Based in Sydney, Australia, Foundry is a blog by Rebecca Thao. Her posts explore modern architecture through photos and quotes by influential architects, engineers, and artists.

Using Reflection Podcast - Open Ended Problems

Using Reflection Podcast - Open Ended Problems

Max Sklar
Podcast

The audio is no longer available, but the transcript and notes are pasted below!

This is my talk with Mark Weiss, host of the podcast Using Reflection, and a former coworker of mine at Wireless Generation (now Amplify). The interview was recorded way back in April 2018 when I was just launching The Local Maximum (remember the 10 episode goal??)

Listen and get the show notes as usingreflection.com

Show notes include links to all my related works and episodes, plus a full transcript!

From Mark Weiss:

Happy to announce a new episode of my podcast "Using Reflection," the podcast where engineers tell their stories! (Listen here: https://usngr.com/2H9jmjo) The last decade has seen a sea change in how we use machine learning, and Max Sklar has had a rare vantage point to watch the rise, having spent almost 8 years at Foursquare starting in 2011. He started before the term Machine Learning Engineer was a thing (!) and helped the company build on its innovative model combining gamification and location. More recently Max launched a podcast, "The Local Maximum," which mixes interviews with professionals, authors and thought leaders with Max's own opinion pieces. His goal is to join the dialogue of ideas and inspire others to do the same. Since the interview, Max has moved on and is some sense putting both strands together working on recommendation for the podcast app Luminary Media. This is a fun and insightful conversation, full of front-row insight and buoyed by Max's enthusiasm for open-ended problems in ML, and in life beyond the screen. Enjoy!

About Using Reflection

Using Reflection is a podcast where Mark Weiss talks to engineers about what they’ve learned in their careers. The podcast is of interest to me for several reasons. First of all, it was great to participate in it and share my story. But it’s also a place to learn from others with varying perspectives, and I’d like to have Mark on The Local Maximum soon to - among other things - learn about interviewing engineers.

One to check out it Mark’s conversation with the first person who hired me - Aaron Boyd. I particularly enjoyed his insights into working productively with engineers and designers who have different ways of looking at things.
http://www.usingreflection.com/Aaron_Boyd_-_I_Dont_Care_if_Youre_a_Genius/#debac944

Full Show Notes from UsingReflection.com

Show Notes

Topics: Alpha Go, Data Engineering, Data Science, Information Systems, Machine Learning, Local Search, Location-Based Apps, Monetization, Natural Language Processing, Objective Functions, Optimization, Podcasting, Self-driving Cars, Sentiment Analysis, Social Apps

Companies and Organizations

Introduction

NOTE: This episode was recorded in April 2018. Since then Max has moved on from Foursquare to join Luminary Media as a Machine Learning Engineer.

After working for a few years as a software engineer, Max Sklar found himself exposed to and fascinated by more open-ended problems while studying machine learning and data mining in the MS Information Systems program at NYU. He created a location-based app StickyMap so he and friends could put “markers” naming locations onto Google Maps. This was fun but the app didn’t take off. After a stint working on local search at Yodle, he discovered Foursquare, recognized their gamified user participation would drive their map app to large scale, and talked his way into a job there as a machine learning engineer. Max spent the next 7 years doing machine learning and helping to build products at the leading edge of the industry. More recently he launched the podcast “The Local Maximum” as a forum for sharing his own opinions about the impact of AI and machine learning and for interviewing guests working in the field. It’s a treat to talk to someone who can actually claim to be an ML veteran, and we dig into the challenges of and differences between data science, machine learning engineering and data engineering, and talk about the crucial role of humans in ML-driven products. It’s equally fun to talk to a fellow podcaster about his goal of using his show to foster a dialogue with his audience and to push them to explore ideas outside their “local maximum.”

Guest Bio

Max Sklar is the host of The Local Maximum Podcast, a weekly podcast with interviews and analysis covering ideas in AI, emerging technology, and current events. The most recent episodes and the archive can be found at localmaxradio.com.

