The following is a conversation with Rajat Manga.
He's an engineering director at Google,
leading the TensorFlow team.
TensorFlow is an open source library
at the center of much of the work going on in the world
in deep learning, both the cutting edge research
and the large scale application of learning based approaches.
But it's quickly becoming much more
than a software library.
It's now an ecosystem of tools for the deployment
of machine learning in the cloud, on the phone,
in the browser, on both generic and specialized hardware.
TPU, GPU, and so on.
Plus, there's a big emphasis on growing
a passionate community of developers.
Rajat, Jeff Dean, and a large team of engineers at Google
Brain are working to define the future of machine learning
with TensorFlow 2.0, which is now in alpha.
I think the decision to open source TensorFlow
is a definitive moment in the tech industry.
It showed that open innovation can be successful
and inspire many companies to open source their code,
to publish, and in general engage in the open exchange
of ideas.
This conversation is part of the Artificial Intelligence
podcast.
If you enjoy it, subscribe on YouTube, iTunes,
or simply connect with me on Twitter
at Lex Friedman, spelled F-R-I-D.
And now, here's my conversation with Rajat Manga.
You were involved with Google Brain
since its start in 2011 with Jeff Dean.
It started with disbelief, the proprietary machine learning
library, and turned into TensorFlow 2014,
the open source library.
So what were the early days of Google Brain like?
What were the goals, the missions?
How do you even proceed forward once there's
so much possibilities before you?
It was interesting back then when I started,
or when you were even just talking about it.
The idea of deep learning was interesting
and intriguing in some ways.
It hadn't yet taken off, but it held some promise.
It had shown some very promising and early results.
I think the idea where Andrew and Jeff had started
was what if we can take this, what people are doing in research,
and scale it to what Google has in terms of the compute power,
and also put that kind of data together, what does it mean?
And so far, the results had been if you scale the computer,
scale the data, it does better, and would that work?
And so that was the first year or two,
can we prove that out, right?
And with disbelief, when we started the first year,
we got some early wins, which is always great.
What were the wins like?
What was the wins where there are some problems to this?
This is going to be good.
I think the two early wins were one was speech
that we collaborated very closely with the speech research
team who was also getting interested in this.
And the other one was on images where
the cat paper, as we call it, that was covered by a lot of folks.
And the birth of Google Brain was around neural networks.
So it was deep learning from the very beginning.
That was the whole mission.
So in terms of scale, what was the sort of dream
of what this could become?
Were there echoes of this open source TensorFlow community
that might be brought in?
Was there a sense of TPUs?
Was there a sense of machine learning
is now going to be at the core of the entire company?
Is going to grow into that direction?
Yeah, I think so that was interesting.
And if I think back to 2012 or 2011,
and first was can we scale it in the year or so,
we had started scaling it to hundreds and thousands
of machines.
In fact, we had some runs even going to 10,000 machines.
And all of those shows great promise.
In terms of machine learning at Google,
the good thing was Google's been doing machine learning
for a long time.
Deep learning was new.
But as we scale this up, we showed that, yes, that was
possible, and it was going to impact lots of things.
Like, we started seeing real products wanting to use this.
Again, speech was the first.
There were image things that photos came out of
and then many other products as well.
So that was exciting.
As we went into with that a couple of years,
externally also academia started to,
there was lots of push on, OK, deep learning's
interesting, we should be doing more, and so on.
And so by 2014, we were looking at, OK, this is a big thing.
It's going to grow.
And not just internally, externally as well.
Yes, maybe Google's ahead of where everybody is,
but there's a lot to do.
So a lot of this start to make sense and come together.
So the decision to open source, I was just chatting with Chris
Flattner about this, the decision to go open source
with TensorFlow, I would say for me personally,
seems to be one of the big seminal moments in all
of software engineering ever.
I think that when a large company like Google
decides to take a large project that many lawyers might argue
has a lot of IP, just decide to go open source with it.
And in so doing, lead the entire world in saying,
you know what, open innovation is a pretty powerful thing.
And it's OK to do.
That was, I mean, that's an incredible moment in time.
So do you remember those discussions happening?
Are there open sources should be happening?
What was that like?
I would say, I think, so the initial idea came from Jeff,
who was a big proponent of this.
I think it came off of two big things.
One was research-wise.
We were a research group.
We were putting all our research out there if you wanted to.
We were building on other's research,
and we wanted to push the state of the art forward.
And part of that was to share the research.
That's how I think deep learning and machine learning
has really grown so fast.
So the next step was, OK, now word software help for that.
And it seemed like they were existing a few libraries out
there, Tiano being one, Torch being another, and a few others.
But they were all done by academia,
and so the level was significantly different.
The other one was, from a software perspective,
Google had done lots of software that we used internally,
and we published papers.
Often, there was an open source project
that came out of that that somebody else picked up
that paper and implemented, and they were very successful.
Back then, it was like, OK, there's
Hadoop, which has come off of tech that we've built.
We know that tech we've built is way better
for a number of different reasons.
We've invested a lot of effort in that.
And turns out, we have Google Cloud,
and we are now not really providing our tech,
but we are saying, OK, we have Bigtable,
which is the original thing.
We are going to now provide HBase APIs on top of that, which
isn't as good, but that's what everybody's used to.
So there's like, can we make something that is better
and really just provide?
Helps the community in lots of ways,
but it also helps push the right, a good standard forward.
So how does Cloud fit into that?
There's a TensorFlow open source library,
and how does the fact that you can use so many of the resources
that Google provides and the Cloud fit into that strategy?
So TensorFlow itself is open, and you can use it anywhere,
and we want to make sure that continues to be the case.
On Google Cloud, we do make sure that there's lots
of integrations with everything else,
and we want to make sure that it works really, really well there.
You're leading the TensorFlow effort.
Can you tell me the history and the timeline of TensorFlow
project in terms of major design decisions,
like the open source decision, but really, what to include
and not?
There's this incredible ecosystem that I'd
like to talk about as all of these parts.
But if you just some sample moments that
defined what TensorFlow eventually became through its,
I don't know if you were allowed to say history when it's just,
but in deep learning, everything moves so fast
in just a few years, it's already history.
Yes, yes.
So looking back, we were building TensorFlow,
I guess we open sourced it in 2015, November 2015.
We started on it in summer of 2014, I guess.
