Kirill Eremenko: This is episode number 201, a panel on emerging technologies. Welcome to the Super Data Science podcast. My name is Kirill Eremenko data science coach and lifestyle entrepreneur. And each week we bring you inspiring people and ideas to help you build your successful career in data science. Thanks for being here today. And now let’s make the complex simple.
Kirill Eremenko: Welcome back to the Super Data Science podcasts. Ladies and gentlemen, today we’ve got not just one guest, not two guests. Today we’ve got four guests on the show who are together in a panel. So as you probably know over the past weekend we had the Data Science GO 2018 event in San Diego, which was a total blast. So if you were there I’m sure we caught up and a huge, huge thank you for coming and making this networking and community so amazing.
Kirill Eremenko: And as a part of this event we had a panel on emerging technologies. And today you will hear the audio from that panel. There was some very interesting discussions, some very interesting questions from the audience, and today you’ll get to know all about it. So the guests on the panel were as follows, we had Mark Skinner, who is the senior solutions architect at NVIDIA. Then we had Rachel Shuyan Wang, who is the manager of data science at TrueCar. We had Ben Taylor, who you know from the podcast, he’s been on the podcast several times now and he is the chief AI officer at Ziff.ai. And we also had Pablos Holman who is a world renowned speaker, inventor, hacker, an entrepreneur. His ted talks have over 20 million views and he was our keynote speaker at the event, and he was also on this panel on emerging technologies.
Kirill Eremenko: In this panel we discussed many topics including blockchain, AI and deep learning, machine learning and disruption startups and much, much more. So as you can imagine, I’m very excited and pumped for you to relive this experience with us. One quick note in the recording you’ll hear me say, “Throw the microphone,” to the audience. And indeed that’s what we were doing. We had this soft microphone. It looks like a box. And whenever somebody was asking a question we would throw it like a basketball across the whole room. So don’t be confused when you hear, “Throw the microphone.” Apart from that let’s dash straight into it. Can’t wait for you to check this out. And without further ado, I bring to the panel from Data Science GO 2018 on Emerging Technologies. Let’s get started.
Kirill Eremenko: All right everybody. Welcome back. So we’re going to have a panel discussion. Who’s excited for a panel discussion? Panel discussion on emerging technologies. And I’d like to welcome to the stage our panelists. Please guys. So we got Rachel from TrueCar, Mark from NVIDIA, Ben from Ziff and Pablos [inaudible 00:03:46] our, keynote speaker. All right. We’re going to start off with a brief introduction guys, like half a minute each please. Mark, where do you work? What do you do?
Mark Skinner: My name is Mark Skinner. I work at NVIDIA. Hopefully some of you guys know about us. We sponsored the room next door. My job is to make your lives easier by bringing you the hardware and the software tools to make data science much easier than it’s been in the past.
Kirill Eremenko: Fantastic. Rachel.
Rachel Wang: Hey guys. It’s working. I’m the data science manager at TrueCar. If you haven’t heard about TrueCar before we are a online car buying platform and we’re currently doing pretty well. So any time you need to buy a car feel free to check [inaudible 00:04:30]. Yes. I’ve been in the data science field for a while and I’m very happy to see you guys.
Ben Taylor: Ben Taylor. Mike on?
Kirill Eremenko: I think so.
Ben Taylor: Okay. Okay. Sorry. So Ben Taylor. I am a chemical engineer. That’s what I studied. I worked for Intel, a hedge fund. I helped build out a data science group [inaudible 00:04:51] company called HireVue, and right now I’m working on a deep learning startup. So deep learning is a service through REST APIs for image, audio, and video. So that’s been our focus, trying to get that into the enterprise for business.
Kirill Eremenko: Good. Pablos?
Pablos Holman: Hey. My name is Pablos. I work at the Intellectual Ventures Lab for Nathan Myhrvold and Bill Gates trying to invent new technologies, next generation technologies. We don’t really make any products or anything. We just invent. I work on a company called Data.world. Do you guys know Data World? Where we’re trying to solve some of the tool chain that you need to do some of the things we really want to be able to do in the future sharing a bunch of data. So Data.world is kind of like a GitHub for data. You guys should all use it. It’s free. So get on there. Actually the deeper it goes it’s super cool because you’re able to load your own data but then crosspollinate it with other people’s data and it all starts to get pretty exciting.
Kirill Eremenko: Awesome. Thank you guys. Can we start the timer please? So the format of the panel, we had a quick chat beforehand to understand how we can serve you guys better in this conversation. That’s what usually happens at panels. Panels get together and trying to understand, “What can we deliver?” And we found that by just coming up with a list of questions beforehand we kind of started delving very deep into either technical algorithm stuff.
Kirill Eremenko: So the way this is going to work is we don’t have any preset questions. It’s all going to be one hour of your questions, whatever is bothering you, whatever you want to know. And we’ve got really cool experts here on the panel. So just for you information Mark and Ben are most experienced in the hardware side of things and deep learning and chips and stuff like that, where the world’s going there. Rachel is going to be here for algorithms, machine learning, AI, and how you can use that. And Pablos is here for inventions-
Pablos Holman: Common commentary.
Kirill Eremenko: And making this fun. So yeah. So get your questions ready. We’re going to start throwing the ball. This is your opportunity. We’ve got 59 minutes left, the timer stopped again, to get going. All right. First question.
There we go.
Audience: Mark.
Kirill Eremenko: All right. You know the drill.
Audience: Here we go. Mark, please-
Mark Skinner: Yes, sir.
Audience: … intersection of hardware and-
Kirill Eremenko: Closer.
