WEBVTT

00:00:04.208 --> 00:00:07.128
So, I'm from Guatemala.

00:00:07.503 --> 00:00:09.381
This is a public service announcement,

00:00:09.381 --> 00:00:10.714
that is where Guatemala is.

00:00:10.756 --> 00:00:13.176
(Laughter)

00:00:13.634 --> 00:00:15.887
Also, that is not
where they keep the prisoners.

00:00:15.887 --> 00:00:17.222
That is called Guantanamo.

00:00:17.263 --> 00:00:18.390
(Laughter)

00:00:18.431 --> 00:00:19.599
Not the same place.

00:00:20.015 --> 00:00:22.226
So Guatemala's right below Mexico.

00:00:22.227 --> 00:00:24.478
And for the Americans in the audience,

00:00:24.521 --> 00:00:27.399
and let this sink in,
because it really applies in most ways,

00:00:27.399 --> 00:00:29.025
for the Americans in the audience,

00:00:29.067 --> 00:00:31.027
you can think of it as Mexico's Mexico.

00:00:31.068 --> 00:00:33.320
(Laughter)

00:00:33.362 --> 00:00:36.783
Just like the US doesn't want
illegal immigration from Mexico,

00:00:36.783 --> 00:00:39.953
Mexico doesn't want
illegal immigration from Guatemala.

00:00:39.953 --> 00:00:41.580
It's a smaller country.

00:00:41.621 --> 00:00:43.331
It's a poor country.

00:00:43.331 --> 00:00:45.834
And well, what can I tell you,

00:00:45.875 --> 00:00:47.418
it has much better Mexican food.

00:00:47.418 --> 00:00:50.921
(Laughter)

00:00:51.881 --> 00:00:53.716
Guatemala is a very poor country.

00:00:53.716 --> 00:00:56.051
And a lot of people talk about education

00:00:56.051 --> 00:00:58.972
as something that brings equality
to different social classes.

00:00:59.014 --> 00:01:02.475
But I always saw it as the opposite,
as something that brings inequality.

00:01:02.475 --> 00:01:04.019
Because what happens in practice

00:01:04.019 --> 00:01:07.856
is that people who have a lot of money
can buy themselves a really good education

00:01:07.897 --> 00:01:10.025
and therefore continue
having a lot of money.

00:01:10.025 --> 00:01:12.152
Whereas people who don't have
very much money

00:01:12.152 --> 00:01:13.778
barely learn how to read and write

00:01:13.778 --> 00:01:15.697
and therefore never make a lot of money.

00:01:15.739 --> 00:01:17.948
And this is especially true
in poor countries.

00:01:17.990 --> 00:01:21.453
Now, I was fortunate that I received
a rich person's education

00:01:21.493 --> 00:01:23.454
even though I didn't grow up rich.

00:01:23.495 --> 00:01:25.331
And it’s because I’m an only child.

00:01:25.331 --> 00:01:27.250
And my mother, who was a single mother,

00:01:27.292 --> 00:01:30.085
spent all of her resources
on my education.

00:01:30.462 --> 00:01:33.465
And this allowed me to come
to college to the US

00:01:33.506 --> 00:01:36.009
and eventually get a PhD
in computer science.

00:01:36.468 --> 00:01:39.888
Now because of all of this

00:01:39.888 --> 00:01:41.097
about 10 years ago,

00:01:41.096 --> 00:01:42.723
I decided I wanted to do something

00:01:42.766 --> 00:01:46.019
that would give equal access
to education to everyone.

00:01:46.061 --> 00:01:49.022
Oh, by the way, this is what I want
to talk to you about today,

00:01:49.022 --> 00:01:51.274
giving equal access
to education to everyone.

00:01:51.316 --> 00:01:53.902
At the time, I was a professor
of computer science

00:01:53.942 --> 00:01:55.819
at Carnegie Mellon University,

00:01:55.861 --> 00:02:00.158
and I decided to work on this
with my PhD student, Severin.

00:02:01.117 --> 00:02:02.368
The way my brain works,

00:02:02.409 --> 00:02:05.496
all of education is just
too general of a problem.

00:02:05.537 --> 00:02:08.500
So I decided, let's start
by teaching one thing.

00:02:08.957 --> 00:02:12.086
And then I started thinking,
OK, well, what should we teach first?

