The following is a conversation with Chris Latner,
his second time in the podcast.
He's one of the most brilliant engineers
in modern computing, having created
LLVM compiler infrastructure project,
the Clang compiler, the Swift programming language,
a lot of key contributions to TensorFlow and TPUs
as part of Google.
He served as vice president of autopilot software at Tesla,
was a software innovator and leader at Apple,
and now is at SyFive as senior vice president
of platform engineering,
looking to revolutionize chip design
to make it faster, better, and cheaper.
Quick mention of each sponsor,
followed by some thoughts related to the episode.
First sponsor is Blinkist,
an app that summarizes key ideas from thousands of books.
I use it almost every day to learn new things
or to pick which books I want to read or listen to next.
Second is Neuro,
the maker of functional sugar-free gum and mints
that I use to supercharge my mind
with caffeine, Althianine, and B vitamins.
Third is Masterclass,
online courses from the best people in the world
on each of the topics covered from rockets
to game design, to poker, to writing, and to guitar.
And finally, Cash App,
the app I use to send money to friends for food,
drinks, and unfortunately lost bets.
Please check out the sponsors in the description
to get a discount and to support this podcast.
As a side note, let me say that Chris
has been an inspiration to me on a human level
because he is so damn good as an engineer
and leader of engineers and yet he's able to stay humble,
especially humble enough to hear the voices
of disagreement and to learn from them.
He was supportive of me and this podcast from the early days
and for that, I'm forever grateful.
To be honest, most of my life,
no one really believed that I would amount to much.
So when another human being looks at me,
it makes me feel like I might be someone special,
it can be truly inspiring.
That's a lesson for educators.
The weird kid in the corner with a dream
is someone who might need your love and support
in order for that dream to flourish.
If you enjoy this thing, subscribe on YouTube,
review it with five stars on Apple Podcast,
follow on Spotify, support on Patreon,
or connect with me on Twitter at Lex Freedman.
And now, here's my conversation with Chris Latner.
What are the strongest qualities of Steve Jobs, Elon Musk,
and the great and powerful Jeff Dean
since you've gotten the chance to work with each?
You're starting with an easy question there.
These are three very different people.
I guess you could do maybe a pair-wise comparison
between them instead of a group comparison.
So if you look at Steve Jobs and Elon,
I worked a lot more with Elon than I did with Steve.
They have a lot of commonality.
They're both visionary in their own way.
They're both very demanding in their own way.
My sense is Steve is much more human factor focused
where Elon is more technology focused.
What does human factor mean?
Steve's trying to build things that feel good,
that people love, that affect people's lives, how they live.
He's looking into the future a little bit
in terms of what people want.
Where I think that Elon focuses more on
learning how exponentials work and predicting
the development of those.
Steve worked with a lot of engineers.
That was one of the things that reading the biography
and how can a designer essentially talk to engineers
and get their respect?
I think so.
I did not work very closely with Steve.
I'm not an expert at all.
My sense is that he pushed people really hard,
but then when he got an explanation that made sense to him,
then he would let go.
He did actually have a lot of respect for engineering.
But he also knew when to push.
And when you can read people well,
you can know when they're holding back
and when you can get a little bit more out of them.
And I think he was very good at that.
I mean, if you compare the other folks,
so Jeff Dean, right?
Jeff Dean's an amazing guy.
He's super smart as are the other guys.
Jeff is a really, really, really nice guy.
Well-meaning, he's a classic Googler.
He wants people to be happy.
He combines it with brilliance.
So he can pull people together in a really great way.
He's definitely not a CEO type.
I don't think he would even want to be that.
He's the deal if he still programs.
Oh, yeah, he definitely programs.
Jeff is an amazing engineer today, right?
And that has never changed.
So it's really hard to compare Jeff to either of those two.
He, I think that Jeff leads through technology
and building it himself and then pulling people in
and inspiring them.
And so I think that that's one of the amazing things about Jeff.
But each of these people, you know,
with their pros and cons,
all are really inspirational
and have achieved amazing things.
So it's been a, it's been,
I've been very fortunate to get to work with these guys.
For yourself, you've led large teams.
You've done so many incredible,
difficult technical challenges.
Is there something you've picked up from them
about how to lead?
Yeah, well, so I mean, I think leadership is really hard.
It really depends on what you're looking for there.
I think you really need to know what you're talking about.
So being grounded on the product,
on the technology, on the business,
on the mission is really important.
Being understanding what people are looking for,
why they're there.
One of the most amazing things about Tesla
is the unifying vision, right?
People are there because they believe in clean energy
and electrification, all these kinds of things.
The other is to understand what really motivates people,
how to get the best people,
how to build a plan that actually can be executed, right?
There's so many different aspects of leadership
and it really depends on the time, the place, the problems,
you know, there's a lot of issues
that don't need to be solved.
And so if you focus on the right things
and prioritize well, that can really help move things.
Two interesting things you mentioned.
One is you really have to know what you're talking about.
How you've worked on a lot of very challenging technical
things.
So I kind of assume you were born technically savvy,
but assuming that's not the case.
How did you develop technical expertise?
Like even at Google, you worked on,
I don't know how many projects,
but really challenging, very varied.
Compilers, TPUs, hardware, cloud stuff,
a bunch of different things.
The thing that I've become comfortable
is I've more comfortable with as I've gained experience
is being okay with not knowing.
And so a major part of leadership is actually,
it's not about having the right answer,
it's about getting the right answer.
And so if you're working in a team of amazing people, right?
And many of these places, many of these companies
all have amazing people.
It's the question of how do you get people together?
How do you build trust?
How do you get people to open up?
How do you get people to be vulnerable sometimes
with an idea that maybe isn't good enough,
but it's the start of something beautiful?
How do you provide an environment
where you're not just like top down,
don't do the thing that I tell you to do, right?
But you're encouraging people to be part of the solution
and providing a safe space
where if you're not doing the right thing,
they're willing to tell you about it, right?
So you're asking dumb questions.
Yeah, dumb questions are my specialty, yeah.
Well, so I've been in the harbor room recently
and I don't know much at all about how chips are designed.
I know a lot about using them.
I know some of the principles
and the ars technical level of this,
but it turns out that if you ask a lot of dumb questions,
you get smarter really, really quick.
And when you're surrounded by people that want to teach
and learn themselves, it can be a beautiful thing.
So let's talk about programming languages if it's okay.
Sure, sure.
At the highest absurd philosophical level,
because I- Don't get romantic on me, Lex.
I will forever get romantic and torture you, I apologize.
Why do programming languages even matter?
Okay, well, thank you very much.
So you're saying why should you care
about any one programming language
or why do we care about programming computers or?
No, why do we care about programming language design,
creating effective programming languages,
choosing a, you know, one program languages
versus another program language,
why we keep struggling and improving
through the evolution of these programming languages?
Sure, sure, sure.
Okay, so I mean, I think you have to come back
to what are we trying to do here, right?
So we have these beasts called computers
that are very good at specific kinds of things.
And we think it's useful to have them do it for us, right?
Now, you have this question of how best to express that
because you have a human brain still
that has an idea in its head
and you want to achieve something, right?
So, well, there's lots of ways of doing this.
You can go directly to the machine
and speak assembly language.
And then you can express directly
what the computer understands, that's fine.
You can then have higher and higher
and higher levels of abstraction
up until machine learning.
And you're designing a neural net to do the work for you.
The question is where, where along this way
do you want to stop
and what benefits do you get out of doing so?
And so programming languages in general,
you have C, you have Fortran and Java
and Ada, Pascal, Swift, you have lots of different things.
They all have different trade-offs
and they're tackling different parts of the problems.
Now, one of the things that most programming languages do
is they're trying to make it
so that you have pretty basic things
like portability across different hardware.
So you've got, I'm gonna run on an Intel PC,
I'm gonna run on a RISC-5 PC,
I'm gonna run on an ARM phone or something like that, fine.
I wanna write one program and have it portable.
And this is something the assembly doesn't do.
Now, when you start looking
at the space of programming languages,
this is where I think it's fun
because programming languages all have trade-offs
and most people will walk up to them
and they look at the surface level of syntax
and say, oh, I like curly braces or I like tabs
or I like, you know, semicolons or not or whatever, right?
Subject, fairly subjective, very shallow things.
But programming languages, when done right,
can actually be very powerful.
And the benefit they bring is expression, okay?
And if you look at programming languages,
there's really kind of two different levels to them.
One is the down-in-the-dirt nuts and bolts
of how do you get the computer to be efficient,
stuff like that, how they work,
type systems, compiler stuff, things like that.
The other is the UI.
And the UI for programming language
is really a design problem
and a lot of people don't think about it that way.
And the UI, you mean all that stuff with the braces
and the actual...
Yeah, all that stuff's the UI.
And what it is, and UI means user interface.
And so what's really going on
is it's the interface between the guts and the human.
And humans are hard, right?
Humans have feelings, they have things they like,
they have things they don't like.
And a lot of people treat programming languages
as though humans are just kind of abstract creatures
that cannot be predicted.
But it turns out that actually there are,
there is better and worse.
Like people can tell when a programming language is good
or when it was an accident, right?
And one of the things with Swift in particular
is that a tremendous amount of time
by a tremendous number of people
have been put into really polishing and making it feel good.
But it also has really good nuts and bolts underneath it.
You said that Swift makes a lot of people feel good.
How do you get to that point?
So how do you predict that tens of thousands,
hundreds of thousands of people are going to enjoy
using this user experience of this programming language?
Well, you can look at it in terms of better and worse, right?
So if you have to write lots of boilerplate
or something like that, you will feel unproductive.
And so that's a bad thing.
You can look at it in terms of safety.
If like C, for example,
is what's called a memory unsafe language.
And so you get dangling pointers
and you get all these kinds of bugs
that then you have spent tons of time debugging.
It's a real pain in the butt and you feel unproductive.
And so by subtracting these things from the experience,
you get, you know, happier people.
But again, keep interrupting.
I'm sorry.
It's so hard to deal with.
If you look at the people that are most productive
on Stack Overflow, they have a set of priorities
that may not always correlate perfectly
with the experience of the majority of users.
You know, if you look at the most upvoted,
quote unquote, correct answer on Stack Overflow,
it's usually really sort of prioritized
as like safe code, proper code, stable code,
you know, that kind of stuff.
As opposed to like, if I want to use go-to statements
in my basic, right, I want to use go-to statements.
What if 99% of people want to use go-to statements?
So you use completely improper, you know, unsafe syntax.
I don't think that people actually,
like if you boil it down and you get below the surface level,
people don't actually care about go-to's
or if statements or things like this.
They care about achieving a goal.
Yeah.
And so the real question is,
I want to set up a web server and I want to do a thing
and whatever, like how quickly can I achieve that?
Right.
And so from a programming language perspective,
there's really two things that matter there.
One is what libraries exist.
And then how quickly can you put it together?
And what are the tools around that look like, right?
And when you want to build a library
that's missing, what do you do?
Okay, now this is where you see huge divergence
in the force between worlds, okay?
And so you look at Python, for example,
Python is really good at assembling things,
but it's not so great at building all the libraries.
And so you get, because of performance reasons,
other things like this,
as you get Python layered on top of C, for example.
And that means that doing certain kinds of things,
well, it doesn't really make sense to do in Python.
Instead, you do it in C and then you wrap it
and then you have, you're living in two worlds
and two worlds never is really great
because tooling and the debugger doesn't work right
and like all these kinds of things.
Can you clarify a little bit what you mean by
Python is not good at building libraries,
meaning it doesn't make it conducive?
Certain kinds of libraries.
No, but it's just the actual meaning of the sentence.
Meaning like it's not conducive to developers
to come in and add libraries,
or is it the duality of the,
it's a dance between Python and C and you can never.
Well, so Python's amazing.
Python's a great language.
I do not mean to say that Python is bad for libraries.
What I meant to say is there are libraries
that Python's really good at that you can write in Python,
but there are other things
like if you wanna build a machine learning framework,
you're not gonna build a machine learning framework
in Python because of performance, for example,
or you want GPU acceleration or things like this.
Instead, what you do is you write a bunch
of C or C++ code or something like that,
and then you talk to it from Python, right?
And so this is because of decisions
that were made in the Python design,
and those decisions have other counterbalancing forces,
but the trick when you start looking at this
from a programming language perspective,
you start to say, okay, cool,
how do I build this catalog of libraries
that are really powerful?
And how do I make it so that then they can be assembled
into ways that feel good
and they generally work the first time?
Because when you're talking about building a thing,
you have to include the debugging, the fixing,
the turnaround cycle, the development cycle,
all that kind of stuff into the process
of building the thing.
It's not just about pounding out the code.
And so this is where things like catching bugs
at compile time is valuable, for example.
But if you dive into the details in this,
Swift, for example, has certain things like value semantics,
which is this fancy way of saying that
when you treat a variable like a value,
it acts like a mathematical object would.
Okay, so you have used PyTorch a little bit.
In PyTorch, you have tensors.
Tensors are n-dimensional grid of numbers.
Very simple.
You can do plus and other operators on them.
It's all totally fine.
But why do you need to clone a tensor sometimes?
Have you ever run into that?
Yeah.
Okay, and so why is that?
Why do you need to clone a tensor?
It's the usual object thing that's in Python.
So in Python, and just like with Java
and many other languages, this isn't unique to Python.
In Python, it has a thing called reference semantics,
which is the nerdy way of explaining this.
And what that means is you actually have a pointer
do a thing instead of the thing.
Now, this is due to a bunch of implementation details
that you don't want to go into.
But in Swift, you have this thing called value semantics.
And so when you have a tensor in Swift, it is a value.
If you copy it, it looks like you have a unique copy.
And if you go change one of those copies,
then it doesn't update the other one
because you just made a copy of this thing, right?
So that's like highly error prone
in at least computer science, math-centric disciplines
about Python, that like the thing you would expect
to behave like math.
Like math, it doesn't behave like math.
And in fact, quietly it doesn't behave like math
and then can ruin the entirety of your math.
Exactly.
Well, and then it puts you into debugging land again.
Yeah.
Right, now you just want to get something done
and you're like, wait a second.
Where do I need to put clone?
In what level of the stack, which is very complicated,
which I thought I was reusing somebody's library
and now I need to understand it
to know where to clone a thing, right?
And hard to debug, by the way.
Exactly, right.
And so this is where programming languages really matter, right?
So in Swift, having values semantics,
so that both you get the benefit of math,
working like math, right?
But also the efficiency that comes with certain advantages
there, certain implementation details there,
really benefit you as a programmer, right?
And so-
You clarify the values semantics, like how do you know
that a thing should be treated like a value?
Yeah, so Swift has a pretty strong culture
and good language support for defining values.
And so if you have an array,
so tensors are one example that the machine learning folks
are very used to, just think about arrays, same thing,
where you have an array, you create an array,
you put two or three or four things into it,
and then you pass it off to another function.
What happens if that function adds some more things to it?
Well, you'll see it on the side that you pass it in, right?
This is called reference semantics.
Now, what if you pass an array off to a function?
It scrolls it away in some dictionary
or some other data structure somewhere, right?
Well, it thought that you just handed it that array,
then you return back and that reference to that array
still exists in the caller,
and they go and put more stuff in it, right?
The person you handed it off to may have thought
they had the only reference to that,
and so they didn't know that this was gonna change
underneath the covers.
And so this is where you end up having to do clone.
So like, I was passed a thing,
I'm not sure if I have the only version of it,
so now I have to clone it.
So what value semantics does is it allows you to say,
hey, I have a, so in Swift, it defaults to value semantics.
For most of it.
Also defaults to value semantics,
and then because most things end up being like this,
then it makes sense for that to be the default.