Audience: Intersection of hardware and … is NVIDIA have a team that can address small needs for say a small engineering teams and we need a specific set of hardware who do I get in touch with at NVIDIA to make that happen?
Mark Skinner: Well, the good news is that you don’t have to go directly to NVIDIA. You can’t generally purchase anything from us, but you always purchase through a reseller. Could be HP or IBM or Dell or our local person.
Audience: But I’m talking about pretty unique stuff like cutting edge … How do I get in touch with the right people at NVIDIA to-
Mark Skinner: If you want to get to the engineering team?
Audience: Perhaps.
Mark Skinner: Okay. And you have an idea for a particular algorithm that would be best fit in hardware or [crosstalk 00:08:04].
Audience: Perhaps.
Mark Skinner: Okay. Then the best thing to do is reach me afterwards. I’ll give you a card and I’ll-
Audience: Thank you.
Mark Skinner: … introduce you to somebody in the engineer staff.
Audience: Who next?
Kirill Eremenko: Okay. Next question. All right. We’ve got one at the top. Yeah, throw it. There. Miss.
Audience: Nice shot.
Speaker 7: It’s in the [inaudible 00:08:26] right there.
Audience: This is for the gentle man from NVIDIA. I’m sorry. I didn’t catch your name.
Mark Skinner: There are other people here too, just so you know.
Audience: So I’ve been-
Kirill Eremenko: How about you just ask the question and then we’ll decide who will answer it, okay?
Audience: Sure. Sounds good. So we have seen GPUs take over, go deep and do deep learning and then [inaudible 00:08:48] a lot of insights. When you see GPU becoming the mainstream … or I’ve not seen a lot of adoptions from big companies. Do you think that it’s going to happen very soon or is it going to take some more time for companies to adopt GPUs?
Mark Skinner: I think we’re at the beginning of the knee curve for GPU accelerated applications. If you just look at how much Amazon has been spending buying GPUs to put into their facilities, Google also in the cloud it has GPUs. They also want to buy their own product and put their own product insights. So we have the TPU as competition. So there are other people that are seeing GPUs as a really good target to go after, what GPUs do. And so we’ll end up with some competition, but I think every company right now is investigating where are we at with our data center, plans for a GPU accelerated servers, and who can use them in our company.
Mark Skinner: So the good news is I just think we’re cracking the shell on this. We’re just opening it up. And all of you guys, if you don’t know what a GPU is, it’s that thing that does parallel map better. So we can talk about it some more. But basically that’s it.
Audience: Ben [inaudible 00:10:02].
Ben Taylor: Yeah. So I think right now there’s a gap between what businesses see as being useful AI and what they can do. And the spend for some of this new hardware is expensive. So at HireVue we built out a Gamey box. I think we spent $15000. If our first box was $40000, you get into the bigger stuff you’re spending hundreds of thousands of dollars. No business is going to sign up for that as a science project. So you really need the implied applications. And there is a crawl, walk, run. So if you want to crawl immediately go get an EGPU, an external GPU USBC, play with it, find some value, upgrade, find value. And then bigger businesses, it’s easier. They can commit to bigger machines to support a whole R & D team. So that’s kind of the hesitation. The hardware is more expensive, but the value is definitely there. There’s just a closing gap between what’s applied, what’s useful.
Kirill Eremenko: Got you. Thanks. Okay.
Audience: So there’s a lot of-
Kirill Eremenko: Could you stand up please.
Audience: Yep.
Kirill Eremenko: Thank you.
Audience: So there’s a lot of business leaders out there that are totally uninformed about how machine learning and data science is going to transform the industry. So as kind of data scientists practitioners here what do you think are the keys to evangelizing the value behind what data science provides to companies?
Kirill Eremenko: Maybe Rachel. You’ve got yours.
Ben Taylor: Yours works, yeah.
Rachel Wang: It works. Does it? Okay. Cool. So can talk about that from a organizational standpoint because our company recently went through a process of going from a centralized data science team to splitting the team up into two parts. Part of the people will be machine learning engineers and that team will grow, and the other part of people will be product analytics data scientists, and through the machine learning engineers will be more focused on implementing algorithms directly in the [inaudible 00:12:04], and they are kind of like engineers who know machine learning.
Rachel Wang: And I think part of that reason is because of the emerging technologies now that has allowed people who are not from a computer science background to be implementing code directly into the production database. And the other part of the people are doing kind of like getting insights and doing researches to improve products.
Rachel Wang: So I think that’s part of the trend I’m seeing. I know like Google or Facebook, once you get large it’s very hard to keep a centralized data science team, and they are kind of like splitting the role up. And I think that’s a good thing for the industry too, because that means there are going to be more jobs and also jobs that cater to your interests because in the past you probably have to do a little bit of this and a little bit of that. But in the future you can be more focused on machine learning or a product or like other roles.
Kirill Eremenko: Cool. Thank you.
Pablos Holman: I think it’s hard for companies if they don’t already have a culture that values making decisions based on data then getting to machine learning and some of these more advanced tools is probably a ways off. Look, a lot of companies, they don’t make their decisions even with a spreadsheet. They make their decisions by getting in a meeting and sort of baton something around and then deciding what to do and then going off until the next meeting and doing it. And those are the companies that I think you’re going to have the hardest time in, right? Because they don’t … or the key thing to look for is the culture of the company making decisions based on data, and if they are then it’s a pretty natural progression to go to ever more advanced tools and more sophisticated models.
Kirill Eremenko: So if a company isn’t making decisions based on data yet how would you recommend somebody to influence the management and directors to start doing that?
Pablos Holman: I don’t know. That’s not something I’ve ever tried to do.
Kirill Eremenko: Yeah.
