Cal Newport is a computer scientist who also writes about productivity. Please support this podcast by checking out our sponsors:
- ExpressVPN: https://expressvpn.com/lexpod and use code LexPod to get 3 months free
- Linode: https://linode.com/lex to get $100 free credit
- Sun Basket: https://sunbasket.com/lex and use code LEX to get $35 off
- SimpliSafe: https://simplisafe.com/lex and use code LEX to get a free security camera

EPISODE LINKS:
A World Without Email (book): https://amzn.to/3blXyjv
Deep Work (book): https://amzn.to/3c0npMM
Digital Minimalism (book): https://amzn.to/3kJPMmx
Cal's Website: https://www.calnewport.com/
Deep Questions (podcast): https://www.calnewport.com/podcast/

PODCAST INFO:
Podcast website: https://lexfridman.com/podcast
Apple Podcasts: https://apple.co/2lwqZIr
Spotify: https://spoti.fi/2nEwCF8
RSS: https://lexfridman.com/feed/podcast/
Full episodes playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOdP_8GztsuKi9nrraNbKKp4
Clips playlist: https://www.youtube.com/playlist?list=PLrAXtmErZgOeciFP3CBCIEElOJeitOr41

OUTLINE:
0:00 - Introduction
2:15 - Deep work
7:00 - Focus
12:43 - Time blocking
19:38 - Deadlines
29:13 - Do less, do better, know why
31:55 - Clubhouse
45:58 - Burnout
52:25 - Boredom
1:00:10 - Quit social media for 30 days
1:10:04 - Social media
1:35:12 - How email destroyed our productivity at work
1:44:57 - How we fix email
1:51:59 - Over-optimization
1:56:14 - When to use email and when not to
2:03:57 - Podcasting
2:08:33 - Alan Turing proving the impossible
2:12:32 - Fragility of math in the face of randomness
2:21:21 - Neural networks
2:30:06 - What will the P=NP proof look like?
2:33:46 - Is math discovered or invented?
2:37:53 - Book publishing
2:47:59 - Love
2:51:21 - Death
2:54:17 - Meaning of life

SOCIAL:
- Twitter: https://twitter.com/lexfridman
- LinkedIn: https://www.linkedin.com/in/lexfridman
- Facebook: https://www.facebook.com/LexFridmanPage
- Instagram: https://www.instagram.com/lexfridman
- Medium: https://medium.com/@lexfridman
- Reddit: https://reddit.com/r/lexfridman
- Support on Patreon: https://www.patreon.com/lexfridman

Human, despite what his comedy sounds like, he's the sweetest human being.

And he's a big proponent of just pick up the phone.

And it makes me so uncomfortable people call me.

It's like, I don't know what to do with this thing.

But it kind of gets everything done quicker, I think, if I remove the anxiety from that.

Or is email could still be the most efficient way to do this?

No, I mean, look, if you have to interact with someone, there's a lot of efficiency in synchrony, right?

And this is something from distributed system theory where you know if you go from synchronous to asynchronous networks, there's a huge amount of overhead to the asynchrony.

So actually the protocols required to solve things in asynchronous networks are significantly more complicated and fragile than synchronous protocols.

So if we can just do real time, it's usually better.

And also from a interaction, like social connection standpoint, there's a lot more information in the human voice and the back and forth.

Yeah, if you just call so very generational, right?

Like our generation will be comfortable talking on the phone in a way that like a younger generation isn't, but an older generation is more comfortable with, well, you just call people.

Whereas we, so there's a happy medium, but most of my good friends, we just talk, we have regular phone calls.

Yeah, it's not, I don't just call them, we schedule it, we schedule it, yeah.

Just on text, like, yeah, you wanna talk sometime soon.

Do you ever have a process around friends?

Well, you have a lot of interesting friend possibilities.

Really interesting people you can talk to.

The other 1 is the introversion, where I'm just afraid of people and get really stressed.

But it's almost like there's a, it has to do with the timetables thing and the deep work that the nice thing about the processes is it not only automates sort of automates away the context switching, It ensures you do the important things too.

So the thing is with email, because everything is done over email, you can be lazy in the same way with like social networks and do the easy things first, that are not that important.

So the process also enforces that you do the important things.

And for me, the important things is like, okay, that sounds weird, but like social connection.

No, that's 1 of the most important things in all of human existence.

Yeah and doing it the the paradoxical thing I got into this for digital minimalism The more you sacrifice on behalf of the connection the stronger that connection feels right?

So Sacrificing non-trivial time and attention on behalf of someone is what tells your brain that this is a serious relationship, which is why social media had this paradoxical effect of making people feel less social, because it took the friction out of it.

The brain just doesn't, like, yeah, you've been commenting on this person's whatever, you've been retweeting them or sending them some text, you haven't, it's not hard enough.

