Latest / Agentic Conversations (formally mlops.community) / Reliable ML // Niall Murphy & Todd Underwood // Coffee Sessions #127

Transcript

Reliable ML // Niall Murphy & Todd Underwood // Coffee Sessions #127 · Agentic Conversations (formally mlops.community)

• 0:00
  Hi. My name is Niall Murphy.
• 0:01
  I'm running a small start-up called stanza systems and I am a
• 0:05
  coffee refuse neck and if you want to get me via caffeine you
• 0:10
  have to do tea and milk and sugar.
• 0:14
  Hey, I'm tired, I ran. MLS sorry for Google.
• 0:19
  I drink coffee by the leader and these days.
• 0:21
  I drink it in this cup that my daughter destroyed in the back.
• 0:25
  At least she slammed the door on it.
• 0:27
  Like these are supposed to be indestructible, but Anyway, I
• 0:30
  generally have between like one and a half and two liters of
• 0:33
  coffee a day. So they drink it black because
• 0:36
  it's coffee. I drink coffee.
• 0:38
  Not non-coffee people have coffee and non-coffee.
• 0:41
  That seems like an error gv2 punted, long time.
• 0:46
  No see man. Hello?
• 0:48
  Hello. Is everything I love that the
• 0:52
  beard. You're looking real good.
• 0:53
  Real scruffy. Mmm.
• 0:55
  Yeah, I'm actually in Greece right now.
• 0:57
  Trying to walk up the last Rays of Summer.
• 1:02
  How is it super hot over there? I'm curious.
• 1:04
  Yep, it is an. I am going to the beach every
• 1:07
  day still. It's really nice.
• 1:10
  I'm not really looking forward to getting back to Germany
• 1:13
  because I feel like it's going to be a very extreme change.
• 1:16
  From one. Yeah.
• 1:17
  Flying to the next but you know how it goes anyway.
• 1:20
  Dude, where ya Goin? We had Nile Murphy and Todd
• 1:27
  Underwood on here. Both of these.
• 1:29
  All's we interviewed before prior to this step.
• 1:33
  So if you have not seen their first interviews, I highly
• 1:35
  recommend checking them out. We're going to leave the link to
• 1:38
  those in the description, but I mean these guys are highly
• 1:42
  decorated, like, Todd is the head of ml infrastructure SRE at
• 1:49
  Google? That is huge.
• 1:52
  He's got else might've heard of that company.
• 1:55
  Nile also is, the guy that wrote the book on being In a sorry.
• 2:00
  So like and he was at gue better heard of that, he then went to
• 2:04
  Microsoft. He then became a founder and he
• 2:09
  is starting this new startup that he's got right now.
• 2:12
  So it's really cool to see this but the reason we got him back
• 2:15
  on here is because of their new book that they just came out
• 2:17
  with which is reliable machine learning.
• 2:19
  And we mentioned it a few times. We talked about this in the
• 2:22
  episode, we go through a few chapters we asked them to
• 2:26
  comment on it what were some of your takeaways?
• 2:29
  Yeah, I think, first off, it was great that they wrote this book
• 2:33
  in the first place. I think, you know, in a space
• 2:35
  that's really new and where lots of people entering into entering
• 2:39
  into it, have lots of questions. I feel like this book clarified,
• 2:43
  a few things and answered some of those questions.
• 2:46
  Something in particular I thought was really interesting
• 2:48
  was, you know, just their focus on the cultural side of things
• 2:51
  and we have mentioned this number a number of times in
• 2:54
  previous episodes, you know, but it but I think they really like
• 2:58
  they You know, they actually put that into practice by writing,
• 3:02
  you know, chapters dedicated to some of these things and they
• 3:06
  were thoughtfully, you know, put together, you know, introducing
• 3:09
  things that I in my opinion weren't, you know, exactly
• 3:12
  obvious. But we maybe have thought about.
• 3:14
  But again, just the valley, there's a lot of value in
• 3:16
  putting it all in one place so that people can benefit from it.
• 3:19
  And so I just really one takeaway that I had was it was
• 3:23
  great to have all of their experiences, kind of documented
• 3:27
  or encoded into this one book. Look and I'm sure there's so
• 3:30
  many more things that are not there, but it has a good gist of
• 3:34
  like what they think the spaces is about and how to best manage
• 3:39
  it. Yeah.
• 3:41
  Can I tell you a little secret from?
• 3:42
  And this is just what I heard from Todd, but at the end of the
• 3:48
  episode that we had discussed Leon, do you remember when we
• 3:50
  interviewed him back? I think it was like a little
• 3:53
  over a year ago now. Yeah.
• 3:55
  Well, time flies, for those of you that do not know who D
• 3:58
  Scully is, You probably read some of his work.
• 4:01
  He is one of the authors of these Monumental papers around
• 4:06
  ml Ops like the ml test score and the high interest credit
• 4:10
  card debt of machine learning. And so at the end of that, we
• 4:14
  asked him. Hey so you're going to write a
• 4:16
  book anytime soon and he these goalie was like oh hell no, I'm
• 4:21
  not doing it. We should grab that clip and
• 4:25
  yeah in the description but basically D was like nope.
• 4:28
  Nope not doing that. No reason to do that at all.
• 4:31
  And he shot me down hard when I ask that question to so much so
• 4:34
  that I was like who like I shouldn't have asked him.
• 4:37
  That I felt kind of like an idiot.
• 4:39
  Anyway Todd watched that. And then reached out to D
• 4:43
  because they both work at Google and then taught asked him if he
• 4:48
  wants to help them write this book and he said, yes, so nice
• 4:52
  Kelly is one of the authors of the reliable machine learning
• 4:57
  glory and I like to think that I played a part in that it was.
• 5:02
  It was all because yeah, I mean you did help them get all
• 5:06
  together. Right.
• 5:07
  You you but I think yeah. Like the fact that they like
• 5:11
  they have this Rich network of people that they can reach out
• 5:14
  to, to help them. You know, could contribute to
• 5:16
  this book is really cool and it's a play.
• 5:19
  It's I feel really privileged to be you know, in the same Circle
• 5:22
  as some of these guys you know and and learning from them in
• 5:24
  person you know on a podcast which is this episode we hear a
• 5:28
  lot of that. That is just even their tone,
• 5:30
  their perspective, towards things are very common and and
• 5:33
  have a good. They don't take themselves too,
• 5:36
  seriously, which I think also is an important thing to think
• 5:39
  about, you know, because this is a complicated space.
• 5:41
  And, and when you're trying to learn about it, you need, it's
• 5:43
  nice. When someone can put it in a
• 5:45
  way, that's digestible, it's friendly.
• 5:47
  And yeah, you just see that all throughout the book and
• 5:50
  definitely in this episode as well.
• 5:51
  They're funny guys to be clear dese Kelly and Todd knew each
• 5:56
  other before they Todd reached out to him, actually Todd was
• 6:00
  the reason taught introduced me to do, need to get him on the
• 6:03
  podcast. So the story, the plot thickens
• 6:06
  as they say. But I really appreciated the
• 6:09
  piece where you said, where you ask them about incident response
• 6:12
  and how complicated that can be when it comes to machine
• 6:15
  learning. So I encourage everyone to keep
• 6:18
  an eye out or an ear out for that piece because like you
• 6:22
  said, they have so much experience in this field and the
• 6:25
  way that they're able to frame things is lightning And that's
• 6:29
  it. Let's we might as well jump into
• 6:31
  it for anyone that is listening, and if you are not already
• 6:36
  subscribed, please it would mean the world to us.
• 6:39
  If you yes, do a subscribe to this podcast, video cast.
• 6:43
  And if you're listening on iTunes or Spotify, leave us a
• 6:46
  review that would be huge, if your little appreciate, you
• 6:49
  throw a comment and all of that goes a long way to get this,
• 6:53
  podcast out to more people. So let's jump into it with now,
• 6:57
  Murphy and Todd Underwood. Let's do it.
• 7:05
  So, how am I supposed to follow that one, dude, Todd.
• 7:09
  I mean, maybe we should just start with Pat Hague alien
• 7:14
  reference that you did. I don't even know, man.
• 7:16
  I mean, no, here's what. We'll start with your we go,
• 7:19
  fellas. How's it feel to be back on the
• 7:22
  envelopes? Community podcast, we've missed
• 7:25
  you. I don't know if you've
• 7:26
  reciprocated those kind of feelings for us, but we
• 7:29
  definitely have missed you. Woohoo, feel good.