Ben Taylor: So that’s something we’ve paid tuition on the last year. It’s been a painful process. So one of the issues we feel like 90% … this isn’t a hard number. It’s just finger in the wind. We feel like 90 percent of AI projects fail. They never get shipped. So far a project to succeed it needs to be shipped and consumed, a validation report. Conversation, that’s not a win. That’s not true.
Ben Taylor: So to get the executive buy in no one wants a science project. And so you really have to … One of the reasons these projects fail they worked on the wrong problem. You need to find a problem that has massive impact. Labeling an image and saying there’s a swimming pool, no swimming pool in this property, is that useful? No. No, it’s probably not useful. So finding the right project is key. And so I think that’s important. You have to have that conversation where there’s accountability on the project. It’s not a science project.
Kirill Eremenko: Love it. Actually I wanted on this point to touch what we were discussing yesterday, Pablos, where you mentioned … I think that that reminded me that … We were chatting with Pablos yesterday and he said that organizations that don’t use data and don’t use AI, machine learning, it’s not their intention or job to go in and change them. These big companies are very slow rolling. Tell us a bit about that, what you described about
[crosstalk 00:15:31].
Pablos Holman: Yeah. I mean, the way I think about it any successful business or institution of any description, it could be a business, could be healthcare, democracy, or education, whatever, any large scale successful institution has partly evolved an immune system, right? And it’s the job of the immune system to suppress risk. And if anything looks like risk it’s change. And this is why where I come from in Silicon Valley the way we think about things we will not try to fix your business. We will not try to fix your industry. We will not even try. We will make a new parallel industry or business and make an end run around you using all the superpowers we get from our computers. And that’s what disruption really means. That’s what you see with Uber and Airbnb and Facebook and these cases were a couple of assholes from Silicon Valley made an end run around a previously massive industry.
Pablos Holman: And what I think about it is if you are in a context where it feels like pulling teeth trying to get somebody to buy into your idea for how to actually make our decisions based on some real data changing people is hard. Changing companies and cultures is hard. And this is why what I look for situations where you have opportunity to run experiments and kill off the failures. And sometimes that’s an entire company. So we’re going to just leave you to die. And we’re going to make a startup that performs better. That’s why we do it. That’s why we have thousands of startups. Every one of them is a little million dollar experiment that we can try and sometimes they outperform.
Pablos Holman: It’s like why TrueCar exists. I mean, it’s not a startup anymore, but eight years ago or nine years ago or something TrueCars was a little startup that figured out how to give people actual real data about what’s going on with car buying and make a decision and make it accessible and easy for people to consume. Complex market data, that’s why TrueCar exists. And it made an end run around what was essentially the old shitty go in and talk to a salesman, see if maybe they’re lying to you, “I don’t know,” and do you want the spray on the bottom of the car to keep it from rusting or how much is that really worth, all these kind of screwy stuff.
Pablos Holman: Now you don’t do that, right? You just know the actual price. You know what it’s going for, you know if you’re getting screwed or not. I mean, those are the kinds of things where I wouldn’t stay too long. I mean, this is a beautiful thing right now. It’s like you guys are in demand. Every business on earth wants data scientists and can’t afford them. So you can vote with your feet, go where people will listen to you, where you’ll be valued, and let those old businesses die. We’ll make new ones. It’s not a problem.
Kirill Eremenko: This is going to completely discourage executives from sending people to this conference. Thank you very much, Pablos.
Mark Skinner: I would like to add that-
Kirill Eremenko: Please do.
Mark Skinner: … if you do want to stay with your company, the lady asked the question, you can get started with your own company’s data. It doesn’t cost more than like or seven dollars an hour to run some jobs iteratively in Amazon or Azure or Baidu or Oracle, whichever cloud provider. The barrier to entry for doing this kind of research is pennies on the dollar. I mean, it really is an acceleration from just 10 years ago. So you can throw some of your own department’s data in a model that you think might give you some insight and just slide it across the table at the next meeting that people are kicking things around. Maybe they’ll kick around your analysis and then you’ll provide value. People say, “How did you get that?” “Well, I spent $25 on Amazon. I hope that’s okay.” If they give you a promotion stay for a while.
Kirill Eremenko: To Mark’s point there’s the saying, better to seek forgiveness than ask permission. Not always true, right? Sometimes you’ll get fired for it.
Mark Skinner: I think that was my Angelou.
Kirill Eremenko: Probably.
Mark Skinner: [inaudible 00:20:07].
Kirill Eremenko: Please.
Audience: One of the applications were GPU gets used a lot is in crypto mining and lately the profitability in crypto mining with GPU has been declining a lot.
Mark Skinner: Thank God.
Audience: So this is a two part question. First part is do you see NVIDIA getting into the ASIC chips business if it’s not already? That’s the first question. Second question is a more general question. Where do you guys see FPGA … I mean FPGA-
Mark Skinner: FPGAs. Yeah.
Audience: … versus ASIC in the future when it comes to either data science or artificial intelligence or machine learning or any of its subcomponents?
Mark Skinner: Okay. Well, the first thing is for people who don’t know crypto mining, the people who were trying to make a bazillion dollars on bitcoin or ethereum or whatnot, they found that they could get a lot of the work done for mining, just solving the equations in GPUS. That caused a spike in the demand for gamer GPUS. And a lot of G-Force cards didn’t go into kids’ gaming boxes. They went into minds that literally sometimes we’re minds with thousands and thousands of computers just running these hashing algorithms. So I’m really pleased that the complexity of the algorithm’s got harder, so GPUS weren’t a cost effective manner anymore. So that allows our number one customer base, the gamers, to get their cards back at relatively a normal price. So the good news is GPUS for crypto currency is a thing of the past. Don’t recommend it.