And one of the important things about that is that arrays
and dictionaries and all these other collections
that are aggregations of other things
also have value semantics.
And so when you pass this around
to different parts of your program,
you don't have to do these defensive copies.
And so this is great for two sides, right?
It's great because you define away the bug,
which is a big deal for productivity,
the number one thing most people care about,
but it's also good for performance.
Because when you're doing a clone,
so you pass the array down to the thing,
it was like, I don't know if anybody else has it,
I have to clone it.
Well, you just did a copy of a bunch of data,
it could be big.
And then it could be that the thing that called you
is not keeping track of the old thing.
So you just made a copy of it,
and you may not have had to.
And so the way the value semantics work in Swift
is it uses a thing called copy on write,
which means that you get the benefit of safety
and performance, and it has another special trick
because if you think certain languages like Java,
for example, they have immutable strings.
And so what they are trying to do
is they provide value semantics
by having pure immutability.
Functional languages have pure immutability
in lots of different places,
and this provides a much safer model
and it provides value semantics.
The problem with this is if you have immutability,
everything is expensive.
Everything requires a copy.
For example, in Java,
if you have a string X and a string Y,
you append them together,
we have to allocate a new string to hold X, Y.
If they're immutable.
Well, and the strings in Java are immutable,
and there's optimizations for short runs,
and it's complicated,
but generally think about them as a separate allocation.
And so when you append them together,
you have to go allocate a third thing
because somebody might have a point
or two either of the other ones, right?
And you can't go change them.
So you have to go allocate a third thing
because of the beauty
of how the Swift value semantics system works out.
If you have a string and Swift,
and you say, hey, put in X, right?
And they say, append on Y, Z, W, W, what?
It knows that there's only one reference to that.
And so it can do an in place update.
And so you're not allocating tons of stuff on the side.
You don't have all those problems.
When you pass it off,
you can know you have the only reference.
If you pass it off to multiple different people,
but nobody changes it,
they can all share the same thing.
So you get a lot of the benefit of purely immutable design.
And so you get a really nice sweet spot
that I haven't seen in other languages.
Yeah, that's interesting.
I thought there was going to be a philosophical narrative here
that you're gonna have to pay a cost for it.
Because it sounds like, I think value semantics
is beneficial for easing of debugging
or minimizing the risk of errors,
like bringing the errors closer to the source,
bringing the symptom of the error closer
to the source of the error, however you say that.
But you're saying there's not a performance cost either
if you implement it correctly.
Well, so there's trade-offs with everything.
And so if you are doing very low-level stuff,
then sometimes you can notice a cost.
But then what you're doing is you're saying,
what is the right default?
So coming back to user interface,
when you talk about programming languages,
one of the major things that Swift does
that makes people love it that is not obvious
when it comes to designing a language
is this UI principle of progressive disclosure of complexity.
So Swift, like many languages, is very powerful.
The question is, when do you have to learn the power
as a user?
So Swift, like Python, allows you to start
with like print hello world, right?
Certain other languages start with like public static,
void, main, class, like all the ceremony, right?
And so you go to teach a new person,
hey, welcome to this new thing.
Let's talk about public access control classes.
Wait, what's that?
String system.out.println, packages, like, yeah.
Right, and so instead if you take this and you say,
hey, we need packages, you know, modules,
we need powerful things like classes,
we need data structures, we need like all these things.
The question is how do you factor the complexity
and how do you make it so that the normal case scenario
is you're dealing with things that work the right way
in the right way, give you good performance
by default.
But then as a power user, if you wanna dive down to it,
you have full c performance, full control
over low level pointers, you can call malloc
if you wanna call malloc.
This is not recommended on the first page of every tutorial,
but it's actually really important
when you wanna get work done, right?
And so being able to have that is really the design
in programming language design.
And design is really, really hard.
It's something that I think a lot of people
kind of outside of UI, again,
a lot of people just think is subjective,
like there's nothing, you know, it's just like curly braces
or whatever, it's just like somebody's preference,
but actually good design is something that you can feel.
And how many people are involved with good design?
So if we look at Swift, we'll look at historically,
I mean, this might touch like,
there's almost like a Steve Jobs question too,
like how much dictatorial decision-making
is required versus collaborative?
And we'll talk about how all that can go wrong or right,
but yeah, well, Swift, so I can't speak to in general,
all design everywhere.
So the way it works with Swift is that there's a core team.
And so core team is six or seven people ish,
something like that, that is people that have been working
with Swift since very early days.
And by early days is not that long ago.
Okay, yeah, so it became public in 2014,
so it's been six years public now,
but so that's enough time that there's a story arc there.
And there's mistakes have been made that then get fixed
and you learn something and then you, you know,
and so what the core team does is it provides continuity.
And so you wanna have a,
okay, well, there's a big hole that we wanna fill.
We know we wanna fill it.
So don't do other things that invade that space
until we fill the hole, right?
There's a boulder that's missing here.
We wanna do, we will do that boulder
even though it's not today, keep out of that space.
And the whole team remembers, remembers the myth
of the boulder that's there.
Yeah, yeah, there's a general sense
of what the future looks like in broad strokes
and a shared understanding of that,
combined with a shared understanding
of what has happened in the past that worked out well
and didn't work out well.
The next level out is you have the,
what's called the Swift Evolution community.
And you've got in that case, hundreds of people
that really care passionately about the ways Swift evolves.
And that's like an amazing thing to, again,
the core team doesn't necessarily need to come up
with all the good ideas.
You got hundreds of people out there
that care about something and they come up
with really good ideas too.
And that provides this like tumbling,
rock tumbler for ideas.
And so the evolution process is, you know,
a lot of people in a discourse forum
that are like hashing it out and trying to like talk about,
okay, well, should we go left or right?
Or if we did this, what would be good?
And, you know, here you're talking about hundreds of people.
So you're not gonna get consensus necessarily.
You're not obvious consensus.
And so there's a proposal process
that then allows the core team
and the community to work this out.
And what the core team does is it aims
to get consensus out of the community
and provide guardrails, but also provide long-term,
make sure we're going the right direction kind of things.
So does that group represent like the,
how much people will love the user interface?
Like, do you think they're able to capture that?
Well, I mean, it's something we talk about a lot.
It's something we care about.
How well we do that's up for debate,
but I think that we've done pretty well so far.
Is the beginner in mind?
Like you said, the progressive disclosure.
Yeah, so we care a lot about that,
a lot about power, a lot about efficiency,
a lot about there are many factors to good design,
and you have to figure out a way to kind of
work your way through that.
So if you like think about like a language I love is Lisp,
probably still because I use Emacs,
but they haven't done anything, any serious working Lisp,
but it has a ridiculous amount of parentheses.
Yeah.
I've also, you know, with Java and C++, the braces,
you know, I enjoyed the comfort of being between braces,
you know, and then Python is really, sorry to interrupt,
just like, and last thing to me,
as a design, if I was a language designer, God forbid,
is I would be very surprised that Python with no braces
would nevertheless somehow be comforting also.
So like, I can see arguments for all of these.
But look at this, this is evidence
that it's not about braces versus tabs.
Right, exactly.
You're good, it's a good point.
Right, so like, you know, there's evidence that-
But see like, it's one of the most argued about things.
Oh yeah, of course, just like tabs and spaces,
which it doesn't, I mean, there's one obvious right answer,
but it doesn't actually matter.
What's that?
Come on, we're friends.
Like, come on, what are you trying to do to me here?
People are gonna, yeah, half the people are gonna tune out.
Yeah.
So do you're able to identify things
that don't really matter for the experience?
Well, no, no, it's always a really hard,
so the easy decisions are easy, right?
I mean, fine, those are not the interesting ones.
The hard ones are the ones that are most interesting, right?
The hard ones are the places where,
hey, we want to do a thing,
everybody agrees we should do it,
there's one proposal on the table,
but it has all these bad things associated with it.
Well, okay, what are we gonna do about that?
Do we just take it?
Do we delay it?
Do we say, hey, well, maybe there's this other feature
that if we do that first, this will work out better.
How does this, if we do this,
are we paying ourselves into a corner, right?
And so this is where, again,
you're having that core team of people
that has some continuity and has perspective,
has some of the historical understanding
is really valuable because you get,
it's not just like one brain,
you get the power of multiple people coming together
to make good decisions,
and then you get the best out of all these people,
and you also can harness the community around it.
And what about like the decision of whether,
like in Python, having one type
or having, you know, strict typing?
Yeah, okay.
Yeah, let's talk about this.
So I like how you put that, by the way.
Like, so many people would say
that Python doesn't have types.
Doesn't have types, yeah.
But you're right, I've listened to you enough
to where I'm a fan of yours
and I've listened to way too many podcasts
and videos that you talk about this.
Oh yeah, so I would argue that Python has one type.
And so, like when you import Python and Swift,
which by the way works really well,
you have everything comes in as a Python object.
Now, here they're trade-offs because,
you know, it depends on where you're optimizing for.
And Python is a super successful language
for a really good reason.
Because it has one type,
you get duck typing for free and things like this.
But also, you're pushing,
you're making it very easy to pound out code on one hand,
but you're also making it very easy
to introduce complicated bugs, the after-debug.
And you pass the string into something
that expects an integer and it doesn't immediately die.
It goes all the way down the stack trace
and you find yourself in the middle of some code
that you really didn't wanna know anything about
and it blows up and you're just saying,
well, what did I do wrong, right?
And so, types are good and bad and they have trade-offs
that are good for performance and certain other things,
depending on where you're coming from,
but it's all about trade-offs.
And so, this is what design is, right?
Design is about weighing trade-offs
and trying to understand the ramifications
of the things that you're weighing,
like types or not, or one type or many types.
But also within many types,
how powerful do you make that type system
is another very complicated question
with lots of trade-offs.
It's very interesting, by the way.
But that's one dimension and there's a bunch
of other dimensions, jit-compiled versus static-compiled,
garbage-collected versus reference-counted,
versus manual memory management,
versus all these different trade-offs
and how you balance them
or what make the program language good.
Good currency, so in all those things,
I guess when you're designing the language,
you also have to think
of how that's gonna get all compiled down to...
If you care about performance, yeah.
Well, and go back to Lisp, right?
So Lisp also, I would say JavaScript
is another example of a very simple language, right?
And so one of the...
So I also love Lisp, I don't use it
as much as maybe you do or you did.
No, I think we're both...
Everyone who loves Lisp, it's like, you love...
It's like, I don't know, I love Frank Sinatra,
but how often do I seriously listen to Frank Sinatra?
Sure, sure.
But you look at that or you look at JavaScript,
which is another very different,
but relatively simple language.
And there's certain things that don't exist in the language,
but there is inherent complexity
to the problems that we're trying to model.
And so what happens to the complexity?
In the case of both of them, for example,
you say, well, what about large-scale software development?
Okay, well, you need something like packages.
Neither language has a language affordance for packages.
And so what you get is patterns.
You get things like NPN, you get things like,
these ecosystems that get built around.
And I'm a believer that if you don't model
at least the most important inherent complexity
in the language, then what ends up happening
is that complexity gets pushed elsewhere.
And when it gets pushed elsewhere,
sometimes that's great because often building things
as libraries is very flexible and very powerful
and allows you to evolve and things like that.
But often it leads to a lot of unnecessary divergence
in the force and fragmentation.
And when that happens, you just get kind of a mess.
And so the question is, how do you balance that?
Don't put too much stuff in the language
because that's really expensive
and it makes things complicated.
But how do you model enough of the inherent complexity
of the problem that you provide the framework
and the structure for people to think about?
Well, so the key thing to think about
with programming languages,
and you think about what a programming language is there for,
is it's about making a human more productive, right?
And so like there's an old,
I think it's Steve Jobs quote about,
it's a bicycle for the mind, right?
You can definitely walk,
but you'll get there a lot faster
if you can bicycle on your way.
And a programming language is a bicycle for the mind.
Yeah.
It's basically, wow,
that's a really interesting way to think about it.
By raising the level of abstraction now,
you can fit more things in your head.
By being able to just directly leverage somebody's library,
you can now get something done quickly.
In the case of Swift, SwiftUI is this new framework
that Apple has released recently for doing UI programming.
And it has this declarative programming model
which defines away entire classes of bugs.
It builds on value semantics
and many other nice Swift things.
And what this does is allows you to just get way more done
with way less code.
And now your productivity as a developer is much higher,
right?
And so that's really what programming languages
should be about,
is it's not about tabs versus spaces
or curly braces or whatever,
it's about how productive do you make the person.
And you can only see that when you have libraries
that were built with the right intention
that the language was designed for.
And with Swift, I think we're still a little bit early,
but SwiftUI and many other things that are coming out now
are really showing that.
And I think that they're opening people's eyes.
It's kind of interesting to think about like how
that the knowledge of something,
of how good the bicycle is,
how people learn about that.
So I've used C++,
now this is not going to be a trash talking session
about C++,
but you see C++ for a really long time.
I could go there if you want.
I have the scars.
I feel like I spent many years without realizing,
like there's languages that could,
for my particular life style,
brain style, thinking style,
there's languages that could make me a lot more productive
in the debugging stage,
in the just the development stage
and thinking like the bicycle for the mind
that I could fit more stuff into my.
Python's a great example of that, right?
I mean, a machine learning framework in Python
is a great example of that.
It's just very high abstraction level.
And so you can be thinking about things
on a very high level algorithmic level,
instead of thinking about,
okay, well, am I copying this tensor to a GPU or not?
Right?
It's not what you want to be thinking about.
And as I was telling you,
I mean, I guess the question I had is,
how does a person like me or general people discover
more productive languages?
Like how I was, as I've been telling you offline,
I've been looking for like a project to work on in Swift
so I can really try it out.
I mean, my intuition was like doing a hello world
is not going to get me there
to get me to experience the power of language.
Yeah, you need a few weeks to change your metabolism.
Exactly, beautifully put.
That's one of the problems with people with diets.
Like I'm actually currently to go in parallel,
but in a small tangent is,
I've been recently eating only meat.
Okay.
And okay.
So most people are like,
think that's horribly unhealthy or whatever.
You have like a million,
whatever the science is, it just doesn't sound right.
Well, so back when I was in college,
we did the Atkins diet, that was a thing.
Similar, but you have to always give these things a chance.
I mean, with dieting, always not dieting,
but just the things that you like,
if I eat personally, if I eat meat,
just everything, I can be super focused or more focused
than usual, I just feel great.
I've been running a lot of, you know,
doing push-ups and pull-ups and so on.
I mean, Python is similar in that sense for me.
Where are you going with this?
No, I mean, literally I just felt ahead
like a stupid smile on my face
when I first started using Python.
I could code up really quick things.
Like I would see the world,
I'll be empowered to write a script to, you know,
to do some basic data processing,
to rename files on my computer, right?
And like Perl didn't do that for me.
A little bit.
Well, and again, like none of these are about
which is best or something like that,
but there's definitely better and worse here.
But it clicks, right?
Well, yeah.
And if you look at Perl, for example,
you get bogged down in scalers versus arrays
versus hashes versus type globs
and like all that kind of stuff.
And Python's like, yeah, let's not do this, right?
And some of it is debugging.
Like everyone has different priorities,
but for me it's, can I create systems for myself
that empower me to debug quickly?
Like I've always been a big fan,
even just crewed like asserts.
Like always stating things that should be true,
which in Python I found in myself do more
because of type all these kinds of stuff.
Well, you could think of types in a programming language
as being kind of assert.
They could check to compile time, right?
So how do you learn a new thing?
Well, so this or how do people learn new things, right?
This is hard.
People don't like to change.
People generally don't like change around them either.
