Hal Varian on "Computer Mediated Transactions" (AEA 2010)

Uploaded by Google on 24.02.2010

Okay, welcome to the 2010 Ely Lecture, featuring Hal Varian. I've known admired Hal Varian
since I was a kid, a System Professor at Berkeley and he was a Graduate student. He sought my
advice about what to study and I said, "We're in McFadden." And he tells me this, I don't
actually recall this encounter, but I credit it. And he tells me, he learned a lot from
McFadden as we all did in those days and we had ever since. But something tells me, he
wouldn't have learned in McFadden at Berkeley in, in the late 1960s even without my advice.
When I was on MIT Faculty in 1973, Hal gave the most amazing and entertaining job market
talk. Later, published in the Journal of Economic Theory on Envy. The key result, which I'm
sure you all know Theorem 2.1 is quote, "If X is a strongly efficient allocation, then
there's some agent who envies no one and there are some agent that no one envies." I thought
that was an amazing result. "Now, thanks to his success at Google, Hal has become the
agent who envies no one," at least, to many economist. So, Hal when on to do path...
>> Louder please, we can't hear you over here. >> Well, I want you to move into the middle,
there's plenty of sits in the middle, okay? I don't think we can do better. The sound
system isn't great, I understand. Okay, well, Hal went on to do path breaking academic research
on generalization to reveal preference on information economics. You know, in the slue
of other applied theory topics. He has exercised his talents in a remarkable range of spheres.
He has to be the only economist who's ever published in Byte Magazine, which is not a
food magazine, okay? He wrote wonderful informative and here I say it, non-tendentious columns
for the New York Times, which contrasted nicely with those of another high-profile economists.
He brought the Berkeley Library School into the 21st century. So, I taught Information
Management in place of Paper Management. He's the author of two widely used textbooks in
Micro Theory, much influenced by that, Dan McFadden. And the wonderful "Information Rules"
and Carl Shapiro, I'm happy to say is here the co-author of that wonderful book, it's
a popular book that's been translated into every known human language and Klingon. Starting
seven years ago, he became a force at Google where he is now a Chief Economist, showing
as we all suspect, to be failed to demonstrate ourselves, that good economic is good business.
I'm delighted to welcome Hal Varian as the 2010, to the Ely Lecture, speaking on Computer
Mediated Transactions. >> VARIAN: Well, thank you very much for that
wonderful introduction. It brings back a lot of good memories for me. And hope I'll be
least as entertaining now as in my job talk, 35 years ago. So, I want to talk about computer
mediated transactions. And there's a little outline here. I'm going to talk a bit about
innovation, waves of innovation. And then, I'll talk about the computer mediated transactions
directly. And then, at the end a little bit about Collaborative Computing. And the general
theme of the talk is--this is what's happening in technology. What are the implications for
Commerce? What are the implications for Economics? So, we start out with a pretty obvious observation.
There's been this huge innovation on the Web and on the Internet for the last 15 years.
We've got webpages and search engines, or wikis and docs, and all sorts of other new
innovations being coming out. So, why is there been so much and why is it been so rapid?
Well, my explanation for this is this is an example of what I like to call "combinatorial
innovation." The idea there is that there's a set of component technologies that becomes
available, different times in history. And then, there's a wave of innovation as inventors
combined and recombines those components to create new innovations. So back in the 1800s,
the new set of component parts was interchangeable mechanical parts. And you may know the story
of how Thomas Jefferson went to France, to restock the White House or maybe to stock
the White House wine cellar. But along the way, he saw that the French ammunitions makers
were so skilled. They could build weapons and had interchangeable parts. Came back to
United States and said, "We should set up a type of program to do these ourselves."
The result was the Springfield Armory, where they worked on building interchangeable parts,
mechanical parts and it was the availability of those components, those office shelves,
police wheels, gears, cans, et cetera, that provided the raw materials for innovators
in the 19th century to build all those great innovations; the sewing machine and the washing
machine, and the bicycle; and then towards the end of the century, the automobile. And
then around 1900, we saw the gasoline engines. That found their way into automobiles, and
tractors, and pumps, and motorcycles, and all sorts of things. Setting off another wave
of innovation, think about the Wright Brothers. They took kite technology, bicycle technology,
and the gasoline engine. And combine those components to create a new innovation. In
1960s, we saw integrated circuits there must be thousands of integrated circuits in this
room alone. And then in the early '90s, mid-'90s and 2000s, we saw this range of combinatorial
innovation surrounding the Internet. Now, this is not a new idea at all. There are many
other economists who've observed this kind of a phenomenon. Often this process takes
years, decades, even a century in the case of interchangeable parts. But this time we
saw this huge innovation in just a few years. And I think what's different this time, is
the components, the component parts, they're all bits. They're all protocols and languages
and software, there's TCPIP and HTTP and HTML, all of these different kinds of non-material
components. And then the bits, and the protocols and the languages can be combined to make
webpages and wikis and auctions and exchanges and search engines and on, and on, and on.