Mark Skinner: As far as FPGAs and ASICs we kind of already do that. We did it with Tensor Cores for doing accelerated deep learning algorithms and matrix multiplication. We said there’s an algorithm that’s A plus B equal C matrix and we put it into a chip and we put that into our PGU and then we did it again for ray tracing with touring. So we kind of already do those kinds of come take a math problem and put it into hardware. That’s kind of the example of what an ASIC does.
Mark Skinner: An FPGA is a field programmable version of an ASIC, but there are other companies like Intel and others that have their own and they’re saying, “Well, there’s the demand for this type of math to do your data science problem. So let’s compete with NVIDIA. It’s a good news, bad news for us. It means it proved the market so we’re doing something right, but the bad news is like you have choice. But the good news is that you see your choice against our complete stack and platform from A to Z versus somebody who puts out a product said, “This is five times faster than what you can do in a GPU.”
Mark Skinner: But this is only the PCIU card. You have to do all of the development and the sources and the tools and maybe it’s not such a bargain at first sight. So you know, it’s good that we have competition. Somebody else asked me about AMD. They do a good job doing what they do and they make us look good when we come out with something better. So I’m a firm believer of competition. I’m still going to wear and go home and play on my G-Force cards, but I want all of you guys to understand that we are not stopping with just the equipment that we’ve done and the chips that we’ve built. It was a $3 billion a procedure to get Volta, which is the one with the Tensor Cores which you guys are going to love if you get a chance to play with it.
Mark Skinner: So we’re not stopping … we’re not sitting on our laurels. So yeah. I mean, I don’t have a dog in this fight really, but from a fundamental perspective NVIDIA is advancing. Other companies are also advancing on performance related to use the things that are relevant, machine learning. And there’s a bunch of new technologies coming and I wouldn’t worry about that. I’m worried about getting good at framing up the problems you’re trying to solve and creating models. But we’re going to be able to give you an absurd amount of performance in this life. And a way to think about it is for most of your life to date we just make computers faster by making smaller transistors, which is largely what we refer to as Moore’s law.
Mark Skinner: Now we can’t really make smaller transistors because they’re about the size of an electron. And so there’s all these other ways you could make computers faster that had been neglected, that nobody’s been working on. And now we finally have demand to go after them and that’s kind of what Tensor Cores are as an example. But from a fundamental perspective, like massive increases in computational ability are coming, but we need you to do is figure out what the hell to do with it when we give you those machines because it’s all [crosstalk 00:25:27]. Yeah.
Kirill Eremenko: That’s cool. Thank you. While we’re on this cryptocurrency I want to raise the topic of blockchain because the panel is emerging technologies. It’s one of the emerging technology we’ve seen really spike at the beginning of this year and now it’s kind of like floating there. And so I think it’s pretty high up still, even though the cryptocurrencies are dropping off. What are your thoughts, Pablos? Maybe let’s start with you. What are your thoughts on blockchain? Does it have a future?
Pablos Holman: Well, I worked on cryptocurrency in the late 90s. What we were trying to do is … what we saw was that there were a lot more people coming on the internet. We were going to have to put some effort into architecting this thing to preserve our values. And specifically where I come from, what we believed, is that nobody should get an asymmetric advantage on the network. We believed it should be a level playing field for everybody. And that means you don’t get to decide what I do on the internet. So we tried to build protocols that would support that. And that’s a big difference from building services, right?
Pablos Holman: So in the 90s when people started to get online they were leaving this big walled garden called AOL, which is a service that controlled what you could do and how you could transact, right? TCP/IP is that not a service. It’s a protocol and anyone can jump on that network and do whatever they want. And then once we all got on to the internet we ran into a big walled garden service that’s like Facebook and Google. So what’s happening now is people are starting to clue in that these services don’t represent their values, that they’re messing with your ability to do what you want, transact with who you want, communicate with who you want, those kinds of things. And so what needs to happen is we have to replace them with protocols.
Pablos Holman: And so that’s kind of the philosophical view of how we been doing it. So we started with systems to try and make it possible for everybody to use the internet evenly. That’s where things like … We have some success stories, so like BitTorrent. Lots of companies hate BitTorrent, but it made it so anyone could share massive amounts of data online, not just the companies who had huge infrastructure. We made the onion routers, which is what also makes the dark web very controversial. Lots of people hate it, countries hate it, but you can’t shut it down because these are protocols, not services. And we made bitcoin, which again is the first currency ever for humans that’s not a service, it’s not centralized, it’s protocol. And that’s why despite lots of people hating it, lots of countries hating it, you can’t shut it down.
Pablos Holman: And so what blockchain has done is given us the ability to create a decentralized mint for the first time. That was the whole point. For everything that you don’t like about Bitcoin we don’t care because we’ve architected the next hundred versions already for every possible problem. Cryptocurrencies are here to stay. I’m not telling you whether to invest in them. I don’t give a shit about cryptocurrency speculation. I care about the fact that it makes it a level playing field for everybody. No one can stop me from transacting with you. And that’s the point of these systems.
Pablos Holman: Now, blockchain is only one tool in that tool kit. We have a huge toolkit in the crypto toolkit of things that you could use to architect protocols. And that’s very rarely done. It’s hard work but we’ve attracted an entire, because of the success of Bitcoin, we have attracted an entire generation of coders to the crypto toolkit and now we can go reinvent some of these things. How hard would it be to make … If you want to bitch about Facebook messing with your feed and controlling what you see, well, how hard could it be to make a Facebook feed that’s decentralized where you get to control the knob that decides what you see? Same thing with Instagram, same thing with Twitter, all these things. We need to change them from services to protocols, and we have the tools to do that at our disposal. So blockchain gets all the attention, but the truth is that’s just the newest tool in the bag.