00:02:12.127 --> 00:02:13.295
Should we teach math?

00:02:13.295 --> 00:02:14.588
I mean, we both love math.

00:02:14.631 --> 00:02:16.549
Me and my cofounder, we both love math.

00:02:16.549 --> 00:02:19.718
And, you know, we didn't
decide to do math.

00:02:19.760 --> 00:02:23.097
Then we started thinking, well, maybe
we should teach computer science.

00:02:23.098 --> 00:02:26.518
But eventually, and this is maybe
surprising to people in North America,

00:02:26.558 --> 00:02:29.603
eventually, we decided
that the best subject to start with

00:02:29.645 --> 00:02:31.480
was teaching foreign languages.

00:02:31.897 --> 00:02:33.316
And let me tell you why.

00:02:33.358 --> 00:02:34.734
There's a number of reasons.

00:02:35.193 --> 00:02:37.654
One of them is that there's
a huge audience for it.

00:02:37.695 --> 00:02:41.199
There’s about two billion people
in the world learning a foreign language,

00:02:41.199 --> 00:02:43.034
both in school and outside of schools.

00:02:43.033 --> 00:02:45.620
Most of these people, by the way,
are learning English.

00:02:45.662 --> 00:02:47.831
About 80 percent of them
are learning English.

00:02:47.830 --> 00:02:49.248
In this weird map right here,

00:02:49.248 --> 00:02:50.500
all the countries in green

00:02:50.542 --> 00:02:53.628
are countries in which people
are predominantly learning English.

00:02:53.627 --> 00:02:54.795
And the reason for that

00:02:54.837 --> 00:02:57.132
is because English can truly
transform your life.

00:02:57.132 --> 00:02:58.633
In most countries in the world,

00:02:58.633 --> 00:03:02.262
knowledge of English can significantly
increase your income potential.

00:03:02.302 --> 00:03:05.764
So this is another reason why we decided
to start with foreign languages.

00:03:05.806 --> 00:03:08.143
And it's because you can directly
make more money

00:03:08.183 --> 00:03:10.687
if you learn another language,
in particular English.

00:03:10.687 --> 00:03:13.481
And see, this is pretty different
than many other subjects.

00:03:13.480 --> 00:03:14.649
For example, take math.

00:03:14.691 --> 00:03:15.859
In the case of math,

00:03:15.900 --> 00:03:18.819
just knowledge of math does not
increase your income potential

00:03:18.861 --> 00:03:21.698
because usually you have to learn math
to then learn physics

00:03:21.698 --> 00:03:24.701
to then become a civil engineer,
that's how you make more money.

00:03:24.742 --> 00:03:27.870
Whereas with languages,
if you are a waiter and you learn English,

00:03:27.871 --> 00:03:30.915
now you can be a waiter at a hotel
and make more money.

00:03:30.956 --> 00:03:34.126
So we decided, well, let's get started
by teaching languages.

00:03:34.127 --> 00:03:36.421
Now, we also convinced ourselves

00:03:36.462 --> 00:03:40.549
that the only way
to really reach a lot of people

00:03:40.550 --> 00:03:43.927
was by using a mobile phone,
or a smartphone in particular.

00:03:44.304 --> 00:03:47.390
See, building schools all over the world
is simply too expensive.

00:03:47.806 --> 00:03:49.141
On the other hand,

00:03:49.141 --> 00:03:52.269
most of the world's population
already has access to a smartphone,

00:03:52.270 --> 00:03:55.231
and the trend is that that fraction
is only going to increase.

00:03:55.230 --> 00:03:56.649
So we decided at the time

00:03:56.649 --> 00:04:00.694
that we would make a way to learn
foreign languages on a mobile phone

00:04:00.694 --> 00:04:02.530
that was accessible to everyone.

00:04:02.572 --> 00:04:04.616
And then we called it Duolingo.

00:04:04.949 --> 00:04:06.117
Thank you.

00:04:06.116 --> 00:04:09.286
(Applause)

00:04:09.329 --> 00:04:13.081
Now in order to truly be accessible
to everyone, rich and poor,

00:04:13.123 --> 00:04:16.627
Duolingo uses a freemium model
to support itself.

