Jim Keller: Moore’s Law, Microprocessors, Abstractions, and First Principles
嘉宾：Jim Keller
全文文稿 · 按对话顺序保留完整内容，深入理解上下文。
本篇目录12 章节
§01Transcript · 00:00–09:06§02Transcript · 09:08–17:16§03Transcript · 17:19–26:27§04Transcript · 26:30–34:46§05Transcript · 34:47–43:07§06Transcript · 43:08–51:42§07Transcript · 51:44–1:00:20§08Transcript · 1:00:22–1:08:52§09Transcript · 1:08:54–1:17:04§10Transcript · 1:17:05–1:25:07§11Transcript · 1:25:08–1:34:12§12Transcript · 1:34:15–1:34:37§01Transcript · 00:00–09:06
The following is a conversation with Jim Keller,
legendary microprocessor engineer
who has worked at AMD, Apple, Tesla, and now Intel.
He's known for his work on AMD K7, K8, K12,
and Zen microarchitectures, Apple A4 and A5 processors,
and co-author of the specification
for the x86-64 instruction set
and HyperTransport Interconnect.
He's a brilliant first principles engineer
and out of the box thinker,
and just an interesting and fun human being to talk to.
This is the Artificial Intelligence Podcast.
If you enjoy it, subscribe on YouTube,
give it five stars on Apple Podcast,
follow on Spotify, support it on Patreon,
or simply connect with me on Twitter,
Alex Friedman, spelled F-R-I-D-M-A-N.
I recently started doing ads
at the end of the introduction.
I'll do one or two minutes after introducing the episode
and never any ads in the middle
that can break the flow of the conversation.
I hope that works for you
and doesn't hurt the listening experience.
This show is presented by Cash App,
the number one finance app in the App Store.
I personally use Cash App to send money to friends,
but you can also use it to buy, sell,
and deposit Bitcoin in just seconds.
Cash App also has a new investing feature.
You can buy fractions of a stock, say $1 worth,
no matter what the stock price is.
Brokers services are provided by Cash App Investing,
a subsidiary of Square and member SIPC.
I'm excited to be working with Cash App
to support one of my favorite organizations called FIRST,
best known for their FIRST Robotics and Lego competitions.
They educate and inspire hundreds of thousands of students
in over 110 countries
and have a perfect rating at Charity Navigator,
which means that donated money
is used to maximum effectiveness.
When you get Cash App from the App Store or Google Play
and use code LEXPODCAST,
you'll get $10 and Cash App will also donate $10 to FIRST,
which again is an organization
that I've personally seen inspire girls and boys
to dream of engineering a better world.
And now here's my conversation with Jim Keller.
What are the differences and similarities
between the human brain and a computer
with the microprocessor at its core?
Let's start with the philosophical question perhaps.
- Well, since people don't actually understand
how human brains work, I think that's true.
- I think that's true.
- So it's hard to compare them.
Computers are, you know, there's really two things.
There's memory and there's computation, right?
And to date, almost all computer architectures
are global memory, which is a thing, right?
And then computation where you pull data
and you do relatively simple operations on it
and write data back.
So it's decoupled in modern computers.
And you think in the human brain,
everything's a mesh, a mess that's combined together.
- What people observe is there's, you know,
some number of layers of neurons
which have local and global connections.
And information is stored in some distributed fashion.
And people build things called neural networks in computers
where the information is distributed
in some kind of fashion.
You know, there's a mathematics behind it.
I don't know that the understanding of that is super deep.
The computations we run on those
are straightforward computations.
I don't believe anybody has said
a neuron does this computation.
So to date, it's hard to compare them, I would say.
- So let's get into the basics before we zoom back out.
How do you build a computer from scratch?
What is a microprocessor?
What is a microarchitecture?
What's an instruction set architecture?
Maybe even as far back as what is a transistor?
- So the special charm of computer engineering
is there's a relatively good understanding
of abstraction layers.
So down at the bottom, you have atoms.
And atoms get put together in materials
like silicon or dope silicon or metal.
And we build transistors on top of that.
We build logic gates, right?
And then functional units, like an adder, a subtractor,
an instruction parsing unit.
And then we assemble those into processing elements.
Modern computers are built out of probably 10 to 20
locally organic processing elements
or coherent processing elements.
And then that runs computer programs, right?
So there's abstraction layers.
And then software, there's an instruction set.
You run, and then there's assembly language, C,
C++, Java, JavaScript.
There's abstraction layers,
essentially from the atom to the data center, right?
So when you build a computer,
first there's a target, like what's it for?
Like how fast does it have to be?
Which today there's a whole bunch of metrics
about what that is.
And then in an organization of a thousand people
who build a computer,
there's lots of different disciplines
that you have to operate on.
Does that make sense?
And so--
- So there's a bunch of levels of abstraction.
In an organization like Intel, and in your own vision,
there's a lot of brilliance that comes in
at every one of those layers.
Some of it is science, some of it is engineering,
some of it is art.
What's the most, if you could pick favorites,
what's the most important, your favorite layer
on these layers of abstractions?
Where does the magic enter this hierarchy?
- I don't really care.
That's the fun, you know, I'm somewhat agnostic to that.
So I would say, for relatively long periods of time,
instruction sets are stable.
So the x86 instruction set, the ARM instruction set.
- What's an instruction set?
- So it says, how do you encode the basic operations?
Load, store, multiply, add, subtract, conditional branch.
There aren't that many interesting instructions.
Like if you look at a program and it runs,
90% of the execution is on 25 opcodes, 25 instructions.
And those are stable, right?
- What does it mean, stable?
- Intel architecture has been around for 25 years.
- It works.
- It works.
And that's because the basics are defined a long time ago.
Now, the way an old computer ran,
is you fetched instructions and you executed them in order.
Do the load, do the add, do the compare.
The way a modern computer works,
is you fetch large numbers of instructions, say 500.
And then you find the dependency graph
between the instructions.
And then you execute in independent units,
those little micrographs.
So a modern computer, like people like to say,
computers should be simple and clean.
But this turns out the market for simple,
complete, clean, slow computers is zero, right?
We don't sell any simple, clean computers.
Now you can, there's how you build it can be clean,
but the computer people want to buy,
that's say in a phone or a data center,
fetches a large number of instructions,
computes the dependency graph,
and then executes it in a way that gets the right answers.
- And optimize that graph somehow.
- Yeah, they run deeply out of order.
And then there's semantics around how memory ordering works
and other things work.
So the computer sort of has a bunch of bookkeeping tables.
It says, what order should these operations finish in
or appear to finish in?
But to go fast, you have to fetch a lot of instructions
and find all the parallelism.
Now there's a second kind of computer,
which we call GPUs today.
And I call it a difference.
There's found parallelism,
like you have a program
with a lot of dependent instructions.
You fetch a bunch and then you go figure out
the dependency graph and you issue instructions out of order.
That's because you have one serial narrative to execute,
which in fact can be done out of order.
- You call it a narrative?
- Yeah.
- Wow.
- Yeah, so humans think in serial narrative.
So read a book, right?
There's a sentence after sentence after sentence
and there's paragraphs.
Now you could diagram that.
Imagine you diagrammed it properly and you said,
which sentences could be read in any order
without changing the meaning, right?
- That's a fascinating question to ask of a book, yeah.
- Yeah, you could do that, right?
So some paragraphs could be reordered,
some sentences can be reordered.
§02Transcript · 09:08–17:16
You could say, he is tall and smart and X, right?
And it doesn't matter the order of tall and smart.
But if you say the tall man is wearing a red shirt,
what colors, you know, like you can create dependencies.
And so GPUs on the other hand,
run simple programs on pixels,
but you're given a million of them.
And the first order, the screen you're looking at
doesn't care which order you do it in.
So I call that given parallelism.
Simple narratives around the large numbers of things
where you can just say it's parallel
because you told me it was.
- So found parallelism where the narrative is sequential
but you discover like little pockets of parallelism versus--
- Turns out large pockets of parallelism.
- Large, so how hard is it to discover?
- Well, how hard is it?
That's just transistor count, right?
So once you crack the problem, you say,
here's how you fetch 10 instructions at a time.
Here's how you calculated the dependencies between them.
Here's how you describe the dependencies.
Here's, you know, these are pieces, right?
- So once you describe the dependencies,
then it's just a graph,
sort of it's an algorithm that finds,
well, what is that?
I'm sure there's a graph theory,
theoretical answer here that's solvable.
In general, programs, modern programs
that human beings write,
how much found parallelism is there in them?
- About 10x.
- What does 10x mean?
- So if you execute it in order--
- Versus, yeah.
- You would get what's called cycles per instruction
and it would be about, you know, three instructions,
three cycles per instruction
because of the latency of the operations and stuff.
And in a modern computer,
excuse it, but like 0.2, 0.25 cycles per instruction.
So it's about, we today find 10x.
And there's two things.
One is the found parallelism in the narrative, right?
And the other is to predictability of the narrative, right?
