WEBVTT - How the Nerds Conquered the NBA 0:00:15.316 --> 0:00:24.036 Pushkin. About a decade ago, special cameras in the rafters 0:00:24.036 --> 0:00:27.196 of NBA arenas started following all the players around the 0:00:27.236 --> 0:00:30.196 court and tracking the ball everywhere it went. It was 0:00:30.236 --> 0:00:34.516 this huge new trove of data, but nobody really knew 0:00:34.516 --> 0:00:37.396 what to do with it. This is a very familiar 0:00:37.436 --> 0:00:40.516 problem in the modern world, tons of data that just 0:00:40.636 --> 0:00:44.596 kind of sits there. Then a few computer scientists came 0:00:44.636 --> 0:00:47.916 along and had an idea for how to solve the problem. 0:00:48.196 --> 0:00:50.676 They thought they knew how to take all that data 0:00:50.756 --> 0:00:55.116 and turn it into profoundly useful information, and so they 0:00:55.156 --> 0:00:58.236 started a company called Second Spectrum to see if their 0:00:58.276 --> 0:01:04.996 idea would work. I'm Jacob Goldstein, and this is What's 0:01:05.036 --> 0:01:07.916 Your Problem, the show where entrepreneurs and engineers talk about 0:01:07.916 --> 0:01:10.116 how they're going to change the world once they solve 0:01:10.156 --> 0:01:14.236 a few problems. My guest today is Regeev Mahiswaran. He's 0:01:14.316 --> 0:01:17.676 the co founder and president of Second Spectrum. The company 0:01:17.796 --> 0:01:20.996 started out turning video of NBA games into information that 0:01:21.076 --> 0:01:24.756 coaches could use. Today, they work with every NBA team 0:01:24.836 --> 0:01:27.596 and with Major League Soccer teams in the US and 0:01:27.796 --> 0:01:31.436 with Premier League teams in the UK. Regiev's problem is 0:01:31.476 --> 0:01:35.156 this how do you teach a computer to understand sports? 0:01:38.236 --> 0:01:41.196 Regiev and his co founders started Second Spectrum back in 0:01:41.276 --> 0:01:44.836 twenty thirteen. Last year, the company was acquired by a 0:01:44.876 --> 0:01:48.676 sports analytics company called Genius for two hundred million dollars. 0:01:49.196 --> 0:01:52.556 My conversation with Regieve focused on his work with the NBA. 0:01:52.636 --> 0:01:55.476 That's where Second Spectrum has been working the longest. But 0:01:55.756 --> 0:01:58.876 he started out talking about this bigger idea, how do 0:01:58.916 --> 0:02:02.476 you turn data, really any kind of data into something useful. 0:02:02.716 --> 0:02:05.076 So I think this happens everywhere. There's a lot of 0:02:05.116 --> 0:02:08.796 companies that collect data, faster data, more data. But what 0:02:09.156 --> 0:02:11.876 generally happens is the big piles of data. I feel 0:02:11.876 --> 0:02:13.636 like they're like grain. They just if people don't know 0:02:13.636 --> 0:02:14.956 what to do with them, so they just stick them 0:02:14.996 --> 0:02:16.836 in the closet and it's like, yep, I got piles 0:02:16.836 --> 0:02:18.996 of grain in there, and it keeps piling up. And 0:02:19.076 --> 0:02:21.636 I think what we said was, you know, we brought 0:02:21.676 --> 0:02:23.396 a bunch of machine learning, a bunch of other stuff 0:02:23.396 --> 0:02:25.836 to and said, you know, the data, this massive amount 0:02:25.876 --> 0:02:28.196 of coordinate information. No one can work with it. Coaches 0:02:28.196 --> 0:02:30.556 can't work with it. The leagues that have a tough 0:02:30.596 --> 0:02:32.196 time working with the media has a tough time worker 0:02:32.236 --> 0:02:34.556 that we are going to sort of grind that grain 0:02:34.676 --> 0:02:38.436 into story elements that people can actually use. So we 0:02:38.516 --> 0:02:41.436 turn it into oh, that's a pass, or that's a shot, 0:02:41.556 --> 0:02:43.236 or that's a pick and roll, or that's a between 0:02:43.236 --> 0:02:45.156 the lines pass, or that's a blitz, and so we 0:02:45.196 --> 0:02:48.756 start turning them into words by which people can tell stories. 