WEBVTT

00:00:00.000 --> 00:00:00.500
[TEASER]   
[MUSIC PLAYS UNDER DIALOGUE] 

00:00:00.500 --> 00:00:06.600
BEHNAZ ARZANI: I guess the thing I'm seeing is 
that we are freed up to dream more—in a way.  

00:00:06.600 --> 00:00:11.760
Maybe that's me being too … I'm a little bit of a 
romantic, so this is that coming out a little bit,  

00:00:11.760 --> 00:00:18.960
but it's, like, because of all this, we have 
the time to think bigger, to dream bigger,  

00:00:18.960 --> 00:00:25.484
to look at problems where maybe five years 
ago, we wouldn't even dare to think about. 

00:00:25.484 --> 00:00:27.400
[TEASER ENDS]
GRETCHEN HUIZINGA: You’re listening to Ideas,

00:00:27.400 --> 00:00:32.800
a Microsoft Research Podcast that dives deep into 
the world of technology research and the profound  

00:00:32.800 --> 00:00:45.600
questions behind the code. I'm Dr. Gretchen 
Huizinga. In this series, we'll explore the  

00:00:45.600 --> 00:00:50.490
technologies that are shaping our future and 
the big ideas that propel them forward.  

00:00:50.490 --> 00:00:51.880
[MUSIC FADES] 

00:00:51.880 --> 00:00:57.440
My guest today is Behnaz Arzani. Behnaz is a 
principal researcher at Microsoft Research,  

00:00:57.440 --> 00:01:02.360
and she's passionate about the systems and 
networks that provide the backbone to nearly all  

00:01:02.360 --> 00:01:08.280
our technologies today. Like many in her field, 
you may not know her, but you know her work:  

00:01:08.280 --> 00:01:13.040
when your networks function flawlessly, you 
can thank people like Behnaz Arzani. Behnaz,  

00:01:13.040 --> 00:01:17.505
it's been a while. I am so excited to catch 
up with you today. Welcome to Ideas! 

00:01:17.505 --> 00:01:19.880
BEHNAZ ARZANI: Thank you. And 
I'm also excited to be here. 

00:01:19.880 --> 00:01:26.080
HUIZINGA: So since the show is about ideas and 
leans more philosophical, I like to start with  

00:01:26.080 --> 00:01:30.760
a little personal story and try to tease out 
anything that might have been an inflection  

00:01:30.760 --> 00:01:38.960
point in your life, a sort of aha moment, or 
a pivotal event, or an animating “what if,” we  

00:01:38.960 --> 00:01:44.600
could call it. What captured your imagination and 
got you inspired to do what you're doing today? 

00:01:44.600 --> 00:01:50.080
ARZANI: I think that it was a little bit of an 
accident and a little bit of just chance, I guess,  

00:01:50.080 --> 00:01:57.040
but for me, this happened because I don't like 
being told what to do! [LAUGHTER] I really hate  

00:01:57.040 --> 00:02:02.640
being told what to do. And so, I got into research 
by accident, mostly because it felt like a job  

00:02:02.640 --> 00:02:07.920
where that wouldn't happen. I could pick what I 
wanted to do. So, you know, a lot of people come  

00:02:07.920 --> 00:02:13.840
talking about how they were the most curious kids 
and they all—I wasn't that. I was a nerd, but I  

00:02:13.840 --> 00:02:20.320
wasn't the most curious kid. But then I found 
that I'm attracted to puzzles and hard puzzles  

00:02:20.320 --> 00:02:25.760
and things that I don't know how to answer, and 
so that gravitated me more towards what I'm doing  

00:02:25.760 --> 00:02:29.840
today. Things that are basically difficult 
to solve … I think are difficult to solve. 

00:02:29.840 --> 00:02:33.970
HUIZINGA: So that's your inspiring 
moment? “I'm a bit of a rebel, and …” 

00:02:33.970 --> 00:02:35.152
ARZANI: Yup!
HUIZINGA: … I like puzzles … ”?

00:02:35.152 --> 00:02:39.800
ARZANI: Yup! [LAUGHTER] Which is not really a 
moment. Yeah, I can't point to a moment. It's  

00:02:39.800 --> 00:02:43.920
just been a journey, and it's just, like, been 
something that has gradually happened to me,  

00:02:43.920 --> 00:02:49.320
and I love where I am, but I can't really pinpoint 
to like this, like this inspiring awe-drop—no. 

00:02:49.320 --> 00:02:51.280
HUIZINGA: OK. So let me ask you this:  

00:02:51.280 --> 00:02:54.992
is there nobody in this building that 
tells you what to do? [LAUGHS] 

00:02:54.992 --> 00:02:58.990
ARZANI: There are people who 
have tried, [LAUGHS] but … 

00:02:58.990 --> 00:02:59.660
HUIZINGA: Oh my gosh! 

00:02:59.660 --> 00:03:03.180
ARZANI: No, it doesn't work. And I think if you 
ask them, they will tell you it hasn't worked. 

00:03:03.180 --> 00:03:09.940
HUIZINGA: OK. The other side question is, have you 
encountered a puzzle that has confounded you? 

00:03:09.940 --> 00:03:14.750
ARZANI: Have I encountered a puzzle? 
Yes. Incident management. [LAUGHTER] 

00:03:14.750 --> 00:03:18.640
HUIZINGA: And we'll get there in the 
next couple of questions. Before we do,  

00:03:18.640 --> 00:03:23.800
though, I want to know about who might 
have influenced you earlier. I mean,  

00:03:23.800 --> 00:03:29.432
it's interesting. Usually if you don't have a 
what, there might not be a who attached to it … 

00:03:29.432 --> 00:03:32.186
ARZANI: No. But I have a who. I 
have multiple “whos” actually. 

00:03:32.186 --> 00:03:36.660
HUIZINGA: OK! Wonderful. So tell us a little bit 
about the influential people in your life. 

00:03:36.660 --> 00:03:42.040
ARZANI: I think the first and foremost is my mom. 
I have a necklace I'm holding right now. This is  

00:03:42.040 --> 00:03:46.760
something my dad gave my mom on their wedding day. 
On one side of it is a picture of my mom and dad;  

00:03:46.760 --> 00:03:51.000
on the other side is both their names on it. 
And I have it on every day. To my mom’s chagrin.  

00:03:51.000 --> 00:03:58.120
[LAUGHTER] She is like, why? But it's, like, it 
helps me stay grounded. And my mom is a person  

00:03:58.120 --> 00:04:03.840
that … she had me while she was an undergrad. She 
got her master’s. She got into three different PhD  

00:04:03.840 --> 00:04:08.840
programs in her lifetime. Every time, she gave 
it up for my sake and for my brother's sake. But  

00:04:08.840 --> 00:04:13.240
she's a woman that taught me you can do anything 
you set your mind to and that you should always be  

00:04:13.240 --> 00:04:17.320
eager to learn. She was a chemistry teacher, 
and even though she was a chemistry teacher,  

00:04:17.320 --> 00:04:22.040
she kept reading new books. She came to the 
US to visit me in 2017, went to a Philadelphia  

00:04:22.040 --> 00:04:25.720
high school, and asked, can I see your chemistry 
books? I want to see what you're teaching your  

00:04:25.720 --> 00:04:32.160
kids. [LAUGHTER] So that's how dedicated she is 
to what she does. She loves what she does. And I  

00:04:32.160 --> 00:04:37.320
could see it on her face on a daily basis. And at 
some point in my life a couple of years ago, I was  

00:04:37.320 --> 00:04:41.200
talking to my mom about something, and she said, 
tell yourself, “I'm stronger than my mom.” 

00:04:41.200 --> 00:04:42.060
HUIZINGA: Oh my gosh. 

00:04:42.060 --> 00:04:46.280
ARZANI: And that has been, like, the most 
amazing thing to have in the back of my  

00:04:46.280 --> 00:04:49.760
head because I view my mom as one of 
the strongest people I've ever met,  

00:04:49.760 --> 00:04:52.180
and she's my inspiration for everything I do. 

00:04:52.180 --> 00:04:56.060
HUIZINGA: Tell yourself you're 
stronger than your mom. … Did you? 

00:04:56.060 --> 00:04:58.741
ARZANI: I'm not stronger than my 
mom, I don't think … [LAUGHS] 

00:04:58.741 --> 00:05:00.260
HUIZINGA: [LAUGHS] You got 
to change that narrative! 

00:05:00.260 --> 00:05:03.200
ARZANI: But, yes, I think it's just this thing of,  

00:05:03.200 --> 00:05:06.780
like, “What would Mom do?” is a great 
thing to ask yourself, I think. 

00:05:06.780 --> 00:05:11.560
HUIZINGA: I love that. Well, and so 
I would imagine, though, that post-,  

00:05:11.560 --> 00:05:16.200
you know, getting out of the house, you've 
had instructors, you've had professors,  

00:05:16.200 --> 00:05:20.992
you've had other researchers. I 
mean, anyone else that's … ? 

