(background music plays)

Welcome to the SuperDataScience 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.

(background music plays)

Hello and welcome back to the SuperDataScience podcast. We have got a very exciting guest again on the show, we've got Ben Taylor. So Ben previously appeared in episode 29, make sure to check that out if you haven't yet. And also, good news, this episode is available in video as well as just audio. So if you're near a computer, or if you have access to your phone and you have time to check out a video on there, then go to www.superdatascience.com/85 and you will find this exact recording, but in its video version and you can check us out there as well. At the same time, I know that you might be listening to this while running, or driving, in that case, it's totally cool. Sit back and relax and just enjoy this audio version of the podcast.

So what we are talking about here, a very interesting conversation that we had with Ben, we talked about many things including polyphasic sleep, intermittent fasting, starting startups (Ben just started his own startup with some business partners called Ziff), patents and trade secrets, and the future of technology, and of course, our favourite subject with Ben, drones. We talked about different applications of drones.

So a very exciting episode ahead with one of the thought leaders in the space of data science and artificial intelligence today, it's something that you definitely cannot miss. And make sure to stay tuned up to the very end, because there we will share with you our secret tips for getting jobs in the space of data science. Can't wait for you to dig in, and without further ado, I bring to you Ben Taylor, Chief Data Officer at Ziff.

(background music plays)

Welcome everybody to the SuperDataScience podcast. We've got a super exciting guest again on the show, Ben Taylor. Ben, welcome back. Great to have you here.

Ben: Thanks. Thanks for meeting up again.

Kirill: So what have you been up to? Since the last time we spoke, you've done so many new things!

Ben: Yeah, so I'm now full time on my startup, which, for anyone out there, including yourself, doing a startup is a lot more stressful than you think it's going to be, because you have different knobs you can turn, so one of them would be people. You may not have money to hire more people. And then you've got money, you can inject money. Then the other resource you have is time, which usually means you end up working ridiculous hours. And so one of the things we've been doing to work these 80, 90, 100 hour weeks is we've been doing something called polyphasic sleeping.

Kirill: Oh, wow, no way, that's so cool!

Ben: Yeah. So I've been doing it for the last year, and I love it. So I go to bed at 11, I wake up at 2:30 am, or 3 am, and then I work. And then I need to sneak in two 20-minute naps during the day, and sometimes I can get away with something like 13 or 14 minutes. So if you really optimize your naps, you can do it. And then one other thing we’ve been doing which people think is absolutely crazy is we’re doing intermittent fasting. You’re thinking, like, “Well, is that because you don’t have enough money coming to the startup?” No, no, we actually intentionally do it. We do it every week and we’ll fast, and we won’t eat for sometimes two days or longer, and the thing I really like about it is the first day sucks, it’s terrible, you’re grumpy. But the second and third day you’re no longer hungry.

Like right now, I didn’t eat yesterday so right now I have a heightened sense of alertness so I really like it when I’m programming. That’s kind of the body hacking. So I’ve been doing this startup, we have customers, we have a really good pipeline, we haven’t been able to hire our third employee yet, but we think we’re close and we’ve been talking to salespeople. So the startup this year will land somewhere between supporting our families, which I think we’re barely there, to very successful. So, somewhere in between, hopefully it’s closer to this and not that.

And then one of the things I was mentioning to you before the call, I did get invited to go out to D.C. to present to the EEOC. They’re the department of government in the U.S. that oversees racism and sexism and bias in the workplace and in the community. They do a lot of the discovery for the class action lawsuits and they carry them all the way to trial. I presented to them on racism and how AI can actually fix racism. I love this story because it goes across so many different verticals and disciplines, where there’s no amount of training you can give a human to fix this problem. And everyone is racist, myself included, and you, and your grandma. Everyone is racist, and as much as we try to resist that and say, “I’m not racist,” if you put 100,000 applicants through me, you’ll see it can’t be balanced. It’s not.

You’ll have some stereotype based on your upbringing and your personal background. But AI can fix it. And one last point for that story which I found really discouraging and it kind of amplifies this point, in the U.S. we have allowed people to do criminal background checks on you. So let’s say I’m a data science startup and I want to hire you – which is probably true, I want to hire you (Laughs) – and I find out that you’ve gone to jail for methamphetamines and robbery….

Kirill: Which is not true, just for the record. It’s not true. (Laughs)

Ben: I can ask you that during the job screening and legally, in the past, you were required to say yes, you’d mark the felony box and you wouldn’t continue in the application process. They’ve been discouraging that because they’ve noticed that it causes adverse impact against minorities because in the U.S. we have a terrible prison system cycle. I think if you look at the numbers, it’s really depressing how many people are thrown in the jail in the U.S. That has a very heavy skew towards blacks. So they’ve been encouraging people to remove the box, remove the felony check, because you having a felony three years ago actually has no bearing on your ability to be a good data scientist. That sounds like a good idea, right? I think it sounds like a good idea. Like, this shows you how terrible humans are. What they’ve noticed is, by removing the box, it’s made the racism worse.

Kirill: How’s that?

Ben: Because before, you were allowed to ask me if I have had a felony. Now you’re not, so you’re just going to assume I have had one. Isn’t that terrible? This is why we need AI. The way we do it with AI is normally with AI, we’re so focused on teaching the computer to remember, be the generalist, so the approach that HireVue came up with, and lots of other people do that, is you actually teach it two things at once. You teach it to remember what you care about, but at the same time, you’re teaching it to forget. So you’re rewarding it to forget race and to forget gender.

And if this is confusing to people, the simple way to explain it is if I gave you a resume, you could look at the name and where I’m from and you could predict my race and gender. But if I start taking that away, eventually I can remove enough features that you have no chance of predicting my race and gender and now you’re focused more on my credentials. In a similar fashion, that’s what a computer does, but on a much deeper scale.

Kirill: Okay. That’s really cool. So many interesting things right there. I love this. This is like a very interesting dynamic and saturated start to the conversation, so now we’re going to dig into these. And just for those out there who haven’t heard the first podcast with Ben, Ben is a super data scientist, like a genius — I shouldn’t say ‘super’ because that’s the name of the course. I really think he’s a genius data scientist, he’s got amazing articles on LinkedIn, you guys have to check them out. You were the head data scientist for HireVue, is that right?

Ben: Yeah, I was the Chief Data Scientist for HireVue and now I’m their Data Adviser.

