Latest / What's Your Problem? / Using Computer Vision to See What Coaches Can’t

Transcript

Using Computer Vision to See What Coaches Can’t · What's Your Problem?

0:08
Speaker 1: Pushkin.
0:16
Speaker 2: Because I didn't know that sports biomechanics could be a career.
0:20
Speaker 2: I decided to go to grad school and initially start
0:22
Speaker 2: working on prosthetic limbs. So that was the first two
0:26
Speaker 2: years of grad school. Then about two years in I
0:31
Speaker 2: discovered baseball pitching biomechanics research. The funny thing that happened
0:36
Speaker 2: there was I gave a PhD committee meeting where I
0:40
Speaker 2: spent most of the time talking about prosthetic limbs, and
0:44
Speaker 2: then I spent the last few minutes as an aside
0:47
Speaker 2: on the baseball pitching research I found. And my committee
0:50
Speaker 2: was like, Jimmy, the last few minutes were way better
0:53
Speaker 2: than the first forty fives.
0:55
Speaker 3: So I was like, all right.
0:57
Speaker 2: So then they were like, we'll let you do baseball
1:00
Speaker 2: pitching biomechanics as your PhD work, but just so you know,
1:04
Speaker 2: it might be really hard to have a career doing that.
1:08
Speaker 2: But you know, they were like, if you want to
1:09
Speaker 2: go for it, you can go for it. And so
1:10
Speaker 2: I was like, you know what, Sure, I'll go for it.
1:19
Speaker 1: I'm Jacob Goldstein, and this is what's your problem. Today
1:22
Speaker 1: we have the second episode in our series about people
1:25
Speaker 1: who are working at the frontiers of technology to help
1:29
Speaker 1: a lead athletes perform better. My guest today is Jimmy Buffy,
1:33
Speaker 1: and as it happened, the concerns of his grad school
1:36
Speaker 1: advisors were unfounded. Jimmy has in fact made a career
1:40
Speaker 1: out of the biomechanics of pitching in baseball and sports
1:44
Speaker 1: biomechanics more broadly. When he finished grad school, he got
1:48
Speaker 1: a job with the Los Angeles Dodgers, and he went
1:51
Speaker 1: on to co found a company called Reboot Motion that
1:54
Speaker 1: works with teams in Major League Baseball and the NBA.
1:57
Speaker 1: Jimmy's problem is this, how do you take massive amounts
2:01
Speaker 1: of data about how professional athletes move and turn all
2:05
Speaker 1: that data into information that actually helps those athletes perform better.
2:12
Speaker 1: You end up doing your dissertation research on the on
2:15
Speaker 1: the biomechanics of pitching, Yes, of baseball, pitching and baseball. Yeah,
2:19
Speaker 1: and and then you get hired by the Dodgers.
2:23
Speaker 3: Yeah.
2:23
Speaker 2: Yeah, that was That was awesome because I originally didn't again,
2:27
Speaker 2: didn't realize that that could be a thing that could happen.
2:30
Speaker 1: Well, and it kind of wasn't right, Like, you're kind
2:33
Speaker 1: of just coming into this field as it's becoming a
2:36
Speaker 1: field where you can get a job where it's a.
2:38
Speaker 3: Field, right exactly. Yeah, there.
2:40
Speaker 2: I mean the challenge then was there wasn't a lot
2:42
Speaker 2: of options for actually even getting the data that you
2:45
Speaker 2: need to analyze.
2:47
Speaker 1: So ten years ago, like what is the state of
2:51
Speaker 1: play in this sort of nasson field that you're in
2:54
Speaker 1: helping to create?
2:56
Speaker 2: So the field is, I would say, is like sports biomechanic,
3:00
Speaker 2: and what that is is being able to analyze the
3:03
Speaker 2: movement of athletes for lots of purposes, help them reduce
3:07
Speaker 2: injury risk, help them improve performance.
3:10
Speaker 1: And to be clear, like folk, sports biomechanics has been
3:14
Speaker 1: around forever, right, that's what coaches do. They stay out
3:16
Speaker 1: there and they watch it. Yeah, And so it's kind
3:18
Speaker 1: of becoming it's becoming more technical, right, the field is
3:21
Speaker 1: becoming more technical.
3:22
Speaker 3: Yeah.
3:22
Speaker 2: So the state of the art relied on what is
3:26
Speaker 2: called marker based motion capture, which is where you literally
3:31
Speaker 2: put reflective markers like little balls, you stick them all
3:36
Speaker 2: over somebody's body. Usually the person has to like strip
3:39
Speaker 2: their clothes off because you want the markers like literally
3:41
Speaker 2: like on the skin, on the joints, and then you
3:45
Speaker 2: have these special cameras that track those markers.
3:48
Speaker 1: So you've got like a picture in his underwear with
3:50
Speaker 1: a bunch of little metal balls, takes them and they're like,
3:54
Speaker 1: just pitch like you always pitch.
3:55
Speaker 3: Right, And that's the challenge.
3:57
Speaker 2: That's why that wasn't That's why it wasn't very wide
3:59
Speaker 2: SPA is a thing people did because it was so
4:02
Speaker 2: hard to collect the data because ultimately, what you would
4:04
Speaker 2: need is you need the data on how someone is moving.
4:08
Speaker 3: You need to track.
4:09
Speaker 2: Where their elbow is, where their wrist is, where their
4:11
Speaker 2: knees are so that you can analyze it. And the
4:13
Speaker 2: state of the art for tracking it was an awful
4:16
Speaker 2: experience for the people you were trying to track.
4:18
Speaker 1: That presumably would mean they didn't pitch like they usually exactly,
4:22
Speaker 1: because they don't usually stand there in their underwear with
4:25
Speaker 1: met Were they really in their underwear by the way,
4:27
Speaker 1: I'm saying it because it sounds absurd, but is that
4:28
Speaker 1: actually what they were doing.
4:29
Speaker 2: That's actually you strip down to your your boxers, your
4:32
Speaker 2: boxer briefs, and that's it, that's all you're wearing.
4:35
Speaker 1: So it basically didn't work, and it basically wasn't very
4:38
Speaker 1: widely used as a result.
4:39
Speaker 2: Right exactly, Yeah, you yeah, you look at studies in
4:43
Speaker 2: that field, and people would be throwing like several miles
4:47
Speaker 2: an hour slower than they would be throwing when they
4:49
Speaker 2: weren't wearing all that stuff. So what changes computer vision?