So certain operations, they do a bunch of calculations
and if greater than one, do this, else do that.
That decision is predicted in modern computers
to high 90% accuracy.
So branches happen a lot.
So imagine you have a decision
to make every six instructions,
which is about the average, right?
But you want to fetch 500 instructions,
figure out the graph and execute them all in parallel.
That means you have, let's say,
if you fix 600 instructions and it's every six,
you have to fetch,
you have to predict 99 out of 100 branches correctly
for that window to be effective.
- Okay, so parallelism,
you can't parallelize branches, or you can?
- No, you can predict--
- What does predict a branch mean?
Or what does predict a branch mean?
- So imagine you do a computation over and over,
you're in a loop.
So while n is greater than one, do.
And you go through that loop a million times.
So every time you look at the branch,
you say, it's probably still greater than one.
- And you're saying you could do that accurately.
- Very accurately.
Modern computer--
- My mind is blown.
How the heck do you do that?
Wait a minute.
- Well, you want to know?
This is really sad.
20 years ago, you simply recorded
which way the branch went last time
and predicted the same thing.
- Right.
- Okay.
- What's the accuracy of that?
- 85%.
So then somebody said, hey, let's keep a couple of bits
and have a little counter.
So when it predicts one way,
we count up and then pins.
So say you have a three bit counter.
So you count up and then you count down.
And if it's, you can use the top bit as the sign bit.
So you have a sign two bit number.
So if it's greater than one, you predict taken
and less than one, you predict not taken, right?
Or less than zero, whatever the thing is.
And that got us to 92%.
- Oh.
- Okay, now it gets better.
This branch depends on how you got there.
So if you came down the code one way,
you're talking about Bob and Jane, right?
And then said, is just Bob like Jane?
It went one way.
But if you're talking about Bob and Jill,
does Bob like Jane?
You go a different way, right?
So that's called history.
So you take the history and a counter.
That's cool.
But that's not how anything works today.
They use something that looks a little
like a neural network.
So modern, you take all the execution flows
and then you do basically deep pattern recognition
of how the program is executing.
And you do that multiple different ways.
And you have something that chooses what the best result is.
There's a little supercomputer inside the computer.
- That's trying to predict branching.
- That calculates which way branches go.
So the effective window that is worth finding grass
and gets bigger.
- Why was that gonna make me sad?
'Cause that's amazing.
- It's amazingly complicated.
- Oh, well.
- Well, here's the funny thing.
So to get to 85% took a thousand bits.
To get to 99% takes tens of megabits.
So this is one of those, to get the result,
to get from a window of say 50 instructions to 500,
it took three orders of magnitude
or four orders of magnitude more bits.
- Now, if you get the prediction of a branch wrong,
what happens then?
- You flush the pipe.
- You flush the pipe, so it's just the performance cost.
- But it gets even better.
So we're starting to look at stuff that says,
so they executed down this path
and then you had two ways to go,
but far, far away there's something
that doesn't matter which path you went.
So you took the wrong path, you executed a bunch of stuff.
Then you had the mispredicting, you backed it up,
but you remembered all the results you already calculated.
Some of those are just fine.
Like if you read a book and you misunderstand a paragraph,
you're understanding the next paragraph,
sometimes it's invariant to their understanding.
Sometimes it depends on it.
- And you can kind of anticipate that invariance.
- Yeah, well, you can keep track
of whether the data changed.
And so when you come back to a piece of code,
should you calculate it again or do the same thing?
- Okay, how much of this is art
and how much of it is science?
'Cause it sounds pretty complicated.
- Well, how do you describe a situation?
So imagine you come to a point in the road
where you have to make a decision.
And you have a bunch of knowledge about which way to go,
maybe you have a map.
So you wanna go the shortest way
or do you wanna go the fastest way
or you wanna take the nicest road.
So there's some set of data.
So imagine you're doing something complicated
like building a computer
and there's hundreds of decision points
all with hundreds of possible ways to go.
And the ways you pick interact in a complicated way.
- Right.
- And then you have to pick the right spot.
- Right, so that's-- - Or that's art or science,
I don't know.
- You avoided the question,
you just described the Robert Frost problem
of "Road Less Taken."
- Described the Robert Frost problem?
(laughs)
- That's what we do as computer designers,
it's all poetry.
- Okay. - Great.
Yeah, I don't know how to describe that
because some people are very good
at making those intuitive leaps.
It seems like there's combinations of things.
Some people are less good at it
but they're really good at evaluating the alternatives.
Right, and everybody has a different way to do it.
And some people can't make those leaps
but they're really good at analyzing it.
So when you see computers are designed by teams of people
who have very different skill sets
§03Transcript · 17:19–26:27
and a good team has lots of different kinds of people.
I suspect you would describe some of them as artistic.
- Right. - But not very many.
- Unfortunately, or fortunately.
- Unfortunately.
(laughs)
Well, you know, computer design's hard,
it's 99% perspiration.
- And-- - The 1% inspiration
is really important.
- But you still need the 99.
- Yeah, you gotta do a lot of work.
And then there are interesting things to do
at every level of that stack.
- At the end of the day,
if you run the same program multiple times,
does it always produce the same result?
Is there some room for fuzziness there?
- That's a math problem.
So if you run a correct C program,
the definition is every time you run it,
you get the same answer.
- Yeah, well, that's a math statement.
- But that's a language definitional statement.
So for years when people did,
when we first did 3D acceleration of graphics,
you could run the same scene multiple times
and get different answers.
- Right.
- Right, and then some people thought that was okay,
and some people thought it was a bad idea.
And then when the HPC world used GPUs for calculations,
they thought it was a really bad idea, okay?
Now, in modern AI stuff,
people are looking at networks
where the precision of the data is low enough
that the data is somewhat noisy.
And the observation is the input data is unbelievably noisy.
So why should the calculation be not noisy?
And people have experimented with algorithms
that say can get faster answers by being noisy.
Like as a network starts to converge,
if you look at the computation graph,
it starts out really wide and then it gets narrower.
And you can say, is that last little bit that important?
Or should I start the graph on the next rev
before we whittle it all the way down to the answer?
Right, so you can create algorithms that are noisy.
Now, if you're developing something
and every time you run it, you get a different answer,
it's really annoying.
And so most people think even today,
every time you run the program, you get the same answer.
- No, I know, but the question is,
that's the formal definition of a programming language.
- There is a definition of languages
that don't get the same answer, but people who use those.
You always want something 'cause you get a bad answer
and then you're wondering, is it because
of something in the algorithm or because of this?
And so everybody wants a little switch
that says no matter what, do it deterministically.
And it's really weird 'cause almost everything
going into modern calculations is noisy.
So why do the answers have to be so clear?
- Right, so where do you stand?
- I design computers for people who run programs.
So if somebody says, I want a deterministic answer,
most people want that.
- Can you deliver a deterministic answer,
I guess is the question.
- Yeah, hopefully, sure.
What people don't realize is you get a deterministic answer
even though the execution flow is very undeterministic.
So you run this program 100 times,
it never runs the same way twice, ever.
And the answer, it arrives at the same answer.
- But it gets the same answer every time.
- It's just amazing.
Okay, you've achieved in the eyes of many people,
a legend status as a chip architect.
What design creation are you most proud of?
Perhaps because it was challenging, because of its impact,
or because of the set of brilliant ideas
that were involved in bringing it to life.
- I find that description odd.
And I have two small children, and I promise you,
they think it's hilarious.
- This question, I do it for them.
- I'm really interested in building computers.
And I've worked with really, really smart people.
I'm not unbelievably smart.
I'm fascinated by how they go together,
both as a thing to do, and as an endeavor that people do.
- How people and computers go together?
- Yeah, like how people think and build a computer.
And I find sometimes that the best computer architects
aren't that interested in people,
or the best people managers aren't that good
at designing computers.
- So the whole stack of human beings is fascinating.
So the managers, the individual engineers.
- Yeah, yeah, so yeah, I said I realized
after a lot of years of building computers,
we sort of build them out of transistors,
logic gates, functional units, computational elements,
that you could think of people the same way.
So people are functional units.
And then you could think of organizational design
as a computer architectural problem.
And then it was like, oh, that's super cool,
'cause the people are all different,
just like the computational elements are all different.
And they like to do different things.
And so I had a lot of fun reframing
how I think about organizations.
- Just like with computers,
we were saying execution paths,
you can have a lot of different paths
that end up at the same good destination.
So what have you learned about the human abstractions,
from individual functional human units
to the broader organization?
What does it take to create something special?
- Well, most people don't think simple enough.
- All right, so do you know the difference
between a recipe and the understanding?
There's probably a philosophical description of this.
So imagine you're gonna make a loaf of bread.
The recipe says, get some flour, add some water,
add some yeast, mix it up, let it rise,
put it in a pan, put it in the oven.
It's a recipe.
Understanding bread, you can understand biology,
supply chains, grain grinders, yeast,
physics, thermodynamics.