Kirill: Because you have this startup, right?

Ben: Yeah. I’m the Chief Data Officer for ziff.ai.

Kirill: Ziff with double F?

Ben: Yeah, double F.

Kirill: Okay, ziff.ai, you’ve got your own startup. Yeah, that’s some really exciting stuff. Of course, I want to drill more into that, but I would love to start with the polyphasic sleep. That’s so cool, because after our podcast number 2 with Hadelin de Ponteves, who is my business partner, a lot of people —because he also does polyphasic, he actually does three hours a day and then also naps.

Ben: Oh, yeah, I love that.

Kirill: He’s been doing that for like three years. And a lot of people came back and some people said that’s impossible, others were just like, “Take care of yourself. You’re going to die. Be careful. This is insane, this is un-human,” and now we’ve got an amazing example of a second person who not only does just that, but also you do the intermittent fasting, which is another stress on your body, and you seem to be fine. So tell us about that. What made you start this polyphasic cycle?

Ben: It was actually suggested to me from HireVue CTO Loren Larsen. When he first told me about it, I thought it was stupid, I thought it was crazy. I didn’t really pay attention. But then later, when I started to become really stressed for time, because I’m working full time for HireVue and I’m working equivalent full time on the startup, as much as I can, I kind of came back around and realized that “Wait, you’re getting time for free? So there’s a Ben in the parallel universe that has 15 fewer hours per week than I do now? I want to have the 15 hours of extra time.” So I talked to him again about it and I tried it. The first week was really rough, I felt impaired. When I was driving, I was like, “Ay...” I should not be driving right now.

But I figured it out and I feel great. The trick is you have to be really strict on your schedule and you have to perfect your naps. So when I sleep at night, I sleep on my stomach or my side, but when I take my naps, I always sleep on my back, my mouth is probably open, I probably look like a fool because I need maximum oxygen, and then I have ear protection, and I use a sleep mask so there’s no distractions and then my alarm will wake me up in 15-20 minutes depending on the day and the priorities.

Usually it’s really easy to justify a 15-minute nap. HireVue actually has a nap room, so we have a nap room for that, I think Google does that too, but if it’s busy or if you don’t have a slot, I’m fine to go sleep in my car for 15 minutes and come back. But it’s amazing how refreshed you feel in 15 minutes. If I sleep longer than that, if I sleep longer than 20 minutes, there is no gain and I actually feel worse. So let’s say it’s the middle of the day and I took a 2-hour nap. I would feel like crap. I treat my naps like going to the restroom, so I have number one, number two, and then I have my brain drain, which is my nap. I love it and I feel great, I feel better than I normally feel. Most people feel groggy in the afternoon. Like, you’re not going to invent the light bulb at 3:00 P.M. Maybe you will at 9:00 A.M., but not 3:00 P.M., so having a short nap brings me back.

Kirill: That’s really cool. And do you still drink coffee?

Ben: Yeah. Loren, he was saying to eliminate coffee and caffeine, but I still do that. I don’t notice a problem with it. I like to eat after my naps because that kind of wakes you up a little bit.

Kirill: Yeah, not before the nap. Then you don’t restore your brain.

Ben: Yeah. But the one ‘gotcha’ that is not very well-known is what the long-term consequences are. Because there are long-term consequences with sleep deprivation where you’re more likely to get Alzheimer’s due to your brain—your brain actually cleans itself. They’ve done MRI studies and you can see your brain is cleaning itself when it sleeps. So the assumption is you’re not doing that well enough, but the problem is there isn’t a big enough group of polyphasic sleepers.

I don’t have any sleep deprivation side effects at all. When I started I did. I had major ones. But once I figured out the naps, I feel great. And the one selling point here is there’s no one competing for your time from 2:30 A.M. until 6:30 A.M., there’s no one banging down your door demanding that time, so it’s great for you to work on your passion projects.

Krill: That’s so cool. What about for people with families, for people who have kids or—

Ben: That’s why I do it. I have a family, I have three kids. They actually don’t notice. The reason I started doing it is before, I had so much pressure with my startup that I had to make my wife be a single mother essentially, where in the evening I was working and I’d have to tell her, “I’m so sorry, sweetie, can you put the kids to bed?” again and again. But now, from my 5:00 P.M. sharp until 9:00, I’m a fantastic dad. I’m 100% dad where I can put the kids to bed every single night, I’m not stressed out because I’m going to wake up tomorrow at 2:30, work on whatever I need to work on, and I still get like 9:00 to 11:00 if I had to kick off some job on the cloud. Like, usually we’re running stuff through the night so it’s important for us to queue up our jobs before we go to bed.

So I’d recommend it even more for people who are family-oriented. It’s totally worth it for my wife to let me sneak in a nap in the afternoon for 15 minutes. That’s nothing. I think I spend more time going to the bathroom, especially with your phone, right?

Kirill: Yeah, totally. That’s really cool. I’m glad you figured it out. From my experience, for those listening, I tried, I don’t do it consistently, but after I got inspired by Hadelin, I did it for two or three weeks. It wasn’t as extreme as yours, I did like 5 hours of sleep every day and then a nap. But then one weekend on a Sunday, I clearly remember for some reason I was just put into circumstances where I had to have unwarranted sleep for two or three hours in the middle of the day, and then after than my whole next week was ruined. I was so exhausted, I didn’t know what was going on, I was sleeping 9 hours a day, my body was not responding, I couldn’t think or anything. So you’re right, you have to be very, very strict about it. As soon as you break the pattern you’re going to be in trouble.

Ben: And it’s okay to have—like, it’s rare now, but maybe once a month I will sleep an 8 or 9-hour sleep just to kind of see what that’s like. But I can usually recover pretty quickly. Even if I miss a nap, I can get back into it. The other thing I’ve noticed is my self-esteem is tied to my productivity, so is my overall happiness. The other thing we didn’t talk about is I’m a backcountry skier. A lot of people think I’m crazy, but crazy is relative. So if I’m not waking up at 2:30 to go work and program, I’m waking up at 2:30 to hike some mountain somewhere and go ski it and then I’m back home before most people are awake. So that’s huge for me, for my overall mental happiness. But yeah, you should definitely get back into it and just try to commit to it.

Kirill: We’ll create a little club – data scientists who do polyphasic sleep cycles. That’ll be fun.

Ben: Yeah. If we’re in the same time zone, we could slack each other at 3 in the morning.