• 7:34
  I think like, what I would say is that our absence is not for
• 7:39
  lack of fondness for you. And further I will say that if
• 7:43
  anybody this is an open offer anybody ever hears, me, say the
• 7:47
  phrase that book writes itself, you have permission to smack me
• 7:52
  with the nearest macro you can find because I will tell you,
• 7:56
  the book does not write itself. Someone writes the book and it
• 7:59
  turns out in this case, it was, uh, so yeah.
• 8:02
  I think that we have spent a huge chunk of the last year and
• 8:05
  a half working on this Gamba. Yes, that's pretty big.
• 8:10
  That doesn't know that book is reliable machine learning.
• 8:13
  We are going to leave a link to it in the comments and that's
• 8:17
  one of the reasons we got you both here.
• 8:18
  I mean it's been about a year since I mean a little over for
• 8:23
  you. Todd you came in at the end of
• 8:26
  2020, I think. And so what do you been up to
• 8:32
  besides writing books You go, no fair enough.
• 8:38
  Yeah. The past while in common with
• 8:41
  kind of most things out there in the world past while has been
• 8:45
  pretty frantic pretty fast moving.
• 8:47
  I now find myself the CEO of a stealth mode startup in the
• 8:51
  Estuary in ml space called stanza systems, which may I
• 8:55
  think actually the first official announcement.
• 8:57
  Anyway, I'm doing that prior to that, I suppose I I homeschooled
• 9:05
  my eldest child through his entrance to University.
• 9:10
  Those are more or less. The that's the story of the past
• 9:13
  one. That's a lot.
• 9:15
  Yeah. Very busy.
• 9:17
  Very productive. Yes, I find.
• 9:19
  I'm unable to recommend generally as the thing you
• 9:22
  should do, but if you find yourself doing it, then you have
• 9:26
  my sympathy. Yeah.
• 9:28
  I think like you know, that Niles putting together this this
• 9:32
  company that's you know, globally, distributed company
• 9:37
  and we all know like trying to trying to figure out strategy
• 9:40
  and trying to figure out try to get people on the same page in a
• 9:44
  distributed fashion when they can't spend time together really
• 9:47
  hard. It's interesting that as you
• 9:50
  know, as people who live on computers, we do form lasting
• 9:54
  and strong relationships across computers, but I still find it
• 9:57
  baffling that people sort of assert willfully.
• 10:00
  Nobody ever needs to see anyone. Like that's just nonsense.
• 10:03
  Like we do back. Yeah.
• 10:04
  We see each other. It doesn't mean we have to.
• 10:06
  It doesn't mean we can't work around it but it means it's a
• 10:09
  lot more work. So you know I've certainly watch
• 10:12
  Nile doing that really effectively but it's a lot more
• 10:14
  work and you know similarly I've been running the machine
• 10:17
  learning essary Organization at Google, you know, I would say
• 10:21
  now close to half of the people who work for me.
• 10:25
  I've never met because, you know, it's been two and a half
• 10:27
  years and we haven't been traveling much.
• 10:30
  And, and that's just tough because you're just like, I
• 10:32
  don't actually know, you know, like I don't think it's harder
• 10:36
  to remember who people are. It's harder to like, put that
• 10:39
  connection into, they're like oh you moved to this place.
• 10:42
  Under these difficult circumstances but you're having
• 10:44
  this. Oh, you used to be in Seattle.
• 10:46
  But now you're in Zurich, I should probably know.
• 10:49
  Where you are in the world within like three or four time
• 10:52
  zones. So I think that's been toughen
• 10:55
  and really trying to figure out, you know, in particular during
• 10:59
  pandemic I think a lot of us were able to do the simple stuff
• 11:02
  like do the next thing, do the next thing to the next thing.
• 11:05
  But it's been it's been difficult and be ambitious, and
• 11:08
  do big new things because to do that.
• 11:10
  You just have to Future. And so that's for me, has been
• 11:14
  the struggle is getting people. Really if you believe in the
• 11:18
  future, then You can invest in the future and then you can do
• 11:20
  big stuff. But if you're a constantly
• 11:22
  convinced that everything's about to fall apart, it's easier
• 11:25
  to just, you know, answer the email go to the meeting, but not
• 11:28
  build the thing for 20-25. What are we doing for 20-25
• 11:32
  today? Yeah, so that's, that's been a
• 11:34
  struggle. I think any tips on running
• 11:38
  these different distributed teams.
• 11:41
  I mean, I've heard some people talk about how they like to do
• 11:45
  check ins with the team via slack in the And just ask what
• 11:50
  your goals are and what you plan to accomplish today.
• 11:53
  And then at the end of the day, when you check out, you can say,
• 11:56
  what you actually got accomplished.
• 11:59
  Do you like that? Is that to micro managerial?
• 12:04
  Anything that you found that has helped particularly?
• 12:07
  Well, yeah, so we spend a lot of time in slack and we have a
• 12:12
  couple of X Slackers on the staff.
• 12:15
  Actually. I'm not sure if that's the
• 12:16
  plural noun, but you know what I mean?
• 12:18
  The The I suppose there's, there's trade-offs to all of
• 12:21
  these things, right? If you're in a fully remote
• 12:23
  situation, you get to I suppose cast a very wide hiring Nat if
• 12:27
  you're in a completely local situation.
• 12:29
  You get to kind of take advantage of the intensity of
• 12:33
  feeling of being in the same room etcetera.
• 12:35
  But from a tips point of view, we definitely live in slack and
• 12:39
  there is something that comes from that interactive.
• 12:42
  You know, X Y Zed is typing style, feelings.
• 12:45
  That's great. But conversely, Is not very good
• 12:49
  for searchable material retrievable material.
• 12:53
  So you need to pair that kind of tooling with a Wiki or dogs or
• 12:57
  some kind of information management strategy because you
• 12:59
  think that uttering something into electrons is the same as it
• 13:04
  being preserved for all time and it turns out that's not the
• 13:07
  case. So that's one.
• 13:09
  One tip I would have I have I had some teams that found one
• 13:14
  thing that worked, which is there's a particular problem
• 13:17
  which is bringing new people. To a team.
• 13:19
  So now sort of describing the formation of a new team where
• 13:22
  everyone's new and that's one set of challenges.
• 13:24
  The other challenge is bringing two or three, new people into a
• 13:27
  team of six or eight people. And in that case, what we found
• 13:31
  is that especially less experienced or more or more,
• 13:35
  Junior employees, but anyone really coming into a team, just
• 13:39
  has trouble figuring out like, what are we doing?
• 13:41
  What's going on? What are the expectations?
• 13:43
  How do I do this? And so one of the strategies I
• 13:46
  saw which is really simple strategy is You just get on you
• 13:50
  just have the experience people of the team one day per week for
• 13:54
  an hour work on a video conference, they just have a
• 13:58
  they just have a window up there, just working.
• 14:01
  No one has to come by and in some cases they cast their
• 14:04
  screen because they're like, hey, I'm doing this migration or
• 14:06
  I'm trying to write this code or I'm working on this design, doc,
• 14:10
  or whatever. The thing that's awesome.
• 14:12
  And in some cases, they're just working on the screen.
• 14:14
  And they're like, if you got questions show up, and what we
• 14:16
  found is that the junior folks Didn't want to make an
• 14:19
  appointment. They didn't want to set up a 30
• 14:21
  minute needing to ask a question because they were embarrassed,
• 14:24
  but if they know that each of these four hours of the weeks,
• 14:27
  someone who knows, what the heck is going on?
• 14:29
  Is going to be there. They might pop by and say like,
• 14:32
  oh hey how do you do this thing or where are these files or
• 14:35
  explain the source? Treating me one more time
• 14:37
  because I don't get it and that helps a lot one.
• 14:40
  Follow-up question for that is is I feel like a lot of teams
• 14:44
  are organizations will have one person that's like really
• 14:46
  skilled in this area. And they rely the organization
• 14:50
  relies on them to kind of, you know, take care of things at the
• 14:52
  strategy and teach others. How do you prevent those type of
• 14:55
  people from getting burnt out? Because that sort of scenario
• 14:58
  where they're constantly available, I could see that for
• 15:00
  myself. Getting a little bit tiring.
