00:00:04
When I think about the rise of AI,
00:00:07
I'm reminded by the rise of literacy.
00:00:10
A few hundred years ago,
00:00:11
many people in society thought
00:00:13
that maybe not everyone needed
to be able to read and write.
00:00:17
Back then, many people were
tending fields or herding sheep,
00:00:20
so maybe there was less need
for written communication.
00:00:23
And all that was needed
00:00:24
was for the high priests
and priestesses and monks
00:00:26
to be able to read the Holy Book,
00:00:28
and the rest of us could just go
to the temple or church
00:00:31
or the holy building
00:00:32
and sit and listen to the high priest
and priestesses read to us.
00:00:35
Fortunately, it was since figured out
that we can build a much richer society
00:00:39
if lots of people can read and write.
00:00:42
Today, AI is in the hands
of the high priests and priestesses.
00:00:46
These are the highly skilled AI engineers,
00:00:48
many of whom work
in the big tech companies.
00:00:51
And most people have access
only to the AI that they build for them.
00:00:55
I think that we can build
a much richer society
00:00:58
if we can enable everyone
to help to write the future.
00:01:03
But why is AI largely concentrated
in the big tech companies?
00:01:08
Because many of these AI projects
have been expensive to build.
00:01:11
They may require dozens
of highly skilled engineers,
00:01:14
and they may cost millions
or tens of millions of dollars
00:01:17
to build an AI system.
00:01:19
And the large tech companies,
00:01:20
particularly the ones
with hundreds of millions
00:01:22
or even billions of users,
00:01:24
have been better than anyone else
at making these investments pay off
00:01:28
because, for them,
a one-size-fits-all AI system,
00:01:33
such as one that improves web search
00:01:35
or that recommends better products
for online shopping,
00:01:38
can be applied to [these] very
large numbers of users
00:01:41
to generate a massive amount of revenue.
00:01:44
But this recipe for AI does not work
00:01:47
once you go outside the tech
and internet sectors to other places
00:01:52
where, for the most part,
00:01:53
there are hardly any projects
that apply to 100 million people
00:01:57
or that generate comparable economics.
00:02:00
Let me illustrate an example.
00:02:03
Many weekends, I drive a few minutes
from my house to a local pizza store
00:02:09
to buy a slice of Hawaiian pizza
00:02:11
from the gentleman
that owns this pizza store.
00:02:14
And his pizza is great,
00:02:15
but he always has a lot
of cold pizzas sitting around,
00:02:19
and every weekend some different flavor
of pizza is out of stock.
00:02:23
But when I watch him operate his store,
00:02:25
I get excited,
00:02:27
because by selling pizza,
00:02:29
he is generating data.
00:02:31
And this is data
that he can take advantage of
00:02:34
if he had access to AI.
00:02:37
AI systems are good at spotting patterns
when given access to the right data,
00:02:43
and perhaps an AI system could spot
if Mediterranean pizzas sell really well
00:02:47
on a Friday night,
00:02:48
maybe it could suggest to him
to make more of it on a Friday afternoon.
00:02:53
Now you might say to me,
"Hey, Andrew, this is a small pizza store.
00:02:56
What's the big deal?"
00:02:58
And I say, to the gentleman
that owns this pizza store,
00:03:01
something that could help him
improve his revenues
00:03:03
by a few thousand dollars a year,
that will be a huge deal to him.
00:03:08
I know that there is a lot of hype about
AI's need for massive data sets,
00:03:14
and having more data does help.
00:03:17
But contrary to the hype,
00:03:19
AI can often work just fine
00:03:21
even on modest amounts of data,
00:03:23
such as the data generated
by a single pizza store.
00:03:26
So the real problem is not
00:03:28
that there isn’t enough data
from the pizza store.
00:03:30
The real problem is
that the small pizza store
00:03:33
could never serve enough customers
00:03:34
to justify the cost of hiring an AI team.
00:03:39
I know that in the United States
00:03:41
there are about half a million
independent restaurants.