4:56
Speaker 2: That was the big That was the big inflection point. Now,
5:00
Speaker 2: to be fair, like, even when I was finishing my PhD,
5:04
Speaker 2: and I'll give them, I'll give them a shout out,
5:06
Speaker 2: there was a company that was already like working really
5:10
Speaker 2: hard to solve this problem for baseball teams that I
5:13
Speaker 2: was getting that I got to be familiar with, called Kinnetrax.
5:18
Speaker 2: But yeah, the big inflection point was computer vision, basically
5:21
Speaker 2: using artificial intelligence to identify where those joints are in
5:26
Speaker 2: a camera image rather than needing to paste those reflective markers.
5:32
Speaker 1: So computer vision takes off. You're working at the Dodgers,
5:35
Speaker 1: and then eventually in twenty nineteen, right, you leave the
5:39
Speaker 1: Dodgers and you start your company a reboot motion, what
5:43
Speaker 1: does your company do?
5:45
Speaker 2: We do what we call biomechanics as a service, So
5:49
Speaker 2: we try to analyze this computer vision data at a
5:52
Speaker 2: very large scale to help teams and coaches make use
5:56
Speaker 2: of it to help athletes get better. Bas Yeah, that says.
6:03
Speaker 1: But bess Yah and and who are your customers?
6:06
Speaker 2: Our customers are Major League Baseball teams actually NBA teams,
6:14
Speaker 2: so we've gotten into basketball also sort of like league
6:19
Speaker 2: wide data providers. So yeah, leagues teams is basically our
6:23
Speaker 2: sweet spot.
6:24
Speaker 1: So let's talk. Let's talk in like a little more
6:27
Speaker 1: detail about about what you actually do, right, tell me
6:30
Speaker 1: the story of what you do.
6:33
Speaker 2: So, Evan, my co founder, Evan Demchik, he likes to
6:37
Speaker 2: call this the biomechanics Trainkay.
6:41
Speaker 1: Let's take a ride on the biomechanics trend.
6:43
Speaker 2: And we call it that because with the way our
6:45
Speaker 2: product works is we let people get on the train
6:49
Speaker 2: at whatever stop works for them, and get off the
6:51
Speaker 2: train at whatever stop works for them.
6:53
Speaker 1: How far are we going to go with this metaphor
6:55
Speaker 1: of a little o'b nervous about it?
6:57
Speaker 3: That might be as far as we go.
6:59
Speaker 1: Okay, good, good, So just tell it to me. Start
7:04
Speaker 1: at whatever seems like the beginning of a you know,
7:07
Speaker 1: a of an encounter, and so I'd like to understand
7:10
Speaker 1: how that works. That's really kind of a way to
7:12
Speaker 1: think about it. So let's start with the data, right,
7:16
Speaker 1: what is a basic what is the basic thing that's.
7:18
Speaker 2: Happening So the very first thing that happens is you
7:22
Speaker 2: record videos of the athlete doing the athletic motion. So
7:27
Speaker 2: we'll talk about pitching, So you record videos of a
7:30
Speaker 2: pitcher pitching. Our product, we actually have implemented our own
7:34
Speaker 2: computer vision models, so we can do that if people want.
7:38
Speaker 2: But generally speaking, people have systems like Kinetrax is one
7:43
Speaker 2: I've mentioned that, Hawkeye is another popular one where there
7:47
Speaker 2: is a system in place that has the cameras that
7:49
Speaker 2: record the videos and then runs those computer vision models.
7:53
Speaker 2: What those computer vision models do is they extract the
7:56
Speaker 2: locations in three dimensional space of all of like the
8:00
Speaker 2: joint centers. So where's my elbow, where's my wrist, where's
8:03
Speaker 2: my knee? In three dimensional space? That's what comes out
8:07
Speaker 2: of these computer vision systems.
8:08
Speaker 1: Okay, so pretty much everybody has that at this point,
8:12
Speaker 1: Like every professional baseball team has that for every pitch
8:16
Speaker 1: in every game at this point.
8:17
Speaker 3: Yes, exactly.
8:18
Speaker 2: Okay, So it's a ton a ton of data. So
8:21
Speaker 2: that's another challenge that we've solved, is not only how
8:24
Speaker 2: do you do this, but how do you do this
8:25
Speaker 2: at a very large scale?
8:27
Speaker 1: Okay, so this data, everybody's got it now, and you're not.
8:31
Speaker 1: You can sort of process the data, but that's not
8:33
Speaker 1: your special sauce. That's not your secret sauce, right, So
8:36
Speaker 1: typically they'll send you that that data of like, here's
8:39
Speaker 1: all the body points. Here is how they're moving in
8:41
Speaker 1: physical space. Then what do you do with it?
8:44
Speaker 2: So, yeah, this is where the special sauce comes in.
8:47
Speaker 2: So the first step to that is how do you
8:49
Speaker 2: turn those key points into a human skeleton. So you've
8:53
Speaker 2: got to figure out, like where do the bones connect,
8:55
Speaker 2: what sort of degrees of freedom do those bones have,
8:58
Speaker 2: So then you can figure out how do those key
9:01
Speaker 2: points animate a human skeleton.
9:05
Speaker 1: So you're sort of rebuilding it. It's like you start
9:07
Speaker 1: with the picture of a person and then you turn
9:08
Speaker 1: it into a bunch of data points, and now you've
9:10
Speaker 1: got to kind of build the person back up again from.
9:13
Speaker 2: The day exactly exactly. So now you have an actual
9:17
Speaker 2: human skeleton where the shoulder is rotating, the elbow is flexing,
9:21
Speaker 2: the knee is flexing, the hips are rotating. And now
9:24
Speaker 2: once you do that, now you can understand that data
9:27
Speaker 2: in the context of how the body works. Okay, so
9:30
Speaker 2: once we've done that, then we calculate how energy flows
9:34
Speaker 2: through the body, we calculate how momentum flows through the body,
9:37
Speaker 2: and once we've done that, we can analyze how efficiently
9:40
Speaker 2: the athlete is moving. Are they generating energy and momentum
9:44
Speaker 2: in the direction that they want to generate in? What
9:47
Speaker 2: is that desired direction? So we then we calculate all
9:50
Speaker 2: sorts of metrics around movement efficiency and direction. Then once
9:54
Speaker 2: we calculate all those metrics, now we can understand how
9:58
Speaker 2: those relate to what you're trying to do. Throw the
10:00
Speaker 2: ball as hard as possible.
10:02
Speaker 1: So so presumably with the picture, what you want to
10:05
Speaker 1: optimize for is is having as much of the picture's
10:09
Speaker 1: energy of their body go toward making the ball go
10:13
Speaker 1: toward home plate. Exactly right, I mean that is that? Yeah,
10:15
Speaker 1: they mail optimization problem mail it.