The bulk of his work as a machine learning engineer was at Foursquare, where his worked there included building Foursquare City Guide’s critically acclaimed 10-point venue rating system and the Marsbot app. More recently he led the development of a causality model for Foursquare’s Ad Attribution product.

He currently works at Luminary Media where he is transferring these skills to build a recommendation engine for podcasts.

Links

“The Local Maximum” Episodes Related to Topics Covered in this Episode

Transcript

Host
Hello and welcome to “Using Reflection,” a podcast about humans engineering. And we’re joined today by Max Sklar and Max is a machine learning engineer at Foursquare. Why don’t you go ahead and introduce yourself Max and tell us a little about your background and we’ll get into the conversation. We’re gonna have a great conversation today about your career in machine learning and how things have changed and also get into your awesome podcast, “The Local Maximum” and it should be a lot of fun. Take it away.

Max Sklar
Great. Yeah. Hi Mark. Thanks for having me on. I am a machine learning engineer at Foursquare about coming up on seven years now, which is like crazy. And you’re right, I have a podcast and this is my first time doing an audio recording on someone else’s show since I started the podcast. So very excited about that. I’m already finding a lot of things due to this podcast. Like I found your podcast. I like you have a lot of interesting conversations. It’s really great.

Host
So, you know, one reason I thought it was interesting to talk to you was you, you’ve been in the machine learning space and had that title and been doing that kind of engineering work while the entire field essentially blew up around us, while the term data scientist was invented, while big data infrastructure, uh, became a huge part of cloud platform offerings and moved into the business mainstream. All of these trends, you know, in these last seven years, you could really say. So I was interested to start with, you know, your perspective on that evolution and maybe you could start with your personal story as we were talking about before we got on mic, just how you moved in that direction, and then maybe you could talk about what you’ve seen changes in the field around you.

Max Sklar
Let’s start with how I got into machine learning. I, you know, uh, it was, well, I did take machine learning as an Undergrad and that was, that was fun. I was into it, but I wasn’t able, I, you know, I wasn’t able to find any particular work in that. I guess what we were doing at Wireless Generation – and I think we should probably mention that we both worked together at Wireless Generation way back in probably I want to say 2008 – so that was a long time ago. Yeah. And so it was, it kind of seems like we might be going in that direction at Wireless Generation where we were doing data analysis in order to do smart instruction plans, but we never kind of got there. At least I personally never got the chance to work on that stuff.

Host
Can we just interrupt for a second and maybe frame that by saying what kinds of, what kinds of products Wireless was making and who the audience was so people have a little context there.

Max Sklar
Wireless Generation was, oh my God, it’s 10 years ago, Mark. I’m trying, we were doing the front end a lot for assessments, like reading and math assessments. Mostly for a young children, if I remember correctly, it was like K [kindergarten] through three [third grade] for the most part. We had other grades too. And the point was to do those assessments. We had a good front end for it and the Palm Pilot. And then the back end would do some data analysis and you know, would tell teachers how to do instructions with the kids in order to kind of maximize their effectiveness as teachers. So that sounded really good, but it wasn’t, you know, it wasn’t very large scale machine learning type stuff. It was more like, you know, kind of handmade rules type stuff that we’re already given to us by the sort of, by the education industry or, you know, if I remember correctly, like the testing companies and things like that.

Host
So you’d have these domain experts who were the ones formulating the rules and then the data analysis was identifying patterns that match the rules.

Max Sklar
Yes. Yeah. And so it wasn’t, you know, there was nothing, there was nothing to do on the engineering and product building side other than kind of taking those specs and executing on, which there’s some interesting problems there, but it wasn’t the kind of open-ended problem that I wanted to work on. And I sort of didn’t know at the time. But when I got to Grad school and I started learning about more about machine learning and natural language processing, I was like, oh, these are the problems that I think are going to capture my interest because they’re so open-ended and they’re so, you know, you don’t know. You don’t know where. You have to kind of figure out where the stopping point is. You have to kind of figure out what it is you’re going to get the machine to learn.