There's so many levels of understanding there.
And then when people build and design things,
they frequently are executing some stack of recipes.
And the problem with that is the recipes
all have limited scope.
Like if you have a really good recipe book
for making bread, it won't tell you anything
about how to make an omelet.
But if you have a deep understanding of cooking,
then bread, omelets, sandwich,
there's a different way of viewing everything.
And most people, when you get to be an expert at something,
you're hoping to achieve deeper understanding,
not just a large set of recipes to go execute.
And it's interesting to walk groups of people
because executing recipes is unbelievably efficient
if it's what you want to do.
If it's not what you want to do, you're really stuck.
And that difference is crucial.
And everybody has a balance of, let's say,
deeper understanding of recipes.
And some people are really good at recognizing
when the problem is to understand something deeply.
Deeply.
Does that make sense?
- It totally makes sense.
Does every stage of development,
deep understanding on the team needed?
- Well, this goes back to the art versus science question.
- Sure.
- If you constantly unpack everything
for deeper understanding, you never get anything done.
And if you don't unpack understanding when you need to,
you'll do the wrong thing.
And then at every juncture, like human beings
are these really weird things because everything you tell
them has a million possible outputs.
And then they all interact in a hilarious way.
And then having some intuition about what do you tell them,
what do you do, when do you intervene, when do you not,
it's complicated.
- Right, so--
- It's essentially computationally unsolvable.
- Yeah, it's an intractable problem, sure.
Humans are a mess.
But with deep understanding, do you mean also
sort of fundamental questions of things like
what is a computer?
Or why?
Like the why question is why are we even building this?
Like of purpose?
Or do you mean more like going towards
the fundamental limits of physics,
sort of really getting into the core of the science?
- Well, in terms of building a computer,
think a little simpler.
So common practice is you build a computer.
And then when somebody says, I want to make it 10% faster,
you'll go in and say, all right,
I need to make this buffer bigger.
And maybe I'll add an ad unit.
Or I have this thing that's three instructions wide,
I'm gonna make it four instructions wide.
And what you see is each piece
§04Transcript · 26:30–34:46
gets incrementally more complicated.
And then at some point you hit this limit.
Like adding another feature or buffer
doesn't seem to make it any faster.
And then people say, well, that's because
it's a fundamental limit.
And then somebody else will look at it and say,
well, actually the way you divided the problem up
and the way that different features are interacting
is limiting you and it has to be rethought, rewritten.
So then you refactor it and rewrite it.
And what people commonly find is the rewrite
is not only faster, but half as complicated.
- From scratch?
- Yes.
- So how often in your career,
but just have you seen as needed,
maybe more generally to just throw the whole thing out?
- This is where I'm on one end of it,
every three to five years.
- Which end are you on?
- Rewrite more often.
- Rewrite, and three to five years is?
- If you want to really make a lot of progress
on computer architecture, every five years
you should do one from scratch.
- So where does the x86-64 standard come in?
How often do you?
- I wrote the, I was the co-author of that spec in '98.
That's 20 years ago.
- Yeah, so that's still around.
- The instruction set itself has been extended
quite a few times.
- Yes.
- And instruction sets are less interesting
than the implementation underneath.
There's been, on x86 architecture,
Intel's designed a few, AMD's designed a few,
very different architectures.
And I don't want to go into too much of the detail
about how often, but there's a tendency
to rewrite it every 10 years,
and it really should be every five.
- So you're saying you're an outlier in that sense?
- Rewrite more often.
Rewrite more often.
- Well, and here's the problem.
- Isn't that scary?
- Yeah, of course.
Well, scary to who?
- To everybody involved, because like you said,
repeating the recipe is efficient.
Companies want to make money,
well, no, individual engineers want to succeed,
so you want to incrementally improve,
increase the buffer from three to four.
- Well, this is where you get into
diminishing return curves.
I think Steve Jobs said this, right?
So every, you have a project,
and you start here, and it goes up,
and you have diminishing return.
And to get to the next level, you have to do a new one,
and the initial starting point will be lower
than the old optimization point, but it'll get higher.
So now you have two kinds of fear,
short-term disaster and long-term disaster.
- And you're-- - So grown-ups.
- Grown-ups.
- Right, like, you know, people with a quarter-by-quarter
business objective are terrified about changing everything.
And people who are trying to run a business
or build a computer for a long-term objective
know that the short-term limitations
block them from the long-term success.
So if you look at leaders of companies
that had really good long-term success,
every time they saw that they had to redo something,
they did.
- And so somebody has to speak up?
- Or you do multiple projects in parallel.
Like, you optimize the old one while you build a new one.
But the marketing guys are always like,
promise me that the new computer
is faster on every single thing.
And the computer architect says,
well, the new computer will be faster on the average.
But there's a distribution of results and performance,
and you'll have some outliers that are slower.
And that's very hard, 'cause they have one customer
who cares about that one.
- So speaking of the long-term, for over 50 years now,
Moore's Law has served, for me and millions of others,
as an inspiring beacon of what kind of amazing future
brilliant engineers can build.
I'm just making your kids laugh all of today.
- Yeah, that's great.
- So first, in your eyes, what is Moore's Law,
if you could define for people who don't know?
- Well, the simple statement was, from Gordon Moore,
was double the number of transistors every two years.
Something like that.
And then my operational model is,
we increase the performance of computers by 2x
every two or three years.
And it's wiggled around substantially over time.
And also, in how we deliver performance has changed.
- So, right, so the--
- The foundational idea was 2x the transistors
every two years.
The current cadence is something like,
they call it a shrink factor.
Like .6 every two years, which is not .5.
- But that's referring strictly, again,
to the original definition of just--
- Yeah, of transistor count.
- A shrink factor, just getting 'em smaller,
smaller, smaller.
- Well, it's for a constant chip area.
If you make the transistors smaller by .6,
then you get one over .6 more transistors.
- So can you linger on it a little longer?
What's the broader, what do you think should be
the broader definition of Moore's Law?
When you mentioned how you think of performance,
just broadly, what's a good way to think about Moore's Law?
- Well, first of all, I've been aware of Moore's Law
for 30 years.
- In which sense?
- Well, I've been designing computers for 40.
- You're just watching it before your eyes, kind of thing.
- Well, and somewhere where I became aware of it,
I was also informed that Moore's Law was gonna die
in 10 to 15 years.
And I thought that was true at first,
but then after 10 years, it was gonna die
in 10 to 15 years.
And then at one point, it was gonna die in five years,
and then it went back up to 10 years,
and at some point, I decided not to worry
about that particular prognostication
for the rest of my life, which is fun.
And then I joined Intel, and everybody said,
Moore's Law is dead.
And I thought, that's sad, 'cause it's the Moore's Law
company, and it's not dead, and it's always been gonna die.
And humans, like these apocryphal kind of statements,
like, we'll run out of food, or we'll run out of air,
or run out of room, or run out of something.
- Right, but it's still incredible that it's lived
for as long as it has.
And yes, there's many people who believe now
that Moore's Law is dead.
- You know, they can join the last 50 years
of people who had the same idea.
- Yeah, there's a long tradition.
But why do you think, if you can try to understand it,
why do you think it's not dead currently?
- Well, first, let's just think,
people think Moore's Law is one thing,
transistors get smaller.
But actually, under the sheets, there's literally
thousands of innovations, and almost all those innovations
have their own diminishing return curves.
So if you graph it, it looks like a cascade
of diminishing return curves.
I don't know what to call that.
But the result is an exponential curve.
Well, at least it has been.
So, and we keep inventing new things.
So if you're an expert in one of the things
on a diminishing return curve, right,
and you can see its plateau,
you will probably tell people, "Well, this is done."
Meanwhile, some other pile of people
are doing something different.
So that's just normal.
So then there's the observation of how small
could a switching device be?
So a modern transistor is something like
a thousand by a thousand by a thousand atoms, right?
And you get quantum effects down around two to 10 atoms.
So you can imagine a transistor as small
as 10 by 10 by 10.
So that's a million times smaller.
And then the quantum computational people
are working away at how to use quantum effects.
So--
- A thousand by a thousand by a thousand.
- Atoms.
- That's a really clean way of putting it.
- Well, a fan, like a modern transistor,
if you look at the fan, it's like 120 atoms wide,
but we can make that thinner.
And then there's a gate wrapped around it,
and then there's spacing.
There's a whole bunch of geometry.
And a competent transistor designer
could count both atoms in every single direction.
(laughing)
§05Transcript · 34:47–43:07
Like there's techniques now to already put down atoms
in a single atomic layer.
And you can place atoms if you want to.
It's just, you know, from a manufacturing process,
if placing an atom takes 10 minutes
and you need to put 10 to the 23rd atoms together
to make a computer, it would take a long time.
So the methods are both shrinking things
and then coming up with effective ways
to control what's happening.
- Manufacture stably and cheaply.
- Yeah.
So the innovation stack's pretty broad.