0:02:48.756 --> 0:02:50.476 So I would say, like the grain is not useful, 0:02:50.516 --> 0:02:52.676 you don't want to make like bread and donuts. Is 0:02:52.716 --> 0:02:56.356 there some particular example you can give me from the 0:02:56.436 --> 0:03:00.116 early days of the company of basically of a problem 0:03:00.116 --> 0:03:01.756 you solved, of a thing you set out to do. 0:03:01.996 --> 0:03:03.516 Maybe it was hard to do, maybe it didn't work 0:03:03.556 --> 0:03:04.916 the way you thought it would, but in the end 0:03:04.956 --> 0:03:08.956 you made some useful thing for coaches, teams, whatever. There 0:03:09.156 --> 0:03:10.676 are two things that we did that sort of a 0:03:10.716 --> 0:03:13.676 big we're a big game changer, right. So one was 0:03:13.756 --> 0:03:17.556 the idea that people were taking shots and you couldn't 0:03:17.556 --> 0:03:19.116 tell if they were a good shot or a bad shot. 0:03:19.156 --> 0:03:20.556 You would use some rules. It's like, oh, if you 0:03:20.556 --> 0:03:23.476 took it here, you know, if people weren't guarding you 0:03:23.556 --> 0:03:25.276 too closely, it was good. But you know, if you 0:03:25.436 --> 0:03:27.356 dribbled a lot, and you jumped in somebody's face. It 0:03:27.396 --> 0:03:29.516 was bad. I think we were able to use math 0:03:29.556 --> 0:03:32.156 and say, oh, that shot should go in forty two 0:03:32.236 --> 0:03:34.196 percent of the time. That shot goes in forty nine 0:03:34.196 --> 0:03:36.116 percent of the time for an average player, it goes 0:03:36.156 --> 0:03:38.396 in for this player. So we were able to quantify 0:03:38.476 --> 0:03:41.476 the quality of a shot, which was basically the core 0:03:41.596 --> 0:03:43.916 thing in basketball is you want to get high quality 0:03:43.956 --> 0:03:46.236 shots and prevent high quality shots, and there was no 0:03:46.276 --> 0:03:48.676 way of measuring it. And we basically said, like, no, no, 0:03:48.676 --> 0:03:50.996 we know exactly what the quality of the shot is, 0:03:51.196 --> 0:03:53.756 and then we also know which players add value beyond 0:03:53.836 --> 0:03:57.236 that shot. So players like Steph Curry and Kevin Durant 0:03:57.276 --> 0:03:59.196 can take a shot that it might be a forty 0:03:59.196 --> 0:04:00.756 five for an average player, and for them it goes 0:04:00.756 --> 0:04:03.076 in fifty five. And that was a big deal. That 0:04:03.156 --> 0:04:05.956 was the first big thing, just to step through that 0:04:06.076 --> 0:04:09.516 a little bit more slowly. So before you came along, 0:04:10.236 --> 0:04:13.396 what did people know? What did coaches know about sort 0:04:13.396 --> 0:04:17.996 of percentage probabilities that a shot will go in? They 0:04:17.996 --> 0:04:21.116 would just track it much more coarsely, so they would say, oh, 0:04:21.276 --> 0:04:24.396 from the corner three, or you would shoot this, and 0:04:24.476 --> 0:04:26.356 from above the break you would shoot this, or in 0:04:26.396 --> 0:04:28.396 the lane you would shoot this. But they didn't know 0:04:28.476 --> 0:04:32.196 things like, hey, there's a really tall defender running straight 0:04:32.236 --> 0:04:34.716 at you, or you're moving sideways. You know, and these 0:04:34.756 --> 0:04:38.556 things also matter important. Yeah, and then the raw data 0:04:38.596 --> 0:04:40.476 allowed us to know all that, and then what we 0:04:40.516 --> 0:04:44.116 did was start to build models that basically predicted the 0:04:44.116 --> 0:04:46.716 likelihood of various shots going in. So one of the 0:04:46.756 --> 0:04:49.556 things you did is coming up with a better metric 0:04:49.716 --> 0:04:53.476 for shot quality that has more inputs. The other thing 0:04:53.516 --> 0:04:56.356 you did, is it right, was around the pick and roll. 0:04:56.596 --> 0:04:58.796 So I think there are all these events in basketball, 0:04:58.836 --> 0:05:01.076 and you can get pretty esoteric. There's a pick and roll, 0:05:01.156 --> 0:05:03.916 but there's many types of pick and rolls. There's many 0:05:03.916 --> 0:05:06.076 types of ways to defend pick and rolls. Just to 0:05:06.116 --> 0:05:07.636 be clear, let me let me just say we won't 0:05:07.636 --> 0:05:10.436 go into detail, but pick and roll it's just a play. 0:05:10.516 --> 0:05:13.436 It's an offensive. It's a play. Yeah, it's somewhat complicated. 