So, there's no time to manufacture, there's no inventory problems. There's no logistics,
the bits can be shipped around the world in seconds. And you can have all of these innovators
working in parallel, of the same set of component parts. And so, you get this very, very rapid
evolution and technological progress. You never run out of HTML, just like you never
run out of email. You can get as much as you want, or maybe even much more than you want.
So, the question I want to explore today is then, what are the implications of this kind
of combinatorial innovation for Commerce? And I want to start with a point which is
so obvious and so mundane, it hardly seems worth mentioning. But the point is there's
a computer in the middle of almost every economic transaction. So, between every buyer and seller,
essentially there's a computer. And it could be a very simple computer. If you go to the
corner store and buy a cup of coffee, they'll ring up the sale on a computer because cash
registers nowadays are just computers with a different user interface. And if you go
on the Web and buy something, there'll be a database involved, when you bought your
ticket to fly here; you had a computer mediated transaction, using very complicated set of
algorithms to determine the prices you paid. Now, that itself is an interesting story,
because the original reason for putting systems like "Sabre" together for transportation or
a systems like any kind of web data warehouse, well, that's just to do the accounting. So,
you just to keep track of the transactions. But once that system is put in place, then
there been many other perhaps surprising and unattended uses that evolve for those who
are computer mediated transactions. So the general story here is what are enabled by
computer mediated transactions; one, I claim better contracts; two, the data extraction
and analysis; three, controlled experimentation; and four, the personalization and customization.
I want to spend the next half hour or so talking about these different ways that you can use
that computer in the middle of the transaction to perform useful tasks, so just a review
a little bit. Contracts are fundamental to Commerce. So, the simplest contract would
be, you know, "I'll do X if you do Y." And so, we'd use this for exchange of goods, services,
labor. If you give me $2, I'll give you a cup of coffee. Now, the major problem comes
in monitoring the contract. Because sometimes you can directly observe the contract, you
can observe what's happening. But oftentimes, the quality of the goods, the services, the
labor, whatever is not directly observed. And there's a very large literature in contract
area, the mechanism design that works off of this idea, of what's observed, and what's
not observed, what's enforceable, what's not enforceable, and so on. And so as an example,
I gave you the contract, "If you give me $2, I'll give you a cup of coffee." Well allegedly,
Abraham Lincoln was once served the beverage, he took a sip, and said, "Madam, if this is
coffee, please bring me tea. If this is tea, please bring me coffee." And I'm sure we've
all had a similar experience at some time or another. So, what happens when the computers
come in, well, they make things more observable, and thus more contractible; and even though,
I'm talking about computer mediated transactions, I really should be thinking in broader terms
of information technology in general, because information technology has been improving
monitoring change in transactions, actually for thousands and thousands of years. So,
just to give a summary of what I've been saying. Since the computer serves as an intermediary
in these transactions or in the middle of all these transactions, they not only could
be an accountant, they can also verify contractual performance. And that allows us to structure
more elaborate contracts and potentially improve economic efficiency. So the gentleman you
see on the slide is a famous accountant. You might not have known, there was such a thing
as a famous accountant. But this is Francesco di Marco Datini, who invented--made several
advances in accounting in the middle-ages. And just think of what would happen if he
had a computer. In fact, he could engage in accounting plus computers gives you a better
monitoring and therefore potentially, potentially better contracts. So, let me give you some
examples, so a simplest example, just to motivate things. Think about rental cars. So, what
happens supposed, since we're an economists, we have to make some assumptions. Supposed
insurance for rental cars would cost less if the people drove more slowly, that's assumption
one. Drivers might be willing to drive more slowly if they paid less for the rental car.
