DataTopics Unplugged

#34 Thoughts on Sports, Data & AI

January 29, 2024 DataTopics Episode 34
DataTopics Unplugged
#34 Thoughts on Sports, Data & AI
Show Notes Transcript Chapter Markers

Welcome to the cozy corner of the tech world where ones and zeros mingle with casual chit-chat. Datatopics Unplugged is your go-to spot for relaxed discussions around tech, news, data, and society.

In episode #34, "Thoughts on Sports, Data & AI", we lace up our virtual running shoes and dribble into the intersection of sports, technology, and data. This episode is not just a playground for tech enthusiasts but a stadium where data geeks and sports fans unite. So, grab your favorite sports drink, relax, and let's dive into the heart of data and sports, unplugged style!

  • AI-Driven Fitness Coaching: Exploring the latest in AI-based running coaches like Trenara and Ultra Trail Coaching.
  • The Science of Soccer: Delving into VAEP, a cutting-edge framework for valuing soccer player actions.

Intro music courtesy of fesliyanstudios.com.

Speaker 1:

Yes, let's hit it, I'll do it. Hi everyone, welcome to Data Topics Unplugged. It's a casual, light-hearted, cozy corner of the web where we have a weekly short discussion on what's new in data, from sports to browsers again, anything goes. Today is 26th of January of 2024. My name is Marillo. I'm your host for today. I'm joined by my partner in crime, bart Hi, and I also have. We have some really special guests today. We have Tim.

Speaker 3:

Hello there.

Speaker 1:

And we have Thibault, hello. So Tim is the one that keeps the show running, the Bizdev show running here at Data Roots, the King of Bizdev, and Thibault is one of our expert data strategists sitting at maybe are you taller than Sandy Brun. No.

Speaker 4:

I don't know 196.

Speaker 1:

Close enough, close enough, but I'm going to be tired. Yes, yes, we also live streaming. We're on YouTube, we're on LinkedIn, we're on Twitch, we are on X, so feel free to leave a question there and we can actually tackle them in the recording. Maybe, tim, would you like to introduce yourself for starters. Anything you'd like to share with everyone?

Speaker 3:

The last time I was on, you had something prepared, so I feel like I'm not the only one that's unprepared today. Wow, Waiting for me on the spot. Maybe first take some time to. I feel like we're all looking a bit funny that is true.

Speaker 2:

I think it's only really obvious for you, Tim and Marilla.

Speaker 3:

Yeah, you guys usually look a bit weird.

Speaker 1:

But maybe for the people that are just listening, we are all in a different outfit let's say, some of us Because today we have a topic. We have a sports analytics topic, so we thought it would be a fun idea to just dress up accordingly. So Tim is throwing all his puffiness on his arms his NBA shirt. But yeah, what's about? What's with the shirt?

Speaker 3:

Wait, it's actually one of the first basketball jerseys I've ever bought and, knowing that I played basketball for 20 years, this is saying something. It didn't grow up at all.

Speaker 1:

It's the same height.

Speaker 3:

Yeah, I'm standing as tall as I was when I was six years old. No, no, this is a very, very retro San Antonio Philadelphia 76ers shirt. So it's Al 9%. Who has been out of the league, the NBA, for, I think, about 13 years now. I think his DICE was 2011. So it's actually one of the t-shirts I'm more proud of.

Speaker 1:

You can see that Tim is a business guy because he has like really nice pants, but then he has like a NBA shirt, you know, showing off his puffiness in his arms, and then like a dress pants, you know to this is how I usually visit clients. Yes, I see.

Speaker 3:

Yeah. So it's to throw them off, like always, keep them on their toes, Never let them know you're next. So they first see this part and then they're like, oh what? And then he's he did. Oh yeah, that's a nice guy.

Speaker 2:

So Tim's head of our business development team, is also a very good basketball player. Right Long history in basketball.