You know, there's equipment, there's optics,
there's chemistry, there's physics,
there's material science, there's metallurgy.
There's lots of ideas about when you put
different materials together, how do they interact?
Are they stable?
Is it stable over temperature?
Like are they repeatable?
You know, there's like literally thousands
of technologies involved.
- But just for the shrinking, you don't think
we're quite yet close to the fundamental limits of physics.
- I did a talk on Moore's Law and I asked
for a roadmap to a path of 100 and after two weeks,
they said we only got to 50.
- 100 what, sorry?
- 100x shrink.
- 100x shrink?
We only got to 50?
- To 50 and I said, why don't you give it
another two weeks?
Well, here's the thing about Moore's Law, right?
So I believe that the next 10 or 20 years
of shrinking is gonna happen, right?
Now, as a computer designer, you have two stances.
You think it's going to shrink, in which case
you're designing and thinking about architecture
in a way that you'll use more transistors.
Or conversely, not be swamped by the complexity
of all the transistors you get, right?
You have to have a strategy, you know?
- So you're open to the possibility and waiting
for the possibility of a whole new army
of transistors ready to work.
- I'm expecting--
- Expecting.
- More transistors every two or three years
by a number large enough that how you think
about design, how you think about architecture
has to change.
Like imagine you build buildings out of bricks
and every year the bricks are half the size
or every two years.
Well, if you kept building bricks the same way,
you know, so many bricks per person per day,
the amount of time to build a building
would go up exponentially.
- Right.
- Right.
But if you said, I know that's coming,
so now I'm going to design equipment
that moves bricks faster, uses them better,
'cause maybe you're getting something out
of the smaller bricks, more strength, thinner walls,
you know, less material, efficiency out of that.
So once you have a roadmap with what's going to happen,
transistors, we're going to get more of them,
then you design all this collateral around it
to take advantage of it and also to cope with it.
Like that's the thing people don't understand,
it's like if I didn't believe in Moore's law
and then Moore's law transistors showed up,
my design teams were all drowned.
- So what's the hardest part of this in flood
of new transistors?
I mean, even if you just look historically
throughout your career, what's the thing,
what fundamentally changes when you add more transistors
in the task of designing an architecture?
- Well, there's two constants, right?
One is people don't get smarter.
- By the way, there's some science showing
that we do get smarter because of nutrition, whatever.
Sorry to bring that up.
- The Flint effect.
- Yes.
- Yeah, I'm familiar with it.
Nobody understands it, nobody knows if it's still going on.
So that's a--
- Or whether it's real or not, but yeah.
- I sort of--
- Anyway, but not exponentially.
- I would believe for the most part,
people aren't getting much smarter.
- The evidence doesn't support it.
- That's right.
- And then teams can't grow that much.
- Right.
- Right, so human beings, we're really good in teams of 10,
up to teams of 100, they can know each other.
Beyond that, you have to have organizational boundaries.
So you're kind of, you have,
those are pretty hard constraints, right?
So then you have to divide and conquer.
Like as the designs get bigger,
you have to divide it into pieces.
The power of abstraction layers is really high.
We used to build computers out of transistors.
Now we have a team that turns transistors into logic cells
and another team that turns them into functional units
and another one that turns them into computers, right?
So we have abstraction layers in there.
And you have to think about when do you shift gears on that?
We also use faster computers to build faster computers.
So some algorithms run twice as fast on new computers,
but a lot of algorithms are N squared.
So, you know, a computer with twice as many transistors
in it might take four times as long to run.
So you have to refactor the software.
Like simply using faster computers
to build bigger computers doesn't work.
So you have to think about all these things.
- So in terms of computing performance
and the exciting possibility
that more powerful computers bring,
is shrinking the thing we've just been talking about,
one of the, for you,
one of the biggest exciting possibilities
of advancement in performance,
or is there other directions
that you're interested in?
Like in the direction of sort of enforcing given parallelism
or like doing massive parallelism
in terms of many, many CPUs,
you know, stacking CPUs on top of each other,
that kind of parallelism or any kind of parallelism?
- Well, think about it in a different way.
So old computers, you know, slow computers,
you said A equal B plus C times D.
Pretty simple, right?
And then we made faster computers with vector units
and you can do proper equations and matrices, right?
And then modern like AI computations
or like convolutional neural networks,
where you convolve one large data set against another.
And so there's sort of this hierarchy of mathematics,
you know, from simple equation to linear equations
to matrix equations to deeper kind of computation.
And the data sets are getting so big
that people are thinking of data as a topology problem.
You know, data is organized in some immense shape.
And then the computation,
which sort of wants to be get data from immense shape
and do some computation on it.
So what computers have allowed people to do
is have algorithms go much, much further.
So that paper you referenced, the Sutton paper,
they talked about, you know, like when AI started,
it was apply rule sets to something.
That's a very simple computational situation.
And then when they did first chess thing,
they solved deep searches.
So have a huge database of moves and results, deep search,
but it's still just a search, right?
Now we take large numbers of images
and we use it to train these weight sets
that we convolve across.
It's a completely different kind of phenomena.
We call that AI.
Now they're doing the next generation.
And if you look at it,
they're going up this mathematical graph, right?
And then computations, both computation and data sets
support going up that graph.
- Yeah, the kind of computation that might,
I mean, I would argue that all of it is still a search,
right?
Just like you said, a topology problem of data sets,
you're searching the data sets for valuable data.
And also the actual optimization of neural networks
is a kind of search for the--
- I don't know, if you had looked at the inner layers
of finding a cat, it's not a search.
It's a set of endless projections.
So, you know, a projection,
here's a shadow of this phone, right?
And then you can have a shadow of that on the something
and a shadow on that of something.
If you look in the layers, you'll see,
this layer actually describes pointy ears
and round eyeness and fuzziness.
And, but the computation to tease out the attributes
is not search.
- Right, I mean--
- Like the inference part might be search,
but the training's not search.
- Okay, well--
§06Transcript · 43:08–51:42
- And then in deep networks, they look at layers
and they don't even know it's represented.
And yet if you take the layers out, it doesn't work.
- Okay, so--
- So I don't think it's search.
- All right, well--
- But you'd have to talk to a mathematician
about what that actually is.
- Well, we could disagree, but the,
it's just semantics, I think, it's not,
but it's certainly not--
- I would say it's absolutely not semantics, but--
- Okay.
All right, well, if you wanna go there.
So optimization to me is search,
and we're trying to optimize the ability
of a neural network to detect cat ears.
And the difference between chess
and the space, the incredibly multi-dimensional,
100,000 dimensional space that neural networks
are trying to optimize over
is nothing like the chessboard database.
So it's a totally different kind of thing.
Okay, in that sense, you can say--
- Yeah, yeah.
- It loses the meaning.
- I can see how you might say, if you,
the funny thing is, is the difference between
given search space and found search space.
- Right, exactly.
- Yeah, maybe that's a different way to describe it.
- That's a beautiful way to put it, okay.
But you're saying, what's your sense
in terms of the basic mathematical operations
and the architectures, computer hardware
that enables those operations?
Do you see the CPUs of today still being
a really core part of executing
those mathematical operations?
- Yes.
Well, the operations continue to be add, subtract,
load, store, compare, and branch.
It's remarkable.
So it's interesting that the building blocks
of computers or transistors, under that, atoms.
So you got atoms, transistors, logic gates, computers,
right, functional units of computers.
The building blocks of mathematics at some level
are things like adds and subtracts and multiplies,
but the space mathematics can describe
is, I think, essentially infinite.
But the computers that run the algorithms
are still doing the same things.
Now, a given algorithm might say,
I need sparse data, or I need 32-bit data,
or I need a convolution operation
that naturally takes eight-bit data,
multiplies it, and sums it up a certain way.
So the data types in TensorFlow imply an optimization set,
but when you go right down and look at the computers,
it's and and or gates doing adds and multiplies.
Like, that hasn't changed much.
Now, the quantum researchers think
they're gonna change that radically,
and then there's people who think about analog computing,
'cause you look in the brain,
and it seems to be more analogish.
You know, that maybe there's a way
to do that more efficiently.
But we have a million X on computation,
and I don't know the relationship
between computational, let's say, intensity
and ability to hit mathematical abstractions.
I don't know any ways to describe that,
but just like you saw in AI,
you went from rule sets to simple search
to complex search to, say, found search.
Like, those are orders of magnitude
more computation to do.
And as we get the next two orders of magnitude,
like a friend, Roger Godori, said,
every order of magnitude changes the computation.
- Fundamentally changes what the computation is doing.
- Yeah.
Oh, you know the expression,
the difference in quantity is the difference in kind.
You know, the difference between ant and anthill, right?
Or neuron and brain.
You know, there's this indefinable place
where the quantity changed the quality, right?
And we've seen that happen in mathematics multiple times,
and my guess is it's gonna keep happening.
- So, in your sense, is it, yeah,
if you focus head down and shrinking the transistor.