And I just always liked the idea of, you know, rather than having us program the computer and trying to figure out every single rule that comes up, just kind of figure out a way to have the computer program itself. So that way we can kind of like sit back and have it have to do its work. And it just seems to me in the long run that’s going to be, that’s going to get us to a place where we’re much more effective. And it turns out that in machine learning, there’s a lot of stuff that, you know, humans have to do all the time. It’s not that, um, you know, it’s not that computers just go off on their own and figure it all out. Although I have built some systems that do get smarter over time without much intervention. And that’s always really cool. That’s always when I feel like, you know, wow. I built this, you know, a few years ago and now it’s still working and it’s getting smarter because it has more data. I’m always just like amazed when that happens.

Host
And is that related to the open-ended idea? Are those two things connected?

Max Sklar
Yeah, I mean, the open-ended idea is that it’s not just, you know, hey, you have a client or you know a company that wants a specific thing done and then you have to execute on that specific thing. It’s kind of an open-ended research question. You know, for example, some of the things we do on an ad just to bring in some of the things we started doing at Foursquare … what’s the sentiment of a tip? Is someone’s text and it’s like a tip and a tip at Foursquare is like two or three sentences that tell you whether a place is good or not. Little reviews that people leave, and can we automatically detect whether it’s positive or negative. You know, I don’t, nobody’s coming at you with the rules. It’s just like what, how, how accurate can you make it? And also like, what’s the, you know, what’s your metric for how well you’re doing?

Host
Right. And even that’s … what’s interesting about what you’re saying there as part of the openendedness is, you know, as engineers, we need to make things precise and quantitative, right? And that’s like how do you make that qualitative judgement quantitative enough to even define success engineered to a goal, and then measure whether you’re reaching the goal?

Max Sklar
Yeah, yeah. I mean, it’s oftentimes I think there’s an arms race between the machine learning algorithms that figure out how to optimize for a certain objective function. And then there’s kind of working on the objective function itself and what that should be doing because, and this is probably true when you are managing an organization as well. I think if you kind of set very specific goals for people to hit, you know, let’s say, you know, you say this quarter we’re going to work on maximizing our revenue, let’s say. And then, you know, you have all your rewards and bonuses based around that particular goal. Well, you might end up overdoing it. You might end up having people you know do things that are good for the short term but bad for the long term.

And the analogy I’m drawing is it’s the same thing for the machine. If you have a certain objective function, sometimes it’ll get too smart that it does well on the objective function, but it doesn’t do very well on what you’re trying to accomplish. And so the objective function sometimes gets more and more complicated. A good example like that recently that we came up with at Foursquare is, you know, we have our Pilgrim technology, which is our “snap to place.” So that takes the data from your phone and tries to figure out exactly which venue you’re in. So to give an example, like, are you in the Starbucks? Are you in, are you at the Dunkin Donuts? Are you at the mall? Well there’s a certain issue with the mall where it was always, you know, it would snap people to the mall, but it was very bad at telling you which place exactly you are inside the mall. Even though we have those places in our database just because it was always more accurate when it just says mall. So it, and we wanted the data and the places inside the mall. And so that was a change. They had to make a change on that team and you have to make that change by, you know, fiddling with the objective function, not with the algorithm itself.

Host
I see. So a large part of the design then and the human aspect of this is trying to capture the goal, which is human-defined in the objective function, which is what the system is learning toward and optimizing toward.

Max Sklar
Yeah. Yeah. It would be nice if you had one goal to work towards and then you are, we’re constantly researching the algorithm and the data and the features to get to that goal. But one thing you learn is life is never like that.

Host
Right. And in fact, as you were talking about this, I was just thinking how reductionist and challenging it is to have just one objective function. You know how like you were talking about managing an organization, well, your example was the objective was motivating too much behavior towards short term gain, but of course like that’s an example where very quickly you see there are multiple goals and they have multiple timeframes. And how do you capture that in one objective function? The nice thing about human interaction is you could set more than one goal for people and humans can think about that and, and work with that.

Max Sklar
Yeah. And oftentimes there’s a lot unsaid with humans and human learning that they pick up on that. But that’s why, you know, that’s why experience is valued and that’s why, you know, human intelligence is still valued. Thank God

Host
We’ve got another minute before obsolescence.