And then then the perceived strength of that social connection diminishes, where if you talk to them or go spend time with them or whatever, you're going to feel better about it.

I have a thing with some of my friends where at the end of each call, we take a couple minutes to schedule the next.

Then you never have to, it's like I do with haircuts or something, right?

Like if I don't schedule it then, I'm never going to get my haircut, right?

And so we, it's like, okay, when do you want to talk next?

I just don't call friends and like every 10 years, I do something dramatic for them so that we maintain the friendship.

Like I'd murder somebody that they really don't like.

Careful man, Joey might ask you to do that.

Lex, I need you to come down to New Jersey.

I don't know if you listen to Joe Rogan, all those folks.

They kind of are doing something interesting for MIT and academia.

They're shaking up this world a little bit.

Like podcasting, because comedians are paving the way for podcasting.

And so you have like Andrew Huberman, who's a neuroscientist at Stanford, a friend of mine now.

He's like into podcasting now, and you're into podcasting.

Of course, you're not necessarily podcasting about computer science currently, right?

But that, it feels like you could have a lot of the free spirit of the comedians implemented by the people who are academically trained.

Yeah, and then that results, I mean, who knows what the experiment looks like, but that results me being able to talk about robotics with Joey Diaz when he says, you know, drops F-bombs every other sentence.

And I, the world is, like, I've seen actually a shift within colleagues and friends within MIT where they're becoming much more accepting of that kind of thing.

Because they're seeing how popular it is.

I don't know how they think about it at Georgetown, for example.

It's interesting, but I think what happens is the popularity of it combined with just good conversations with people they respect.

And this is more fun to listen to than a shitty Zoom lecture about their work.

And we don't, nobody actually knows what that is.

Just like with like Clubhouse or something, nobody's figured out like, where is this medium take?

Is this a legitimate medium of education?

Well, that's your innovation, I think, was we can bring on professors.

And I know Joe Rogan did some of that too, but you know, but your professors in your field.

You bring on all these MIT guys who I remember.

Well, that's been the big challenge for me is I don't, is I feel I would ask big philosophical questions of people like yourself that are really well, like so for example, you have a lot of excellent papers that has a lot of theory in it, right?

And there is some temptation to just go through papers.

And I think it's possible to actually do that.

I haven't done that much, but I think it's possible.

And I can probably only do that with things that I'm actually in the field I'm aware of.

But there's a dance that I would love to be able to try to hit right, where it's actually getting to the core of some interesting ideas as opposed to just talking about philosophy.

At the same time, there's a large audience of people that just want to be inspired by disciplines where they don't necessarily know the details.

But there's a lot of people that are like, I'm really curious.

I've been thinking about pivoting careers into software engineering.

They would love to hear from people like you about computer science, even if it's like theory.

Yeah, but just like the idea that you can have big ideas, you push them through and it's interesting, you fight for it,

Well, there's some, there's, what is it, Computerphile and Numberphile, these YouTube channels.

There's channels I watch on chess, exceptionally popular, where I don't understand maybe 80% of the time what the hell they're talking about, because they're talking about why this move is better than this move, but I love the passion and the genius of those people and just overhearing it.

Or do you look at like Scott Aronson's blog at all, the Shuttle to Optimize?

Yeah, it's like the hardcore complexity theory, but it's just an enthusiasm or like Terry Tao's blog.

He would, and it would just be, I'm going all in on, you know, here's a new affine group with which you can do whatever.

Well, in the case of Scott Harrison, he's good.

He's able to turn on like the inner troll and comedian and so on.

He keeps the fun, which is the best of kinds

And he's a philosophical guy, he wrote that book.

Yeah, so we're exploring these different ways of communicating science and exciting the world.

Speaking of which, I gotta ask you about computer science.

So, I mean, a lot of your work is what inspired this deep thinking about productivity from all the different angles because some of the most rigorous work is mathematical work and in computer science, the theoretical computer science.

Let me ask the Scott Aronson question of like, is there something to you that stands out in particular that's beautiful or inspiring or just really insightful about computer science or maybe mathematics?

And in particular, what I've always liked in theory is the notion of impossibilities.

So within the context of distributed algorithms, my specialty is impossibility results.

So the idea that you can argue, nothing exists that solves this, or nothing exists that can solve this faster than this.

And that's, I think that's really interesting.

And that goes all the way back to Turing.

There's his original paper on computable numbers with their connection to the, it's in German, the Eichler-Tung problem, but basically the German name that Hilbert called the decision problem.

This was pre-computers, but he was, you know, he's English, so it's written in English, so it's a very accessible paper.

And it lays the foundation for all of theoretical computer science.