And so we're all very slow to adapt and change.
And usually there's a catalyst that's required
to force yourself over this.
So for learning a programming language
is really comes down to finding an excuse,
like build a thing that the language is actually good for
that the ecosystem is ready for.
And so if you were to write an iOS app, for example,
that'd be the easy case.
Obviously you would use Swift for that, right?
There are other...
Android.
So Swift runs on Android.
Oh, does it?
Oh, yeah.
Yeah, Swift runs in lots of places.
How does that work?
Okay, so Swift is built on top of LLVM.
LLVM runs everywhere.
LLVM, for example, builds the Android kernel.
Oh, wow.
Okay.
So yeah.
I didn't realize this.
Yeah, so Swift is very portable.
Runs on Windows.
There's...
It runs on lots of different things.
And Swift...
Sorry to interrupt that.
Swift UI.
And then there's a thing called UIKit.
So can I build an app with Swift?
Well, so that's the thing is the ecosystem
is what matters there.
So Swift UI and UIKit are Apple technologies.
Okay, got it.
So Swift UI happens to be written in Swift,
but it's an Apple proprietary framework
that Apple loves and wants to keep on its platform,
which makes total sense.
You go to Android and you don't have that library.
Yeah.
And so Android has a different ecosystem of things
that hasn't been built out
and doesn't work as well with Swift.
And so you can totally use Swift
to do arithmetic and things like this,
but building a UI with Swift on Android
is not a great experience right now.
So if I wanted to learn Swift,
what's the...
I mean, the one practical different version of that
is Swift for TensorFlow, for example.
And one of the inspiring things for me
with both TensorFlow and PyTorch
is how quickly the community can like switch
from different libraries.
You could see some of the communities
switching to PyTorch now,
but it's very easy to see.
And then TensorFlow is really stepping up its game.
And then there's no reason why.
I think the way it works is basically,
it has to be one GitHub repo,
like one paper steps up.
It gets people excited, yeah.
It gets people excited and they're like,
ah, I have to learn this at Swift for...
What's Swift again?
And then they learn and they fall in love with it.
I mean, that's what happened with PyTorch.
There has to be a reason, a catalyst.
And so, and there, I mean, people don't like change,
but it turns out that once you've worked
with one or two programming languages,
the basics are pretty similar.
And so, one of the fun things about learning programming
languages, even maybe Lisp, I don't know
if you agree with this, is that when you start doing that,
you start learning new things.
Because you have a new way to do things
and you're forced to do them.
And that forces you to explore
and it puts you in learning mode.
And when you get in learning mode,
your mind kind of opens a little bit
and you can see things in a new way,
even when you go back to the old place.
Right.
Yeah, so with Lisp, it's functional stuff.
But I wish there was a kind of window.
Maybe you can tell me if there is, there you go.
This is a question to ask,
what is the most beautiful feature
in a programming language?
Before I ask it, let me say like with Python,
I remember when I saw Lisp Comprehensions,
was like, when I really took it in.
I don't know, I just loved it.
It was like fun to do, like,
it was fun to do that kind of,
it was something about it to be able to filter
through a list and to create a new list
on a single line was elegant.
I could all get into my head
and it just made me fall in love with the language.
So it's there, let me ask you a question.
Is there, what do you use the most beautiful feature
in a programming languages
that you've ever encountered in Swift maybe
and then outside of Swift?
I think the thing that I like the most
from a programming language.
So I think the thing you have to think about
with a programming language, again, what is the goal?
You're trying to get people to get things done quickly.
And so you need libraries, you need high quality libraries,
and then you need a user base around them
that can assemble them and do cool things with them.
And so to me, the question is,
what enables high quality libraries?
Yeah.
And there's a huge divide in the world
between libraries who enable high quality libraries
versus the ones that put special stuff in the language.
So programming languages that enable
high quality libraries?
High quality libraries, got it.
So and what I mean by that is expressive libraries
that then feel like a natural integrated part
of the language itself.
So an example of this in Swift is that int and float
and also ray and string, things like this,
these are all part of the library.
Like int is not hard coded into Swift.
And so what that means is that because int is just
a library thing defined in the standard library,
along with strings and rays and all the other things
that come with the standard library.
Well, hopefully you do like int.
But anything that any language features
that you needed to define int,
you can also use in your own types.
So if you wanted to find a quaternion
or something like this, right?
Well, it doesn't come in the standard library.
There's a very special set of people
that care a lot about this,
but those people are also important.
It's not about classism, right?
It's not about the people who care about instant fluids
are more important than the people who care about quaternions.
And so to me, the beautiful things about programming languages
is when you allow those communities
to build high quality libraries that feel native,
that feel like they're built into the compiler
without having to be.
What does it mean for the int to be part
of not hard coded in?
So is it like how, so what isn't, what is an int?
Okay.
Int is just a integer in this case.
It's like a, you know, like a 64 bit integer
or something like this.
But so like the 64 bit is hard coded or no?
No, none of that's hard coded.
So int, if you go look at how it's implemented,
it's just a struct and Swift.
And so it's a struct.
And then how do you add two structs?
Well, you define plus and so you can define plus on int
and well, you can define plus on your thing too.
You can define int has like an, is the odd method
or something like that on it.
And so yeah, you can add methods on a thing.
So you can, you can define operators
like how it behaves.
That's used beautiful when there's something
about the language which enables others
to create libraries which are not hacky.
Yeah, they feel native.
And so one of the best examples of this is Lisp, right?
Because in Lisp, like all the libraries
are basically part of the language, right?
You write, turn rewrite systems and things like this.
And so can you as a counter example
provide what makes it difficult
to write a library that's native?
Is it the Python C?
Well, so one example, I'll give you two examples,
Java and C plus plus or Java and C.
They both allow you to define your own types.
But int is hard coded in the language.
Okay, well, why?
Well, in Java, for example,
coming back to this whole reference semantic,
value semantic thing, int gets passed around by value.
Yeah, that.
But if you make like a pair or something like that,
a complex number, right?
It's a class in Java
and now it gets passed around by reference by pointer.
And so now you lose value semantics, right?
You lost math, okay?
Well, that's not great, right?
If you can do something with int,
why can't I do it with my type, right?
So that's the negative side of the thing I find beautiful
is when you can solve that,
when you can have full expressivity
where you as a user of the language
have as much or almost as much power
as the people who implemented all the standard built-in stuff
because what that enables
is that enables truly beautiful libraries.
You know, it's kind of weird
because I've gotten used to that.
That's one, I guess, other aspect of program language design.
You have to think, you know,
the old first principles thinking,
like, why are we doing it this way?
By the way, I mean, I remember,
because I was thinking about the Waller's operator
and I'll ask you about it later,
but it hit me that like the equal sign for assignment.
Yeah.
Like, why are we using the equal sign for assignment?
It's wrong.
And that's not the only solution, right?
So if you look at Pascal,
they use colon equals for assignment
and equals for equality.
And they use like less and greater than
instead of the not equal thing.
Like there are other answers here.
So, but like, and yeah, like I ask you all,
but how do you then decide to break convention to say,
you know what?
Everybody's doing it wrong.
We're gonna do it right.
Yeah.
So it's like an ROI, like return on investment trade off,
right?
So if you do something weird,
let's just say like not like colon equal
instead of equal for assignment,
that would be weird with today's aesthetic, right?
And so you'd say, cool, this is theoretically better,
but is it better in which ways?
Like, what do I get out of that?
Do I define away class of bugs?
Well, one of the class of bugs that C has
is that you can use like, you know,
if X equals without equals equals,
if X equals Y, right?
Well, turns out you can solve that problem in lots of ways.
Clang, for example, GCC, all these compilers
will detect that as a likely bug, produce a warning.
Do they?
Yeah.
I feel like they didn't, or Clang does.
GCC didn't, and it's like,
one of the important things about
programming language design is like,
you're literally creating suffering in the world.
Okay, like, like I feel like,
I mean, one way to see it is the bicycle for the mind,
but the other way is to like, minimizing suffering.
Well, you have to decide if it's worth it, right?
And so, let's come back to that.
Okay.
But if you look at this, and again,
this is where there's a lot of detail
that goes into each of these things.
Equal and C returns a value.
Yeah.
That's messed up.
That allows you to say X equals Y equals Z,
like that works in C.
Yeah.
Is it messed up?
You know, most people think it's messed up, I think.
It is very, by messed up, what I mean is,
it is very rarely used for good,
and it's often used for bugs.
Yeah.
Right?
And so, I mean-
That's a good definition of messed up, yeah.
You could use, you know, it's a,
in hindsight, this was not such a great idea, right?
No.
One of the things with Swift that is really powerful,
and one of the reasons it's actually good,
versus it being full of good ideas,
is that when we launched Swift 1,
we announced that it was public,
people could use it, people could build apps,
but it was gonna change and break, okay?
When Swift 2 came out, we said,
hey, it's open source, and there's this open process
which people can help evolve and direct the language.
So the community, at large, like Swift users,
can now help shape the language as it is.
And what happened is that as part of that process is,
a lot of really bad mistakes got taken out.
So for example, Swift used to have the C-style plus-plus
and minus-minus operators.
Like, what does it mean when you put it before
or versus after, right?
Well, that got cargo-culted from C into Swift early on.
What's cargo-culted?
Cargo-culted means brought forward
without really considering it.
Okay.
This is maybe not the most PC term.
But-
I have to look it up in Urban Dictionary, yeah.
Yeah, so it got pulled into C without,
or it got pulled into Swift without very good consideration.
And we went through this process,
and one of the first things got ripped out was
plus-plus and minus-minus because they lead to confusion.
They have very low value over saying x plus equals one.
And x plus equals one is way more clear.
And so when you're optimizing for teachability
and clarity and bugs and this multi-dimensional space
that you're looking at, things like that really matter.
And so being first principles on where you're coming from
and what you're trying to achieve
and being anchored on the objective is really important.
Well, let me ask you about the most sort of this podcast
isn't about information, it's about drama.
Okay.
Let me talk to you about some drama.
So you mentioned Pascal and colon equals.
There's something that's called the Walrus operator.
Okay.
And Python 3.8 added the Walrus operator.
And the reason I think it's interesting
is not just because of the feature,
it has the same kind of expression feature
that you can mention to see that it returns
the value of the assignment.
And then maybe you can comment on that in general,
but on the other side of it,
it's also the thing that toppled the dictator.
So it finally drove Guido to step down
from the DFL, the toxicity of the community.
So maybe what do you think about the Walrus operator
in Python, is there an equivalent thing in Swift
that really stress tested the community?
And then on the flip side,
what do you think about Guido stepping down over it?
Yeah.
Well, if I look past the details of the Walrus operator,
one of the things that makes it most polarizing
is that it's syntactic sugar.
Okay.
What do you mean by syntactic sugar?
It means you can take something that already exists
in language and you can express it in a more concise way.
So, okay, I'm gonna play devil's advocate.
So, this is great.
Is that objective or subjective statement?
Like, can you argue that basically anything
is syntactic sugar or not?
No, not everything is syntactic sugar.
So for example, the type system,
like can you have classes versus,
do you have types or not, right?
So one type versus many types is not something
that affects syntactic sugar.
And so, if you say, I wanna have the ability
to define types, I have to have all this
like language mechanics to define classes
and oh, now I have to have inheritance
and I have like, I have all this stuff.
That's just making the language more complicated.
That's not about sugaring it.
Swift has sugar.
So like Swift has this thing called Iflet
and it has various operators
that are used to concisify specific use cases.
So the problem with syntactic sugar when you're talking about,
hey, I have a thing that takes a lot to write
and I have a new way to write it.
You have this horrible trade-off
which becomes almost completely subjective,
which is how often does this happen and does it matter?
And one of the things that is true about human psychology,
particularly when you're talking about
introducing a new thing,
is that people overestimate the burden of learning something
and so it looks foreign when you haven't gotten used to it.
But if it was there from the beginning,
of course, it's just part of Python.
Like unquestionably, like this is just the thing I know
and it's not a new thing that you're worried about learning.
It's just part of the deal.
Now with Guido, I don't know Guido well.
Yeah, have you passed cost much?
Yeah, I've met him a couple of times
but I don't know Guido well.
But the sense that I got out of that whole dynamic
was that he had put not just the decision maker
weight on his shoulders,
but it was so tied to his personal identity
that he took it personally
and he felt the need and he kind of put himself
in the situation of being the person
instead of building a base of support around him.
I mean, this is probably not quite literally true,
but by-
Yeah, too much, so there's too much-
Too much concentrated on him, right?
And so-
And that can wear you down?
Well, yeah, particularly because people then say,
Guido, you're a horrible person.
I hate the saying blah, blah, blah, blah, blah, blah,
and sure it's like maybe 1% of the community
that's doing that,
but Python's got a big community
and 1% of millions of people is a lot of hate mail
and that just from human factor will just wear on you.
Well, to clarify, it looked from just what I saw
in the messaging for the,
let's not look at the million Python users,
but at the Python Core developers,
it feels like the majority,
the big majority on a vote were opposed to it.
Okay, I'm not that close to it, so I don't know.
So this, okay, so the situation is like literally,
yeah, I mean, the majority of the core developers are against-
Were opposed to it?
So, and they weren't even like against it.
It was, there was a few, well, they were against it,
but the against it wasn't like, this is a bad idea.
They were more like, we don't see why this is a good idea.
And what that results in is there's a stalling feeling,
like you just slow things down.
Now, from my perspective, that you could argue this,
and I think it's very interesting
if we look at politics today and the way Congress works,
it's slowed down everything.
It's a dampener.
Yeah, it's a dampener, but like that's a dangerous thing too,
because if it dampens things like, you know-
Well, what are you talking about?
Like it's a low-pass filter,
but if you need billions of dollars injected
into the economy or trillions of dollars,
then suddenly stuff happens, right?
And so-
For sure, so you're talking about-
I'm not defending our political situation,
just to be clear.
But you're talking about like a global pandemic.
Well, I was hoping we could fix like the healthcare system
and the education system, like, you know-
I'm not a politics person, I don't know.
When it comes to languages, the community's kind of right
in terms of it's a very high burden
to add something to a language.
So as soon as you add something,
you have a community of people building on it
and you can't remove it, okay?
And if there's a community of people
that feel really uncomfortable with it,
then taking it slow, I think, is an important thing to do.
And there's no rush, particularly with something
that's 25 years old and is very established.
And, you know, it's not like coming into its own.
What about features?
Well, so I think that the issue with Guido
was that maybe this is a case
where he realized that had outgrown him
and it went from being the language.
So Python, I mean, Guido's amazing.
But Python isn't about Guido anymore.
It's about the users.
And to a certain extent, the users own it.
And, you know, Guido spent years of his life,
a significant fraction of his career on Python.
And from his perspective, I imagine he's like,
well, this is my thing,
I should be able to do the thing I think is right.
But you can also understand the users
where they feel like, you know, this is my thing.
I use this like, and I don't know, it's a hard thing.
But what, if we could talk about leadership in this,
because it's so interesting to me,
I'm gonna make, hopefully somebody makes it,
if not, I'll make it a water-stopper shirt,
because I think it represents to me,
maybe it's my Russian roots or something.
You know, it's the burden of leadership.
Like, I feel like to push back,
I feel like progress can only,
like most difficult decisions, just like you said,
there'll be a lot of divisiveness over,
especially in the passionate community.
It just feels like leaders need to take those risky decisions
that if you like, listen,
that with some non-zero probability,
maybe even a high probability would be the wrong decision,
but they have to use their gut and make that decision.
Well, this is like one of the things
where you see amazing founders.