00:03:44
And collectively, these restaurants
do serve tens of millions of customers.
00:03:48
But every restaurant is different
with a different menu,
00:03:51
different customers,
different ways of recording sales
00:03:53
that no one-size-fits-all AI
would work for all of them.
00:03:58
What would it be like
if we could enable small businesses
00:04:01
and especially local businesses to use AI?
00:04:05
Let's take a look
at what it might look like
00:04:07
at a company that makes
and sells T-shirts.
00:04:10
I would love if an accountant working
for the T-shirt company
00:04:14
can use AI for demand forecasting.
00:04:16
Say, figure out what funny memes
to prints on T-shirts
00:04:19
that would drive sales,
00:04:20
by looking at what's trending
on social media.
00:04:23
Or for product placement,
00:04:25
why can’t a front-of-store manager
take pictures of what the store looks like
00:04:29
and show it to an AI
00:04:30
and have an AI recommend
where to place products to improve sales?
00:04:34
Supply chain.
00:04:35
Can an AI recommend to a buyer
whether or not they should pay 20 dollars
00:04:39
per yard for a piece of fabric now,
00:04:41
or if they should keep looking
00:04:43
because they might be able to find
it cheaper elsewhere?
00:04:46
Or quality control.
00:04:47
A quality inspector
should be able to use AI
00:04:50
to automatically scan pictures
of the fabric they use to make T-shirts
00:04:55
to check if there are any tears
or discolorations in the cloth.
00:04:59
Today, large tech companies routinely
use AI to solve problems like these
00:05:04
and to great effect.
00:05:06
But a typical T-shirt company
or a typical auto mechanic
00:05:11
or retailer or school or local farm
00:05:15
will be using AI for exactly zero
of these applications today.
00:05:19
Every T-shirt maker is sufficiently
different from every other T-shirt maker
00:05:24
that there is no one-size-fits-all AI
that will work for all of them.
00:05:28
And in fact, once you go outside
the internet and tech sectors
00:05:33
in other industries, even large companies
00:05:35
such as the pharmaceutical companies,
00:05:37
the car makers, the hospitals,
00:05:39
also struggle with this.
00:05:42
This is the long-tail problem of AI.
00:05:46
If you were to take all current
and potential AI projects
00:05:50
and sort them in decreasing
order of value and plot them,
00:05:55
you get a graph that looks like this.
00:05:57
Maybe the single most valuable AI system
00:05:59
is something that decides what ads
to show people on the internet.
00:06:02
Maybe the second most valuable
is a web search engine,
00:06:05
maybe the third most valuable is an online
shopping product recommendation system.
00:06:09
But when you go
to the right of this curve,
00:06:12
you then get projects
like T-shirt product placement
00:06:15
or T-shirt demand forecasting
or pizzeria demand forecasting.
00:06:20
And each of these is a unique project
that needs to be custom-built.
00:06:24
Even T-shirt demand forecasting,
00:06:26
if it depends on trending memes
on social media,
00:06:29
is a very different project
than pizzeria demand forecasting,
00:06:34
if that depends
on the pizzeria sales data.
00:06:37
So today there are millions of projects
00:06:39
sitting on the tail of this distribution
that no one is working on,
00:06:43
but whose aggregate value is massive.
00:06:46
So how can we enable
small businesses and individuals
00:06:49
to build AI systems that matter to them?
00:06:52
For most of the last few decades,
00:06:54
if you wanted to build an AI system,
this is what you have to do.
00:06:58
You have to write pages
and pages of code.
00:07:00
And while I would love
for everyone to learn to code,
00:07:03
and in fact, online education
and also offline education
00:07:06
are helping more people
than ever learn to code,
00:07:09
unfortunately, not everyone
has the time to do this.
00:07:13
But there is an emerging new way
00:07:16
to build AI systems
that will let more people participate.
00:07:20
Just as pen and paper,
00:07:22
which are a vastly superior technology
to stone tablet and chisel,
00:07:26
were instrumental to widespread literacy,
00:07:29
there are emerging new
AI development platforms
00:07:32
that shift the focus from asking you
to write lots of code
00:07:35
to asking you to focus on providing data.