10:17
Speaker 2: Yeah, Yeah, that's exactly it. And that's the problem that
10:20
Speaker 2: we try to understand. So when we build our sort
10:22
Speaker 2: of models regarding like how does a pitcher create efficient
10:25
Speaker 2: fastball velocity, one of the most important things that comes
10:28
Speaker 2: out of those models is lining up the direction of
10:31
Speaker 2: your torsore rotation with the direction of your arm rotation.
10:34
Speaker 1: The pitching bution is crazy complex, right, Like they're they
10:37
Speaker 1: kick their leg up and they got their front arms
10:39
Speaker 1: doing something and their back arms going back like a
10:42
Speaker 1: lot is happening. Yeah, the sort of platonic ideal is
10:46
Speaker 1: every little millimeter of every motion is going toward maximizing
10:50
Speaker 1: the energy of the ball going toward home plate.
10:52
Speaker 2: Yes, and not just not just maximizing like in a vacuum,
10:56
Speaker 2: but doing it in the most efficient way, because if
10:59
Speaker 2: you just sort of maximize in a vacuum, maybe you're
11:02
Speaker 2: transferring that energy in a way that hurts your elbow.
11:04
Speaker 2: You're transferring that energy in a way that hurts your shoulder.
11:07
Speaker 2: So not only do we figure out how can a
11:09
Speaker 2: pitcher maximize it, but we try to figure out how
11:11
Speaker 2: they can have that energy and momentum go in a
11:14
Speaker 2: direction that doesn't hurt their joints, So try to have
11:18
Speaker 2: them throw the ball a little harder while also reducing
11:21
Speaker 2: their injury risk a little bit.
11:23
Speaker 1: So you're whatever doing the math at reboot HQ, and
11:30
Speaker 1: then what are you sending back to the.
11:33
Speaker 2: Team, So some teams so it sort of so again,
11:38
Speaker 2: all right, I told you that was the end of
11:39
Speaker 2: the biomechanics train analogy.
11:41
Speaker 3: I'm going to bring it back for a brief second.
11:43
Speaker 1: Okay, I'm ready.
11:45
Speaker 2: So we go all the way to building a report
11:48
Speaker 2: that has a bunch of suggestions. So there's a report
11:51
Speaker 2: that's like, this is how efficient you are. You can
11:53
Speaker 2: like tilt your torso a little bit to be a
11:55
Speaker 2: little bit more efficient. You can tilt your arm a
11:57
Speaker 2: little bit to be a little bit more efficient. So
11:59
Speaker 2: we go all the way to generating a report that's
12:02
Speaker 2: like the final stop on the train.
12:04
Speaker 1: And that's a report for for one pitcher.
12:06
Speaker 3: Yeah, that's a report on a pitcher for whatever.
12:09
Speaker 1: And and in a way that's like the nerdiest it's
12:13
Speaker 1: what a pitching coach does, but just in a way
12:15
Speaker 1: nerdier way.
12:16
Speaker 3: Yeah.
12:16
Speaker 2: Well, so we try to sort of like give the
12:20
Speaker 2: coaches superpowers, you know, even though that you know, the
12:23
Speaker 2: coaches like can look at an athlete and understand a
12:25
Speaker 2: lot about the athlete just by looking at them. We
12:29
Speaker 2: try to make a report that can really amplify what
12:33
Speaker 2: the coach is already doing, maybe help them discover some
12:36
Speaker 2: things they weren't thinking about, or measure some things that
12:39
Speaker 2: they were thinking about, but now they can track those
12:41
Speaker 2: a little bit easier. So that's you know, that's the
12:44
Speaker 2: ultimate thing that we produces a report that can sort
12:47
Speaker 2: of like we say, give a coach superpowers.
12:51
Speaker 1: Where was the trained metaphor doing that? Is there an
12:53
Speaker 1: earlier station of disembarkation exactly?
12:57
Speaker 2: So a lot of teams now are hiring people with
13:01
Speaker 2: biomechanics and analytics backgrounds, so rather than just use our
13:06
Speaker 2: reports out of the box, they want to build their
13:09
Speaker 2: own reports and their own statistical models and their own
13:11
Speaker 2: AI models. So we also get let people get off
13:14
Speaker 2: the train a little bit earlier and build whatever they
13:17
Speaker 2: want on top of the data that we're generating.
13:19
Speaker 1: Are they not going to disintermediate you once those people
13:22
Speaker 1: are there? Does that reduce the value you provide to
13:25
Speaker 1: the team.
13:26
Speaker 2: Now because we still have to process all that data.
13:29
Speaker 1: Do you have some like IP or like why can't
13:32
Speaker 1: somebody like you who works for a team just do
13:35
Speaker 1: that without you?
13:36
Speaker 2: That is a great question. We answered this question all
13:39
Speaker 2: the time. Is because it's a very complex engineering problem.
13:44
Speaker 2: Not only to do all the math that the physics
13:47
Speaker 2: based math, to calculate the energy, calculate the momentum, like
13:50
Speaker 2: all of that math is really hard, but also to
13:54
Speaker 2: do it at a very large scale. So someone like
13:58
Speaker 2: me in grad school learned how to do that on
14:00
Speaker 2: a sample size. You know, my PhD was actually really
14:03
Speaker 2: just one pitch, but lots of people do it on
14:06
Speaker 2: like ten pitches or maybe one hundred pitches, So we
14:10
Speaker 2: do it on several thousand pitches like and swings like
14:16
Speaker 2: every morning.
14:18
Speaker 3: You know.
14:18
Speaker 2: There's just every team has like seven affiliates, so there's
14:23
Speaker 2: you know, one hundred and fifty games every day that
14:26
Speaker 2: need to be.
14:26
Speaker 1: So you're doing this on farm teams as well.
14:28
Speaker 3: Yeah, exactly.
14:29
Speaker 2: So not only did we solve the problem of doing
14:32
Speaker 2: it for like one pitcher in a way that's really actionable,
14:35
Speaker 2: but we solved the problem doing it for every game
14:38
Speaker 2: every day, you know, so that you have the data
14:41
Speaker 2: when you wake up.
14:44
Speaker 1: So you guys are doing it at scale. The answer
14:46
Speaker 1: to why it is that that there is in fact
14:48
Speaker 1: an economy of scale, a benefit of scale.
14:51
Speaker 3: Yes, which you have, Yes, exactly, Yeah. Yeah.
14:54
Speaker 1: What's what's a specific example of a thing that a
14:58
Speaker 1: coach might tell a pitcher in response to your report
15:01
Speaker 1: to try and get them to throw whatever differently?
15:05
Speaker 3: A really.