• 15:03
  What do you do to balance that? So it's not, I mean, I love that
• 15:06
  idea. Let me just first say, I think
• 15:07
  that's really great because it does make people more available,
• 15:09
  more accessible, right? And the more likely for someone
• 15:12
  to ask a question, maybe when they wouldn't do it through like
• 15:15
  a one-on-one meeting, but how do you do that in such a way where
• 15:17
  they're not overwhelmed with? Like, you know, I'm embodiment
• 15:19
  of questions and people wanting to see how they do things, you
• 15:22
  know. Yeah, I think in our case, I
• 15:26
  think one of the things you have to do is you have to reward, you
• 15:29
  have to consciously identify this as a set of behaviors that
• 15:33
  you're requesting of your senior Engineers or if you're you know,
• 15:36
  particular class of experienced employees and you have to reward
• 15:39
  it. So one of the things that I
• 15:41
  think a lot of organizations found in pandemic is they
• 15:45
  explicitly rewarded a set of things, like productivity, like
• 15:48
  ship the code, fix the bugs, launch, the product, and they
• 15:52
  did not. Explicitly reward mentoring
• 15:55
  like, teaching encouraging. And so what happened during
• 15:58
  pandemic is a bunch of the senior employees had an
• 16:01
  opportunity to figure out just how much more productive they
• 16:05
  were at the things. They were being formally
• 16:07
  evaluated on at home and how much less productive they were
• 16:12
  at those things at work but at work they were much more
• 16:15
  productive on the set of things that the companies said they
• 16:18
  cared about but didn't actually evaluate or reward on.
• 16:21
  And so I think to some extent this is Some management problem.
• 16:24
  Like if Senior Management doesn't want to train up new
• 16:26
  employees, then they should not prioritize training up new
• 16:29
  employees and they'll have to deal with the fact that like
• 16:31
  they have a high failure rate of new employee.
• 16:33
  I think that's a bad strategy, I should say, but it is uh,
• 16:35
  strategy. The other strategy is you say
• 16:37
  like, oh, mentoring is actually a part of your job or we're
• 16:40
  going to carve out 25 to 30 percent of everyone's job to do
• 16:43
  that. Maybe you trade those off a
• 16:45
  little bit, but to your point David like I do think that some
• 16:48
  people are just bad at this. Like some people don't like, I
• 16:51
  don't like other people. I don't like to talk to them.
• 16:53
  Um I don't I don't I don't like to talk to you all the truth
• 16:56
  comes out, its people against their interest.
• 16:59
  This is like interviewing to me like most companies need to
• 17:02
  interview there are some people that are terrible at
• 17:05
  interviewing or hate it so much that it ruins their entire week.
• 17:08
  Every time they have an interview to probably find a way
• 17:11
  to not have those people interview for your company.
• 17:14
  Yeah, fair enough. And I think this is a night, I
• 17:16
  don't know if I don't want to jump to ahead into the book, but
• 17:19
  in the book, you do mention like this idea of process rewards in
• 17:22
  people. And I found that Really
• 17:23
  insightful to consider rewarding people because you do need to
• 17:27
  keep people incentivize, right? So that they are not just doing
• 17:31
  the bare minimum but we'll do a little bit more and it's really
• 17:34
  hard to get people to do that when they feel like it's like I
• 17:37
  don't know you have to frame it in a certain way to encourage
• 17:41
  people to want to do that. So can we talk a little bit
• 17:43
  about how you guys thought about that for your book?
• 17:47
  Sure, I can start off, I can kick off if you want and just I
• 17:53
  suppose I'd opened by saying it's so strange that in the year
• 17:58
  of Our Lord 2022, we are saying it is a revolutionary management
• 18:02
  technique to reward people like in the first case it doesn't it
• 18:07
  seem relatively intuitive that if you want to sort of inspire
• 18:12
  certain behavior and people that you might positively reward
• 18:15
  though. So I I refuse to regard that as
• 18:18
  a revolutionary technique, I can say that I think the subtlety
• 18:24
  and management has always been precisely, what do you reward?
• 18:28
  The more kind of narrow and Metric.
• 18:30
  Based you get about it. The more difficulty you often
• 18:33
  encounter because people find a way, people are very, you know,
• 18:37
  ingenious in many ways and they find all kinds of ways to make
• 18:41
  the metric go up. I think there's a relatively
• 18:44
  famous case of a bank. Rewarding.
• 18:47
  The The number of new accounts being opened, you know, in an
• 18:50
  attempt to drive new business. Oh, yes.
• 18:53
  Yes, we can definitely open lots of new accounts.
• 18:56
  That is true. Yeah.
• 18:58
  So anyway, leaving aside questions of metrics, I think,
• 19:03
  in some cases may be in many cases, introducing machine,
• 19:07
  learning to an organization that hasn't really heard this before,
• 19:10
  or hasn't had it in a kind of a thorough way.
• 19:13
  In anything other than individual experiments is like
• 19:16
  any other change. Eject that you have in any other
• 19:18
  organization, people need to understand why they're doing it
• 19:22
  and they need to in some sense believe in it I suppose and
• 19:26
  process and rewards and persuasion and all of that are
• 19:31
  huge part of driving change in any organization.
• 19:34
  I think one of the things we emphasize in the book is
• 19:38
  precisely how for want of a better term tentacle-like.
• 19:43
  Machine learning is and the sinews of data.
• 19:47
  Find their way into everywhere in your organization, more or
• 19:52
  less. So it's a lot tougher to take
• 19:55
  the traditional, kind of computer science or even
• 20:00
  organizational. Approach of we will Silo this
• 20:02
  into this tang and they'll be somebody in charge of it and we
• 20:05
  can just just and quote give all of that thing to this particular
• 20:10
  group. And that seems to me to be just
• 20:13
  kind of less tractable in the ml case.
• 20:15
  All kinds of people need to know, All kinds of things about
• 20:19
  how ml is done. Well, the goal is how successful
• 20:23
  it's being and the degree of kind of cross communication.
• 20:27
  I think we argue is, is strongly increased.
• 20:30
  What do you think Todd? Yeah, so I think that's, that's
• 20:35
  an important point. My first I want to say, like,
• 20:37
  David, like I love that you asked this question because when
• 20:40
  we, so, one of the things I really loved about the first, a
• 20:43
  storybook that Niall did was the section that I think, almost no
• 20:47
  one. Read.
• 20:47
  But was my favorite section, which is like, how do you manage
• 20:50
  these? A sorry teams.
• 20:51
  Like there's a whole section on like how do you run a production
• 20:54
  meeting? And I was like to those of us
• 20:56
  who are SRE managers and like, this is the gold.
• 20:58
  This is the hard one, this stuff matters.
• 21:01
  Like, how do you talk about outages?
• 21:03
  How do you like, coordinate the activities of a team?
• 21:06
  But most people are, like, show me the tech.
• 21:08
  How do I do this? Like what is the thing I type
• 21:10
  into the command line? Well, the thing you're hacking
• 21:13
  is people not computers and so, so I love that section and so
• 21:17
  when we To write this book on reliable ml.
• 21:20
  Like we want to try to address organizational leaders who
• 21:25
  themselves are not into ml yet but no, they should be there.
• 21:28
  Like they think this stuff is going to matter or maybe they
• 21:31
  already know it matters, but they're not sure what to do
• 21:34
  about it. And already we get these
• 21:36
  questions semi constantly even from relatively technically,
• 21:39
  sophisticated companies of how do I either?
• 21:42
  Introduce machine learning into my business into my
• 21:45
  organization, or how do I hard I've assessed the energy that
• 21:49
  I've got, how do I harness the different teams that I already
• 21:52
  have in a sensible way and so when we were in sure, rate,
• 21:55
  those chapters I seriously did think nobody's going to read
• 21:57
  this but if they do, it's going to be great.
• 22:00
  And so I was really it's exciting that you actually look
• 22:02
  through those. Because Yeah, you mentioned, I
• 22:04
  was really good for managers but I also think for individual
• 22:07
  contributors themselves to get a sense of like how to mature and
• 22:10
  grow, I think myself. I don't really see myself being
• 22:13
  a manager, but as an individual contributor, I do want to think
• 22:16
  more maturely and And part of that is, is thinking broader you
• 22:20
  mentioned in your book, I think it was about like staff
• 22:22
  engineer. So they should be very cross
• 22:24
  functional right or work across the York, instead of just being
• 22:27
  so focused in their narrow domain.