- Well, it's not just head down,
and we're aware of the software stacks
that are running the computational loads,
and we're kind of pondering,
what do you do with a petabyte of memory
that wants to be accessed in a sparse way
and have, you know, the kind of calculations
AI programmers want?
So, there's a dialogue and interaction,
but when you go in the computer chip,
you know, you find adders and subtractors and multipliers.
- So, if you zoom out then with,
as you mentioned, Rich Sutton,
the idea that most of the development
in the last many decades in AI research
came from just leveraging computation
and just simple algorithms
waiting for the computation to improve.
- Well, software guys have a thing that they call it
the problem of early optimization.
- Right.
- So, you write a big software stack,
and if you start optimizing,
like, the first thing you write,
the odds of that being the performance limiter is low.
But when you get the whole thing working,
can you make it 2X faster by optimizing the right things?
Sure.
While you're optimizing that,
could you have written a new software stack,
which would have been a better choice?
Maybe.
Now you have creative tension.
So--
- But the whole time as you're doing the writing,
that's the software we're talking about.
The hardware underneath gets faster and faster.
- This goes back to the Moore's Law.
If Moore's Law is gonna continue,
then your AI research should expect that to show up,
and then you make a slightly different set of choices,
then we've hit the wall, nothing's gonna happen,
and from here, it's just us rewriting algorithms.
Like, that seems like a failed strategy
for the last 30 years of Moore's Law's death.
- So, can you just linger on it?
I think you've answered it,
but I'll just ask the same dumb question
over and over.
So, why do you think Moore's Law is not going to die?
Which is the most promising, exciting possibility
of why it won't die in the next five, 10 years?
So, is it the continued shrinking of the transistor,
or is it another S-curve that steps in,
and it totally sort of--
- Well, shrinking the transistor
is literally thousands of innovations.
- Right, so there's stacks of S-curves in there.
- There's a whole bunch of S-curves
just kind of running their course
and being reinvented and new things.
You know, the semiconductor fabricators
and technologists have all announced
what's called nanowires,
so they took a fin which had a gate around it
and turned that into little wires
so you have better control of that,
and they're smaller,
and then from there, there's some obvious steps
about how to shrink that.
So, the metallurgy around wire stacks and stuff
has very obvious abilities to shrink,
and there's a whole combination of things there to do.
- Your sense is that we're gonna get a lot
if this innovation from just that shrinking.
- Yeah, like a factor of 100, it's a lot.
- Yeah, I would say.
That's incredible.
And it's totally unknown.
- It's only 10 or 15 years.
- Now, you're smart, and you might know,
but to me, it's totally unpredictable
of what that 100x would bring
in terms of the nature of the computation
that people would be doing.
- Yeah, you're familiar with Bell's Law.
So, for a long time, it was mainframes,
minis, workstation, PC, mobile.
Moore's Law drove faster, smaller computers.
And then, when we were thinking about Moore's Law,
Rajagirdari said, "Every 10x generates a new computation."
So, scalar, vector, matrix, topological computation.
And if you go look at the industry trends,
there was mainframes and minicomputers and PCs,
and then the internet took off,
and then we got mobile devices,
and now we're building 5G wireless
with one millisecond latency.
And people are starting to think about the smart world
where everything knows you, recognizes you.
The transformations are gonna be unpredictable.
- How does it make you feel that you're one
of the key architects of this kind of future?
So, we're not talking about the architects
of the high-level people who build the Angry Bird apps.
- What's wrong with Angry Bird apps?
§07Transcript · 51:44–1:00:20
Who knows?
Maybe that's the whole point of the universe.
- I'm gonna take a stand at that,
and the attention-distracting nature of mobile phones.
I'll take a stand.
But anyway, in terms of--
- I don't think that matters much.
- The side effects of smartphones
or the attention-distraction, which part?
- Well, who knows where this is all leading?
It's changing so fast.
- Wait, so back to the--
- My parents used to yell at my sisters
for hiding in the closet with a wired phone
with a dial on it.
Stop talking to your friends all day.
- Right.
- Now my wife yells at my kids
for talking to their friends all day on text.
I don't know, it looks the same to me.
- It's always, it echoes of the same thing.
Okay, but you are one of the key people
architecting the hardware of this future.
How does that make you feel?
Do you feel responsible?
Do you feel excited?
- So we're in a social context,
so there's billions of people on this planet.
There are literally millions of people
working on technology.
I feel lucky to be doing what I do
and getting paid for it, and there's an interest in it.
But there's so many things going on in parallel.
Like the actions are so unpredictable.
If I wasn't here, somebody else would do it.
The vectors of all these different things
are happening all the time.
You know, there's a, I'm sure some philosopher
or meta-philosophers, you know,
wondering about how we transform our world.
- So you can't deny the fact that these tools,
whether, that these tools are changing our world.
- That's right.
- So do you think it's changing for the better?
- I read this thing recently, it said
the two disciplines with the highest GRE scores
in college are physics and philosophy.
And they're both sort of trying to answer the question,
why is there anything?
And the philosophers are on the kind of theological side,
and the physicists are obviously on the material side.
And there's 100 billion galaxies with 100 billion stars.
It seems, well, repetitive at best.
So, you know, there's, on our way to 10 billion people.
I mean, it's hard to say what it's all for,
if that's what you're asking.
- Yeah, I guess I am.
- Things do tend to significantly increase in complexity.
And I'm curious about how computation,
like our world, our physical world,
inherently generates mathematics.
It's kind of obvious, right?
So we have XYZ coordinates.
You take a sphere, you make it bigger,
you get a surface that falls, you know, grows by R squared.
Like it generally generates mathematics,
and the mathematicians and the physicists
have been having a lot of fun talking
to each other for years.
And computation has been, let's say, relatively pedestrian.
Like computation in terms of mathematics
has been doing binary algebra,
while those guys have been gallivanting
through the other realms of possibility, right?
Now, recently, the computation lets you do
mathematical computations that are sophisticated enough
that nobody understands how the answers came out, right?
- Machine learning.
- Machine learning.
- Yeah, yeah.
- It used to be, you get data set,
you guess at a function.
The function is considered physics
if it's predictive of new functions, new data sets.
Modern, you can take a large data set
with no intuition about what it is
and use machine learning to find a pattern
that has no function, right?
And it can arrive at results that I don't know
if they're completely mathematically describable.
So computation has kind of done something interesting
compared to A equal B plus C.
- There's something reminiscent of that step
from the basic operations of addition
to taking a step towards neural networks
that's reminiscent of what life on Earth
at its origins was doing.
Do you think we're creating sort of the next step
in our evolution in creating artificial intelligence systems
that will--
- I don't know.
I mean, there's so much in the universe already,
it's hard to say.
- Where we stand in this whole thing.
- Are human beings working on additional abstraction layers
and possibilities?
Yeah, it appears so.
Does that mean that human beings don't need dogs?
You know, no.
Like there's so many things
that are all simultaneously interesting and useful.
- Well, you've seen, throughout your career,
you've seen greater and greater level abstractions
built in artificial machines, right?
Do you think, when you look at humans,
do you think that the look of all life on Earth
as a single organism building this thing,
this machine with greater and greater levels of abstraction,
do you think humans are the peak,
the top of the food chain
in this long arc of history on Earth?
Or do you think we're just somewhere in the middle?
Are we the basic functional operations of a CPU?
Are we the C++ program, the Python program?
Are we the neural network?
Like somebody's, you know, people have calculated
like how many operations does the brain do?
And something, you know, I've seen the number 10
to the 18th a bunch of times, arrived different ways.
So could you make a computer
that did 10 to the 20th operations?
- Yes. - Sure.
- So you think-- - We're gonna do that.
Now, is there something magical
about how brains compute things?
I don't know.
You know, my personal experience is interesting
'cause, you know, you think you know how you think
and then you have all these ideas
and you can't figure out how they happened.
And if you meditate, you know,
like what you can be aware of is interesting.
So I don't know if brains are magical or not.
You know, the physical evidence says no.
Lots of people's personal experience says yes.
So what would be funny is if brains are magical
and yet we can make brains with more computation.
You know, I don't know what to say about that, but.
- Well, do you think magic is an emergent phenomena?
What-- - It could be.
I have no explanation for it.
I'm an engineer. - Let me ask Jim Keller
of what in your view is consciousness?
- What's consciousness?
- Yeah, like what, you know, consciousness, love,
things that are these deeply human things
that seems to emerge from our brain.
Is that something that we'll be able to make,
encode in chips that get faster and faster
and faster and faster?
- That's like a 10 hour conversation.
Nobody really knows.
- Can you summarize it in a couple of sentences?
- Many people have observed that organisms run
at lots of different levels, right?
If you had two neurons, somebody said
you'd have one sensory neuron and one motor neuron, right?
So we move towards things and away from things
and we have physical integrity and safety or not, right?
And then if you look at the animal kingdom,
you can see brains that are a little more complicated
and at some point there's a planning system
and then there's an emotional system
that's happy about being safe
or unhappy about being threatened, right?