0:05:13.436 --> 0:05:16.236 It involves two people on offense and two people on defense. 0:05:16.316 --> 0:05:18.916 Let's just say that's that's all you need to know. Okay, 0:05:18.956 --> 0:05:20.996 So what did you do with the pick and roll? 0:05:21.516 --> 0:05:23.436 So I think the first thing we did was we 0:05:23.436 --> 0:05:26.876 had a machine be able to identify it by staring 0:05:26.916 --> 0:05:30.316 at the data. So before us, it was basically humans 0:05:30.356 --> 0:05:34.036 would watch the video and just start counting how many 0:05:34.076 --> 0:05:36.716 happened and counting what the type of defense was, and 0:05:36.756 --> 0:05:39.876 basically they would just collate all that information. But I 0:05:39.876 --> 0:05:42.676 think the big thing that we discovered is when machines 0:05:42.756 --> 0:05:45.356 did it. The big example was there was a year 0:05:45.476 --> 0:05:48.796 where the human collection said Chris Paul led the league 0:05:48.836 --> 0:05:51.316 running eight hundred pick and rolls, and we said, well, 0:05:51.396 --> 0:05:54.756 Chris Paul led the league running four thousand pick and rolls, 0:05:55.036 --> 0:05:58.156 and one of these is not right. But then we 0:05:58.156 --> 0:06:01.036 could just say like, well, here's our four thousand, and 0:06:01.076 --> 0:06:03.516 then you're saying, oh, okay, you're missing sort of you know, 0:06:03.556 --> 0:06:06.236 eighty percent of them. And so I think then I 0:06:06.276 --> 0:06:08.876 think once we did that, we could automate a lot 0:06:08.956 --> 0:06:12.356 more complicated things, like having the machine understand all the 0:06:12.396 --> 0:06:14.676 defenses that you could play against various types of pick 0:06:14.716 --> 0:06:17.236 and rolls, and then once we solve those, we sort 0:06:17.236 --> 0:06:20.156 of built this engine that could generate words accurately, and 0:06:20.196 --> 0:06:22.836 then we started growing teams and each one of them 0:06:22.836 --> 0:06:24.996 would say, hey, can you do these words can you 0:06:25.036 --> 0:06:26.956 do these words? Can you do these words? Like say, 0:06:26.956 --> 0:06:28.956 some of the words they were asking for. So they're 0:06:28.996 --> 0:06:31.756 sort of like you know, flares and loops and zippers 0:06:31.756 --> 0:06:34.916 and jams and trails and whips and you know veer backs, 0:06:34.956 --> 0:06:38.036 and I mean the sort these things are basically what 0:06:38.636 --> 0:06:41.836 plays their their defenses, their moves. That's right. So there's 0:06:41.836 --> 0:06:44.716 like sort of you know, the particular kind of screen 0:06:44.796 --> 0:06:47.356 that they used to get Steph Curry free off in 0:06:47.436 --> 0:06:50.196 the side versus the name of a defense that the 0:06:50.316 --> 0:06:53.316 person running after him might choose to play while he's 0:06:53.356 --> 0:06:56.116 doing that thing, right, so to avoid that screen. Yeah, 0:06:57.156 --> 0:06:59.556 and when they say learn a work, they really mean 0:06:59.796 --> 0:07:03.516 can you teach your machine or get your machine to 0:07:03.756 --> 0:07:08.916 learn to recognize this play this thing? Because they want 0:07:08.916 --> 0:07:12.356 to know every time they do this two questions. Every 0:07:12.396 --> 0:07:14.516 time they do this and we do this, how efficient 0:07:14.636 --> 0:07:16.916 is it? And or every time they do this and 0:07:16.956 --> 0:07:19.716 we do this, show me every time that that a 0:07:19.836 --> 0:07:22.116 video of every time that happened over the last several years. 