He's like, well, if we think about like an algorithm, I mean, he figures out like all effective procedures or Turing machines are basically algorithms.

We could really describe a Turing machine with a number, which we can now imagine with like computer code, you could just take a source file and just treat the binary version of the file as like a really long number, right?

But he's like, every program is just a finite number.

And then he realized like 1 way to think about a problem is you have, and this is like kind of the Mike Sipser approach, but you have a sort of, it's a language.

So of infinite number of strings, some of them are in the language and some of them aren't, but basically you can imagine a problem is represented as an infinite binary string, where in every position, like a 1 means that string is in the language and a 0 means it isn't.

And then he applied Cantor from the 19th century and said, okay, the natural numbers are countable, so it's countably infinite, and infinite binary strings, you can use a diagonalization argument and show they're uncountable.

So there's just vastly more problems than there are algorithms.

So basically anything you can come up with for the most part almost certainly is not solvable by a computer.

You know, and then he was like, let me give a particular example, and he figured out the very first computability proof.

And he said, Let's just walk through with a little bit of simple logic.

The halting problem can't be solved by an algorithm.

And that kicked off the whole enterprise of some things can't be solved by algorithms.

Some things can't be solved by computers.

And we've just been doing theory on that since that was the 30s he wrote that.

So proving that something is impossible is sort of a more, a stricter version of that.

Is it like proving bounds on the performance of different algorithms?

So you say this algorithm does at least this well and no worse than this, but you're looking at a particular algorithm.

And possibility proofs say no algorithm ever could ever solve this problem.

So no algorithm could ever solve the halting problem.

It's making something different, making a conclusive statement about the problem.

And that's somehow satisfying because it's- It's just philosophically interesting.

I mean, it all goes back to, you get back to Plato, it's all reducto ad absurdum.

The only way to do it is because there's an infinite number of solutions you can't go through them.

You say, let's assume for the sake of contradiction that there existed something that solves this problem, and then you turn to crank a logic until you blow up the universe.

And then you go back and say, okay, our original assumption that this solution exists can't be true.

I just think philosophically, it's like a really exciting kind of beautiful thing.

It's what I specialize in within distributed algorithms is more like time bound impossibility results.

Like no algorithm can solve this problem faster than this in this setting.

Of all the infinite number of ways you might ever do it.

So you have many papers, but the 1 that caught my eye is Smooth Analysis of Dynamic Networks, in which you write, a problem with the worst case perspective is that it often leads to extremely strong lower bounds.

These strong results motivate a key question.

Is this bond robust in the sense that it captures the fundamental difficulty introduced by dynamism?

Or is the bond fragile in the sense that the poor performance it describes depends on an exact sequence of adversarial changes?

Fragile lower bounds leave open the possibility of algorithms that might still perform well in practice.

That's in the sense of the impossible and the bounds discussion presents the interesting question.

I just like the idea of robust and fragile bounds, but what do you make about this kind of tension between what's provably, like what bounds you can prove that are like robust and something that's a bit more fragile.

And also by way of answering that for this particular paper, can you say what the hell are dynamic networks?

Yeah, well, OK, so smooth analysis, it's, so it wasn't my idea.

So Spielman and Tang came up with this in the context of sequential algorithms.

So just like the normal world of an algorithm that runs on a computer.

And they were looking at, there's a well-known algorithm called the simplex algorithm, but basically you're trying to find a hole around a group of points.

And there was an algorithm that worked really well in practice.

But when you analyze it, you would say, I can't guarantee it's gonna work well in practice because if you have just the right inputs, this thing could run really long, right?

But in practice, it seemed to be really fast.

So smooth analysis is they came in and they said, let's assume that a bad guy chooses the inputs.

It could be anything like really bad ones.

And all we're gonna do, because in simplex, they're numbers.

We're gonna just randomly put a little bit of noise on each of the numbers.

And they show if you put a little bit of noise on the numbers, suddenly simplex algorithm goes really fast.

Like, oh, that explains this lower bound, this idea that it could sometimes run really long was a fragile bound because it could only run a really long time if you had exactly the worst pathological input.

So then my collaborators and I brought this over to the world of distributed algorithms.

We brought them over the general lower bounds, right?

So in the world of dynamic networks, so distributed algorithm is a bunch of algorithms on different machines talking to each other, trying to solve a problem.

So you imagine them connected with network links and a dynamic network, those can change, right?

So I was talking to you, but now I can't talk to you anymore.

It's a really hard environment, mathematically speaking.

And there's a lot of really strong lower bounds, which you can imagine if the network can change all the time and a bad guy's doing it, it's like hard to do things well.

So there's an algorithm running on every single node in the network.

And then you're trying to say something of any kind that makes any kind of definitive sense about the performance of that algorithm.