They might agree to that choice but the contract can't be made because the speed of operation
isn't observed. So you end up with this inefficiency where yes, you sign the contract that says,
"You'll operate the cars safely." And then, you peel out of the driveway and never to
be seen again. But now of course, you can enforce this contract. You just put a computer
in the trunk of the car; the computer monitors the vehicle or system. And then, you can commit
to paying a certain rate if you operate the car in certain way. The contract becomes feasible
unless very, very simple example. Everybody's made better off. So you get a parade (ph)
of improvement. Now that wouldn't happen in general if you had heterogeneous taste, it
might be that some people who had a very strong taste for speeding. One really wanted to do
this, but if enough people wanted to sign the contract, you would have kind of selection
problem. That forced you into a solution that was better overall in some sense, but not
necessarily better for every party. But, that's kind of thing, I'm thinking about, you have
a contract previously infeasible, you bring in some information technology that probes
the monitoring, then you end up with a more efficient contract. So, I want to give a few
examples of these. And I'm going to start at the beginning, 3300 BC, might as well get
there early. And then, talk about the cash registers in 1880s, semi trucks in the '80s,
video stores in the '90s, and online advertising in the 2000s. I don't really mean to imply
that online advertising is the combination of 5,000 years of economic activity. But,
it's just something I've learned a little about and thought it'd be nice to share with
you today. All right, well, let's talk about the Mediterranean shipping which is--I think,
really just a wonderful example, it's about the problem of how do you ensure that your
shipment is received at the other end of the voyage when there's no written language and
there's no--people are illiterate and innumerate. There's no numbers and there's no letters.
So, what they did is they had this very, very clever solution. These little clay tokens
known as bullae and there's a couple of pictures of them over there. They were tokens that
were shape like the item that was being shipped. So, you might have a handful of olive oil
or you might have a sheep skin, that's what those two bullae are that you see on the screen.
And then, the harbor master each time a barrel of oil was loaded under the ship, the harbor
master would move the token from one side to the accounting desk to the other. And then,
when they're all loaded on the ship, those were sealed up in an envelope and stamp with
a special official seal. It was baked in an oven, given to the captain. The captain delivered
the cargo, the other end of the voyage. They took off the envelope broke it open. And every
time a barrel of oil was unloaded, they moved the token from one side of the counter to
the other. So, they'll be able to verify that the tokens where in one to one to one correspondents
with the shipment. And the contract was then completed. So, adding this little bit of information
technology, very, very trivial kind of information technology allowed you to in fact carry a
transaction or otherwise would have been totally infeasible. And in fact, if you get a look
at archeology museums, there--every shipwreck in the Mediterranean from this period has
a set of a bullae in it and many scholars believe that this is what lead to writing,
because subsequently they would inscribe marks of the outside of the clay to indicate how
many tokens were supposed to be inside. And then they said, "Well gee, we've got the marks,
we've the clay, do we really need the tokens," and that was the invention of writing in numbers.
So I think it's a very important story in human history. Now, skip forward a few thousand
years, so there's some examples of bullae that we used in that period. So skip forward
a few thousand years and think about poor Mr. James Ritty, who owned a saloon in Cincinnati
and would occasionally catches employees pilfering money out of the cash store. So, he had a
monitoring problem as well and went to work and try to solve it, 1883, he invented a device
known as the "Incorruptible Cashier," I love the name the "incorruptible cashier." He had
two innovations; one innovation was that when you open the cash drawer , the cash register
went, "Ka ching." And that was noticed that the saloon keeper should look over and make
sure the things were being on the up, and up. And the other aspect was that it recorded
each transaction a paper tape and the--end of the day, the owner could go through and
verify that the amount of transactions on the tape were supposed to be in accord with
amount of money in the cash drawer. So, there are some scholars that argued that this invention
of the cash register was actually quite important because it allow businesses to expand beyond
just a family. And you can have strangers who actually operating, holding the cash which
is something it was very difficult to do without that capability for monitoring. And then go
forward in few more years in the 1980s, we saw these vehicular monitoring systems that
were put on trucks; virtually every semi that you see driving down the road has a computer
in it. And most of those computers are actually connected to a satellite. So, you can monitor
the performance of the truck. It was put in place to improve logistics in shipping. But
it had a very beneficial effect on honesty as well, because a common fraud had been for
drivers to fill up their buddy's tank in the company credit card, once you had a record
of the mileage that could no longer be done. So, these are nice examples of how adding
the technology had improved the performance of the industry. Another great example I think
is this video store rental. So many in this audience are old enough to remember the bad
old days and the bad old days is where when walked into the video store, wanted to rent
the most recent video. And what had happened, well the recent video, of course, had been
rented by someone else. The prices were high, $60 a video. The store would buy only a few