And somewhere like three to six late 2014,
by then we had decided that, OK, there's
a high likelihood we'll open source it.
So we started thinking about that and making sure
that we're heading down that path.
By that point, we had seen a few lots of different use cases
at Google.
So there were things like, OK, yes,
you want to run in at large scale in the data center.
Yes, we need to support different kind of hardware.
We had GPUs at that point.
We had our first GPU at that point
or was about to come out roughly around that time.
So the design included those.
We had started to push on mobile.
So we were running models on mobile.
At that point, people were customizing code.
So we wanted to make sure TensorFlow could support that
as well so that that became part of that overall
design.
When you say mobile, you mean like pretty complicated
algorithms of running on the phone?
That's correct.
So when you have a model that you
deploy on the phone and run it the right time.
So already at that time, there was ideas of running machine
learning on the phone?
That's correct.
We already had a couple of products
that were doing that by then.
And in those cases, we had basically
customized handcrafted code or some internal libraries
that we're using.
So I was actually at Google during this time in a parallel,
I guess, universe.
But we were using Theano and CAFE.
Was there some degree to which you were bouncing,
like trying to see what CAFE was offering people,
trying to see what Theano was offering
that you want to make sure you're delivering on whatever that
is, perhaps the Python part of thing.
Maybe did that influence any design decisions?
Totally.
So when we built this belief, and some of that
was in parallel with some of these libraries
coming up, I mean, Theano itself is older.
But we were building this belief focused on our internal thing
because our systems were very different.
By the time we got to this, we looked
at a number of libraries that were out there.
Theano, there were folks in the group
who had experience with Torch, with Lua.
There were folks here who had seen CAFE.
I mean, actually, Yang Cheng was here as well.
There's, what other libraries?
I think we looked at a number of things.
Might even have looked at Jane and her back then.
I'm trying to remember if it was there.
In fact, yeah, we did discuss ideas around, OK,
should we have a graph or not?
And they were, so putting all these together
was definitely, you know, there were key decisions
that we wanted.
We had seen limitations in our prior disbelief things.
A few of them were just in terms of research
was moving so fast.
We wanted the flexibility.
We want the hardware was changing fast.
We expected to change that so that those probably
were two things.
And yeah, I think the flexibility in terms
of being able to express all kinds of crazy things
was definitely a big one then.
So what the graph decisions, though,
with moving towards TensorFlow 2.0, there's more, by default,
there'll be eager execution.
So sort of hiding the graph a little bit
because it's less intuitive in terms of the way
people develop and so on.
What was that discussion like with in terms of using graphs?
It's kind of the theano way.
Did it seem the obvious choice?
So I think where it came from was our disbelief,
had a graph like thing as well.
It wasn't a general graph.
It was more like a straight line thing.
More like what you might think of cafe, I guess, in that sense.
But the graph was, and we always cared
about the production stuff, like even with disbelief,
we were deploying a whole bunch of stuff in production.
So graph did come from that when we thought of, OK,
should we do that in Python and re-experiment with some ideas
where it looked a lot simpler to use,
but not having a graph meant, OK, how do you deploy now?
So that was probably what tilted the balance for us.
And eventually, we ended up with the graph.
And I guess the question there is, did you?
I mean, production seems to be a really good thing to focus on.
But did you even anticipate the other side of it
where there could be, what is it?
What are the numbers?
Something crazy, 41 million downloads?
Yep.
I mean, was that even like a possibility in your mind
that it would be as popular as it became?
So I think we did see a need for this a lot
from the research perspective and early days
of deep learning in some ways.
41 million?
No, I don't think I imagined this number then.
It seemed like there's a potential future where lots more
people would be doing this.
And how do we enable that?
I would say this kind of growth, I probably
started seeing somewhat after the open sourcing where it was
like, OK, deep learning is actually
growing way faster for a lot of different reasons.
And we are in just the right place to push on that
and leverage that and deliver on lots of things
that people want.
So what changed once the open source?
Like how this incredible amount of attention
from a global population of developers,
what, how did the projects start changing?
I don't even actually remember it during those times.
I know looking now, there's really good documentation.
There's an ecosystem of tools.
There's a YouTube channel now.
It's very, very community driven.
Back then, I guess 0.1 version.
Is that the version?
I think we called 0.6 or 5, something like that.
I forgot about that.
What changed leading into 1.0?
It's interesting.
I think we've gone through a few things there.
When we started out, when we first came out,
people loved the documentation we have.
Because it was just a huge step up from everything else.
Because all of those were academic projects people doing
who don't think about documentation.
I think what that changed was instead of deep learning
being a research thing, some people who were just developers
could now suddenly take this out and do some interesting things
with it, who had no clue what machine learning was before then.
And that, I think, really changed
how things started to scale up in some ways and pushed on it.
Over the next few months, as we looked at how do we stabilize
things, as we look at not just researchers,
now we want stability.
People want to deploy things.
That's how we started planning for 1.0.
And there are certain needs for that perspective.
And so again, documentation comes up.
Designs, more kinds of things to put that together.
And so that was exciting to get that to a stage
where more and more enterprises wanted to buy in
and really get behind that.
And I think post 1.0 and with the next few releases,
their enterprise adoption also started to take off.
I would say between the initial release and 1.0,
it was, OK, researchers, of course.
Then a lot of hobbies and early interest,
people excited about this who started to get on board.
And then over the 1.x thing, lots of enterprises.
I imagine anything that's below 1.0 gets
pressured to be an enterprise problem or something
that's stable.
Exactly.
And do you have a sense now that TensorFlow is stable?
It feels like deep learning in general
is extremely dynamic field.
So much is changing.
And TensorFlow has been growing incredibly.
Do you have a sense of stability at the helm of it?
I mean, I know you're in the midst of it.
I think in the midst of it, it's often
easy to forget what an enterprise wants
and what some of the people on that side want.
There are still people running models
that are three years old, four years old.
So inception is still used by tons of people.
Even less than 50 is what?
A couple of years old now or more.
But there are tons of people who use that, and they're fine.
They don't need the last couple of bits of performance
or quality.
They want some stability in things that just work.
And so there is value in providing that with that kind
of stability and making it really simpler,
because that allows a lot more people to access it.
And then there's the research crowd, which wants, OK,
they want to do these crazy things exactly
like you're saying.