And then our brains have massive numbers of structures,
you know, like planning and movement and thinking
and feeling and drives and emotions.
And we seem to have multiple layers of thinking systems.
And we have a brain, a dream system
that nobody understands whatsoever,
which I find completely hilarious.
And you can think in a way that those systems
are more independent and you can observe,
you know, the different parts of yourself can observe them.
I don't know which one's magical.
I don't know which one's not computational.
So.
- Is it possible that it's all computation?
- Probably.
Is there a limit to computation?
I don't think so.
- Do you think the universe is a computer?
- I don't know, it seems to be.
It's a weird kind of computer
because if it was a computer, right?
Like when they do calculations on what it,
how much calculation it takes to describe quantum effects
is unbelievably high.
So if it was a computer,
§08Transcript · 1:00:22–1:08:52
wouldn't you have built it out of something
that was easier to compute?
Right, that's a funny, it's a funny system.
But then the simulation guys have pointed out
that the rules are kind of interesting.
Like when you look really close, it's uncertain.
And the speed of light says you can only look so far
and things can't be simultaneous
except for the odd entanglement problem
where they seem to be.
Like the rules are all kind of weird.
And somebody said physics is like having 50 equations
with 50 variables to define 50 variables.
Like, you know, it's, you know,
like physics itself has been a shit show
for thousands of years.
It seems odd when you get to the corners of everything.
You know, it's either uncomputable
or undefinable or uncertain.
- It's almost like the designers of the simulation
are trying to prevent us from understanding it perfectly.
- But also the things that require calculations
require so much calculation
that our idea of the universe of a computer is absurd
because every single little bit of it
takes all the computation in the universe to figure out.
So that's a weird kind of computer.
You know, you say the simulation is running in the computer
which has by definition infinite computation.
- Not infinite.
Oh, you mean if the universe is infinite?
- Yeah, well, every little piece of our universe
seems to take infinite computation to figure out.
- Just a lot.
- Well, a lot's a pretty big number.
Compute this little teeny spot takes all the mass
in the local one light year by one light year space.
It's close enough to infinite.
- Oh, it's a heck of a computer if it is one.
- I know, it's a weird description
'cause the simulation description seems to break
when you look closely at it.
But the rules of the universe seem to imply something's up.
That seems a little arbitrary.
- The universe, the whole thing, the laws of physics,
it just seems like how did it come out to be the way it is?
- Well, lots of people talk about that.
Like I said, the two smartest groups of humans
are working on the same problem.
- From different sides.
- Different aspects and they're both complete failures.
So that's kind of cool.
- They might succeed eventually.
- Well, after 2,000 years, the trend isn't good.
- Oh, 2,000 years is nothing in the span
of the history of the universe.
So we have some time.
- But the next 1,000 years doesn't look good either.
- That's what everybody says at every stage.
But with Moore's Law, as you've just described,
not being dead, the exponential growth of technology,
the future seems pretty incredible.
- Well, it'll be interesting, that's for sure.
- That's right.
So what are your thoughts on Ray Kurzweil's sense
that exponential improvement in technology
will continue indefinitely?
Is that how you see Moore's Law?
Do you see Moore's Law more broadly
in the sense that technology of all kinds
has a way of stacking S-curves on top of each other
where it'll be exponential
and then we'll see all kinds of--
- What does an exponential of a million mean?
That's a pretty amazing number.
And that's just for a local little piece of silicon.
Now let's imagine you say decided to get
1,000 tons of silicon to collaborate in one computer
at a million times the density.
Like now you're talking, I don't know,
10 to the 20th more computation power
than our current already unbelievably fast computers.
Like nobody knows what that's gonna mean.
The sci-fi guys call it computronium.
Like when a local civilization turns
the nearby star into a computer.
Like I don't know if that's true.
- So just even when you shrink a transistor,
- That's only one dimension.
- The ripple effects of that.
- Like people tend to think about computers
as a cost problem, right?
So computers are made out of silicon
and minor amounts of metals.
And you know, this and that.
None of those things cost any money.
Like there's plenty of sand.
Like you could just turn the beach
and a little bit of ocean water into computers.
So all the cost is in the equipment to do it.
And the trend on equipment is once you figure out
how to build the equipment, the trend of cost is zero.
Elon said first you figure out what configuration
you want the atoms in and then how to put them there.
Right?
- Yeah.
- But here's the, you know, his great insight is
people are how constrained.
I have this thing, I know how it works.
And then little tweaks to that will generate something
as opposed to what do I actually want
and then figure out how to build it.
It's a very different mindset.
And almost nobody has it, obviously.
- Well, let me ask on that topic.
You were one of the key early people
in the development of autopilot,
at least in the hardware side.
Elon Musk believes that autopilot and vehicle autonomy,
if you just look at that problem,
can follow this kind of exponential improvement.
In terms of the how question that we're talking about,
there's no reason why it can't.
What are your thoughts on this particular space
of vehicle autonomy and your part of it
and Elon Musk's and Tesla's vision for--
- Well, the computer you need to build is straightforward.
And you could argue, well, does it need to be
two times faster or five times or 10 times?
But that's just a matter of time or price in the short run.
So that's not a big deal.
You don't have to be especially smart to drive a car.
So it's not like a super hard problem.
I mean, the big problem with safety is attention,
which computers are really good at, not skills.
- Well, let me push back on one.
You say everything you said is correct,
but we as humans tend to take for granted
how incredible our vision system is.
So--
- You can drive a car with 20/50 vision
and you can train a neural network to extract
the distance of any object and the shape of any surface
from a video and data.
- Yeah, but-- - It's really simple.
- No, it's not simple.
- That's a simple data problem.
- It's not simple.
It's because it's not just detecting objects,
it's understanding the scene and it's being able to do it
in a way that doesn't make errors.
So the beautiful thing about the human vision system
and our entire brain around the whole thing
is we're able to fill in the gaps.
It's not just about perfectly detecting cars,
it's inferring the occluded cars.
It's trying to, it's understanding the physics--
- I think that's mostly a data problem.
So you think what data would compute
with improvement of computation,
with improvement in collection--
- Well, there is a, you know, when you're driving a car
and somebody cuts you off, your brain has theories
about why they did it.
You know, they're a bad person, they're distracted,
they're dumb, you know, you can listen to yourself.
- Right.
- So, you know, if you think that narrative is important
to be able to successfully drive a car,
then current autopilot systems can't do it.
But if cars are ballistic things with tracks
and probabilistic changes of speed and direction,
and roads are fixed and given by the way,
they don't change dynamically, right?
You can map the world really thoroughly.
You can place every object really thoroughly, right?
You can calculate trajectories of things really thoroughly.
Right?
- But everything you said about really thoroughly
has a different degree of difficulty.
So--
- And you could say at some point,
computer autonomous systems will be way better
at things that humans are lousy at.
Like, they'll be better at attention,
they'll always remember there was a pothole in the road
that humans keep forgetting about.
They'll remember that this set of roads
has these weirdo lines on it
that the computers figured out once.
And especially if they get updates
so if somebody changes a given,
like the key to robots and stuff,
somebody said is to maximize the givens.
Right?
- Right.
- So having a robot pick up this bottle cap
is way easier to put a red dot on the top.
'Cause then you have to figure out,
if you wanna do a certain thing with it,
§09Transcript · 1:08:54–1:17:04
maximize the givens is the thing.
And autonomous systems are happily maximizing the givens.
Like humans, when you drive someplace new,
you remember it 'cause you're processing it the whole time.
And after the 50th time you drove to work,
you get to work, you don't know how you got there.
Right?
You're on autopilot.
Right?
Autonomous cars are always on autopilot.
But the cars have no theories about why they got cut off
or why they're in traffic.
- So that's--
- They also never stop paying attention.
- Right.
So I tend to believe you do have to have theories,
meta models of other people,
especially with pedestrian and cyclists,
but also with other cars.
So everything you said is actually essential to driving.
Driving is a lot more complicated than people realize,
I think.
So sort of to push back slightly, but--
- So to cut into traffic, right?
- Yep.
- You can't just wait for a gap.
You have to be somewhat aggressive.
You'd be surprised how simple a calculation for that is.
- I may be on that particular point, but there's--
Maybe I actually have to push back.
I would be surprised.
You know what?
I'll say where I stand.
I would be very surprised,
but I think you might be surprised how complicated it is.
- I tell people, it's like progress disappoints
in the short run and surprises in the long run.
- It's very possible.
Yeah.
- I suspect in 10 years, it'll be just taken for granted.
- Yeah, probably.
But you're probably right.
Now it look like--
- It's gonna be a $50 solution that nobody cares about.
It's like GPS is like, wow, GPS is,
we have satellites in space
that tell you where your location is.
It was a really big deal.
Now everything has a GPS in it.
- Yeah, that's true.
But I do think that systems that involve human behavior
are more complicated than we give them credit for.
