Thanks for joining us.

My name's Josh Gordon, and I'm very, very

happy to be here today with Jeremy Howard, who

is the co-founder of a course I'm sure many of you

have heard of called fast.ai, and he's faculty at USF.

So, Jeremy, thank you so much for joining us today.

JEREMY HOWARD: Oh, it's a pleasure.

JOSH GORDON: And, today, I'd love to talk about a new course

that you're working on for Swift.

So I am totally new to Swift.

So could you tell me a little bit maybe first about fast.ai

in general, and then maybe a little bit

about your plans for Swift?

JEREMY HOWARD: Yeah, absolutely.

So fast.ai, as you know, is a self-funded nonprofit.

We do a few things.

We do research, we do teaching, we do software development,

and we do community development.

So the research we do is all about how

do we make world class deep learning more

accessible to regular people.

So we do lots of algorithm development and testing

and curation.

That research ends up in courses that we

provide online through fast.ai.

There are a couple of main ones people know.

One is Practical Deep Learning for Coders,

which is taking that research and saying, how can that

be used to make somebody with a year of coding background,

but no particular math background,

into an effective deep learning practitioner.

And then there's a more advanced course,

which is about research level deep learning.

And then we have our community online, which

we're super excited about.

But then there's a really interesting thing we do,

which is to help all that stuff become easier,

we have a software library called

fast.ai, which currently sits on top of PyTorch and Python.

But we're very interested by what Chris Lattner and his team

are doing with Swift now and Swift for TensorFlow.

We think there are big opportunities to help all four

of those areas by embracing Swift for TensorFlow

as well as PyTorch and Python.

JOSH GORDON: OK, so moving on to Swift, so traditionally,

almost all of machine learning development

is done in Python, and of course languages

like R and now JavaScript.

So why Swift?

Could you tell me a little bit about it?

JEREMY HOWARD: Sure.

So I'm a real programming language nerd.

I like studying programming languages,

and I've been involved in lots of different programming

languages.

I actually was the chair of the working group that

tried to bring scientific programming

capabilities to Perl 6 back in the late '90s, early 2000s.

And my programming language nerdship

is driven by a deep discontent with any programming language

when it comes to numerical programming.

I did a lot of work in C++ in the early 2000s and got really

turned off by the long compile times and the complexity

of expression templates.

I quite like C# and F#, but there's always been--

cross platform took a long time to come,

and then cross platform has come with some performance

regressions in some way.

So I've always been looking for a really good programming

language for numerical computing.

Python isn't that language.

Python is a great glue language to sit on top of other

languages, so like NumPy is basically C.

And TensorFlow Python is just wrapping up C++,

which sits on top of Cuda C libraries and Icon and the C++

library.

And, in the end, it's dissatisfying for a student

and for a teacher, because at some point,

I have to say, this is the point at which we can stop finding

out what's going on, because beneath here,

it's kind of assembly code, or machine code,

or compiled stuff.

It's very frustrating, as a researcher,

like an NLP, where you keep on wanting

to do more stuff with RNNs.

And we keep on hitting the point where it's like, oh,

we can't really implement this different RNN cell,

because if we do, then it's not going to use cuDNN anymore,

and the performance is going to fall apart.

Swift offers a way past all this.

It lets us write on a language that's

a thin layer over the amazing LLVM infrastructure, compiler

infrastructure, which can get lots

of different bits of an algorithm pulled together,

optimize the whole thing, for CPU and for GPU.

And, at the same time, because it

has the view of everything going on from top to bottom,

it can tell me all the times I've screwed up.

So, hey, you thought that this dimension was batch size,

but over here, you use this other dimension as batch size,

and I tried to multiply those two tensors.

At compile time, I could tell the shapes

aren't going to match.

JOSH GORDON: It's huge.

JEREMY HOWARD: Yeah.

So I'm very excited to learn how it's

going to make me more productive as a programmer.

It's going to allow me to turn my research ideas into code.

At the moment, I keep butting up against things

where I just can't do it.

And it's going to let us teach things

deeper, which is what I'm always trying to do.

JOSH GORDON: Right.

So how does this relate to fast.ai?

JEREMY HOWARD: So then, the next course, which

is going to be in a couple of weeks

at the University of San Francisco,

we'll be recording it, and then that recording will

become our next MOOC in June.