Kirill: Yeah, totally. Let’s get back into the data science side of things. We’ve given people a bit of inspiration to test out the polyphasic sleep cycle, especially, given that you have a family, that’s even more inspiration, because a lot of people think you have to be single to do it. But let’s talk a little bit more about data science. Last time we talked a lot about deep learning, about how you use it at HireVue and so on. Tell us a bit more about your startup. What do you guys do?

Ben: So, the startup, what we’re doing in there, this kind of goes back to one of the main reasons I left to do this full-time, is I felt there really wasn’t anyone in our space that was doing this very well, with helping a lay person get state-of-the-art AI into production. These big companies are still a little confusing for engineers. Our goal is to be the simplest way for you to get world class machine learning, mostly deep learning, into production. We have these workflows and you essentially gather the data and then you get a model that can be deployed on the cloud.

But the fun thing that I’ve learned doing the startup is once you really get neck deep in these production deep learning models, your appreciation of size changes dramatically. Before, I thought ImageNet was impressive. So if you came to me and said, “Hey, I spent the week solving ImageNet and 1.3 million images on my own deep net,” I’d say “Wow! I’m impressed, good job.” Where now, that feels more like EMNIST. Because we have models that have over 10 million images and we have some that have over 30 million images where they’re so big, the idea of solving a 1 million image dataset makes me laugh, where five months ago that was not true. Like, five months ago, if you came to me with, like, “Hey, I have a 50 million image dataset, I need you to build it,” [inaudible], I wouldn’t know how to do it. So we focused a lot of time and attention on that.

And this is a quick side note for people doing startups. I have a lot of background in patents. I’ve done seven patents for HireVue in machine learning, and the interesting perspective I have now is you have — a lot of people feel like you need patents to protect your IP, especially in the U.S. So if you have a good idea you’re going to feel, “Oh, I need to patent it because that’s what’s going to protect me.” The thing I’ve learned is patents—it’s really comparable to building a war chest. So if you and I are peers but we’re competitive, it’s useful for me to have a patent, especially a couple of them, because if you try to put pressure on me, hopefully I have more patents that I can actually scare you away or put pressure on you.

But if I’m competing against a big company, like an Oracle or an IBM or a Microsoft or a Google, the patent’s value kind of evaporates because your war chest is too big. Like, if I have one patent that you’re infringing on and you have a thousand, you can definitely find something that I’ve infringed on, and then your pockets are deeper, so if it comes to a lawsuit, the lawyers always win in a lawsuit, but am I going to bleed more than you do? So I think patents are great for companies that are smaller than you, but I’ve kind of come 180 where I’m really excited about trade secrets. I love trade secrets.

So if you have a product and you have a trade secret and there’s no way for me to reverse engineer it, that’s fantastic. Because if I’m a big company and you have a trade secret, I have no choice but to buy you for a ridiculous amount of money. This is what I’ve seen looking at different case studies, different companies. At Ziff, we feel like three months ago we had no trade secrets and now we have two.

Kirill: What trade secrets? Are you going to tell us? (Laughs)

Ben: I’ll whiteboard what they are right here. So, one of the things I think we’ve really nailed is our auto scaling capability. We can run essentially billions of inferences per day without losing sleep where four months ago that was not true. So, we have very high reliability and very high throughput on thousands and thousands of recurrent connections and we’re really good at bursting traffic.

And then the other thing that we’ve nailed, which I’m glad we identified as a trade secret, because I probably would have just started telling everyone, but my co-founder, David Gonzalez, there’s a reason he’s the CEO. He’s good at that stuff. So one of the things that we realized we’re really good at is dealing with mixed deep nets. If you have a deep net where you have text and image and structured data and you need a single model that maximizes the information from all of those, apparently that’s confusing for a lot of people. And our approach, we’re packaging that up as a trade secret.

One of the models we used is we have a home pricing model where we predict the price of your home, but we use all of the images of your home and we use the unstructured description and we use all of the structured data, so that would include square footage, number of beds, baths, when was it built, home style. And what we’ve noticed is, taking the classical approach, we can take the R-value on the price prediction from a .6 up to a .92, leveraging the information in the images and in the description. We’re offering that as a workflow. So if you have datasets that are complicated and they’re in that format, then we can handle combining the audio/image/video/text into a single deep net. So, find those trade secrets and keep them secret.

Kirill: Yeah. And with patents, you have to apply for the patent. With a trade secret, you just don’t tell anybody. That’s all.

Ben: Yeah. And the patents we’ve written, you go into so much information in the patent that an expert can recreate what you’ve done. Sometimes that’s bad. Like, if you’re a huge multibillion dollar company and I made something really interesting, if you can just read my patents… And sometimes there’s really creative ways to sidestep a patent because I’ve worked on the side of protecting my patents and making them very broad, so someone who’s had similar experience can figure out loopholes and ways to kind of sidestep your patent and get similar value.

I am a huge fan of trade secrets, but I think patents are still important. We’re actually consulting with a company right now where we’re going to go and help them write a patent next week. We’re doing it for two reasons. Your board loves it, your investors really like patents, it makes them feel better, and they’re relatively cheap, so you might as well file a patent on the stuff that isn’t mission critical. But if you can keep it a trade secret, then it really is hard to reverse engineer it.

Kirill: Gotcha. Thanks a lot. It sounds exciting, what you’re doing at Ziff. By the way, for those listening, if you’re a bit lost in this discussion, make sure to check out the first podcast with Ben because there we talked more about deep learning and what it is and how it works and why it’s better than machine learning. We’re not going to go into detail on that again so not to repeat ourselves, but with that in mind this all 

makes more sense, even though some of this stuff is still over the top of my head. (Laughs)

Let’s talk about some cool stuff. You got some great feedback from the first podcast and you even met some people who before they met you had heard your first podcast, which I was really excited to hear about. So that’s really cool. And you mentioned that on this one you wanted to brainstorm some ideas. Let’s do that. Have you heard of the show “Black Mirror”?

Ben: Yeah, I love it. I had not heard of it and I was giving a talk to an MBA course and I must have been in a bad mood or something because they were asking me questions about the future of AI and I told a very, very dark future. I was supposed to speak for an hour and it ended up being an hour and a half and they just kept asking questions. I kept talking about all these very disturbing ethical scenarios and one of the students raised his hand and said, “That’s actually an episode in ‘Black Mirror.’” I was like, “Really? I need to watch this show.” Yeah, there are some scenarios in “Black Mirror” that could definitely come to life.