and the result is that there were many unsatisfied customers. And then, in the late '80s, some
brilliant persons came up with the revenue sharing model, where the video store was able
to buy the videos for $4 from the distributor. And then, share in the each transaction, a
fraction of the rental charge. So, you had a revenue sharing arrangement replace this
purchase and selling arrangements. So now, the store owner had no reason to economize
on videos. He wanted as many as he could. Store owners are better off, the distributors
are better off, the users are better off. And some economists Julie Mortimer, Dana and
Spier have documented this in nice papers. What allows this to happen was in fact in
advance in the information technology. There were barcodes that you could scan. There was
a cash register that records everything. And then in the middle of the night, they would
call up to the warehouse and send in the record of transactions. Just like James Ritty saloon,
they could go over that record in transactions and verify that everybody was getting the
appropriate share. And verify the correctness, the contractual performance, really, on a
daily basis. So now, we come up to the present, the 2000s and talk about incentives and online
advertising. Well, what happens there is there's a kind of a misalignment of incentives in
any kind of advertising. Because the publisher, the person that has the content has space
to put an ad. That's true whether it's T.V., whether it's magazine, whether it's online,
whatever. They've got some space where they can stick an ad. And of course, they want
to sell that space to the highest bidder. The advertiser on the other hand, doesn't
directly care about the ad impressions. Doesn't necessarily directly care about how many people
see this ad, but they care about how many clicks they get. How many people go to their
store, whether it's an online store or whether it's a walk-in store? And normally (ph) of
course, they care about the sales they make and the profit they make. So fundamentally,
the publisher wants to sell impressions but the advertiser wants to buy clicks. So, they've
got to align these two different incentives. And it's a little bit like somebody who wants
to sell something in euros and somebody else wants to pay in dollars. So what do you need
to make it work, you need an exchange rate, okay? And in this case, exchange rate online
is called a Click-Through Rate. Because you can look at the value per click, how much
is worth to the advertiser to get that click. And then, times that clicks per impression,
how many clicks that a particular ad gets each time it's shown. And that can allow you
to convert the value per click to the value per impression. So all you have to do is just
calculate the Click-Through Rate, just like you estimate an exchange rate in other circumstances.
Now, in this case it's a huge statistical machine learning problem, it was nice that
you've mentioned Dan McFadden, this is a logistic regression with a trillion observations, and
hundreds of millions of explanatory variables. You could imagine, it's not sold by a metrics
in version but it is a quite an impressive problem in machine learning. And coming out
with the right exchange rate is of course a very useful competitive of advantage. And
it aligns incentives to the publisher who wants to sell the impression. The advertiser
who wants to buy the click, were in fact, the conversion and then the user who wants
to see relevant ad, which is also part of the entire system. And then, just as a footnote,
there's also a revenue sharing model there, where every time somebody clicks on the ad,
the revenue is shared between a publisher and the ad network provided--that the ad network
provider and so, you have the same kind of accounting problem that I described earlier.
So, you have a double example of computer mediated transactions in that case. So, just
to drive this home a little bit, let me give you an example. This is actually one of the
examples I put on the Google blog. You think of two advertisers, Joe's Jets and Moe's Models.
They both want to buy a keyword, Jet Airplane. So, everybody, somebody who types in a query
about Jet Airplane, they want their ad to come up, but supposed there's only space for
one of them. So, Joe, who sells the Jets, he's willing to pay a lot because he has a
very high profit margin and those Jets offer a lot. But the problem is not very many people
can afford a jet. I mean, Larry and Sergey could afford a jet, but they've already got
one. So, there's a much of a market there. On the other hand, you look at Moe's Models.
He sells model airplanes. You know, he sells it for $25 or something. And it doesn't make
very much money for sell, but he gets a lot of clicks. So, which ad do you want to put
there, the one with a high-value and few sells? Or the one with more sells and well value?
Well, you want to compute the revenue, right? What matter is the revenue? So, you look at
the expected revenue that you would get from the high-price low-volume versus the low-price
high-volume. And so, to figure out which ad should be shown or who's going to win that
particular auction? You have to be able to estimate how many clicks per impression their
ads are going to get. And so, that's what the service, the ad network provides and if
you look at all the major search engines are selling search engine advertising, critical
part of their business just to be able to estimate those Click-Through Rates. So, they
can show the right ads at the right time. Now, furthermore this kind of computer mediated
transactions, once you have the computer in the middle of the transaction, you can make
advertising accountable. So, everybody knows the old story of Henry Wanamaker who said,
"Half of my ad budget is wasted, I just don't know which half." Well, now you can find out
which half because you can link the ad, they ad click, they ad impression back to a conversion
or actual transaction, at least, on a statistical basis. And that allows you to optimize and
to net purchase a process. So of course, this is very easy to do in search advertising where
you have very, very high relevance, since the ad that people see are key to their queries.
