000
times a day in every industry, testing out different options and collecting data and so on. You want algorithm that they don't know the optimal
action apriori. They have to try things out. But if they try
things out too much, they don't hone in on the one that gives them a lot of reward so there's a trade-off,
exploration and exploitation. Their optimal outcomes for this. One of them is known as UCB, not University of
California, Berkeley. It's upper confidence bound. You maintain a statistical
confidence interval on each of these objects. The mean rewards. You update
that interval over time. Now you take the upper bound on the confidence
interval and you pick the arm that has the
highest upper bound. If you take our classes,
you'd learn all about why. This is a reasonable
thing to do. Some ways it's obvious that
if it has a high upper bound, that likely means it
has a high reward. It's not a bad thing to
choose or it could have a very big uncertainty
so you should choose it to knock
down your uncertainty. Anyway, it has a
optimal regret bound. It converges quickly and so on and so forth lots
to say about that. That was the learning side. Again, it's not the
ChatGPT learning, it's a different kind,
but we studied this. There's just as much as we
do the gradient algorithms. On the economic side, there have been Nobel
prizes given for matching markets, Gale-Shapley
and others. You have buyers on one side
and sellers on the other. I think you all know
about these things. You write down your
preferences are priori on both sides, and then there's a
matching algorithm, works out a stable
match, an equilibrium. It's not an optimum,
is at equilibrium. Great. This has been applied in lots and lots
of real-world problems. Kidney matching and
college admissions on. But the problem is, for a lot of the problems
we're interested in, you have to write down all
your preferences are priori. Who wants to do that? I can
maybe do it for colleges, but even there it's hard. I can't do it for
restaurants in Shanghai. The first time I
go to Shanghai, or books I'd like to read or
whatever, it's just crazy. The only way out here is to have an algorithm explores
that exploits together, but in a market context. That's actually an
advantage because of lots of people were explored
and is pulling together. We can share information and we can converge more quickly. I hope you can see, we want to have multiple agents in these matching markets. We have a human and a bear and most audiences of people who would
go for the human, but here, I don't know. Let's suppose they both
pick arm to at some point. Now we have competition. That's the real life. A lot of the modern AI people don't think
about competition. They think there's going to
be surplus in ad infinitum. We don't have to
worry about scarcity ever again, nonsense. There's always
gonna be scarcity. If we both take the
same arm, who wins? Well, let's suppose that
way you don't get a double. We're not going to suddenly
generate more value. R2 has some say in the matter, and suppose they pick the bear. The bear gets the reward
and the human gets nothing. Human says, oh, I like that arm, but I see when I pick that arm, the bear also seems like that arm and the bear seems to win because their arm prefers
the bear. What should I do? I should explore more
than I otherwise would. I should try some other
arms a little more. That says I will have high
regret because of competition. Now as a mathematics person, Lydia and Horia and
me sat down and said, well, can we get the
regret bound and characterize how much you
lose from competition? How can you mitigate all of that and what are
the trade-offs? We did all of that.
There are papers on it. There's this notion,
abandoned markets. I'm going to show
one equation here, which is this is a regret bound. It goes only up as the logarithm
of the number of trials. That's fantastic.
That's the optimal. If there's a small gap
between multiple agents, that number is small and you get a larger regret. But
it's only a constant. It's not a function
of the number of trials, it's only a constant. Competition hardship, but
only in a constant sense. Anyway, so there's
lots more work with this kind of do in this topic. It's somehow classical, but pretty interesting
and allows you to start talking about social
networks blended together with market mechanisms
put together learning. Again, there's almost no literature
on most of these things. Then last topic
we're doing here. I don't know if any of
you in the audience, but here are four of the
current people in my group. Anastasios, Stephen
again, Clara and Tijana. I'm just going to
briefly tell you about a topic called
prediction-powered inference. It's again trying to bring engineering care to
prediction systems. We have all these
prediction systems that are just being
thrown out there. They're not calibrated and they may be highly
accurate in some sense, they may be striking to look at, but they're not calibrated,
what does that mean? Well, here's an example, your proteins and we have
a system called AlphaFold, which won the competition
better than anything else, it does the prediction of protein structure amazingly
well, that is progress. It's amazing. It's great. But now, how are you going
to use that in real life? How should biologists use it? Well, instead of having to spend a lot time in the lab, we now, after all these years have
hundreds of thousands of crystallized or of
amino acid sequence with their structures known. You can now get
hundreds of millions of structures predicted today. That sounds great. It is
great, but here's the problem. Here's a pretty
interesting paper, it was published in 2004, it's studying the
relationship between intrinsic disorder and a protein that's where
the quantum effects are big enough that you
see some still vibrations. It's not a full structure. That's been known for a
long time they exist. Is it important biologically?