Not just deep learning in the straight-up models
that used to be there.
They want RNNs.
And even RNNs are maybe old.
They are transformers now.
And now it needs to combine with RL and GANs and so on.
So there's definitely that area, like the boundary that's
shifting and pushing the state of the art.
But I think there's more and more of the past
that's much more stable.
And even stuff that was two, three years old
is very, very usable by lots of people.
So that part makes it a lot easier.
So I imagine maybe you can correct me if I'm wrong.
One of the biggest use cases is essentially
taking something like ResinF50 and doing
some kind of transfer learning on a very particular problem
that you have.
It's basically probably what majority of the world does.
And you want to make that as easy as possible.
So I would say, for the hobbyist perspective,
that's the most common case.
In fact, the apps on phones and stuff
that you'll see, the early ones, that's the most common case.
I would say there are a couple of reasons for that.
One is that everybody talks about that.
It looks great on slides.
That's a great presentation.
Exactly.
What enterprises want is that is part of it,
but that's not the big thing.
Enterprises really have data that they
want to make predictions on.
This is often what they used to do with the people who
were doing ML was just regression models, linear
regression or just a regression, linear models,
or maybe gradient booster trees and so on.
Some of them still benefit from deep learning,
but they weren't that that's the bread and butter,
like the structured data and so on.
So depending on the audience you look at,
they're a little bit different.
And they just have, I mean, the best of enterprise probably
just has a very large data set where deep learning can
probably shine.
That's correct.
That's right.
And then I think the other pieces that they want, again,
with 2.0 or the developer summit we put together
is that the whole TensorFlow Extended
piece, which is the entire pipeline,
they care about stability across doing their entire thing.
They want simplicity across the entire thing.
I don't need to just train a model.
I need to do that every day again, over and over again.
I wonder to which degree you have a role in, I don't know.
So I teach a course on deep learning.
I have people like lawyers come up to me and say,
when is machine learning going to enter legal realm?
The same thing in all kinds of disciplines,
immigration, insurance, often when
I see what it boils down to is these companies are often
a little bit old school in the way they organize the day.
So the data is just not ready yet.
It's not digitized.
Do you also find yourself being in the role of an evangelist
for like, let's get, organize your data, folks.
And then you'll get the big benefit of TensorFlow.
Do you have those conversations?
Yeah, yeah.
I get all kinds of questions there from, OK, what can I,
what do I need to make this work, right?
Do we really need deep learning?
I mean, there are all these things.
I already used this linear model.
Why would this help?
I don't have enough data, let's say.
Or I want to use machine learning,
but I have no clue where to start.
So it varies.
Start to all the way to the experts who are very specific
things, so it's interesting.
Is there a good answer?
It boils down to oftentimes digitizing data.
So whatever you want automated, whatever data
you want to make prediction based on,
you have to make sure that it's in an organized form.
Like with an intensive flow ecosystem,
there's now you're providing more and more data sets
and more and more pre-trained models.
Are you finding yourself also the organizer of data sets?
Yes.
I think with TensorFlow data sets that we just released,
that's definitely come up where people want these data sets.
Can we organize them?
And can we make that easier?
So that's definitely one important thing.
The other related thing I would say is I often tell people,
you know what, don't think of the most fanciest thing
that the newest model that you see.
Make something very basic work, and then you can improve it.
There's just lots of things you can do with it.
Yeah, start with the basics.
Sure.
One of the big things that makes TensorFlow even more
accessible was the appearance whenever
that happened of Keras, the Keras standard outside
of TensorFlow.
I think it was Keras on top of Tiano at first only,
and then Keras became on top of TensorFlow.
Do you know when Keras chose to also add TensorFlow
as a back end, was it just the community that
drove that initially?
Do you know if there was discussions, conversations?
Yeah, so Franco started the Keras project
before he was at Google.
And the first thing was Tiano.
I don't remember if that was after TensorFlow was created
or way before.
And then at some point, when TensorFlow started becoming
popular, there were enough similarities
that he decided to create this interface
and put TensorFlow as a back end.
I believe that might still have been before he joined Google.
So we weren't really talking about that.
He decided on his own and thought
that was interesting and relevant to the community.
In fact, I didn't find out about him being at Google
until a few months after he was here.
He was working on some research ideas
and doing Keras on his nights and weekends project and stuff.
Oh, interesting.
So he wasn't like part of the TensorFlow.
He didn't join initially.
He joined research, and he was doing some amazing research.
He has some papers on that and research.
He's a great researcher as well.
And at some point, we realized, oh, he's doing this good stuff.
People seem to like the API, and he's right here.
So we talked to him, and he said, OK,
why don't I come over to your team
and work with you for a quarter, and let's
make that integration happen, and we talked to his manager,
and he said, sure, wait, quarter's fine.
And that quarter's been something like two years now.
So he's fully on this.
So Keras got integrated into TensorFlow in a deep way.
And now with TensorFlow 2.0, Keras
is kind of the recommended way for a beginner
to interact with TensorFlow, which
makes that initial sort of transfer learning
or the basic use cases, even for an enterprise,
super simple, right?
That's correct.
That's right.
So what was that decision like?
That seems like it's kind of a bold decision as well.
We did spend a lot of time thinking about that one.
We had a bunch of APIs some bit by us.
There was a parallel layers API that we were building.
And when we decided to do Keras in parallel,
so they were like, OK, two things that we are looking at.
And the first thing we was trying to do
is just have them look similar, like be
as integrated as possible, share all of that stuff.
There were also like three other APIs
that others had built over time, because we
didn't have a standard one.
But one of the messages that we kept hearing from the community,
OK, which one do we use?
And they kept saying, OK, here's a model in this one,
and here's a model in this one, which should I pick?
So that's sort of like, OK, we had to address that straight
on with 2.0.
The whole idea was we need to simplify.
We had to pick one.
Based on where we were, we were like, OK, let's
see what are the people like.
And Keras was clearly one that lots of people loved.
There were lots of great things about it.
So we settled on that.
Organically, that's kind of the best way to do it.
It was great.
It was surprising, nevertheless, to bring in and outside.
I mean, there was a feeling like Keras
might be almost like a competitor in a certain kind
of a two-tensor flow.
And in a sense, it became an empowering element
of tensor flow.
That's right.
Yeah, it's interesting how you can put two things together
which can align, right?