The founders understand exactly what's happened
and how the company got there,
and are willing to say, we have been doing thing X
the last 20 years, but today we're gonna do thing Y.
And they make a major pivot for the whole company.
The company lines up behind them, they move,
and it's the right thing.
But then when the founder dies, the successor
doesn't always feel that agency
to be able to make those kinds of decisions.
Even though they're a CEO, they could theoretically do whatever.
There's two reasons for that, in my opinion,
or in many cases, it's always different.
But one of which is they weren't there
for all the decisions that were made,
and so they don't know the principles
in which those decisions were made.
And once the principles change,
you should be obligated to change what you're doing
and change direction, right?
And so if you don't know how you got to where you are,
it just seems like gospel,
and you're not gonna question it.
You may not understand that it really is
the right thing to do, so you just may not see it.
That's so brilliant.
I never thought of it that way.
Like it's so much higher burden
when as a leader you step into a thing
that's already worked for a long time.
Yeah, yeah.
And if you change it and it doesn't work out,
now you're the person who screwed it up.
People always second guess it.
Yeah.
And the second thing is that
even if you decide to make a change,
even if you're theoretically in charge,
you're just a person that thinks they're in charge.
Meanwhile, you have to motivate the troops.
You have to explain it to them in terms they'll understand.
You have to get them to buy into it and believe in it
because if they don't,
then they're not gonna be able to make the turn
even if you tell them their bonuses are gonna be curtailed.
They're just not gonna like buy into it.
And so there's only so much power you have as a leader
and you have to understand what those limitations are.
Are you still BDFL?
You've been BDFL of some stuff.
You're very heavy on the B, the benevolent,
benevolent, dictated for life.
I guess LVM.
Yeah, so I still lead the LVM world.
I mean, what's the role of,
so then on Swift you said that there's a group of people.
Yeah.
So if you contrast Python with Swift, right?
One of the reasons,
so everybody on the core team takes the role really seriously
and I think we all really care about where Swift goes,
but you're almost delegating the final decision making
to the wisdom of the group.
And so it doesn't become personal.
And also when you're talking with the community,
so yeah, some people are very annoyed
at certain decisions that get made.
There's a certain faith in the process
because it's a very transparent process
and when a decision gets made,
a full rationale is provided, things like this.
These are almost defense mechanisms
to help both guide future discussions
and provide case lock and like Supreme Court does
about this decision was made for this reason
and here's the rationale
and what we wanna see more of or less of.
But it's also a way to provide a defense mechanism
so that when somebody's griping about it,
they're not saying that person did the wrong thing.
They're saying, well, this thing sucks
and later they move on and they get over it.
Yeah, the analogy is Supreme Court,
I think is really good.
But then, okay, not to get person on the SWIFT team,
but like is there, it just seems like it's impossible
for division not to emerge.
Well, each of the humans on the SWIFT core team,
for example, are different
and the membership of the SWIFT core team changes slowly
over time, which is I think a healthy thing.
And so each of these different humans
have different opinions.
Trust me, it's not a singular consciousness
by any stretch of the imagination.
You've got three major organizations,
including Apple, Google and Sci-Fi,
all kind of working together
and it's a small group of people, but you need high trust.
You need, again, it comes back to the principles
of what you're trying to achieve
and understanding what you're optimizing for.
And I think that starting with strong principles
and working towards decisions is always a good way
to both make wise decisions in general,
but then be able to communicate them to people
so that they can buy into them.
And that is hard.
And so you mentioned LVM.
LVM is gonna be 20 years old this December.
So it's showing its own age.
Do you have like a dragon cake plan or do you have a?
No, I should definitely do that.
Yeah, if we can have a pandemic cake.
Pandemic cake.
Everybody gets a slice of cake and it gets sent through email.
But LVM has had tons of its own challenges over time too.
And one of the challenges that the LVM community has,
in my opinion, is that it has a whole bunch of people
that have been working on LVM for 10 years,
because this happens somehow.
And LVM has always been one way,
but it needs to be a different way.
And they've worked on it for like 10 years.
It's a long time to work on something.
And you suddenly can't see the faults
in the thing that you're working on.
And LVM has lots of problems
and we need to address them and we need to make it better.
And if we don't make it better,
then somebody else will come up with a better idea.
And so it's just kind of of that age
where the community is in danger of getting too calcified.
And so I'm happy to see new projects joining
and new things mixing it up.
Fortran is now a new thing in the LVM community,
which is hilarious and good.
I've been trying to find, on this little tangent,
find people who program in Cobalt or Fortran,
Fortran especially, to talk to it.
They're hard to find.
Yeah, look to the scientific community.
They still use Fortran quite a bit.
Well, interesting thing you kind of mentioned with LVM
or just in general that as something evolved,
you're not able to see the faults.
So do you fall in love with the thing over time
or do you start hating everything about the thing over time?
Well, so my personal folly is that I see maybe not all,
but many of the faults and they grate on me
and I don't have time to go fix them.
Yeah, and they get magnified over time.
Well, and they may not get magnified,
but they never get fixed.
And it's like sand underneath,
that's just like grating against you
and it's like sand underneath your fingernails or something.
It's just like, you know, it's there,
you can't get rid of it.
And so the problem is that if other people don't see it,
right, nobody ever, like I can't go,
I don't have time to go write the code and fix it anymore,
but then people are resistant to change.
And so you say, hey, we should go fix this thing.
They're like, oh yeah, that sounds risky.
It's like, well, is it the right thing or not?
Are the challenges, the group dynamics
or is it also just technical?
I mean, some of these features like,
I think as an observer, it's almost like a fan
in the, you know, as a spectator of the whole thing.
I don't often think about, you know,
some things might actually be technically difficult
to implement.
An example of this is we built this new compiler framework
called MLIR.
Yes.
MLIR is this a whole new framework.
It's not, many people think it's about machine learning.
The ML stands for multi-level
because compiler people can't name things very well, I guess.
Can we dig into what MLIR is?
Yeah, so when you look at compilers,
compilers have historically been solutions
for a given space.
So LLVM is a, it's really good for dealing with CPUs,
let's just say, at a high level.
You look at Java, Java has a JVM.
The JVM is very good for garbage collected languages
that need dynamic compilation.
And it's very optimized for a specific space.
And so hotspot is one of the compilers
that gets used in that space.
And that compiler's really good at that kind of stuff.
Usually when you build these domain specific compilers,
you end up building the whole thing from scratch
for each domain.
What's a domain?
So what's the scope of a domain?
Well, so here I would say, if you look at Swift,
there's several different parts to the Swift compiler,
one of which is covered by the LLVM part of it.
There's also a high level piece that's specific to Swift.
And there's a huge amount of redundancy
between those two different infrastructures
and a lot of re-implemented stuff
that is similar but different.
What is LLVM defined?
LLVM is effectively an infrastructure.
So you can mix and match it in different ways.
It's built out of libraries.
You can use it for different things.
But it's really good at CPUs and GPUs.
CPUs and the tip of the iceberg on GPUs.
It's not really great at GPUs.
But it turns out that then use it to talk to CPUs.
And so it turns out there's a lot of hardware out there
that is custom accelerators.
So machine learning, for example.
There are a lot of matrix multiply accelerators
and things like this.
There's a whole world of hardware synthesis.
So we're using MLIR to build circuits.
And so you're compiling for a domain of transistors.
And so what MLIR does is it provides
a tremendous amount of compiler infrastructure
that allows you to build these domain-specific compilers
in a much faster way and have the result be good.
If we're thinking about the future,
now we're talking about ASICs, so anything.
So if we project into the future,
it's very possible that the number of these kinds of ASICs,
very specific infrastructure, architecture things,
multiplies exponentially.
I hope so.
So that's MLIR.
So what MLIR does is it allows you to build these compilers
very efficiently.
Now, one of the things that, coming back to the LLVM thing,
and then we'll go to hardware, is LLVM
is a specific compiler for a specific domain.
MLIR is now this very general, very flexible thing
that can solve lots of different kinds of problems.
So LLVM is a subset of what MLIR does.
So MLIR is, I mean, it's an ambitious project then.
Yeah, it's a very ambitious project, yeah.
And so to make it even more confusing,
MLIR has joined the LLVM umbrella project.
So it's part of the LLVM family.
But where this comes full circle is now
folks that work on the LLVM part, the classic part
that's 20 years old, aren't aware of all the cool new things
that have been done in the new thing that MLIR was built by me
and many other people that knew a lot about LLVM.
And so we fixed a lot of the mistakes that lived in LLVM.
We have this community dynamic where it's like, well,
there's this new thing, but it's not familiar.
Nobody knows it.
It feels like it's new.
And so let's not trust it.
And so it's just really interesting
to see the cultural, social dynamic that comes out of that.
And I think it's super healthy because we're
seeing the ideas percolate and we're
seeing the technology diffusion happen
as people get more comfortable with it.
They start to understand things in their own terms.
And this just gets to the, it takes a while for ideas
to propagate, even though they may be very different than what
people are used to.
Maybe let's talk about that a little bit, the world of ASICS.
And well, actually, you have a new role at SciFive.
What's that place about?
What is the vision for their vision for, I would say,
the future of computer?
Yeah.
So I lead the engineering and product teams at SciFive.
SciFive is a company who was founded
with this architecture called RISC-5.
RISC-5 is a new instruction set.
Instruction sets are the things inside of your computer
that tell how to run things.
x86 from Intel and ARM from the ARM company and things
like this or other instruction sets.
I've talked to Dave Patterson, who's
super excited about RISC-5.
Dave is awesome.
He's brilliant.
Yeah.
The RISC-5 is distinguished by not being proprietary.
And so x86 can only be made by Intel and AMD.
ARM can only be made by ARM.
They sell licenses to build arm ships to other companies,
things like this.
MIPS is another instruction set that
is owned by the MIPS company, now Wave.
And then it gets licensed out, things like that.
And so RISC-5 is an open standard
that anybody can build chips for.
And so SciFive was founded by three
of the founders of RISC-5 that designed and built it
in Berkeley, working with Dave.
And so that was the genesis of the company.
SciFive today has some of the world's best RISC-5 cores.
And we're selling them.
And that's really great.
They're going to tons of products.
It's very exciting.
So they're taking this thing that's open source
and just trying to be, or are, the best in the world
at building these things.
Yeah, so here it's the specifications open source.
It's like saying TCPIP is an open standard,
or C is an open standard.
But then you have to build an implementation of the standard.
And so SciFive, on the one hand, pushes forward and defined
and pushes forward the standard.
On the other hand, we have implementations
that are best in class for different points in the space,
depending on if you want a really tiny CPU,
or if you want a really big, beefy one that is faster,
but it uses more area and things like this.
What about the actual manufacturer chip?
So where does that all fit?
I'm going to ask a bunch of dumb questions.
That's OK.
This is how we learn, right?
And so the way this works is that there's generally
a separation of the people who design the circuits
and then the people who manufacture them.
And so you'll hear about fabs, like TSMC and Samsung
and things like this, that actually produce the chips.
But they take a design coming in and that design
specifies how you turn code for the chip
into little rectangles that then use photo lithography
to make mask sets and then burn transistors onto silicon.
So and we're talking about mass manufacturing.
Yeah, they're talking about making hundreds of millions
of parts and things like that.
And so the fab handles the volume production, things
like that.
But when you look at this problem,
the interesting thing about the space when you look at it
is that the steps that you go from designing a chip
and writing the quote unquote code for it and things
like fair log and languages like that down to what
you hand off to the fab is a really well studied, really
old problem.
Tons of people have worked on it.
Lots of smart people have built systems and tools.
These tools then have generally gone through acquisitions.
And so they've ended up at three different major companies
that build and sell these tools.
They're called EDA tools like for electronic design
automation.
The problem with this is you have huge amounts of fragmentation.
You have loose standards.
And the tools don't really work together.
So you have tons of duct tape and you
have tons of lost productivity.
Now these are tools for design, so the RISC-5 is an instruction.
What is RISC-5?
How deep does it go?
How much does it touch the hardware?
How much does it define?
How much are the hardware?
Yeah, so RISC-5 is all about given a CPU,
so the processor and your computer.
How does the compiler, like the Swift compiler, the C
compiler, things like this, how does it make it work?
So what is the assembly code?
And so you write RISC-5 assembly instead of XA6 assembly,
for example.
But it's a set of instructions as opposed to instructions.
What do you say?
It tells you how the compiler works.
Sorry, it's what the compiler talks to.
OK.
And then the tooling you mentioned, the disparate tools,
are for what?
For when you're building a specific chip.
So RISC-5.
In hardware.
In hardware, yeah.
So RISC-5, you can buy a RISC-5 core from Sci-5
and say, hey, I want to have a certain number of run,
a certain number of gigahertz.
I want it to be this big.
I want it to have these features.
I want floating point or not, for example.
And then what you get is you get a description of a CPU
with those characteristics.
Now, if you want to make a chip, you
want to build an iPhone chip or something like that.
You have to take both the CPU, but then you
have to talk to memory.
You have to have timers, IOs, a GPU, other components.
And so you need to pull all those things together
into what's called an ASIC, an application
specific in a great circuit, so a custom chip.
And then you take that design, and then you
have to transform it into something
that the fabs, like TSMC, for example,
know how to take to production.
Got it.
So yeah.
And so that process, I can't help but see it as,
is a big compiler.
It's a whole bunch of compilers written without thinking about
it through that lens.
Isn't the universe a compiler?
Yeah, compilers do two things.
They represent things and transform them.
And so there's a lot of things that end up being compilers.
But this is a space where we're talking about design
and usability.
And the way you think about things,
the way things compose correctly, it matters a lot.
And so Sci-Fi is investing a lot into that space.
And we think that there's a lot of benefit
that can be made by allowing people to design chips faster,
get them to market quicker, and scale out.
Because at the end of Moore's Law,
you've got this problem of you're not
getting free performance just by waiting another year
for a faster CPU.
And so you have to find performance in other ways.
And one of the ways to do that is with custom accelerators
and other things in hardware.
And so we'll talk a little more about ASICs.
But do you see that a lot of people, a lot of companies
will try to have a different sets of requirements
that this whole process to go for?
So almost different car companies
might use different and different PC manufacturers.
So is risk five in this whole process,
is it potentially the future of all computing devices?
Yeah, I think that so if you look at risk five
and step back from the silicon side of things,
risk five is an open standard.
And one of the things that has happened over the course
of decades, if you look over the long arc of computing,
somehow it became decades old.
So you have companies that come and go,
and you have instruction sets that come and go.
Like one example of this, out of many, is Sun with Spark.
Sun one way.
Spark still lives on it if you just do.
But we have HP had this instruction set called PA Risk.
So PA Risk was its big server business
and had tons of customers.
They decided to move to this architecture called
Itanium from Intel.
This didn't work out so well.
And so you have this issue of you're
making many billion dollar investments on instruction
sets that are owned by a company.
And even companies as big as Intel
don't always execute as well as they could.
They have their own issues.
HP, for example, decided that it wasn't
in their best interest to continue investing in the space
because it was very expensive.
And so they make technology decisions.
They make their own business decisions.
And this means that as a customer, what do you do?
You've sunk all this time, all this engineering,
all this software work, all these.
You've built other products around them, and now you're stuck.
What risk five does is it provides you
more optionality in the space.
Because if you buy an implementation of risk five from sci-fi,
and you should, they're the best ones.
But if something bad happens to sci-fi in 20 years,
well, great, you can turn around and buy risk five
core from somebody else.
And there's an ecosystem of people
that are all making different risk five cores with different
trade-offs, which means that if you have more than one
requirement, if you have a family of products,
you can probably find something in the risk five space that
fits your needs.