So we can do incredible things with technology
that don't involve humans, but when you--
- I think humans are less complicated than people
frequently ascribed.
- Maybe I--
- We tend to operate out of large numbers of patterns
and just keep doing it over and over.
- But I can't trust you because you're a human.
That's something a human would say.
But my hope is on the point you've made is,
even if, no matter who's right,
I'm hoping that there's a lot of things
that humans aren't good at
that machines are definitely good at.
Like you said, attention and things like that.
Well, they'll be so much better
that the overall picture of safety and autonomy
will be obviously cars will be safer,
even if they're not as good at it.
- I'm a big believer in safety.
I mean, there are already the current safety systems
like cruise control that doesn't let you run into people
and lane keeping.
There are so many features that you just look at the Pareto
of accidents and knocking off like 80% of them
is super doable.
- Just to linger on the autopilot team
and the efforts there,
it seems to be that there's a very intense scrutiny
by the media and the public in terms of safety,
the pressure, the bar put before autonomous vehicles.
What are your sort of as a person there
working on the hardware and trying to build a system
that builds a safe vehicle and so on,
what was your sense about that pressure?
Is it unfair?
Is it expected of new technology?
- Yeah, it seems reasonable.
I was interested, I talked to both American
and European regulators,
and I was worried that the regulations
would write into the rules technology solutions
like modern brake systems imply hydraulic brakes.
So if you read the regulations
to meet the letter of the law for brakes,
it sort of has to be hydraulic, right?
And the regulator said,
they're interested in the use cases,
like a head-on crash, an offset crash,
don't hit pedestrians, don't run into people,
don't leave the road, don't run a red light or a stoplight.
They were very much into the scenarios.
And they had all the data about which scenarios
injured or killed the most people.
And for the most part, those conversations were like,
what's the right thing to do to take the next step?
Now Elon's very interested also in the benefits
of autonomous driving or freeing people's time
and attention as well as safety.
And I think that's also an interesting thing,
but building autonomous systems so they're safe
and safer than people seemed,
since the goal is to be 10x safer than people,
having the bar to be safer than people
and scrutinizing accidents seems philosophically correct.
So I think that's a good thing.
- It's different than the things you worked at,
the Intel, AMD, Apple, with autopilot chip design
and hardware design.
What are interesting or challenging aspects
of building this specialized kind of computing system
in the automotive space?
- I mean, there's two tricks to building
like an automotive computer.
One is the software team, the machine learning team
is developing algorithms that are changing fast.
So as you're building the accelerator,
you have this worry or intuition
that the algorithms will change enough
that the accelerator will be the wrong one.
And there's a generic thing,
which is if you build a really good general purpose computer
say its performance is one,
and then GPU guys will deliver about 5x to performance
for the same amount of silicon,
because instead of discovering parallelism,
you're given parallelism.
And then special accelerators get another two to 5x
on top of a GPU, because you say,
I know the math is always eight bit integers
into 32 bit accumulators,
and the operations are the subset
of mathematical possibilities.
So, AI accelerators have a claim performance benefit
over GPUs because in the narrow mass space,
you're nailing the algorithm.
Now, you still try to make it programmable,
but the AI field is changing really fast.
So there's a little creative tension there of,
I want the acceleration afforded by specialization
without being over specialized
so that the new algorithm is so much more effective
that you'd have been better off on a GPU.
So there's a tension there.
To build a good computer for an application like automotive,
there's all kinds of sensor inputs and safety processors
and a bunch of stuff.
So one of Elon's goals to make it super affordable.
So every car gets an autopilot computer.
So some of the recent startups you look at,
and they have a server in the trunk,
because they're saying,
I'm gonna build this autopilot computer
or replaces the driver.
So their cost budget's 10 or $20,000.
And Elon's constraint was, I'm gonna put one in every car,
whether people buy autonomous driving or not.
So the cost constraint he had in mind was great.
And to hit that, you had to think about the system design.
That's complicated, it's fun.
You know, it's like, it's craftsman's work.
Like a violin maker, right?
You can say Stradivarius is this incredible thing,
the musicians are incredible.
But the guy making the violin, you know,
picked wood and sanded it, and then he cut it,
you know, and he glued it, you know,
and he waited for the right day
so that when he put the finish on it,
it didn't, you know, do something dumb.
That's craftsman's work, right?
You may be a genius craftsman
'cause you have the best techniques
and you discover a new one,
but most engineering is craftsman's work.
And humans really like to do that.
You know the expression-- - Smart humans.
- No, everybody.
- All humans. - I don't know.
I used to, I dug ditches when I was in college.
I got really good at it, satisfying.
- Yeah.
Digging ditches is also craftsman work.
- Yeah, of course.
So there's an expression called complex mastery behavior.
So when you're learning something,
that's fun 'cause you're learning something.
When you do something and it's rote and simple,
§10Transcript · 1:17:05–1:25:07
it's not that satisfying.
But if the steps that you have to do are complicated
and you're good at 'em, it's satisfying to do them.
And then if you're intrigued by it all,
as you're doing them, you sometimes learn new things
that you can raise your game.
But craftsman's work is good.
And engineers, like engineering is complicated enough
that you have to learn a lot of skills
and then a lot of what you do is then craftsman's work,
which is fun.
Autonomous driving, building a very
resource-constrained computer,
so a computer has to be cheap enough
to put in every single car,
that essentially boils down to craftsman's work.
It's engineering, it's--
- Yeah, you know, there's thoughtful decisions
and problems to solve and trade-offs to make.
Do you need 10 camera in ports or eight?
You know, you're building for the current car
or the next one.
You know, how do you do the safety stuff?
You know, there's a whole bunch of details.
But it's fun.
It's not like I'm building a new type of neural network
which has a new mathematics and a new computer to work.
You know, that's, like there's more invention than that.
But the rejection to practice,
once you pick the architecture, you look inside
and what do you see?
Adders and multipliers and memories and, you know,
the basics.
So computers is always this weird set of abstraction layers
of ideas and thinking that reduction to practice
is transistors and wires and, you know, pretty basic stuff.
And that's an interesting phenomenon.
By the way, like factory work,
like lots of people think factory work
is road assembly stuff.
I've been on the assembly line.
Like the people who work there really like it.
It's a really great job.
It's really complicated.
Putting cars together is hard, right?
And the car is moving and the parts are moving
and sometimes the parts are damaged
and you have to coordinate putting all the stuff together
and people are good at it.
They're good at it.
And I remember one day I went to work
and the line was shut down for some reason
and some of the guys sitting around were really bummed
'cause they had reorganized a bunch of stuff
and they were gonna hit a new record
for the number of cars built that day
and they were all gung-ho to do it.
And these were big, tough buggers.
(Luke laughs)
But what they did was complicated and you couldn't do it.
- Yeah, and I mean--
- Well, after a while you could,
but you'd have to work your way up
'cause, you know, like putting the bright,
what's called the brights,
the trim on a car on a moving assembly line
where it has to be attached 25 places
in a minute and a half is unbelievably complicated.
And human beings can do it, it's really good.
I think that's harder than driving a car, by the way.
- Putting together, working--
- Working in a factory.
- Too smart people can disagree.
- Yay.
- I think driving a car--
- Well, we'll get you in the factory someday
and then we'll see how you do.
- No, not for us humans driving a car is easy.
I'm saying building a machine
that drives a car is not easy.
Okay, driving a car is easy for humans
because we've been evolving for billions of years.
- To drive cars, yeah, I noticed that.
The paleolithic cars are super cool.
- Oh, now you join the rest of the internet in mocking me.
- Okay.
(Luke laughs)
I wasn't mocking, I was just, you know,
intrigued by your anthropology.
- Yeah, it's--
- I'll have to go dig into that.
- There's some inaccuracies there, yes.
Okay, but in general,
(Luke laughs)
what have you learned in terms of,
thinking about passion, craftsmanship,
tension, chaos, you know--
- Jesus.
- The whole mess of it,
what have you learned, have taken away from your time
working with Elon Musk, working at Tesla,
which is known to be a place of chaos,
innovation, craftsmanship, and all those things.
- I really like the way he thought.
Like, you think you have an understanding
about what first principles of something is,
and then you talk to Elon about it,
and you didn't scratch the surface.
You know, he has a deep belief
that no matter what you do, it's a local maximum.
Right, and I had a friend,
he invented a better electric motor,
and it was a lot better than what we were using.
And one day he came by, he said,
"You know, I'm a little disappointed,
"'cause this is really great,
"and you didn't seem that impressed."
And I said, "You know, when the super intelligent aliens
"come, are they gonna be looking for you?"
Like, where is he?
The guy who built the motor.
Probably not.
But doing interesting work that's both innovative,
and let's say craftsman's work on the current thing,
it's really satisfying, and it's good.
And that's cool.
And then Elon was good at taking everything apart.
Like, what's the deep first principle?
Oh, no, what's really, no, what's really?
You know, that ability to look at it without assumptions,
and how constraint is super wild.