• 22:30
  I'm curious from your experience.
• 22:32
  Why why does that work? Well, why does it work well to
• 22:34
  be? Yeah, I guess how does that
• 22:38
  work? Well, because it seems like kind
• 22:39
  of my earlier question. We're getting burnt out being
• 22:41
  all over the place. See, I feel like they're
• 22:43
  related, but maybe in your experience, that's not the case.
• 22:46
  It just naturally works and there's Has no issues getting
• 22:48
  people to kind of collaborate across orange there.
• 22:51
  Absolutely is like I think that there's an analogy and
• 22:54
  something, I'm not sure this is a Perfect Analogy, but an
• 22:57
  analogy and something I'll said about.
• 22:59
  So when you go back it's hard for us.
• 23:01
  But like, when you go back in time to the introduction of
• 23:05
  Information, Technology into Enterprise.
• 23:08
  Like so we're wandering through the 50s and 60s, and people have
• 23:12
  big businesses and they're moving pieces of paper all over
• 23:15
  the place and they're sending memos and they Steno Pools full
• 23:18
  of almost entirely women who type things and male things up
• 23:22
  like this is how business runs and computer start to come in
• 23:26
  and no one's really sure what to do with them.
• 23:28
  And I think there's an interesting analogy here because
• 23:31
  they pick a few things that they're positive, that computers
• 23:34
  work really well for and have them do those things but that's
• 23:37
  not where it matters. It matters.
• 23:39
  It turns out everywhere not somewhere, not in some places
• 23:43
  but everywhere and if you look at the 80s and the 90s and the
• 23:46
  2000s is people are trying to decide like do we have
• 23:49
  technology divisions per business group?
• 23:52
  Do we have an IT organization? Do it?
• 23:55
  Like, how do we think about and is technology of cost center or
• 23:58
  a profit Center? Like and so like this is all ml
• 24:02
  is going to do the same thing. And I think the reason it's
• 24:04
  going to do the same thing as the data comes from everywhere,
• 24:08
  like the data are already everywhere.
• 24:10
  Like if you do a thing that matters to your business, that
• 24:13
  thing you're doing produces data, and those data could Used
• 24:18
  to make that part and other parts of your business better.
• 24:21
  So if you sell Airline seats and you know what, airline seats you
• 24:26
  sell, and who you sell them to that can make the catering
• 24:29
  division of your airline more efficient.
• 24:32
  But only, if you use the data from the seats, you sell to
• 24:35
  predict the meals, you're supposed to be producing.
• 24:38
  But like you would have to share those between those business
• 24:41
  groups, and similarly, for magazines, and for car
• 24:45
  manufacturing. And for You know, serving ads on
• 24:48
  search Pages or whatever the heck you're doing.
• 24:50
  So I think that's really where it comes about.
• 24:52
  David. Yeah, go ahead and oh yeah, it's
• 24:55
  just going to say. I agree with what Tom said
• 24:58
  there. Except the data is everywhere
• 25:00
  rather than the data are everywhere.
• 25:02
  But anyway that is a plural word and they bullied me in this
• 25:10
  book, we use data as a collective singular and so it's
• 25:14
  wrong. I know.
• 25:18
  As if making this mistake on purpose because Nile is wrong
• 25:23
  and Kathy, T is wrong and D. Scully, they're all wrong.
• 25:27
  Okay, sorry. Go ahead.
• 25:29
  Data is everywhere in your organization.
• 25:31
  Even if you don't know it, there are probably places where it
• 25:34
  doesn't escape the context of the local team, or the person
• 25:37
  who's keeping it or whatever. But ml is a way to
• 25:41
  instrumentalize that dated to way to do something with it.
• 25:45
  It's a way to Surface, its actual significance.
• 25:50
  May be in in different places than it would otherwise be.
• 25:54
  I mean to come back to Todd's Point about your lines and
• 25:59
  catering and so on so forth. Probably there's some kind of
• 26:02
  rough rule of thumb, where the catering Department orders,
• 26:05
  approximately statistically about the same amount of
• 26:08
  vegetarian meals that somebody told them to do and 2011, or
• 26:12
  something like that and then maybe they changed it every
• 26:16
  year. But Being geared on the data and
• 26:20
  surfacing that every day and processing.
• 26:22
  That would allow you to make much stronger and more relevant
• 26:26
  predictions and maybe you have to spend less or maybe you can
• 26:31
  predict with more accuracy and so on.
• 26:34
  Yeah, that's a good point. Yeah.
• 26:35
  And you know, I guess what I assumed that maybe we can kind
• 26:38
  of finish the train of thought, but Todd made a point about like
• 26:41
  it's yeah, what I thought was, it seems like tooling can really
• 26:44
  help like not only harness, the value.
• 26:47
  Of data but allow other people to collect.
• 26:50
  Like, I'm thinking of like a feature story like this whole
• 26:52
  idea of like, you know, across functional data right.
• 26:55
  Getting people to share data and you reuse it.
• 26:57
  But I guess what I want to maybe we can the kind of wouldn't
• 26:59
  talk. I'm back on and is that that
• 27:01
  interesting relationship between like tooling and also at the
• 27:06
  same time, like, I guess enabling people or productivity
• 27:10
  or like, like what Todd was saying, harnessing, the power of
• 27:14
  what's there, I feel like it's a, there is a there there.
• 27:17
  Very closely related, you know, and I feel like there's, like,
• 27:20
  you mentioned data is everywhere.
• 27:21
  I always feel like machine learning is mostly about data
• 27:24
  like a lot of the time and I feel like there there that's, I
• 27:26
  just, I think about that, you know, how to why it?
• 27:29
  Why that is and how they're related and how that shapes the
• 27:32
  tooling that we have and the process around that tooling.
• 27:34
  So I'm curious to hear you and thoughts thoughts around that
• 27:37
  because it seems like it is an important part, right?
• 27:39
  Like the so I guess a lot of people say it's about the
• 27:42
  business value or doing, what's best for the business but I
• 27:45
  guess something underneath that simplicity.
• 27:47
  Is like productivity and being able to like really extract
• 27:50
  value from something that everyone can do but only a few
• 27:54
  people do well because there's like all these complexity
• 27:57
  complexities around. It nice.
• 27:59
  This has been perfect because I appreciate the ideas around how
• 28:05
  we need to look at a machine learning.
• 28:08
  As so close to the business side of things.
• 28:11
  Not just like this software side of things that you can.
• 28:15
  Like you said, Todd like it. Be siloed off and just do
• 28:19
  something. It's an IOU, mentioned it as
• 28:21
  what, like an octopus and its tentacles are everywhere.
• 28:25
  And then the other piece of this is the cultural aspect that
• 28:30
  doesn't get talked about enough with ML Ops and really ml ops's.
• 28:36
  Borrowing from devops practices heavily and devops has so much
• 28:42
  based on the cultural side of things.
• 28:44
  And so I appreciate you bringing Ringing both of these pieces up.
• 28:49
  Now, I do want to just jump into a few things because as David
• 28:55
  was saying with the staff engineers and thinking, broader,
• 28:59
  and being able to then utilize that in the business and and how
• 29:04
  they're able to utilize data in new ways, does this also work
• 29:10
  for smaller startups and especially Nile?
• 29:12
  I mean, you're living and breathing this.
• 29:14
  Now, right? You went from And two companies,
• 29:18
  where you had these staff engineers and what over 100,000
• 29:23
  people that were working in the companies.
• 29:25
  And now you're leading the ship of a small start-up.
• 29:29
  How do you think about the two roles in that aspect?
• 29:35
  The two roles of machine learning and stuff engineers.
• 29:39
  The sorry, it's really one role in two different places, like
• 29:43
  the staff engineer or just the machine learning engineer, in
• 29:46
  general, being able to be broad but at a start-up.
• 29:51
  Are you still like what are the use cases in the difficulties
• 29:55
  that you're facing there as opposed to when you're at a
• 29:57
  Google? Sure.
• 29:59
  So I mean there's a lot we could say here organizational, theory
• 30:05
  has many, many books, and column inches, and so on.
• 30:08
  Written about it. But to my mind, those kind of
• 30:11
  like a couple of differences at the start.