Kirill: Yeah. And I saw one old one from Season 1 or something which I didn’t really like, but just yesterday I watched one from Season 3, Episode 1, where they have these things in the eyes and they always give each other ratings. Have you seen that one?

Ben: Yeah, I saw that one. I liked that one. I really like the one where someone was able to — their spouse died, so this woman, her husband died and she was able to recreate him using AI. And that’s actually a pretty natural — like, that was the idea I was proposing to the class, that if I have all of your phone conversations from these podcasts and if you die – heaven forbid – I can have you say whatever I want you to say. So using deep learning, I can have a conversation with you and if I mine your social feed and all your e-mails, I could potentially create a personality that feels very much like you. Is that going to help my mourning process? So if I could wake up in the morning and have a conversation with you on the computer, is that going to make me feel better? You realize that the answer is, “Yeah, I’ll feel a little better,” but there’s something very wrong about this. I love that episode.

Kirill: That’s so cool. Let’s talk about the one with the ratings. What did you think of that? For me it was very interesting to see that that’s where the world is going. And for those who haven’t seen “Black Mirror,” there are lots of episodes and they are completely unlinked, so we’re not going to ruin the whole show for you if we talk a little bit about one of the episodes. You know how in Uber you swipe, you give a number of stars to somebody? Well, in this one, every interaction you have with any person, you rate them from 0 stars to 5 stars and then in your eye you have this lens which tells you who you’re talking with and you can get some more information on them.

To me it felt like that is where the world is going. I was saying to Hadelin later that what we’re teaching in AI, this is the world that we’re helping create. I don’t know if I really want the world to go that way. What are your thoughts on that? Are we really going to get super-hyperconnected and really rely that much on technology that we’re going to just lose ourselves and just be all about how we interact, not only on social networks, but in real life and what ratings we have?

Ben: Yeah. I feel we’re already there a little bit because I’ve got a buddy in L.A., he works for a big brand media company, he’s an executive there, and for a lot of the parties he goes to, no one cares about who you are unless you have tens of millions of followers, so it’s all about how many people follow you and how many people care about your day-to-day. If you have less than 1,000 followers, then no one cares about you, at least as far as marketing dollars.

But I’m glad we’re doing a video so I can actually show you — this little thing is the Raspberry Pi Zero. It’s essentially flat, it has a little camera on it, so if I could actually do deep learning on this device—it’s actually a lot smaller than it looks because I can take the case away. So I could actually have this embedded in my clothing and I could have multiple cameras in my clothing where I could really record every interaction I have with everyone. And that’s usually information I have for safety reasons and for tracking, but I could score our interaction.

Kirill: Can you already do that?

Ben: No, we’re just playing with this. This isn’t like a real business driver for us, but just a curiosity. And why not have the business pay for it? So, we’re running deep nets on here that are these tiny, tiny little compressed nets that are optimized for this type of architecture and they work fairly well. So if I wanted to grab faces and try to identify people for the day, I could do that on my clothing.

Kirill: Wow! So you’re going to be the one introducing this to the world sometime soon?

Ben: Yeah.

Kirill: And what do you think of those shows in general? It’s really hard at this stage to create those types of technologies, but it’s really easy to create a show about it and to portray that.

Ben: I think the shows are actually more important than people realize because it raises awareness. Elon Musk has been harping on how dangerous AI can be and how we need to be thinking about it in the future, and he’s right, but some people feel like it’s a little too soon. Like, “Let’s not start having this conversation until things get weird in 20 years. Why should we have it now?” You actually mentioned a guy last time that has all those predictions. What’s his name?

Kirill: Ray Kurzweil.

Ben: Yeah. If you go and read his predictions, they’re scary.

Kirill: Yeah, totally.

Ben: Like, in the year 2060, I’m not that excited for my grandkids.

Kirill: 2060 is like 30, 35 years away — no, a bit more.

Ben: Yeah, 43 years away. His predictions aren’t spot on, he’s already been off on some of them, but I’d say he’s more accurate than we are and his predictions are impressive. I think he’s predicting that you can buy a computer equivalent to the human consciousness in flops for $1,000 by the year 2019. I don’t think that will be true. It won’t be $1,000, it will be like $50,000 or $100,000. I don’t know when he made that prediction, but the fact that he made it in the past, that’s… And you have the Volta chip from NVidia which is a huge breakthrough.

That’s a whole other thing for us to talk about, is we’re kind of on the cusp of a deep learning revolution. So as soon as you have these Volta chips coming out this year and Intel and Micron are working on their phase-change memory, I think they call it Optane memory, so you’re going to see a deep learning revolution next year unlike anything we’ve seen before. So we’re definitely on track to have the hardware and the learning capabilities.

Kirill: Tell us more about the Volta chip. What is this chip coming out from NVidia?

Ben: Yeah. A lot of times when chip manufacturers make improvements — I worked for Intel and Micron making NAND Flash — so when they make improvements it will be a 30 or 40% reduction of size, or for people doing computing, it’s a 30 to 40% reduction on computing capacity. So the chip next year, the upgrade, it will be 30% faster. It’s never going to be 1000% faster.

Kirill: It’s like iPhone 7 versus iPhone 6.

Ben: Exactly, it’s just like that.

Kirill: It’s maybe about 30 to 50% better.

Ben: Yeah. So imagine going from the original iPhone to iPhone 7 in one generation. That’s what NVidia has done. They have their Volta chip, it’s a 12x speedup. I don’t remember the exact number, but I think they spent $3 billion – it’s in the billions – developing it. And they made a major breakthrough on how they calculate this deep learning.

So, you have teraflops, which is a measure of your compute ability, cycles per second, and now they’re talking about floating point operations per second. Now they’re talking about a DEFLOP, which is specific to deep learning. So if you’re doing something that’s not deep learning, you won’t realize these speedups, but if you’re using deep learning, they came up with a 12x speedup this year. 12x is amazing.

And now you can buy their new DGX-1. They have a single computer that is so close to a petaflop that they should have just made it a petaflop. The engineers should have just spent another billion dollars and just made it a petaflop. So one computer can have a petaflop of computing and it’s going to change a lot of things, ImageNet will literally be the new EMNIST. You will solve ImageNet in less than an hour or maybe less than 10 minutes.