And in this case, I think Francois, the team,
and a bunch of us have chatted.
And I think we all want to see the same kind of things.
We all care about making it easier
for the huge set of developers out there.
And that makes a difference.
So Python has Guido Van Rossum, who
until recently held the position of benevolent
dictator for life.
Right?
So there's a huge successful open source project
like tensor flow.
Need one person who makes a final decision.
So you did a pretty successful tensor flow Dev Summit
just now, the last couple of days.
There's clearly a lot of different new features
being incorporated, an amazing ecosystem, so on.
How are those design decisions made?
Is there a BDFL in tensor flow?
And or is it more distributed and organic?
I think it's somewhat different, I would say.
I've always been involved in the key design directions.
But there are lots of things that are distributed
where their number of people, Martin Wick being one who
has really driven a lot of our open source stuff,
a lot of the APIs.
And there are a number of other people
who have been pushed and been responsible for different parts
of it.
We do have regular design reviews.
Over the last year, we've really spent a lot of time opening up
to the community and adding transparency.
We're setting more processes in place.
So RFCs, special interest groups, really grow that community
and scale that.
I think the kind of scale that ecosystem is in,
I don't think we could scale with having me
as the lone point of decision maker.
I got it.
So yeah, the growth of that ecosystem,
maybe you can talk about it a little bit.
First of all, when I started with Andre Karpathi
when he first did ComNet.js, the fact
that you can train in your own network
and the browser in JavaScript was incredible.
So now TensorFlow.js is really making
that a serious, a legit thing, a way
to operate, whether it's in the back end or the front end.
Then there's the TensorFlow Extended, like you mentioned.
There's TensorFlow Lite for mobile.
And all of it, as far as I can tell,
it's really converging towards being
able to save models in the same kind of way.
You can move around, you can train on the desktop,
and then move it to mobile, and so on.
That's right.
So there's that cohesiveness.
So can you maybe give me whatever
I missed, a bigger overview of the mission of the ecosystem
that's trying to be built, and where is it moving forward?
Yeah.
So in short, the way I like to think of this
is our goals to enable machine learning.
And in a couple of ways, one is we
have lots of exciting things going on in ML today.
We started with deep learning, but we now
support a bunch of other algorithms too.
So one is to, on the research side,
keep pushing on the state of the art.
How do we enable researchers to build
the next amazing thing?
So BERT came out recently.
It's great that people are able to do new kinds of research.
There are lots of amazing research
that happens across the world.
So that's one direction.
The other is, how do you take that
across all the people outside who want to take that research
and do some great things with it and integrate it
to build real products, to have a real impact on people?
And so if that's the other axes in some ways.
And at a high level, one way I think about it
is there are a crazy number of computer devices
across the world.
And we often used to think of ML and training and all of this
as, OK, something you do either in the workstation
or the data center or cloud.
But we see things running on the phones.
We see things running on really tiny chips.
And we had some demos at the developer summit.
And so the way I think about this ecosystem
is, how do we help get machine learning on every device that
has a compute capability?
And that continues to grow.
And so in some ways, this ecosystem
has looked at various aspects of that
and grown over time to cover more of those.
And we continue to push the boundaries.
In some areas, we've built more tooling and things
around that to help you.
I mean, the first tool we started was TensorBoard.
You want to learn just the training piece, the effects
for TensorFlow Extended to really do your entire ML
pipelines if you care about all that production stuff,
but then going to the edge, going to different kinds of things.
And it's not just us now.
We are a place where there are lots of libraries being built
on top.
So there are some for research, maybe things
like TensorFlow Agents or TensorFlow Probability that
started as research things or for researchers
for focusing on certain kinds of algorithms,
but they're also being deployed or reduced by production folks.
And some have come from within Google, just teams
across Google who wanted to do the build these things.
Others have come from just the community
because there are different pieces
that different parts of the community care about.
And I see our goal as enabling even that.
It's not, we cannot and won't build every single thing.
That just doesn't make sense.
But if we can enable others to build the things
that they care about, and there's
a broader community that cares about that,
and we can help encourage that, and that's great.
That really helps the entire ecosystem, not just those.
One of the big things about 2.0 that we're pushing on
is, OK, we have these so many different pieces, right?
How do we help make all of them work well together?
There are a few key pieces there that we're pushing on,
one being the core format in there
and how we share the models themselves through SAVE model
and what TensorFlow Hub and so on.
And a few of the pieces that we really put this together.
I was very skeptical that that's, when TensorFlow.js came out,
it didn't seem, or deep learning.js.
Yeah, that was the first.
It seemed like technically a very difficult project.
As a standalone, it's not as difficult.
But as a thing that integrates into the ecosystem,
it seems very difficult.
So I mean, there's a lot of aspects of this
you're making look easy.
But on the technical side, how many challenges
have to be overcome here?
A lot.
And still have to be overcome.
That's the question here, too.
There are lots of steps to it.
I think we've iterated over the last few years,
so there's a lot we've learned.
I, yeah, and often when things come together well,
things look easy, and that's exactly the point.
It should be easy for the end user.
But there are lots of things that go behind that.
If I think about still challenges ahead,
there are, you know, we have a lot more devices coming on board,
for example, from the hardware perspective.
How do we make it really easy for these vendors
to integrate with something like TensorFlow, right?
So there's a lot of compiler stuff
that others are working on.
There are things we can do in terms of our APIs
and so on that we can do.
As we, you know, TensorFlow started
as a very monolithic system.
And to some extent, it still is.
There are lots of tools around it,
but the core is still pretty large and monolithic.
One of the key challenges for us to scale that out
is how do we break that apart with clear interfaces.
It's, you know, in some ways, it's software engineering 101.
But for a system that's now four years old, I guess, or more,
and that's still rapidly evolving
and that we're not slowing down with,
it's hard to, you know, change and modify
and really break apart.
It's sort of like, as people say, right,
it's like changing the engine with a car running
or fixed benefits, that's exactly what we're trying to do.
So there's a challenge here because the downside
of so many people being excited about TensorFlow
and becoming to rely on it in many other applications
is that you're kind of responsible.
It's the technical debt.
You're responsible for previous versions
to some degree still working.
So when you're trying to innovate, I mean,
it's probably easier to just start from scratch
every few months.
Absolutely.
So do you feel the pain of that?
2.0 does break some back compatibility, but not too much.