• 30:15
  The first thing I'd say is as my friend, Tanya Riley writes in
• 30:19
  her stuff engineering book, which is due to be released
• 30:22
  soon. She basically talks a lot about
• 30:25
  how being a staff engineer in the larger organization, you're
• 30:29
  essentially modeling Behavior, So there's a human kind of
• 30:34
  modeling aspect and setting culture and so on, which is, of
• 30:38
  course, present, in a start because it is present wherever
• 30:41
  there are human beings, but it's much less a question of scale
• 30:45
  and much more question of can we come to a rough agreement about
• 30:47
  how they should work? And then let's go.
• 30:50
  But the second piece I suppose that a staff engineer does is if
• 30:56
  you're looking around wondering who should do something about
• 30:59
  this problem, the answer, Is probably you because you're the
• 31:04
  stuff engineer and that's a very similar situation to a start-up
• 31:08
  where, in fact, if this thing isn't getting done and it needs
• 31:12
  to be done and usually there's a fairly hard and fast rule that
• 31:16
  allows you to decide whether or not it should be done.
• 31:19
  Then first person to get there, you know, particular leaving
• 31:25
  kind of questions of individual specialization aside but the
• 31:28
  first person to get their should should do that.
• 31:31
  You know, we Could pick up after ourselves.
• 31:33
  So thank you. I think there's a, there's a
• 31:36
  connection between that point and the point that David was
• 31:38
  making earlier which is it is characteristic of the staff
• 31:45
  engineer to have a broader scope to be sort of undaunted by
• 31:52
  organizational boundaries or technical boundaries which is
• 31:55
  tough I think. But so that sort of like if
• 31:57
  you're wondering who's supposed to figure out how that division
• 32:00
  is supposed to work. With this Division, if you're
• 32:04
  not talking to anybody else who's asking the same question,
• 32:06
  that's probably you. And that's a really nice fit for
• 32:09
  machine learning because often it is the case that the best the
• 32:13
  best data or the best idea or the best collaboration happens
• 32:17
  at the interspaces of these organizations like frequently,
• 32:22
  you know the data that you need to make the prediction or the
• 32:25
  data that you need to give the users, the value, whatever it is
• 32:29
  isn't in one place yet. Because that's expensive and
• 32:32
  because it requires a lot of privacy oversight and it
• 32:35
  requires a lot of governance and a lot of Technology like.
• 32:37
  So most companies can't say, oh, all the data, they're right
• 32:41
  here. Here's where they are, we're
• 32:43
  good. Like, just look at what you
• 32:44
  want. I mean that sounds nice.
• 32:45
  And it's actually like but it has some downsides, right?
• 32:48
  Like if all the data are right here, can anybody take them are
• 32:51
  like they exposed or they lockdown is access restricted in
• 32:55
  appropriate ways and so both by inertia and by concerns about
• 33:00
  governance most Organizations have a bunch of different data.
• 33:04
  We don't talk. Everybody's like assume the
• 33:06
  perfectly spherical data Lake all in one place, but that's not
• 33:09
  what like nobody has that. And so, that's why these these
• 33:13
  inter in intra-organizational. But inter division roles like
• 33:18
  the staff engineer, so important for machine learning because
• 33:21
  they have to look around and say, like look your data plus my
• 33:25
  data. That's the magic.
• 33:26
  So let's figure out a way to make this work.
• 33:28
  I love that. Yeah, I would I just I'll get
• 33:31
  Back. I had a question but I'll ask in
• 33:33
  a second good to meet, you know, I was gonna jump into something
• 33:37
  that Todd said last time he was on here, unless you remember
• 33:41
  what you were going to say David?
• 33:42
  Yeah, so something I thought about when you're saying how
• 33:45
  important it is that, you know, they, the intersection of these
• 33:48
  different teams is where a lot of the magic happens.
• 33:51
  Can we talk about that in the context of like, you know,
• 33:54
  incident response and and why it's particularly important
• 33:57
  there that interaction between these teams goes, well, and
• 34:01
  that. They can work together
• 34:02
  effectively in your experience. First up.
• 34:07
  What is that typically look like at most companies, we're how
• 34:09
  they interacted you with incident specifically and maybe
• 34:14
  some best practices that you've learned things that work.
• 34:16
  Well and things that don't work because I'm thinking along the
• 34:19
  lines of, you know, into interaction is, I should be have
• 34:21
  a meeting who goes into that meeting, who gets assigned what?
• 34:26
  And and like, again, like these interesting questions of like
• 34:29
  who's who has ownership of something I think.
• 34:31
  Is really it comes up when something fails and when only
• 34:36
  one person, maybe knows how to fix it.
• 34:38
  And and and I just wondering like, what does this look like,
• 34:40
  you know, Nile and the startup context had a larger context and
• 34:44
  what are some things that we can learn from that that someone
• 34:47
  listening can benefit from? So, if you're talking about the
• 34:51
  question of in, in a world gone, mad in a world where kind of, no
• 34:55
  one owns everything or the first person to get to a problem fixes
• 34:59
  that problem, how do Little methods of problem resolution
• 35:04
  work. And the answer is well, they
• 35:06
  don't work because the first person who got their fixed, the
• 35:09
  problem, or they ask for help or whatever, like at the at the
• 35:13
  small scale, there's only so much kind of escalation you can
• 35:18
  do. So it's much less of a question
• 35:22
  as it is in larger organizations of the organization.
• 35:25
  Needs to find the person with the right knowledge to fix this
• 35:27
  problem. And the problem is essentially
• 35:29
  finding that person, Which I think is a lot of how incident
• 35:35
  response works and kind of larger companies and folks, with
• 35:38
  the significant production estate.
• 35:40
  But if you're coming back, or if you're digging deep on questions
• 35:44
  about organizational scope and communication, and so on and
• 35:48
  saying that actually, is it all that realistic to say that
• 35:52
  everyone needs to know everything and can't you have
• 35:55
  some partitions somewhere and so on?
• 35:57
  Yes of course. I mean there's this apocryphal
• 35:59
  story which I choose Leave is true but anyway of Jeff Bezos.
• 36:05
  Apparently getting annoyed, some relatively early Amazon meeting
• 36:10
  where lots of folks are kind of standing up managers and p.m.
• 36:14
  saying we have to communicate more we have to send memos to
• 36:18
  each other or whatever and Jeff says no this is stupid you
• 36:21
  should be finding out ways that you don't have to talk to each
• 36:24
  other because that will mean you can go faster within your local
• 36:28
  scope. And this is the to beat sodium
• 36:30
  and the there shall be Be nothing except an RPC layer team
• 36:33
  and so on, all of which are fantastic kind of rules of
• 36:35
  thumb, for running organizations with kind of well known
• 36:39
  characteristics. I suppose.
• 36:42
  So I would say that my experience suggests that, yes,
• 36:49
  you do need to put some partitions in.
• 36:51
  Once you go past a certain scale, or once you have pii
• 36:55
  concerns or regulatory concerns various kinds, there are all
• 36:58
  kinds of reasons too, Block various kinds of communication
• 37:03
  flow or having to know stuff. But what I see in mainstream
• 37:08
  business culture is that we still assume we have way fewer
• 37:12
  reasons to talk to each other than we actually do.
• 37:15
  And we're at least in the in the stove we have written.
• 37:19
  We're trying to get people to realize that actually ML and
• 37:23
  liberalizing the value of data should cause you to think again
• 37:29
  about that question. When I think, yeah, I think all
• 37:33
  of that is true and in the context of machine learning like
• 37:37
  we wrote this whole chapter on incident response in in the book
• 37:41
  and it was a fun chapter 2 rate because it like, you know, like
• 37:45
  for people who do production engineering like thinking about
• 37:48
  stuff breaking like that's that's like that's the fun stuff
• 37:51
  like their stories, they're fun to read but they're actually
• 37:54
  interesting to learn a ton you do, right?
• 37:59
  And all of those were inspired by Real.
• 38:02
  Like, none of them is an actual exact outage that happen
• 38:05
  somewhere, but all of them are inspired by real outages that
• 38:09
  some of us have seen. But I think the challenge with
• 38:13
  machine learning, which is, I think this is sometimes true in
• 38:17
  other parts of the business, but it's always true in machine
• 38:19
  learning, is there is no works or doesn't work.