00:05:20.992 --> 00:05:25.000
ARZANI: Many! And in different stages of your 
life, different people step into that role,  

00:05:25.000 --> 00:05:30.840
I feel like. One of the first people 
for me was Jen Rexford, and she is  

00:05:30.840 --> 00:05:37.280
just an amazing human being. She's an amazing 
researcher, hands down. Her work is awesome,  

00:05:37.280 --> 00:05:42.100
but also, she's an amazing human being, as 
well. And that just makes it better.  

00:05:42.100 --> 00:05:42.680
HUIZINGA: Yeah. 

00:05:42.680 --> 00:05:47.160
ARZANI: And then another person is Mohammad 
Alizadeh, who's at MIT. And actually, let’s see,  

00:05:47.160 --> 00:05:50.720
I'm going to keep going a little with 
people—Mark Handley. When I was a PhD  

00:05:50.720 --> 00:05:55.360
student, I would read their papers, and I'd 
be like, wow! And, I want to be like you! 

00:05:55.360 --> 00:05:59.640
HUIZINGA: So linking that back to your love 
of puzzles, were these people that you admired  

00:06:00.200 --> 00:06:02.112
good problem solvers or … ? 

00:06:02.112 --> 00:06:07.880
ARZANI: Oh, yeah! I think Jen is one of those 
who … a lot of her work is also practical,  

00:06:07.880 --> 00:06:12.120
like, you know, straddles a line between both 
solving the puzzle and being practical and being  

00:06:12.120 --> 00:06:16.480
creative and working with theorists and working 
with PL people. So she's also collaborative,  

00:06:16.480 --> 00:06:20.560
which is, kind of, my style of work, as 
well. Mohammad is more of a theorist,  

00:06:20.560 --> 00:06:26.160
and I love … like more the theoretical aspect 
of problems that I solve. And so, like,  

00:06:26.160 --> 00:06:30.520
just the fact that he was able to look at those 
problems and thinks about those problems in those  

00:06:30.520 --> 00:06:35.480
ways. And then Mark Handley’s intuition about 
problems—yeah, I can't even speak to that! 

00:06:35.480 --> 00:06:39.160
HUIZINGA: That's so fascinating because you've 
identified three really key things for a  

00:06:39.160 --> 00:06:45.000
researcher. And each one is embodied in a person. 
I love that. And because I know who you are,  

00:06:45.000 --> 00:06:48.920
I know we're going to get to each of those things 
probably in the course of all these questions  

00:06:48.920 --> 00:06:54.440
that I'll ask you. [LAUGHTER] So we just spent a 
little time talking about what got you here and  

00:06:54.440 --> 00:06:59.720
who influenced you along the way. But your life 
isn't static. And at each stage of accomplishment,  

00:06:59.720 --> 00:07:03.480
you get a chance to reflect and, sort 
of, think about what you got right,  

00:07:03.480 --> 00:07:08.160
what you got wrong, and where you want to go 
next. So I wonder if you could take a minute  

00:07:08.160 --> 00:07:12.560
to talk about the evolution of your 
values as a researcher, collaborator,  

00:07:12.560 --> 00:07:17.024
and colleague and then a sort of “how 
it started/how it's going” thing. 

00:07:17.024 --> 00:07:24.080
ARZANI: Hmm … For me, I think what I've learned 
is to be more mindful—about all of it. But I  

00:07:24.080 --> 00:07:27.680
think if I talk about the evolution, when 
you're a PhD student, especially if you're  

00:07:27.680 --> 00:07:31.960
a PhD student from a place that's not MIT, 
that's not Berkeley, which is where I was from,  

00:07:31.960 --> 00:07:36.280
my main focus was proving myself. I mean, for 
women, always, we have to prove ourselves. But,  

00:07:36.280 --> 00:07:40.080
like, I think if you're not from one of those 
schools, it's even more so. At least that's how  

00:07:40.080 --> 00:07:47.640
I felt. That might not be the reality, but that's 
how you feel. And so you're always running to show  

00:07:47.640 --> 00:07:53.400
this about yourself. And so you don't stop 
to think how you're showing up as a person,  

00:07:53.400 --> 00:07:58.080
as a researcher, as a collaborator. You're 
not even, like, necessarily reflecting on,  

00:07:58.080 --> 00:08:02.840
are these the problems that I enjoy solving? 
It’s more of, will solving this problem help  

00:08:02.840 --> 00:08:08.840
me establish myself in this world that requires 
proving yourself and is so critical and all of  

00:08:08.840 --> 00:08:16.600
that stuff? I think now I stop more. I think more, 
is this a problem that I would enjoy solving?  

00:08:16.600 --> 00:08:21.800
I think that's the most important thing. Would 
other people find it useful? Is it solving a hard  

00:08:21.800 --> 00:08:29.440
technical question? And then, in collaborations, 
I'm being more mindful that I show up in a way  

00:08:30.120 --> 00:08:35.920
that basically allows me to be a good person 
the way I want to be in my collaboration. So  

00:08:35.920 --> 00:08:40.800
as researchers, we have to be critical because 
that's how science evolves. Not all work is  

00:08:40.800 --> 00:08:46.640
perfect. Not all ideas are the best ideas. That's 
just fundamental truth. Because we iterate on each  

00:08:46.640 --> 00:08:52.800
other's ideas until we find the perfect solution 
to something. But you can do all of these things  

00:08:52.800 --> 00:08:58.200
in a way that's kind, in a way that's mindful, 
in a way that respects other people and what they  

00:08:58.200 --> 00:09:03.080
bring to the table. And I think what I've learned 
is to be more mindful about those things. 

00:09:03.080 --> 00:09:07.080
HUIZINGA: How would you define mindful? 
That's an interesting word. It has a lot  

00:09:07.080 --> 00:09:11.560
of baggage around it, you know, in terms 
of how people do mindfulness training.  

00:09:11.560 --> 00:09:14.800
Is that what you're talking about, or 
is it more, sort of, intentional? 

00:09:14.800 --> 00:09:19.120
ARZANI: I think it's both. So I think one of the 
things I said—I think when I got into this booth  

00:09:19.120 --> 00:09:23.040
even—was, I'm going to take a breath before I 
answer each question. And I think that's part  

00:09:23.040 --> 00:09:29.640
of it, is just taking a breath to make sure you're 
present is part of it. But I think there is more  

00:09:29.640 --> 00:09:34.640
to it than that, which is I don't think we even 
think about it. I think if I … when you asked me  

00:09:34.640 --> 00:09:39.000
about the evolution of how I evolved, I never 
thought about it. I was just, like, running to  

00:09:39.000 --> 00:09:43.680
get things done, running to solve the question, 
running to, you know, find the next big thing,  

00:09:43.680 --> 00:09:48.800
and then you're not paying attention to how 
you're impacting the world in the process. 

00:09:48.800 --> 00:09:49.400
HUIZINGA: Right. 

00:09:49.400 --> 00:09:54.680
ARZANI: And once you start paying attention, then 
you're like, oh, I could do this better. I can do  

00:09:54.680 --> 00:09:59.800
that better. If I say this to this person in 
that way, that allows them to do so much more,  

00:09:59.800 --> 00:10:01.860
that encourages them to do so much more. 

00:10:01.860 --> 00:10:02.680
HUIZINGA: Yeah, yeah.  

00:10:02.680 --> 00:10:03.430
ARZANI: So …  

00:10:03.430 --> 00:10:07.360
HUIZINGA: You know, when you started out, 
you said, is this a problem I would enjoy  

00:10:07.360 --> 00:10:10.840
solving? And then you said, is this 
a problem that somebody else needs to  

00:10:10.840 --> 00:10:15.880
have solved? Which is sort of like “do I like 
it?"—it goes back to Behnaz at the beginning:  

00:10:15.880 --> 00:10:21.120
don't tell me what to do; I want to 
do what I want to do. Versus—or and  

00:10:21.120 --> 00:10:26.640
is this useful to the world? And I feel like 
those two threads are really key to you. 

00:10:26.640 --> 00:10:31.520
ARZANI: Yes. Basically, I feel like that defines 
me as a researcher, pretty much. [LAUGHS] Which  

00:10:31.520 --> 00:10:38.160
is, you know, I was one of the, you know, early 
people … I wouldn’t say first. I'm not the first,  

00:10:38.160 --> 00:10:42.560
I don't think, but I was one of the early people 
who was talking about using machine learning in  

00:10:42.560 --> 00:10:45.840
networking. And after a while, I stopped 
because I wasn't finding it fun anymore,  

00:10:45.840 --> 00:10:50.680
even though there was so much hype about, you 
know, let's do machine learning in networking.  