Max Sklar
Yeah, yeah, sure. We do. We have plenty of time.

Host
If we want to backtrack and sort of, you’re at Wireless and you’re noticing that there’s this opportunity to attack these more open-ended problems with data. You go to Grad school, you are exposed to those problems and you’re, you know, it sounds like you were engaged by those ideas. So then how did you keep moving in that direction?

Max Sklar
Well, a few things came together at once. So I was, the grad school program I was in was the Information Systems program at NYU. What it was was just a half the classes were from the Computer Science department and half the classes were from the Business school. So it was kinda nice. I got to exercise both parts of my brain. I didn’t want to do a program that was 100% Computer Science for three years because I already did that as an undergrad. And I already had a job as an engineer for several years before that. So it was kind of like, what’s the point? Even that summer, the summer that I was in grad school, I did a design and research internship at Columbia because, you know, I was like, I don’t want to do coding. I’m going to do years of coding. You know, I’ve done years of coding.

But I think when I was there, I took some classes on data mining and machine learning, both from a business school perspective and from a, and from an engineering perspective. So I took, you know, for example, Yan Lecun’s class at NYU on machine learning. But before that I took the data mining class at Stern to get an idea of, okay, you know, what are, how do people in the business school think about this? What are their, you know, what are their goals? And it was really great to kind of get both. And then at the same time I was sort of, I was learning about, okay, I had been doing a lot of, a lot in the field of local search before that. So when I was an undergrad, I had a website called stickymaps.com. It’s still up.

Host
That was a site where what people would do is it, it was, okay. The Google maps API had just come out and I was kind of thinking, what can I do with this? And one of the things you can do with this is you can have people place little stickers on the map, I guess we call them stickers now because Foursquare Swarm has stickers. But back then we called them markers, little icons on the map, and leave messages and that was a lot of fun. Uh, I was never able to figure out how to turn a business into that. But then I also worked on local search at a Yodle, which was a search engine marketing company that I did after Wireless Generation. And so I had a little bit of experience there and I had some interest there. And so, and I got with StickyMap I really got a kick out of building something like a consumer application that a lot of people were actually using. Not that a lot of people were using that, but you know, my friends were using it. So that was kind of exciting.

Max Sklar
I discovered Foursquare in the summer of 2010. Uh, the way I discovered, actually discovered it through a class that I was taking, it was called electronic communities. It was at Stern and I was assigned to work on a local business in Nolita on how to get their word out on, on the Internet, on social media. It was like a women’s beauty salon in Nolita. They did like the nail polish art. So I was like, I don’t really know too much about this, but I’m going to give it a, I’m going to give it a try.

You know, we’ve talked about Facebook, we talked about Yelp and I looked at Foursquare and I was like, oh, this sounds really interesting. Are these the people who are like competing to be mayor of, you know, the different locations around the city? Uh, I think I like overheard something about that in my internship. So I decided to look more into it. And as I looked into it, I was like, oh, this is what I should have been doing with StickyMap. You know, this is why, you know, Yodle’s product was not, was not as inspiring because it was, it was an enterprise product. It wasn’t really a product to be the world’s best, a local recommendation engine, but you know, what Dennis Crowley had done, and the team at the time had done with Foursquare was just really interesting. They made it a lot of fun and I was like, this is the way to go.

Um, and so I saw Dennis give a talk at NYU and I tried to talk to him in person, but he immediately got swarmed by a bunch of groupies asking questions. And so I just got his card and I emailed him later and it’s kind of embarrassing email because it’s about like, you know, it’s, it’s much longer than I think an email should be. But then he writes back and he, and one of his questions was, are you like a machine learning superstar by any chance? And I was like, okay, well, I guess this is, I’ve got to say yes, right. I’ve got to say yes.

Host
It sounds like the innovations you saw at Foursquare where he figured out relative to Yodle, right, was he figured out how to make this a consumer-facing rather than her a business product. And the gamification aspect made it appealing. So people were participating, which was kind of the product question you hadn’t been able to figure out with StickyMap. Right?