Whereas if you're talking about X86, for example,
Intel's only going to bother to make certain classes of devices.
I see.
So maybe a weird question.
But if sci-fi is infinitely successful in the next 20,
30 years, what does the world look like?
So how does the world of computing change?
So too much diversity in hardware instruction
sets I think is bad.
We have a lot of people that are using
lots of different instruction sets, particularly
in the embedded very tiny microcontroller space, the thing
in your toaster, that are just weird and different
for historical reasons.
And so the compilers and the tool chains and the languages
on top of them aren't there.
And so the developers for that software
have to use really weird tools because the ecosystem that
supports is not big enough.
So I expect that will change.
People will have better tools and better languages,
better features everywhere that then can
serve as many different points in the space.
And I think risk five will progressively
eat more of the ecosystem because it can scale up.
It can scale down sideways.
Left, right, it's very flexible and very well-considered
and well-designed instruction set.
I think when you look at sci-fi of tackling silicon
and how people build chips, which is a very different space,
that's where you say, I think we'll
see a lot more custom chips.
And that means that you get much more battery life.
You get better tuned solutions for your IoT thingy.
You get people to move faster.
You get the ability to have faster time to market,
for example.
So how many custom?
So first of all, on the IoT side of things,
do you see the number of smart toasters
increasing exponentially?
So and if you do, how much customization per toaster
is there?
Do all toasters in the world run the same silicon,
like the same design?
Or is it different companies have different design?
Like how much customization is possible here?
Well, a lot of it comes down to cost.
Right, and so the way that chips work
is you end up paying by the, one of the factors
is the size of the chip.
And so what ends up happening just
from an economic perspective is there's
only so many chips that get made in a year of a given design.
And so often what customers end up having to do
is they end up having to pick up a chip that
exists that was built for somebody else
so that they can then ship their product.
And the reason for that is they don't
have the volume of the iPhone.
They can't afford to build a custom chip.
However, what that means is they're now
buying an off-the-shelf chip that isn't a perfect fit
for their needs.
So they're paying a lot of money for it
because they're buying silicon that they're not using.
Well, if you now reduce the cost of designing the chip,
now you get a lot more chips.
And the more you reduce it, the easier it is to design chips.
The more the world keeps evolving,
and we get more AI accelerators, we get more other things,
we get more standards to talk to, we get 6G, right?
You get changes in the world that you
want to be able to talk to these different things.
There's more diversity in the cross product
of the features that people want.
And that drives differentiated chips
in another direction.
And so nobody really knows what the future looks like.
But I think that there's a lot of silicon in the future.
Speaking of the future, you said Moore's law allegedly is dead.
So do you agree with Dave Patterson and many folks
that Moore's law is dead?
Or do you agree with Jim Keller, who's standing
at the helm of the pirate ship, saying it's still alive?
It's still alive.
Yeah.
Well, so I agree with what they're saying,
and different people are interpreting the animal's law
in different ways.
So Jim would say, there's another 1,000x left in physics.
And we can continue to squeeze the stone
and make it faster and smaller and smaller geometries
and all that kind of stuff.
He's right.
So Jim is absolutely right that there's
a ton of progress left.
And we're not at the limit of physics yet.
That's not really what Moore's law is, though.
If you look at what Moore's law is,
is that it's a very simple evaluation of, OK, well,
you look at the cost per area and the most economic point
in that space.
And if you go look at the now quite old paper that
describes this, Moore's law has a specific economic aspect
to it.
And I think this is something that Dave and others often
point out.
And so on a technicality, that's right.
I look at it from so I can acknowledge
both of those viewpoints.
They're both right.
They're both right.
I'll give you a third wrong viewpoint
that may be right in its own way, which is single threaded
performance doesn't improve like it used to.
And it used to be back when you got a Pentium 66 or something
and the year before you had a Pentium 33.
And now it's twice as fast.
Well, it was twice as fast at doing exactly the same thing.
Like literally the same program ran twice as fast.
You just wrote a check and waited a year and a half.
Well, so that's what a lot of people think about Moore's law.
And I think that is dead.
And so what we're seeing instead is we're pushing people
to write software in different ways.
And so we're pushing people to write CUDA
so they can get GPU compute and the thousands of cores on GPU.
We're talking about C programmers having to use
P threads because they now have 100 threads or 50 cores
in a machine or something like that.
You're now talking about machine learning accelerators
that are now domain specific.
And when you look at these kinds of use cases,
you can still get performance.
And Jim will come up with cool things
that utilize the silicon in new ways for sure.
But you're also going to change the programming model.
And now when you start talking about changing
the programming model, that's when you come back to languages
and things like this too, because often what you see
is like you take the C programming language, right?
The C programming language is designed for CPUs.
And so if you want to talk to a GPU,
now you're talking to its cousin CUDA.
CUDA is a different thing with a different set of tools,
a different world, a different way of thinking.
And we don't have one world that scales.
And I think that we can get there.
We can have one world that scales in a much better way.
And a small tangent then, I think most programming languages
are designed for CPUs, for a single core,
even just in their spirit, even if they allow for parallelization.
So what does it look like for a programming language
to have parallelization or massive parallelization
as it's like first principle?
So the canonical example of this is the hardware design world.
So Verilog, VHDL, these kinds of languages,
they're what's called a high-level synthesis language.
This is the thing people design chips in.
And when you're designing a chip, it's kind of like a brain,
where you have infinite parallelism.
You're like laying down transistors.
Transistors are always running.
And so you're not saying, run this transistor,
then this transistor, then this transistor.
It's like your brain.
Your neurons are always just doing something.
They're not clocked, right?
They're just doing their thing.
And so when you design a chip, or when you design a CPU,
when you design a GPU, when you design,
when you're laying down the transistors, similarly,
you're talking about, well, OK, well,
how do these things communicate?
And so these languages exist.
Verilog is a kind of mixed example of that.
Now, these languages are really great.
Yeah, at a very low level, yeah.
At a very low level.
And abstraction is necessary here.
And there's different approaches with that.
And it's itself a very complicated world.
But it's implicitly parallel.
And so having that as the domain that you program towards
makes it so that by default, you get parallel systems.
If you look at CUDA, CUDA is a point halfway in the space,
where in CUDA, when you write a CUDA kernel for your GPU,
it feels like you're writing a scalar program.
So you have ifs, you have for loops, stuff like this.
You're just writing normal code.
But what happens outside of that in your driver
is that it actually is running you on 1,000 things at once.
And so it's parallel, but it has pulled it
out of the programming model.
And so now you, as a programmer, are working in a simpler world.
And it's solved that for you.
How do you take the language like Swift?
You know, if we think about GPUs, but also ASICs,
maybe if we can dance back and forth between hardware
and software, how do you design for these features
to be able to program, make it a first class citizen,
to be able to do Swift for TensorFlow,
to be able to do machine learning on the current hardware,
but also future hardware like GPUs and all kinds of ASICs
that I'm sure will be popping up more and more.
Yeah, well, so a lot of this comes down to this whole idea
of having the nuts and bolts underneath the covers that
work really well.
So if you're talking to TPUs, you need MLIR or XLA,
or one of these compilers that talks to TPUs to build on top of.
And if you're talking to circuits,
you need to figure out how to lay down the transistors
and how to organize it and how to set up clocking
and all the domain problems that you get with circuits.
Then you have to decide how to explain it to a human.
What is the UI?
And if you do it right, that's a library problem, not
a language problem.
And that works if you have a library or a language which
allows your library to write things that feel native
in the language by implementing libraries.
Because then you can innovate in programming models
without having to change your syntax again.
And you have to invent new code formatting tools
and all the other things that languages come with.
And this gets really interesting.
And so if you look at the space, the interesting thing,
once you separate out syntax, becomes,
what is that programming model?
And so do you want the CUDA style?
I write one program and it runs many places.
Do you want the implicitly parallel model?
How do you reason about that?
How do you give developers, chip architects, the ability
to express their intent?
And that comes into this whole design
question of how do you detect bugs quickly?
So you don't have to tape out a chip to find out it's wrong,
ideally.
And this is a spectrum.
How do you make it so that people feel productive,
so their turnaround time is very quick?
All these things are really hard problems.
And in this world, I think that not a lot of effort
has been put into that design problem
and thinking about the layering in other pieces.
Well, you've, on the topic of concurrency,
you've written the Swift concurrency manifest.
I think it's kind of interesting.
Anything that has the word manifesto in it is very interesting.
Can you summarize the key ideas of each of the five parts
you've written about?
So what is a manifesto?
Yes.
How about we start there?
So in the Swift community, we have this problem,
which is on the one hand, you want
to have relatively small proposals
that you can kind of fit in your head.
You can understand the details at a very fine-grained level
that move the world forward.
But then you also have these big arcs.
And often when you're working on something
that is a big arc, but you're tackling it in small pieces,
you have this question of, how do I
know I'm not doing a random walk?
Where are we going?
How does this add up?
Furthermore, when you start that first small step,
what terminology do you use?
How do we think about it?
What is better and worse than space?
What are the principles?
What are we trying to achieve?
And so what a manifesto in the Swift community
does is it starts to say, hey, well, let's
step back from the details of everything.
Let's paint a broad picture to talk
about what we're trying to achieve.
Let's give an example design point.
Let's try to paint the big picture
so that then we can zero in on the individual steps
and make sure that we're making good progress.
And so the Swift concurrency manifesto
is something I wrote three years ago.
It's been a while, maybe more, trying
to do that for Swift in concurrency.
It starts with some fairly simple things,
like making the observation that when
you have multiple different computers or multiple different
threads that are communicating, it's best for them
to be asynchronous.
And so you need things to be able to run separately
and then communicate with each other.
And this means asynchronous.
And this means that you need a way
to modeling asynchronous communication.
Many languages have features like this.
Async await is a popular one.
So that's what I think is very likely in Swift.
But as you start building this tower of abstractions,
it's not just about how do you write this.
You then reach into the, how do you get memory safety?
Because you want correctness.
You want debugability and sanity for developers.
And how do you get that memory safety into the language?
So if you take a language like Go or C or any of these languages,
you get what's called a race condition
when two different threads or Go routines or whatever
touch the same point in memory, right?
This is a huge like maddening problem to debug
because it's not reproducible generally.
And so there's tools, there's a whole ecosystem of solutions
that built up around this, but it's a huge problem
when you're writing concurrent code.
And so with Swift, this whole value semantics thing
is really powerful there because it turns out that math
and copies actually work even in concurrent worlds.
And so you get a lot of safety just out of the box,
but there are also some hard problems and it talks
about some of that.
When you start building up to the next level up
and you start talking beyond memory safety,
you have to talk about what is a programmer model?
How does a human think about this?
So a developer that's trying to build a program
think about this and it proposes a really old model
with a new spin called Actors.
Actors are about saying we have islands of single threadedness
logically.
So you write something that feels like it's one programming,
one program running in a unit.
And then it communicates asynchronously
with other things.
And so making that expressive and natural feel good
be the first thing you reach for and being safe by default
is a big part of the design of that proposal.
When you start going beyond that, now you start to say, cool.
Well, these things that communicate asynchronously,
they don't have to share memory.
Well, if they don't have to share memory
and they're sending messages to each other,
why do they have to be in the same process?
These things should be able to be in different processes
on your machine, and why just processes?
Well, why not different machines?
And so now you have a very nice gradual transition
towards distributed programming.
And of course, when you start talking about the big future,
the manifesto doesn't go into it.
But accelerators are things you talk to asynchronously
by sending messages to them.
And how do you program those?
Well, that gets very interesting.
That's not in the proposal.
And how much do you want to make the control of that whole process
explicit to the programmer?
Yeah, good question.
So when you're designing any of these kinds of features,
or language features, or even libraries,
you have this really hard trade-off
that you have to make, which is, how much is it magic?
Or how much is it in the human's control?
How much can they predict and control it?
What do you do when the default case is the wrong case?
And so when you're designing a system, I won't name names.
But there are systems where it's really easy to get started.
And then you jump it.
So let's pick logo.
So something like this.
So it's really easy to get started.
It's really designed for teaching kids.
But as you get into it, you hit a ceiling.
And then you can't go any higher.
And then what do you do?
Well, you have to go switch to a different world
and rewrite all your code.
And this logo is a silly example here.
This exists in many other languages.
With Python, you would say concurrency.
So Python has the global interpreter lock.
So threading is challenging in Python.
And so if you start writing a large-scale application
in Python, and then suddenly you need concurrency,
you're stuck with a series of bad trade-offs.
There's other ways to go where you say,
like, foist all the complexity on the user all at once.
And that's also bad in a different way.
And so what I prefer is building a simple model
that you can explain that then has an escape hatch.
So you get in, you have guardrails.
Memory safety works like this in Swift,
where you can start with, by default,
if you use all the standard things, it's memory safe.
You're not going to shoot your foot off.
But if you want to get a C-level pointer to something,
you can explicitly do that.
But by default, there's guardrails.
There's guardrails.
OK.
But whose job is it to figure out which part of the code
is paralyzable?
So in the case of the proposal, it is the human's job.
So they decide how to architect their application.
And then the runtime in the compiler is very predictable.
And so this is in contrast to, like,
there's a long body of work, including on Fortran,
for auto-paralyzing compilers.
And this is an example of a bad thing.
So as a compiler person, I can rag on compiler people.
Often compiler people will say, cool,
since I can't change the code, I'm
going to write my compiler that then takes this unmodified code
and makes it go way faster on this machine.
OK.
Application development.
And so it does pattern matching.
It does really deep analysis.
Compiler people are really smart.
And so they want to do something really clever and tricky.
And you get 10x speed up by taking an array of structures
and turning it into a structure of arrays or something,
because it's so much better for memory.
There's tons of tricks.
They love optimization.
Yeah, you love optimization.
Everyone loves optimization.
Everyone loves it.
And it's this promise of build with my compiler
and your thing goes fast.
But here's the problem.
Lex, you write a program.
You run it with my compiler.
It goes fast.
You're very happy.
Wow, it's so much faster than the other compiler.
Then you go and you got a feature to your program,
or you refactor some code.
And suddenly you got a 10x loss in performance.
Well, why?
What just happened there?
What just happened there is the heuristic,
the pattern matching, the compiler, or whatever
analysis it was doing just got defeated because you didn't
inline a function or something.
As a user, you don't know.
You don't want to know.
That was the whole point.
You don't want to know how the compiler works.
You don't want to know how the memory hierarchy works.
You don't want to know how it got
parallelized across all these things.
You wanted that abstract away from you.
But then the magic is lost as soon as you did something
and you fall off a performance cliff.
And now you're in this funny position where what do I do?
I don't change my code.
I don't fix that bug.
It costs 10x performance.
Now what do I do?
Well, this is the problem with unpredictable performance.
If you care about performance, predictability
is a very important thing.
And so what the proposal does is it
provides architectural patterns for being
able to lay out your code, gives you full control over that,
makes it really simple so you can explain it.
And then if you want to scale out in different ways,
you have full control over that.
So in your sense, the intuition is
for a compiler is too hard to do automated parallelization.
Because the compilers do stuff automatically
that's incredibly impressive for other things.
But for parallelization, we're not close to there.
Well, it depends on the programming model.
So there's many different kinds of compilers.
And so if you talk about a C compiler, a Swift compiler,
or something like that where you're writing imperative code,
parallelizing that and reasoning about all the pointers
and stuff like that is a very difficult problem.
Now, if you switch domains, so there's
this cool thing called machine learning.
So machine learning nerds, among other endearing things
like solving cat detectors and other things like that,
have done this amazing breakthrough
of producing a programming model, operations
that you can pose together that has raised the level
of abstraction high enough that suddenly you
can have auto-parallelizing compilers.