00:10:50.680 --> 00:10:53.520
And it's not because there's not a lot 
of technical stuff left to do. You can  

00:10:53.520 --> 00:10:57.820
do a lot of other things there. There's room 
to innovate. It's just that I got bored. 

00:10:57.820 --> 00:11:03.480
HUIZINGA: I was just going to say, it's still 
cool, but Behnaz is bored! [LAUGHTER] OK, well,  

00:11:03.480 --> 00:11:08.360
let's start to talk a little bit about some of the 
things that you're doing. And I like this idea of  

00:11:08.360 --> 00:11:12.520
a researcher, even a person, having a North Star 
goal. It sounds like you've got them in a lot of  

00:11:12.520 --> 00:11:18.880
areas of your life, and you've said your North 
Star goal, your research goal, is to make the life  

00:11:18.880 --> 00:11:25.360
of a network operator as painless as possible. So 
I want to know who this person is. Walk us through  

00:11:25.360 --> 00:11:31.060
a day in the life of a network operator and tell 
us what prompted you to want to help them. 

00:11:31.060 --> 00:11:36.280
ARZANI: OK, so it's been years since I actually, 
like, sat right next to one of them for a long  

00:11:36.280 --> 00:11:39.200
extended period of time because now we're in 
different buildings, but back when I was an  

00:11:39.200 --> 00:11:43.360
intern, I was actually, like, kind of, like 
right in the middle of a bunch of, you know,  

00:11:43.360 --> 00:11:48.120
actual network operators. And what I observed 
… and see, this was not, like, I've never lived  

00:11:48.120 --> 00:11:51.540
that experience, so I'm talking about somebody 
else's experience, so bear that in mind … 

00:11:51.540 --> 00:11:52.832
HUIZINGA: Sure, but at least you saw it … 

00:11:52.832 --> 00:11:59.400
ARZANI: Yeah. What they do is, there's a lot of, 
“OK, we design the network, configure it.” A lot  

00:11:59.400 --> 00:12:03.760
of it goes into building new systems to manage it. 
Building new systems to basically make it better,  

00:12:03.760 --> 00:12:08.120
more efficient, all of that. And then they 
also have to be on call so that when any of  

00:12:08.120 --> 00:12:12.480
those things break, they're the ones who have 
to look at their monitoring systems and figure  

00:12:12.480 --> 00:12:18.720
out what happened and try to fix it. So they 
do all of this in their day-to-day lives. 

00:12:18.720 --> 00:12:20.592
HUIZINGA: That's tough … 

00:12:20.592 --> 00:12:21.200
ARZANI: Yeah. 

00:12:21.200 --> 00:12:27.080
HUIZINGA: OK. So I know you have a story about 
what prompted you, at the very beginning,  

00:12:27.080 --> 00:12:31.992
to want to help this person. And it had 
some personal implications. [LAUGHS] 

00:12:31.992 --> 00:12:35.400
ARZANI: Yeah! So my internship 
mentor, who's an amazing person,  

00:12:35.400 --> 00:12:40.800
I thought—and this is, again, my perception 
as an intern—the day after he was on call,  

00:12:40.800 --> 00:12:46.720
he was so tired, I felt. And so grumpy … 
grumpier than normal! [LAUGHTER] And, like,  

00:12:46.720 --> 00:12:51.120
my main motivation initially for working in this 
space was just, like, make his life better! 

00:12:51.120 --> 00:12:52.460
HUIZINGA: Make him not grumpy. 

00:12:52.460 --> 00:12:53.950
ARZANI: Yeah. Pretty much. [LAUGHS] 

00:12:53.950 --> 00:12:57.200
HUIZINGA: Did you have success at that 
point in your life? Or was this just,  

00:12:57.200 --> 00:12:59.860
like, setting a North Star goal 
that I'm going to go for that? 

00:12:59.860 --> 00:13:03.920
ARZANI: I mean, I had done a lot of work 
in monitoring space, but back then—again,  

00:13:03.920 --> 00:13:07.880
going back to the talk we were having 
about how to be mindful about problems  

00:13:07.880 --> 00:13:12.160
you pick—back then it was just like, oh, this 
was a problem to solve, and we'll go solve it,  

00:13:12.160 --> 00:13:16.440
and then what's the next thing? So there was not 
an overarching vision, if you will. It was just,  

00:13:16.440 --> 00:13:21.440
like, going after the next, after the next. I 
think that's a point where, like, it all came  

00:13:21.440 --> 00:13:26.840
together of like, oh, all of the stuff that I'm 
doing can help me achieve this bigger thing. 

00:13:26.840 --> 00:13:32.360
HUIZINGA: Right. OK, Behnaz, I want to 
drop anchor, to use a seafaring analogy,  

00:13:32.360 --> 00:13:38.880
for a second and contextualize the language that 
these operators use. Give us a “networking for  

00:13:38.880 --> 00:13:43.480
neophytes” overview of the tools they 
rely on and the terminology they use in  

00:13:43.480 --> 00:13:47.640
their day-to-day work so we're not lost 
when we start to unpack the problems,  

00:13:47.640 --> 00:13:50.180
projects, and papers that 
are central to your work. 

00:13:50.180 --> 00:13:55.360
ARZANI: OK. So I'm going to focus on my pieces of 
this just because of the context of this question.  

00:13:55.360 --> 00:14:00.720
But a lot of operators … just because a lot of 
the problems that we work on these days to be able  

00:14:00.720 --> 00:14:05.960
to manage our network, the optimal form of these 
problems tend to be really, really hard. So a lot  

00:14:05.960 --> 00:14:12.920
of the times, we use algorithms and solutions that 
are approximate forms of those optimal solutions  

00:14:12.920 --> 00:14:17.520
in order to just solve those problems faster. And 
a lot of these heuristics, some of them focus on  

00:14:17.520 --> 00:14:23.160
our wide area network, which we call a WAN. Our 
WANs, basically what they do is they move traffic  

00:14:23.160 --> 00:14:29.120
between datacenters in a way that basically fits 
the capacity of our network. And, yeah, I think  

00:14:29.120 --> 00:14:33.320
for my work, my current work, to understand it, 
that's, I think, enough networking terminology. 

00:14:33.320 --> 00:14:38.960
HUIZINGA: OK. Well, so you've used the 
term heuristic and optimal. Not with  

00:14:38.960 --> 00:14:43.312
an “s” on the end of it. Or you do 
say “optimals,” but it's a noun … 

00:14:43.312 --> 00:14:47.320
ARZANI: Well, so for each problem 
definition, usually, there's one way  

00:14:47.320 --> 00:14:52.720
to formulate an optimal solution. There 
might be multiple optima that you find,  

00:14:52.720 --> 00:14:56.880
but the algorithm that finds the optimum 
usually is one. But there might be many,  

00:14:56.880 --> 00:14:59.800
I guess. The ones that I've worked 
on generally have been one. 

00:14:59.800 --> 00:15:05.560
HUIZINGA: Yeah, yeah. And so in terms of how 
things work on a network, can you give us just  

00:15:05.560 --> 00:15:10.360
a little picture of how something moves 
from A to B that might be a problem? 

00:15:10.360 --> 00:15:17.160
ARZANI: So, for example, we have these datacenters 
that generate terabytes of traffic and—terabytes  

00:15:17.160 --> 00:15:22.600
per second of traffic—that wants to move 
from point A to point B, right. And we only  

00:15:22.600 --> 00:15:28.400
have finite network capacity, and these, what we 
call, “demands” between these datacenters—and you  

00:15:28.400 --> 00:15:33.760
didn't see me do the air quotes, but I did the 
air quotes—so they go from point A to point B,  

00:15:33.760 --> 00:15:40.120
and so in order to fit this demand in the pipes 
that we have—and these pipes are basically links  

00:15:40.120 --> 00:15:45.920
in our network—we have to figure out how to 
send them. And there's variations in them. So,  

00:15:45.920 --> 00:15:48.360
like, it might be the case that 
at a certain time of the day,  

00:15:48.360 --> 00:15:53.280
East US would want to send more traffic to West 
US, and then suddenly, it flips. And that's why  

00:15:53.280 --> 00:15:58.040
we solve this problem every five minutes! Now 
assume one of these links suddenly goes down.  

00:15:58.040 --> 00:16:03.160
What do I do? I have to resolve this problem 
because maybe the path that I initially picked  

00:16:03.160 --> 00:16:07.600
for traffic to go through goes exactly through 
that failed link. And now that it's disappeared,  

00:16:07.600 --> 00:16:11.800
all of that traffic is going to fall on the floor. 
So I have to re-solve that problem really quickly  

00:16:11.800 --> 00:16:16.880
to be able to re-move my traffic and move it 
to somewhere else so that I can still route  

00:16:16.880 --> 00:16:21.200
it and my customers aren't impacted. What 
we're talking about here is a controller,  

00:16:21.200 --> 00:16:26.480
essentially, that the network operators built. 
And this controller solves this optimization  

00:16:26.480 --> 00:16:32.640
problem that figures out how traffic should move. 
When it's failed, then the same controller kicks  

00:16:32.640 --> 00:16:37.880
in and reroutes traffic. The people who built 
that controller are the network operators. 