Speaker 3:

A long history, especially I've been playing. It's yeah, it's my 20th year this year that I'm playing basketball. I used to play at the for me quite a high level, so second division always been interested in sort of the mental part, the intellectual part of basketball, which I think it's quite an intellectual game. I think everybody thinks this is of this sport, but I truly approached this. So sports, and hence the interest in sports analytics, how it sort of developed in order to kind of improve. As you might see, I'm not, I'm roughly, I think I'm roughly point 80, 60, post all. So I'm quite small for a basketball player. So you have to find other ways to get a competitive advantage.

Speaker 1:

Yeah, tim is also quite the football player as well. Yes, yeah.

Speaker 2:

He's like an engine. I was having a discussion with him yesterday. I used to call him nickname him Dr Duncan Stein. I found that offensive and I didn't know. It's a bit awkward because he's too short to dunk, his hands aren't big enough. That's what he told me yesterday.

Speaker 3:

I'm just saying that I feel like the basket is a bit higher for me to do. Okay, I see it. Yeah, the basket.

Speaker 1:

I kind of thought of that, because I'm about the same high to stim and, yeah, I pretty tall.

Speaker 2:

It would be more something for the ball. But what does the body do in sports?

Speaker 4:

I run, you run. I think that's an understanding.

Speaker 2:

I think the ball is someone that only runs when it's more than 50 kilometers.

Speaker 4:

Now, that's not true.

Speaker 3:

No, you run, you run less this morning. Yeah, this morning I didn't have enough time.

Speaker 2:

But you do extreme stuff.

Speaker 4:

Right, I like to run far because, opposed to basketball, it doesn't require any intellectual effort and just as mindless as possible.

Speaker 1:

And what's the shirt you got going on today?

Speaker 4:

It is Well, opposed to Tim, I don't fit into the soccer shirts I bought when I was a child, so I wasn't able to bring one you grew so a broad one of my running crew here in Leuven which every Tuesday we go make some altitudes, just hills, ups and downs, and here in Leuven you have quite some and we go well, and that we say but that just happened, which is like collect potatoes, and the potato is also one measurement unit which is about 30 centimeters, and I have no idea how they come up with this. But this is also a bit of a mystery behind it. Okay, Cool.

Speaker 1:

And what's the longest distance you've ever ran? Actually 100 miles, Miles.

Speaker 4:

And where did you do this? In the audience. So look at the dark Up down. That's what I trained for 100 miles.

Speaker 3:

When you, when you said collect potatoes, I thought you were going to do like one hill and just every time you go up you bring it down, you bring a potato and you put like stuff, a bad backpack, like every time, and at some point it just becomes impossible to move that would be a really good workout.

Speaker 4:

I think we have a new contest going on.

Speaker 3:

Yeah, yeah. So the backyard ultra is like the potato ultra man 100 miles, that's 160 kilometers.

Speaker 4:

Even though I'm tall, I'm also not able to dunk, and them and I even followed volleyball classes once in university and they jump higher than I do, so it's a bit of a for the thing is better to do basketball classes.

Speaker 1:

I was trying to dunk on the volleyball court. I couldn't find it.

Speaker 3:

I kept throwing it in the net. It was super weird.

Speaker 1:

I also. I have a Brazil shirt on. Going back to my roots, I also have a very stylish hat that I got some heat for already. But I'm not going to hide my true self. I also play tennis. I actually play tennis with you, bo, and actually it's a. You see how he's very like tall, he's very wide, you know like, so he can, but you can tell, you can tell he's not very explosive. So you see, and I think also for dunking, I think the, if you're explosive, you know, because like jumping vertically, you know it's from related to explosive legs, I guess. Right, so I can definitely see that. But I thought it was very interesting. You know, like the guy that runs 100 miles, but I think I'm more explosive than you are. I am, but to me it was a take that. What about you, bart? What kind of shirt you got going on, also running shirt.

Speaker 2:

It's a running shirt.

Speaker 1:

Yeah, but you had some issues with the shirt. No, as you grew older, you said the shirt tends to.

Speaker 2:

How should I explain this? It tends to shave sometimes.

Speaker 1:

Shave. What did you mean?

Speaker 2:

Create a bit of shave on sensitive areas. Can you be more concrete Somewhere around torso, torso Two separate areas that are similar.