Kirill Eremenko: Perfect.
Pablos Holman: That’s long winded. I’m sorry you guys. I’m really not good at panels. Ask them some questions. Ben, Rachel.
Ben Taylor: So I think it’s interesting with some of these crypto blockchain opportunities. Some of them go after Amazon’s business. So they’re specifically a Mechanical Turk offering or there’s S-3 offering and kind of the internal joke was someone needs to come out with a deep learning offering where they can be in a pitch and say, “I’ve a startup. It’s crypto, AI, deep learning. I need $100 million for a seed.” But there is a practical application where companies have a really hard time loading when it’s inference time. They want massive service level agreements. They want distributed computing. So if all these gaming cards, if you had crypto where you get paid one penny for that prediction and you get paid 10 pennies, I think there’s a huge opportunity for someone else to go build out a deep learning infrastructure on these gaming cards where you pay kids with crypto and they get a little bit more than they get with bitcoin. And companies are happy. Of course, you’d have like data privacy and stuff. But it’s funny.
Kirill Eremenko: Cool. Great. Yeah. Rachel, any comments?
Rachel Wang: The way I see blockchain is that I feel like from the technology point it’s at a pretty much horror stage. A couple of years ago calculating in blockchain is expensive in terms of calculation power, but nowadays it’s actually not expensive anymore. I feel like it really can solve a lot of question. But the real question is where does it land, because the way I see it as-
Kirill Eremenko: Pablos.
Rachel Wang: … Pablos said countries don’t like it, government don’t like it. Who likes a decentralized financial system? But that’s definitely something blockchain can already deal with. So I’ll wait and see where it goes.
Kirill Eremenko: Got you. Thank you. All right. Next question. Let’s throw the box over here front. Nice. Over there. Yeah.
Audience: So my question is earlier this year Facebook and Google turned off their AI because it started to create its own language. My question is from your perspective what were the threats there? What are the even potential opportunities of AI creating its own language and what do you see all that piece going?
Ben Taylor: So I hate the media when it comes to AI. So they’ll pull on … if they need to sell something they’re going to sell FUD. So it’s fear, uncertainty, and doubt. And so I think people in the AI community would kind of laugh at that story where the media just, “Here’s another thing we can pull on,” and we see that time and time again. Think of Tesla. Anytime Tesla there’s a fatality on a self driving car that’s all they want to talk about. But Tesla has already surpassed the death rate per 100 million miles driven for humans. So really the thing we should be talking about is, “Good. One person died instead of 5 or 10,” or I don’t know what the number is.
Pablos Holman: No, there’s like one person dies in a Tesla, every day 3000 people die in conventional vehicles. They don’t make the news.
Mark Skinner: I did not die in my Tesla driving down here. I want to make that clear.
Ben Taylor: Yeah. So I would say that’s a FUD example where the media is pushing something that makes me angry.
Pablos Holman: Yeah. You guys look, we don’t have artificial intelligence. It does not exist. All we have is machine learning, which is super cool, but extrapolating from that that robots are going to turn up any day now that want to turn us all into paper clips is irresponsible and lazy. It’s bullshit. We have no reason to believe that’s going to happen. We have no evidence that they’ll be malevolent. And I just think these FUD stories sell and you guys know better. Ignore all that shit. There’s nothing scary going on here.
Rachel Wang: Right? [inaudible 00:34:13] I don’t think Facebook or Google actually turned off their AI. I don’t think there is any tech giant who is not actively investing in AI or wants to get really good in this field, but also the definition of AI is so broad. Are you talking about voice recognition? Are we talking about graphical recognition or are you talking about some robot actually having a sense of self existence? And I think that’s really, really far from the technology we’re today. But stuff like writing an automatic report that’s already happening. Yeah. Driving your own car. That.
Mark Skinner: I didn’t drive. I just rode in.
Kirill Eremenko: Thank you. All right. Next question. Let’s go somewhere in the middle of there. Throw it. Yep. Nice.
Audience: Hello? Going back to the topic of blockchain, one of the intentions or the application of blockchains is to eliminate the middleman by doing a peer to peer connection.
Pablos Holman: Digital [inaudible 00:35:20].
Audience: For example, let’s say in the case of Uber you have a peer to peer connection, but then who really owns the platform or is it going back to Pablos’ point going from service to a-
Kirill Eremenko: Protocol.
Audience: … protocol. Right. So who owns it from a commercialization perspective?
Pablos Holman: Well, the point of … So just using cryptocurrency as an example, obviously you could build other things on a blockchain. Historically, you had currencies like cash, which are single entry bookkeeping. That means I hand you a dollar and now you got it. And I don’t. The transaction settlement is completed. Every other transaction system besides cash and barter and trade in gold is double entry bookkeeping. That means, “No, I’m going to decrement my account at my bank. And you’re going to increment your account at your bank and we’re a better pay a bunch of jerks in the middle to go audit everything and make sure that no one’s cheating.”
Pablos Holman: That’s expensive. That’s a drag on an economy. And this is just economic transactions but any kind of transaction is being taxed by that drag. So what we’re trying to accomplish with protocols like Bitcoin is you get instantaneous settlement but you get it what we now called triple entry bookkeeping. And that’s what shared ledger means. So now instead of me having a bank and you having a bank were all aimed at the same database. And there’s only one place that it has to change.
Pablos Holman: And so yeah, there’s nobody who owns that. Nobody who controls that. That’s what decentralized means. And that’s why nobody can shut it down. Hopefully that answers your question.