You know, he built a rocket ship,
and electric car, and everything.
And that's super fun, and he's into it, too.
Like, when they first landed two SpaceX rockets to Tesla,
we had a video projector in the big room,
and like 500 people came down,
and when they landed, everybody cheered,
and some people cried.
It was so cool.
All right, but how did you do that?
Well, it was super hard.
And then people say, "Well, it's chaotic."
Really?
To get out of all your assumptions?
You think that's not gonna be unbelievably painful?
And is Elon tough?
Yeah, probably.
Do people look back on it and say,
"Boy, I'm really happy I had that experience
"to go take apart that many layers of assumptions?"
Sometimes super fun, sometimes painful.
So it could be emotionally and intellectually painful,
that whole process of just stripping away assumptions?
Yeah, imagine 99% of your thought process
is protecting your self-conception.
And 98% of that's wrong.
Now you got the math right.
How do you think you're feeling
when you get back into that one bit that's useful,
and now you're open and you have the ability
to do something different?
I don't know if I got the math right.
It might be 99.9, but it ain't 50.
Imagining that 50% is hard enough.
Now for a long time I've suspected you could get better.
Like you can think better, you can think more clearly,
you can take things apart.
And there's lots of examples of that, people who do that.
And Elon is an example of that.
Apparently. You are an example.
So-- I don't know if I am.
I'm fun to talk to.
Certainly. I've learned a lot of stuff.
Right. Well, here's the other thing
is like, I joke, like I read books,
and people think, oh, you read books.
Well, no, I've read a couple of books a week for 55 years.
Well, maybe 50, 'cause I didn't learn to read
until I was eight or something.
And it turns out when people write books,
they often take 20 years of their life
where they passionately did something,
reduced it to 200 pages.
That's kind of fun.
And then you go online and you can find out
who wrote the best books and who like, you know,
that's kind of wild.
So there's this wild selection process,
and then you can read it,
and for the most part, understand it.
And then you can go apply it.
Like I went to one company, I thought,
I haven't managed much before.
So I read 20 management books,
and I started talking to them,
and basically compared to all the VPs running around,
I'd read 19 more management books than anybody else.
(laughing)
Wasn't even that hard.
§11Transcript · 1:25:08–1:34:12
And half the stuff worked, like first time.
It wasn't even rocket science.
- But at the core of that is questioning the assumptions,
or sort of entering, thinking first principles thinking,
sort of looking at the reality of the situation,
and using that knowledge, applying that knowledge.
- Yeah, so I would say my brain has this idea
that you can question first assumptions.
But I can go days at a time and forget that,
and you have to kind of circle back that observation.
- Because it is emotionally challenging.
- Well, it's hard to just keep it front and center,
'cause you operate on so many levels all the time,
and getting this done takes priority,
or being happy takes priority,
or screwing around takes priority.
Like how you go through life is complicated.
And then you remember, oh yeah,
I could really think first principles.
Oh shit, that's tiring.
But you do for a while, and that's kind of cool.
- So just as a last question in your sense,
from the big picture, from the first principles,
do you think, you kind of answered it already,
but do you think autonomous driving
is something we can solve on a timeline of years?
So one, two, three, five, 10 years,
as opposed to a century?
- Yeah, definitely.
- Just to linger on it a little longer,
where's the confidence coming from?
Is it the fundamentals of the problem,
the fundamentals of building the hardware and the software?
- As a computational problem,
understanding ballistics, roles, topography,
it seems pretty solvable.
I mean, and you can see this,
like speech recognition for a long time,
people are doing frequency and domain analysis,
and all kinds of stuff,
and that didn't work for at all, right?
And then they did deep learning about it,
and it worked great.
And it took multiple iterations.
And, you know, autonomous driving
is way past the frequency analysis point.
You know, use radar, don't run into things.
And the data gathering is going up,
and the computation is going up,
and the algorithm understanding is going up,
and there's a whole bunch of problems
getting solved like that.
- The data side is really powerful,
but I disagree with both you and Elon.
I'll tell Elon once again, as I did before,
that when you add human beings into the picture,
it's no longer a ballistics problem.
It's something more complicated,
but I could be very well proven wrong.
- Cars are highly damped in terms of rate of change.
Like the steering system's really slow
compared to a computer.
The acceleration, the acceleration's really slow.
- Yeah, on a certain time scale.
On a ballistics time scale, but human behavior,
I don't know.
I shouldn't say--
- Human beings are really slow too.
Weirdly, we operate, you know,
half a second behind reality.
Nobody really understands that one either.
It's pretty funny.
- Yeah, yeah.
We very well could be surprised.
And I think with the rate of improvement
in all aspects on both the compute
and the software and the hardware,
there's gonna be pleasant surprises all over the place.
- Mm-hmm.
- Speaking of unpleasant surprises,
many people have worries about a singularity
in the development of AI.
Forgive me for such questions.
- Yeah.
- When AI improves exponentially
and reaches a point of superhuman level
general intelligence,
you know, beyond the point, there's no looking back.
Do you share this worry of existential threats
from artificial intelligence,
from computers becoming superhuman level intelligent?
- No, not really.
You know, like we already have a very stratified society.
And then if you look at the whole animal kingdom
of capabilities and abilities and interests,
and, you know, smart people have their niche,
and, you know, normal people have their niche,
and craftsmen have their niche,
and, you know, animals have their niche.
I suspect that the domains of interest
for things that, you know, astronomically different,
like the whole something got 10 times smarter than us
and wanted to track us all down because what?
We like to have coffee at Starbucks?
Like, it doesn't seem plausible.
No, is there an existential problem
that how do you live in a world
where there's something way smarter than you,
and you base your kind of self-esteem
on being the smartest local person?
Well, there's what, 0.1% of the population who thinks that?
'Cause the rest of the population's been dealing with it
since they were born.
So the breadth of possible experience
that can be interesting is really big.
And, you know, superintelligence seems likely,
although we still don't know if we're magical,
but I suspect we're not,
and it seems likely that it'll create possibilities
that are interesting for us,
and its interests will be interesting for whatever it is.
It's not obvious why its interests
would somehow wanna fight over some square foot of dirt
or, you know, whatever the usual fears are about.
- So you don't think you'll inherit
some of the darker aspects of human nature?
- Depends on how you think reality's constructed.
So for whatever reason, human beings are in, let's say,
creative tension and opposition
with both our good and bad forces.
Like, there's lots of philosophical understanding of that.
Right?
I don't know why that would be different.
- So you think the evil is necessary for the good?
I mean, the tension.
- I don't know about evil,
but like we live in a competitive world
where your good is somebody else's, you know, evil.
You know, there's the malignant part of it,
but that seems to be self-limiting,
although occasionally it's super horrible.
- But yes, there's a debate over ideas
and some people have different beliefs
and that debate itself is a process
so that arriving at something--
- Yeah, and why wouldn't that continue?
- Yeah.
But you don't think that whole process
will leave humans behind in a way that's painful?
Emotionally painful, yes, for the 0.1%.
There'll be--
- Isn't it already painful for a large percentage
of the population?
And it is.
I mean, society does have a lot of stress in it,
about the 1% and about to this and about to that,
but you know, everybody has a lot of stress in their life
about what they find satisfying
and you know, know yourself seems to be the proper dictum
and pursue something that makes your life meaningful
seems proper.
And there's so many avenues on that.
Like, there's so much unexplored space
at every single level.
You know, I'm somewhat of,
my nephew called me a jaded optimist.
(laughing)
- There's a beautiful tension in that label.
But if you were to look back at your life
and could relive a moment, a set of moments
because there were the happiest times of your life
outside of family, what would that be?
- I don't wanna relive any moments.
I like that.
I like that situation where you have some amount of optimism
and then the anxiety of the unknown.
- So you love the unknown, the mystery of it.
- I don't know about the mystery.
It sure gets your blood pumping.
- What do you think is the meaning of this whole thing?
Of life on this pale blue dot?
- It seems to be what it does.
Like the universe, for whatever reason,
makes atoms which makes us which we do stuff.
And we figure out things and we explore things.
- That's just what it is.
- It's not just.
- Yeah, it is.
Jim, I don't think there's a better place to end it
it's a huge honor.
- Well, that was super fun.
- Thank you so much for talking today.
- All right, great.
- Thanks for listening to this conversation
and thank you to our presenting sponsor, Cash App.
Download it, use code LexPodcast, you'll get $10
and $10 will go to FIRST, a STEM education nonprofit
that inspires hundreds of thousands of young minds
to become future leaders and innovators.
If you enjoy this podcast, subscribe on YouTube,
§12Transcript · 1:34:15–1:34:37
get five stars on Apple Podcast, follow on Spotify,
support on Patreon or simply connect with me on Twitter.
And now let me leave you with some words of wisdom
from Gordon Moore.
If everything you try works, you aren't trying hard enough.
Thank you for listening and hope to see you next time.
(upbeat music)
(upbeat music)