Yeah, so I just submitted a new paper on this a couple weeks ago, and we were looking at a very simple problem.

If the network doesn't change, you can do this pretty well.

There's some algorithms that work, basic algorithms that work really well.

If the network can change every round, there's these lower bounds that says it takes a really long time.

There's a way that like no matter what algorithm you come up with, there's a way the network can change in such a way that just really slows down your progress basically, right?

So smooth analysis there says, yeah, but that seems like a really, you know, really bad luck if your network was changing like exactly in the right way that you needed to screw your algorithm.

So we said, what if we randomly just add or remove a couple edges in every round?

So the adversary is trying to choose the worst possible network.

I mean, it's a blinded submission, so maybe I shouldn't, it's not, whatever.

We basically showed- An anonymous friend of yours

So that even just adding like 1 random edge per round, and here's the cool thing about the simplest possible solution to this problem blows away that lower bound and does really well.

So that's like a very fragile lower bound because we're like, it's almost impossible to actually keep things slow.

I wonder how many lower bounds you can smash open with this kind of analysis and show that they're fragile.

Because in distributed algorithms, there's a ton of really famous strong lower bounds, but things have to go wrong, really, really wrong, for these lower bound arguments to work.

So this whole notion of fragile versus robust, I was like, well, let's go in and just throw a little noise in there.

And if it becomes solvable, then maybe that lower bound wasn't really something we should worry about.

That's really embarrassing to a lot of people.

Because, okay, this is the OCD thing with the spaces, is it feels really good when you can prove a nice bound.

And if you say that that bound is fragile, that's like, there's going to be a sad kid that walks like with their lunchbox back home, like, my lower bottom doesn't matter.

It's all, I don't know, it feels like to me a lot of this theory is just math machismo.

So if you show that something is fragile, that's more important, that's really important in practice, right?

So do you think kind of theoretical computer science is living in its own world, just like mathematics, and their main effort, which I think is very valuable, is to develop ideas.

It's not necessarily interesting, whether it's applicable in the real world.

Yeah, we don't care about the applicability.

And we're terrible with computers and can't do anything useful with computers and we don't know how to code.

And We're not productive members of technological society, but I do think things percolate.

You percolate from the world of theory into the world of algorithm design where it will pull on the theory and now suddenly it's useful.

Then the algorithm design gets pulled into the world of practice where they say, well, actually we can make this algorithm a lot better because in practice, really these servers do XYZ and now we can make this super efficient.

And so I do think, I mean, I tell my, I teach theory to the PhD students at Georgetown.

I show them the sort of funnel of like, okay, we're over here doing theory, but it eventually, some of this stuff will percolate down and affect at the very end, you know, a phone.

But the very question you're asking at the highest philosophical level is fascinating.

Like if you take a system, a distributed system or a network and introduce a little bit of noise into it.

Like how many problems of that nature are fundamentally changed by that little introduction of noise?

Yeah, because it's all, especially in distributed algorithms, the model is everything.

Like the way we work is we're incredibly precise about here's exactly, it's mathematical.

Here's exactly how the network works and it's a state machine, algorithms are state machines.

We're super precise, we can prove lower bounds.

But yeah, often those lower, those impossibility results really get at the hard edges of exactly how that model works.

So we'll see if this, so we published a paper on this, that paper you mentioned, that kind of introduced the idea to distributed algorithms world.

And I think that's got some traction and there's been some follow-ups.

I mean, honestly, the issue with the next is that the result fell out so easily, and this shows the mathematical machismo problem in these fields, is there's a good chance the paper won't be accepted because there wasn't enough mathematical self-flagellation.

So even showing that very few, just very little bit of noise, can have a dramatic, make a dramatic statement about the fragility.

big surprise to us, but once we figured out how to show it, it's not too hard.

And these are venues for theoretical work.

So the fascinating tension that exists in other disciplines, like 1 of them is machine learning, which despite the power of machine learning and deep learning and all like the impact of it in the real world, the main conferences on machine learning are still resistant to application papers.

And I'm not sort of, and application papers broadly defined, meaning like finding almost like you would, like Darwin did by like going around, collecting some information, saying, huh, isn't this interesting?

Like those are some of the most popular blogs and yet as a paper it's not really accepted.

I wonder what you think about this whole world of deep learning from a perspective of theory.

What do you make of this whole discipline of the success of neural networks, of how to do science on them?

Are you excited by the possibilities of what we might discover about neural networks?

Do you think it's fundamental in engineering discipline?

Or is there something theoretical that we might crack open 1 of these days in understanding something deep about how optimization and how systems learn.

See all Lex Fridman transcripts on Youtube

Cal Newport: Deep Work, Focus, Productivity, Email, and Social Media | Lex Fridman Podcast #166