0:07:22.156 --> 0:07:24.036 Those are the questions they tend to ask. Okay, and 0:07:24.116 --> 0:07:26.156 so it had it has a bunch of quirks, you know, 0:07:26.236 --> 0:07:28.836 working with it, But but it was not easy. But 0:07:28.916 --> 0:07:31.316 we've done it, and now we've created sort of five 0:07:31.396 --> 0:07:34.356 hundred of these words that you know, coaches and you 0:07:34.396 --> 0:07:39.156 know players use on a daily basis. And so I 0:07:39.196 --> 0:07:41.316 can think of two parts of this that would have 0:07:41.316 --> 0:07:44.956 been sort of hard problems to solve, one being the 0:07:45.116 --> 0:07:48.996 kind of nathy computer part, like building the machine, and 0:07:49.036 --> 0:07:53.036 the other part being the getting people to believe you 0:07:52.796 --> 0:07:59.156 and use your insu So tell me a hard thing? 0:07:59.516 --> 0:08:01.516 Well from each right, So what's a hard thing from 0:08:01.556 --> 0:08:05.196 the from the MATTHI computer part? I think the mat 0:08:05.356 --> 0:08:07.356 the mat side is just that you don't have a 0:08:07.436 --> 0:08:08.916 lot of data. So it's for example, a lot of 0:08:08.956 --> 0:08:11.316 people who do machine learning will say like, oh, give 0:08:11.356 --> 0:08:13.676 me millions of examples of a thing and I will 0:08:13.756 --> 0:08:16.036 learn it. Where well, there aren't millions of pick and 0:08:16.156 --> 0:08:17.796 rolls that you're going to get this sort of how 0:08:17.796 --> 0:08:20.476 do you solve that? Right? Small sample size? Classic problem, 0:08:20.596 --> 0:08:24.036 small sample size, So you know, how do you solve it? Well, 0:08:24.356 --> 0:08:27.676 there's some mathematical chicanery that we invented. The answer to 0:08:27.876 --> 0:08:30.316 be very clever, like we're really good at bath. Is 0:08:30.356 --> 0:08:33.116 that the unsatisig outs? Yeah, let me let me ask you. 0:08:33.196 --> 0:08:36.276 This was there some like things you tried that didn't work. 0:08:36.316 --> 0:08:38.076 It was just like one equation didn't work and then 0:08:38.076 --> 0:08:40.116 another did. I mean, yeah, yeah, yeah, I know. I 0:08:40.156 --> 0:08:42.556 think we've we we you know, I think we we 0:08:42.556 --> 0:08:45.356 we said over the course of our life, we've used 0:08:45.396 --> 0:08:47.756 everything in the AI textbooks to try and solve the problem, 0:08:47.756 --> 0:08:49.556 and they've evolved over years to try and get them 0:08:49.596 --> 0:08:52.396 get to be better. Um. But you know, I think 0:08:52.396 --> 0:08:54.996 that that's part of the adventure of sort of trying 0:08:55.036 --> 0:08:57.196 to figure that stuff out. It wasn't like we used 0:08:57.476 --> 0:09:00.196 we used method X and the answer popped out, and 0:09:00.236 --> 0:09:02.436 so we've been constantly evolving the ability to do But 0:09:02.716 --> 0:09:05.636 I think I think part of the I mean, whatever 0:09:05.676 --> 0:09:07.436 the secret is, you have to leverage the structure of 0:09:07.436 --> 0:09:08.836 the problem. You have to leverage the fact that you 0:09:08.876 --> 0:09:10.956 know something about the fact that it's a bunch of 0:09:10.956 --> 0:09:13.196 people playing a particular sport. When you don't have enough 0:09:13.276 --> 0:09:15.676 when you don't have enough data, you have to leverage 0:09:15.676 --> 0:09:18.636 structure in some way. Yeah, so you're you're able to 0:09:19.356 --> 0:09:22.916 because it's such a constrained environment, because you know the rules, 0:09:22.956 --> 0:09:25.836 because you know the things you wanted to find. You 0:09:25.876 --> 0:09:27.556 can sort of say this is a pick and roll 0:09:27.596 --> 0:09:28.716 this is a pick and roll. This is a pick 0:09:28.716 --> 0:09:29.996 and roll. This is not a pick and roll. This 0:09:30.036 --> 0:09:33.556 is not a pick and roll that kind of thing. Yeah, good, 0:09:33.596 --> 0:09:36.596 So okay, So that's the math, sort of technical side. 0:09:36.676 --> 0:09:39.396 Then there's the getting people to believe you and you know, 0:09:39.676 --> 0:09:43.356 buy what you're selling ultimately, both metaphorically and literally. Tell 0:09:43.396 --> 0:09:47.436 me about that side, like, were there any interesting I 0:09:47.476 --> 0:09:50.036 don't know, people who didn't want to buy it, you know, 0:09:50.116 --> 0:09:53.716 little stories from that side. Almost every coach that we've 0:09:53.756 --> 0:09:55.796 talked to couldn't believe that we did what we believed. 