It seems like the conversion is pretty straightforward.
And do you think that's still important,
given how quickly deep learning is changing?
Can you just, the things that you've learned,
can you just start over, or is there pressure to not?
It's a tricky balance.
So if it was just a researcher writing a paper
who a year later will not look at that code again,
sure, it doesn't matter.
There are a lot of production systems
that rely on TensorFlow, both at Google
and across the world.
And people worry about this.
I mean, these systems run for a long time.
So it is important to keep that compatibility and so on.
And yes, it does come with a huge cost.
We have to think about a lot of things
as we do new things and make new changes.
I think it's a trade-off, right?
If you can, you might slow certain kinds of things down.
But the overall value you're bringing because of that
is much bigger because it's not just
about breaking the person yesterday.
It's also about telling the person tomorrow that,
you know what, this is how we do things.
We're not going to break you when you come on board
because there are lots of new people who are also
going to come on board.
One way I like to think about this,
and I always push the team to think about as well,
when you want to do new things, you
want to start with a clean slate,
design with a clean slate in mind,
and then we'll figure out how to make sure all the other things
work.
And yes, we do make compromises occasionally.
But unless you're designed with the clean slate
and not worry about that, you'll never get to a good place.
That's brilliant.
So even if you are responsible in the idea stage,
when you're thinking of new, just put all that behind you.
That's really well put.
So I have to ask this because a lot of students, developers
ask me, how I feel about PyTorch versus TensorFlow.
So I've recently completely switched my research group
to TensorFlow.
I wish everybody would just use the same thing.
And TensorFlow is as close to that, I believe, as we have.
But do you enjoy competition?
So TensorFlow is leading in many ways, many dimensions
in terms of the ecosystem, in terms of the number of users,
momentum power, production level, so on.
But a lot of researchers are now also using PyTorch.
Do you enjoy that kind of competition,
or do you just ignore it and focus
on making TensorFlow the best that it can be?
So just like research or anything people are doing,
it's great to get different kinds of ideas.
And when we started with TensorFlow,
like I was saying earlier, it was very important for us
to also have production in mind.
We didn't want just research, right?
And that's why we chose certain things.
Now PyTorch came along and said, you know what?
I only care about research.
This is what I'm trying to do.
What's the best thing I can do for this?
And it started iterating and said, OK,
I don't need to worry about graphs.
Let me just run things.
I don't care if it's not as fast as it can be,
but let me just make this part easy.
And there are things you can learn from that, right?
They, again, had the benefit of seeing what had come before,
but also exploring certain different kinds of spaces.
And they had some good things there,
building on, say, things like Jainer and so on before that.
So competition is definitely interesting.
It made us, you know, this is an area
that we had thought about, like I said, very early on.
Over time, we had revisited this a couple of times.
Should we add this again?
At some point, we said, you know what?
Here's it seems like this can be done well.
So let's try it again.
And that's how we started pushing on eager execution.
How do we combine those two together?
Which has finally come very well together in 2.0,
but it took us a while to get all the things together
and so on.
So let me, I mean, ask, put another way.
I think eager execution is a really powerful thing
that was added.
Do you think it wouldn't have been, you know, Muhammad Ali
versus the Frazier, right?
Do you think it wouldn't have been added as quickly
if PyTorch wasn't there?
It might have taken longer.
No longer.
Yeah.
It was, I mean, we had tried some variants of that before.
So I'm sure it would have happened,
but it might have taken longer.
I'm grateful that TensorFlow is part of the way they did.
That's doing some incredible work last couple of years.
What other things that we didn't talk about?
Are you looking forward in 2.0 that comes to mind?
So we talked about some of the ecosystem stuff,
making it easily accessible to Keras, eager execution.
Is there other things that we miss?
Yeah.
So I would say one is just where 2.0 is and, you know,
with all the things that we've talked about.
I think as we think beyond that, there
are lots of other things that it enables us to do
and that we're excited about.
So what it's setting us up for, OK,
there are these really clean APIs.
We've cleaned up the surface for what the users want.
What it also allows us to do a whole bunch of stuff
behind the scenes once we are ready with 2.0.
So for example, in TensorFlow with graphs
and all the things you could do, you could always get
a lot of good performance if you spent the time to tune it.
And we've clearly shown that.
Lots of people do that.
With 2.0, with these APIs, where we can give you
a lot of performance just with whatever you do.
Because we see these, it's much cleaner.
We know most people are going to do things this way.
We can really optimize for that and get a lot of those things
out of the box.
And it really allows us, both for a single machine
and distributed and so on, to really explore
other spaces behind the scenes after 2.0
in the future versions as well.
So right now, the team is really excited about that.
That over time, I think we'll see that.
The other piece that I was talking about in terms of just
restructuring the monolithic thing into more pieces
and making it more modular, I think
that's going to be really important for a lot of the other
people in the ecosystem, other organizations and so on
that wanted to build things.
Can you elaborate a little bit what you mean by making TensorFlow
more ecosystem or modular?
So the way it's organized today is there's one,
there are lots of repositories in the TensorFlow
organization at GitHub.
The core one where we have TensorFlow,
it has the execution engine.
It has the key backends for CPUs and GPUs.
It has the work to do distributed stuff.
And all of these just work together
in a single library or binary.
There's no way to split them apart easily.
I mean, there are some interfaces, but they're not very clean.
In a perfect world, you would have clean interfaces
where, OK, I want to run it on my fancy cluster
with some custom networking.
Just implement this and do that.
I mean, we kind of support that, but it's hard for people today.
I think as we are starting to see more interesting things
in some of these spaces, having that clean separation
will really start to help.
And again, going to the large size of the ecosystem
and the different groups involved there,
enabling people to evolve and push on things more independently
just allows it to scale better.
And by people, you mean individual developers and organizations.
And organizations.
That's right.
So the hope is that everybody sort of major, I don't know,
Pepsi or something uses major corporations
go to TensorFlow to this kind of.
Yeah, if you look at enterprises like Pepsi or these,
I mean, all of them are already using TensorFlow.
They are not the ones that do the development
or changes in the core.
Some of them do, but a lot of them don't.
I mean, they touch small pieces.
There are lots of these, some of them being, let's say,
hardware vendors who are building their custom hardware
and they want their own pieces.
Or some of them being bigger companies, say IBM.