00:04:16.668 --> 00:04:19.963
What that means is that you can learn
as much as you want

00:04:20.005 --> 00:04:22.007
without ever having to pay.

00:04:22.341 --> 00:04:23.509
But if you don't pay,

00:04:23.509 --> 00:04:25.803
you may have to see an ad
at the end of a lesson.

00:04:25.845 --> 00:04:27.138
Now, if you don't like ads,

00:04:27.180 --> 00:04:30.016
you can also pay to subscribe
to turn off the ads.

00:04:30.057 --> 00:04:33.435
And it turns out that the vast majority
of the revenue for Duolingo

00:04:33.478 --> 00:04:36.396
comes from people who pay
to subscribe to turn off the ads.

00:04:38.233 --> 00:04:41.444
Now, who are these people who pay
to subscribe to turn off the ads?

00:04:41.485 --> 00:04:47.324
Well, they're usually well-off people
in rich countries like the US and Canada.

00:04:47.367 --> 00:04:49.536
Who are the people
who don't pay to subscribe?

00:04:49.536 --> 00:04:53.665
They usually come from poorer countries
like Brazil or Vietnam or Guatemala.

00:04:53.997 --> 00:04:55.500
So what I like about this model

00:04:55.541 --> 00:04:57.960
is that it is a small form
of wealth redistribution

00:04:57.961 --> 00:05:00.170
because we're basically
getting the rich people

00:05:00.170 --> 00:05:02.005
to pay for the education of everyone.

00:05:02.005 --> 00:05:03.173
So I like that.

00:05:03.216 --> 00:05:06.718
(Applause)

00:05:06.718 --> 00:05:09.137
So with smartphones,
we can reach a lot of people

00:05:09.137 --> 00:05:12.182
and we can even get the rich people
to pay for the whole thing,

00:05:12.182 --> 00:05:13.350
which is great.

00:05:13.392 --> 00:05:16.478
However, if you're trying to deliver
education with a smartphone,

00:05:16.478 --> 00:05:18.063
you run into a humongous problem.

00:05:18.398 --> 00:05:20.733
And it is that smartphones come equipped

00:05:20.774 --> 00:05:24.611
with some of the most addictive drugs
that humanity has ever engineered.

00:05:24.612 --> 00:05:27.323
TikTok, Instagram, mobile games.

00:05:27.949 --> 00:05:29.951
See, delivering education
over a smartphone

00:05:29.992 --> 00:05:32.411
is like hoping that people
will eat their broccoli,

00:05:32.411 --> 00:05:35.581
but right next to it, you put
the most delicious dessert ever made.

00:05:35.581 --> 00:05:36.748
(Laughter)

00:05:38.000 --> 00:05:40.460
If you really want to deliver
education to everyone,

00:05:40.502 --> 00:05:42.504
not only do you have
to make it accessible,

00:05:42.504 --> 00:05:46.675
but also you have to make it
so that people want to actually learn.

00:05:46.675 --> 00:05:48.844
And with Duolingo,
we've been able to do this.

00:05:48.887 --> 00:05:51.180
And at the highest level,
the way we've done this

00:05:51.221 --> 00:05:54.099
is by making the broccoli
taste like dessert.

00:05:55.392 --> 00:05:56.560
I'll say it another way.

00:05:56.560 --> 00:05:59.855
What we've done is that we've used
the same psychological techniques

00:05:59.855 --> 00:06:04.026
that apps like Instagram,
TikTok or mobile games use

00:06:04.026 --> 00:06:05.235
to keep people engaged,

00:06:05.278 --> 00:06:07.780
but in this case, we use them
to keep people engaged

00:06:07.779 --> 00:06:09.489
but with education.

00:06:09.490 --> 00:06:11.826
Let me give you some examples
of these techniques.

00:06:11.867 --> 00:06:14.495
One of the most powerful ones
is the notion of a streak.

00:06:14.495 --> 00:06:16.747
What a streak is, is it’s just a counter

00:06:16.788 --> 00:06:18.582
that measures the number of days

00:06:18.625 --> 00:06:21.293
that you've used
the product consecutively.

00:06:21.336 --> 00:06:22.586
You just take that number,

00:06:22.629 --> 00:06:24.713
you put it very prominently
in your product

00:06:24.713 --> 00:06:27.007
and then people come back every day.