Kirill Eremenko: I think that’s pretty clear.
Pablos Holman: Yeah. We want to apply that to lots of things besides financial transactions.
Kirill Eremenko: Next question. Let’s go left, [inaudible 00:37:20] there. Stand up please.
Audience: Yeah. Do you want me to stand up?
Kirill Eremenko: Right.
Audience: How far do you think we’re from just replacing banks with like a bitcoin kind of system? Is that ever going to happen? Probably not but I thought it was-
Pablos Holman: Some people already have. I mean, who here likes their bank? You do?
Mark Skinner: I like San Diego County Credit Union.
Pablos Holman: Okay. That’s not a bank, but also, look, banks suck, right? They aren’t actually doing anything for you. They’re a necessary evil for you to be able to transact. You’re not getting interest off of deposits. They’re charging you fees. There a pain in the ass every time you interact with them. I hate banks. I try to avoid them at all costs, and I can’t because I’m trying to buy and sell shit. So banks are kind of an example of an industry where they haven’t figured out that how to be, despite all their claims, customer focused, right?
Pablos Holman: So I think their days are numbered and they know it. Now cryptocurrencies have a lot of additional challenges, because governments derive a lot of power from issuing a currency. And so it’s hard to ge a lot of progress on this, but that said there’s a lot of different countries and a lot of different governments and some that don’t even issue a currency, and all it’s going to take is for a couple of them to start making it fungible, decide, “Oh, we’re going to issue cryptocurrency instead of paper.” Well, I think these are very irreversible trends.
Pablos Holman: So I think we will ultimately end up on cryptocurrencies. You’ll probably end up with competing cryptocurrencies from different jurisdictions and some of them will be issued by governments and some of them will be decentralized like Bitcoin, but they’re important because they can reduce that transaction cost and get us down to zero percent transaction cost. And that’s good for every economy.
Kirill Eremenko: All right. Let’s do a machine learning question. Who has a machine learning question? Put your hand up. Let’s go over there at the back.
Audience: Where do you think reinforcement learning is going in relation to supervised learning? Where do you think is the application bound? Meaning does reinforcement learning have more applications versus supervised learning and the long haul?
Kirill Eremenko: Rachel?
Rachel Wang: Yeah. I think reinforcement learning definitely has a better future than supervised learning because you cannot imagine how many restrictions we’re currently having because we don’t have the training data set for supervised learning. So I definitely think that’s the future. And algorithms are kind of like evolving in an incredible speed today. So I would probably say maybe in three, five years nobody is talking about supervised learning anymore. I think that’s totally possible.
Audience: That’s what I was thinking. Thank you.
Kirill Eremenko: Ben. Do you agree? Cool. Yes, please over here. Throw it far.
Audience: So you’ve all had a lot of failures. Pablos talked a lot about his company’s failures.
Pablos Holman: Yeah. No company I’ve ever worked for us still exists. That’s why I’m on stage.
Audience: What do you do with all the failures? Do you make them open source for people to use or do you just throw them in the trash bin? Or what are you doing with these failures?
Kirill Eremenko: Ben, let’s start with you. You’ve got a lot of failures as well.
Ben Taylor: The question is around the actual products like open sourcing. So you have failures and mistakes you’ve made, but when it actually comes to catastrophic failure for a company like I feel for us if we failed our advisors would tie a bow on us and we’d go sell to someone. So we would see that as a failure. But I don’t know. That wouldn’t be open source if it was a complete failure. We’d probably would be so depressed I don’t know if we were just open source the code. Probably delete it.
Kirill Eremenko: But that was interesting because we were talking about in the car yesterday that your company and your research is so advanced at this stage that if you fail financially like you can’t meet the budget, what not, and your bank account goes down to zero you can still sell it, right?
Ben Taylor: Yeah. So it’s kind of a wild … So we went from being paid well to being paid zero. And I’d been-
Kirill Eremenko: When you quit your job.
Ben Taylor: Yeah. When I quit my job. So we had meeting with the bank, talking about, “Well, how can you take more money out of our house? Do you have two years worth of self employment income?” “No, but I’ve got an LOI.” And the banks like-
Kirill Eremenko: What’s an LOI?
Ben Taylor: “We can’t do anything with that.” It’s just a letter of intent. So we have an acquisition offer. But the funny thing is when you look at the upside with these startups AI startups aren’t ridiculous. You watch Silicon Valley, Silicon Valley times 10. They are insane. So it makes sense for you to fail financially. 401K, if you think you can see daylight on an acquisition … I don’t know. It’s a weird reality.
Kirill Eremenko: Yeah. Got you. Rachel, what about your failures?
Rachel Wang: I don’t want to talk about my failures, but I want to follow with what Ben was saying about companies. Even if your bank account turns zero we have bought some failed startup into TrueCar, and it’s not only as a bank account, right? Your code base, your talent, and your experience in a certain very specific field, these are all your assets. So, yeah, in a sense I don’t think it’s a failure. It’s just some assets that hadn’t turned into cash yet. And you just have to find a way.
Kirill Eremenko: Got you. Mark, what do you do with your failures?
Mark Skinner: Well, I was part of a company here in San Diego that was going to be bought out by one of the large OEM builders. And that failed. So we found out a way of not repeating that mistake with the next company and the next company. So we tried to really do things more ourselves and not look for some big rescue and bail out. But within NVIDIA some of the products that we’ve launched, sometimes they might’ve looked like a failure for one person, but another group say, “Wait a second. I need a lower power version of that. So let me take those GPUS that didn’t make the cut for the top of the range super duper product. Hey, let’s rebrand those and put them in something and affordably priced them for a job that we do need done.”