You can write and model using TensorFlow
and have it run on 1,024 nodes of a TPU.
Yeah, that's true.
I didn't even think about like, because there's
so much flexibility in the design of architectures
that ultimately boil down to a graph that's
parallelizable for you, parallelized for you.
And if you think about it, that's pretty cool.
That's pretty cool, yeah.
And you think about batching, for example,
as a way of being able to exploit more parallelism.
That's a very simple thing that now is very powerful.
That didn't come out of the programming language nerds,
those people.
That came out of people that are just
looking to solve a problem and use a few GPUs
and organically developed by the community of people
focusing on machine learning as an incredibly powerful abstraction
layer that enables the compiler people to go and exploit that.
And now you can drive supercomputers from Python.
That's pretty cool.
So just to pause on that, because I'm not
sufficiently low level, I forget to admire the beauty
and power of that.
But maybe just to linger on it, what
does it take to run in your network fast?
How hard is that compilation?
It's really hard.
So we just skipped.
You said it's amazing that that's a thing.
But how hard is that of a thing?
It's hard.
And I would say that not all of the systems are really great,
including the ones I helped build.
So there's a lot of work left to be done there.
Is it the compiler nerds working on that,
or is it a whole new group of people?
Well, it's a full stack problem, including compiler people,
including APIs, like Keras and the module API and PyTorch and Jax.
And there's a bunch of people pushing on all the different parts
of these things.
Because when you look at it, it's both,
how do I express the computation?
Do I stack up layers?
Well, cool.
Like setting up a linear sequence of layers
is great for the simple case.
But how do I do the hard case?
How do I do reinforcement learning?
Well, now I need to integrate my application logic in this.
Then it's the next level down of,
how do you represent that for the runtime?
How do you get hardware abstraction?
And then you get to the next level down of saying,
forget about abstraction.
How do I get the peak performance out of my TPU,
or my iPhone accelerator, or whatever,
and all these different things?
And so this is a layered problem with a lot
of really interesting design and work going on in the space.
And a lot of really smart people working on it.
Machine learning is a very well-funded area
of investment right now.
And so there's a lot of progress being made.
So how much innovation is there on the lower level,
so closer to the ASIC, so redesigning the hardware
or redesigning concurrently compilers with that hardware?
Is that if you were to predict the biggest,
the equivalent of Moore's law improvements in the inference
and the training of neural networks and just all of that,
where is that going to come from, you think?
Sure, you get scalability, you have different things.
And so you get Jim Keller shrinking process technology,
you get three nanometers instead of five or seven or 10
or 28 or whatever.
And so that marches forward and that provides improvements.
You get architectural level performance.
And so the TPU with a matrix multiply unit
and a systolic array is much more efficient
than having a scaler or a core doing multiplies and ads
and things like that.
You then get system level improvements.
So how you talk to memory, how you talk
across a cluster of machines, how you scale out,
how you have fast interconnects between machines,
you then get system level programming models.
So now that you have all this hardware, how do you utilize it?
You then have algorithmic breakthroughs where you say,
hey, wow, cool.
Instead of training in a resident 50 and a week,
I'm now training it in 25 seconds.
And it's a combination of new optimizers
and new just training regimens and different approaches
to train.
And all of these things come together
to push the world forward.
That was a beautiful exposition.
But if you were to force to bet all your money
on one of these, would you?
Why do we have to?
That's true.
Unfortunately, we have people working on all of this.
It's an exciting time, right?
So I mean, open the eye, I did this little paper
showing the algorithmic improvement you can get.
It's been improving exponentially.
I haven't quite seen the same kind of analysis
on other layers of the stack.
I'm sure it's also improving significantly.
It's a nice intuition build there.
I mean, there's a reason why Moore's Law,
that's the beauty of Moore's Law is somebody
writes a paper that makes a ridiculous prediction.
And it becomes reality, in a sense.
There's something about these narratives
when Chris Lattener on a silly little podcast
makes, bets all his money on a particular thing,
somehow it can have a ripple effect of actually becoming
real.
That's an interesting aspect of it,
because it might have been, we focused with Moore's Law,
most of the computing industry really, really focused
on the hardware.
I mean, software innovation, I don't
know how much software innovation there was in terms
of efficient.
What Intel give a bill takes away, right?
Yeah.
I mean, compilers improve significantly also.
Well, not really.
So actually, I mean, I'm joking about how software's gotten
slower pretty much as fast as hardware got better,
at least through the 90s.
There's another joke, another law in compilers,
which is called, I think it's called Probstein's Law, which
is compilers double the performance of any given code
every 18 years.
So they move slowly.
Yeah, well, so well, yeah, it's exponential also.
Yeah, you're making progress.
But there, again, it's not about the power of compilers
is not just about how do you make the same thing go faster.
It's how do you unlock the new hardware?
A new chip came out.
How do you utilize it?
You say, oh, the programming model,
how do we make people more productive?
How do we have better error messages?
Even such mundane things like, how do I
generate a very specific error message about your code?
Actually makes people happy because they know how to fix it,
right?
And it comes back to, how do you help people get their job done?
Yeah.
And yeah, and in this world of exponentially increasing
smart toasters, how do you expand computing
to all these kinds of devices?
Do you see this world where just everything's
a computing surface?
You see that possibility?
Just everything's a computer?
Yeah, I don't see any reason that that couldn't be achieved.
Turns out that sand goes into glass,
and glass is pretty useful too.
And why not?
Why not?
So a very important question, then,
if we're living in a simulation and the simulation is running
a computer, what's the architecture of that computer,
do you think?
So you're saying is it a quantum system?
Yeah, this whole quantum discussion, is it needed?
Or can we run it with a RISC 5 architecture, a bunch of CPUs?
I think it comes down to the right tool for the job.
And what's the compiler?
Yeah, exactly.
That's my question.
Did I get that job?
Made the universe compiler.
And so there, as far as we know, quantum systems
are the bottom of the pile of turtles so far.
And so we don't know efficient ways
to implement quantum systems without using quantum computers.
Yeah, and that's totally outside of everything
we've talked about, quantum computer.
But who runs that quantum computer?
Yeah.
Right, so if we really are living in a simulation,
then is it bigger quantum computers?
Is it different ones?
Like, how does that work out?
How does that scale?
Well, it's the same size.
It's the same size.
But then the thought of the simulation
is that you don't have to run the whole thing that we
humans are cognitively very limited.
You do checkpoints.
You do checkpoints, yeah.
And the point at which we human, so you basically
do minimal amount of what is it, the SWIFT does on right,
copy on right.
So you only adjust the simulation every time.
Parallel universe theories, right?
And so every time a decision is made,
somebody opens the shorting in your box, then there's a fork.
And then this could happen.
And then thank you for considering the possibility.
But yeah, so it may not require the entirety of the universe
to simulate it.
But it's interesting to think about as we
create these higher and higher fidelity systems.
But I do want to ask on the quantum computer side,
because everything we've talked about with your work
with sci-fi, with compilers, none of that
includes quantum computers, right?
That's true.
So do you ever thought about this whole serious engineering
work of quantum computers looks like of compilers,
of architectures, all of that kind of stuff?
So I've looked at it a little bit.
I know almost nothing about it, which means that at some point
I will have to find an excuse to get involved,
because that's how it works.
Do you think that's the thing to be,
with your little tingly senses of the timing of one
to be involved, is it not yet?
Well, so the thing I do really well
is I jump into messy systems and figure out how to make them,
figure out what the truth in the situation is,
try to figure out what the unifying theory is,
how to factor the complexity, how
to find a beautiful answer to a problem that has been well
studied and lots of people have bashed their heads against it.
I don't know that quantum computers are mature enough
and accessible enough to be figured out yet.
And I think the open question with quantum computers
is, is there a useful problem that
gets solved with a quantum computer that
makes it worth the economic cost of having one of these things
and having legions of people that set it up?
You go back to the 50s, and there's
the projections of the world will only need seven computers.
Well, and part of that was that people hadn't figured out
what they're useful for.
What are the algorithms we want to run?
What are the problems that get solved?
And this comes back to, how do we make the world better,
either economically or making somebody's life better
or solving a problem that wasn't solved before, things
like this?
And I think that just we're a little bit too early
in that development cycle because it's still
literally a science project, not a negative connotation.
It's literally a science project.
And the progress there is amazing.
And so I don't know if it's 10 years away,
if it's two years away, exactly where
that breakthrough happens.
But you look at machine learning,
we went through a few winners before the AlexNet transition,
and then suddenly it had its breakout moment.
And that was the catalyst that then drove the talent
flocking into it.
That's what drove the economic applications of it.
That's what drove the technology to go faster
because you now have more minds thrown at the problem.
This is what caused a serious knee and deep learning
and the algorithms that we're using.
And so I think that's what quantum needs to go through.
And so right now it's in that formidable finding itself,
getting literally the physics figured out.
And then it has to figure out the application that
makes this useful right now.
But I'm not skeptical that I think that will happen.
I think it's just 10 years away, something like that.
I forgot to ask, what programming language
do you think the simulation is written in?
Probably Lisp.
So not Swift.
Like if you were to bet, I'll just leave it at that.
So I mean, we've mentioned that you
worked with all these companies, we've
talked about all these projects.
If we just step back and zoom out about the way you did that
work, and we look at COVID times,
this pandemic we're living through,
that may, if I look at the way Silicon Valley folks are
talking about it, the way MIT is talking about it,
this might last for a long time.
Not just the virus, but the remote nature.
The economic impact.
I mean, yeah, it's going to be a mess.
Do you think, what's your prediction?
I mean, from sci-fi to Google to just all the places
you worked in, just Silicon Valley,
you're in the middle of it, how is this whole place
going to change?
Yeah, so I mean, I really can only
speak to the tech perspective in that bubble.
I think it's going to be really interesting,
because the zoom culture of being remote and on video chat
all the time has really interesting effects on people.
So on the one hand, it's a great normalizer.
It's a normalizer that I think will
help communities of people that have traditionally
been underrepresented, because now you're
taking, in some cases, a face-off,
because you don't have to have a camera going.
And so you can have conversations
without physical appearance being part of the dynamic,
which is pretty powerful.
You're taking remote employees that have already been remote,
and you're saying you're now on the same level
and footing as everybody else.
Nobody gets whiteboards.
You're not going to be the one person that
doesn't get to be participating in the whiteboard conversation,
and that's pretty powerful.
You've got, you're forcing people
to think asynchronously, in some cases,
because it's hard to just get people physically together,
and the bumping into each other forces people
to find new ways to solve those problems.
And I think that that leads to more inclusive behavior,
which is good.
On the other hand, it's also, it just sucks, right?
And so the actual communication just
sucks being not with people on a daily basis
and collaborating with them.
All of that, right?
I mean, everything, this whole situation is terrible.
What I meant primarily was the, I
think that most humans like working physically with humans.
I think this is something that not everybody,
but many people are programmed to do.
And I think that we get something out of that
that is very hard to express, at least for me.
And so maybe this isn't true of everybody.
But, and so the question to me is,
when you get through that time of adaptation,
you get out of March and April, and you get into December,
and you get into next March, if it's not changed, right?
It's already terrifying.
Well, you think about that, and you think about,
what is the nature of work?
And how do we adapt?
And humans are very adaptable species, right?
We can learn things.
And when we're forced to, and there's
a catalyst to make that happen.
And so what is it that comes out of this,
and are we better or worse off?
I think that you look at the Bay Area,
housing prices are insane.
Well, why?
Well, there's a high incentive to be physically located,
because if you don't have proximity,
you end up paying for it in commute, right?
And there has been huge social pressure in terms of,
like you will be there for the meeting, right?
Or whatever scenario it is.
And I think that's gonna be way better.
I think it's gonna be much more the norm
to have remote employees, and I think
this is gonna be really great.
Do you have friends, or do you hear of people moving?
I know one family friend that moved.
They moved back to Michigan, and they were a family
with three kids living in a small apartment,
and like we're going insane, right?
And they're in tech, husband works for Google.
So first of all, friends of mine
are in the process of, or are,
have already lost the business.
The thing that represents their passion, their dream.
It could be small entrepreneur projects,
but it could be large businesses,
like people that run gyms, like do.
Restaurants, like tons of things, yeah.
But also, people look at themselves in the mirror
and ask the question of like,
what do I wanna do in life?
For some reason, they haven't done it until COVID.
They really ask that question,
and that results often in moving,
or leaving the company with starting your business,
or transitioning to a different company.
Do you think we're gonna see that a lot?
Like in the-
Well, I can't speak to that.
I mean, we're definitely gonna see it
at a higher frequency than we did before,
just because I think what you're trying to say is,
there are decisions that you make yourself,
and big life decisions that you make yourself,
and like I'm gonna like quit my job and start a new thing.
There's also decisions that get made for you.
Like, I got fired from my job.
What am I gonna do, right?
And that's not a decision that you think about,
but you're forced to act, okay?
And so, I think that those,
you're forced to act kind of moments
where like, you know, global pandemic comes
and wipes out the economy,
and now your business doesn't exist.
I think that does lead to more reflection, right?
Because you're less anchored on what you have,
and it's not a what do I have to lose
versus what do I have to gain AB comparison.
It's more of a fresh slate.
Cool, I could do anything now.
Do I wanna do the same thing I was doing?
Did that make me happy?
Is this now time to go back to college
and take a class and learn a new skill?
Is this a time to spend time with family?
If you can afford to do that,
is this time to like, you know,
literally move into the parents, right?
I mean, all these things that were not normative before
suddenly become, I think, very,
the value system has changed.
And I think that's actually a good thing
in the short term, at least,
because it leads to, you know,
there's kind of been an over-optimization
along one set of priorities for the world,
and now maybe we'll get to a more balanced
and more interesting world
where people are doing different things,
and I think it could be good.
I think there could be more innovation
that comes out of it, for example.
What do you think about all the social chaos
when in the middle of like-
It sucks.
All right, you think it's, let me ask you,
I hope, do you think it's all gonna be okay?
Well, I think humanity will survive.
The, from the next extension,
like we're not all gonna kill, yeah, well.
Yeah, I don't think the virus is gonna kill all the humans.
I don't think all the humans are gonna kill all the humans.
I think that's unlikely, but I look at it as
progress requires a catalyst, right?
So you need a reason for people
to be willing to do things that are uncomfortable.
I think that the U.S., at least,
but I think the world in general,
is a pretty unoptimal place to live in for a lot of people.
And I think that what we're seeing right now
is we're seeing a lot of unhappiness.
And because of all the pressure,
because of all the badness in the world that's coming together,
it's really kind of igniting some of that debate
that should have happened a long time ago, right?
I mean, I think that we'll see more progress.
You're asking about offline,
you're asking about politics,
and wouldn't be great if politics moved faster
because there's all these problems in the world
and we can move it.
Well, people are inherently conservative.
And so if you're talking about conservative people,
particularly if they have heavy burdens on their shoulders
because they represent literally thousands of people,
it makes sense to be conservative.
But on the other hand, when you need change,
how do you get it?
The global pandemic will probably lead to some change.
And it's not a directed plan,
but I think that it leads to people
asking really interesting questions,
and some of those questions
should have been asked a long time ago.
Well, let me know if you've observed this as well.
Something that's bothered me
in the machine learning community,
I'm guessing it might be prevalent in other places,
is something that feels like in 2020,
increased level of toxicity.
Like, people are just quicker to pile on,
to just harsh on each other,
to mob, pick a person that screwed up
and make it a big thing.
And is there something that we can,
have you observed that in other places?
Is there some way out of this?
I think there's an inherent thing in humanity
that's kind of an us versus them thing,
which is that you wanna succeed and how do you succeed?