I mean, they're involved in some of our special interest
groups, and they see a lot of users who want certain things
and they want to optimize for that.
So folks like that often.
Autonomous vehicle companies, perhaps.
Exactly, yes.
So, yeah, like I mentioned, TensorFlow
has been downloaded 41 million times, 50,000 commits,
almost 10,000 pull requests, and 1,800 contributors.
So I'm not sure if you can explain it,
but what does it take to build a community like that?
What if, in retrospect, what do you think,
what is the critical thing that allowed for this growth
to happen, and how does that growth continue?
Yeah, yeah, that's an interesting question.
I wish I had all the answers there, I guess,
so you could replicate it.
I think there are a number of things
that need to come together, right?
One, just like any new thing, it is about there's
a sweet spot of timing, what's needed,
does it grow with what's needed.
In this case, for example, TensorFlow is not just
grown because of a good tool, it's
also grown with the growth of deep learning itself.
So those factors come into play.
Other than that, though, I think just hearing,
listening to the community, what they're doing, what they need,
being open to, like in terms of external contributions,
we've spent a lot of time in making sure
we can accept those contributions well,
we can help the contributors in adding those,
putting the right process in place,
getting the right kind of community,
welcoming them, and so on.
Like over the last year, we've really pushed on transparency.
That's important for an open source project.
People want to know where things are going,
and we're like, OK, here's a process where you can do that,
here are our season, so on.
So thinking through, there are lots of community aspects
that come into that you can really work on.
As a small project, it's maybe easy to do
because there's two developers, and you can do those.
As you grow, putting more of these processes in place,
thinking about the documentation,
thinking about what two developers
care about what kind of tools would they want to use,
all of these come into play, I think.
So one of the big things, I think,
that feeds the TensorFlow fire is people
building something on TensorFlow.
And implement a particular architecture
that does something cool and useful,
and they put that on GitHub.
And so it just feeds this growth.
Do you have a sense that with 2.0 and 1.0,
that there may be a little bit of a partitioning like there
is with Python 2 and 3, that there'll be a code base?
And in the older versions of TensorFlow,
they will not be as compatible easily?
Or are you pretty confident that this kind of conversion
is pretty natural and easy to do?
So we're definitely working hard to make that very easy to do.
There's lots of tooling that we talked about at the developer
summit this week.
And we'll continue to invest in that tooling.
It's when you think of these significant version changes,
that's always a risk.
And we are really pushing hard to make that transition
very, very smooth.
I think so at some level, people want
to move and they see the value in the new thing.
They don't want to move just because it's a new thing.
And some people do, but most people want a really good thing.
And I think over the next few months,
as people start to see the value,
we'll definitely see that shift happening.
So I'm pretty excited and confident that we
will see people moving.
As you said earlier, this field is also moving rapidly,
so that'll help because we can do more things.
And all the new things will clearly happen in 2.x,
so people will have lots of good reasons to move.
So what do you think TensorFlow 3.0 looks like?
Is there are things happening so crazily
that even at the end of this year
seems impossible to plan for?
Or is it possible to plan for the next five years?
I think it's tricky.
There are some things that we can expect in terms of, OK,
change.
Yes, change is going to happen.
Are there some things going to stick around
and some things not going to stick around?
I would say the basics of deep learning,
the convolutional models or the basic kind of things,
they'll probably be around in some form still in five years.
Will RL and GAN stay very likely based on where they are?
Will we have new things?
Probably, but those are hard to predict.
And directionally, some things that we
can see is in things that we're starting to do
with some of our projects right now is just
to point out combining eager execution and graphs
where we're starting to make it more like just your natural
programming language.
You're not trying to program something else.
Similarly, with Swift for TensorFlow,
we're taking that approach.
Can you do something round up?
So some of those ideas seem like, OK,
that's the right direction.
In five years, we expect to see more in that area.
Other things we don't know is, will hardware accelerators
be the same?
Will we be able to train with four bits instead of 32 bits?
And I think the TPU side of things is exploring.
I mean, TPU is already on version 3.
It seems that the evolution of TPU and TensorFlow
are sort of their co-evolving almost in terms
of both their learning from each other
and from the community and from the applications
where the biggest benefit is achieved.
That's right.
You've been trying to sort of with eager with Keras
to make TensorFlow as accessible and easy to use
as possible.
What do you think for beginners is the biggest thing
they struggle with?
Have you encountered that?
Or is basically what Keras is solving is that eager?
Like we talked about.
Yeah, for some of them, like you said, right,
the beginners want to just be able to take some image model.
They don't care if it's inception or resonant
or something else and do some training or transfer
learning on their kind of model.
Being able to make that easy is important.
So in some ways, if you do that by providing them
simple models with, say, in Hub or so on,
they don't care about what's inside that box,
but they want to be able to use it.
So we are pushing on, I think, different levels.
If you look at just a component that you get which
has the layers already smushed in,
the beginners probably just want that.
Then the next step is, OK, look at building
layers with Keras.
If you go out to research, then they
are probably writing custom layers themselves
or doing their own loops.
So there's a whole spectrum there.
And then providing the pre-trained models
seems to really decrease the time from you trying to start.
So you could basically, in a Colab notebook,
achieve what you need.
So I'm basically answering my own question,
because I think what TensorFlow delivered on recently
is trivial for beginners.
So I was just wondering if there was other pain points
you're trying to ease, but I'm not sure there would be.
No, those are probably the big ones.
I mean, I see high schoolers doing a whole bunch of things
now, which is pretty amazing.
It's both amazing and terrifying.
Yes.
In a sense that when they grow up,
some incredible ideas will be coming from them.
So there's certainly a technical aspect to your work,
but you also have a management aspect
to your role with TensorFlow, leading the project,
a large number of developers and people.
So what do you look for in a good team?
What do you think?
Google has been at the forefront of exploring
what it takes to build a good team.
And TensorFlow is one of the most cutting edge technologies
in the world.
So in this context, what do you think makes for a good team?
It's definitely something I think a fair bit about.
I think in terms of the team being
able to deliver something well, one of the things that's
important is a cohesion across the team.
So being able to execute together and doing things
that's not an end.
Like at this scale, an individual engineer
can only do so much.
There's a lot more that they can do together,
even though we have some amazing superstars across Google
and in the team.
But there's often the way I see it
is the product of what the team generates
is way larger than the whole or the individual put together.
