Transcript: Bay Area AI: Lukas Biewald Interview

a hello everybody I'm Alexa crabber of the organizer of bi ara I'm it up and here we're on location at crowd flower a very well-known company which does human in the loop computing and we have to look as build founder and chief scientist with us here thanks Thank You Lucas for was not ah so a lot of folks heard about crop flower and recently I've seen you guys are the lot of talks about the I so can you tell us briefly what is crowd for how to vault into the kind of center of the AI discussions you should have right now sure sure um you know so I i started crap flower because i was actually working in AI and it was a tool that i really wanted and it didn't exist right you know so what crowd flower does is help people get training data and then help people do human in the loop mm-hmm and and you know I it was like seven eight years ago I guess now that that I started it and and basically ever I went I I wanted to build real machine learning models it really worked in the real world and every what the bottleneck was just getting the training data right so that's just like you know it's like the most simple thing it's a good thing that you don't think about the most but if you actually want to build machine learning it's it's often your biggest problem and so you know we started it around getting people training data and then we realized that you know that the next biggest thing holding people back is a lot of times they're models only you know a percent accurate which often isn't good enough for a lot of real-world processes mm-hmm but you know if you only use the cases where the model is good if it almost by definition you have a useful model yeah and so we realize you can help a lot of our customers actually like deploy things and really do things by adding a human loop component to our software and so you know so that's that's how we ended up where we are today so I remember using crowthorne levels at cloud you know 2012 it's also ready can difficult to place huh do you know like this kind of specialized you know annotation all right and and then you know obviously Mechanical Turk is kind of the the flagship of this kind of services so can it just a little bit how I guys evolved from you know that those years in today and how you're different from mechanical turk and kind of other sharing sources um well I'm surprised you say Mechanical Turk is the flagship I would have to disagree with that ok historical it just article historical first maybe like the dinosaur ok you know so you know how do we have often from from 2011 you know I actually don't think I think if you you know if you looked at our product you would you know you would see that a lot of the kind of like bones of it are very similar to what was happening in 2011 but I think what's what's really changed is the market has clarified itself right so you know if you're the machine learning in 2011 you know there are probably a lot less people doing it right and so you know at that time we just didn't have the same customer base we didn't have you know the company just wasn't as big because you know the market hadn't hadn't taken off in the way that it has now right so you know I I don't think like you know there's there's like huge differences between you know crowd flower now crowd far in 2011 but but I think like all the little details are the difference between you know kind of the the low end PS that we had now the really high NPS that that we have do have today mhm because you know because of you know we now have the resources to get all the details right and actually make the product really good for our customers mm-hmm so I look at you guys um you basically you become the platform all right so kind of jealous a little bit you know in what sense you are the platform for a hike why wouldn't say with a platform for AAA well I guess we're a platform for ya um you know I think I like to think of ourselves as like the essential missing tools for AI I mean what's funny is like you know the AI is kind of the least important part of our platform earlier right like I think it's the human loop and the the training data that that's really where you no crowd flower really like yes but not evil yea yea enable men they enable them it was a good way of saying it and and you know how we became that platform I guess you know we started off with like a very specific you know solution so it was it was only language and it was is basically collecting training data for language based applications mm-hmm and so you know from that very specific point solution with a small number of customers you know we've expanded its kind of all the different things that that AI does and I think like really you know people say like well you know you solved all these different industries like you know you sell to like finance and retail and hand you know consumer internet and and you know think a lot of people say well you know how can you do that how can you know a company that's only like 70 80 people like crop flower looking forward to do that and I think you know the reason that it works is because we actually sell it to really one buyer and he's like all over the walls here right it's we call them dated an all right basically sell it to data scientists uh-huh and so you know I think what's really interesting about data sciences is like you know a data scientist at Bloomberg has a lot of the same needs as a data scientist at home depot mm-hmm you know you might not guessed that but actually really their lives are you know really similar me that both at the mercy of data yes you know they want clean data right that's like you know something that they both need you know they're both trying to build accurate algorithms like when you know for both of them when they're their arms do different things when the algorithms fail it might cause a problem right like so this is function breaks okay so you know they care a lot about the acuras that goes against my disorder the skin Linda for basically exactly sever and so these guys which i think is really cool about being a data scientist today is is like you know I think these folks you we see them bounce around from into different industries right because like a lot of times our customer you know be surprised like they'll go from you know like working at you know working at like uber to working at you know Thomson Reuters and yeah and it's like well it'll take the tools with them they always have you know they need to clean up some new data that will come and they will turn back to you an exact way to help on the new kind of date exactly exactly yeah yeah yeah interesting interesting so it's actually interesting that you mentioned that you know you start with language because you know this meetup used to be a subtext and what I found that essentially people often don't realize especially newcomers do machine learning data size that you know the machinery algorithms are all the same right and so basically look at this real wealth of a mile expertise about a lot of people did not realize this so once we started calling this data in the eye we suddenly doubled and tripled nice our audiences am and it's kind of cross pollinating but the core is the same so so I'm wondering right like what's your take on this evolution you know it used to be an LP and then that the science and later became cool and let's not