E120: Lukas Biewald, Founder of CrowdFlower – Interview

July 27, 2017

https://www.linkedin.com/in/lbiewald/

This interview is all about gathering and preparing data for advanced machine learning. It’s with Lukas Biewald. Lukas is the Founder and Chief Data Scientist at CrowdFlower. They allow companies to upload data like images, text, audio. For example, an autonomous vehicle company could send them images taken by their cars for labeling. The companies describe what they want done, and then CrowdFlower sends the job to the appropriate people on their platform to finish the job. It’s essentially a training data platform (or human in the loop platform where humans help prepare the data) for machine learning. They’ve raised $38 million dollars so far.

It was great hearing more about Lukas’ background and where he wants to take CrowdFlower.

Here are some other things we talk about:

-Hear about what Lukas likes to do in his free time.
-How does CrowdFlower help autonomous vehicle companies?
-How does their platform work?
-How does Crowdflower ensure quality labeling?
-For a machine learning project, what percent of time is spent on data prep vs. model selection/training?

 
 

Transcript

Dave Kruse: Hey everyone. Welcome to another episode of Flyover Labs. Today we get to talk to Lukas Biewald and Lukas is the Founder and Chief Data Scientist at CrowdFlower, and what they do is pretty cool. They allow companies and people to upload data like images, text and audio and describe what they want done and then CrowdFlower finds the right the people to do the job. And so it’s essentially a training data platform or as they call Human in the Loop Platform, where humans help prepare their data for machine learning. And they have raised $38 million so far, which isn’t too bad. So I am excited to learn about what they do and where they want to take it and also looking forward to hearing more about Lukas’s background. So Lukas, thanks for coming on the show today.

Lukas Biewald: Yeah, thank you for having me.

Dave Kruse: Definitely. And the audience didn’t hear this, but I messed up the intro for the second time in Flyover Labs history, so I had to redo it. So Lukas got to hear part of that twice, so I appreciate his patience.

Lukas Biewald: No, that’s a great overview.

Dave Kruse: Yeah, that’s true. All right, so before we get to talk about CrowdFlower, let’s just – can you tell us a little bit about your background?

Lukas Biewald: Sure, yeah. You know so I – when I was a kid I always wanted to be a programmer and I was super interested in artificial intelligence, yeah but there wasn’t a lot of researches around to learn about it, but I would…

Dave Kruse: At what age?

Lukas Biewald: I remember when…

Dave Kruse: How old where you when you were thinking?

Lukas Biewald: You know I remember being really little. I was like just maybe six or seven. My dad brought home from work a – he borrowed one of those IBM XTs I remember. All like you could do with it was actually program it. It didn’t have any programs on it. You had to write your own basic programs, but I was totally fascinated by it. I would – literally I would copy programs from magazines into it and run it. Yeah, yeah, you know it’s funny actually. And then I learnt a really bad programming style, because I had trouble finding the keys you know, because – so I would optimize for the least amount of typing. Yeah exactly, but you know I keep my variables as short as possible and variable names would like always be one letter and I really didn’t – no one showed me indentation, you know so I really didn’t see the point of indenting inside loops, my least favorite. So I can’t even imagine how the legible the code I was writing was, but that’s pretty amazing and…

Dave Kruse: So even AI…

Lukas Biewald: As I used to…

Dave Kruse: Oh! Go ahead, go ahead.

Lukas Biewald: I got really into like programming AI you know. I was really interested in – I loved games as kid actually and I loved writing programs to make computers play games and I’ll tell you when I got to college I knew I really wanted to study AI and at that time AI kind of moved from these rule based systems that I read about to machine learning. It was kind of just becoming really popular and I worked at the Stanford Robotics Lab building actually language translation models and other kind of natural language processing applications. And so then I went to work at Yahoo in 2004 and we were actually deploying machine learning models even then to do search element to decide how to rank search results, which was super fun. I mean it was really early. I am not sure Google actually was using fully machine learning for their ranking at the time. But it was super cool because you know I remember the engineers noticed that I had checked in more codes than everyone else combined, because the code was computer generated. They were just like, what is going on over there, like what are you doing?