Speaker 4:

Now we're going to see how good the moderation of YouTube works.

Speaker 1:

He's talking about his nipples, If anyone was. Yeah yeah, it's very coarse on the nipples, this shirt, but let's talk data Running aside, yeah, but talking about running, right, thibaut, you brought something AI based running coaches.

Speaker 4:

Yes, so on that topic, recently I got a contact by Bart who created a small app on his own running coach. He has been creating, which he, I think it has been a hobby project for Albert which you've been developing for, I think, months, years, I don't know.

Speaker 2:

No, not years. Well, I tried it a few times over the years, that is true.

Speaker 4:

But now the new equity, the upcoming of chat, gpt. You made some major advancements, so the question I want to bring to this panel if the classic running coach which develops a training plan for you is dead or not, with those AI based running coaches which and I see here you have a train Nara running app, that's one. Yeah, so that's one. It's a Belgium which is feel in Belgium is quite popular. So it just a runner who teamed up with a developer and they developed an app which creates running schedule for you. But what I really like to those apps compared to the classic training schedules which I have from our training coaches, when I once bought one five years ago and was just like a 12 week schedule and then if you miss one session you think, damn it, the complete schedule is gone now. And the big advantage to these apps, I think, is that whenever you miss a session, it just adapts its schedule for you. And I think that's also the approach you took part in. Like you give her textual feedback right, because what I find very hard with those apps is like, which metric are you going to optimize and what do you give us input parameters?

Speaker 2:

Yeah, so maybe explain a bit what it is. So I made a very, very, very simple app that basically wraps open AI's API chat GPT that everybody knows to see if it could facilitate me in building a workout plan for the coming week weeks based on what I did in the past, based on my current status, which is typically something like you get from Garmin, you get from Trinara, you get from a lot of places, right and what I did is that I asked JetGPT basically to provide me with a workout for the coming day, for today actually, basically for today based on what I did in the past and based on some biometric data. I have an ororing. Where I have, an ororing gives you a readiness score, gives you a sleep score. So it also has this context. It also knows if I'm working towards something like a race, a running race or a cycling race. It knows a bit how far in the future are these key points coming, what did I did in the past and also how do I like my typical week to look like. For example, friday is a relative rest day these type of things it keeps into account. I also typically tell that every week, I want to do at least one swift race, which is a cycling race. I give it some input and, based on that, it generates a workout for today, tomorrow, it will generate one for tomorrow and further, which actually worked pretty well. I think we've been using it now since November, I want to say something like that.

Speaker 4:

But you put it at the test at Metalboot drill right.

Speaker 2:

Yeah, I built it and then I actually wanted to do this to work towards a run in December somewhere. The workout went very well up to there. Then I got sick. The result was still okay. It's hard to evaluate. What I noticed is normally I don't have a plan. I just work out a lot, but I don't have a plan. That means that you, in hindsight, compared to now, you do much more of the same than I do now, because now when I do an interval training, it's every time a different type of interval training, sometimes 500 meters and 600 meters. Then it's a fact lack. Actually, you have a much more variation. Well, if I don't have someone that says to me do this, then it's okay, I'm going to do 800 again. You get a bit into this same pattern.

Speaker 1:

Also, do you feel like the personalized aspect of it, because you have a readiness score and a sleep score from the o-ring? Do you think that that's a because you mentioned diversity, but also the fact that he's very personalized on how he's feeling today? Is this also a big?

Speaker 2:

question. It helps me especially to do less. Okay. So when I get I saw the score before as well I had the o-ring already Even before I had the o-ring. I know when I wake up I've had a bad night or I did a lot yesterday, but I tend to do more as an athlete. I tend to overdo my workouts and then underdo, and it helps that you have someone in this case my coach, gpt this says oh no, your readiness is bad. You did a lot of hard work yesterday. Today is the rest day and if you want to go out then it's max 20 minutes at this space.

Speaker 3:

But to combine, like your original question was like is this the end of the coaching plans?

Speaker 2:

And I'll get. The thing is, what I built is extremely simple. It's extremely minimal, like it's a prompt, even for you. It's a black box now, right? Well, yeah, you can argue on what is the effectiveness. That it's hard to evaluate for me at this point.