You could do it, in contextual advertising where the ads that people see are queued,
keyed to the content on the page, display advertising, mobile advertising. It's a little
harder with the T.V., print and radio advertising because you don't have that direct link communication
with the user. You don't see the click. You don't necessarily see the actual conversion.
But with these other forms of advertising, since they're so highly-mediated by the computers,
we are able to do the accounting and attribute the conversion or purchase to the actual ad
impression or click. So, we really reach to kind of solve one of maker's dilemma. Now,
the other thing you do is once you've done this, once you're able, in fact, to observe
and link together these different aspects of the purchased process, you can study the
data for patterns. You can do various kinds of econometric analysis. So for example, suppose
you had advertisers who are trying to sell something, and they might be interested buying
the keyword "diamond", singular, or they might buy the keyword "diamonds", plural. And the
question is, "Which of those would result in more purchases in fact?" "Which is the
more commercial word?" Well it turns out the answer is diamonds. Diamonds is about 20%
more costly than diamond. Why? Because if your kid is writing a report on forms of carbon,
they'll search for diamond, and if you're thinking about getting engaged, you'll search
for diamonds, right? So, plurals, in general, signify this more commercial level of activity.
And so they sell it a premium in these AdWord auctions. You can look at how clicks vary
over time of day. Remember the Labor Department did a study a few years ago about whether
people were working more at home. And it turns out--I believe the conclusion was they were
working more at home but they were also doing leisure activities more at work. And if you
look at Google, Google search activity you could see that very, very clearly that there's
a lot of--a lot of online shopping takes place around 11 o'clock in the morning just like
most telephone calls and other things. And you could use this to build predictive and
causal models. So you can try to formulate a model that indicates what, you know, what
the probability of a purchase is given an AdClick, et cetera, et cetera. So this would
be familiar to anybody hears an econometrics. But the really beautiful thing about it is
you don't have to stop there. In fact, what you can do is you can actually do experiments.
So there's one of our researchers at Google Ads who's doing some of these experiments.
And you can do very nice controlled experiments. So in fact last year on the search side at
Google, they ran 6,000 experiments--6,000 experiments. And on the ad side, they ran
another 4 or 5,000 experiments. So there are probably 10 or 11,000 experiments that were
being run at Google and roughly 500 improvements were made to the search algorithm because
of those experiments. And what you do, basically, is you just take one percent of the traffic
or a half percent or two percent and you divert it to the new algorithm and then you compare
the performance of the new algorithm to the old algorithm using various performance metrics.
So the beautiful thing about it is you can kind of explore the space using statistical
analysis or one sort of another but once you come up with a reasonable hypothesis, you
can actually implement the hypothesis, see what happens, and then improve the system.
So you move from a 1% experiment to 2% experiment, or 5% or 10%, and then you run it all the
way up and deploy the application. So in fact, the Japanese have this term "Kaizen", which
is very popular in the '80s. It referred to continue improvement of the production process.
So you would take the production process and continue refine, tune it, and improve it.