0:09:55.836 --> 0:10:01.036 So the story is generally involved just sitting there in 0:10:01.076 --> 0:10:03.436 front of these coaches. You go up a couple of levels. 0:10:03.796 --> 0:10:07.196 So I think that you know, this happened with two coaches. 0:10:07.236 --> 0:10:10.716 It's a very similar story. They were very well known coaches, 0:10:11.396 --> 0:10:14.436 you know, long resumes, have been around the NBA for 0:10:14.476 --> 0:10:17.676 a long time, won lots of championships. You know, it 0:10:17.676 --> 0:10:20.156 took several meetings to even get an audience with them. 0:10:20.156 --> 0:10:22.276 You would go through sort of layer one, than layer two, 0:10:22.276 --> 0:10:23.796 then layer three, and then you would get to them. 0:10:24.036 --> 0:10:27.196 It's like knitting the Pope or something exactly, and then 0:10:27.196 --> 0:10:29.636 you would sort of get to them and there they were. 0:10:29.876 --> 0:10:32.196 They would just basically say it like, you have no 0:10:32.316 --> 0:10:34.476 idea what I intend for the players to do. There's 0:10:34.516 --> 0:10:37.956 no way this machine can understand what what you need. 0:10:37.996 --> 0:10:40.996 And we would say try us, and we would say, okay, 0:10:41.036 --> 0:10:43.756 show me all the times that the play started with 0:10:43.756 --> 0:10:45.516 the pick and roll and there was three passes and 0:10:45.556 --> 0:10:47.316 I took a shot in the corners like okay, here 0:10:47.316 --> 0:10:48.916 it is like, show me all the times there was 0:10:48.956 --> 0:10:51.676 a screen. I mean like you have a laptop and 0:10:51.716 --> 0:10:54.636 you have your software running or something. Yep, that's right, 0:10:54.676 --> 0:10:56.356 and they would we would have tables and charts and 0:10:56.396 --> 0:10:58.476 they would we would project onto a screen and they 0:10:58.516 --> 0:11:01.596 would basically just give us the quiz, the sort of 0:11:01.596 --> 0:11:03.956 the the They would put us through the ringer for 0:11:03.956 --> 0:11:06.276 for like two hours, so it's like a it's basically 0:11:06.316 --> 0:11:10.796 a impromptu grilling from the coach and make your machine 0:11:10.876 --> 0:11:13.276 dance for me and show me the correct answers. And 0:11:13.356 --> 0:11:14.756 any time you come up with an answer that I 0:11:14.756 --> 0:11:18.196 don't agree with, you know, I wouldn't think you're an idiot. 0:11:18.196 --> 0:11:20.556 But we held up. In fact, sometimes these coaches would say, wait, 0:11:20.596 --> 0:11:22.836 show me this list, and if these two guys aren't 0:11:22.916 --> 0:11:25.236 number one, and number two, I think your thing is wrong, 0:11:25.276 --> 0:11:27.076 and then we would do it, and like thank goodness, 0:11:27.276 --> 0:11:29.156 the right players would end up number one and number 0:11:29.196 --> 0:11:31.876 two and and then normally after sort of two hours 0:11:32.076 --> 0:11:35.636 of grilling, they would be pretty good and they would 0:11:35.636 --> 0:11:37.036 just leave the room, and then we knew we would 0:11:37.036 --> 0:11:45.876 have a contract. And is the output like the thing 0:11:45.956 --> 0:11:48.396 that actually you see on the screen when you're running 0:11:48.396 --> 0:11:51.276 your software. Is it lists? Is it little videos where 0:11:51.276 --> 0:11:53.596 all the players are dots? Is it both of those things? 