00:06:27.050 --> 00:06:30.929
And the reason people come back
every day is because, well,

00:06:30.970 --> 00:06:33.389
if they don't come back,
that number resets to zero

00:06:33.389 --> 00:06:35.516
and people don't want
to lose their streak.

00:06:35.557 --> 00:06:36.683
It works.

00:06:37.100 --> 00:06:38.644
Now, on the one side,

00:06:38.644 --> 00:06:42.439
streaks have been criticized
for, for example,

00:06:42.482 --> 00:06:44.525
getting teens addicted to Snapchat.

00:06:44.942 --> 00:06:47.027
But in the case of an educational app,

00:06:47.069 --> 00:06:49.738
streaks get people to come back
to study every day.

00:06:50.281 --> 00:06:52.574
Now, to give you an idea
of the power of streaks,

00:06:52.617 --> 00:06:53.785
in the case of Duolingo,

00:06:53.826 --> 00:06:55.954
we have over three million
daily active users

00:06:55.954 --> 00:06:58.288
that have a streak longer than 365.

00:06:58.747 --> 00:07:02.584
(Applause)

00:07:02.627 --> 00:07:05.713
That means they haven't missed a day
in the last year or longer.

00:07:05.754 --> 00:07:07.464
Now, fun fact about streaks.

00:07:07.464 --> 00:07:11.635
What country do you think
has the longest average streaks

00:07:11.636 --> 00:07:13.179
for an educational app?

00:07:13.596 --> 00:07:14.764
It's Japan.

00:07:15.598 --> 00:07:16.766
Of course.

00:07:17.307 --> 00:07:18.892
Shortest-ever streaks?

00:07:19.560 --> 00:07:21.019
Latin America, baby.

00:07:21.019 --> 00:07:22.729
(Laughter)

00:07:22.730 --> 00:07:24.232
But we're fun, hey, we're fun.

00:07:24.273 --> 00:07:26.317
(Laughter)

00:07:26.858 --> 00:07:30.278
Another important mechanism to get people
to come back to your product

00:07:30.321 --> 00:07:31.489
are notifications.

00:07:31.488 --> 00:07:34.574
On the one side, notifications
can be really spammy and annoying,

00:07:34.574 --> 00:07:36.618
but in the case of an educational product,

00:07:36.619 --> 00:07:38.746
people actually want
to be reminded to learn.

00:07:38.745 --> 00:07:42.332
In the case of Duolingo, we have
a very sophisticated AI system

00:07:42.375 --> 00:07:46.336
that basically chooses
when to send the notification

00:07:46.379 --> 00:07:48.338
and also what to say in each notification

00:07:48.380 --> 00:07:51.174
to maximize the probability
that people come back.

00:07:51.216 --> 00:07:53.176
Now, interestingly,

00:07:53.177 --> 00:07:55.471
even after all this sophistication,

00:07:55.512 --> 00:07:58.056
it turns out that the algorithm
for choosing what time

00:07:58.057 --> 00:08:00.184
to send you a notification
is pretty simple.

00:08:00.184 --> 00:08:03.187
Do you know what is the best time
to send people a notification?

00:08:03.228 --> 00:08:04.396
I'll tell you.

00:08:04.396 --> 00:08:06.649
It's 24 hours after
they used the product last.

00:08:06.690 --> 00:08:08.442
There's an easy explanation.

00:08:08.442 --> 00:08:10.318
If you were free yesterday at 3pm,

00:08:10.360 --> 00:08:12.612
you’re probably free today
at 3pm as well.

00:08:13.156 --> 00:08:16.908
So this is what a very sophisticated
millions of dollars of AI found.

00:08:16.908 --> 00:08:19.120
(Laughter)

00:08:19.161 --> 00:08:20.787
It's funny.

00:08:20.829 --> 00:08:23.665
Now with notifications
you shouldn't be spamming.

00:08:23.665 --> 00:08:25.500
And we're not spamming, with Duolingo,

00:08:25.543 --> 00:08:28.838
we actually stop sending notifications
after seven days of inactivity.

00:08:28.879 --> 00:08:30.964
So if you don't use Duolingo
for seven days,

00:08:30.964 --> 00:08:32.591
we stop sending you notifications.