Kirill: Let’s just go over this thing really quickly. Petaflop is 10 to the power of 15 floating operations per second or something like that?

Ben: Yeah, gigaflop is 10 to the 6th, teraflop is 10 to the 9th, so it’s a thousand—

Kirill: Yeah, so it’s 10 to the power of 12, right?

Ben: Yeah.

Kirill: And the next one would be exaflop, 10 to the 15. Check Wikipedia, guys, we might be off by a 1,000 here.

Ben: No, I think that’s right.

Kirill: So, 10 to the power of 12 – that means 10 to the power of 12 floating operations like adding two numbers or multiplying or whatever operations per second. That’s how we measure supercomputers. Supercomputers are somewhere in the vicinity of 10 to the power of 15 or something like that. That’s what we’re talking about, that’s how quick these things are getting. And then the EMNIST and ImageNet – EMNIST is basically detecting numbers, “Is this a number 1, is this a number 9,” you know, handwritten numbers, understanding which number so a deep learning network can understand those. ImageNet is much more sophisticated, it’s a classification problem, you have to detect objects and images and classify what they are, like human and whatever.

Ben: Yeah. There’s a thousand different categories of images, over a million, and you need to train. And we’re superhuman accurate on that now.

Kirill: So the ImageNet dataset with these new Voltage chips is going to become as easy as the EMNIST dataset was 5 years ago?

Ben: Yeah. At Ziff, we definitely want to get one of these machines this year because we have a 36 teraflop R&D machine and it’s very performant. I saw someone post on ImageNet they’re getting 1,500 images per second running through during training and we get over 2,000 running through training. We have a brilliant R&D box, but we’re backlogged. We have too many projects to run on it, where if we buy the DGX-1 it will be the reverse. We actually don’t have enough jobs to fill its bandwidth and that’s a much better position to be in.

Kirill: That’s so cool. How much is that going to set you back? Do you already know the price?

Ben: I think full retail is 160 or 170. We’re inception partners with NVidia so we do get a decent discount.

Kirill: $170,000?

Ben: Yeah, but it’s worth it.

Kirill: Okay. That’s so cool. And it’s good timing, right? You just started Ziff and this thing comes out. It couldn’t be better.

Ben: Yeah, exactly. Well, I always have an anxiety thinking that we’re starting a machine learning startup too late because you have companies like DataRobot and these people that have been around a couple of years before, but I’ve realized our timing could not be better. The prospects are actually asking for AI where two years ago no one was asking for it or very, very few people were asking for it. Yeah, so it’s great timing. If we don’t do very well, it’s entirely on us, which is the stress of the startup. Every day we wake up, we have to optimize what we’re working on.

Kirill: Yeah, I totally agree. That kind of leads us into a quote by Andrew Ng, the founder of Coursera, “AI is the new electricity. The new revolution is coming.”

Ben: I completely agree. There have been a few revolutions. You look at the industrial revolution, and then you have the computer revolution, and you have the web. I believe the AI revolution will be bigger than all of them. And you will have protests and riots, which you had with the industrial revolution. There were horse and buggy people protesting and textile workers protesting because their jobs were being made obsolete, and I think we’ll also see that with AI in the next 10 years, where you have entire industries that are made obsolete because of AI.

There are problems with that because you have an uneducated workforce like truck drivers and stuff that have to go back into the workforce. And what are they going to do? Are they going to program? Some people are talking about having a universal income, or a liberal income, where everyone will make $25,000 a year or whatever that is for doing nothing. So if you just want to go sit on a hammock at the beach, you at least know you can feed your family and surf a lot.

Kirill: Yeah. Do you agree or not agree?

Ben: I think I agree with that, because I think for how much money that’s out there with these AI companies, you think you could definitely pay some type of tax into this fund. For the good of society, you don’t want to have a bunch of people that are just drinking beer and surfing because we need to continue to innovate and to stress ourselves and do hard things. I think you would have to find a fine balance between having a good safety net, but not having a long-term bottle for people to be lazy with.

Kirill: And people generally need challenges, right? People need – to be happy, they need to overcome challenges. If you’ve got everything, or you’ve got the minimal that you’re happy with, you’re going to be bored out of your mind and when people get bored they become unhappy and depressed.

Ben: Yeah, that’s another reason for intermittent fasting. Because when you eat, it’s so good when you eat — just kidding, but… Yeah, so the Volta chip is huge. It lines perfectly — you have the memory industry with Samsung, Micron and Intel, they’re doing a massive change with their memory technology right now. At the same time, you have NVidia with their breakthrough, so all that’s going to align next year. And the big breakthrough on the memory side is right now we’re all very limited with how much RAM we can have on our laptop. I just bought a new laptop that has 16 GB of RAM, but we’ve been able to buy laptops with 16 GB of RAM for a while now where what you’ll see 2-3 years from now is you could have a terabyte of RAM.

Kirill: That’s 1,000 gigabytes.

Ben: Yeah, 1,000 gigabytes. And imagine what you can do with that. Even today I get ‘out of memory’ errors that I have to be smart about.

Kirill: What drove that breakthrough?

Ben: The breakthrough is you have DRAM, which is your working memory, so that’s where all your stuff is, your programs when you open it, and it is so much faster than disk, a hard drive, I think it’s 3,000 Mbps or 2,000-3,000 Mbps where your hard drive is like 100. That’s why we run programs here, we don’t run them here. If you use all your memory on Linux, you can start running on hard drive and that’s called swapping, you actually write it and it’s super slow, your whole computer just comes to a screeching halt.

So they invented the solid state drive, which is good, you’ve got demand for solid state drives and they’re now becoming much more popular where they’re faster, but they’re only like 300 Mbps and they have some special ones you can buy that can hit 2,000 Mbps. But the big breakthrough with this new memory is everything is 3,000 Mbps. Now that your long-term storage, which is huge, which is like your 4 TB drive, it’s just as fast as this over here, so why not just make them the same thing? So what’s happening is the difference between volatile memory and non-volatile memory, or working memory, and long-term storage, that difference is going away so you’ll just have one storage. So that’s great for us and for people who like AI it’s even better. That’s a big bottleneck that’s now fixed. They’re already shipping it, but they’re kind of on the cusp of major breakthroughs next year.

Kirill: Okay, cool. Tell us what your mission at Ziff is. You’re working on some exciting projects, but what’s the end goal, what do you want to do for the world?