Mark Skinner: So your chocolate chip cookies with nuts that you don’t bring over to your allergy friends you give them to somebody else. So it’s not always a bad thing to have a failure. You know one way of not doing something. So there’s some lessons learned.
Ben Taylor: I just remembered a possible failure. Euphora.com, have you guys heard of Euophora.com as an AI company? They open source their platform. One of the founders … The media is not pushing failures. So you see all these winds startups just kind of dive behind the scenes and you don’t really know. So I noticed one of the founders is working somewhere else and they had a open source and I don’t know. Does anyone know what happened to Euphora, and did they open source their secret sauce in the open source community? I have no idea.
Kirill Eremenko: Yeah, I guess failures are not as fun to talk about, but there’s a saying that you learned from success but you learn 10 times more from failure. So we had a [crosstalk 00:45:36] discussion with someone out in the hallway that how would you feel if you fail, other people are judging you and so on? And one of the attendees made a really interesting comment that as long … if you fail and you don’t learn anything then I will judge you. Not I but that attendee said there’s a reason to judge you. But if you fail and you learn that’s just part of the process.
Pablos Holman: It’s an important part of the process. I mean, it’s big part of why it’s awesome to be in America on the West Coast in California because we’ll give you the benefit of the doubt. You can fail and be like, “Cool man. What do you do now?” If you have a startup in Italy and it fails they open a criminal investigation. You’re never going to get another chance. They have a long memory. For generations they’ll remember that you failed and tell everybody.
Pablos Holman: So for us to be good at what we do we have to be able to try things. And by definition you get a lot of failures. I mean, I told you guys yesterday I get about 999 failures for every nuclear reactor I get. And you’ve got to be willing to accept that and not just personally, but for everybody else. I think you guys should do the craziest thing you could possibly bring yourself to do at any given moment. Do that. That’s what we need from you. We don’t need more people conservatively checking the balance on their 401K every whatever. I don’t think I have a 401K. I might need to sleep on your couch.
Audience: With that though, I’ll do a little crazy. So sensor wise-
Kirill Eremenko: Could you-
Audience: … the integration between sensors and deep learning for example [inaudible 00:47:32] visual sensors or hearing or other types of sensors, whether it be … who are the leading companies in the hardware world, besides NVIDIA, that are helping to bridge the gap between getting that information that comes from the outside world directly into these neural networks? Who are the teams that are doing that right now?
Kirill Eremenko: Do you guys understand the question?
Ben Taylor: So you’re talking about for manufacturing just sensory inputs into deep learning?
Audience: It could be sonar information, it could be what cars are picking up on LiDAR. It could be-
Ben Taylor: So I think it’s all scattered. So you look at these different companies that are doing stuff, so there’s a group in LA. They’re called the tech twins, these two identical Israeli tech twins. So they’re just focusing on radar. So with a localized radar consumption, seeing humans through walls, seeing that your baby is still alive, they just do that, and then you’ve got different applications. I don’t really see a big general provider, but they are showing up.
Ben Taylor: I feel like manufacturing is a little behind with some of the sensory stuff, but there are companies coming online that that’s all they do. You’re going to see a lot happen I think in the next couple of decades with the stuff that we consume.
Kirill Eremenko: Yeah. And, Ben, you mentioned that most of your chips are NVIDIA, the ones that use at Ziff.
Ben Taylor: Yes. Yeah, we use all NVIDIA GPUs for training and then we use CPUs for inference.
Kirill Eremenko: Okay. Got you. All right. Next question. Oh, they’re on the right please. Good catch. Nice.
Audience: I’m going to go back to blockchain again. Blockchain has a huge potential to change the blockchain.
Kirill Eremenko: So many questions about blockchain.
Audience: Yeah, I know. So it has potential to transform the world economy and right now even though it’s great but it still have limitation. So what is your thought on the scalability? What might be the solutions that we can solve the scalability of blockchain? And then second thing is there are big bangs, big government, and maybe organizations trying to put out their permission blockchain now. What is your thought on permission blockchain verse permissionless blockchain? Thank you.
Pablos Holman: Well, first of all, the whole point of using a blockchain for a cryptocurrency is to get rid of the centralized mint, to get rid of the guy who’s issuing the currency because historically he’s always been able to mess with the value of the currency. And you can see this with basically every currency on earth. The US dollars is sort of the least screwed with and even ours gets screwed with by the Fed. So that’s why we want a decentralized mint. It’s really easy to solve scalability problems and performance problems in bitcoin if you get rid of that. Build infrastructure for the mint, you’re done.
Pablos Holman: So coming up with next generation protocols that preserve the decentralized nature of bitcoin and also meet the performance requirements we want going forward. There’s lots of projects on that. Zcash is one of them. Bram Cohen who made a BitTorrent is now working on one called Chia where he’s trying to switch the model from proof of work, which is what we’re doing now for the blockchain on bitcoin to a model where it’s proof essentially of storage. So you waste storage resources instead of wasting processing resources. And that cuts down on the power consumption by a lot. So that’s a good idea.
Pablos Holman: But there’s a lot of things we want to do that bitcoin is not a good choice for things like microtransactions. We’re talking about being able to pay somebody like a 10th of a penny to run some calculations on a GPU. Well, you certainly can’t do that practically with bitcoin right now. So you want to be able to have different cryptocurrencies that you might use for different … You might use bitcoin as a value store, switch to some other currency that you’re going to use to pay people microtransactions every 10 nanoseconds, and then when you accumulate a bunch of that you switch it back to bitcoin, which is a stable currency. I know. Hard to believe Bitcoin, stable currency, but that’s probably what it’ll end up being.