Dave Kruse: This guy is amazing.

Lukas Biewald: Yeah, no I felt really proud I remember, I remember that yeah. And so I was largely into ML and then I went and I worked for a company called Powerset that became Microsoft Bing after an acquisition, and kind of what I kept seeing was that the main thing kind of making it tough to deploy machine models in the real world is the absence of training data right. If you have good training data, you can almost always build a useful model, and if you don’t have good training data there is very little that you can do. So you know your options are either moved to unsupervised or semi supervised algorithms, which I think was really interesting but really, really tough. Very, very few companies are actually able to deploy that type of approach or get training data which is, you know is a tough but you know a totally doable approach. And so that’s how I got into CrowdFlower and you know got really excited about helping out customers build training datasets.

Dave Kruse: Got you, okay. And before CrowdFlower was there, did you work on any project that was pretty noteworthy or anything that – I mean obviously Powerset was pretty noteworthy. I remember when that was around; I remember when Microsoft brought that, but anything there that you did that was really fascinating to you?

Lukas Biewald: Yeah, I think there was so many projects, I feel like I did too many projects. I love projects. You know there is one project that I did, that’s back before Powerset, was I built a – I love the game GO. You know I mean there are not many games that go for me. I always loved GO. I remember building a 3D version of GO. You know my friends suggested to do it and on like a grid with a caller type attached, just been like a diamond shaped kind of crystal structure and that was super fun. I mean I was you know like you know kind of – I remember making AI and AI could basically beat all my friends and we got it flashed. You know back then like Flash was like the big thing and some people downloaded it.

Dave Kruse: So even with..

Lukas Biewald: And you know like late – What’s that?

Dave Kruse: Oh no, go ahead, go ahead.

Lukas Biewald: Oh you know just like lately I have been working on these and I got a house with a garage which I was super pumped about, because I never really had room to spread out. My wife never really liked me taking it to the living room with my soldering iron. So I finally got a garage and so I started building robots in there and so I built the – one of the robots I built I got [inaudible] running on it, so I can do. It can do a dry run and do object recognition on a little raspberry pie. So I’ve been writing these series of articles for O’Riley that I thought maybe no one would read, because it’s just kind of you know projects that like entertain me, mostly in robotics, but yeah anyway I have so many projects I could get through an whole podcast about my projects.

Dave Kruse: Nice. And I saw some of those videos on the…

Lukas Biewald: Oh awesome! Cool, cool.

Dave Kruse: My God! This guy likes to do a lot of stuff and so yeah your right, you like to do projects, that’s cool and yeah, I could probably – the one – I saw a different one that I remember is or was it like for $100 bucks you could build a robot that could see. I think it was, yeah…

Lukas Biewald: Yeah, yeah exactly. That’s the one yeah. I think it’s probably gone down to like $80 now. Seriously like every year that gets cheaper every year.

Dave Kruse: That’s amazing what, yeah. But anyways like you said, that’s – we could talk about that the whole time. Oh! I wonder, so with the GO program that you wrote, so did you write AI that could actually be your friends in GO?

Lukas Biewald: Yeah, yeah, yeah, exactly. I mean it’s a pretty simple system, but I mean I think it’s you know…

Dave Kruse: How simple?

Lukas Biewald: The thing is like GO, 2D GO like everyone knows how to do it, but 3D GO, you know no one’s really – probably a 3D GO in the way I have made it work, it’s like no one’s probably played that many games with 2D GO. Another great idea was to get it, just everyone of it is not so bad…

Dave Kruse: All right, that makes sense. All right so let’s talk about CrowdFlower now. Can you – I gave a brief description, but can you give maybe a better one of kind of how it works and maybe how you started out and what it does now?

