LINNE HA: Hi, thank you for having me here.

My name is Linne.

And I am a director of research programs at Google AI.

I was here a couple of years ago,

where I talked about how much being in this position

and reaching this position was really about not me

but everybody behind me, meaning all the people who helped

me to get to this moment.

And afterwards, I thought about this for a little bit.

And of course, all of that is true.

But as I've been giving these talks,

and I give these talks around the world,

I noticed that people were asking me questions

like, how did you get your job?

And after so many of those questions, I started to think,

and I really kind of reacted defensively, in the sense that,

are they questioning that it's a fluke that I have this job?

That I actually don't deserve this job?

And I don't know, whatever reason.

Because they weren't asking all the other panelists with me.

And then I realized, actually, what that was really

an opportunity to tell my story about how I arrived here.

Because that journey is actually quite interesting.

My job at Google is super interesting.

But how I arrived at Google is also another very interesting

life story.

But let me just say that I didn't

plan on working for Google, or working in AI,

or working in tech.

I actually planned my life, since I was 9 years old,

to be a writer.

I spent all of my education going

towards becoming a writer, because I knew without a doubt.

I was the kid who went to the library, back

when libraries were still very popular

and books were very popular, taking out stacks of books.

And because I grew up in Guam and then in Alaska--

so these very remote, far-reaching--

where your creativity and imagination

were really important to transport you to new places.

So wanting to be a writer, I also

knew that I was probably not going

to get a very good job with an MFA.

So what I did as an undergrad is I studied as many things

as possible.

Also, because I'm kind of a nerd,

and I like to learn many things, so I

did double majors and double minors.

And then the ironic part of that is that when I graduated,

the jobs that were offered to me were

at an investment bank, or the CIA, or going into publishing.

And I really wanted to go into publishing.

But they were going to require me to work 50 hours a week,

making 17,000 a year.

And so basically, I was too scared to go to the CIA.

So I went to banking.

[LAUGHTER]

And I went to banking.

And what my plan was, living in New York--

because I came from Alaska.

I came to New York, and it was very magical to be here.

My plan was to work as much as I could, which was usually

about a year, year and a half, save up money, take off, go

to a residency, and write.

So I kept doing this for a long time.

And then one day, the job that I was offered was at Google.

And actually, I had a pretty decent job before that.

I had actually moved to San Francisco,

because I found it very expensive to live in New York.

And back then, San Francisco was very cheap.

[LAUGHTER]

Actually, I didn't apply.

I was already at a job, and the recruiter

found my CV on Monster, or Indeed, or one

of those on online boards.

And they offered to bring me down to the Mountain View

campus, the headquarters, which is about 40 minutes away

from San Francisco, and to interview, and give me

a tour, and a free lunch.

And I was thinking, oh, how else would I possibly

get a tour to Google unless I went through this whole thing?

I was like, oh, sure.

Why not?

And because I had a job, it was totally

fine whether the interview failed or not.

It didn't matter to me.

But I actually ended up meeting some of the most

interesting, smart people.

But I was still happy with my job.

And I didn't really want to commute down to Mountain View.

Then one day, my manager started questioning

and started micromanaging my work.

And I was like, you know what?

I don't really need to take this,

because I have a job offer from Google.

[LAUGHTER]

And back then, it was actually not as big of a deal

as it is now.

But it seemed more exciting.

So my first job was with a division

of Google called International.

So we didn't have all of these offices around the world.

The offices were on the Google campus,

which is in Mountain View, as I was saying.

And it was in one particular building where--

and they had flags--

all the marketing people, the product managers,

the localization people, the QA people,

were all in one building specifically dedicated

to international products.

So other than my first job, which

was really within the localization department.

And it was called International Program Manager.

Other than that first job, every subsequent one,

I created, actually.

And this is sort of the beauty of Google.

What happens is you start working on something.

And you see that there's something

wrong with the product.

And other people see it, too.

But you have an idea about how you can fix it.

And you convince everybody else that it's a good idea.

And then once it actually starts to flourish,

it becomes a viable product.