But now you have product Kaizen because if you have a software product, particularly
a website, it can evolve in real-time. And so the Google--well for that matter the Bing
that you used two weeks ago is different than the Bing you used yesterday, and it'll be
different from the one you're using tomorrow, because in fact those products are continuously
evolving and continuously improving because you have this continue-controlled experiments
that are going on. And ultimately, when you come down to make these decisions you have
this choice between relying on data or relying on HiPPOs. And what's a HiPPO? HIPPO is a
Highly-Paid Person's Opinion. And even if the highly-paid people are highly-paid for
a reason, they may--suppose they have good opinions, you don't really want them sitting
around and debating whether the ad background should be yellow or green, or whether the
font should be this or that, or what the spacing should be in a webpage, or what the details
or the algorithm should be. What you want to do is you want the highly-paid people to
say, "Go run an experiment and report back what the results, we'll deploy the feature
that ends up performing best." Now that's really the model that I think we're going
to see more and more in the future, is you're going to try to get rid of these highly-paid
people's opinions and substitute data to try to make more informed choices. Another thing
you could do with computers is you will--you have this mass customization where the transactions
could be optimized at an individual level. So we see there are some purchases, you go
to Amazon that tells you, "People have bought this, also bought that," or "How you might
be interested in this book?" You see this in searches where you have personalized searches
that are keyed, the kinds of searches you've done recently, social interaction on Facebook,
all sorts of cases where you can highly customize the interaction with the service provider
because there's a computer in the middle there that's able to capture the history of your
interactions and then assign thing or it make--provide you with services that are key to your individual
interest. So in fact, it's not commonly known but when you look at a webpage, you could
go look at the New York Times or the Wall Street Journal or any online content provider,
that webpage is assembled dynamically just for you. Those ads are stuck in there at the
time the webpage is assembled, there could be three different ad networks, three or four
different ad networks, there could be several different pieces of content and it's all pulled
together at the time the ad is being served. So, just try it some time or look at what's
going on, that page is assembled for you and it can be, in fact, customize or personalize
to your particular interest. Now, of course, this raises the issue of privacy. You've got
the benefits of personalization, you've got the potential cause of privacy. But as an
economist, I look at this problem and I say, "Well, those aren't really diametrically oppose
interest." So if you think about a price, somebody wants a high price, the other side
of the transaction wants a low price, that's a diametrically opposed interest. But in this
case, what happens is the provider is trying to give better service to the users and the
unintended uses are security breaches of one form or another. And so, providing a better
security information, transparency and user control can go a long way in aligning interest.
And I think when you go to these services you have to ask, "Well, what are the benefits?
What are the cost?" And in many ways I think there are attracted tradeoffs that can be
made to provide better service while still respecting privacy in a serious way. Up until
now, I've talked mostly about advertising, but advertising is just the beginning because
these kinds of computer mediated transactions allow for other kinds of optimization. Not
just the buying and selling transactions I've been talking about so far, but all sorts of
logistic optimization and customer feedback, product design, recommender systems, I mean,
there's long list of things where you're going to prove the processes across the board. And
what I want to talk about next for the remainder of the time is worker-to-worker transactions
namely, the organization of work. And to do that, I also want to use a historic example
as an example from Paul David. In fact, he talked about this exactly 10 years ago or
almost 10 years ago at AEA meeting on productivity of electricity. And he described the evolution
of the factory. So back in the 1880s, factories were powered by waterwheels. And the way it
worked is there was a shaft down at the center of the factory. All the machinery are connected
to that central shaft and the machinery tended to be clustered by types. So you've had all
the saws in one place and the drills in another place and the lathes somewhere else, because
after all if you ran a drill, you'd like to talk about drills with the other guys. So,
and if you ran a lathe, you talk about lathes. It's kind of like labor economist and trade
economist, you know, they like to talk about their work. And you can--then carry the work
around from station-to-station to get it manufactured. Well, along comes steam engines. They connected
that onto the shaft. Along comes electric motors, they connected that on the shaft.
Then the electric motor got to be smaller and they put the electric motors on each piece
of equipment. But they still kept them organized the same way. Because that's how you do it,
right? That's the way you design a factory. You put all the lathes in one place, all the
saws in the other place. We've all shown it this way. And as Paul David shows, it was
really Henry Ford in the assembly line that offered the breakthrough instead of putting
all the equipment where they've normally been, you try to arrange the equipment in a way
that most facilitate of getting the work done. So you could arrange the assembly line that
allowed you to put the machine where it was most useful, not where it always had been.
So that's a little story about mechanical production in the 1910, but I think you could
do the same thing for knowledge workflow. So back in 1909, you were worried about optimizing
the flow of physical product across the factory floor and how could you use technology, do
that most effectively. But now in the 2000s, and in 2009, well, you wanted to optimize
the flow of ideas through the organization. And how do you do that? You can do the same
sorts of things. You can do the separation, distribution, and optimization of tasks so
you can have multi-authored documents, collaborative computing of one sort or another, do your
version tracking control. And you can do that within the organization, not just in one physical
position but you can actually do that at all locations once you've got the date available
in the database. So, some of you have probably seen a McKinsey presentation, right? They
have these gorgeous PowerPoint slides and I said, "How do you get such nice slides?"
So it took me two beers to find out. But finally they confessed, what happened is they do their
presentation, put it together. And then in the evening, they had send it off by email
to Bangalore and they'd come back the next morning, they'd have a beautiful presentation.