Lukas Biewald: Yeah, totally. I mean so the vision of CrowdFlower was always to enable machine learning by helping companies collect charting data and do a human loop. So you know I think it’s really best explained with examples. You know people in the field, they always get, you know oh, I need training data. Like that’s kind of the number one paying point, the people doing machine or anything. But you know if you are outside machine learning you know you might not realize how much work goes into collecting training data. So you know maybe a good simple example is one of our customers, big tech company gets tons of support requests right, and so you know they want to classify the requests based on the level of priority. So you know if you use zendesk or you use you know some of these custom support systems, they have tools to kind of automatically flag stuff. But if you are a big company, you probably have your own set of categories, right that aren’t the stocks best in zendesk or you know desk like company. You probably want to like classify support figures space for the category. So what they do is they send you know maybe 10 or a hundred thousand of these tickets or emails that come in; they send it through CrowdFlower and they set up a bunch of rules. Just say you know if you see this type of thing that’s a safety issue, so then we want that to be really high priority. You know if you see type of thing, its spam, so you know maybe we just do an automated response there. So they label these tickets and then from that they build a model right, and you know the thing about making money is it’s really, it’s often really easy to build an 80%, 85% accurate model; often impossible to build a 100% accurate model. So what they do is they run the machine learning and where the machine learning is confident they use the answer. But now what’s not confident they send it back to their original task that they’ve set up to get it labeled confidently by a human being and then they can use that label to decide if they want to prioritize the support set or not. They can also use that label to retrain the model, so that it gets better and better over time. And so lots and lots of people use this approach, but what CrowdFlower does it makes it all really easy to set up. So we set up the task. We find people in the world that can work on the task and do it well. Our software makes sure that the quality of work is high and then we have API’s that can make people use the automatic. So once you start this process, you can let it run forever and your models just get better and better over time and so you know lots and lots of you know big tech companies use this. I mean everyone from you know kind of YouTube and Yahoo to Bloomberg and Thompson Reuters and hundreds of other data science teams use CrowdFlower to make machine learning really work for them.

Dave Kruse: Yeah, those are some big names. All right, so let’s dive into a little more detail on that example. So I am curious, so let’s say somebody sent over 100,000 you know ticket requests and you know what percentage of those do you think will be kicked over to human to be – I guess what confidence level do people usually pick to you know have it sent over to human and how many of those, how many of the 100,000 would actually be sent over and how long does it take and how many people will have to deal with it. I am just kind of curious how the whole process works.

Lukas Biewald: Sure. I mean I guess this is why you know our software is designed for a data science team to use directly. You know it’s not something that we’re selling to people that are not technical. So our core audience is data sciences. So we give them more of the control right, so you know for some applications like – you know you look at a lot of the self driving car companies right and in that case you know you really want your car to be really sure, right, if there is a pedestrian in the image or you know even though it says report ticket routing, you know if there is any chance that something is a safety issue, you want to put that at the top of your queue, right. So your tolerance for error in that realm is extremely low, but you know we also work with you know consumer package goods brands that are going through social media trying to figure out what people are saying about their brands. Now in that application you know 95% confidence might be good enough, right. So you might say well, you know if above 95% just take the results, but if it’s really not sure, really confusing about what somebody is saying, then get it labeled. You know typically the first 10 to 100,000 paths are done completely by CrowdFlower’s contributors, because you know you need to build up a set of training data and you know we are really good at scaling up to that. You know you can get that done in weeks if not days and what’s pretty cool is you know even data scientists can set up tasks you know quite often. People will begin working on it immediately all around the world and then you know if you think the results coming back are wrong or you kind of want to change something about the way you set up a job, you can go right in and change it yourself. It’s pretty different than you know the experience that you might have if you worked directly with an outsourcing company, you know because we give you really good software and tools to make sure that your immediately getting results and you can see exactly how much money your spending and you can also see the level of quality that you think your getting. So you can you know put it in sessions where you know what the answer is or your pretty selective with the answer. You can use those to teach people doing the task that you want. You can also use that to judge the people and are doing the tasks and make sure they are doing it in the way that you want. So it provides us a lot of data and that’s why data science can use this, because we’re able to get really high quality results which I guess you really need for training machine learning algorithms.