00:16:37.880 --> 00:16:42.680
HUIZINGA: And so who does the problem-solving 
or the troubleshooting on the fly? 

00:16:42.680 --> 00:16:45.840
ARZANI: So hopefully—and this, most of the times,  

00:16:45.840 --> 00:16:50.400
is the case—is we have monitoring systems in 
place that the operators have built that, like,  

00:16:50.400 --> 00:16:54.700
kind of, signal to this controller that, oh, OK, 
this link is down; you need to do something. 

00:16:54.700 --> 00:17:02.472
[MUSIC BREAK] 

00:17:02.472 --> 00:17:05.840
HUIZINGA: Much of your recent work 
represents an effort to reify the  

00:17:05.840 --> 00:17:11.640
idea of automated network management and 
to try to understand the performance of  

00:17:11.640 --> 00:17:16.120
deployed algorithms. So talk about the 
main topics of interest here in this  

00:17:16.120 --> 00:17:21.720
space and how your work has evolved in an era 
of generative AI and large language models. 

00:17:21.720 --> 00:17:26.240
ARZANI: So if you think about it, 
what generative AI is going to enable,  

00:17:26.240 --> 00:17:31.240
and I'm using the term “going to enable” a little 
bit deliberately because I don't think it has yet.  

00:17:31.240 --> 00:17:36.240
We still have to build on top of what we have to 
get that to work. And maybe I'll reconsider my  

00:17:36.240 --> 00:17:41.760
stance on ML now that, you know, we have these 
tools. Haven't yet but might. But essentially,  

00:17:41.760 --> 00:17:48.120
what they enable us to do is take automated action 
on our networks. But if we're allowing AI to do  

00:17:48.120 --> 00:17:55.120
this, we need to be mindful of the risks because 
AI in my, at least in my head of how I view it,  

00:17:55.120 --> 00:18:00.680
is a probabilistic machine, which, what that means 
is that there is some probability, maybe a teeny  

00:18:00.680 --> 00:18:05.120
tiny probability, it might get things wrong. 
And the thing that you don't want is when it  

00:18:05.120 --> 00:18:11.240
gets things wrong, it gets things catastrophically 
wrong. And so you need to put guardrails in place,  

00:18:11.240 --> 00:18:16.480
ensure safety, figure out, like, for each action 
be able to evaluate that action and the risks it  

00:18:16.480 --> 00:18:21.040
imposes long term on your network and whether 
you're able to tolerate that risk. And I think  

00:18:21.040 --> 00:18:25.360
there is a whole room of innovation there 
to basically just figure out the interaction  

00:18:25.360 --> 00:18:30.114
between the AI and the network and where … and 
actually strategic places to put AI, even. 

00:18:30.114 --> 00:18:30.614
HUIZINGA: Right. 

00:18:30.614 --> 00:18:34.520
ARZANI: The thing that for me has evolved is 
I used to think we just want to take the human  

00:18:34.520 --> 00:18:38.840
out of the equation of network management. 
The way I think about it now is there is a  

00:18:38.840 --> 00:18:44.520
place for the human in the network management 
operation because sometimes human has context  

00:18:44.520 --> 00:18:50.200
and that context matters. And so I think what 
the, like, for example, we have this paper in  

00:18:50.200 --> 00:18:57.600
HotNets 2023 where we talk about how to put an LLM 
in the incident management loop, and then there,  

00:18:57.600 --> 00:19:02.640
we carefully talk about, OK, these are the places 
a human needs to be involved, at least given where  

00:19:02.640 --> 00:19:07.060
LLMs are right now, to be able to ensure 
that everything happens in a safe way.  

00:19:07.060 --> 00:19:11.920
HUIZINGA: So go back to this “automated 
network management” thing. This sounds  

00:19:11.920 --> 00:19:17.371
to me like you're in a space where it 
could be, but it isn't ready yet … 

00:19:17.371 --> 00:19:18.146
ARZANI: Yeah.  

00:19:18.146 --> 00:19:21.600
HUIZINGA: … and without, sort of, asking 
you to read a crystal ball about it,  

00:19:21.600 --> 00:19:25.480
do you feel like this is something 
that could be eventually? 

00:19:25.480 --> 00:19:31.300
ARZANI: I hope so. This is the best thing 
about research. You get to be like, yeah! 

00:19:31.300 --> 00:19:32.260
HUIZINGA: Yeah, why not? 

00:19:32.260 --> 00:19:36.440
ARZANI: Why not? And, you know, 
maybe somebody will prove me wrong,  

00:19:36.440 --> 00:19:38.800
but until they do, that's 
what I'm working towards! 

00:19:38.800 --> 00:19:40.700
HUIZINGA: Well, right now it's 
an animating “what if?” 

00:19:40.700 --> 00:19:41.260
ARZANI: Yeah.  

00:19:41.260 --> 00:19:41.733
HUIZINGA: Right? 

00:19:41.733 --> 00:19:42.233
ARZANI: Yeah. 

00:19:42.233 --> 00:19:45.100
HUIZINGA: This is a problem Behnaz is 
interested in right now. Let's go! 

00:19:45.100 --> 00:19:48.070
ARZANI: Yeah. Pretty much. [LAUGHTER] 

00:19:48.070 --> 00:19:51.400
HUIZINGA: OK. Behnaz, the systems and 
networks that we've come to depend on  

00:19:51.400 --> 00:19:54.400
are actually incredibly complex. But 
for most of us, most of the time,  

00:19:54.400 --> 00:19:58.800
they just work. There's only drama when they 
don't work, right? But there's a lot going on  

00:19:58.800 --> 00:20:04.760
behind the scenes. So I want you to talk a 
little bit about how the cycle of configuring,  

00:20:04.760 --> 00:20:09.600
managing, reconfiguring, etc., 
helps keep the drama at bay. 

00:20:09.600 --> 00:20:14.000
ARZANI: Well … you reminded me of something! So 
when I was preparing my job … I'm going to tell  

00:20:14.000 --> 00:20:19.280
this story really, really quickly. But when I was 
preparing my job talk, somebody showed me a tweet.  

00:20:19.280 --> 00:20:27.400
In 2014, I think, people started calling 911 
when Facebook was down! Because of a networking  

00:20:27.400 --> 00:20:33.360
problem! [LAUGHS] Yeah. So that's a thing. 
But, yeah, so network availability matters,  

00:20:33.360 --> 00:20:39.080
and we don't notice it until it's actually down. 
But that aside, back to your question. So I think  

00:20:39.080 --> 00:20:45.200
what operators do is they build systems in 
a way that tries to avoid that drama as much  

00:20:45.200 --> 00:20:50.080
as possible. So, for example, they try to build 
systems that these systems configure the network.  

00:20:50.080 --> 00:20:54.480
And one of my dear friends, Ryan Beckett, works on 
intent-driven networking that essentially tries to  

00:20:54.480 --> 00:20:59.480
ensure that what the operators intend with their 
configurations matches what they actually push  

00:20:59.480 --> 00:21:04.560
into the network. They also monitor the network 
to ensure that as soon as something bad happens,  

00:21:04.560 --> 00:21:09.600
automation gets notified. And there's automation 
also that tries to fix these problems when they  

00:21:09.600 --> 00:21:14.360
happen as much as possible. There's a couple of 
problems that happen in the middle of this. One  

00:21:14.360 --> 00:21:19.560
of them is our networks continuously change, 
and what we use in our networks changes. And  

00:21:19.560 --> 00:21:24.760
there's so many different pieces and components of 
this, and sometimes what happens is, for example,  

00:21:24.760 --> 00:21:29.400
a team decides to switch from one protocol 
to a different protocol, and by doing that,  

00:21:29.400 --> 00:21:35.640
it impacts another team's systems and monitoring 
and what expectations they had for their systems,  

00:21:35.640 --> 00:21:39.520
and then suddenly it causes things to go bad. 
And they have to develop new solutions taking  

00:21:39.520 --> 00:21:44.200
into account the changes that happened. And 
so one of the things that we need to account  

00:21:44.200 --> 00:21:49.760
for in this whole process is how evolution 
is happening. And like evolution-friendly,  

00:21:49.760 --> 00:21:52.600
I guess, systems, maybe, is how 
you should be calling it. 

00:21:52.600 --> 00:21:53.320
HUIZINGA: Right. 

00:21:53.320 --> 00:21:57.240
ARZANI: But that's one. The other 
part of it that goes into play is,  

00:21:57.240 --> 00:22:00.840
most of the time you expect a particular 
traffic characteristic, and then suddenly,  

00:22:00.840 --> 00:22:06.110
you have one fluke event that, kind of, throws 
all of your assumptions out the window, so … 

00:22:06.110 --> 00:22:09.552
HUIZINGA: Right. So it's a never-ending job … 

00:22:09.552 --> 00:22:10.480
ARZANI: Pretty much. 