Speaker 1:

Did you use these apps as well before.

Speaker 2:

In the past I've tried it a few times. But what is today? Only to some extent. To some extent it is in my prompt. But I think what you can very much enrich this prompt, that to say I want to use this type of workout, I want to make sure that it's a faced approach that gives so many weeks of build up and then a rest week. So in theory, you can add all this domain knowledge to the prompt as well. Yeah, definitely.

Speaker 4:

But that's where I think, like those classical coaches, like every amateur running coach, they charge you 30, 40 bucks for a 12 week schedule, which I think is ridiculous, because they just give the same schedule to everybody and just adjust the basis, which I think this is much more valuable. But what you just said that as an athlete I tend to over train I think having this psychological effect of a trainer who actually checks you if you stick to your schedule, that's where they're still valuable. But, creating the schedule.

Speaker 2:

I think that's a dead end and it's also, I'd argue, that it depends a bit on like, if you have a trainer that has hundreds of athletes and is hands out of schedule, that's not very personal, but this is actually very personal. So I had an example where I had a very long insurance run and I had a. I didn't eat too much. I ate a bit of you say this in English I had a bit of hypoglycemic and I didn't eat before and it didn't bring anything. So I after the workout so I also post workout some very lightweight description I set next time in the description. Next time, make sure to remind me that I eat something before and that it brings something. And because it has access to the whole context of everything that I did in the past, every time it gives me a long insurance run. Now it gives me also this input make sure to fuel. It takes something with you. Well, yeah, if you would have created this app even if you would have created this app before, it would have taken a lot of development to give these things, while this is super minimal to set up for this.

Speaker 4:

And does it also tell you to bring some food on your recovery runs and short runs, because that's when I would know only on the insurance runs. Okay, and so it explicitly doesn't say so. On your short runs.

Speaker 2:

No, it just takes a little all the time because you enter once but I also made it explicit so next time I do a long endurance run, make sure to maybe a question then, because I feel like you also you understand how, like, prompt engineering works and all that.

Speaker 1:

But do you think if it was just anyone, do you think it would be as effective? Because maybe the person would be like, oh, just remind me to bring food next time, but then for them it's like they.

Speaker 2:

Yeah, that's a fair point, but I think the I think the question here is not can everybody build this? But I think the question is the new era of workout training apps will they use this as a basis or more traditional ones? And I think the new era of training apps will have to to embrace this technology. Yeah, it brings a lot, I think.

Speaker 1:

Maybe. So you also mentioned like that. A lot of the times the coach gives the same training to everyone. But this is an app. So I guess the app is just a database of different workouts and, based on what you click, they just almost give you the same papers. That's that kind of like. Because I guess, if it's really that standard, how come can you build a whole app around it?

Speaker 4:

you know, like that Should, I'm saying that's, but I think you have some basic training. Consumptions, yeah, and there are apps.

Speaker 2:

I think the Trenard one is an example like. It's also precise some extent that you can also give input and it will correct. And if you were able to do it one day, it will push some stuff to later.

Speaker 1:

So definitely there's like a follow up aspect to it.

Speaker 2:

Yeah, and I don't know how they like. There are a lot of these apps, right, but today I would have been developed typically before. The LLM era is very deterministic and like all these things, all these exceptions that happen, you need to think about, you need to program, you need to put out rules to it and like what is the impact on the training schedule? Like, and it takes a lot of time versus like what I just made is in natural language, is super easy.

Speaker 1:

I was also thinking about your approach as well is that you have very specialized data, and I think it's not today we like Apple health and all these things you know is like you can have a bit of a more personalized follow up, right? And I think for me, even if I kind of relate to what you're saying I don't know if that's the reason why you tend to overwork, but sometimes it's like if there is some health metric that is saying, oh, you should probably rest and my first thinking is like well, because I don't want to get injured, because my body is showing signs that I'm getting to a limit somehow, you know, like I didn't recover well or something. So I feel like the fear of injuring yourself, because the thing is, I don't want to not push myself because I'm cool. I'm cool being lazy, or because I don't feel like it or because of this. But I think when you say, well, no, I don't think you should put yourself because you may injure yourself, I think that's a more compelling argument.