Ben: There’s lots of reasons why we’re doing this. One reason is we want to work with the people that we want to work with, so we want to work with very, very talented people and control who’s around us and also control that environment. We want it to be a fun place to work. There’s a lot of perks that you can do once you have a business where if you want to have a climbing wall in the office or a trampoline or arcades or Xboxes or you want to have a private chef. We see these companies that do that. That’s fun, but that’s not the main reason why we’re doing it.

The other reason we’re doing it is we want to work on the problems that we want to work on when we want to work on them. And sometimes, like with Google, with DeepMind, they have a luxury of working on problems where there’s no market appetite. Them teaching a computer to play Go, there’s no market for that right now. We know that that will open up markets and there are side benefits of them doing that.

So having a company that’s profitable, that’s growing, we can hire really smart people, we can work on really hard problems and we can work on what we want to work on. But the main goal is we want to — I joke that we are the AI missionaries from Utah. So we are out there spreading the word about AI and our goal is to get AI into as many products as possible in 2017 and 2018. So, as many products as we can, we want AI to be levelling them up and differentiate them.

And then the final part is impact. We want to have as much impact as we can. And this sounds — we always think of impact as being good, but impact could also be bad to some companies. Some of that impact could be reducing cost. If you’re a company and you’re selling inferences and we cut your legs off, that’s actually good. It’s maybe not good for you, but it’s good for society. AI everywhere and AI that is cost effective is really the focus.

Kirill: Do you guys think you will ever get to general artificial intelligence?

Ben: I’m glad you brought that up because the thing that doesn’t scale—I think this is fair to discuss. One of the issues that’s going on right now is there is a major skills gap for these hardware devices. Let’s say you’re company A, you’re well respected, people know who you are, you’re in the Fortune 1500 largest companies in the world. If I come to you and I say, “Hey, you need to buy this $200,000 deep net server,” I can probably show you a pitch, you’d be impressed by it, but you’re not going to buy it. And the reason you’re not going to buy it is there is a major skills gap where you don’t have the expertise to leverage that and you can’t connect the dots between that and value.

In the end, we have to make money as a business and if I can’t connect the dots, I’m not going to buy it. So, one of the things that we’ve noticed is that bridging that gap is very, very helpful. Providing examples that are closer to the company, for what they need, and then filling that gap where — if you can’t hire that $200,000 plus a year deep learning expert, then you can use this general workflow to provide value. So the thing that we’ve been working on right now, and I feel like we’ve done it, is that we have a general workflow for training models where there’s not a human in the loop.

So, if you have images that are grouped into different folders, into a folder structure whether that’s on the cloud or it’s in a .zip or .tar file, essentially you’re done with the deep learning problem. So what you do is you upload it and then it goes through our platform where we use best practices and we give you a very competitive model. And we’re really pleased with how we’re doing. So the thing that we take away from you, which is a good thing, is we take away from allowing you to decide what the validation set should look like because most people are not qualified to make that decision. So the validation set is the set that we remove from your dataset to make sure it’s working.

And then the other thing we take away is network architecture decisions. Some of those people, unless they’re full-time deep learning expert, they’re not qualified. And then we take away the training headache, so training massive datasets on the cloud, managing the optimization and the learning rate decay. So we’ve rolled that all up into a pipeline that we offer our customers. It’s not exactly general AI, but it’s close because it’s taking care of image classification, audio classification, text classification.

Kirill: So AI in the box?

Ben: AI in the box, but as far as having Jarvis in your house that you can talk to, it’s not Jarvis in your house yet. It’s kind of the stepping stone.

Kirill: Okay, that’s pretty cool. So you’re on the way? You’re on the way to bring a new consciousness to this planet eventually?

Ben: Yeah. We’re working as fast as we can to improve AI and how it’s used in society and we’ll keep working until we’re scared.

Kirill: (Laughs) I actually had an interesting question on this. There are so many companies that are racing towards AI. For instance, we’ve got Google DeepMind, you’ve got OpenAI, a new project by Elon Musk, Neuralink, a similar kind of concept to expand our brains, you’ve got Ziff, you’ve got quite a few startups and more mature companies as well in this race. What role do you think government will play in all of this? Because ultimately, AI is like a superpower, it’s like a super weapon you have. You know, there’s regulation around nuclear weapons, which countries can and which countries can’t have them, individuals can’t go and create them. AI is not a nuclear weapon, but it’s close. So what are your thoughts on what role governments will play and how quickly they will start playing this role?

Ben: This is something I feel pretty strongly about. It’s really sad how much the U.S. spends on their military right now. They spend trillions of dollars on technology that is decades old. So you have these aircraft carriers and you may remember the news the U.S. shot 52 or 54 Predator missiles into Syria. We kind of have a crazy president right now, unfortunately.

Kirill: I didn’t hear that, but okay.

Ben: Okay. But those missiles are a million dollars each and the technology is over a decade old. So you just spent $52 million to destroy some jets in Syria. Imagine this scenario where I drop 100 drones out to the sky, they free-fall, their arms come out and their blades turn on 100 feet before they hit the ground, and they go searching for jets. And all it is, it’s just a little deep learning camera and it’s got a payload of C-4 explosive. It’s going to find the jet in the hangar and it’s going to find the most critical part on the jet. So the drone knows, “If I explode right here, that’s terrible.” They may not destroy the whole jet, but they’re not salvaging anything.

Kirill: When you say blades come out, you mean the propellers?

Ben: Oh, yeah.

Kirill: So they don’t cut anything, they have the explosive on them?

Ben: Yeah. So you have a quadcopter. Imagine like a DGX-1 or something. You have a quadcopter with four blades. It falls out of the sky, it’s just free-falling so it’s totally silent, and right before it hits the ground the props turn on and it goes searching for targets.

Kirill: And it has a C-4 explosive attached to it?

Ben: Yeah. So all the technology components are there, but that had been demonstrated instead. I don’t know if the U.S. is working on stuff like that, but if you had given me or you a million dollars of funding, we would have pulled that off.

Kirill: Yeah, not 52. Like, one million is enough.

Ben: Yeah. And you would have had fewer civilian casualties and that entire spend may have been $300,000 or half a million to execute that. So I would love it if the U.S. would take a big risk — I don’t see it as a risk, but I’d love it if they said, “You know what? Half of our military spend has to be AI.” Because there are secondary benefits that come from that, society benefits, and it keeps the U.S. on edge with their military power which I think is not required, but… So you’re talking about responsibility and I’m kind of pushing the gas on the pedal saying we need to really amplify what we’re doing as far as AI.