Max Sklar
Yeah, yeah, that’s exactly true. And I mean, Yodle wasn’t trying to be a consumer app and that was one of the things that was maybe a little bit frustrating to work on it because, you know, of course it was shut down a little bit after I left, maybe a couple of years after I left. But it was just like, you know, the purpose of it was to drive leads to their clients. So it was never going to be the, you know, one of the greatest recommender systems in the world. I hope I’m not throwing them under the bus too much, but it’s just, it wasn’t, the purpose of the product wasn’t to be the best recommender system, whereas Foursquare’s was. So I think with my experience on StickyMap and with Yodle, I was able to identify the Foursquare product more readily than maybe other people could.

Host
Right. And you could see the strengths and sort of the cleverness of the fit found in the solution they had found in the market and you’re like, yeah, this, I’ve been in this space and this is solving these problems. And it makes sense. And I think this can work. And early too. So for people who maybe didn’t use Foursquare in the early days, this is like what, seven, eight years ago now already and in its beginnings, it was this idea where you would go to these different places and be the first right and kind of just say, I’m the mayor of this place. And so there was this gamification aspect that drove usage. And then the second thing people are doing right from the beginning is sort of posting where they were and it was this sort of peer to peer network of all your contacts so they would know where you were. To facilitate social interaction. Is that an accurate description?

Max Sklar
Yeah, yeah, that is accurate. And we still do that. We still have our swarm app, which is the app that we have that does, you know, that’s the gamification aspect of it. That’s where people check into places, tell their where they are and they … there’s kind of games and stickers and mayorships and stuff. And the, one of the things that I added as like a hack day project a couple of years ago was you get extra points in Swarm if you actually created the venue and added to our venue database or if one of your friends did or one of your friends of friends did. And so the reason why I like that is because it encourages people to add places to our database and be first. So I’m always really excited from the first one to like a grand opening of a place and then I add it to the Foursquare database and then later on I see, oh, you know, a thousand people have checked it.

Well maybe I haven’t gotten anything that big, but maybe like 300 people have checked in there. Wow. And I was the one who created it. So I get kind of a kick out of that and then I kind of see and then sort of, I see sometimes my friends, you know, created the bagel shop down the street and I was like, ‘What? You made this place? No way.’ You know? So that’s always, it’s always a lot of fun. It kind of encouraged people to create the database. So it was sort of a fun way to get the power of the crowd, which we, was kind of, we talked a lot about back then, but it’s hard to actually get the crowd to work together in a productive way.

Host
Yeah. And it’s, in fact, as you were saying that, I was thinking the same thing. It’s hard to think of another Internet business that really you could argue solved the or, or figured out a way to tap into this crowd-sourcing idea very concretely earlier than that. Like they really did help sort of figure out that model, right?

Max Sklar
Yeah. I mean, one example is kind of, you know, Wikipedia maybe,

Host
Well, Wikipedia is a much more, it’s a much more narrow participation model than like the Foursquare model is actually the social graph model. You scale everyone, you get everyone to participate, you get this entire social graph to participate, and it keeps scaling. Uh, and they’re, and they’re not just talking to each other, but they’re building up the value of the data that they’re providing you. It is actually essentially feeding back to them as the value they’re getting from the service. So this was kind of a really good fit for you in terms of you found a place that saw the world and you could say the same way you did where they were looking at this open-ended problem and had to actually devised a product which could itself help them evolve toward solutions, toward features that were valuable to the customers, toward a circular interaction with the customers. And really it’s become a platform, right. For multiple products. So, that’s really interesting and I can see why you stayed there all this time. You know, maybe this is a good way to segue into kind of how you, you could, maybe we can talk about it through some specific examples of things you’ve worked on there, but how your approach to machine learning has changed, maybe what you’ve learned. And then back to our original question, how the world has changed around you, how the field has changed around you.

Max Sklar
Yeah, well, definitely in terms of just having experience building and shipping machine learning models. Um, you know, one of the things that I have to be careful about, which is probably the same as general software engineers too, but for some reason I feel like it’s a, it’s an even bigger problem when you’re working on machine learning models is you know, these kind of what we call rabbit holes that you could fall into where something’s not working quite right. And then, you know, you spend months and months trying to fix it and you’re trying to fix this one thing and you sort of lose track of the big picture. And it’s not something that, it’s not something that’s easy to identify when it’s happening or it’s not as easy to identify as you as you would think it would be.