00:22:10.480 --> 00:22:16.240
HUIZINGA: It's about now that I ask all 
my guests what could possibly go wrong if,  

00:22:16.240 --> 00:22:21.200
in fact, you got everything right. And so 
for you, I'd like to earth this question  

00:22:21.200 --> 00:22:26.120
in the broader context of automation and the 
concerns inherent in designing machines to  

00:22:26.120 --> 00:22:29.960
do our work for us. So at an earlier 
point in your career—we talked about  

00:22:29.960 --> 00:22:34.280
this already—you said you believed you could 
automate everything. Cool. Now you're not so  

00:22:34.280 --> 00:22:39.660
much on that. Talk about what changed your 
thinking and how you're thinking now. 

00:22:39.660 --> 00:22:42.920
ARZANI: OK, so the shallow answer to 
that question—there's a shallow answer,  

00:22:42.920 --> 00:22:47.120
and there's a deeper answer—the shallow answer 
to that question is I watched way too many  

00:22:47.120 --> 00:22:53.960
movies where robots took over the world. 
And honestly speaking, there's a scenario  

00:22:53.960 --> 00:22:58.720
that you can imagine where automation starts 
to get things wrong and then keeps getting  

00:22:58.720 --> 00:23:01.520
things wrong, and wrong, not by the 
definition of automation. Maybe they're  

00:23:01.520 --> 00:23:06.590
doing things perfectly by the objectives and 
metrics that you used to design them … 

00:23:06.590 --> 00:23:07.468
HUIZINGA: Sure.  

00:23:07.468 --> 00:23:11.120
ARZANI: … but they're screwing things up in 
terms of what you actually want them to do. 

00:23:11.120 --> 00:23:11.692
HUIZINGA: Interesting.  

00:23:11.692 --> 00:23:16.320
ARZANI: And if everything is automated and 
you don't leave yourself an intervention plan,  

00:23:16.320 --> 00:23:18.720
how are you going to take control back? 

00:23:18.720 --> 00:23:22.480
HUIZINGA: Right. So this goes back to the 
humans-in-the-loop/humans-out-of-the-loop.  

00:23:22.480 --> 00:23:26.522
And if I remember in our last podcast, we 
were talking about humans out of the loop. 

00:23:26.522 --> 00:23:27.100
ARZANI: Yeah. 

00:23:27.100 --> 00:23:31.000
HUIZINGA: And you've already talked a 
bit about what the optimal place for a  

00:23:31.000 --> 00:23:36.800
human to be is. Is the human always going to 
have to be in the loop, in your opinion? 

00:23:36.800 --> 00:23:42.240
ARZANI: I think it's a scenario where you 
always give yourself a way to interrupt. Like,  

00:23:42.240 --> 00:23:49.240
always put a back door somewhere. When we notice 
things go bad, we have a way that's foolproof that  

00:23:49.240 --> 00:23:53.960
allows us to shut everything down and take control 
back to ourselves. Maybe that's where we go. 

00:23:53.960 --> 00:23:56.960
HUIZINGA: How do you approach 
the idea of corner cases? 

00:23:56.960 --> 00:24:01.280
ARZANI: That's essentially what my research 
right now is, actually! And I love it,  

00:24:01.280 --> 00:24:06.260
which is essentially figuring out, in a 
foolproof way, all the corner cases.  

00:24:06.260 --> 00:24:06.780
HUIZINGA: Yeah?  

00:24:06.780 --> 00:24:11.200
ARZANI: Can you build a tool that will tell 
you what the corner cases are? Now, granted,  

00:24:11.200 --> 00:24:14.640
what we focus on is performance 
corner cases. Nikolaj Bjørner,  

00:24:15.320 --> 00:24:19.120
in RiSE—so RiSE is Research in 
Software Engineering—is working on,  

00:24:19.120 --> 00:24:24.200
how do you do verification corner cases? But all 
of them, kind of, have a hand-in-hand type of,  

00:24:24.200 --> 00:24:27.660
you know, Holy Grail goal, which is, 
how do you find all the corner cases? 

00:24:27.660 --> 00:24:30.800
HUIZINGA: Right. And that, kind of, 
is the essence of this “What could  

00:24:30.800 --> 00:24:34.542
possibly go wrong?” question, 
is looking in every corner … 

00:24:34.542 --> 00:24:35.186
ARZANI: Correct. 

00:24:35.186 --> 00:24:41.040
HUIZINGA: … for anything that could go wrong. 
So many people in the research community have  

00:24:41.040 --> 00:24:46.840
observed that the speed of innovation in 
generative AI has shrunk the traditional  

00:24:46.840 --> 00:24:50.880
research-to-product timeline, and some people 
have even said everyone's an applied researcher  

00:24:50.880 --> 00:24:56.520
now. Or everyone's a PM. [LAUGHS] Depends 
on who you are! But you have an interesting  

00:24:56.520 --> 00:25:01.120
take on this Behnaz, and it reminds me 
of a line from the movie Nanny McPhee:  

00:25:01.120 --> 00:25:06.520
“When you need me but do not want me, then I will 
stay. When you want me but no longer need me,  

00:25:06.520 --> 00:25:10.240
I have to go.” So let's talk a little 
bit about your perspective on this  

00:25:10.240 --> 00:25:17.320
idea-to-ideation pipeline. How and where are 
researchers in your orbit operating these days,  

00:25:17.320 --> 00:25:20.940
and how does that impact what we might 
call “planned obsolescence” in research? 

00:25:20.940 --> 00:25:28.800
ARZANI: I guess the thing I'm seeing is that we 
are freed up to dream more—in a way. Maybe that's  

00:25:28.800 --> 00:25:34.240
me being too … I'm a little bit of a romantic, so 
this is that coming out a little bit, but it's,  

00:25:34.240 --> 00:25:40.680
like, because of all this, we have the time 
to think bigger, to dream bigger, to look at  

00:25:40.680 --> 00:25:46.640
problems where maybe five years ago, we wouldn't 
even dare to think about. We have amazingly,  

00:25:46.640 --> 00:25:52.280
amazingly smart, competent people in our product 
teams. Some of them are actually researchers. So  

00:25:52.280 --> 00:25:58.360
there's, for example, the Azure systems research 
group that has a lot of people that are focused  

00:25:58.360 --> 00:26:03.200
on problems in our production systems. And 
then you have equivalents of those spread  

00:26:03.200 --> 00:26:09.120
out in the networking sphere, as well. And so 
a lot of complex problems that maybe like 10  

00:26:09.120 --> 00:26:12.880
years ago Microsoft Research would look 
at nowadays they can handle themselves.  

00:26:12.880 --> 00:26:17.480
They don't need us. And that's part of what 
has allowed us to now go and be like, OK,  

00:26:17.480 --> 00:26:21.560
I'm going to think about other things. Maybe 
things that, you know, aren't relevant to you  

00:26:21.560 --> 00:26:25.220
today, but maybe in five years, you'll come 
in and thank me for thinking about this! 

00:26:25.220 --> 00:26:31.240
HUIZINGA: OK. Shifting gears here! In a recent 
conversation, I heard a coelleague refer to you  

00:26:31.240 --> 00:26:35.840
as an “idea machine.” To me, that's one of the 
greatest compliments you could get. But it got  

00:26:35.840 --> 00:26:41.760
me wondering, so I'll ask you: how does your 
brain work, Behnaz, and how do you get ideas? 

00:26:41.760 --> 00:26:47.880
ARZANI: Well, this has been, to my chagrin, one of 
the realities of life about my brain apparently.  

00:26:47.880 --> 00:26:51.720
So I never thought of this as a strength. I always 
thought about it as a weakness. But nowadays, I'm  

00:26:51.720 --> 00:26:58.360
like, oh, OK, I'm just going to embrace this now! 
So I have a random brain. It’s completely ran—so,  

00:26:58.360 --> 00:27:02.360
like, it actually happens, like, you're talking, 
and then suddenly, I say something that seems to  

00:27:02.360 --> 00:27:07.120
other people like it came out of left field. I 
know how I got there. It’s essentially kind of  

00:27:07.120 --> 00:27:10.360
like a Markov chain. [LAUGHTER] So a Markov 
chain is essentially a number of states,  

00:27:10.360 --> 00:27:13.960
and there's a certain probability you can 
go from one state to the other state. And,  

00:27:13.960 --> 00:27:17.520
actually, one of the things I found out 
about myself is I think through talking  

00:27:17.520 --> 00:27:22.680
for this exact reason. Because people see 
this random Markov chain by what they say,  

00:27:22.680 --> 00:27:26.760
and it suddenly goes into different places, 
and that's how ideas come about. Most of  

00:27:26.760 --> 00:27:29.360
my ideas have actually come through 
when I've been talking to someone. 