Well, who knows? But there was a paper
that thought maybe it's related to phosphorylation, which is a very important
biological notion. Structure on proteins. But they didn't have enough
data. They couldn't do it. That's all the structures
they had in 2004. They couldn't actually, I'll say this statistical
hypothesis tests, yes or no, there's
an association. Suddenly we have all these structures
coming from AlphaFold. Someone wrote a paper
that pumped that into an analysis
and interestingly, didn't even use the real data at all because they
had all of these, just so many of these
now off a whole section, they are still good so
they pumped them in there. What do I mean pumping them
in there? Well, they did a hypothesis test is
there relationship to the intrinsic disorder and phosphorylation based on
the results from AlphaFold, all the protein structures? They got a result. Here's a statistical entity
they're trying to test. This is the population
functional probability of intrinsic disorder given
phosphorylation and not that. You replace this now with predictions and this
is not happening, not just in this field, but
all throughout science. Astrophysicists and so on, replaced instead of data, put in prediction,
hoping that it works. I hope you can see
there's a problem here. Some of the predictions
are actually wrong. How does that feed into
the rest of the issue? We've done a number
of experiments and I'm going to show
you a number of results. They all have the
following forum. We did large
Monte-Carlo simulations to get a notion of ground
truth we can test against. Here's the ground truth
of that IDR ratio. If it was one, there's
no association. If it's bigger than one,
there is an association. There is the ground
truth. It looks like there really is an
association in this data. Here is the confidence interval from the AlphaFold predictions. Just look at that, you don't
know the ground truth. You look at that and say, wow, I have nailed the problem. Look how small my
confidence intervals. I'm totally confident
I'm far away from one. That's what they did in
this paper of course. Now if you're a
careful statistician, you look at this and say, no, don't do that. You
can't trust that stuff. Just take the stuff
where you have ground truth data and do your confidence
interval on that. That's the gray region. It's huge and worrisome,
it covers one. You can't assert that there's actually a significant
difference. We've developed a
new procedure called prediction power
to influence which gets the best of both worlds. Our intervals cover the truth, but they make use of this data. It's a really easy little idea. I may run out of time and
leave you, there's a paper. Now we're preparing
weights on the archive. All this is on the archive. But I'm going to try to give you a quick flavor of what it is, but I like the examples
as much as anything. Just briefly here's the setup. It looks like semi-supervised
learning, but it's not. You have some labeled data
and you have predictions, and you have vast amounts of unlabeled data and you
have these predictions. You'd like to design a
confidence interval that covers the truth with some
asserted probability. The classical approach
would be to throw away their predictions because
you can't trust them. The imputer approach would be to trust all the predictions and we don't think that either is the
correct thing to do. We want the best of both worlds. Here's another example. This was a vote in San
Francisco a few years ago, Matt Haney against somebody. I don't remember what
it's about. This is just mean estimation
who had the most votes. Some of the ballots
are messed up. You run computer vision
algorithms on this to make a prediction about
what the actual vote was. People, you could do that. It's the same problem.
Here's what happens. If you use all the computer
vision labeled stuff, you get this confidence interval.