Dave Kruse: Yeah, that’s awesome. And do you share pricing – not pricing. I was curious how much do you – and if its fine if you don’t – but like you pay per instance or somebody classifying a piece of data or does it just depend on how much they do or how does that work?

Lukas Biewald: Well, so you pay CrowdFlower, you pay CrowdFlower before using our software and then optionally you’ll typically pay someone to do the task right. So you know our software costs about $1,500 a month and that will give you up to 100,000 rows, which is typically enough to build a good machine learning model and then we have plans that go beyond 100,000 if you need that and then you also have to pay the people working on the tasks, but that’s completely up to you and we’re super transparent about that, right. So you know you can say you know hey, for every label you know I’ll pay the person $1 and then you can see if people are willing to do it for $1 and it depends a lot on you know how easy your task is and how willing – if you are willing to let people you know outside the US do the task, because you have access to a lot more stuff. You know if you need someone that’s you know fluent in Spanish and also understands how to read financial statements that can be you know significantly tougher to find. So that’s kind of an open market place that you should get access to. But a lot of customers use our software to organize their own teams or you know organize as a company to have an existing relationship with as well.

Dave Kruse: Interesting, okay. And when somebody comes, when a new tester comes on or a new classifier I guess, what do you call your people to do all the classifying or labeling?

Lukas Biewald: Oh! Contributors.

Dave Kruse: Contributors, all right. So when a new contributor comes on how do you make sure they are good and they are not messing up. Do you have people to like double check their work or how does that work?

Lukas Biewald: Yeah, exactly. So we do a whole bunch of things. You know we actually again are all about the giving the customer control and transparency in the process, so when someone new comes into our system they don’t have a level yet and we’re constantly monitoring peoples work. So we look for some aggregate statistics, right. Like overall we think that their distribution methods should be similar to other people. You think that you know how long they do, how long the exponent task should be similar to other people and then we’ll also, you know we’ll double up questions asked to go with and question and see if they agree and you know if they don’t we might ask a third, fourth, fifth, sixth person to lay in. You know some questions are naturally ambiguous, but some are pretty clear and then you know finally most of our customers when they create a job, they will hide in some questions where they know what the answer is. And so you know we look at all the statistics on how people are doing and then you know as people do good work for our customers over time, we move them up a level. So you know if you target your task to the highest level of contributors in the system you will always get really, really good results. Now you know over time you might want to scale to you know tens or hundreds of thousands of people, in that case you kind of need to make your tasks more robust and you can you know target a wider pool of people. What that does is it sort of gives you, it gives you control over you know sort of the costs, scalability and quality tradeoff that you are willing to make as the customer.

Dave Kruse: And how many contributors do you have on the platform?

Lukas Biewald: You know I find that people don’t always like me to say it, because it’s a little – I mean it gets just a little misleading.

Dave Kruse: Okay.

Lukas Biewald: I think we’ve had over 20 million people come in and do a task. It doesn’t mean that you can access 20 million people on a given day, do you know what I mean? Like so just to be like super clear about that.

Dave Kruse: Sounds good.

Lukas Biewald: But there are hundreds of thousands of people active on a platform anytime and you know we’ve – what the people on our platform want is they want more customers to come in and work on a task. So you know every customer, they are always like, oh! You know are you scalable enough for us and to-date and we are always scalable enough, as long as you know people are going to read it and will pass it in English. Now I support English and you support a wide variety of languages, but in some of the rarer languages or languages where we have less work coming in, the scale really can be in Cuba in English. In English it’s very, very hard at this point to affect their market. I mean you have to be spending millions of dollars on CrowdFlower to you know to that sort of trouble finding people working on the task, yeah.

Dave Kruse: Nice. And how do you find – how do people find you? I suppose by now they just kind of know you. Was it hard finding contributors initially?