15:08
Speaker 2: Common low hanging fruit type of piece of feedback that
15:12
Speaker 2: often comes out of the reports is how a pitcher
15:16
Speaker 2: is using their lead arm in a typical pitching motion,
15:21
Speaker 2: the pitcher will reach forward with their lead arm.
15:24
Speaker 1: So, just to be clear, the lead arm is the
15:25
Speaker 1: arm that is not holding the ball.
15:27
Speaker 3: Right right, yeah, the arm in front of you.
15:29
Speaker 1: Yeah.
15:29
Speaker 3: Yeah.
15:30
Speaker 2: A pitcher will reach forward with that lead arm while
15:32
Speaker 2: the rear arm is holding the baseball, and they'll rotate
15:35
Speaker 2: that lead arm really hard, and that's the thing that
15:39
Speaker 2: kind of initiates the torso rotation. So a very common
15:43
Speaker 2: flaw that we see is if a pitcher has a
15:47
Speaker 2: very vertical pitching arm, they're pitching the ball way over
15:51
Speaker 2: the top of their head, but their lead arm when
15:54
Speaker 2: they pull it through, when they swing it through, they
15:56
Speaker 2: swing it in a very flat.
15:58
Speaker 1: Plane as horizontal.
16:01
Speaker 2: Horizontal, right, yeah, That is not a very efficient plane
16:05
Speaker 2: to use when you're throwing the ball on a very
16:07
Speaker 2: vertical plane. So a very common low hang piece of
16:10
Speaker 2: fruit feedback that comes out of the reports is having
16:15
Speaker 2: pitchers just try to rotate their lead arm pull with
16:19
Speaker 2: their lead arm in a more vertical plane to better
16:21
Speaker 2: match what their torso is doing, a better match with
16:24
Speaker 2: their pitching arm is doing.
16:26
Speaker 1: That's a good one. I feel like that one's so
16:28
Speaker 1: simple that you don't want it to get out that
16:30
Speaker 1: everybody ell to start looking at it.
16:31
Speaker 2: No, I mean really, I mean like it's this has
16:34
Speaker 2: happened when I went to talk to a team and
16:37
Speaker 2: we talked about some pieces, you know, some low hanging fruit,
16:40
Speaker 2: and they're like, okay, great, we'll take the lead arm thing,
16:43
Speaker 2: will implement it everywhere. And I'm like, well, what about
16:45
Speaker 2: like the other ten pages of the report.
16:49
Speaker 3: Good with the lead arm thing.
16:52
Speaker 1: Thanks by so for you. The end of the train
16:56
Speaker 1: is the report. But that report goes to the coach, right,
17:00
Speaker 1: and so presumably the meaningful change hasn't happened yet, right,
17:03
Speaker 1: It has to somehow get from the coach to the picture.
17:06
Speaker 1: And like, I know that piece of it is not
17:09
Speaker 1: your business now, but it was kind of your business
17:11
Speaker 1: when you were at the Dodgers. Presumably you're familiar with
17:14
Speaker 1: it now, Like, how does that piece of it work?
17:16
Speaker 1: Is it like the pitching coach is like reading from
17:18
Speaker 1: the report to the picture. I imagine not, but I
17:20
Speaker 1: don't know.
17:20
Speaker 2: No, definitely, definite, definitely not. Even at the Dodgers, my
17:27
Speaker 2: role was not being the one to coach the players.
17:32
Speaker 1: It's like, whatever you do, don't talk to the pictures. Man,
17:34
Speaker 1: go back to your computer.
17:36
Speaker 2: Yeah no, no, I mean thankfully I got to be
17:40
Speaker 2: in the in the room as you know, dark Walld,
17:43
Speaker 2: the interaction is happening. But I think that is the
17:47
Speaker 2: the art of coaching that is so important is understanding
17:52
Speaker 2: the picture and how the pitcher thinks about themselves and
17:56
Speaker 2: giving the right feedback to have the picture do the
18:00
Speaker 2: thing that you that you want them to do.
18:02
Speaker 1: Uh huh. Knowing how to talk to a player in
18:06
Speaker 1: a way that is not generic. Presumably different pitchers need
18:09
Speaker 1: to hear different things, even if the outcome is the same, right.
18:12
Speaker 2: I mean, there's a there's a classic like debate in
18:15
Speaker 2: baseball of like do you swing down on the ball
18:19
Speaker 2: or do you swing with an uppercut? And in reality,
18:22
Speaker 2: like the batpath is an arc, the path goes down
18:25
Speaker 2: and then the path goes up. But some coach, some
18:28
Speaker 2: players respond better when a coach will say swing down
18:32
Speaker 2: on the ball, and some players spawn better when a
18:35
Speaker 2: coach will say, you know, have get a little bit
18:37
Speaker 2: more uppercut, do your swing. But in reality, you know,
18:39
Speaker 2: the bat goes down and the bat goes up. So
18:41
Speaker 2: it's like really understanding what is the player what helps
18:44
Speaker 2: the player the most?
18:45
Speaker 1: So I heard you use this phrase in another interview
18:48
Speaker 1: that I think is kind of what you're talking about here,
18:50
Speaker 1: And it's feel versus real? What is that? What is
18:54
Speaker 1: feel versus real?
18:55
Speaker 2: So it's exactly what we're talking about. Is sometimes with
18:59
Speaker 2: the apt let feels like they're doing is not.
19:03
Speaker 3: What's actually happening.
19:05
Speaker 2: But if you understand what the athlete feels like they're doing,
19:08
Speaker 2: you can give feedback that interacts with how they're feeling.
19:12
Speaker 2: So if they feel like they're swinging down on the
19:16
Speaker 2: ball and you give them feedback related to that, even
19:18
Speaker 2: if they're actually swinging with an uppercut, you know, the
19:21
Speaker 2: important thing is like how do they feel and how
19:24
Speaker 2: do you give them feedback related to how they feel,
19:26
Speaker 2: which then impacts is real. But it's understanding the interplay
19:30
Speaker 2: between the feel and the real.
19:32
Speaker 1: Yeah, it's wild that, like there is there is a
19:36
Speaker 1: human being doing a thing, and then there's this huge
19:39
Speaker 1: industrial machinery of your company and the team and all
19:44
Speaker 1: of these scientists and all these cameras and computers that
19:47
Speaker 1: are trying to get this human being to change their
19:50
Speaker 1: behavior in a very subtle way.
19:52
Speaker 3: Yeah, exactly.
19:53
Speaker 1: It's a behavior that is in many ways intuitive right,
19:57
Speaker 1: it's sort of partly conscious but partly intuitive him like
20:00
Speaker 1: there's a really interesting human being at the center.