00:27:29.360 --> 00:27:30.160
HUIZINGA: Really?  

00:27:30.160 --> 00:27:30.670
ARZANI: Yeah.  

00:27:30.670 --> 00:27:32.107
HUIZINGA: Them talking or you talking? 

00:27:32.107 --> 00:27:32.673
ARZANI: Both. 

00:27:32.673 --> 00:27:33.306
HUIZINGA: Really? 

00:27:33.306 --> 00:27:36.560
ARZANI: So it's, like, basically, I think 
the thing that has recently … like, I've just  

00:27:36.560 --> 00:27:41.480
noticed more—again, being more mindful does that 
to you—it's like I'm talking to someone. I'm like,  

00:27:41.480 --> 00:27:44.960
I have an idea. And it's usually they said 
something, or I was saying something that  

00:27:44.960 --> 00:27:49.400
triggered that thought coming up. Which doesn't 
happen when … I'm not one of those people that you  

00:27:49.400 --> 00:27:53.320
can put in a room for three days—somebody actually 
once told me this— [LAUGHTER] like, I'm not one  

00:27:53.320 --> 00:27:57.560
of those people you can put in a room for three 
days and I come out with these brilliant ideas.  

00:27:57.560 --> 00:28:01.499
It's like you put me in a room with five other 
people, then I come out with interesting ideas. 

00:28:01.499 --> 00:28:02.920
HUIZINGA: Right. … It's the interaction.  

00:28:02.920 --> 00:28:03.700
ARZANI: Yeah.  

00:28:03.700 --> 00:28:07.760
HUIZINGA: I want to link this idea of the 
ideas that you get to the conversations  

00:28:07.760 --> 00:28:13.240
you have and maybe go back to linking it 
to the work you've recently done. Talk  

00:28:13.240 --> 00:28:19.225
about some of the projects, how they came 
from idea to paper to product even … 

00:28:19.225 --> 00:28:23.160
ARZANI: Mm-hm. So like one of the works 
that we were doing was this work on, like,  

00:28:23.160 --> 00:28:29.360
max-min fair resource allocation that recently got 
published in NSDI and is actually in production.  

00:28:29.360 --> 00:28:35.600
So the way that came out is I was working with 
a bunch of other researchers on risk estimation,  

00:28:35.600 --> 00:28:39.720
actually, for incident management of all things, 
which was, how do you figure out if you want to  

00:28:39.720 --> 00:28:44.320
mitigate a particular problem in a certain way, 
how much risk it induces as a problem. And so one  

00:28:44.320 --> 00:28:49.000
of the people who was originally … one of the 
original researchers who built our wide-area  

00:28:49.000 --> 00:28:53.480
traffic engineering controller, which we were 
talking about earlier, he said, “You're solving  

00:28:53.480 --> 00:28:58.880
the max-min fair problem.” We're like, really? 
And then this caused a whole, like, one-year  

00:28:58.880 --> 00:29:04.320
collaboration where we all sat and evolved this 
initial algorithm we had into a … So initially  

00:29:04.320 --> 00:29:09.200
it was not a multipath problem. It had a lot of 
things that didn't fully solve the problem of  

00:29:09.200 --> 00:29:13.680
max-min fair resource allocation, but it evolved 
into that. Then we deployed it, and it improved  

00:29:13.680 --> 00:29:19.040
the SWAN solver by a factor of three in terms 
of how fast it solved the problem and didn't  

00:29:19.040 --> 00:29:24.080
have any performance impact, or at least very 
little. And so, yeah, that's how it got born. 

00:29:24.080 --> 00:29:28.440
HUIZINGA: OK. So for those of us 
who don't know, what is max-min  

00:29:28.440 --> 00:29:31.620
fair resource allocation, and 
why is it such a problem? 

00:29:31.620 --> 00:29:36.880
ARZANI: Well, so remember I said that in our wide 
area network, we route traffic from one place to  

00:29:36.880 --> 00:29:41.200
the other in a way that meets capacity. So one 
of the objectives we try to meet is we try to be  

00:29:41.200 --> 00:29:46.840
fair in a very specific metric. So max-min is just 
the metric of fairness we use. And that basically  

00:29:46.840 --> 00:29:52.360
means you cannot improve what you allocated 
to one piece of traffic in a way that would  

00:29:52.360 --> 00:29:58.160
hurt anybody who has gotten less. So there's 
a little bit of a, like, … it's a mind bend  

00:29:58.160 --> 00:30:02.920
to wrap your head a little bit around the max-min 
fair definition. But the reason making it faster  

00:30:02.920 --> 00:30:08.120
is important is if something fails, we need to 
quickly recompute what the paths are and how we  

00:30:08.120 --> 00:30:12.160
route traffic. So the faster we can solve this 
problem, the better we can adapt to failures. 

00:30:12.160 --> 00:30:16.400
HUIZINGA: So talk a little bit about 
some of the work that started as an  

00:30:16.400 --> 00:30:19.980
idea and you didn't even maybe know that 
it was going to end up in production. 

00:30:19.980 --> 00:30:24.560
ARZANI: There was this person from Azure 
Networking came and gave a talk in our group. And  

00:30:24.560 --> 00:30:27.960
he's a person I've known for years, so I was like, 
hey, do you want to jump on a meeting and talk?  

00:30:28.600 --> 00:30:32.160
So he came into that meeting, and I was like, 
OK, what are some of the things you're curious  

00:30:32.160 --> 00:30:35.400
about these days? You want to answer these 
days? And it was like, yeah, we have this  

00:30:35.400 --> 00:30:41.840
heuristic we're using in our traffic engineering 
solution, and essentially what it does is to make  

00:30:41.840 --> 00:30:47.680
the optimization problem we solve smaller. If a 
piece of traffic is smaller than a particular,  

00:30:47.680 --> 00:30:51.960
like, arbitrary threshold, we just send it on 
a shortest path and don't worry about it. And  

00:30:51.960 --> 00:30:57.120
then we optimize everything else. And I just want 
to know, like, what is the optimality gap of this  

00:30:57.120 --> 00:31:03.040
heuristic? How bad can this heuristic be? And 
then I had worked on Stackelberg games before,  

00:31:03.040 --> 00:31:07.680
in my PhD. It never went anywhere, but it was an 
idea I played around with, and it just immediately  

00:31:07.680 --> 00:31:13.200
clicked in my head that this is the same problem. 
So Stackelberg games are a leader-follower game  

00:31:13.200 --> 00:31:18.680
where in this scenario a leader has an objective 
function that they're trying to maximize, and they  

00:31:18.680 --> 00:31:25.560
control one or multiple of the inputs that their 
followers get to operate over. The followers,  

00:31:25.560 --> 00:31:30.320
on the other hand, don't get to control anything 
about this input. They have their own objective  

00:31:30.320 --> 00:31:34.480
that they're trying to maximize or minimize, 
but they have other variables in their control,  

00:31:34.480 --> 00:31:38.800
as well. And what their objective is, is 
going to control the leader's payoff. And  

00:31:38.800 --> 00:31:42.640
so this game is happening where the leader 
has more control in this game because it's,  

00:31:42.640 --> 00:31:47.200
kind of, like the followers are operating 
in subject to whatever the leader says,  

00:31:47.200 --> 00:31:51.880
right. But the leader is impacted by what the 
followers do. And so this dynamic is what they  

00:31:51.880 --> 00:31:57.200
call a Stackelberg game. And the way we map the 
MetaOpt problem to this is the leader in our  

00:31:57.200 --> 00:32:01.880
problem wants to maximize the difference between 
the optimal and the heuristic. It controls the  

00:32:01.880 --> 00:32:06.920
inputs to both the optimal and the heuristic. 
And now this optimal and heuristic algorithms  

00:32:06.920 --> 00:32:10.680
are the followers in that game. They don't 
get to control the inputs, but they have other  

00:32:10.680 --> 00:32:14.500
variables they control, and they have objectives 
that they want to maximize or minimize.  

00:32:14.500 --> 00:32:15.000
HUIZINGA: Right.  

00:32:15.000 --> 00:32:20.120
ARZANI: And so that's how the Stackelberg-game 
dynamic comes about. And then we got other  

00:32:20.120 --> 00:32:23.880
researchers in the team involved, and then 
we started talking, and then it just evolved  

00:32:23.880 --> 00:32:27.840
into this beast right now that is a tool, MetaOpt, 
that we released, I think, a couple of months ago.  

00:32:29.480 --> 00:32:33.880
And another piece that was really cool was 
people from ETH Zürich came to us and were like,  

00:32:33.880 --> 00:32:38.360
oh, you guys analyzed our heuristic! We have a 
better one! Can you analyze this one? And that  

00:32:38.360 --> 00:32:42.870
was a whole fun thing we did where we analyzed 
their heuristics for them. And, then, yeah ... 