0:11:53.916 --> 0:11:56.476 It's everything. So we have sort of ranking stables that 0:11:56.516 --> 0:11:58.316 can answer who's the best at X, Y and Z. 0:11:58.476 --> 0:12:01.796 We have sort of a variety of visualizations that show 0:12:02.596 --> 0:12:05.836 breakdowns of various actions. You can always click on anything 0:12:05.916 --> 0:12:08.356 and show all the video of every moment you know 0:12:08.516 --> 0:12:11.196 anything that you asked it. So we've built lots of 0:12:11.236 --> 0:12:14.516 visualizations and data formats to make it easy for various 0:12:14.556 --> 0:12:17.196 people in the organizations to use them. There are instances 0:12:17.196 --> 0:12:19.956 in the world in different domains where like a data 0:12:20.036 --> 0:12:23.156 driven approach suggests doing one thing, but that thing is 0:12:23.276 --> 0:12:26.756 contrary to conventional wisdom, right, But like in football, it 0:12:26.756 --> 0:12:30.876 seems pretty clear that the data suggests coaches should go 0:12:30.916 --> 0:12:32.756 for it on fourth down more often than they do, 0:12:32.876 --> 0:12:35.116 and it seems like as the result, coaches have started 0:12:35.156 --> 0:12:36.996 going for it more on fourth down. But there is 0:12:37.036 --> 0:12:38.836 this thing where if a coach goes for it on 0:12:38.876 --> 0:12:41.796 fourth down and doesn't make it, the team doesn't get 0:12:41.836 --> 0:12:45.076 the first down, then the coach gets pillaried. Right. So 0:12:45.556 --> 0:12:48.396 even though the data says it makes sense to go 0:12:48.516 --> 0:12:51.436 for it in a kind of a personal incentives way, 0:12:51.556 --> 0:12:53.996 it might not make sense right in a social way, 0:12:54.036 --> 0:12:56.476 which is real and important, it might not make sense. 0:12:57.196 --> 0:12:59.396 Is there a basketball version of that? Is there a 0:12:59.476 --> 0:13:02.156 version of that you have observed where your data suggests 0:13:02.156 --> 0:13:04.436 that coaches should be doing something, but they're reluctant to 0:13:04.476 --> 0:13:08.636 do it because it's contrary to sort of conventional wisdom. No, 0:13:08.716 --> 0:13:10.596 I think there's a lot more nuance and what we do. 0:13:10.636 --> 0:13:14.596 I think there's the public basically is aware that they say, oh, 0:13:14.596 --> 0:13:16.956 people should take more threes. But I think the reality 0:13:17.076 --> 0:13:19.476 is much more complex than that, because I think it's 0:13:19.556 --> 0:13:22.116 not about just simply take a three. What you want 0:13:22.116 --> 0:13:24.476 to do is get the best possible shot for yourself. 0:13:24.516 --> 0:13:27.116 Sometimes that's a three, sometimes that's something else. Different kinds 0:13:27.156 --> 0:13:29.476 of threes are not the same, and your talent is 0:13:29.476 --> 0:13:30.916 not the same. In fact, if you look at all 0:13:30.956 --> 0:13:33.196 the you know that you know, I'm just gonna sort 0:13:33.236 --> 0:13:35.676 of Lebron James and Kevin Durant and Steph Curry and 0:13:35.756 --> 0:13:38.156 Janis and Yo Kitchen, all these players, and you know 0:13:38.516 --> 0:13:41.876 they are very different players, and teams are built around 0:13:41.916 --> 0:13:44.116 these players, and what you want to do is optimize 0:13:44.156 --> 0:13:46.596 the team built around these very unique players. And so 0:13:46.956 --> 0:13:48.556 a lot of it is in the details of how 0:13:48.596 --> 0:13:51.356 you structure your offense and your defense to sort of 0:13:51.476 --> 0:13:54.116 use them to get your team the best possible shot. 0:13:54.316 --> 0:13:55.916 So a lot of it is that these coaches do 0:13:55.956 --> 0:13:58.636 a lot of work on basically on the micro level 0:13:58.676 --> 0:14:00.876 to sort of create these plays. That it's not just 0:14:00.876 --> 0:14:03.596 sort of walk up the court and take a three. 0:14:03.676 --> 0:14:05.076 They do a lot of work to try and put 0:14:05.116 --> 0:14:07.996 their position their players in positions to have a lot 0:14:08.036 --> 0:14:10.716 of good options to get the best possible shot. So 0:14:10.916 --> 0:14:13.316 I think that it's it's really a lot more nuanced 0:14:13.356 --> 0:14:21.196 in the actual execution of it. Fans would be really 0:14:21.236 --> 0:14:26.996 surprised at the level of sophistication entirely across the league 0:14:27.036 --> 0:14:29.676 and many leagues in sports, especially the top ones. There's 0:14:29.716 --> 0:14:32.236 there's a degree of sophistication and how they use data 0:14:32.276 --> 0:14:37.796 and video that I think would surprise almost everyone. You 0:14:37.916 --> 0:14:41.036 think that the sort of play calling side of coaching 0:14:41.076 --> 0:14:43.796 will become more and more delegated to the machine. I mean, 0:14:43.836 --> 0:14:49.116 I could see coaches as essentially psychologists, being persistently useful. 