I think Elon is worried about a nuclear arms race equivalent with AI. The drones I just described, they can also target soldiers or military-capable individuals and take them out, too, and you have a thousand of those drones and I have a million and someone has a billion and these drones are charged with solar. That suddenly becomes really scary because if you have drones and they’re given a mission and they can kind of go off and they do this Lévy search pattern through your country where they go and do an objective and then they charge and do an objective, then it becomes an arms race.

I’ve heard that some of these leaders like Elon, they’re pushing for the U.S. to agree to never allow AI to make a kill decision without a human in the loop. So if the drone is flying around and decides you have an AK-47 and it needs to take you out, they want a human soldier somewhere through a satellite to make that final call. That’s kind of the responsibility piece that people are pushing. But I think it could be pretty easy to show type 1/type 2 errors.

Getting back to humans doing things poorly, if you’re making a split decision where your life is on the line, you’re more likely to kill a civilian or to not save a hostage, but if you allow AI to just do it, then you’re going to get headshots right here and no civilian casualties. We can definitely beat the human, but that becomes quite scary.

Kirill: Yeah. It’s kind of like with the driving. Yes, there will be accidents of automated cars. It’s inevitable, you cannot get 100% all the time. But how many accidents right now do we have on the roads? Millions, right? Every year so many people die on the roads. So for the politicians it’s not a question of introducing self-driving cars that never have accidents. Their challenge is to convince society that we need self-driving cars and we need to understand that yes, there will be accidents that are not human error-related but AI error-related, but they will happen like one millionth of a time with frequency as the humans currently have.

Ben: Yeah. The vanity metric there is deaths per 100 million miles, and Tesla has already beaten that. So Tesla is better than that. And the other key point is, here in my local town we have people that die every week driving, but no one learns. The fact that you fell asleep because of your polyphasic sleep experiment and you crashed and died, no one is going to learn from that sad experience. But the beauty with AI is it can literally recover the cameras, potentially it can recovery the memory, and all of that can go into a big learning algorithm where it’s constantly getting smarter.

Right now, I feel like all the driving-related deaths out there are completely wasted. There’s no value that society gets, it’s a complete loss. But with self-driving cars there will be some measurable — of course we don’t want accidents, but it’s not a total loss. We will learn from these.

Getting back to the government, right now I actually see the government as an AI limiter rather than an AI assist because you have these classical organizations where they’re limiting AI’s integration in society. One example happening right now in the U.S. is with the FDA, where you’re bringing up an important point. The important point is, “Tell me about the human equivalent. I know that self-driving car killed your relative, but tell me about the human equivalent,” and the human equivalent is worse. And when it comes to the FDA – that’s the Federal Drug Administration for the U.S. – they kind of safeguard a lot of drugs that are rolled out and some of the technologies and they’re being a major roadblock right now for AI in healthcare.

And the reason is they’re not holding the doctors to the same standard, they’re holding AI to a higher standard. So if I have an algorithm that can outpredict you as a dermatologist, it should be diagnosing all of your patients. But they’re holding AI to such a high standard where the AI rates have to be like above 99% or something absurd. What they need to be doing is say, “What’s the type 1/type 2 error rate of the physician?” because it’s quite high. “The physicians have a major error rate that is being beaten by the AI, so how about we just double down on the AI?” Right now, I think Europe is actually leading the U.S. with some of this research because they don’t have as many roadblocks.

Kirill: Interesting. Why do you think that’s happening in the U.S.? Because I’ve even heard there’s an app already on your phone that you can get and take photos of your skin – maybe we even discussed this in our previous podcast – and it can tell you if it’s cancer or not. What’s up with the limitations that the FDA is imposing in this case?

Ben: It’s hard to know what the main drivers are, but it’s not just the FDA. What I see is you have this new injection with a big, old, massive organization and it’s hard for you to move and/or impact it. We saw the same thing in HR with I/O psychology. So it’s this huge body where they kind of control the assessment space and the predictive analytics. So when you come and try to inject deep learning and data science and machine learning, that’s foreign to them and they’re not very receptive to that and they’re very slow to change.

And in the U.S. you don’t really look to the governments to lead the charge on innovation, you look to the private sector. So the private sector innovates and then sells it to the government. The government is not going to innovate internally at the rate that they need to. Maybe that’s one of the drivers, is you’re dealing with organizations that aren’t equipped to innovate as quickly as the private sector.

I don’t know the exact reason for the FDA, but it’s discouraging. I’m to the point now where I would actually be angry if one of my kids started to have symptoms and they’d have to get a brain scan to see if there’s a tumour, now I feel angry. I would want deep learning to be run on that image to assist the radiologist. This is a pretty personal thing for me right now because my wife has chronic kidney disease. She’s had it for 10 years and her kidneys operate at 40% of what ours do, and when they get below 15%, she will need a kidney transplant. She is still undiagnosed. She’s had a kidney biopsy, she’s met a with these nephrologists, she’s undiagnosed. They don’t know why her kidneys are functioning at this level, and they don’t know if it’s going to decrease.

We’re hoping they’ve gone down based on the challenges of just having kids, the childbirth, but when you go and meet with a nephrologist, he’s literally printing off a piece of paper, which is her blood work, and he’s scanning it and it just makes me want to yell because he’s looking for two or three numbers. And as a human, in your head, tell me what kind of complex calculations can you do.

Kirill: Like, 10 times 5. (Laughs)

Ben: Yeah, exactly, 10 times 5. That was hard. Times like 10.3 times 10.7 and the fact that you’re Caucasian and your BMI is this and your patient history is this – there’s two things that are pretty upsetting here. One is the human brain is not equipped to comprehend that data, and the other thing that’s upsetting is the human lacks the experience when it comes to these rare outlier diseases that there’s a good chance that he may have never seen her disorder, but if you combine all the nephrologists on the West Coast, they have seen it.