Well, it's relative to somebody else.
And so what's happening,
at least in some part, is that with the internet
and with online communication,
the world's getting smaller, right?
And so we're having some of the social ties
of like my town versus your town's football team, right?
Turn into much larger and yet shallower problems.
And people don't have time, the incentives.
So if the clickbait and like all these things
kind of really, really feed into this machine,
and I don't know where that goes.
Yeah, I mean, the reason I think about that,
I mentioned to you this offline a little bit,
but I have a few difficult conversations scheduled,
some of them political related,
some of them within the community,
difficult personalities that went through some stuff.
I mean, one of them I've talked before,
I will talk again is Yalma Koon.
You gotta have a little crap on Twitter
for talking about a particular paper
and the bias within a data set.
And then there's been a huge in my view,
and I'm only comfortable saying it,
irrational over exaggerated pile on on his comments
because he made pretty basic comments
about the fact that if there's bias in the data,
there's going to be bias in the results.
So we should not have bias in the data,
but people piled on to him because he said
he trivialized the problem of bias,
like it's a lot more than just bias in the data,
but like, yes, that's a very good point,
but that's-
That's not what he was saying.
That's not what he was saying.
And the response, like the implied response
that he's basically sexist and racist
is something that completely drives away
the possibility of nuanced discussion.
One nice thing about like a pocket long form
conversation is you can talk it out,
you can lay your reasoning out,
and even if you're wrong,
you can still show that you're a good human being
underneath it.
You know, your point about,
you can't have a productive discussion.
Well, how do you get to that point where people can turn,
they can learn, they can listen, they can think,
they can engage versus just being a shallow like,
like and then keep moving, right?
And I don't think that progress really comes from that, right?
And I don't think that one should expect that.
I think that you'd see that as reinforcing
individual circles and the us versus them thing.
And I think that's fairly divisive.
Yeah, I think there's a big role in,
like the people that bother me most on Twitter
when I observe things is not the people
who get very emotional and angry like over the top.
It's the people who like prop them up.
It's all the, I think what should be the,
we should teach each other is to be sort of empathetic.
The thing that it's really easy to forget,
particularly on like Twitter or the internet or the email,
is that sometimes people just have a bad day, right?
You have a bad day or you're like,
I've been in this situation where it's like between meetings,
like fire off a quick response in emails
because I want to like help get something unblocked.
Phrase it really objectively wrong.
I screwed up.
And suddenly this is now something that sticks with people.
And it's not because they're bad.
It's not because you're bad.
Just psychology of like you said a thing,
it sticks with you.
You didn't mean it that way,
but it really impacted somebody
because the way they interpreted it.
And this is just an aspect of working together as humans.
And I have a lot of optimism in the long term,
the very long term about what we as humanity can do,
but I think that's gonna be, it's just always a rough ride.
And you came into this by saying like,
what does COVID and all the social strife
that's happening right now mean?
And I think that it's really bad in the short term,
but I think it will lead to progress.
And for that, I'm very thankful.
Yeah, it's painful in the short term though.
Well, yeah, I mean, people are out of jobs,
like some people can't eat, like it's horrible.
And but it's progress, so we'll see what happens.
I mean, the real question is when you look back 10 years,
20 years, 100 years from now,
how do we evaluate the decisions
that are being made right now?
I think that's really the way you can frame that
and look at it and you say,
you integrate across all the short term
horribleness that's happening
and you look at what that means.
And is the improvement across the world
or the regression across the world,
significant enough to make it a good or a bad thing?
I think that's the question.
Yeah, and for that, it's good to study history.
I mean, one of the big problems for me right now
is I'm reading the rise and fall of the third Reich.
Okay, light reading.
So it's everything is just, I just see parallels
and I mean, you have to be really careful
not to overstep it,
but just the thing that worries me the most
is the pain that people feel when a few things combine,
which is like economic depression,
which is quite possible in this country.
And then just being disrespected in some kind of way,
which the German people were really disrespected
by most of the world,
like in a way that's over the top,
that something can build up
and then all you need is a charismatic leader
to go either positive or negative
in both work, as long as they're charismatic.
And there's...
It's taking advantage of, again,
that inflection point that the world's in
and what they do with it could be good or bad.
And so it's a good way to think about times now,
like on the individual level,
what we decide to do is when history is written,
30 years from now, what happened in 2020,
probably history is gonna remember 2020.
Yeah, I think so.
Either for good or bad.
And it's up to us to write it, so it's good.
Well, one of the things I've observed
that I find fascinating is most people act
as though the world doesn't change.
You make decisions knowingly, right?
You make a decision where you're predicting the future
based on what you've seen in the recent past.
And so if something's always been,
it's rained every single day,
then of course you expect it to rain today too, right?
On the other hand, the world changes all the time,
constantly, like for better and for worse,
right?
And so the question is, if you're interested
in something that's not right,
what is the inflection point that led to a change?
And you can look to history for this.
Like what is the catalyst that led to that explosion
that led to that bill that led to the,
like you can kind of work your way backwards from that.
And maybe if you pull together the right people
and you get the right ideas together,
you can actually start driving that change
and doing it in a way that's productive
and hurts fewer people.
Yeah, like a single person, a single event can turn
all of this. Yeah, absolutely.
Everything starts somewhere and often
it's a combination of multiple factors,
but yeah, these things can be engineered.
That's actually the optimistic view that...
I'm a long-term optimist on pretty much everything
and human nature, you know,
we can look to all the negative things that humanity has,
all the pettiness and all the like self-servingness
and just the cruelty, right?
The biases, just humans can be very horrible.
But on the other hand, we're capable of amazing things.
And the progress across, you know,
100-year chunks is striking.
And even across decades, we've come a long ways
and there's still a long ways to go,
but that doesn't mean that we've stopped.
Yeah, the kind of stuff we did in the last 100 years
is unbelievable.
It's kind of scary to think what's gonna happen
in this 100 years.
It's scary, like exciting.
Like scary in a sense that it's kind of sad
that the kind of technology is gonna come out
in 10, 20, 30 years.
We're probably too old to really appreciate
because you don't grow up with it.
It'll be like kids these days with their virtual reality
and their TikToks and stuff like this.
Like, how does this thing?
And like, come on, give me my, you know, static photo.
My Commodore 64, yeah.
Yeah, exactly.
Okay, sorry, we kind of skipped over.
But let me ask on, you know,
the machine learning world has been kind of inspired,
their imagination captivated with GPT-3
and these language models.
I thought it'd be cool to get your opinion on it.
What's your thoughts on this exciting world of,
it connects to computation actually,
is of language models that are huge
and take multiple, many, many computers,
not just to train, but to also do inference on.
Sure.
Well, I mean, it depends on what you're speaking to there.
But I mean, I think that there's been
a pretty well understood maximum deep learning
that if you make the model bigger
and you shove more data into it,
assuming you train it right
and you have a good model architecture,
that you'll get a better model out.
And so on one hand, GPT-3 was not that surprising.
On the other hand, a tremendous amount of engineering
went into making it possible.
The implications of it are pretty huge.
I think that when GPT-2 came out,
there was a very provocative blog post from OpenAI
talking about, we're not gonna release it
because of the social damage it could cause
if it's misused.
I think that's still a concern.
I think that we need to look at how technology is applied
and well-meaning tools can be applied in very horrible ways
and they can have very profound impact on that.
I think that GPT-3 is a huge technical achievement.
And what will GPT-4 be?
It will probably be bigger, more expensive to train.
Really cool architectural tricks.
What do you think?
Is there, I don't know how much thought you've done
on distributed computing.
Is there some technical challenges that are interesting
that you're hopeful about exploring in terms of,
a system that, like a piece of code that,
would GPT-4 that might have, I don't know,
hundreds of trillions of parameters
which have to run on thousands of computers.
Is there some hope that we can make that happen?
Yeah, well, I mean, today you can write a check
and get access to a thousand TPU cores
and do really interesting large-scale training
and inference and things like that in Google Cloud,
for example, right?
So I don't think it's a question about scales,
it's a question about utility.
And when I look at the Transformer series
of architectures that the GPT series is based on,
it's really interesting to look at that
because they're actually very simple designs.
They're not recurrent.
The training regimens are pretty simple.
And so they don't really reflect like human brains, right?
But they're really good at learning language models
and they're unrolled enough that you can simulate
some recurrence, right?
And so the question I think about is, where does this take us?
So we can just keep scaling it,
have more parameters, more data, more things,
we'll get a better result for sure.
But are there architectural techniques
that can lead to progress at a faster pace, right?
This is when, you know, how do you get,
instead of just like making it a constant time bigger,
how do you get like an algorithmic improvement
out of this, right?
And whether it be a new training regimen,
if it becomes sparse networks, for example,
human brains sparse, all these networks are dense.
The connectivity patterns can be very different.
I think this is where I get very interested
and I'm way out of my league on the deep learning side
of this, but I think that could lead to big breakthroughs.
When you talk about large scale networks,
one of the things that Jeff Dean likes to talk about
and he's given a few talks on is this idea
of having a sparsely gated mixture of experts
kind of a model where you have, you know,
different nets that are trained
and are really good at certain kinds of tasks.
And so you have this distributor across a cluster
and so you have a lot of different computers
that end up being kind of locally specialized
in different demands.
And then when a query comes in, you gait it
and use learn techniques to route
to different parts of the network.
And then you utilize the compute resources
of the entire cluster by having specialization within it.
And I don't know where that goes
or when it starts to work,
but I think things like that
could be really interesting as well.
And then on the data side too, if you can think of data
selection as a kind of programming.
Yeah.
I mean, essentially, if you look at like
Karpathi talked about software 2.0,
I mean, in a sense, data is the programming.
Yeah, yeah.
So let me try to summarize Andre's position
really quick before I disagree with it.
Yeah.
So Andre Karpathi is amazing.
So this is nothing personal with him.
He's an amazing engineer.
And also a good blog post writer.
Yeah.
Well, he's a great communicator.
I mean, he's just an amazing person.
He's also really sweet.
So his basic premise is that software is suboptimal.
I think we can all agree to that.
He also points out that deep learning
and other learning-based techniques are really great
because you can solve problems in more structured ways
with less like ad hoc code that people write out
and don't write test cases for in some cases.
And so they don't even know if it works in the first place.
And so if you start replacing systems of imperative code
with deep learning models, then you get a better result.
And I think that he argues that software 2.0
is a pervasively learned set of models
and you get away from writing code.
And he's given talks, he talks about swapping over
more and more and more parts of the code to being learned
and driven that way.
I think that works.
And if you're predisposed to liking machine learning,
then I think that that's definitely a good thing.
I think this is also good for accessibility in many ways
because certain people are not gonna write
C code or something.
And so having a data-driven approach
to do this kind of stuff, I think can be very valuable.
On the other hand, there are huge trade-offs
and it's not clear to me that software 2.0 is the answer.
And probably Andre wouldn't argue
that it's the answer for every problem either.
But I look at machine learning
as not a replacement for software 1.0.
I look at it as a new programming paradigm.
And so programming paradigms, when you look across
across domains is structured programming
where you go from go-tos to if-then-else
or functional programming from Lisp.
And you start talking about higher order functions
and values and things like this.
Or you talk about object-oriented programming.
You're talking about encapsulation,
subclassing, inheritance.
You start talking about generic programming
where you start talking about code reuse
through specialization and different type instantiations.
When you start talking about differentiable programming,
something that I am very excited about
in the context of machine learning,
talking about taking functions
and generating variants like the derivative
of another function.
Like that's a programming paradigm
that's very useful for solving certain classes of problems.
Machine learning is amazing
at solving certain classes of problems.
Like you're not gonna write a cat detector
or even a language translation system by writing C code.
That's not a very productive way to do things anymore.
And so machine learning is absolutely
the right way to do that.
In fact, I would say that learned models
are really one of the best ways to work
with the human world in general.
And so anytime you're talking about sensory input
of different modalities, anytime that you're talking
about generating things in a way that makes sense to a human,
I think that learned models are really, really useful.
And that's because humans are very difficult to characterize.
Yeah.
And so this is a very powerful paradigm
for solving classes of problems.
But on the other hand, imperative code is too.
You're not gonna write a bootloader for your computer
and with a deep learning model.
Deep learning models are very hardware intensive.
They're very energy intensive
because you have a lot of parameters
and you can provably implement any function
with a learned model like this has been shown.
But that doesn't make it efficient.
And so if you're talking about caring
about a few orders and magnitudes worth of energy usage,
then it's useful to have other tools in the toolbox.
It's also robustness too.
I mean, as a...
Yeah, exactly.
All the problems of dealing with data and bias in data,
all the problems of, you know, software 2.0
and one of the great things that Andre is arguing towards,
which I completely agree with him,
is that when you start implementing things
with deep learning, you need to learn from software 1.0
in terms of testing, continuous integration,
how you deploy, how do you validate all these things
and building systems around that
so that you're not just saying, like,
ooh, it seems like it's good, ship it, right?
Well, what happens when I regress something?
What happens when I make a classification that's wrong
and now I heard somebody, right?
I mean, all these things you have to reason about.
Yeah, but at the same time,
the bootloader that works for us humans
is looks awfully a lot like a neural network, right?
Yeah.
It's messy and you can cut out different parts of the brain.
There's a lot of this neural plasticity work
that shows that it's gonna adjust.
It's a really interesting question.
How much of the world's programming
could be replaced by software 2.0?
Like with...
Oh, well, I mean, it's provably true
that you could replace all of it.
Right, so then it's a question of the trade-offs.
Right, so anything that's a function, you can.
So it's not a question about if,
I think it's an economic question,
it's a what kind of talent can you get?
What kind of trade-offs in terms of maintenance, right?
Those kinds of questions,
I think what kind of data can you collect?
I think one of the reasons that I'm most interested in
machine learning as a programming paradigm
is that one of the things
that we've seen across computing in general
is that being laser focused on one paradigm
often puts you in a box, it's not super great.
And so you look at object-oriented programming.
Like it was all the rage in the early 80s
and like everything has to be objects
and people forgot about functional programming
even though it came first.
And then people rediscovered that,
hey, if you mix functional and object-oriented
and structure, like you mix these things together,
you can provide very interesting tools
that are good at solving different problems.
And so the question there is,
how do you get the best way to solve the problems?
It's not about whose tribe should win, right?
It's not about, that shouldn't be the question.
The question is, how do you make it
so that people can solve those problems the fastest?
And they have the right tools in their box
to build good libraries and they can solve these problems.
And when you look at that, that's like,
you look at reinforcement learning
as one really interesting sub-domain of this.
Reinforcement learning,
often you have to have the integration of a learned model
combined with your Atari
or whatever the other scenario it is
that you're working in.
You have to combine that thing
with the robot control for the arm, right?
And so now it's not just about that one paradigm,
it's about integrating that with all the other systems
that you have, including often legacy systems
and things like this, right?
And so to me, I think that the interesting thing to say
is like, how do you get the best out of this domain?
And how do you enable people to achieve things
that they otherwise couldn't do
without excluding all the good things
we already know how to do?
Right, but okay, this is a crazy question,
but we talked a lot about GPT-3,
but do you think it's possible that these language models
that in essence, in the language domain,
software 2.0 could replace some aspect of compilation,
for example, or do program synthesis,
replace some aspect of programming?
Yeah, absolutely.
So I think that the learned models in general
are extremely powerful
and I think that people underestimate them.
Maybe you can suggest what I should do.
So of access to the GPT-3 API,
would I be able to generate Swift code, for example?
Do you think that could do something interesting
and would work?
So GPT-3 is probably not trained on the right corpus.