It’s like, well, wasting three months of time. That sounds like, it sounds like it would be easy to identify. It’s usually in hindsight it’s very easy, but it’s, there’s a lot of, you know, well, I just want to add one more feature to my model or I just want to, you know, make it a little more sophisticated. So it could capture more and which are all good things to do, but you do need to zoom out from time to time and figure out, okay, what is the problem that I’m trying to solve? We constantly asked that and then what’s the best thing I need to do next to solve that problem? So that’s just something that comes with more experience and screwing up a bunch. And I think it’s something that’s true for software products in general, but …

Host
Right. And, and this, I was going to tie this back to two things you said earlier. So one is the open-endedness of the problem and then the other is the objective function, the definition of a correct objective function or a useful one.

Max Sklar
And then being able to ship it. Like oftentimes you get a model that you really like and then you want to deploy it in production. It’s like, okay, that takes a lot of other work. A lot more work than, than you would think.

Host
Yeah, that’s been a lot of emphasis in the field and in the service cloud services the last couple of years is, you know, these solutions for production scale and production deployment and, and the, the idea of, you know, how do you train against huge datasets and then make predictions very quickly against the model in production. These are very different use cases, right? Kind of coming from the same source. So it’s a really, it’s a lot of challenges there. Right?

Max Sklar
Right. I want to get something deployed nice and early because you know, then you know what the problems are going to be in terms of deploying it.

Host
So, interesting. So you feel like you’ve gained experience in being able to, define the problem more clearly and define objective functions more clearly and be … and, define what is, what is the correct place to sort of, what is the correct kind of quality level set of features to reach, to be able to ship as early as possible. Those are things you feel you’ve gotten better at through experience. That completely makes sense.

Max Sklar
Absolutely. Yeah. And I, it’s interesting now that you mentioned it, you know, defining the objective function is one of those things that, if I were to make a prediction about the future of AI, when people think, ‘Oh, AI is just going to take over and there’ll be no jobs for humans,’ I think defining the objective function, at least at the highest level, is always going to be, the human, you know, that’s the human job.

Host
Right. And, and if you look at the extremely successful current examples, and then now we’re absolutely in a danger zone of being futurists and looking stupid later. And me speaking from …

Max Sklar
I do that on my show all the time. So I mean that’s one of the things that I do a lot and, but it’s, well we can talk about it later, but it’s, it’s sort of one of the things that I want to try to do, make predictions. And one of the great things is you can go back later and see where you went wrong and then hopefully do better in the next round.

Host
If you look at the current examples there, like Alpha Go and a self-driving cars let’s say, the objective function is … first of all, the objective function in all games is predefined and, in a certain sense, trivial, right? And in, in self driving cars, there’s a lot of complexity, but the higher level objective function is essentially trivial also. Everyone can describe what happens when they get in a car and go from one place to another, and why they’re doing it.

Max Sklar
You know, the, the types of risks that you can and can’t take is a very open question. And I mean, we still need humans to solve all the other problems too. That’s not automated yet either, but, but the objective function I think is it would be the last to go.

Host
Right. I see. Right. That’s sort of the codification of the human goal that the system’s trying to achieve.

Max Sklar
Right. And I don’t think, yeah, and human goals are very complicated. It’s, again, like I said, it’s an arms race. So as the machines get more complicated, our goals are going to have to get more complicated.

Host
Maybe we could also talk a little about, uh, was some changes in the field at large that have happened around you and you know, your perspective on those. We were talking before we got started about your title as a machine learning engineer relative to data scientist and that term has arisen in the last few years of what you may be see are the differences between the two. You know, some larger changes you’ve seen in the field. Maybe it would be an interesting way to go on this from your point of view as a practitioner for seven years and having had the same title and having have this title of machine learning engineer, all of which are relatively unique attributes from which you can speak

Max Sklar
Right, machine learning engineer really isn’t my official title. It was just when I joined I wanted to have a good way to, something to put on my LinkedIn page that would describe what, accurately described what I was doing in relation to what I was doing before. So it was like, okay, I’m still an engineer. I’m still doing the, I’m still doing the server side work. I can still ship my code, but I am sort of specializing a little bit more and building machine learning models and you know, solving all of the types of problems that go along with that. So that, that was just that, I don’t know if it was a description that was popular back in 2011 when I added it. I kind of just made it up, but now I see it all over the place.