20:05
Speaker 3: Of all of this.
20:06
Speaker 2: Yeah, And I think that's also why it's so important
20:09
Speaker 2: to have the coach in the loop, because they understand
20:13
Speaker 2: the human being even beyond just like the feel versus
20:15
Speaker 2: real aspect of like did the athlete not get a
20:19
Speaker 2: good night's sleep last night, then maybe today is not
20:21
Speaker 2: a good day to give them feedback. Are they going
20:24
Speaker 2: through challenges, you know, with a significant other, are they
20:27
Speaker 2: going through challenges in other ways? Or today are they
20:30
Speaker 2: really fired up? So today is a really good day
20:31
Speaker 2: to give them feedback. It's like understanding the human being.
20:34
Speaker 2: I'll give him a shout out. Connor McGuinness is the
20:38
Speaker 2: assistant pitching coach right now for the Dodgers, and he
20:41
Speaker 2: was so good at this. He would always talk about
20:45
Speaker 2: he never felt comfortable truly coaching a player until the
20:49
Speaker 2: player trusted him, until he felt like he had a
20:52
Speaker 2: good enough relationship with the player with the player would
20:54
Speaker 2: feel comfortable taking his feedback.
20:57
Speaker 1: Like start with the player as a human beings, Start
20:59
Speaker 1: with the player and get to get to.
21:02
Speaker 2: And this is why I think it's going to be
21:04
Speaker 2: really really difficult to have a product that goes direct
21:10
Speaker 2: to a player. We've dabbled with that, you know, and
21:13
Speaker 2: lots of companies have dabbled with film yourself with your
21:15
Speaker 2: iPhone and you get some feedback, you know, on your movement.
21:20
Speaker 2: But I think there's so much subjective stuff that goes
21:23
Speaker 2: into what we're talking about, how does the athlete move
21:26
Speaker 2: and how do they feel like they're moving, that I
21:29
Speaker 2: think it's going to be really hard to solve the
21:31
Speaker 2: challenge of giving an athlete feedback without a coach.
21:36
Speaker 1: In a minute, Jimmy's work in the NBA and why
21:40
Speaker 1: figuring out how to help NBA players shoot better is
21:43
Speaker 1: actually a really hard problem. What are you doing in basketball?
21:55
Speaker 2: Basketball is very very cool because it's a slightly different
21:59
Speaker 2: channe In baseball, like we've been talking about, the challenge
22:05
Speaker 2: for a pitcher is mostly just maximizing efficiency, or even
22:10
Speaker 2: for a hitter, you know, they're reacting to the pitcher,
22:13
Speaker 2: but they're still trying to swing as hard as they
22:15
Speaker 2: can and hit the ball as far as they can.
22:18
Speaker 2: So basketball is very different because you got a target.
22:22
Speaker 2: You're not trying to do this thing as hard as
22:24
Speaker 2: humanly possible.
22:25
Speaker 1: The target is the hoop. The targets to be clear, the.
22:28
Speaker 3: Target is the hoop.
22:29
Speaker 2: So that's a really interesting kind of like motor control
22:34
Speaker 2: problem of no matter where you are on the court,
22:38
Speaker 2: how good are you at getting the ball to do
22:41
Speaker 2: what you want it to do. We've been finding that
22:44
Speaker 2: there seems to be lots of trade offs that good
22:49
Speaker 2: shooters are making regarding like when do they release the
22:53
Speaker 2: ball in the course of their jump, how high of
22:55
Speaker 2: an arc do they use when they release the ball,
22:59
Speaker 2: how high they have their release point when they.
23:01
Speaker 3: Release the ball.
23:02
Speaker 2: So there's all sorts of trade offs that these shooters
23:05
Speaker 2: seem to be making related to how good they are
23:09
Speaker 2: at controlling their own jump, controlling their own velocity, where
23:13
Speaker 2: they are on the court, how close the nearest defender is.
23:16
Speaker 2: It's it's a very different problem.
23:18
Speaker 1: Sounds way harder, sounds way way harder, is that right
23:22
Speaker 1: for you? Harder for you to sort of solve to
23:25
Speaker 1: write a report that says, do this different thing and
23:27
Speaker 1: you'll hit a higher percentage of your shots.
23:29
Speaker 2: Right exactly? Yeah, And which makes it which makes it fun.
23:33
Speaker 2: I love solving hard podcasts.
23:35
Speaker 1: What have you solved in basketball so far?
23:38
Speaker 2: The big the first challenge that we had to solve
23:40
Speaker 2: was the scale challenge in basketball. There isn't quite as
23:44
Speaker 2: much there aren't quite as many games in basketball, but
23:49
Speaker 2: the data is a lot harder to parse because the
23:53
Speaker 2: events aren't as distinct. It's not like it's not like
23:57
Speaker 2: this is a shot, I mean only for free throws.
23:59
Speaker 1: It's a more continuous.
24:00
Speaker 2: Continuous motion. So the first big challenge was how do
24:03
Speaker 2: we even just isolate the events that we care about.
24:06
Speaker 1: Like what is time? When does a shot begin? When
24:09
Speaker 1: does a time equals zero for shot?
24:11
Speaker 2: But kind of debatable, right exactly. So that, yeah, that
24:15
Speaker 2: was the first big challenge that we had to solve.
24:17
Speaker 2: Was just sort of like the data engineering challenge. And now,
24:21
Speaker 2: like I said, our reports are more than telling you,
24:26
Speaker 2: you know, how to be more or less efficient. It's
24:29
Speaker 2: trying to surface the trade offs that you're making. Where
24:32
Speaker 2: in your jump are you are you releasing the ball,
24:34
Speaker 2: how high is your release point? What kind of arc
24:37
Speaker 2: are you using? And how does that compare to other
24:40
Speaker 2: arcs you could be using?
24:41
Speaker 1: What like what what are the trade offs?
24:44
Speaker 2: A really really interesting trade off to me is how
24:50
Speaker 2: much arc are you putting on the basketball? Not just
24:53
Speaker 2: to like evade a defender. But there's a trade off
24:57
Speaker 2: where if you shoot the ball at a highhigher arc,
25:01
Speaker 2: you have to use more velocity to get the ball
25:04
Speaker 2: to go all the way to the rim.
25:07
Speaker 1: Right, because it's going to travel farther in total in space.
25:10
Speaker 2: Right, And if you are not the most coordinated human,
25:16
Speaker 2: it might be harder for you to add more velocity
25:19
Speaker 2: in a really in a really precise way. So the
25:22
Speaker 2: more arc you have, the more prone you can be
25:25
Speaker 2: to what we would call like velocity errors overshooting undershooting.