Lukas Biewald: Yeah, you know there aren’t many good contributors. I mean lots of people come in and they start working on tasks and the key is like you know who really wants to do good work and who is really reliable. There are going to be a lot of like kind of scamy sites in there that they are like hey, you know make money online, you know make money from your home and you know so it’s a little bit tough to standout in our environment. But you know we’ve been doing it for so long and I think we are consistently doing good work for so long that we have a good reputation among contributors and then we also do deals with existing outsourcing companies. So we’ll route past existing outsourcing companies and you know that gives our customers who want to do that with a little more trust that you know there is a contract and that it’s going to someone like in an environment that’s more controlled.

Dave Kruse: Interesting, okay. Yeah, I mean you built such an interesting platform right. Where do you kind of see things headed? Like do you want to – I don’t know I could see I guess different visions for you. I’m curious, before I say what I think you could do, I’m curious where you want to take things with what you are doing now?

Lukas Biewald: You know in principle I think there’s a long way to go with exactly what we’re doing today. You know I think more and more people need training day as they are doing machine learning and I think CrowdFlower is becoming that platform that you know everyone knows they can use to get high quality training data. So you know there’s like a million important and challenging projects within that vision. I mean one place that we’re really starting to expand into is you know first of all the human loop system right. So keeping more data coming in overtime and then also we now offer the service of building the machine learning models for you. So we didn’t want to start with machine learning models, because we saw a lot of people building kind of machine learning APIs and you know the markets crowded and actually you know for a lot of our customers building the model is really their expertise and so they love doing it, so you know we didn’t want our customers to feel like we’re kind of like taking away the fun part of their job. We want them to know like we’re you know, we’re helping about with the hard part of their job. So – but you know lately actually we’ve been seeing a lot of success actually where our system will let you build the machine learning model directly off the data that you’ve collected and so you can actually use that – you can use that to have a complete ML systems, you know suits and nuts that you can plug an API into and know that it’s going to be really robust. But you know I think like where we live is sort of the intersection of hybrid, human and computer systems and yes, I think that like always CrowdFlower has changed our effect as you know being a section of humans and machines.

Dave Kruse: Yeah, and that’s a good leeway into my next question which was and I saw another one of your videos about where you’re talking about how – and I feel it was either chess or – I forgot which game it was, but how a machine plus a human was better than either one alone and…

Lukas Biewald: Totally.

Dave Kruse: And do you see CrowdFlower kind of going into that space a little bit. Like I was curious like even more advanced projects you worked on beyond let’s just say labeling or yeah, I mean maybe that’s – to me that question is off base, but just curious to get your thoughts on it.

Lukas Biewald: Well, you know I mean I think like the realm that makes sense for us is a bit on supervised learning. So you know it’s probably not even like the GO playing systems anytime soon. Yeah, that’s what we have tied up with what we do, but you know almost everything that machine learning really does today is essentially to supervise learning problem. You know you have some variables that you’re trying to fill in based on some data that you showed the systems previously. So you know for example there is these forms in computer vision, right. Like you know if you’re trying to do self driving cars, well you know first of all you need a pedestrian detector right. You know you need to label out the pedestrians and then you know like what if you see like a baby in a stroller right, you might need a model for that. I mean you definitely don’t want to do that – you don’t want anybody to know that, about that baby carriage right and then you know and then like people on bicycles right and then you know what about dogs right you know, yeah you want to know what’s going on there too. So you know I think there’s sort of an endless stream of these types of tasks like where you know its these kind of clear your objectives, you know machine learning tasks where you need a lot of training data and you don’t want to screw it up, and then inside businesses where you think about all the business process that could stand to be automated right. I mean routing support requests, what about routing in the sales fleet. I mean you know every big company has a pretty complicated system for that and if you automate that you could save them a lot of trouble right. I mean you even start to see like you know HR stuff. At a big company you know like Clique sorting through resumes and deciding you know which ones are the most promising to look at is actually a pretty good application for machine learning. So you know it sounds like CrowdFlower like has enough applications today with all these details.

Dave Kruse: Right and you know I was like more advanced, but I mean you’re actually talking about some of the most advanced machine learning. Like you work with the self driving car companies, like what else more am I looking for?

Lukas Biewald: Totally.