0:14:49.156 --> 0:14:51.796 But do you see a time when the sort of 0:14:51.836 --> 0:14:55.796 core strategic decision making will just be be done better 0:14:55.796 --> 0:14:58.156 by a machine than by coach. I don't. I don't 0:14:58.156 --> 0:15:01.556 think so. Our thesis has always been we build iron 0:15:01.556 --> 0:15:04.996 Man suits and and different people will want iron you know, 0:15:05.116 --> 0:15:07.556 different iron Man suits, and you will want because you 0:15:07.596 --> 0:15:09.276 want you need somebody in the middle of the iron 0:15:09.356 --> 0:15:12.196 Man suit directing it, using everything they know about the world. 0:15:12.356 --> 0:15:13.956 But you want to be a lot more powerful. And 0:15:14.036 --> 0:15:16.676 I think that that's the model we've always used, is 0:15:17.036 --> 0:15:19.356 we build Iron Man suits for everybody, and then different 0:15:19.396 --> 0:15:21.356 coaches and different assistant coaches are going to use them, 0:15:21.636 --> 0:15:23.596 you know, to be a lot more powerful. That's what 0:15:23.636 --> 0:15:25.156 we want to do. We want to build Iron Man 0:15:25.196 --> 0:15:30.076 suits for everybody. Does basketball look different? Is it played 0:15:30.116 --> 0:15:35.636 differently because of your work? I certainly know that there 0:15:35.676 --> 0:15:37.876 have been examples where you know, in a couple of 0:15:37.996 --> 0:15:41.716 NBA Finals strategies were changed because of the data that 0:15:41.796 --> 0:15:45.036 we provided. You know, close ones are where one team 0:15:45.076 --> 0:15:48.596 basically discovered a pick and roll defense strategy that was 0:15:48.636 --> 0:15:51.996 effective and employed that and it to you know, very 0:15:51.996 --> 0:15:53.796 good effect. Another one it was one of these sort 0:15:53.796 --> 0:15:57.756 of off ball screen defense strategies where they found the 0:15:57.836 --> 0:16:00.476 data show that there were particular strategies because that would 0:16:00.476 --> 0:16:02.316 shave off a couple of points of efficiency from the 0:16:02.316 --> 0:16:04.956 other team, and they employed that very heavily and they 0:16:04.996 --> 0:16:06.996 won a very close series. So I think, can you 0:16:07.036 --> 0:16:10.956 say what teams it was? I should Okay, I get 0:16:10.996 --> 0:16:13.996 people looking at they could probably figure it out. The 0:16:14.076 --> 0:16:15.476 point that I want to realize, like there are a 0:16:15.476 --> 0:16:16.916 lot of people who come out and think like, oh, 0:16:16.956 --> 0:16:19.676 these these teams are sort of the geniuses and these 0:16:19.676 --> 0:16:21.916 teams are the Luddites, and that's not the case. Like 0:16:22.196 --> 0:16:25.196 almost every team is. Like people would be surprised that 0:16:25.396 --> 0:16:27.556 at the minimum, that's fine. I mean, everybody in the 0:16:27.676 --> 0:16:30.596 NBA is every player in the NBA is good at 0:16:30.636 --> 0:16:33.596 playing basketball, but some are better than others. And so similarly, 0:16:33.596 --> 0:16:36.276 you might imagine that every sort of quant team in 0:16:36.316 --> 0:16:38.956 the NBA is good at at being you know, doing 0:16:39.036 --> 0:16:42.196 quantitative analysis, but some are probably better than others. So 0:16:42.236 --> 0:16:43.716 I would say, you're right, there are some who are 0:16:43.756 --> 0:16:45.716 better than others, but the minimum level is a lot 0:16:45.796 --> 0:16:48.876 higher than when everyone would sure, yeah, yeah, And can 0:16:48.916 --> 0:16:53.276 you say who's better than others? I cannot. I I 0:16:53.476 --> 0:16:56.236 this is our lifeblood of being the reason that all 0:16:56.276 --> 0:16:59.476 these coaches trust us and is that we don't. We 0:16:59.476 --> 