It's going to be a seven part course.

It's going to be called Deeper Deep Learning.

And Deeper Deep Learning is going

to be all about how do we take a practical practitioner

of deep learning and turn them into somebody who

can go further, like do cutting edge research,

get things into production, make things run fast.

So of the seven lessons in the next course,

the last two are going to be about Swift for TensorFlow.

And it's going to be really cool, because actually, I'm

going to be co-teaching it with Chris Lattner.

JOSH GORDON: And so for people that don't know,

so Chris Lattner is the inventor of Swift.

JEREMY HOWARD: The inventor of Swift.

So, for me, I'm kind of having a little flashy moment now,

where I'm like, oh my god, I can't

believe I'll be standing next to Chris Lattner, the Chris

Lattner.

So I'm very excited about that.

And I'm very excited about what it

means in terms of what we can say

to these students who are at the very

advanced level at this point.

They've done 100 plus hours of study,

and they're competent coders.

And we will be able to say, here's

a deeply well-designed, thoughtful, fast, brilliant

language, which has had very, very little numerical computing

so far, but has the might now of Google, and Chris Lattner,

and a brilliant team behind it.

The world's your oyster.

Almost nothing's been implemented.

So the class projects are going to be things like, create

this layer that hasn't been implemented yet,

or implement this architecture, or be

the first person that's run this model end to end on ImageNet.

So the class projects-- and there'll also

be things like, help us create the fast.ai library for Swift

for TensorFlow.

JOSH GORDON: Yes.

JEREMY HOWARD: So I can tell you the fast.ai library for Swift

for TensorFlow even has a code name now.

JOSH GORDON: Awesome.

JEREMY HOWARD: The code name is harebrain.

JOSH GORDON: OK.

JEREMY HOWARD: Harebrain, because it's a crazy idea.

It's this amazingly crazy idea, which

is, we can take something that's so early,

but it's going to be such potential,

and actually start writing the dev UX layer from the start.

JOSH GORDON: You're right.

So this is all new territory.

JEREMY HOWARD: So it's this crazy idea.

It's super exciting.

But also hare, because hares are fast.

And Swift's fast, and fast.ai is fast, right.

JOSH GORDON: It sounds cool.

JEREMY HOWARD: And then brain, because we're working

with Google Brain on this.

So that's going to be our little internal code name for this.

So anybody who's getting involved in the next course,

through the MOOC or in person, can actually

help contribute to this codename harebrain library

from the very earliest days.

And so for a language nerd like me, it's so exciting.

JOSH GORDON: Awesome.

And what's the best way for people

to find out about the new course and follow it, keep updated,

sign up?

JEREMY HOWARD: Yeah, so if you want

to sign up for the in-person course, do it right now.

So just go to the University of San Francisco Data Institute.

But that's in San Francisco, so, obviously, a lot of people

won't be able to make it in person.

Otherwise, keep an eye on fast.ai in June,

where we'll be launching that course as a MOOC.

And yeah, at that point, it's still

going to be super early days for everything.

So there'll be lots of opportunities

for people to become part of what

I think is going to be a very, very, very

impactful project for scientific programming,

and for deep learning, and for differentiable computing

more generally, being Swift for TensorFlow.

JOSH GORDON: I'm really looking forward to it.

I have one last question for you, too.

JEREMY HOWARD: OK, hit me.

JOSH GORDON: So will Swift be primarily

for researchers, novices?

Long-term, how do you see the evolution of Swift

for deep learning?

JEREMY HOWARD: Because it's so versatile,

I plan to do a lot of research in Swift myself,

to get past that boundary I described of places where

you can't go with Python as a researcher.

I also expect it to be great in production,

because it's going to allow you to take the stuff that you

wrote at prototyping time and have something that's

super fast straight away, particularly because, for most

people--

maybe not most Googlers, but most normal people--

production inference means CPU, not GPU.

And Swift will be a really great option for that

because we're going to get all that compilation

niceness of LLVM to make that CPU code super fast.

JOSH GORDON: Awesome.

So thank you so much for joining us.

And thank you, everyone.

I learned a lot from this talk, and I'm really

looking forward to your course.

JEREMY HOWARD: Cheers.

JOSH GORDON: OK.

Thanks very much.

[MUSIC PLAYING]