00:32:42.870 --> 00:32:47.800
HUIZINGA: Yeah. So all these things that you're 
mentioning, are they findable as papers? Were  

00:32:47.800 --> 00:32:52.840
they presented at conferences, and where 
are they in anybody's usability scenario? 

00:32:52.840 --> 00:32:57.840
ARZANI: So the MetaOpt tool that 
I just mentioned, that one is in …  

00:32:57.840 --> 00:33:03.000
it’s an open-source tool. You can go online and 
search for MetaOpt. You'll find the tool. We're  

00:33:03.000 --> 00:33:06.400
here to support anything you need; if you 
run into issues, we’ll help you fix it. 

00:33:06.400 --> 00:33:10.920
HUIZINGA: Great. You can probably find 
all of these papers under publications  

00:33:10.920 --> 00:33:15.330
on your bio page on the website, 
Microsoft Research website. 

00:33:15.350 --> 00:33:22.080
HUIZINGA: Cool. If anyone wants to do that. So, 
Behnaz, the idea of having ideas is cool to me,  

00:33:22.080 --> 00:33:25.880
but of course, part of the research 
problem is identifying which ones you  

00:33:25.880 --> 00:33:29.680
should go after [LAUGHS] and which 
ones you shouldn't. So, ironically,  

00:33:29.680 --> 00:33:33.640
you've said you're not that good at that part 
of it, but you're working at getting better.  

00:33:33.640 --> 00:33:34.320
ARZANI: Yes. 

00:33:34.320 --> 00:33:37.000
HUIZINGA: So first of all, why 
do you say that you're not very  

00:33:37.000 --> 00:33:39.600
good at it? And second of all, 
what are you doing about it? 

00:33:39.600 --> 00:33:44.760
ARZANI: So I, as I said, get attracted to 
puzzles, to hard problems. So most of the  

00:33:44.760 --> 00:33:49.900
problems that I go after are problems I have no 
idea how to solve. And that tends to be a risk. 

00:33:49.900 --> 00:33:50.700
HUIZINGA: Yeah. 

00:33:50.700 --> 00:33:55.800
ARZANI: Where I think people who are better at 
selecting problems are those who actually have  

00:33:55.800 --> 00:34:01.960
an idea of whether they'll be able to solve this 
problem or not. And I never actually asked myself  

00:34:01.960 --> 00:34:06.680
that question before this year. [LAUGHTER] 
So now I'm trying to get a better sense of,  

00:34:06.680 --> 00:34:12.440
how do I figure out if a problem is solvable 
or not before I try to solve it? And also,  

00:34:12.440 --> 00:34:16.800
just what makes a good research problem? 
So what I'm doing is, I'm going back to the  

00:34:16.800 --> 00:34:21.800
era that I thought had the best networking 
papers, and I'm just trying to dissect what  

00:34:21.800 --> 00:34:26.200
makes those papers good, just to understand 
better for myself, to be like, OK, what do  

00:34:26.200 --> 00:34:31.260
I want to replicate? Replicate, not in terms 
of techniques, but in terms of philosophy. 

00:34:31.260 --> 00:34:36.000
HUIZINGA: So what you're looking at is 
how people solve problems through the work  

00:34:36.000 --> 00:34:42.272
that they did in this arena. So what are you 
finding? Have you gotten any nuggets of … 

00:34:42.272 --> 00:34:47.600
ARZANI: So a couple. So one of my favorite 
papers is Van Jacobson's TCP paper. The  

00:34:47.600 --> 00:34:54.000
intuition is amazing to me. It’s almost like he 
has a vision of what's happening, is the best I  

00:34:54.000 --> 00:35:03.480
can describe it. And another example of this is 
also early-on papers by people like Ratul Mahajan,  

00:35:03.480 --> 00:35:10.680
Srikanth Kandula, those guys, where you see that 
they start with a smaller example that, kind of,  

00:35:10.680 --> 00:35:16.080
shows how this problem is going to happen and 
how they're going to solve it. I mean, I did this  

00:35:16.080 --> 00:35:20.080
in my work all the time, too, but it was never 
conscious. It's more of like that goes to that  

00:35:20.080 --> 00:35:24.360
mindfulness thing that I said before, too. It's 
like you might be doing some of these already,  

00:35:24.360 --> 00:35:28.240
but you don’t notice what you're doing. It 
more of is, kind of, like putting of like,  

00:35:28.240 --> 00:35:31.600
oh, this is what they did. And I do this, 
too. And this might be a good habit to keep  

00:35:31.600 --> 00:35:35.863
but cultivate into a habit as opposed to an 
unconscious thing that you're just doing. 

00:35:35.863 --> 00:35:39.880
HUIZINGA: Right. You know, this whole idea 
of going back to what's been done before,  

00:35:39.880 --> 00:35:44.840
I think that's a lesson about looking at 
history, as well, and to say, you know,  

00:35:44.840 --> 00:35:49.240
what can we learn from that? What are we 
trying to reinvent that maybe doesn't need  

00:35:49.240 --> 00:35:57.200
to be reinvented? Has it helped you to get more 
targeted on the kinds of problems that you say,  

00:35:57.200 --> 00:35:59.460
“I'm not going to work on that. 
I am going to work on that”? 

00:35:59.460 --> 00:36:03.926
ARZANI: To be very, very, very fair, I haven't 
done this for a long time yet! This has been … 

00:36:03.926 --> 00:36:05.305
HUIZINGA: A new thing.
ARZANI: I started this this month, yeah.

00:36:05.305 --> 00:36:05.832
HUIZINGA: Oh my goodness!  

00:36:05.832 --> 00:36:09.604
ARZANI: So we’ll see how far I get and 
how useful it ends up being! [LAUGHS] 

00:36:09.604 --> 00:36:18.190
[MUSIC BREAK] 

00:36:18.190 --> 00:36:23.200
HUIZINGA: One of my favorite things to talk about 
on this show is what my colleague Kristina calls  

00:36:23.200 --> 00:36:27.840
“outrageous” lines of research. And so I've 
been asking all my guests about their most  

00:36:27.840 --> 00:36:32.680
outrageous ideas and how they turned out. 
So sometimes these ideas never got off the  

00:36:32.680 --> 00:36:38.800
ground. Sometimes they turned out great. And 
other times, they've failed spectacularly. Do  

00:36:38.800 --> 00:36:42.640
you have a story for the “Microsoft 
Research Outrageous Ideas” file? 

00:36:42.640 --> 00:36:49.000
ARZANI: I had this question of, if language has 
grammar, and grammar is what LLMs are learning,  

00:36:49.000 --> 00:36:54.560
which, to my understanding of what people who are 
experts in this field say, this maybe isn't that,  

00:36:54.560 --> 00:36:59.600
but if it is the case that grammar is what 
allows these LLMs to learn how language works,  

00:36:59.600 --> 00:37:03.400
then in networking, we have the equivalent 
of that, and the equivalent of that is  

00:37:03.400 --> 00:37:07.840
essentially network protocols. And 
everything that happens in a network,  

00:37:07.840 --> 00:37:13.200
you can define it as an event that happens in a 
network. You can think of those, like, the events  

00:37:13.200 --> 00:37:19.280
are words in a language. And so, is it going to 
be the case, and this is a question which is,  

00:37:19.280 --> 00:37:23.960
if you take an event abstraction and encode 
everything that happens in a network in that  

00:37:23.960 --> 00:37:28.520
event abstraction, can you build an equivalent 
of an LLM for networks? Now what you would use  

00:37:28.520 --> 00:37:33.920
it for—this is another reason I've never worked 
on this problem—I have no idea! [LAUGHTER] But  

00:37:33.920 --> 00:37:38.880
what this would allow you to do is build 
the equivalent of an LLM for networking,  

00:37:38.880 --> 00:37:44.320
where actually you just translate that network's 
events into, like, this event abstraction,  

00:37:44.320 --> 00:37:48.720
and then the two understand each other. So 
like a universal language of networking,  

00:37:48.720 --> 00:37:56.000
maybe. It could be cool. Never tried it. 
Probably a dumb idea! But it's an idea. 

00:37:56.000 --> 00:37:58.960
HUIZINGA: What would it take to try it? 

00:37:58.960 --> 00:38:04.360
ARZANI: Um … I feel like bravery is, I 
think, one because with any risky idea,  

00:38:04.360 --> 00:38:06.260
there's a probability that you will fail. 

00:38:06.260 --> 00:38:12.520
HUIZINGA: As a researcher here at Microsoft 
Research, when you have this idea,  

00:38:12.520 --> 00:38:16.320
um … and you say, well, I'm not brave 
enough … even if you were brave enough,  

00:38:16.320 --> 00:38:19.312
who would you have to convince 
that they should let you do it? 