• 38:23
  I mean, if you think about what it is to make a model, like
• 38:26
  we're making a model and it's better than not having a model
• 38:30
  you're like, but is it right? Right.
• 38:32
  You like? Well, that's not a super
• 38:34
  well-defined question. Is it right?
• 38:36
  I mean, it's pretty good for lots of stuff.
• 38:39
  Sometimes you're like, okay, but so that's fine.
• 38:42
  When you're thinking about model development, you're thinking
• 38:45
  about like building this into your organization or your
• 38:47
  application. But now, when you come to do
• 38:50
  reliability engineering or production engineering, you're
• 38:52
  doing ml Ops you like the models less good than it used to be.
• 38:58
  Is that okay? I don't know less good for whom
• 39:00
  under what circumstances and this makes the incident response
• 39:04
  super hard. I think it's super interesting
• 39:06
  because you're like do we because sometimes you just have
• 39:09
  an outage you're like oh we loaded the model and just
• 39:11
  serving in all of the serving layer crashed.
• 39:14
  Okay. Well that's an out it like
• 39:15
  that's fine. We don't have a model but like a
• 39:17
  lot of the times like we have a new model and in some ways we
• 39:21
  think it might be somewhat less good for certain kinds of people
• 39:24
  than the old model was. Okay.
• 39:27
  Is that an outage? I don't know, it sounds kind of
• 39:29
  like an outage, right? How are you going to know when
• 39:31
  it's done? And so anyway, like David going
• 39:33
  back to your point like that creates these communication, you
• 39:36
  know, part of the point that now was talking about his, now you
• 39:39
  need to communicate with the people building the model, with
• 39:42
  the people who are evaluating the quality of the model.
• 39:44
  With all of the people who produce all of the data that
• 39:47
  goes into that model because like these data in the ways that
• 39:51
  their combined, and the way that they're trained into this model,
• 39:55
  like maybe they were joined wrong.
• 39:56
  Maybe somebody like hasn't produced New data for six weeks.
• 39:59
  And that's why this model is that it you don't know that by
• 40:02
  the time it gets to serving because you just have a model
• 40:04
  that says you know what it says. So I think that's what's, that's
• 40:08
  what makes the ml incident response way trickier and much
• 40:11
  more fun. Yeah, that fun part sometimes.
• 40:15
  It's sometimes it's stressful. You know what?
• 40:18
  You're trying to fix something and knowing no one really
• 40:21
  understands why something is gone.
• 40:22
  But I think that's something that I thought about is if that
• 40:25
  if those are the type of questions that you're asking
• 40:27
  that, I would dig into that. Why do you not know who is
• 40:30
  responsible for that or or why don't we have a system in place
• 40:34
  to you know frame you the right questions.
• 40:35
  It almost seems like framing. The questions is almost the
• 40:39
  important part and that seems related to team structure and
• 40:41
  goals and what your priorities are and things of that nature.
• 40:46
  You should be aware of the Todd uses the word fun in the context
• 40:49
  of someone who has been in a sari for well over a decade and
• 40:51
  therefore, their adrenaline, glands are close to useless but
• 40:57
  I do think like one of the things I appreciate about my
• 41:00
  current employer is I've never been on call and felt alone like
• 41:05
  abandoned and I think that's what gives the stress if you are
• 41:09
  in it and you have people to help you with it and you are
• 41:12
  kind of lost. But you know what to how to ask
• 41:15
  for help. It's not so bad.
• 41:16
  You're like yes, this is bad, but we're in this together.
• 41:19
  We're going to figure it out. I think, for most of the people,
• 41:21
  I talked to they're like, they're at an organization that
• 41:24
  either because of its size or its maturity or it's bad
• 41:28
  culture. Just like you're on calling her,
• 41:30
  get it man it's just you good luck to holding the hordes off
• 41:35
  and you're like I don't like I just got here I could barely
• 41:38
  spell LS what is happening in this conversation rain?
• 41:42
  So in those cases like if you don't know where the dashboard
• 41:44
  is you don't know What the model supposed to do and you don't
• 41:47
  understand how qualities being evaluated and you didn't
• 41:50
  understand the data joining system that went into it.
• 41:52
  It's time to like get some help and that's should be fine.
• 41:55
  And so, so that one thing that I'm wondering John and I hope my
• 42:01
  bad Nile go ahead. No I was just going to say you
• 42:05
  haven't lost me. I'm still here.
• 42:07
  Okay so but Todd I mean as someone who Had to write your
• 42:16
  title a few times in the last time that you were on here and
• 42:23
  then when you were in the apply conference, it's a mouthful.
• 42:27
  Do you think that your or we as a ml Ops or ml Community?
• 42:32
  Are going to see something like the MRE actually take off.
• 42:39
  Yeah, so honestly I think maybe I don't understand, you know
• 42:45
  this is weird having worked at Google for like 13 years.
• 42:48
  But I will just say, like, I don't understand the way big
• 42:51
  corporate Tech deals with job titles and professionalism.
• 42:54
  Like I just don't understand how it works or why it works the way
• 42:57
  and I know this is like a little bit ridiculous thing.
• 43:00
  Like I'm a senior director at a giant Tech Corp and I'm just
• 43:03
  asserting. I don't understand why or why we
• 43:05
  do that or how that works. So I don't know like what I know
• 43:09
  is that as high Ring manager. Like I know the kind of skills I
• 43:13
  evaluate for but I also I do appreciate that there is this
• 43:18
  there is this like social professional wave that we all
• 43:22
  are creating and participating in and I know it matters because
• 43:25
  we saw in essary like before we had a sorry before.
• 43:30
  Like, you know, Google popularize the term and other
• 43:33
  people adopted it and we started collectively thinking about what
• 43:37
  site reliability engineering would be.
• 43:39
  Versus systems Administration versus some other kind of like
• 43:43
  you know versus production engineering or devops lake.
• 43:46
  So before we had that it was harder to communicate what
• 43:50
  skills we were looking for. So I do think we're going to get
• 43:53
  to something that's a little bit closer to that but what I wonder
• 43:56
  I would flip it on you Demetrius and say like well why do we
• 44:00
  think that that won't be all of the production engineer?
• 44:03
  Like what is the thing looking out?
• 44:05
  10 or 15 years? What is the set of?
• 44:08
  Because like you don't go Back, you know, you don't say like do
• 44:11
  you think we're going to like have software skills required to
• 44:15
  work on computers? Like what do you think about
• 44:17
  that? And like, yes, I think we are
• 44:19
  and I think we're going to have some machine learning skills
• 44:22
  required to work on computers in the future.
• 44:25
  Now we're going to have five to 15 years of that transition, but
• 44:29
  I don't see a world 15 years from now where people are using
• 44:32
  computers and the computers aren't configuring themselves
• 44:35
  and the cute the systems that we deploy aren't fundamentally.
• 44:39
  And Inextricably dependent upon machine learning.
• 44:42
  Like I just that just doesn't seem possible to me in which
• 44:46
  case why do you like? You're just like inventing a new
• 44:49
  word for the thing that everyone is?
• 44:51
  So maybe the answer is like, no you don't get like ml
• 44:54
  reliability, you don't KML reliability Engineers, you just
• 44:57
  have reliability Engineers, who have to know?
• 44:59
  I'm out. So yeah.
• 45:00
  Now what do you think? So Todd has moved to occupy The
• 45:02
  aspirational High Ground. So I'll come in with a little
• 45:05
  bit of low ground here and I'll say that to the extent.
• 45:09
  And that s re has obtained. A good reputation in the in the
• 45:15
  industry as a whole people are going to want it bad of that in
• 45:19
  there org. So somebody comes up to them and
• 45:22
  says oh you can have an SRE except for this ml tang and
• 45:26
  nurseries are really good. Didn't you hear that?
• 45:28
  Oh yes I heard that. Yes we'd better hire and Mallory
• 45:32
  or equivalent. So like there's just kind of
• 45:35
  linguistic degradation overall is people attempt to get the The
• 45:39
  good things for themselves. So that would be my, my
• 45:44
  estimation, what's going to happen.
• 45:47
  So I love this, especially because it does feel like, yes,
• 45:53
  the and this is something that Todd, you've talked about a few
• 45:56
  times, and I really appreciate the idea of how the hard
• 46:01
  problems aren't specific machine learning problems, right?
• 46:04
  And it's like distributed systems problems and distributed
• 46:08
  computing. And I kind of thing.