Pablos Holman: And so there’s just work to do to advance on those protocols, but there’s massive potential in that, lots of work being done, lots of possible ways you could go with that. So I think these are technical problems that we can solve.
Kirill Eremenko: Got you. So in terms of … just to give a perspective how that indeed blockchain is emerging technology where we have so many questions on it the amount of power consumed by blockchain miners worldwide right now is more than the mileage of power Iceland consumes on an annual basis. That’s how much.
Pablos Holman: Iceland is tiny.
Kirill Eremenko: It’s still a country.
Ben Taylor: Lots of minings in Iceland.
Kirill Eremenko: All right. So we’re going to stop off the blockchain questions. Okay. Non blockchain question please. Let’s throw it right over here to the front. Yep. Big throw. Whoa!
Ben Taylor: One less data scientist.
Audience: All right. I [crosstalk 00:53:28]-
Ben Taylor: You’re allowed to catch it if you don’t want to answer ask a question.
Audience: So my question is in each of your respective fields what technologies are you most excited about and why?
Ben Taylor: I’m the most excited about GANs. I see huge applications in business for encoding and some of the … Yeah, I think you’re gonna see a lot more GANs showing up in business.
Mark Skinner: Everybody know-
Ben Taylor: A GAN is an adversarial network. It can generate fake content. So it’ll study your data and if you want to see a dog it’ll produce a fake dog or the image I showed up my slide, the fake human faces they’re pretty awesome.
Mark Skinner: Yeah. They’re so realistic too.
Ben Taylor: Yeah.
Kirill Eremenko: Mark.
Mark Skinner: I’m most excited for, well, it’s kind of self indulgent, but the self driving ability of my car. I got to play with it this week. I’ve only had it for three weeks. So bear with me. It’s still a new toy. But the fact that I drove to the airport the other day and I only had to change lanes by doing this twice to get to the freeway where I got off to get to the airport. I mean, I’m most excited about other people who don’t have a facility for driving or the ability to drive to have that mobility, and we have so many different car companies sprouting up to do that Uber and Lyft and others and Waymo, a subset of Google. I’m most excited about people being able to be a mobile.
Kirill Eremenko: Got you. Rachel?
Pablos Holman: Yeah. And if you want to save lives in America-
Mark Skinner: Yeah, the benefit of that.
Pablos Holman: Get rid of human drivers.
Rachel Wang: Personally, I’m really, really excited about VR. Every night before I go to sleep I’m like, “Come on, I want to play the sims in VR,” and will build my own kingdom and get in it. But-
Ben Taylor: [inaudible 00:55:23].
Rachel Wang: … from the business standpoint, I’m very excited about cloud computing. It’s already happening, but it’s still amazing to see how it’s transforming business like us. We used to maintain a very expensive data center and our own server and it’s costing us millions of dollars and it fails constantly. Right now you’ll be amazed by how reliable and cheap AWS is, and also for data science AWS is [inaudible 00:55:56], but they are putting out this thing called SageMaker where you can do tons off models on it and with cloud computing. So it’s pretty amazing. Yeah.
Kirill Eremenko: Pablos?
Pablos Holman: What technology might excited about? I’m obsessed right now with figuring out how to make sewing robots.
Kirill Eremenko: Sewing?
Pablos Holman: Yeah. It’s like if you look at all those factory dungeon jobs in Asia, a lot of them are sewing. And no miracles are required. If we can make self driving cars and we can make surgical robots how come I can’t make a self driving sewing machine? And I think that’s an interesting one because right now that entire industry apparel is $2 trillion a year. Every human on earth is a customer. And if you could build a sewing robot you could re-architect that entire industry without chasing the cheapest labor on earth the way they do it now. So I don’t know. I’m obsessed with that.
Ben Taylor: I just thought of another idea of longer term … So right now we’re all constrained by our ability to type and eventually a lot of us will get carpal tunnel, a lot to worry about that. But I think with automatic speech recognition breakthroughs we’re looking at right now were going to see speech programming and the next couple of years. I’m not talking about toy, “Ha ha ha look what someone did with Python.” I mean, useful speech programming where you’re writing Python and you’re kind of coming up with some custom shortcuts. So I think the dream in the next 5 years, 10 years as we develop new things a lot of that will be more visual and audio and less tactical like typing. So I think I’m excited about that. I think it’s very doable. I don’t think it’s science fiction. I think it’s something that’ll happen sooner than later.
Kirill Eremenko: Perfect. All right. We’re going to wrap it up there. So a good question to wrap up on. Let’s give it up for our panel. Thank you guys.
Speaker 8: Thank you guys [inaudible 00:57:56].
Mark Skinner: Yeah, it’s good talk.
Kirill Eremenko: Thank you. Thank you [inaudible 00:57:59]. Pablos? Okay. So there you have it. So that was the panel from Data Science GO 2018 on emerging technologies. If you were there then I hope you enjoyed this recap. If you missed it then I’m sure you picked up lots of interesting ideas and thoughts from here. We talked about blockchain, AI, deep learning, machine learning, disruption, startups, and many, many more topics. I would love to know what’s your favorite one was.
Kirill Eremenko: Leave your comments in the show notes at www.superdatascience.com/201. We’re kicking off our next hundred of episodes very strong and I look forward to seeing you for the next 100 episodes. And of course if you enjoyed this, if you want to be part of this next year at Data Science GO 2019 then head on over to www.datasciencego.com and guess you’re super early bird tickets there. Thanks so much guys for being part of the Super Data Science podcast. I look forward to seeing you next time, and until then happy analyzing.