So what a great service. You know, what you've done is you can outsource that specialized
task and if you think about it, everybody knows about Adam Smith's pin factory for assembling
mechanical work, or you can do the same thing with knowledge-work products. You can take
those and do this kind of specialization-optimization, not only within the organization or within
your particular occasion but potentially globally--globally. And it's all enabled by Cloud Computing. So
you have this evolution of computing that we've all seen and loved to the mainframe,
to the network workstation, to the personal computer, and network computing. So what happens
is with the Cloud Computing, you have direct access. The data lives in the Cloud, you could
store once and you could read it anywhere, write anywhere using the Web, access it from
any device, at any time, by any authorized user. So that means you're going to increase
the productivity of knowledge work. Now, in this decade, in this century the same way
that they increase the flow of physical labor 100 years ago. So let me extend Paul David's
example of how factory production worked to something that's relevant to us all here today
namely, how do we do--produce documents. So think about the old days where you dictated
a document to a stenographer, was typewritten. Maybe there were carbon copies that were circulated
for comment. There were notes scribbled in margins. There was cutting. There was pasting.
And somehow that final document got assembled. Now, there were some productivity enhancements
in the 20th Century site. We saw white out, we saw a Post-it notes, we saw other ways
to make this process work a little more smoothly. When the personal computer came along, it
became, yet easier. You could produce the copy of the document, circulate it via email
or file transfer for comments and then assemble all those different pieces into the final
piece of work that you wanted to produce. But nowadays, we have a much better model.
Instead of circulating the document to a bunch of people, there's one master copy of the
document that lives in the Cloud that is some data center on the Internet. And that document
can be accessed anytime, from anywhere, on any device, by any authorized user. And so,
that means that all of these comments and notes and collaboration, version control,
checkpoints, document restore, all those capabilities have now been made much, much easier and you
can produce the final piece of knowledge work at a much more efficient and much more controlled
way than you could by using the old technology of circulating documents via email. So what
it means, I think, is that the barriers to entry for online businesses have really been
dramatically reduced. Because what was previously a fixed cause having to assemble the infrastructure
and the knowledge. So the important part is the knowledge to keep that infrastructure
all running. First, you have to assemble that and then you could build your business on
top of it. That's the way things worked 10 years ago. But now, you can purchase computation
in the data center, storage on demand from Amazon, a development environment and database
access all from Cloud Computing providers. And it's all priced in a variable cost basis.
So instead of having to get the fixed cost at the end of the business, you can purchase
the required infrastructure and the required knowledge and the required operation to keep
your business scaled up as your revenue gross. And I think what that does to circle back
to the beginning of the talk is it really pushes this [INDISTINCT] innovation to new
level because not only can you combine the parts to create the prototype or do the innovation,
but you can now actually do the deployment. It's like being able to buy a factory in modular
pieces and assemble your product in a way that was just not feasible previously. So
it might be that this fosters a whole burst of creative activity and one of the things
that leads to, which I kind of like, it's a term I thought I coined but then I Googled
it and found out that other people had used it. Micromultinationals. And the idea there
is that you think about it, a small or medium-sized enterprise has access to communication infrastructure
that only the biggest companies could afford 10 or 15 years ago and for virtually nothing.
You have Voice Over IP, you have email, you have Wikis, you have Docs, you have all these
services lying in a coordinated activity across the business--across distance so you don't
have to rely on coordinating activity only in one particular physical location. So some
friends of mine at the school that Bob mentioned, we had several people went off and became
entrepreneurs and many of them are, in fact, micromultinationals. I ran acrossed a friend.
I said, "What are you doing these days?" She says, "I have a company." I said, "Where is
your company?" She says, "I have four people in San Francisco, two people in the Czech
Republic and three people in New Delhi, and they work around the clock because when they
end up closing down in San Francisco, the work has just got sent over to India. It's
picked up there and you have this 24-hour work cycle." And her business was in fact
born international--it was international at the beginning? Created by group of students
who'd study together. And when they dispersed to their home country, they were able to,
in fact, continue to communicate in a much more easy way than they had been available
a decade ago and they were able, in fact, to engage in productive activity. So in my
view, this is really just the beginning. I think there's a huge burst of productivity
that's going to come from this availability of information technology that allows you
to coordinate productive activity not just internally but externally and globally, and
we're only at the beginning. So, thank you very much for your time.
>> [INDISTINCT]. >> VARIAN: So we have a few minutes for questions.
I'm acutely aware that I'm standing between you and your cocktail, so we'll try to limit
this--limit this, yes. >> [INDISTINCT].