25:29
Speaker 2: The advantage you get when you create more arc is
25:33
Speaker 2: if you if you imagine the ball coming down from
25:35
Speaker 2: that arc and the angle the angle with which it
25:38
Speaker 2: approaches the rim, it approaches at a steeper angle, which
25:41
Speaker 2: means you literally have more rim to aim at.
25:44
Speaker 3: So if you're hot, so if you shoot.
25:46
Speaker 2: Higher, you know this is a steph Curry thing. He
25:49
Speaker 2: has a really high arc. Presumably this is the hypothesis,
25:52
Speaker 2: because he's one of the most coordinated humans on the.
25:55
Speaker 1: Planet, so he sea.
25:59
Speaker 2: So he can use a really high arc because he's
26:01
Speaker 2: really good at controlling his velocity output. Oh huh, so
26:05
Speaker 2: he can really dial in how hard he releases the ball,
26:07
Speaker 2: which means he gets the advantage of having more rimmed
26:10
Speaker 2: to aim at, where as somebody who's bad at controlling
26:13
Speaker 2: their velocity output when they try to aim higher, they'll
26:16
Speaker 2: just overshoot an undershoot.
26:17
Speaker 1: So is the notion, and this is something of an oversimplification,
26:21
Speaker 1: but that for any given level of coordination, there is
26:26
Speaker 1: some optimal arc, and the more coordinated you are, the
26:29
Speaker 1: higher the optimal arc would be for you setting aside defense.
26:33
Speaker 3: Maybe that's the hypothesis.
26:35
Speaker 1: Yeah, I mean that seems like where what you were
26:38
Speaker 1: saying goes exactly.
26:40
Speaker 2: Yeah, that's the hypothesis. So that's what we're trying to
26:42
Speaker 2: look into.
26:43
Speaker 1: Okay, I'll be curious to see what you figure out.
26:47
Speaker 1: Are free throws easier? Did you think of starting with
26:50
Speaker 1: free throws?
26:51
Speaker 3: Yeah?
26:52
Speaker 2: Right, right now. Honestly, the phase we're at in basketball
26:57
Speaker 2: is mostly just collecting a lot of data. So you know,
27:01
Speaker 2: a starting pitcher will you know, throw ninety pitches in
27:04
Speaker 2: a game, and now you have a sample size of
27:05
Speaker 2: ninety pitches that are all the same, whereas a basketball
27:08
Speaker 2: shooter only has a couple of free throws in a game.
27:12
Speaker 2: You know, we're working on ways with teams of collecting
27:15
Speaker 2: data in a practice setting, so getting a lot of
27:18
Speaker 2: this data in a bigger chunks. But really at this
27:22
Speaker 2: point it's collecting a lot of data so we can
27:24
Speaker 2: do some of this research to explore some of these hypotheses.
27:27
Speaker 1: Huh. So it seems like in basketball you're where you
27:31
Speaker 1: were ten years ago or something in.
27:33
Speaker 3: Baseball, right exactly.
27:35
Speaker 1: So basketball is sort of one kind of frontier. It
27:40
Speaker 1: seems like one thing you're trying to figure out and
27:42
Speaker 1: haven't really cracked yet. What are some of the other
27:45
Speaker 1: frontiers the other things you're figuring out, whether it's in
27:48
Speaker 1: baseball or in the fundamental technology or whatever. What are
27:51
Speaker 1: you working on?
27:54
Speaker 2: We want to try to have a computer vision be
27:59
Speaker 2: even more accessible. You know, there's been a lot and
28:04
Speaker 2: a lot, a lot a lot of improvements over the
28:07
Speaker 2: last ten years in computer vision where you can do
28:11
Speaker 2: really good motion capture with just your iPhone, but still
28:17
Speaker 2: for certain specialized movements, pitching, shooting, things like that, there's
28:25
Speaker 2: still have little ways to go to get really really
28:27
Speaker 2: good data straight from your iPhone. So that's one of
28:30
Speaker 2: the frontiers, is like continuing to try to help understand
28:35
Speaker 2: how to make computer vision more accessible. Another one is
28:40
Speaker 2: more fitness based analysis. You know, it's interesting to think
28:46
Speaker 2: about companies like Mirror and companies that have tried to
28:50
Speaker 2: give people feedback and like a fitness environment, But how
28:55
Speaker 2: do you give someone or how do you give like
28:58
Speaker 2: a strength and conditioning coach good feedback that can be
29:00
Speaker 2: used in a weight room setting. Yeah, and then continue
29:03
Speaker 2: to explore other emotions and other sports, Like football is
29:07
Speaker 2: a really interesting one because a lot of football is
29:12
Speaker 2: interacting with other humans. Indeed, yeah, I mean that's obvious,
29:17
Speaker 2: but like, how do you get a takeaway from like
29:21
Speaker 2: two linemen interacting?
29:24
Speaker 1: Ye, things, but that one I wonder. I mean, it
29:26
Speaker 1: seems like if you think of the line, you know,
29:29
Speaker 1: in football, it's so they're so on top of each other,
29:33
Speaker 1: the defensive line, of the defensive line, that I feel
29:35
Speaker 1: like vision might not be what you want. You might
29:37
Speaker 1: want sensors, right, you might want pressure sensors in the
29:40
Speaker 1: lineman's clothes or something. I don't know, I'm just making
29:43
Speaker 1: that up, but like it seems like that might be
29:45
Speaker 1: more useful, just because it's hard to see what's going
29:47
Speaker 1: on in the line.
29:47
Speaker 3: Yeah.
29:48
Speaker 2: Yeah, yeah, except professional athletes don't like wearing random.
29:52
Speaker 1: Things, well are aren't people trying to make sensors like
29:57
Speaker 1: woven into the clothes. I mean, I feel like there
29:59
Speaker 1: are ways you would just be putting on your jersey
30:02
Speaker 1: or putting on your paths or whatever and they would
30:03
Speaker 1: have the sensors built in.
30:05
Speaker 3: Yeah. Yeah, one hundred percent.
30:06
Speaker 2: There's lots of really cool technology of that is you know,
30:10
Speaker 2: you know, microscopic sensors that are just woven, woven into
30:13
Speaker 2: clothing for sure.
30:16
Speaker 1: So so let's talk for a second more about the
30:20
Speaker 1: consumer side. You sort of touched on it and moved on.
30:23
Speaker 1: I mean, is that are you actively working on that
30:26
Speaker 1: or is that just like yeah, it kind of seems interesting,
30:28
Speaker 1: but not for now. We're too busy, like what's what's
30:30
Speaker 1: happening on the consumer side.
30:33
Speaker 2: We're not actively tackling in the consumer side right now.