And so how do we have all of them work together,
the culture of the team itself?
Hiring good people is important, but part of that
is it's not just that, OK, we hire a bunch of smart people
and throw them together and let them do things.
It's also people have to care about what they're building.
People have to be motivated for the right kind of things.
That's often an important factor.
And finally, how do you put that together
with a somewhat unified vision of where we want to go?
So are we all looking in the same direction
or just going all over?
And sometimes it's a mix.
Google's a very bottom-up organization in some sense.
Also research even more so.
And that's how we started.
But as we've become this larger product and ecosystem,
I think it's also important to combine that well with a mix
of, OK, here's the direction we want to go in.
There is exploration we'll do around that,
but let's keep staying in that direction, not just
all over the place.
And is there a way you monitor the health of the team?
Sort of like, is there a way you know you did a good job?
The team is good.
I mean, you're saying nice things,
but it's sometimes difficult to determine how aligned.
Because it's not binary.
It's there's tensions and complexities and so on.
And the other element of this is the mesh of superstars.
There's so much, even at Google, such a large percentage
of work is done by individual superstars too.
So there's a, and sometimes those superstars
could be against the dynamic of a team and those tensions.
I mean, I'm sure TensorFlow might be a little bit easier,
because the mission of the project is so beautiful.
You're at the cutting edge, so it's exciting.
But have you had struggle with that?
Has there been challenges?
There are always people challenges
in different kinds of ways.
That said, I think we've been both good
about getting people who care and have the same kind of culture.
And that's Google in general to a large extent.
But also, like you said, given that the project has had
so many exciting things to do, there's
been room for lots of people to do different kinds of things
and grow, which does make the problem a bit easier, I guess.
And it allows people, depending on what they're doing,
if there's room around them, then that's fine.
But yes, we do care about whether a superstar or not
that they need to work well with the team across Google.
That's interesting to hear.
So it's like superstar or not, the productivity broadly
is about the team.
Yeah.
I mean, they might add a lot of value,
but if they're supporting the team, then that's a problem.
So in hiring engineers, it's so interesting, right?
The hiring process, what do you look for?
How do you determine a good developer or a good member
of a team from just a few minutes or hours together?
Again, no magic answers, I'm sure.
Yeah.
And Google has a hiring process that we've refined
over the last 20 years, I guess, and that you've probably
heard and seen a lot about.
So we do work with the same hiring process,
and that's really helped.
For a mean particular, I would say,
in addition to the core technical skills,
what does matter is their motivation in what they want to do.
Because if that doesn't align well with where we want to go,
that's not going to lead to long-term success
for either them or the team.
And I think that becomes more important the more senior
the person is, but it's important at every level.
Like even the junior most engineer,
if they're not motivated to do well at what they're trying to do,
however smart they are, it's going to be hard for them to succeed.
Does the Google hiring process touch on that passion?
So trying to determine, because I think, as far as I understand,
maybe you can speak to it, that the Google hiring process helps.
The initial determines the skill set there,
is your puzzle solving ability, problem solving ability good.
But I'm not sure, but it seems that determining
whether the person is fire inside them.
That burns to do anything, really doesn't really matter.
It's just some cool stuff, I'm going to do it.
Is that something that ultimately ends up
when they have a conversation with you,
or once it gets closer to the team?
So one of the things we do have as part of the process
is just a culture fit, like part of the interview process itself,
in addition to just the technical skills.
And each engineer or whoever the interviewer is,
is supposed to rate the person on the culture,
and the culture fit with Google, and so on.
So that is definitely part of the process.
Now, there are various kinds of projects
and different kinds of things.
So there might be variants in the kind of culture
you want there, and so on.
And yes, that does vary.
So for example, TensorFlow has always
been a fast-moving project, and we
want people who are comfortable with that.
But at the same time now, for example,
we are at a place where we are also a very full-fledged product,
and we want to make sure things that work really, really work.
You can't cut corners all the time.
So that balancing that out and finding the people who
are the right fit for those is important.
And I think those kind of things do vary a bit
across projects and teams and product areas across Google,
and so you'll see some differences there
in the final checklist.
But a lot of the core culture, it
comes along with just engineering, excellence, and so on.
What is the hardest part of your job?
I'll take your pick, I guess.
It's fun, I would say.
Hard, yes.
I mean, lots of things at different times.
I think that does vary.
So let me clarify that difficult things are fun
when you solve them, right?
It's fun in that sense.
I think the key to a successful thing across the board,
and in this case, it's a large ecosystem now,
but even a small product, is striking that fine balance
across different aspects of it.
Sometimes it's how fast you go versus how perfect it is.
Sometimes it's how do you involve this huge community?
Who do you involve?
Or do you decide, OK, now is not a good time
to involve them because it's not the right fit.
Sometimes it's saying no to certain kinds of things.
Those are often the hard decisions.
Some of them you make quickly because you
don't have the time.
Some of them you get time to think about them,
but they're always hard.
So both choices are pretty good, those decisions.
What about deadlines?
Is this defined TensorFlow to be driven by deadlines
to a degree that a product might?
Or is there still a balance to where it's less deadline?
You had the Dev Summit, they came together incredibly.
It looked like there's a lot of moving pieces and so on.
So did that deadline make people rise to the occasion,
releasing TensorFlow 2.0 Alpha?
I'm sure that was done last minute as well.
I mean, up to the last point.
Again, it's one of those things that you
need to strike the good balance.
There's some value that deadlines bring
that does bring a sense of urgency
to get the right things together.
Instead of getting the perfect thing out,
you need something that's good and works well.
And the team definitely did a great job
in putting that together.
So it was very amazed and excited by everything,
how that came together.
That said, across the year, we try
not to put artificial deadlines.
We focus on key things that are important,
figure out how much of it's important.
And we are developing in the open, internally and externally,
everything's available to everybody.
So you can pick and look at where things are.
We do releases at a regular cadence.
So fine if something doesn't necessarily
end up with this month, it'll end up
in the next release in the month or two.
And that's OK.
But we want to keep moving as fast as we can
in these different areas.
Because we can iterate and improve on things.
Sometimes it's OK to put things out that aren't fully ready.
If you make sure it's clear that, OK, this is experimental.
But it's out there if you want to try and give feedback.
That's very, very useful.