So AI is so much better to tackle this problem. And I’m not saying that AI has to replace the nephrologist, but the AI could return me a top 10 ranking of what’s the likelihood that my wife has this and there’s a really good chance that her disease or her issue would be top 3 or top 5 just by doing a Bayesian approach on a really big dataset. Yeah, so it’s really upsetting that it’s—

Kirill: And to your point, IBM Watson is already doing that, right? It diagnosed a case of rare type of cancer for some lady, she was going around the doctors for years and they thought it was one thing and they were treating her for one thing. And within a few split seconds, or like a minute or so, it diagnosed her with a specific type of cancer that nobody else had thought of just because it has access to all of the cases that ever existed and that are on the Internet. I agree, within 5-10 years from now, whenever we go to a hospital, we are going to be like, “Do you want a robot surgeon or do you want a human surgeon?” “No, thank you, I want a robot surgeon.”

Ben: Yeah. And even you going to a hospital may go down because you may just have your smart device where your kid’s coughing, you pull it out and it actually diagnoses the disease based on the cough sound based on millions of kids coughing, or the rash example that you brought up.

Kirill: Yeah, totally. Cool. Ben, we’re kind of coming close to time. I want to keep this about an hour. There’s so much more I want to talk to you about, so you definitely come back on the show again. By then you will probably have some more major breakthroughs at Ziff and any other projects that you’re working on. Yeah, thank you so much for coming on and sharing all of these insights.

Ben: Yeah. Shall we end with one last tip to the listeners on how to get a job?

Kirill: Oh, yeah, that’s right, that’s right.

Ben: One secret tip — you probably have an opinion and I have an opinion, I’m sure they complement each other, so what’s your one tip? I’m a new data science candidate, I show you my resume, it’s kind of boring, and you’re going to give me some advice. What’s that advice going to be?

Kirill: What’s my tip when you come to an interview?

Ben: Or even before the interview. Like, you’re a data science hiring manager and I’ve talked to you at a meetup. So I’m not going to give you an interview yet.

Kirill: Okay, I got it. I’ve got one. I saw this once when I was at Deloitte, there was this super high competition for this one role, and this guy completely smashed it. This is that advice I can give to anybody. If you really, really care about the company you want to work for and you really want to work there, register a website, for instance WhyDropboxShouldHireBenTaylor.com, and set up a really cool flash animation with all your skills and moving through the whole corporate world and ladder of Dropbox or whatever company you want to work for and show them how you bring value all through this animation. Add some programming and maybe some data science into this animation, and show how data is coming in, coming out, what you’re doing.

So when you send it to the hiring manager, or whoever you’re contacting with, don’t send them your resume, send them that one link and it already has the explanation and the name of the link – WhyDropboxShouldHireBenTaylor.com – and not to click on that link would be really unwise of a person. Just out of curiosity, people click on it, and as soon as they’re there they see this amazing animation. You need to show how much work you’re willing to put in, how badly you want it. People would argue if they see you want—you’re not just applying it to everybody, you specifically want to work for them. If they see that, they will talk to you, at least they will talk to you. Most likely they will already prioritize you over the other candidates because you’ve put in that effort, you’ve shown them that you want to work at that company.

Ben: Yeah. That makes me laugh because it actually kind of brings up a competency that David and I were talking about in L.A., because I feel like the number one reason we hire someone is passion, like if you’re passionate enough, you are willing to do the work. But David and I were sitting down for dinner and we’re trying to really flush out how are we different than the rest of the people breaking into the space, what’s unique to us. And I think the thing that we settled on was we’re crazy. We both met each other through separate career paths and different things, but the level of crazy that we both have, I think that distinguishes us and that kind of leads into your example.

So for you to go and do that, you’re a little crazy. You’re a little crazy to spend that time. So, being passionate and crazy and motivated — because most of the people we get in an application process, they’re all the same. They all did Coursera, they all did Northwestern, they all did this program. And I’ve talked to people who have done undergrad and graduate data science programs at universities, and they’re all unremarkable. They’re all just kind of the same. I mean, it’s not saying anything mean about them, it’s just like there’s no one exceptional, there’s no one I’m excited about. I’m not excited to hire anyone.

But the people that stand out — I’m going to give a shout-out to Adam Rogers, he’s at Microsoft right now. I was presenting deep learning at a local university. Most of the students don’t know anything about deep learning, so it’s kind of entry level, I’m talking through it, talking about what’s going on. He is in the front row and he’s asking me about Siamese nets and Torch. The questions he’s asking me, your eyebrows go up. He’s in the front row and he’s asking me these questions and I turn to him and at the end, I tell them why people get jobs, but in the back of my mind I’m saying, “Well, whoever this kid is in the front row, I don’t know who he is, but he’s going to have no problem getting a job.”

And sure enough, like two weeks later, he got a job where he was competing against PhD level data scientists, but I think he dropped out of his graduate program. So it’s not about your titles or your career, it’s that intrinsic motivation. If I tell you to go do A, B, C, and I see you at the next meetup a year from now or a month from now and you haven’t done that, I have no faith that you’re going to get any data science job. But if I found out that you stayed up all night this weekend and you actually sent me a code example and I can tell that you’ve put a lot of time in this, you’re going to be just fine.

Kirill: Yeah. So, to summarize, your one tip is…

Ben: Intrinsic motivation and passion. You need to fall in love with it. And then a little crazy garnish.

Kirill: Sprinkle a little crazy on the top. Sounds fantastic. Thanks a lot, Ben. It’s a pleasure and I’m sure a lot of people will get tons of value out of it.

Ben: Okay. It’s good to see you. Talk to you soon.

Kirill: So there you have it. That was Ben Taylor, Chief Data Officer at Ziff. I really hope you enjoyed this episode. For me personally, I probably found most interesting the journey that Ben has gone through himself, how in these five months since his previous episode on this show he’s now started a startup, he’s creating amazing products, he’s working with fascinating technology, he’s building a team, he’s working with clients. So some huge steps that he’s taken in his own career and his own life and that I think is very inspiring and I think we can all learn from that on how to push ourselves out of our comfort zone. Because even five months ago, Ben was already one of the thought leaders of data science and he was already up as high as he could go in the company where he was. Like, he could just sit back and relax and enjoy life at that point, and he was already doing the work he loved to do, but he decided to push it even further, start his own business and drive massive change into the world.

And I hope that that will also encourage you to push yourself to your limits. And on that note, you can get all of the links to all the materials we mentioned at www.superdatascience.com/85. You can also find the video version of this episode there. If you haven’t seen it yet, you can check it out there. Plus make sure to connect with Ben. Hit him up on LinkedIn, connect, and see how his career progresses even further in the future. And I’ll see you back here next time. Until then, happy analyzing.