00:08:32.633 --> 00:08:34.427
Now, at some point it occurred to us,

00:08:34.427 --> 00:08:36.637
if we're stopping
to send people notifications,

00:08:36.678 --> 00:08:37.846
we should let them know.

00:08:37.889 --> 00:08:40.600
So we started sending
this notification to people saying,

00:08:40.600 --> 00:08:42.810
"Hey, these reminders
don't seem to be working.

00:08:42.809 --> 00:08:44.394
We'll stop sending them for now."

00:08:44.437 --> 00:08:47.065
You know what people do
when they get this notification?

00:08:47.065 --> 00:08:48.232
They come back.

00:08:48.274 --> 00:08:49.609
(Laughter)

00:08:50.359 --> 00:08:51.694
Passive aggressive.

00:08:51.693 --> 00:08:54.072
(Laughter)

00:08:54.113 --> 00:08:56.074
Works for my mother, works for Duolingo.

00:08:56.073 --> 00:09:02.288
(Laughter and applause)

00:09:02.996 --> 00:09:04.831
These passive aggressive notifications

00:09:04.831 --> 00:09:07.043
are really good at getting
people to come back

00:09:07.043 --> 00:09:10.337
back because they feel like our green
owl mascot has given up on them,

00:09:10.379 --> 00:09:11.547
so they come back.

00:09:11.547 --> 00:09:13.841
And speaking of our
green owl mascot, by the way,

00:09:13.883 --> 00:09:16.719
because all our notifications come
from our green owl mascot

00:09:16.760 --> 00:09:19.387
and well, he's passive aggressive
and also pretty pushy,

00:09:19.388 --> 00:09:21.891
this has given rise to a lot
of memes on the internet

00:09:21.890 --> 00:09:24.644
that make fun of the great lengths
that he will go through

00:09:24.644 --> 00:09:26.687
to get you back to learn a language.

00:09:26.729 --> 00:09:28.272
Here's one of my favorite ones.

00:09:28.272 --> 00:09:30.817
This is it's a meme,
it’s one of my favorite ones.

00:09:30.817 --> 00:09:32.025
(Laughter)

00:09:32.025 --> 00:09:35.153
Basically looks like you forgot
your Spanish lessons.

00:09:35.196 --> 00:09:36.906
And then there's an intruder alert,

00:09:36.905 --> 00:09:40.450
presumably the owl broke into your house
to get you to learn language.

00:09:40.493 --> 00:09:43.079
Now, Duolingo has entered the zeitgeist.

00:09:43.078 --> 00:09:46.331
And there's thousands of memes,
there's SNL skits about it.

00:09:46.331 --> 00:09:50.794
And it's because we've managed
to get people to want to learn a language

00:09:50.836 --> 00:09:52.504
by using the same techniques

00:09:52.504 --> 00:09:57.384
that mobile games and social media use
to get people engaged.

00:09:57.802 --> 00:10:00.303
And this is a really important point,

00:10:00.346 --> 00:10:01.596
let me say this.

00:10:02.264 --> 00:10:05.768
I don't actually believe that there's
a way to make an educational app

00:10:05.809 --> 00:10:09.980
be as engaging as something like TikTok
or Instagram or mobile games.

00:10:11.023 --> 00:10:12.399
But the good news is that --

00:10:12.399 --> 00:10:14.610
And by the way, the reason
I don't believe that

00:10:14.652 --> 00:10:17.363
is because ultimately you have
to teach people something.

00:10:17.363 --> 00:10:20.074
And it's hard to compete with,
like, cats and celebrities.

00:10:20.073 --> 00:10:22.617
But the good news is
that I don't think you have to.

00:10:22.618 --> 00:10:23.828
See, here's the thing.

00:10:23.870 --> 00:10:26.621
When you're learning something,
you get meaning out of it.

00:10:26.621 --> 00:10:29.332
Whereas when you're scrolling
for two hours on Instagram,

00:10:29.332 --> 00:10:32.502
a lot of times afterwards, you feel
like you just wasted your time.

00:10:32.503 --> 00:10:36.674
So I think it's actually OK
if your educational product

00:10:36.673 --> 00:10:39.759
is only 80 or 90 percent as engaging
as something like TikTok,

00:10:39.802 --> 00:10:41.637
because the other 10 or 20 percent

00:10:41.678 --> 00:10:43.972
will be provided
by people's internal motivation,

00:10:44.014 --> 00:10:46.057
though of course, not much more than that.