I think that quick cycle and quick iteration is important.
That's what we often focus on rather than here's
a deadline where you get everything else.
Is 2.0, is there pressure to make that stable?
Or like, for example, WordPress 5.0 just came out.
And there was no pressure to.
It was a lot of build updates that delivered way too late.
And they said, OK, well, we're going
to release a lot of updates really quickly to improve it.
Do you see TensorFlow 2.0 in that same kind of way?
Or is there this pressure to once it's 2.0,
once you get to the release candidate
and then you get to the final, that that's
going to be the stable thing?
So it's going to be stable in just like when
NodeX was where every API that's there
is going to remain in work.
It doesn't mean we can't change things under the covers.
It doesn't mean we can't add things.
So there's still a lot more for us to do.
And we continue to have more releases.
So in that sense, I don't think we'd
be done in like two months when we release this.
I don't know if you can say, but is there, you know,
there's not external deadlines for TensorFlow 2.0,
but is there internal deadlines, artificial or otherwise,
that you're trying to set for yourself?
Or is it whenever it's ready?
So we want it to be a great product, right?
And that's a big, important piece for us.
TensorFlow is already out there.
We have 41 million downloads for one NodeX.
So it's not like we have to have this.
Yeah, exactly.
So it's not like a lot of the features
that we've really polishing and putting them together
are there.
We don't have to rush that just because.
So in that sense, we want to get it right
and really focus on that.
That said, we have said that we are
looking to get this out in the next few months,
in the next quarter.
And as far as possible, we'll definitely
try to make that happen.
Yeah, my favorite line was, spring is a relative concept.
I love it.
Yes.
Spoken like a true developer.
So something I'm really interested in,
and your previous line of work is, before TensorFlow,
you led a team at Google on search ads.
I think this is a very interesting topic on every level,
on a technical level.
Because if their best ads connect people
to the things they want and need,
and that they're worse, they're just
these things that annoy the heck out of you
to the point of ruining the entire user experience of whatever
you're actually doing.
So they have a bad rep, I guess.
And so on the other end, so that this connecting users
to the thing they need and want is a beautiful opportunity
for machine learning to shine.
Like huge amounts of data that's personalized,
and you map to the thing they actually won't get annoyed.
So what have you learned from this Google that's
leading the world in this aspect?
What have you learned from that experience?
And what do you think is the future of ads?
Take you back to that point.
Yes, it's been a while.
But I totally agree with what you said.
I think the search ads, the way it was always looked at,
and I believe it still is, is it's an extension of what
search is trying to do.
The goal is to make the information
and make the world's information accessible.
With ads, it's not just information,
but it may be products or other things
that people care about.
And so it's really important for them
to align with what the users need.
And in search ads, there's a minimum quality level
before that ad would be shown.
If we don't have an ad that hits that quality bar,
it will not be shown, even if we have it.
And OK, maybe we lose some money there.
That's fine.
That is really, really important.
And I think that that is something
I really liked about being there.
Advertising is a key part.
As a model, it's been around for ages, right?
It's not a new model.
It's been adapted to the web and became
a core part of search in many other search engines
across the world.
I do hope, like I said, there are aspects of ads
that are annoying.
And I go to a website.
And if it just keeps popping an ad in my face,
not to let me read, that's going to be annoying, clearly.
So I hope we can strike that balance between showing
a good ad where it's valuable to the user
and provides the monetization to the service.
And this might be search, this might be a website,
all of these, they do need the monetization for them
to provide that service.
But if it's done in that good balance between showing
just some random stuff that's distracting
versus showing something that's actually valuable.
So do you see it moving forward as to continue being
a model that funds businesses like Google,
that's a significant revenue stream?
Because that's one of the most exciting things,
but also limiting things on the internet
is nobody wants to pay for anything.
And advertisements, again, coupled at their best,
are actually really useful and not annoying.
Do you see that continuing and growing and improving?
Or is there more Netflix-type models
where you have to start to pay for content?
I think it's a mix.
I think it's going to take a long while for everything
to be paid on the internet, if at all.
Probably not.
I mean, I think there's always going
to be things that are monetized with things like ads.
But over the last few years, I would say
we've definitely seen that transition
towards more paid services across the web
and people are willing to pay for them,
because they do see the value.
I mean, Netflix is a great example.
I mean, we have YouTube doing things.
People pay for the apps they buy.
More people I find are willing to pay for newspaper content,
for the good news websites across the web.
That was, in the case, even a few years ago, I would say.
And I just see that change in myself as well
and just lots of people around me.
So definitely hopeful that we'll transition to that mix model
where maybe you get to try something out for free,
maybe with ads, but then there's a more clear revenue model
that sort of helps go beyond that.
So speaking of revenue, how is it
that a person can use the TPU in a Google Colab for free?
So what's the?
I guess the question is, what's the future of TensorFlow
in terms of empowering, say, a class of 300 students
and amassed by MIT?
What is going to be the future of them
being able to do their homework in TensorFlow?
Where are they going to train these networks?
What's that future look like with TPUs,
with cloud services, and so on?
I think a number of things there,
I mean, any TensorFlow open source,
you can run it wherever.
You can run it on your desktop.
And your desktops always keep getting more powerful,
so maybe you can do more.
My phone is like, I don't know how many times more powerful
than my first desktop.
You'll probably train it on your phone, though.
Yeah, that's true.
Right, so in that sense, the power
you have in your hands is a lot more.
Clouds are actually very interesting
from, say, students or courses perspective,
because they make it very easy to get started.
I mean, Colab, the great thing about it
is go to a website, and it just works.
No installation needed, nothing to do.
You're just there, and things are working.
That's really the power of cloud as well.
And so I do expect that to grow.
Again, Colab is a free service.
It's great to get started, to play with things,
to explore things.
That said, with free, you can only get so much.
Yeah.
So just like we were talking about free versus paid.
Yeah, there are services you can pay for and get a lot more.
Great.
So if I'm a complete beginner interested in machine
learning and TensorFlow, what should I do?
Probably start with going to a website and playing there.
Just go to TensorFlow.org and start clicking on things.
Yep.
Check out tutorials and guides.
There's stuff you can just click there and go to a Colab
and do things.
No installation needed.
You can get started right there.
OK, awesome.
Roger, thank you so much for talking today.
Thank you, Lex.
Fun.
It's great.