Dave Kruse: Yes, so that’s interesting. And is that with machine learning, with self driving companies and I had the CEO of NuTonomy on here and we talked a little bit about that and I asked about snow because he is in Boston and we are in Madison, Wisconsin. So he said that snow is not the easiest and so I am curious and maybe you can’t disclose this, you know what type of labeling do they want done on – because they must have so much data and so many images. So I guess it’s what the confidence level is below a certain threshold is when it gets popped over. Are they trying to say, is this a person or is this a sign or is this a stripe on the road. You know I feel like just something is…

Lukas Biewald: Yeah, there’s something variable and I think each self driven car company right now has a pretty different take. You know a lot of them worked with us, right. They tend to be more secretive than others. Like they definitely talk generally about what they look for right. I mean you see all kinds of classifications. Different companies frame it in different ways right and some companies are just like hey, you know we want to know like anything that might move. We want to have a classifier for that, right. So anything that’s sort of not like a tree or a rock you know like where it kind of likes changes quickly, so you know a cyclist or a dog like we were talking about. And then other companies they want to say you know hey, like what’s the terrain here that actually we could drive over right. So you know you can drive on a road, but then you can also drive on you know like a gravel road and maybe in an emergency you can drive on a sidewalk but you could never drive into a building. So you know there’s all different kinds of labeling tasks that people want and actually there’s different kinds of sensors you know. So you know some cars have what they call LIDAR right that gets you sort of the distance, not just the particular. Other car companies really focus on cameras and you know when we test it we haven’t shown on it right. Yeah, so these labeling tests can get pretty complicated right where you are actually like labeling things in 3D and then there’s even labeling things in video right. So some of our customers they want to label every frame of the video with you know where the cars are for example and that’s the type of thing that we really need high quality tools to do that. Because if you know you are really going through every image and labeling stuff you’ll never finish, you’ll never get a whole series of – we get labeled on a few of these smart things like we do, like interpolate between the video. So you know you label a car in one video, in one frame of the video and then 30 seconds later you label it again, you know it’s probably going to be in between the two places that you thought it was, in between the video. So even if you have a whole suite of tools for image annotation and I would say the primary buyers of image annotation is these self driven car companies. I mean maybe number two is satellite and drone companies, which are also taking off and have pretty similar needs, but that’s more looking downwards, you know cars are kind of looking across at other things. But pretty similar – pretty interesting and pretty similar types of things.

Dave Kruse: So what the image recognition for let’s say self driving cars, you know there might be a lot of stuff in that. Well, actually I haven’t seen those images, so I don’t know, but there could be a lot of stuff. You know there could be multiple cars, there could be a street sign.

Lukas Biewald: Totally.

Dave Kruse: So how is it labeling? Do they draw around the object or yeah, how does that…

Lukas Biewald: Totally, yeah. So you know again it depends on what the customer is trying to do with it right. So this is again why we really believe in sort of the labeling power in the hands of the people doing the machine learning, because you know if you collect the label for machine learning you really want to have your machine learning people involved in the labeling process. So we don’t take a strong point of view on this I guess. You see some of our customers they want to build boxes around things because they have bounding box tools that are then trying to find bounding boxes around objects, that’s super common. You know another approach that’s gotten really popular like this whole semantic segmentation and this is where you try to label literally every pixel in an image, it’s what it is. So we have a tool that we call a pixel label tool, where its – you really just Photoshop like a magic wand. Things where you can kind of drag it and it sort of guesses the edges of stuff. I’m not sure if you’ve ever seen that, but we basically have a set of tools that are yeah, exactly. So we have a set of tools that drive really smart about that, because like you know if you’re going in, in Microsoft Pen or something and trying to like label every pixel, which is part or a road or part of a tree or part of a pedestrian, it would take you years right to label images. So we put together a set of tools just for this sort of like dense pixel labeling use case and then – you know and then video has its own, like I was saying, its own set of stuff. So you know I think all the stuff that people are learning in images is really exciting. It’s been great for our business. I mean a couple of years ago, almost everything we’re doing with tech space and you know we were actually looking at the numbers and over half of our new customers in 2017 so far have been doing some type of image or video label.