• 46:10
  And then this Vision that you have of the next wave is going
• 46:16
  to be everything will be touched with machine learning.
• 46:19
  So it's not going to be a thing. Like machine learning won't be
• 46:23
  such a, it won't be as hype as it is now and it'll just be
• 46:26
  taken for granted. And potentially what we call
• 46:30
  software in the future will have machine learning inside of that
• 46:34
  just it's assumed that machine learning is part of software.
• 46:39
  So, So the one thing that I want to touch on that is something
• 46:44
  you mentioned last time, you're on the podcast, Todd around,
• 46:48
  trustworthiness being a real barrier to implement machine
• 46:52
  learning and over the span of writing the book.
• 46:56
  I know that you all touched on ethics and trustworthiness and
• 47:00
  things like that. And I'm really wondering like
• 47:04
  how has, like, how have things changed?
• 47:08
  If at all your Point on the trustworthiness of not only the
• 47:12
  data but like if someone wants to do machine learning, that is
• 47:16
  a big risk for the company and so they have to trust that this
• 47:22
  initiative that they're going to embark on is going to work.
• 47:28
  Yeah, so I think we take these points, super seriously.
• 47:31
  And I think that some of the most interesting macros
• 47:34
  structural organizational societal questions that we have.
• 47:38
  I think there's a narrow technical answers about explain
• 47:41
  ability and about certain kinds of governance.
• 47:43
  And I think those are important answers, but I think that, you
• 47:47
  know, so like just to say, like explain ability is a real thing.
• 47:51
  Like many models have certain explain ability features.
• 47:55
  I think that the people doing work in Ability are doing
• 47:58
  important work. However, I also think it's the
• 48:01
  case that there. They do not see the gap between
• 48:04
  the answers, they give and what society's want.
• 48:08
  And so, what I think is pretty hopeful, is, you know, if we go
• 48:13
  about I am not a policy expert and like my company will kill
• 48:16
  me. If I say important things about
• 48:18
  like big policy, but I will say, like, we look at data protection
• 48:23
  rules in the European Union going back, you know, five, six,
• 48:27
  seven. Years.
• 48:29
  Those were super bumpy super problematic.
• 48:32
  They looked almost impossible, but in the end, I think a lot of
• 48:36
  people, look at those and say like wait, those are not, those
• 48:39
  are actually not as a person, maybe not as a tech company.
• 48:42
  But as a person like those are not too bad.
• 48:45
  Wait, you're saying, I have certain rights to control of my
• 48:48
  data. I have rights to know.
• 48:49
  When you're getting my data, I can make you get rid of it, and
• 48:53
  I can know what you're doing with it.
• 48:54
  That sounds pretty reasonable to me, right?
• 48:57
  So, That went from unfathomable to, you know, Tech in the leg,
• 49:02
  early 2000's to like, seeming, pretty reasonable to those of us
• 49:06
  here in 2022. I think we're also starting to
• 49:09
  see some of those things for artificial intelligence and I
• 49:12
  think that's good like as a as a technologist.
• 49:14
  I'm like, I should just get to do whatever I want to do, but as
• 49:17
  a citizen of like a couple different places.
• 49:19
  I'm like, no, I want. Like I don't want societies to
• 49:22
  have a deep set of trust and investment in these Technologies
• 49:26
  because they're good. Good because they matter.
• 49:28
  And so if we're going to use these two right loans, we should
• 49:32
  probably not be racist about it. For example, like if we're going
• 49:34
  to use these to drive cars, we should not run into people and
• 49:38
  we should know under what conditions were better or worse
• 49:41
  than human drivers. But now, like you spent a lot
• 49:45
  more time and that's section of the book than I did, what are
• 49:49
  your thoughts on? You know, this general question
• 49:52
  about like the risks of AI in organizations and like where
• 49:56
  we're going to go and then it's You years.
• 49:58
  Yeah. I suppose I'd start off by
• 50:01
  saying that the the chapter in the book that we have about.
• 50:04
  This is more or less geared to organizations that aren't maybe
• 50:09
  doing a lot or that much or their take their stepping
• 50:14
  putting their first toe in the water etcetera.
• 50:16
  So it's more or less a guideline to if you aren't doing anything
• 50:19
  at all, please consider doing the following things and they
• 50:22
  are reasonably sensible things. Like maybe you should care about
• 50:25
  who has access to your data. And here's a technique for
• 50:28
  making sure you don't disadvantage this subgroup of a
• 50:32
  larger set as opposed to this other.
• 50:34
  So group, yada yada yada basically just what would happen
• 50:38
  if you took this seriously kind of questions, right?
• 50:41
  But there's a huge huge amount to talk about here with respect
• 50:47
  to the to the larger question. Of course I think I'll say I
• 50:52
  mean I have no crystal ball here whatsoever but I think what I'll
• 50:55
  say is, I see this kind of regional eyes.
• 50:58
  In the world, I see like in the writing of this book, the
• 51:04
  Chinese government has their statement about how a I should
• 51:09
  be used inside companies and the European Union has not just the
• 51:13
  eat GDP or but also their guidelines about how to be used
• 51:17
  Etc. I see more regulation, kind of
• 51:22
  happening and I see the conflict if that's the right word.
• 51:27
  Word between public policy and machine learning generally and I
• 51:33
  suppose technical development generally as as being very
• 51:37
  definitely, a thing, which is going to keep happening for
• 51:41
  years and years yet. And that gap between the
• 51:46
  conventional kind of public understanding of things,
• 51:50
  particularly in government and politicians, who often have this
• 51:54
  famous background in, you know, for philosophy politics.
• 51:57
  Economics etcetera. Those decision makers aren't
• 52:02
  equipped with the Nuance to understand.
• 52:05
  When somebody says this ml stuff is Magic pixie dust.
• 52:08
  Can pixie dust and you can sprinkle on and everything, and
• 52:11
  it just makes everything better and also it makes everything
• 52:14
  infinitely terrible. Like, those two messages are
• 52:17
  clearly discernible in public, kind of discussions about this
• 52:21
  thing. And my concern is that we will
• 52:25
  regionalize just as A factor of our cultural inertia and the
• 52:31
  direction we're going in in the various different continents,
• 52:34
  right? And we won't ever close that gap
• 52:37
  between the understanding of data science.
• 52:42
  And so on, so forth and key decision makers.
• 52:45
  And I think that that is a gap. I got grob increasingly
• 52:48
  concerned about it. That's it.
• 52:51
  Say, jokes are great time. So, it's just interesting, you
• 52:54
  say that now because like I see what you're talking.
• 52:57
  About. But in some ways, I perceive the
• 52:59
  gap between technologists and policy as as less than it was
• 53:08
  during the early days of data regulation.
• 53:11
  Like, during the early days of data regulation, it felt like
• 53:13
  people were like Doug in fighting tooth and nail but what
• 53:18
  I see on the AI stuff as I see a lot more constructive
• 53:21
  engagement. I see people saying like hold
• 53:23
  on, I get like, I totally understand why you want to do
• 53:26
  some of these things. Can we talk about the details a
• 53:29
  little bit more? So maybe I'm wrong.
• 53:31
  Maybe like, maybe that'll blow up in the same way, but at least
• 53:35
  in the US tech companies and the proposed European oversight for
• 53:41
  AI. I'm seeing a lot more
• 53:42
  constructive. Like let's try to understand
• 53:44
  what you're trying to accomplish and why don't you try to
• 53:47
  understand what constraints we have and how flexible we can be.
• 53:50
  And maybe we can work something out, but maybe I'm wrong.
• 53:54
  I don't think you're wrong. But I do think that the
• 53:56
  separation is The political layer and decision-making and
• 54:00
  not necessarily the administrative layer.
• 54:03
  So, I see the administrative layers of government Staffing up
• 54:06
  with AI regulation, units, and people whose job it is to know
• 54:09
  all about this stuff and there will be local competence is the
• 54:13
  political layer connected to that, in the way, it needs to
• 54:15
  be, like not in all regions in like all of the time, shall we?
• 54:20
  Say, you guys bring up some really interesting points.
• 54:23
  And something that I thought about our question asked myself
• 54:26
  is, how do we prepare the Next generation of practitioners to
• 54:29
  not have this as like an afterthought.
• 54:31
  I feel like with with with tech that's what it's always felt.