cool anymore and the eyes is cool right now so kind of you know the design is like so 2015 so this is the year of AI nvidia right and so so what what is your take right like the methods don't change but like the kind of people change the kind of company support what do you see in terms of my customers kind of turning 22 you where the the trance kind of what's interesting in you new developments well i mean i would say one okay so it one thing like for me you know about the this sort of like hype around AI is like i have to say like i love it I mean that's like the first thing that I really loved and I was kind of disappointed when I came out of school and a lot of stuff that you do like was an AI so I mean I have to say like for me like like AI becoming such a thing is I think it's awesome and I just you know I had a friend who was told me they asked me their day a grad student she was saying you know it like what's difference between you know like I understand the difference in like you know you know the regression I do and machine learning and I was kind of like well you know there's not much difference actually I think that's like I kind of astute observation right then sort of the same thing but that doesn't mean that it's not like powerful and important and interesting so I guess like I'm all for you know people being excited about AI and doing a and I keeping I was a little skeptical about about deep learning and I've totally changed my mind after like trying some of it myself I mean I think like you know neural networks are doing some amazing stuff in vision mm-hmm at the very least and it seems like cool stuff in text as well so you know so I think a lot of this stuff that's happening right now is actually like just legitimately really exciting I mean I think that comes with a whole bunch of irritating hype but you know we've seen other hype cycles and you were okay with that yeah better than a pipe them to you know no hype i'm at least it nice Peter takes like yours ya know this is fill the bowl since it's good for the context right yeah and in the parent everybody is 8 so Sonya look like yeah my here the vesicle essentially every be company going to Tori I i realized the data problems the key problems haha I hear that they all have teams working internal and all of them are probably the your customers huh yeah right so but you know they they don't talk about kind of this date in richmond in richmond is happening I don't know it goes chembur so I under right what do you think of this i mean it's it's possible for business run across basically let's say you know Cooper is kind of gonna do it on their own maps who has the maps Yeah right they called the key to location oh you know Apple basically decoupled from google maps and I will bear wants to do its own maps and son they're all probably going to do the same thing with maps so yeah I'm curious how it reflects on on your business so do you for instance if you're not eight maps for over a hot girl yeah welp experience right so do you reuse this and in kind of and then kind of now you have knowledge let's say how to quickly purify maps huh so what kanokon email make it like a macular purification a separate business ah how does working around dirty well you know so we can't share data between our customers for obvious reasons right um but we do with mood of a program where you can use our software for free if we can open up your data and you know the reason that I did that was I was actually really concerned that you know with with with data not being open source and openly available as data becomes kind of like you know oxygen for companies likes just so essential for companies it made me worried that the best assets were all closed yeah actually is a humongous issue yes I mean I think like open source has done so many good things for software I think there's a kind of analogous like thing going on and data except that like it's all closed and so here what happens is like all the academics they want to work on really big really interesting data sets they end up going into industry you know for that reason right and then and then you know just kind of perpetuates the cycle of more and more clothes data so you know I think that's actually a major issue and and you know we're trying to help that by opening up the data sets that we can mhm um you know in a way it's good for business that people want to keep collecting the same data sets over and over but i think you know it's one of the things like in a kind of like a bigger longer view I think there's like plenty of money to be made in data and you know the more successful people are the more stuff that they'll do yes and Eunice will come with a new date alright so yeah so do you see any solution to this I mean do you think people will start sharing my data I will get it in the initial stage everybody's calling data and then people will realize like this data can share or do you think that will just continue more and more like this and you know big companies will amass preparator data small come to be like looked out haha no have access to this I don't know I think it's really tricky issue actually I mean I think I hope that we find a way to convince companies to open up date I mean I think like you know our governments doing a fantastic job of opening up data I mean I hope that you know people can find incentives you know like like I was pretty excited about you know even this is like a while ago but remember the netflix pride is right with you yeah and that actually caused a lot of data to get open yes but then you know is right to participate like I wrote a fortune for girls openmp nice I mean so I think it's really sad that they didn't do it the next year because there are privacy issues you know a dude animas ation yeah so you know I kind of felt I feel like um I feel like that's a major issue that that that people really need to figure out because I think like there's a major hidden costs to not having open data available like that and you know like everybody is flix dataset even now your peer recommendation engine they always use the netflix data center right and so you know probably like you know if if companies realized that they could get so much work done on their data sets you know maybe they'd open more data set so people would work on them and give them free insights kind of like open source all right yeah I'm wondering you know if kind of data will follow right open sore because open source did not materialise immediately right do you realize maybe 40 years right after a straight kind of developed so some curious also bought you know you know Madonna so obvious as an example but you know I think you know from personal experience that you guys are cold when you know that it is proprietary and you have this NDE haha partners right where you can draw data in companies like Bloomberg right ashcraft information yea rather three so I'm curious right how you guys do this I'd like those has never been a leak at least you know I'm not seen a leak rights for a data analysis gold right so how do you kind of convince companies to give you all this available data how do you sure that this providers will keep that a secret I how do you manage this process and convince people like bloomer and of course higher standard of Kevin challenge to basically give you all the data to eternity um well I think there's no a monastic