So it probably has the ability to generate some Swift,
I bet it does.
It's probably not gonna generate a large enough body
of Swift to be useful,
but like taking the next step further,
like if you had the goal of training something like GPT-3
and you wanted to try to generate source code, right?
It could definitely do that.
Now the question is,
how do you express the intent of what you want filled in?
You can definitely like write a scaffolding of code
and say fill in the hole
and sort of put in some for loops
or put in some classes or whatever
and the power of these models is impressive,
but there's an unsolved question, at least unsolved to me,
which is how do I express the intent of what to fill in?
Right, and kind of what you'd really want to have
and I don't know that these models are up to the task
is you want to be able to say,
here's a scaffolding
and here are the assertions at the end
and the assertions always pass.
And so you want a generative model on the one hand, yes.
Oh, that's fascinating, yeah.
Right, but you also want some loopback,
some reinforcement learning system or something
where you're actually saying like,
I need to hill climb towards something
that is more correct and I don't know that we have that.
So it would generate not only a bunch of the code,
but like the checks that do the testing,
it would generate the tests.
I think the humans would generate the tests, right?
Oh, okay.
But it would be fascinating if-
Well, the tests are the requirements.
Yes, but the, okay, so-
Because you have to express to the model
what you want to, you don't just want gibberish code.
Look at how compelling this code looks.
You want a story about four horned unicorns or something.
Well, okay, so exactly, but that's human requirements,
but then I thought it's a compelling idea
that the GPT-4 model could generate checks
like that are more high fidelity
that check for correctness.
Because the code it generates,
like say I ask it to generate a function
that gives me the Fibonacci sequence.
Sure.
I don't like-
So decompose the problem, right?
So you have two things.
You have, you need the ability to generate
syntactically correct Swift code
that's interesting, right?
I think GPT series of model architectures can do that.
But then you need the ability to add the requirements.
So generate Fibonacci.
The human needs to express that goal.
We don't have that language that I know of.
No, I mean, it can generate stuff.
Have you seen what GPT-3 can generate?
You can say, I mean, there's interface stuff,
like it can generate HTML.
It can generate basic for loops that give you like-
Right, but pick HTML.
How do I say I want Google.com?
Well, no, you could say-
Or not literally Google.com.
How do I say I want a webpage
that's got a shopping cart into this and that?
Yeah, it does that.
I mean, so, okay, so just,
I don't know if you've seen these demonstrations,
but you type in, I want a red button
with the text that says hello,
and you typed that natural language
and it generates the correct HTML.
I've done this demo.
It's kind of compelling.
So you have to prompt it with similar kinds of mappings.
Of course, it's probably hand-picked.
I have to experiment that probably,
but the fact that you can do that once,
even out of like 20, is quite impressive.
Again, that's very basic.
Like the HTML is kind of messy and bad.
But yes, the intent is,
the idea is the intent to specify the natural language.
Yeah, so I have not seen that.
That's really cool.
Yeah, but the question is the correctness of that.
Like visually you can check,
oh, the button is red,
but for more complicated functions
where the intent is harder to check,
this goes into like NP completeness kind of things.
Like I want to know that this code is correct
and generates a giant thing
that does some kind of calculation.
It seems to be working.
It's interesting to think like,
should the system also try to generate checks
for itself for correctness?
Yeah, I don't know.
And this is way beyond my experience.
The thing that I think about
is that there doesn't seem to be
a lot of equational reasoning going on.
There's a lot of pattern matching and filling in
and kind of propagating patterns
that have been seen before into the future
and into the generated result.
And so if you want to get correctness,
you kind of need to improve in kind of things
and like higher level logic.
And I don't know that,
you could talk to Yon about that
and see what the Bright Minds are thinking about right now.
But I don't think the GPT is in that vein.
It's still really cool.
Yeah, it's surprising.
Who knows, maybe reasoning is-
It's overrated.
Yeah, it's overrated.
Right, I mean, do we reason?
Yeah.
How do you tell, right?
Are we just pattern matching
based on what we have
and then reverse justifying it to ourselves?
Yeah, exactly, reverse.
So like I think what the neural networks are missing
and I think GPT form might have
is to be able to tell stories to itself about what it did.
Well, that's what humans do, right?
I mean, you talk about like network explainability, right?
And we give, no, no, that's a hard time about this.
But humans don't know why we make decisions.
We have this thing called the intuition
and then we try to like say this feels like the right thing,
but why, right?
And you know, you wrestle with that
when you're making hard decisions.
And is that science?
Not really.
Let me ask you about a few high level questions, I guess,
is you've done a million things in your life
and been very successful.
A bunch of young folks listen to this, ask for advice
from successful people like you.
If you were to give advice to somebody,
you know, another graduate student
or some high school student about pursuing a career
in computing or just advice about life in general,
is there some words of wisdom you can give them?
So I think you come back to change
and you know, profound leaps happen
because people are willing to believe
that change is possible and that the world does change
and are willing to do the hard thing
that it takes to make change happen.
And whether it be implementing a new programming language
or implementing a new system
or implementing a new research paper,
designing a new thing, moving the world forward
in science and philosophy, whatever,
it really comes down to somebody
who's willing to put in the work, right?
And you have, the work is hard
for a whole bunch of different reasons,
one of which is you, it's work, right?
And so you have to have the space in your life
in which you can do that work,
which is why going to grad school
can be a beautiful thing for certain people.
But also there's a self-doubt that happens.
Like you're two years into a project,
is it going anywhere, right?
Well, what do you do?
Do you just give up because it's hard?
Well, no, I mean, some people like suffering.
And so you plow through it.
The secret to me is that you have to love what you're doing
and follow that passion
because when you get to the hard times,
that's when, you know, if you love what you're doing,
you're willing to kind of push through.
And this is really hard
because it's hard to know what you will love doing
until you start doing a lot of things.
And so that's why I think that,
particularly early in your career, it's good to experiment.
Do a little bit of everything.
Go take the survey class on, you know,
for the first half of every class
in your upper division, you know, lessons
and just get exposure to things
because certain things will resonate with you.
And you'll find out, wow, I'm really good at this.
I'm really smart at this.
Well, it's just because it works with the way your brain.
And when something jumps out,
I mean, that's one of the things that people often ask about
is like, well, I think there's a bunch of cool stuff out there.
Like, how do I pick the thing?
Like, how do you hook in your life?
How did you just hook yourself in and stuck with it?
Well, I got lucky.
I mean, I think that many people forget
that a huge amount of it or most of it is luck, right?
So let's not forget that.
So for me, I fell in love with computers early on
because they spoke to me, I guess.
What language did they speak to?
Basic.
Basic, yeah.
But then it was just kind of following
a set of logical progressions,
but also deciding that something that was hard
was worth doing and a lot of fun, right?
And so I think that that is also something
that's true for many other domains,
which is if you find something that you love doing
that's also hard, if you invest yourself in it
and add value to the world,
then it will mean something, generally, right?
And again, that can be a research paper,
that can be a software system,
that can be a new robot,
that can be, there's many things that that can be,
but a lot of it is like real value comes
from doing things that are hard.
And that doesn't mean you have to suffer.
But it's hard.
I mean, you don't often hear that message.
Talked about it last time a little bit,
but it's one of my, not enough people talk about this.
It's beautiful to hear a successful person.
Well, in self-doubt and imposter syndrome,
and these are all things that successful people
suffer with as well,
particularly when they put themselves
in a point of being uncomfortable,
which I like to do now and then,
just because it puts you in learning mode.
Like if you wanna, if you wanna grow as a person,
put yourself in a room with a bunch of people
that know way more about whatever you're talking about
than you do and ask dumb questions.
And guess what?
Smart people love to teach, often, not always, but often.
And if you listen, if you're prepared to listen,
if you're prepared to grow,
if you're prepared to make connections,
you can do some really interesting things.
And I think that a lot of progress is made by people
who kind of hop between domains now and then,
because they bring a perspective into a field
that nobody else has,
if people have only been working in that field themselves.
We mentioned that the universe is kind of like a compiler.
You know, the entirety of it,
the whole evolution is kind of a kind of a compilation.
Maybe us human beings are kind of compilers.
Let me ask the old sort of question
that I didn't ask you last time,
which is what's the meaning of it all?
Is there a meaning?
Like if you asked a compiler, why?
What would a compiler say?
What's the meaning of life?
What's the meaning of life?
You know, I'm prepared for it not to mean anything.
Here we are, all biological things programmed to survive
and propagate our DNA.
And maybe the universe is just a computer
and you just go until entropy takes over the world
or takes over the universe and then you're done.
I don't think that's a very productive way
to live your life, if so.
And so I prefer to bias towards the other way,
which is saying the universe has a lot of value.
And I take happiness out of other people.
And a lot of times part of that's having kids,
but also the relationships you build with other people.
And so the way I try to live my life is like,
what can I do that has value?
How can I move the world forward?
How can I take what I'm good at
and like bring it into the world?
And how can I, one of these people that likes
to work really hard and be very focused
on the things that I do.
And so if I'm gonna do that,
how can it be in a domain that actually will matter?
Because a lot of things that we do,
we find ourselves in the cycle of like,
okay, I'm doing a thing, I'm very familiar with it.
I've done it for a long time.
I've never done anything else,
but I'm not really learning.
I'm not really, I'm keeping things going,
but there's a younger generation
that can do the same thing,
maybe even better than me, right?
Maybe if I actually step out of this
and jump into something I'm less comfortable with,
it's scary, but on the other hand,
it gives somebody else a new opportunity.
It also then puts you back in learning mode,
and that can be really interesting.
And one of the things I've learned is that
when you go through that,
that first you're deep into imposter syndrome,
but when you start working your way out,
you start to realize,
hey, well, there's actually a method to this.
And now I'm able to add new things
because I bring different perspective.
And this is one of the good things
about bringing different kinds of people together.
Diversity of thought is really important.
And if you can pull together people
that are coming at things from different directions,
you often get innovation.
And I love to see that, that aha moment
where you're like, we've really cracked this.
This is something nobody's ever done before.
And then if you can do it in the context
where it adds value, other people can build on it,
it helps move the world,
then that's what really excites me.
So that kind of description of the magic
of the human experience,
do you think we'll ever create that in like an AGI system?
Do you think we'll be able to create,
give AI systems the sense of meaning
where they operate in this kind of world
exactly in the way you've described,
which is they interact with each other,
they interact with us humans.
Sure, sure.
Well, so I mean, why are you being so a specist, right?
All right, so AGI versus Bionets,
you know, versus Bionets, you know,
what are we but machines, right?
We're just programmed to run our,
we have our objective function
that we were optimized for, right?
And so we're doing our thing.
We think we have purpose, but do we really, right?
I'm not prepared to say that those new fangled AGI's
have no soul just because we don't understand them, right?
And I think that would be, when they exist,
that would be very premature to look at a new thing
through your own lens without fully understanding it.
You might be just saying that
because AI systems in the future will be listening to this.
And then-
Oh yeah, exactly, you don't wanna say anything.
Please be nice to me, you know, when Skynet kills everybody,
please spare me.
So why is, why is look ahead thinking?
Yeah, but I mean, I think that people will spend a lot of time
worrying about this kind of stuff.
And I think that what we should be worrying about is,
how do we make the world better?
And the thing that I'm most scared about with AGI's
is not that necessarily the Skynet
will start shooting everybody with lasers
and stuff like that to use us for our calories.
The thing that I'm worried about is that humanity,
I think, needs a challenge.
And if we get into a mode of not having a personal challenge,
not having a personal contribution,
whether that be like, you know, your kids
and seeing what they grow into and helping guide them,
whether it be your community that you're engaged in,
you're driving forward, whether it be your work
and the things that you're doing and the people
you're working with and the products you're building
and the contribution there.
If people don't have a objective,
I'm afraid what that means.
And I think that this would lead to a rise
of the worst part of people, right?
Instead of people striving together
and trying to make the world better,
it could degrade into a very unpleasant world.
But I don't know, I mean, we hopefully have a long ways
to go before we discover that.
And unfortunately, we have pretty
on the ground problems with the pandemic right now.
And so I think we should be focused on that as well.
Yeah, ultimately, just as you said, you're optimistic.
I think it helps for us to be optimistic
because that's faking until you make it.
Yeah, well, and why not?
I mean, what's the other side, right?
So, I mean, I'm not personally a very religious person,
but I've heard people say like,
oh, yeah, of course I believe in God.
Of course I go to church because if God's real,
I want to be on the right side of that.
And if it's not real, it doesn't matter.
Yeah, it doesn't matter.
And so, you know, that's a fair way to do it.
Yeah, I mean, the same thing with nuclear deterrence,
all, you know, global warming, all these things,
all these threats, natural engineer pandemics,
all these threats we face,
I think it's paralyzing to be terrified
of all the possible ways we could destroy ourselves.
I think it's much better or at least productive
to be hopeful and to engineer defenses
against these things to engineer a future
where like, you know, see like a positive future
and engineer that future.
Yeah, well, and I think that's another thing
to think about as, you know, a human,
particularly if you're young
and trying to figure out what it is
that you want to be when you grow up like I am.
I'm always looking for that.
The question then is, how do you want to spend your time?
And right now there seems to be a norm
of being a consumption culture.
Like I'm going to watch the news
and revel in how horrible everything is right now.
I'm going to go find out about the latest atrocity
and find out all the details of like the terrible thing
that happened and be outraged by it.
You can spend a lot of time watching TV
and watching the news at home or whatever
people watch these days.
I don't know, but that's a lot of hours, right?
And those are hours
that if you're turned into being productive,
learning, growing, experiencing, you know,
when the pandemic's over, going exploring, right?
It leads to more growth.
And I think it leads to more optimism and happiness
because you're building, right?
You're building yourself, you're building your capabilities,
you're building your viewpoints,
you're building your perspective.
And I think that a lot of the consuming
of other people's messages leads
to kind of a negative viewpoint,
which you need to be aware of what's happening
because that's also important,
but there's a balance that I think focusing on creation
is a very valuable thing to do.
Yeah, so what you're saying is people should focus
on working on the sexiest field of all,
which is compiler design.
Exactly.
Hey, you can go work on machine learning
and be crowded out by the thousands of graduates
popping out of school that all want to do the same thing,
or you could work in the place that people overpay you
because there's not enough smart people working in it.
And here at the end of Moore's Law, according to some people,
actually the software's the hard part too.
I mean, optimization is truly, truly beautiful.
And also on the YouTube side or education side,
there's a, it'd be nice to have some material
that shows the beauty of compilers.
Yeah, yeah.
That's something.
So that's a call for people to create
that kind of content as well.
Chris, you're one of my favorite people to talk to.
It's such a huge honor that you would waste your time
talking to me.
I've always appreciated it.
Thank you so much.
Yeah, I'm in the bucket today.
The truth of it is you spent a lot of time talking to me
just on walks and other things like that.
So it's great to catch up with.
Thanks, man.
Thanks for listening to this conversation with Chris Latner.
And thank you to our sponsors.
Blinkist, an app that summarizes key ideas
from thousands of books.
Neuro, which is a maker of functional gum and mints
that supercharged my mind.
Masterclass, which are online courses from world experts.
And finally, Cash App,
which is an app for sending money to friends.
Please check out these sponsors in the description
to get a discount and to support this podcast.
If you enjoy this thing, subscribe on YouTube,
review it with five stars on Apple Podcast,
follow on Spotify, support on Patreon,
connect with me on Twitter, Alex Friedman.
And now let me leave you with some words from Chris Latner.
So much of language design is about trade-offs.
And you can't see those trade-offs
unless you have a community of people
that really represent those different points.
Thank you for listening and hope to see you next time.