So either it, it was there or it was just, it’s, it’s gone from sort of made up to real now. But it was really just descriptive and we were hiring for a data scientist at the time. And so what, what is the difference between the two then? Right. I mean it’s, it’s always, I don’t think, I think every company does it differently, which is really hard because when you post something for data scientists, you’re going to get a lot of different types of people. So there’s sort of the data analyst role, which is someone who doesn’t necessarily go into the, your backend server code all the time. But, um, what they do is they take a lot of your data sources and they might go like in the middle of your data pipeline and take some of your data sources and they run analysis on that offline.

Um, and so, you know, a lot of them will use R and Python as well. And they’re oftentimes really important people to have at your company. We have a bunch of them and they sort of, they can diagnose problems really well and they’re actually more like scientists in the sense that, you know, they’ll, they’ll go through the scientific method, you know, they’ll come up with a hypothesis. They’ll test it by writing some queries on the code and then they’ll, they’ll see if that’s a, that’s right or wrong. That’s, that’s sort of one role. And then you have my role, which is more of a machine learning engineer where you’re actually, you know, like I said, building, building the ML models and, uh, and deploying them. Um, and the other one is sort of a data engineer who just takes the entire data pipeline and make sure that the pipeline is a whole works, kind of deploys our technology, whether it’s, you know, MapReduce or Scalding or Spark and, you know, they’re doing a lot of server engineering and a lot of, I mean, it’s not like, you know, I’ll build the, I’ll build the pipeline and write the code for it. But there, there are a lot of people who were working on that stuff who don’t necessarily get into the machine learning stuff, but they understand kind of how much, how much memory it’s going to take and how it’s deployed and where it fits into the whole system. So, I mean, all of those people, all of those groups are in roles that are kind of similar and sometimes you have to do all three jobs. But, um, I think that that the role of data scientist, if you’re hiring for data scientist or if you’re applying for a job as a data scientist, you have to ask very clearly. You know, what the mix of those tasks is going to be, and try to find out what the company actually needs.

Host
Right. So it sounds like you’re saying it’s multidisciplinary and different companies kind of have a different idea about this or different emphasis. Yeah. Right. That makes sense. Because you need all three of these aspects.

Max Sklar
Yeah. And they’re, they’re probably more than three. I probably shouldn’t do … that was just three that came to mind, but …

Host
No but then it’s like, okay, so there’s like qualitative understanding of the data, right? And a set of goals and business goals you want to achieve, by, from, from the insights you can get from the data, right? Uh, what are the features and then what predictive value do they have? What, and then of course, what do we want to predict? So there’s business goals there, right?

Max Sklar
Yeah.

Host
And, and, and an understanding of the data as an asset to achieve those goals. That’s kind of the first piece. You were describing the second piece of someone with, uh, the mathematical and algorithmic expertise and machine learning experience you were describing that you’ve gained to, you know, turn that into an objective function and a learning system, right?

Max Sklar
Yep.

Host
And third piece you were describing earlier and then return to again is how do we make this repeatably deployable, scalably deployable and run, run reliably and run every day and being able to add new pipelines. How do we integrate all the data we need at scale and how do we make this something that runs really smoothly for everyone else so they don’t have to care about that? Right, right. So maybe this is a good chance to, uh, switch gears a little and we can talk about your podcast. So maybe just start with what motivated you to start to do that? What, what did you think sort of you had to say or add to the conversation? You know, what interested you about pursuing that?