00:07:39
And this turns out to be much easier
for a lot of people to do.
00:07:43
Today, there are multiple companies
working on platforms like these.
00:07:47
Let me illustrate a few of the concepts
using one that my team has been building.
00:07:51
Take the example of an inspector
00:07:54
wanting AI to help
detect defects in fabric.
00:07:58
An inspector can take
pictures of the fabric
00:08:00
and upload it to a platform like this,
00:08:03
and they can go in to show the AI
what tears in the fabric look like
00:08:07
by drawing rectangles.
00:08:09
And they can also go in to show the AI
00:08:11
what discoloration
on the fabric looks like
00:08:14
by drawing rectangles.
00:08:16
So these pictures,
00:08:17
together with the green
and pink rectangles
00:08:19
that the inspector's drawn,
00:08:21
are data created by the inspector
00:08:23
to explain to AI how to find
tears and discoloration.
00:08:28
After the AI examines this data,
00:08:30
we may find that it has seen
enough pictures of tears,
00:08:32
but not yet enough pictures
of discolorations.
00:08:35
This is akin to if a junior inspector
had learned to reliably spot tears,
00:08:39
but still needs to further hone
their judgment about discolorations.
00:08:43
So the inspector can go back
and take more pictures of discolorations
00:08:47
to show to the AI,
00:08:48
to help it deepen this understanding.
00:08:50
By adjusting the data you give to the AI,
00:08:53
you can help the AI get smarter.
00:08:56
So an inspector using
an accessible platform like this
00:09:00
can, in a few hours to a few days,
00:09:03
and with purchasing
a suitable camera set up,
00:09:07
be able to build a custom AI system
to detect defects,
00:09:11
tears and discolorations in all the fabric
00:09:13
being used to make T-shirts
throughout the factory.
00:09:16
And once again, you may say,
00:09:19
"Hey, Andrew, this is one factory.
00:09:22
Why is this a big deal?"
00:09:23
And I say to you,
00:09:25
this is a big deal to that inspector
whose life this makes easier
00:09:28
and equally, this type of technology
can empower a baker to use AI
00:09:32
to check for the quality
of the cakes they're making,
00:09:35
or an organic farmer to check
the quality of the vegetables,
00:09:39
or a furniture maker to check
the quality of the wood they're using.
00:09:44
Platforms like these will probably
still need a few more years
00:09:47
before they're easy enough to use
for every pizzeria owner.
00:09:51
But many of these platforms
are coming along,
00:09:53
and some of them
are getting to be quite useful
00:09:56
to someone that is tech savvy today,
00:09:58
with just a bit of training.
00:10:00
But what this means is that,
00:10:02
rather than relying
on the high priests and priestesses
00:10:04
to write AI systems for everyone else,
00:10:07
we can start to empower every accountant,
00:10:10
every store manager,
00:10:11
every buyer and every quality inspector
to build their own AI systems.
00:10:17
I hope that the pizzeria owner
00:10:19
and many other small
business owners like him
00:10:22
will also take advantage
of this technology
00:10:24
because AI is creating tremendous wealth
00:10:28
and will continue to create
tremendous wealth.
00:10:30
And it's only by
democratizing access to AI
00:10:33
that we can ensure that this wealth
is spread far and wide across society.
00:10:39
Hundreds of years ago.
00:10:41
I think hardly anyone
understood the impact
00:10:44
that widespread literacy will have.
00:10:47
Today, I think hardly anyone understands
00:10:50
the impact that democratizing
access to AI will have.
00:10:54
Building AI systems has been
out of reach for most people,
00:10:58
but that does not have to be the case.
00:11:01
In the coming era for AI,
00:11:03
we’ll empower everyone to build
AI systems for themselves,
00:11:06
and I think that will be
incredibly exciting future.
00:11:10
Thank you very much.
00:11:11
(Applause)