30:37
Speaker 2: And that the reason why we started with professional sports
30:40
Speaker 2: is one it's because like it's what I know, you know,
30:42
Speaker 2: I work for the Dodgers, but also because the value
30:47
Speaker 2: proposition of what we're doing is so impactful. You know, yeah,
30:53
Speaker 2: you know, you understand, you try to understand, like the
30:54
Speaker 2: relationship between people have tried to put numbers on this,
30:58
Speaker 2: it's the people have estimated, it's in the millions of dollars.
31:01
Speaker 2: But the value of adding one mile an hour of
31:04
Speaker 2: fastball velocity to a picture, you know, people have valued that,
31:08
Speaker 2: We have valued that in millions of dollars.
31:10
Speaker 1: Well, sure you will tell you millions of dollars, But
31:13
Speaker 1: what is the like, I don't know what's a what's
31:16
Speaker 1: a what's a top picture make? These days? I don't
31:18
Speaker 1: even know anymore.
31:20
Speaker 3: I mean, you want to talk about show.
31:23
Speaker 1: Half a billion, hundreds of millions, half a billion? Yeah,
31:28
Speaker 1: so so right, so so the marginal benefit has a
31:32
Speaker 1: very large value. Yah a million dollars against one hundred
31:35
Speaker 1: million dollars, like one percent better. If they're making a
31:37
Speaker 1: hundred million dollars, it's worth a million dollars presumably.
31:40
Speaker 2: Yeah, And we start to get the pro sports teams
31:43
Speaker 2: to believe.
31:43
Speaker 1: That, well, how many how many pro sports teams are
31:48
Speaker 1: paying you at this point, about uh.
31:51
Speaker 2: Close to ten in Major League Baseball and a couple
31:54
Speaker 2: in the NBA.
31:55
Speaker 3: The NBA is a lot newer.
31:57
Speaker 1: And how big is the field? Like, what's sort of
31:59
Speaker 1: the state of play in the field of I don't
32:03
Speaker 1: even know what to say. Biomechanics as a service seems
32:06
Speaker 1: like a niche construction of it. But what would you
32:09
Speaker 1: say sports analytics? I mean, I guess that's not quite right.
32:12
Speaker 1: What how do you construct the broader field?
32:15
Speaker 2: Sports analytics is definitely a part of it. The data
32:18
Speaker 2: that we provide is novel or is different than like
32:25
Speaker 2: traditional sports analytics. So we don't really have a lot
32:28
Speaker 2: of companies as competitors. Honestly, our biggest competitors are teams
32:38
Speaker 2: wanting to try to do this type of thing internally.
32:40
Speaker 2: How you know, hire a bunch of data engineers, hire
32:42
Speaker 2: a bunch of software engineers, hire people with biomechanics backgrounds,
32:46
Speaker 2: and try to build out these pipeline processing pipelines themselves.
32:50
Speaker 1: Does every team have like a somebody with a PhD
32:53
Speaker 1: in biomechanics working for them now?
32:55
Speaker 3: In baseball? Yeah?
32:57
Speaker 1: Wow?
32:58
Speaker 2: In basketball and not Yeah, but they're starting to.
33:02
Speaker 1: So I know people, I mean, I think in general
33:06
Speaker 1: baseball fans like to complain, but so one of the
33:08
Speaker 1: things they've complained about lately is the way analytics more
33:11
Speaker 1: broadly made the game more boring, right, like the shift
33:16
Speaker 1: and changing pictures more frequently and whatever else people complain about.
33:21
Speaker 1: Are you do you fit into that at all? Uh?
33:28
Speaker 3: That's a good that's a good question. Some people.
33:35
Speaker 2: Some people complain about how hard pictures are throwing these
33:38
Speaker 2: days because it creates more strikeouts, and people think strikeouts
33:43
Speaker 2: are boring.
33:45
Speaker 3: So maybe, yeah, so.
33:47
Speaker 1: Maybe you need to get better at helping hitters. Really
33:50
Speaker 1: you're out that way, you can even at back up, all.
33:53
Speaker 2: Right, And that's what we talked about takeaways four Hitters
33:55
Speaker 2: are harder because they're reacting to the picture.
33:59
Speaker 1: So if you think about the field, what your company
34:02
Speaker 1: say in five years, yeah, whatever is your kind of
34:05
Speaker 1: medium term future that you think about, what is the
34:09
Speaker 1: company in the sort of the world, the sports world
34:12
Speaker 1: that you're interacting with look like at that time and
34:15
Speaker 1: say whatever, five years, ten years.
34:18
Speaker 2: What I hope is that what we are trying to
34:23
Speaker 2: foster is an environment where coaches have really incredible tools
34:32
Speaker 2: at their disposal to understand how an athlete moves. So,
34:35
Speaker 2: you know, we talk about the very beginning where sort
34:38
Speaker 2: of the still more or less the state of the
34:40
Speaker 2: art is a coach just looks at an athlete, watches
34:43
Speaker 2: video of an athlete, and tries to give the athlete
34:46
Speaker 2: feedback regarding what they see on the video, what they
34:48
Speaker 2: see with their eyes. We hope that the standard becomes
34:54
Speaker 2: you use an analytical tool to help you understand how
34:59
Speaker 2: the athlete is moving and to really like level up
35:02
Speaker 2: your coaching because now you have this objective information about
35:06
Speaker 2: how the athlete is moving. I kind of I kind
35:09
Speaker 2: of make the analogy related to just radar guns. Before
35:14
Speaker 2: radar guns were a thing, a coach would just like
35:17
Speaker 2: look at a picture and be like, oh, that looks
35:18
Speaker 2: pretty fast, you know, make an adjustment, and I think
35:22
Speaker 2: that looks a little faster. But then radar guns came out,
35:27
Speaker 2: and you could actually measure how fast the picture was throwing,
35:30
Speaker 2: and you could actually measure if the picture is getting
35:33
Speaker 2: throwing the ball harder based on your feedback.
35:35
Speaker 1: Yeah, and now you're you're doing that but in a
35:40
Speaker 1: in a way more complex way.
35:41
Speaker 3: Exactly.
35:42
Speaker 1: Yeah, we'll be back in a minute with the lightning round.
36:01
Speaker 1: Let's do the lightning rounds. Let's start with a few
36:03
Speaker 1: baseball questions. Who's the most underrated picture of all time?
36:08
Speaker 2: Whoa underrated picture of all time? Oh my goodness, I
36:16
Speaker 2: don't know if I can give you one that's like
36:20
Speaker 2: all time, because I don't know if that's fair. I'm
36:22
Speaker 2: sure I'm not thinking of everybody of.