Max Sklar
I have given a lot of talks and kind of done some instructional videos and conference presentations in the past. But it was always, it always had to be in front of a very specific audience and it always had to be, you know, about a specific topic. And I wanted to put something out there where it was a little more free to explore, you know, different issues that are, I am passionate about or interested in and to kind of do it weekly and build kind of an audience that is really interested in new ideas, and I want it to be kind of interactive. So I want to have an audience where I can put out ideas and get lots of feedback. Go back and forth and kind of just to join the marketplace of ideas, I guess.

And I think, you know, a lot of good things are going to come out of it and I think me and my audience are going to learn a ton of stuff. I think that, from each other, I think we’re going to figure out how to express our ideas better. I know I am and my audience and my, my group of guests. And I hope that in the future it sort of inspires people to, you know, to execute on some of, to execute on some of the ideas that they’ve, that they got from listening to the show. And I’m trying to interview people who are building stuff that is, you know, either interesting or some entrepreneurs or someone who looks at a problem in a little different way that you wouldn’t think about otherwise. And you know, as the title of the show is “The Local Maximum,” which kind of suggests that you want to get out of your funk, you want to, you know, you want to find something new, you want to find the better hill to climb. So I’m, I’m trying to build kind of a show in a network of, of audience members and listeners that we’ll do that.

Host
And is the focus on the same kinds of open-ended machine learning problems that, you know, have you been interested in in your career?

Max Sklar
Yeah, well, I, this is, I feel like I can leverage some of the stuff that I’ve worked on, you know, in the last decade or so. And so I have, I mean, I kind of compare it to the show that I had as an undergrad when I was at Yale. I had a radio show for two and a half years, like from 2004 to 2006. And I go back and listen to it sometimes and uh, cause I was pretty good. I had some, uh, I, it was pretty entertaining. That was called “Max and the Wiz.” Um, and it was all, a lot of it was, current events related, whether it was national stories or just stuff that was going on, on campus and in the school. And one part that we did was we would do a science and technology news segment from time to time.

And so that I kind of want to go back to and see what we were talking about then because I’m pretty sure I covered self-driving cars in 2005 and so it’d be interesting to see what we said. I’ve always been much more natural in the audio medium and so this was just a good, a good way to start that. I had so much fun doing it. And so that’s just how I knew that this would be a good way to, you know, to, to build those ideas and get that content out there that I want to.

Host
You had a sort of specific angle or area that you felt could add to the conversation.

Max Sklar
Yeah, yeah. And that’s the difference between then and now. Then I was just kind of making it up as I went along, which was, which was fun. It was a lot of fun. But now I actually have, you know, a little more knowledge yet …

Host
You attract an audience in a sense because they have an idea of what this is and what they can expect.

Max Sklar
Yeah, yeah. And it’s coming back to the name, “The Local Maximum,” why I picked it. It’s kind of a triple, triple play on words there because you know, a local maximum is something that we talk about in machine learning and data science. That’s when, you know, you have a model that is trying to get better and better and then it reaches the top of a hill where any direction that it goes in, it gets worse. So it kind of stops and says, this is, this is as good as I can get. I’ve converged, you know, but then there’s a, it’s actually you’re in a local maximum. It actually has to get worse before it gets better. It has to explore different ideas and that’s a problem that comes up in machine learning. But it’s also a problem that comes up in product design and in entrepreneurship and in, in marketing and all these different areas you get to a local maximum.

And the way to get out of it is you kind of have to stop what you’re doing. Stop climbing that hill and just pursue new ideas and trying to climb a new hill. And that’s you know, that’s what I’m hoping to encourage people to do by listening to all these interesting people that come on my program and to throw out ideas that are maybe partially baked that I’ve been thinking about for a while, but I haven’t shared publicly yet. I feel like I have some good ideas, but I don’t verbalize them for years and years. I just kind of let them sit in my brain and I don’t verbalize them because, you know, I feel like they’re kind of half baked or no one’s going to get it. And this has been a good opportunity to just, you know, go for it, spit it out, see what happens. And I feel I want more people to do that.

Tech 2025 Podcast - Machine Learning Superstar

Tech 2025 Podcast - Machine Learning Superstar
Max on the Machine Yearning Podcast with Shane Mac

Max on the Machine Yearning Podcast with Shane Mac