And many of Google products, including Gmail--

and I can't even think of all the products

that came out of this 20% idea.

So I was working on Google Earth.

And Google Earth had been a recent acquisition.

And the same people who, by the way, are doing the Pokemon Go--

Project Niantic, right?

And with that first job working on Google Earth, which

was a downloadable client application

for regular consumers, prosumers, and professionals--

so enterprise, right?

So there were three different versions.

And then you had the support for the Mac OS.

You had support for PCs, Windows,

and then you also had Linux support.

So it was quite unwieldy.

But in that acquisition, one of the recent mandates we'd had

was that Eric Schmidt, who was our CEO at that time,

wanted all of our products be internationalized

within a month.

So basically, in order to do that,

imagine all the different UIs and so on.

And when you're talking about geography,

it's not just a matter of translation.

In fact, names of cities, or bodies of water,

or mountains, or borders are highly contentious.

And these are, of course, what many wars are based on.

So basically, it was a very hot product,

but one nobody really wanted, because it was also

very challenging.

But I was given this because somebody

was going on maternity leave.

And basically, at Google, at that time,

you just had so much work.

It was one of those situations where

you could just work all day and never really catch up.

So with Google Earth, one of the things

that was really annoying for the nationalization of Google Earth

to rush out was we were going to do FIGS, which is usually

the first set of languages that you launch to, which is French,

Italian, German, Spanish.

So the left hand navigation bar was hardcoded.

It wasn't flexible.

It was hard coded based on the most clicks and queries.

So it was really menu items of food items like fast food.

Fast food was one category.

Pizza was another category.

Barbecue restaurants were another category,

and Italian was another category.

So in theory, that should be fine.

Because everybody eats pizza, and fast food, and barbecue,

and so on.

But what is an Italian restaurant in Italy?

How do you-- so coming up with this ontology

was something that we really wanted to fix.

So I found out that a lot of the data that we were getting

was from Google Maps.

So I'd worked with the team on working with Linguist

to figure out what should those categories actually be.

What are people looking for in these countries?

What are the most popular queries around businesses?

And so with Linguist, we started building up that ontology.

So it was actually flexible for each country

we were launching into.

And then, as we were building up these ontologies, which,

ultimately, ended up to be part of Google Maps,

as well, I moved to Google Maps so

that it would help the rest of the teams, if you will.

Everybody else could use this data.

So working with Google Maps on building these ontologies,

I was traveling around the world.

Because you really want to be able to go locally

to see how people call--

just to give an example, a motel in the US is a motor hotel.

It's something very convenient.

It's budget accommodations.

A motel in Japan or in Brazil, for instance,

is actually a sex hotel.

So you get charged by the hour.

So unless you know this, unless you could actually

get this localized, the information

is not going to be very useful.

So I would go to Japan often.

And one of the annoying things, for me,

was that the Japanese maps was all in Japanese.

And part of that was because most of the data at that time

was based on the license data that other people were

producing.

And it was what was most rich for that country.

So from that, I worked with some other people

who thought, oh, let's try to fix

this in some sort of algorithmic way.

So I worked with a speech team to come up

with some sort of pronunciation.

So the way you pronounce certain labels or certain words,

excuse me, we could actually go ahead and transliterate this

into Latin characters.

And then based on a set, let's say 20,000--

that's sort of a ballpark.

I don't remember if that is exactly what we did.

But based on a small set, we can actually train and test

whether we can generate these automatically.

So in Japanese, for instance, you

have many different character sets.

You have kanji.

You have katakana, hiragana, and romaji.

So you wanted to be able to get all of that transliterated

into Latin characters so that we can then

produce them into other characters like Chinese

or whatnot.

So Linguist would go through, and go and transliterate,

which is how something sounds into Latin characters.

So doing that, we're able to create new labels for maps.

And we did a launch.

And that formed a team called Maps Transliteration.

They still continue to do this, to this day.

The next thing was, because I had been working with a speech

team on those pronunciations, they actually had just formed,

and they were trying to launch--

and they were listening to Eric's mandate

of launching into 40 languages.

But there's a big difference between translating a UI

and actually creating a new language model.