• 54:34
  Like it's just something you got to think about after the fact
• 54:37
  but the real focus is on the tech and the cool stuff.
• 54:40
  Do you feel like it maybe in typical computer science, you
• 54:44
  know programs there needs to be a classroom like the philosophy
• 54:47
  of Technology, the philosophy Fai, I sometimes feel like that
• 54:51
  could help that couldn't help but it also feels related to
• 54:53
  like maybe something maybe we could learn from the past, like,
• 54:56
  in the past, you know? The public had a certain
• 54:59
  perception towards a piece of technology and there was a role
• 55:02
  of government and in shaping that or the technologists also
• 55:05
  shaping that it feels like they're all related.
• 55:07
  Like the perception of things are competence is also related
• 55:11
  to what practitioners think about it.
• 55:13
  Do they value thinking about these things at the right now?
• 55:16
  It feels like they don't, it's an afterthought only when
• 55:18
  they're in trouble or when they're under pressure.
• 55:20
  Does it become really relevant? So, yeah, I just, I don't know
• 55:24
  if that's like the solution, you know.
• 55:26
  Because, you know, we talked about philosophy.
• 55:27
  Sphero that does help you think more broadly think cogently
• 55:30
  about things. How can we do that in this
• 55:32
  space? When, you know, engineer is just
• 55:35
  like to focus on what like they like to focus on.
• 55:37
  You know, it's very hard to kind.
• 55:38
  It's like herding cats. How do you do that?
• 55:40
  How do you bridge that Gap? So I do think I have a thought
• 55:46
  here, and maybe you do too. Now, like I think there has
• 55:49
  there is a very, very long tradition of Engineers being
• 55:53
  responsible for the uses of their creation and this goes
• 55:58
  back generations and generations.
• 56:00
  And like if we think about relatively recent examples, like
• 56:04
  no one thinks the engineers at Volkswagen and Audi are like not
• 56:08
  culpable for the cheating of the emissions.
• 56:11
  And the amount of pollution that they caused in particular, in
• 56:14
  Europe by programming, the computers to recognize when they
• 56:17
  were being tested for emissions at the emissions, testing places
• 56:20
  using the GPS and the specific greatly we know that and as
• 56:24
  Engineers we say like oh yeah you knew what you were doing and
• 56:27
  you were responsible for this. No I know like one of the things
• 56:31
  that happens with some of these creations is the consequences
• 56:34
  are a little bit distant, so if I have a model that learns from
• 56:40
  biased humans, And then the model is biased, is that my
• 56:44
  fault or Humanities faults right?
• 56:46
  And what I think what I'm starting to see is like, a lot
• 56:50
  of organizations are saying, like, well, we'll give you the
• 56:53
  tools to detect when you've reproduced by us.
• 56:57
  And then, then it is your fault because now you can detect like
• 57:01
  did you train a sexist model. If you train to sexist model and
• 57:05
  we give you the tools to detect that it was sexist and you
• 57:07
  launch it anyway that you launched a sexist model and so
• 57:11
  AI in the book. One of the things we did that
• 57:14
  I'm super appreciative of as we have this whole section on
• 57:16
  privacy and ethics like an entire chapter written by an
• 57:20
  independent expert in that which is fantastic.
• 57:22
  But we incorporated those concerns that you every single
• 57:26
  other chapter because you don't, you don't get to like isolate
• 57:29
  that and do to your point David. It's not just a separate class.
• 57:32
  It has to be in everything you do.
• 57:34
  Sorry. Now, did you have any thoughts
• 57:35
  on that? Yeah, I suppose I'd say two
• 57:38
  things. The first one is like as you say
• 57:41
  there's a Long tradition of scientists kind of struggling
• 57:44
  with the consequences of their actions and it's from everything
• 57:47
  from Oppenheimer and the atomic bomb all the way through to like
• 57:53
  a model that somebody pushed yesterday.
• 57:56
  Possibly with slightly different consequences for scale, guns
• 57:59
  concerts. But the thing I'd say is I kept
• 58:04
  coming back to your suggestion as the possessor of a Humanities
• 58:08
  and a computer science degree or whatever.
• 58:11
  I would say I would like to say yes sure we could we should
• 58:16
  expose two people to these complicated ideas of what the
• 58:21
  good means and how it should be defined and so on so forth.
• 58:25
  But actually is that going to be a necessary kind of improvement
• 58:30
  like a monotonic Improvement and what we see today I'm Convinced
• 58:34
  that I think the thing that matters is people see advantage
• 58:37
  in doing this Yeti thing, right? And they do that because there
• 58:43
  is some advantage to it, how can we make that not be an
• 58:47
  advantage? Well today the answer to that is
• 58:50
  regulation or jail or finding, or any one of a series of kind
• 58:56
  of activities like that. Shall we say incentives?
• 58:59
  Like that? I think we can get a Some degree
• 59:04
  by relying on character and culture and those are both very
• 59:09
  strong forces, you know, different Scopes Etc.
• 59:12
  But ultimately if we're trying to say people you should not use
• 59:16
  this thing in this way. Someone else is going to have to
• 59:19
  say that with a loud voice. Yes, so good.
• 59:25
  Well said. So I could sit here talking to
• 59:30
  you guys for another couple hours.
• 59:31
  Sadly, you are very important people and you also have lives
• 59:35
  to lead and we have got to cut it but this has been super,
• 59:42
  super cool. I highly encourage anyone out
• 59:44
  there. If you have not read the book go
• 59:48
  out and get it, you can go on to O'Reilly right now and check it
• 59:51
  out. And the final push, I think is
• 59:56
  for what? When's it going to come out
• 59:58
  officially, so it should be out electronically before you can
• 1:00:01
  edit this, they're talking about pushing it later this week or
• 1:00:04
  early next week. Perfect.
• 1:00:05
  Alright. So by the time you hear this,
• 1:00:07
  unless you guys are like editing, listen, the background
• 1:00:10
  and pushing it later to bed. By the time people read this,
• 1:00:12
  they can get the book in print within a week or two and like
• 1:00:16
  certainly electronically incredible and we're planning on
• 1:00:20
  doing a reading. Club in the MLS community.
• 1:00:23
  So if you go to reading group go and check that out and we'll all
• 1:00:27
  be reading it together and then Todd and Nile are both in the
• 1:00:31
  community and we're so thankful to have you guys in there.
• 1:00:33
  There it's been super cool. So if you have questions for
• 1:00:37
  them just hit him up in slack. Not too many questions because
• 1:00:41
  like I said they got they got things to do also but the last
• 1:00:46
  things I was going to ask you guys are the best ways to get a
• 1:00:52
  hold of. You is it just slack?
• 1:00:53
  Do you prefer Twitter LinkedIn? Let's I hate people.
• 1:00:59
  I don't like, like, I don't know, like Nile and I are both
• 1:01:04
  on Twitter, both on LinkedIn. I think LinkedIn has
• 1:01:07
  increasingly had, I think because of chips, book, chick
• 1:01:12
  wins book, which is fantastic. You should get her on if you
• 1:01:15
  have it already but like I've seen an increasing and some of
• 1:01:18
  the work you've been doing Demetrius over there.
• 1:01:20
  Like I've seen an increasing amount of ml Ops content on
• 1:01:23
  LinkedIn. That's been really useful, so I
• 1:01:25
  think that's a nice focused way there.
• 1:01:28
  I'm on Twitter but it's been a little bit more focused on
• 1:01:30
  LinkedIn. I think.
• 1:01:32
  Absolutely not. I think I'd say if you come to
• 1:01:35
  me on LinkedIn you will contend with the number of people who
• 1:01:39
  want to hire me. If you come to me on Twitter,
• 1:01:42
  you'll contend with the number of people who want to be
• 1:01:44
  sarcastic at me. If you come to me on slack,
• 1:01:48
  that's probably faster, but You know, there's content with your
• 1:01:54
  co-workers exactly who have to continue it.
• 1:01:57
  Well fellas, this was awesome. Thanks again for doing this and
• 1:02:00
  we will hopefully have you back on here sooner than another year
• 1:02:04
  because this is just super cool to talk to you all.
• 1:02:07
  Yeah yeah we're not writing any more books.
• 1:02:10
  Forget it looks do not, do not awesome guys.
• 1:02:15
  Thank you so much. It was such a pleasure.
• 1:02:18
  Thanks so much.