like data security is an issue where there's no silver bullet right i mean there's just a lot of lead bullets as they say all right so i mean you know i think um you know i think that the way you convince a company like Bloomberg to give you data is actually just like consistently not having leaks yeah you know why we don't have leaks I think it's because we have good engineers building good software and then we we work with really good partners I mean you know I think like a data leak would would put any of our NDA channels out of business if it's certainly like one data leak we're never going to send our customers to to one of these partners again and so you know they're highly incentivized to to make sure that stuff isn't leaking and we're feel really proud that we haven't had any leaks but I don't want to encourage any hackers or anything to you have to come try that yeah but I'm hears about the kind of the the culture because it reflects on the culture because obviously you're startups are people know who build them that engineering culture is key to successful startup right and apparently guys managed to build the data culture where the humans involute not will do the right thing so yeah the curious right do you have kind of training materials we should give to the partners how do you kind of certify a partner or do you kind of have long-term relationship is attacking informal kind of people-to-people oh how do you have any process do have any manuals that have any kind of training you send you know you're you know managers of this company is that they serve the people how's work yeah I mean he gets all the above right so you know we you know it's a pretty long process to get started as one of our partners mm-hmm I mean I think the person that we have doing it is really excellent I think you know think one thing that's actually really benefited us is we tend to work with more partners that also have a social mission mhm and and I think that tends to get higher quality companies okay so you know for example one of our partners you know gives work to Muslim moment that in India have trouble like leaving they can't go very far from home right I remember she spoke at the regime yeah trata yeah yeah so yeah she's a longtime partner and now when you talk to these women that do the work they're there they love crowd flower there they feel really connected to crowd fire they even feel connected to our customers it's funny they know so much about American culture because a lot of move works as crowd flower tasks you know night and day for the last four years mhm right so um near they're telling me like men we've learned a lot about American fashion you know we've learned about you know I mean just like all kinds of like interesting things that they that they know and and and you know because you know they're the highest earners in their family you know they're there I mean I guarantee it'll never leak mm-hmm anything on purpose because you know it's it's such a meaningful thing for their lives and their families mhm no this is great yeah it's I think it's really awesome and I remember this from the summit right i think it's great context you guys out there so i think i'll probably you know wrap up with this question this meetup is kind of the first in this kind of reeva grated you know I I cycle we're doing right we're going to actually wrap up the the temple gonna do you know to adopts a month and we also want to do kind of systematic exploration of why I so on the one hand we want to have kind of you know horizontal exploration so different verticals different industries and you guys probably see a lot of them but on the other kind of want to have this kind of build up as they do in science but in the startup context right so we teach people i think is very appropriate that we start with lowest level service data is data right here in the bottom or right of the foundation so but the christian you know is experienced practitioner and what would be your advice to a lot of people get into the area i feel right so we want to kind of you know down with hands-on talks haha right from different partitioner so kind of build kind of a mini course right huh so if you were kind of learning a I you know kind of if you tell your previous self lonely I know in what you know right now and to be the most efficient person in this like least amount of time yeah right like how should you go about it like should you learn you know linear algebra and calculus and all should you just play with our aggression like what is kind of the most efficient way to become a data scientist and stuff playing with all the schedules mmm do a question um I'm actually teaching a lot of data science classes these days it is really fun and I would say you know you don't necessarily see the most efficient thing like I i mean i'd say like I I love math you know I'm uh you know I'm math guy I thought was a mathematician and and so you know for me like i love all the math but i think if if you want to be fishing don't want any math you know any math right i mean actually like I think um you know I think like you know you learn some Python and you learn I mean scikit-learn is such a fantastic piece of software what I mean I just like very practically i would say learn Python hey learn scikit-learn start building models and then focus on the applications because like I think there's so many machine learning people that get so kind of like narrow mindedly maniacally focused on the the algorithms and making it like a percent or two better mhm and I think like the real key to making machine learning work is not like you know can you get from eighty percent to eighty-one percent but it's like you know how do you make eighty percent okay yes you know like those giggle data get yeah you get more data that's one way or more like or maybe like make it in some way that like the mistakes are like okay or like fun you know like I mean you know I think there's just like so much creativity like I actually think like a lot of stuff is like you know like actually it's more human-computer action what you think about like you know I could make like a ten percent accurate self-driving car they could be really useful if that ten percent of the time it's accurate is its parallel parking right right a car that pillow pressed for you that's great i love it right yeah if I make a 99 percent accurate self-driving car that doesn't know when it's gonna crash yeah it's gonna kill me immediately like I said it's like incredibly dangerous oh that's right you know I think like focusing on the accuracy I mean let the like what the nerds like you know nerd out on that and I think if you're if you're kind of just coming into the field I think you can bring your like you know whatever experience like you have to bear by building interesting applications in a field that you know particularly well mm-hmm it's actual device right I mean when you mentioned this i think you know stuff that people know that you know much much better and easier thing to do is just ask the user right right totally the loop right yeah i said okay i'll try to squeeze a percentage from a matrix just get another oh there yeah just ask them right yeah I'm some some some kind of active learning and see what happens totally yeah cool thank you very much we're looking forward to talking on great great two story here awesome thanks