The truth is over here. Here's the throw away all
the predictions and just use only the labels and
here's our new approach. I can tell you which
one I would prefer, especially because there's
a theorem behind this. I don't know if you
can see this, but finding spiral galaxies
with computer vision, a fantastic problem with it. A part of the sky is that
a spiral galaxy or not. You can label a bunch of them, but now you can make huge
numbers of predictions before doing this. Same thing. We did that and here is the computer vision
confidence interval. Here's the truth, and here
is our new procedure. Gene expression, using
the transformer model, trying to decide whether certain promoter region
leads to expression or not. A good classical
biology problem. Just really terrible. We're sending this to
science, by the way, hoping the science
will look at this. Here's another one,
California's census, trying to estimate
some coefficient of income when predicting whether a person has private
health insurance or not. Look how terrible that
is. My favorite example, this is Clara I think, who does lots of marine
biology and other things. A really important
problem is to say how good or how
healthy the ocean is, is how much plankton you have. But if you just start
gathering samples, it's very hard to see interesting
plankton into treatise. I think on the right
is the treatise. You can run algorithms
that make a prediction, they're pretty good
accuracy-wise, but did they give good
confidence intervals, as you probably expect
by this time to talk, not so good, but
not so terrible. Here's interesting,
the throwaway, the predictions thing
is doing pretty bad. We still are not great,
but we're honest. Let me just, on the one picture give you a little
flavor of the idea. It's interesting, somewhat
newish idea in statistics. I'd call it even new. I think the small sample survey literature
had some of this, but it's new and it's not
that hard to think about. There is a truth out
there and we get this prediction version
of that which has a bias. The bias, if I had
the whole population, I could just compute the
bias. It's some number. Some areas of statistics
say, compute the bias, no don't compute it,
estimate the bias, it's just your statistics,
then bias correct. When you do that, often
you get worse results, strangely enough, because
you've added variance. This is known statistics. But it's still a bias correcting is a thing and you should do it. That's not what we're doing. We're doing something different, which I think is pretty cool. All credit is due to my
four colleagues here. We take this object,
this expectation, that's what we're
trying to estimate and we don't just pick a point
estimate of that object, we take a confidence interval on the bias. You can do that. It's perfectly fine in
modern statistics to get confidence
intervals on the bias. That's what's happening here. There's a confidence interval on the bias and that object, or we can call it a
rectifier is an object which takes the point estimate here and builds a confidence
interval for it. Now you pump this prediction through the confidence
interval on the rectifier to get confidence on the
corrected values. Using a confidence
interval to correct. You get a new set we
call CPP over there, which is a confidence
interval on predictions based on the
confidence level on the bias. If some of you are understanding that,
someone you're not. But there's a short paper
on this which you'll see the results and the
proof is not that hard. Let me just say that it's all those little
green things I showed you there are all doing that. The amount of code here
is about this much. I just think it should
be just standard. This should be part of
any pipeline that's doing science with predictive models. Actually here's some
of the math and here's the final theorem that
says these new objects do in fact give you statistical
coverage provably. Also, by the way, this is
the gradient right here. For all of these applications, the entire procedure
is just using gradients and if you're a
machine learning person, you know how important and
good that is. I'm finished. I'm just going to throw that
slide back up there again. This was a slide
that I do want to emphasize that in our era, I think it's important,
especially for younger people to say
what are we doing? I think it's important to be responsible that we are building really exciting new things
that have an impact worldwide. But you need to think
about that bigger object, the collective and what is it and how could we
build new collectives. I didn't spend that much time on the social science thing
about new collectives. But I do want to inspire you
to think more about that. My social science colleagues on that paper spent
some fair amount of time in that paper
thinking about it. There are new deliberative
councils out there. Taiwan has them. Ireland has done them. Where people use data and
interactive protocols and computer analysis to help
deliberation among humans, or shown new ways to think
about democracy that do this. I was talking to my
spouse the other day about what are intelligent
collective things to do. She said, well, how about
intelligent migration? I thought that was
very thoughtful thing to say, intelligent migration. What would that
mean? I don't know, but it sounds very
interesting to think about. No one
thinks like that. I just think migration
is bad or good, but what would that mean? Intelligent, all
things like that. What are the consequences
of thinking that way? I am a little irritated
about our era. I've been doing this
for 30, 40 years. I'm very irritated that
the AI hype wave has come. Not that it's wrong
fundamentally, it's great. There's a lot of great
stuff behind it, but it is completely obscured the clarity of what we're doing and why and what
we can do and not do. We will have yet more ChatGPTs, there'll yet be more
brute-force AI. But we should be thinking about them in the right frame of mind and thinking about what we're really doing in 20 or 30 years, we've took the right path to exploit those in the right way. Thank you. [APPLAUSE] [MUSIC]