00:10:46.100 --> 00:10:47.475
This is really a key point.

00:10:47.518 --> 00:10:50.980
If we want to get people
to do something meaningful,

00:10:51.022 --> 00:10:54.650
you can use the same techniques
that apps like social media use

00:10:54.649 --> 00:10:56.026
to get people to do it.

00:10:56.067 --> 00:10:59.237
And even if you're not as engaging
as those apps are,

00:10:59.238 --> 00:11:02.700
you can still get hundreds of millions
of people to use your product.

00:11:02.741 --> 00:11:04.993
In the case of Duolingo, for example,

00:11:04.994 --> 00:11:08.456
there are more people learning languages
on Duolingo in the United States

00:11:08.456 --> 00:11:12.084
than there are people learning languages
across all US high schools combined.

00:11:12.125 --> 00:11:13.376
(Applause)

00:11:13.418 --> 00:11:16.004
And this is true
in most countries in the world.

00:11:16.047 --> 00:11:20.426
My hope is that -- I know
we can do this, but, you know,

00:11:20.467 --> 00:11:21.885
my hope is that as humanity,

00:11:21.927 --> 00:11:24.554
we can do what Duolingo has done
for learning languages

00:11:24.596 --> 00:11:26.057
but for all other subjects.

00:11:26.474 --> 00:11:29.936
Where we can get people
to learn math with mobile phones,

00:11:29.976 --> 00:11:32.605
like millions of people to learn
math with mobile phones

00:11:32.605 --> 00:11:34.273
or physics or whatever.

00:11:34.315 --> 00:11:38.653
I hope for a future in which
screen time is not a bad thing,

00:11:38.693 --> 00:11:42.155
in which we can deliver high-quality
education to everyone, rich or poor,

00:11:42.155 --> 00:11:43.323
using a mobile phone.

00:11:43.658 --> 00:11:47.702
But the single most important thing
that I can end this talk with,

00:11:47.702 --> 00:11:52.207
is a reminder to please,
pretty please, I beg you,

00:11:52.207 --> 00:11:53.750
do your language lessons today.

00:11:53.750 --> 00:11:55.001
Thank you very much.

00:11:55.001 --> 00:11:56.670
(Laughter)

00:11:56.711 --> 00:11:57.922
Thank you.

00:11:57.964 --> 00:12:04.887
(Applause)

00:12:05.596 --> 00:12:07.682
Helen Walters: Luis, thank you so much.

00:12:07.682 --> 00:12:10.976
I wonder if you can just say
a little bit more about that last point,

00:12:10.976 --> 00:12:14.938
how do you think you can apply
this type of thinking

00:12:14.938 --> 00:12:16.690
to other subjects?

00:12:16.731 --> 00:12:18.817
So you mentioned math,
and things like that.

00:12:18.817 --> 00:12:20.235
How do you do that?

00:12:20.278 --> 00:12:22.071
Luis von Ahn: I think in particular,

00:12:22.113 --> 00:12:24.198
subjects that are learned
through repetition

00:12:24.197 --> 00:12:27.158
and it turns out most things
that are kind of really meaningful

00:12:27.201 --> 00:12:29.327
are learned through
thousands of repetitions.

00:12:29.327 --> 00:12:31.121
You learn to read through repetition,

00:12:31.121 --> 00:12:33.581
you learn elementary school math
through repetition.

00:12:33.581 --> 00:12:35.918
Most things that you can learn
through repetition,

00:12:35.960 --> 00:12:38.879
you can actually gamify
and turn into something like Duolingo,

00:12:38.879 --> 00:12:40.965
where people just
do it a lot and do it fun.

00:12:40.965 --> 00:12:43.301
It's a little harder for things
like explanations.

00:12:43.341 --> 00:12:46.052
That probably is going to require
some really good videos.

00:12:46.095 --> 00:12:48.264
Sal Khan is doing a really
good job with that.

00:12:48.264 --> 00:12:50.515
But for things that require
a lot of repetition,

00:12:50.515 --> 00:12:52.267
I think we can use the same methods.

00:12:52.309 --> 00:12:53.435
(Applause)