0:17:02.156 don't leak. So I'm sorry. It would be fun one day, 0:17:02.276 --> 0:17:04.516 One day, I'll call all the stories. It's not okay, 0:17:04.516 --> 0:17:10.476 I'm sorry. Regiev is not going to name names today Alas, 0:17:10.916 --> 0:17:13.236 but he is going to talk about a big, interesting 0:17:13.276 --> 0:17:16.436 problem that second spectrum is close to solving but has 0:17:16.516 --> 0:17:27.356 not quite nailed yet. That's in a minute. Now back 0:17:27.356 --> 0:17:30.876 to the show. I want to talk about things you 0:17:30.916 --> 0:17:33.676 haven't figured out yet, like what are the next problems 0:17:33.676 --> 0:17:36.156 you're trying to solve, Like what is the frontier? So 0:17:36.236 --> 0:17:38.836 I think the frontier where we're trying to basically basically 0:17:38.876 --> 0:17:40.836 get full body pose of the human, not just a 0:17:40.876 --> 0:17:43.436 little dot. We're trying to get their entire skeleton. So 0:17:43.516 --> 0:17:46.236 the basic sort of product you have, the basic thing 0:17:46.236 --> 0:17:48.716 you do turns each player into a dot, just like 0:17:48.756 --> 0:17:51.156 in the classic like x's and o's diagram or whatever. 0:17:51.236 --> 0:17:52.876 And so you can watch little dots moving around and 0:17:52.876 --> 0:17:54.636 you can see how far the dots are from the other. 0:17:55.076 --> 0:17:57.476 But there's a lot going on that you don't see 0:17:57.476 --> 0:17:59.196 in the dot, right you might want to know, like 0:17:59.276 --> 0:18:01.156 what are the things that are important that you don't 0:18:01.156 --> 0:18:04.396 see in the dot? That's exactly right. So the challenge 0:18:04.396 --> 0:18:06.396 we're working out right now is turning the dot into 0:18:06.436 --> 0:18:10.036 a human skeleton and then having that skeleton and generate 0:18:10.116 --> 0:18:12.596 data in you know, one hundred to two hundred milliseconds. 0:18:12.756 --> 0:18:14.436 So that's the challenge that we're working on right now, 0:18:14.436 --> 0:18:16.356 and we've actually made a lot of progress on that, 0:18:16.436 --> 0:18:19.236 and that's right now. And so what's an example of 0:18:19.516 --> 0:18:21.236 why that would be useful? What are some of the 0:18:21.276 --> 0:18:24.716 ways you could learn from seeing each player as a 0:18:24.756 --> 0:18:28.036 whole body rather than as a dot. So I think 0:18:28.036 --> 0:18:29.316 there are a lot of things you can do. There's 0:18:29.356 --> 0:18:32.836 one is, obviously, you know, with health and fitness, where 0:18:32.836 --> 0:18:34.716 you can figure out did they land on one foot, 0:18:34.716 --> 0:18:36.356 did they land on two feet? You know, did they 0:18:36.436 --> 0:18:38.316 you know they take off from this port of that foot. 0:18:38.916 --> 0:18:40.316 So there's a lot of health and fitness stuff that 0:18:40.316 --> 0:18:43.636 you can figure out health and fitness, meaning learning to 0:18:44.436 --> 0:18:47.636 reduce the risk of injury, learning that certain kinds of 0:18:47.876 --> 0:18:52.476 moves or subtle distinctions in moves make a player more 0:18:52.636 --> 0:18:54.956 or less likely to be injured. That's right, And so 0:18:54.956 --> 0:18:57.156 that's one class of things we can do. Another class 0:18:57.196 --> 0:18:59.716 is sort of just understanding moves better. So if a 0:18:59.796 --> 0:19:03.636 player made, you know, dribble moves between their legs and 0:19:03.676 --> 0:19:06.636 crossovers and step backs and did all all kinds of 0:19:06.636 --> 0:19:08.916 things with their hands and feet to get open, now 0:19:08.916 --> 0:19:12.356 we can categorize those things. There's that, there's health, there's strategy, 0:19:12.636 --> 0:19:15.716 there's media, there's officiating. There's you can help with officiating 0:19:15.756 --> 0:19:17.556 if you start knowing where people are. So there's a 0:19:17.596 --> 0:19:19.876 lot of things that you can do once you understand 0:19:19.916 --> 0:19:22.196 more of the human skeleton. And let me ask you this, 0:19:22.996 --> 0:19:25.556