00:38:19.312 --> 00:38:20.412
ARZANI: I don't think anybody! 

00:38:20.412 --> 00:38:21.072
HUIZINGA: Really? 

00:38:21.072 --> 00:38:22.880
ARZANI: That's the whole … that's 
the whole point of me being here!  

00:38:22.880 --> 00:38:25.830
I don't like being told what to do! [LAUGHS] 

00:38:25.830 --> 00:38:27.440
HUIZINGA: Back to the beginning! 

00:38:27.440 --> 00:38:30.800
ARZANI: Yeah. The only thing is that, maybe, like,  

00:38:30.800 --> 00:38:33.600
people would be like, what have you 
been doing in the past six months? And  

00:38:33.600 --> 00:38:36.693
I wouldn't have ... that's the risk. 
That's where bravery comes in.  

00:38:36.712 --> 00:38:40.440
ARZANI: The bravery is more of there is 
a possibility that I have to devote three  

00:38:40.440 --> 00:38:44.320
years of my life into this, to figuring out 
how to make that work, and I might not be  

00:38:44.320 --> 00:38:48.660
able to. And there's other things. So it's a 
tradeoff also of where you put your time. 

00:38:48.660 --> 00:38:49.260
HUIZINGA: Sure. 

00:38:49.260 --> 00:38:50.320
ARZANI: So there. Yeah. 

00:38:50.320 --> 00:38:54.760
HUIZINGA: And if, but … part of it would be 
explaining it in a way to convince people:  

00:38:54.760 --> 00:38:56.640
if it worked, it would be amazing! 

00:38:56.640 --> 00:38:59.120
ARZANI: And that's the other problem with 
this idea. I don't know what you would  

00:38:59.120 --> 00:39:02.400
use it for. If I knew what you would use it 
for, maybe then it would make it worth it. 

00:39:02.400 --> 00:39:04.703
HUIZINGA: All right. Sounds like you 
need to spend some more time … 

00:39:04.703 --> 00:39:05.440
ARZANI: Yeah.  

00:39:05.440 --> 00:39:12.640
HUIZINGA: …ruminating on it. Um, yeah. The whole 
cliché of the solution in search of a problem.  

00:39:12.640 --> 00:39:15.501
ARZANI: Yeah. 

00:39:15.501 --> 00:39:19.800
HUIZINGA: [LAUGHS] As we close, I want to talk 
a little bit about some fun things. And so,  

00:39:19.800 --> 00:39:24.240
aside from your research life, I was 
intrigued by the fact, on your bio page,  

00:39:24.240 --> 00:39:30.240
that you have a rich artistic life, as well, and 
that includes painting, music, writing, along with  

00:39:30.240 --> 00:39:35.440
some big ideas about the value of storytelling. So 
I'll take a second to plug the bio page. People,  

00:39:35.440 --> 00:39:40.600
go look at it because she's got paintings and 
cool things that you can link to. As we close,  

00:39:40.600 --> 00:39:44.560
I wonder if you could use this time to 
share your thoughts on this particular  

00:39:44.560 --> 00:39:49.160
creative pursuit of storytelling and 
how it can enhance our relationships  

00:39:49.160 --> 00:39:52.640
with our colleagues and ultimately make us 
better researchers and better people? 

00:39:52.640 --> 00:39:57.040
ARZANI: I think it's not an understatement to 
say I had a life-changing experience through  

00:39:57.040 --> 00:40:02.880
storytelling. The first time I encountered it, 
it was the most horrific thing I had ever seen!  

00:40:02.880 --> 00:40:08.320
I had gone on Meetup—this was during COVID—to 
just, like, find places to meet people, build  

00:40:08.320 --> 00:40:12.400
connections and all that, and I saw this event 
called “Storytelling Workshop,” and I was like,  

00:40:12.400 --> 00:40:17.520
good! I'm good at making up stories, and, you 
know, that's what I thought it was. Turns out  

00:40:17.520 --> 00:40:23.000
it's, you go and tell personal stories about 
your life that only involve you, that make you  

00:40:23.000 --> 00:40:27.320
deeply vulnerable. And, by the way, I'm Iranian. 
We don't do vulnerability. It’s just not a thing.  

00:40:27.880 --> 00:40:31.680
So it was the most scary thing I've ever done 
in my life. But you go on stage and basically  

00:40:31.680 --> 00:40:36.640
talk about your life. And the thing it taught 
me by both telling my own stories and listening  

00:40:36.640 --> 00:40:42.400
to other people's stories is that it showed me 
that you can connect to people through stories,  

00:40:42.400 --> 00:40:49.160
first of all. The best ideas come when you're 
actually in it together. Like one of the things  

00:40:49.160 --> 00:40:54.840
that now I say that I didn't used to say, we, 
we're all human. And being human essentially  

00:40:54.840 --> 00:40:58.920
means we have good things about ourselves and 
bad things about ourselves. And as researchers,  

00:40:58.920 --> 00:41:03.440
we have our strengths as researchers, and we 
have our weaknesses as researchers. And so  

00:41:03.440 --> 00:41:09.400
when we collaborate with other people, we bring 
all of that. And collaboration is a sacred thing  

00:41:09.400 --> 00:41:15.160
that we do where we're basically trusting each 
other with bringing all of that to the table  

00:41:15.160 --> 00:41:20.400
and being that vulnerable. And so our job as 
collaborators is essentially to protect that,  

00:41:20.400 --> 00:41:26.640
in a way, and make it safe for everybody 
to come as they are. And so I think that's  

00:41:26.640 --> 00:41:31.560
what it taught me, which is, like, 
basically holding space for that. 

00:41:31.560 --> 00:41:33.900
HUIZINGA: Yeah. How's that working? 

00:41:33.900 --> 00:41:36.960
ARZANI: First of all, I stumbled into it,  

00:41:36.960 --> 00:41:39.972
but there are people who are 
already “that” in this building … 

00:41:39.972 --> 00:41:40.548
HUIZINGA: Really? 

00:41:40.548 --> 00:41:44.920
ARZANI: … that have been for years. It's just 
that now I can see them for what they bring,  

00:41:44.920 --> 00:41:46.871
as opposed to before, I didn't 
have the vocabulary for it.  

00:41:46.871 --> 00:41:48.072
HUIZINGA: Gotcha …  

00:41:48.072 --> 00:41:52.560
ARZANI: But people who don't, it’s like what 
I've seen is almost like they initially look  

00:41:52.560 --> 00:41:57.560
at you with skepticism, and then they think it's 
a gimmick, and then they are like, what is that?  

00:41:57.560 --> 00:42:03.240
And then they become curious, and then 
they, too, kind of join you, which is very,  

00:42:03.240 --> 00:42:07.560
very interesting to see. But, like, again, 
it’s something that already existed. It's  

00:42:07.560 --> 00:42:13.360
just me not being privileged enough to know 
about it or, kind of, recognize it before. 

00:42:13.360 --> 00:42:15.640
HUIZINGA: Yeah. Can that become part of a culture,  

00:42:15.640 --> 00:42:19.512
or do you feel like it is part of the 
culture here at Microsoft Research, or … ? 

00:42:19.512 --> 00:42:23.520
ARZANI: I think this depends on how 
people individually choose to show  

00:42:23.520 --> 00:42:29.960
up. And I think we're all, at the end 
of the day, individuals. And a lot of  

00:42:29.960 --> 00:42:34.840
people are that way without knowing they are 
that way. So maybe it is already part of the  

00:42:34.840 --> 00:42:38.660
culture. I haven't necessarily sat down and 
thought about it deeply, so I can't say. 

00:42:38.660 --> 00:42:44.720
HUIZINGA: Yeah, yeah. But it would 
be a dream to have the ability to be  

00:42:44.720 --> 00:42:48.680
that vulnerable through storytelling 
as part of the research process? 

00:42:48.680 --> 00:42:52.440
ARZANI: I think so. We had a storytelling coach 
that would say, “Tell your story, change the  

00:42:52.440 --> 00:42:59.440
world.” And as researchers, we are attempting 
to change the world, and part of that is our  

00:42:59.440 --> 00:43:05.222
stories. And so maybe, yeah! And basically, what 
we're doing here is, I'm telling my story. So … 

00:43:05.222 --> 00:43:06.228
HUIZINGA: Yeah. 

00:43:06.228 --> 00:43:07.640
ARZANI: … maybe you're changing the world! 

00:43:07.640 --> 00:43:15.040
HUIZINGA: You know, I'm all in! I'm here for 
it, as they say. Behnaz Arzani. It is such a  

00:43:15.040 --> 00:43:20.620
pleasure—always a pleasure—to talk to you. Thanks 
for sharing your story with us today on Ideas.  

00:43:20.620 --> 00:43:22.085
ARZANI: Thank you. 

00:43:22.085 --> 00:43:37.644
[MUSIC]