36:24
Speaker 1: Your lifetime, of your lifetime.
36:26
Speaker 3: I grew up a hardcore Red Sox fan.
36:29
Speaker 2: I grew up in Rhode Island, and the first one
36:33
Speaker 2: that comes to mind mostly in the Red Sox atmosphere,
36:37
Speaker 2: but I wonder if you could make a broader argument
36:39
Speaker 2: was Tim Wakefield, who was a knuckleball pitcher for the
36:43
Speaker 2: Red Sox and what made amazing physics amazing physics and aerodynamics,
36:49
Speaker 2: And the reason being is he just did so many
36:53
Speaker 2: things for the Red Sox, ate up so many innings
36:56
Speaker 2: and was so effective closing, starting, whatever, But he never
37:01
Speaker 2: got incredible recognition because it was a knuckleball that was going,
37:05
Speaker 2: you know, fifty five sixty miles an hour.
37:07
Speaker 1: What's one thing you would change about baseball to make
37:09
Speaker 1: it more popular?
37:13
Speaker 2: I do feel like actually the changes that are being
37:17
Speaker 2: made are good ones, in like reducing the amount of downtime.
37:23
Speaker 1: A pitch clock, it's one, is that the way you're
37:25
Speaker 1: thinking of?
37:26
Speaker 2: Or yeah, yeah, yeah, exactly, yeah, a pitch clock. The
37:29
Speaker 2: thing that's challenging for me as a biomechanist is when
37:32
Speaker 2: you reduced the amount of time that a pitcher has
37:35
Speaker 2: to throw, you theoretically could introduce more fatigue, which theoretically
37:39
Speaker 2: also introduces more injury risk. So this is something that
37:42
Speaker 2: we've been thinking about. Is like, So for me, while
37:46
Speaker 2: I like that change that baseball is making to make
37:50
Speaker 2: the game, to speed the game up, I think the
37:53
Speaker 2: pitchers also need to train a little bit differently. Oh,
37:57
Speaker 2: that's to be able to better withstand the shorter rest time.
38:03
Speaker 1: Are there aspects of your work and the changes you've
38:07
Speaker 1: seen over the course of your career that illuminate sort
38:11
Speaker 1: of broader changes in computer vision and AI more generally.
38:16
Speaker 2: Yeah, In particular, the most important one has been the
38:19
Speaker 2: improvement in computer vision, which because computer vision at its
38:23
Speaker 2: core is artificial and intelligence neural networks, and as those
38:28
Speaker 2: as that that technology has gotten better and better, you know,
38:31
Speaker 2: the latest and greatest people always talk about like the
38:33
Speaker 2: Transformer model really changed AI. I mean that changed computer
38:38
Speaker 2: vision too, Like lots of the modern more modern like
38:41
Speaker 2: computer vision models are based on transformers.
38:44
Speaker 1: And which, to be clear to the Transformer model is
38:47
Speaker 1: what gives us chat GPT, right, The T in GPT
38:50
Speaker 1: is transformed, right, So how has it affected the computer vision.
38:53
Speaker 2: Side making the models more accurate and more efficient? It
38:58
Speaker 2: used to be you know, a couple of years ago
38:59
Speaker 2: when I try to run a computer vision model on
39:02
Speaker 2: my laptop, like it could take an hour just to
39:05
Speaker 2: like analyze one pitch, one video, and now it takes
39:09
Speaker 2: a matter of seconds.
39:12
Speaker 1: And so what is the what is the bigger implication
39:16
Speaker 1: of that? Beyond your beyond your work?
39:19
Speaker 2: More and more data is available regarding how people move,
39:25
Speaker 2: you know, like when when I first started, it was
39:27
Speaker 2: really hard to have data in a baseball game. Now
39:31
Speaker 2: every major league game, every minor league game, every NBA game,
39:35
Speaker 2: maybe every G League game, every w NBA game, you know,
39:39
Speaker 2: every single uh basketball and baseball game, you know, more
39:44
Speaker 2: or less is now like being recorded with computer vision
39:48
Speaker 2: to get the three dimensional data about how people are moving.
39:51
Speaker 2: Just lots and lots of data on how people move.
39:54
Speaker 2: And this is really impacting. I think lots of fields.
39:58
Speaker 2: In particular, I think like self driving cars, robots that
40:02
Speaker 2: are meant to like interact with the world, all rely
40:05
Speaker 2: on computer vision models. I mean, I think one of
40:08
Speaker 2: the coolest things about how cars do this sort of
40:11
Speaker 2: thing is not only do they have to understand where
40:15
Speaker 2: a person is now, but they're really cool models that
40:21
Speaker 2: they take where a person is now and the last
40:25
Speaker 2: like ten seconds of what that person did and try
40:28
Speaker 2: to predict all the different things that the person might do.
40:30
Speaker 2: They might run across the street, they might jump out
40:34
Speaker 2: of the way, they might jump forward, you know, they
40:36
Speaker 2: might run and chase a soccer ball.
40:38
Speaker 1: I mean, my sense is all of the hard edge
40:41
Speaker 1: cases in self driving cars. Basically, the reason we don't
40:44
Speaker 1: truly have self driving cars yet is because people are
40:48
Speaker 1: so hard to understate. Yeah, right, Like if the world
40:50
Speaker 1: was all self driving cars, then it would be a
40:53
Speaker 1: solved problem, right, Like the machine could understand what other
40:57
Speaker 1: machines are going to do. But people, human drivers, human
41:00
Speaker 1: pedestrians are strange and very hard for machines to understand.
41:06
Speaker 3: Yeah.
41:07
Speaker 1: Yeah, it's like they need a coach. It's like the
41:09
Speaker 1: coaching problem again. We found your SoundCloud. Oh yeah, so,
41:18
Speaker 1: Jimmy Buffett fans are called parrot heads. What are Jimmy
41:22
Speaker 1: Buffy fans called?
41:26
Speaker 3: I never thought about it.
41:31
Speaker 1: You have a song called Let's have some Fruit parentheses
41:36
Speaker 1: the fruit song is fruit a metaphor.
41:41
Speaker 3: Leave it up to your imagination. Fair.
41:45
Speaker 1: Jimmy Buffy is the co founder and CEO of Reboot
41:49
Speaker 1: Motion Swanson swe Today's show was produced by Gabriel Hunter Cheng.
42:06
Speaker 1: It was edited by Lyddy Jean Kott and engineered by
42:09
Speaker 1: Sarah Bruguer. You can email us at problem at Pushkin
42:13
Speaker 1: dot Fm. I'm Jacob Goldstein and we'll be back next
42:16
Speaker 1: week with another episode of What's Your Problems