>> VARIAN: Right. So the question was, "What about concentration because Cloud computing
can be done by very large firms and will you end up with a concentration in that industry
even though you might get a lot of innovative, small firm activity to support it by?" Well,
I think it's kind of an example of book publishing. Because it used to be--20 years ago, publishers
printed their own books and then they realized this was a--you know, it's a very capital-intensive
operation. So it made much more sense just to have a few people specialized in the book
production and then you had all of the publishers that outsource their work to the same place.
And you can do pretty much the same thing with this kind of Cloud Computing Infrastructure
I'm talking about. The one thing that I--that I think is interesting is people are very
concerned about proprietary standards and lock-in. So there's a lot of move to make
the standards in these Cloud Computing facilities open to facilitate switching. And indeed if
you look at it, it's kind of natural because mostly they run on Linux and mostly they run
on MySQL and mostly they run on Pythons. So they're all using open source technologies
in a way that makes it easier to do the switching. So I think it's a potential problem but I
don't think it's actually going to turn into a real problem. Let's see, over here.
>> [INDISTINCT] >> VARIAN: So the question was, "Since information
can be personalized, they--won't this destroy--destroy competition?" Well I think--I think the issue
is the information itself isn't personalized. It's just there's some idea of what you're
looking for. If you are looking for Jaguar and your previous search is a bit about cats,
they're going to show you one set of results, but if your previous search is a bit about
cars, they're going to show you another set of results. So indeed, you want to take account
of not only the current activity you're already doing but at least the recent past because--and
it can provide better service for people. So I don't think its going or really have
a negative consequences. Eric. >> [INDISTINCT]
>> This sounds like a set up but it's not. I think there's a huge demand for--oh yeah,
I'll repeat the question, okay. So he said, "What's the application for labor demand of
this picture that I've painted?" And I'll repeat something I said a few months ago that
the trick is to be scarce but complimentary to something that's ubiquitous and cheap,
right? That's what you want to do. So what's ubiquitous and cheap? It's data. There's data
everywhere. And what scarce? Well, it's the analytic ability to understand that data to
make it--tell its tale. You know, they say that if you torture the data long enough,
it'll confess anything. Well, what you need is these waterboarding techniques that you
can use to analyze this data in a way that really makes it tell its tale. And so I think
this is going to be a very, very valuable talent. That's not just the statistics and
the econometrics but it's also the data visualization, it's the database access, it's the experimental
design, everything else that comes along with this. Because I've talked about the high-tech
businesses, the Googles and the Microsofts of the world. But in fact the technology is
moving further and further down. There's more and more data that's available at a lower
and lower level of operation. And everybody is collecting data about their business. It's
just some people have been able to bring the talent to exploit that and other people having
that realized that's possible. So I see that is the really important career for the--for
the next decade. Bob. >> [INDISTINCT]
>> VARIAN: Okay. Well, this is really an excellent question. I won't pretend to have an answer.
It's a great research topic. I'll repeat the question. The question is, if you look back
at Solow's Paradox where we see computers everywhere. But in the productivity statistics,
why did it take so long for the computers to really show up in the productivity statistics?
And why is it the case that it seems to have shown up so quickly--it leads that I claim
it shown up quickly in the examples I've cited. Well, I think it comes down to partly to this
cost issue that Information Technology was very expensive and it's been getting a lot
cheaper. And so you can expand to a much, much broader base. I still think there is
some delay in really being able to exploit that the riches that we do have available
from this data because the big scarcity is not the data, it's the knowledge of what to
do with it that scares. And so I think it's going to take some time for these things to
happen as well but at least the ground work is there. The other thing is although I do
believe that there will be a big boost in a knowledge-worker productivity, I do have
to say that it's going to be very, very hard to measure, because--what do you do? You count
pages, you count citations, you count documents, you count PowerPoint slides--I mean, how is
it that you really measure the benefits of having this better ways to produce knowledge?
It's going to be a challenge. Yes. >> [INDISTINCT]
>> VARIAN: Mm-hmm... >> [INDISTINCT]
>> VARIAN: Yep. Yeah, so the question was about the Semantic Web has act out and as
far as expected. And the answer is, no, it hasn't because it's just a huge gigantic coordination
problem. It's one of the biggest chicken and egg problems around. And it would be nice
of it to--it happened but I think nowadays, few people are really holding out a lot of
prospect for that to happen soon because the coordination difficulties are just so great.
I think this maybe is a good place to wrap up. So thank you again.
>> All right, thank you very much.