Dave Kruse: Oh wow! That’s exciting.

Lukas Biewald: Yeah, it’s a pretty big change for us actually.

Dave Kruse: Yeah, I bet. And the tech space, I suppose that was a variety of use cases. Like you talked about the HR, sport ticket and stuff like that.

Lukas Biewald: Yeah, you know and especially I mean the – I think like people who are in the industry probably – they probably wouldn’t know like what’s the really big stuff. But like one person who sees just a ton of machine learning and then retail companies trying to rank which searches for someone, because that is just like totally effective bottom line. If your Home Depo and you know somebody is searching for like a 15 inch bath tub and you show him like you know a 55 inch bath tub, it’s irrelevant to them right, yeah they are going to go to Lowes and buy it there. So you know search is this really challenging problem that also really affects your top line and so that’s where you see a lot of resources deployed to make it work.

Dave Kruse: Got you, okay. So we’re almost out of time, but a couple of more questions. One is, and you might not have an answer to this, but I was wondering if you had any type of an example or case study that was kind of abnormal I guess. Like a different project that made you laugh or is like that you can share and if you don’t mind.

Lukas Biewald: Yeah, totally. I mean we had so many, I love it. I mean I could do that, I could talk all day about this stuff, but one project I thought was pretty cool. People they asked us recently as they were – they are actually building a classifier to detect elephants and we were like wow! like why are they doing that and I guess it’s to – they have these drones flying around following herds of elephants and the point of it is to stop poaching and then it’s really real. I mean this customer is collecting tons of data and you know has drones out there looking for elephants and looking for humans, to make sure that humans aren’t trying to kill the elephants. I just thought that was such a cool application in a computer embedded in the CrowdFlower.

Dave Kruse: I bet you didn’t expect that one when you started back in 2007?

Lukas Biewald: Did not. I didn’t know that was…

Dave Kruse: You didn’t pitch that to investors.

Lukas Biewald: Yeah, exactly.

Dave Kruse: All right, so one of my last questions is I always try to ask you know, what do you do during the day or the week to you know – that gives you a good energy or makes you happy. I mean it sounds like you got good energy right and it sounds like you work specifically like working on lots of stuff, but anything in particular?

Lukas Biewald: Yeah, I love working on stuff and you know I’ll you the company has gotten bigger. I think like the thing that really rejuvenates me, it’s a little weird, but I think just working alone in my basement on my own projects where I can kind of see how everything works from beginning to end is the best thing. You know whenever I am like end up saying oh no! I feel like you can go back to the basement and do some stuff down there.

Dave Kruse: Go to the basement.

Lukas Biewald: Good to know that those make me happy. I can just like – I feel like I’m just kind of talking to me when I’m building stuff, so that’s one of my favorite things to do.

Dave Kruse: That’s what you grew up doing, so it makes sense.

Lukas Biewald: Yeah.

Dave Kruse: How many people work at CrowdFlower now? About…

Lukas Biewald: About 85 I think today.

Dave Kruse: Wow!

Lukas Biewald: It might be a little more. It has grown a lot lately, yeah.

Dave Kruse: Wow! Nice. Well, I think we’re just about out of time unfortunately, but Lukas I really appreciate your time and thoughts and what you’ve built is quite useful for the machine learning industries, so yeah.

Lukas Biewald: Thanks so much. I really appreciate it.

Dave Kruse: And I guess for catching poachers, I mean that’s amazing. You should make a long list of different projects like that in your sight that are kind of a little different. Maybe you do, maybe I missed it.

Lukas Biewald: Definitely yeah. That will be cool. I’d love to do that.

Dave Kruse: But yeah, so I definitely appreciate your time and thanks everyone for listening to another episode of Flyover Labs. As always I greatly appreciate it and we’ll see you next time. Thanks everyone. Thanks Lukas. I appreciate it.

Lukas Biewald: Take care.

Dave Kruse: Bye.