Speaker 4:

But that brings us to the transparency of such algorithms. Right, like why do they recommend certain workouts, but do you know?

Speaker 2:

but do you know that on these apps, for example, no, what you want in these apps is that there is a scientific team behind this that has implemented it. That's what you assume. Yeah Right, that there is a.

Speaker 3:

To put a question on the thing that you developed. Is it sort of like a like? Do you have a rag based thing where it really tries to match, or is it simply prompt answer?

Speaker 2:

And I have the database of everything that I did in the past.

Speaker 3:

You have a line Just inject it in the prompt. It is inject like the full database is injected in a prompt.

Speaker 2:

Yeah, that's why I need to use the GPT-V4 Turbo, right With the very big context, the most expensive one. Okay, I had to switch after a few weeks because it didn't fit in the context anymore.

Speaker 3:

This year I just found another company.

Speaker 2:

But we're still talking a few euros a month to use this. Any app would have been more expensive.

Speaker 3:

Could you improve on what you developed simply by fine-tuning or Just a little bit of fine-tuning? The model with very good training programs and very good knowledge?

Speaker 2:

Yeah, I think that was an excellent?

Speaker 3:

Yeah, because what I'd be concerned about is that. So my the idea and feel free to challenge me on this, but like JetGPT and all the LLMs, they're trained on like a huge amounts of data and I think there's. If you go on the internet right now and you look up training programs, you'll find just every 100,000 websites that have training programs and knowledge on, and I don't think that everything is scientifically based. From what I know of like personally, you know of coaches, not everything is scientifically based and you can do a lot of stuff wrong with training. Right, you have to. There's this principle of like you train. There's a you feel worse than you feel better and if you start training too soon, then you create this effect of gradually getting worse With the current thing is your app. The value there is mostly on having the structure, having something that structures you and that challenges you and gives you variety and these kind of things, but it's not on the, on a scientific like. We know that this is what it's going to do with your body. This is.

Speaker 1:

Yeah. Yeah, that's the thing I think. Sometimes there's some variability between people, right yeah, but I'm not trying to do with my own.

Speaker 2:

It's purely a very small experiment that I'm doing personally. What I'm not trying to do is to build something that is the best solution for the endurance run, right, I'm just having something, a tool, a utility to help me. That's what I'm building and it's super. It was super easy to build. And maybe also what I think where there is a very big measure of value in LLMS is that the optimization part around workouts is something that is hard to grasp, because when you talk about optimization, what do I do best next? Very much depends on. You need to think about, like, what are my inputs? And inputs are hard in sports. Like I can say I did a work, I did an interval, for example. I can say that interval law, that I did 600 meters, did it five times and the rest of the six minutes, and then I did it again, and then I did some, did two hill repeats to close, and I can write this out, and so you have this description, but you end up in the end with numbers. So you have what is the duration, what was your heart rate? You have maybe your perceived exertion rates, which is yourself like, how hard was it? You have a number of metrics and you go from this description. This description is already very hard to put into something that is understandable by computer. Just the intervals, these data points like the duration and heart rate. That's very easy to understand but it is very hard for a computer to understand what was the effect of that. If you didn't determine this equation, because your perceived exertion rate is like yeah, okay, it's an eight, what does that mean? It was very hard. But from the moment that you say if in your post-workout description, if you say something yeah, it was hard, but I felt fine, there was a bit of a snow on the ground, so that's why I didn't reach my speed set that I had targeted for. So it's actually a training is normal, that's what it should have been and every coach, every person that reads that, understands what it means. But for a computer to understand that is super hard. And with LLMs you come much closer to understanding this natural language input which, in my eyes, you do need to understand how is a person feeling when they do a workout? And that to me is the biggest factor to make something that is hard to capture in very deterministic, very specific metrics. Make it understandable by computer, and I think that is where LLMs will shine in this regard.

Speaker 3:

I think that the sport of running, and running fast it also lends itself a bit to that. I've been thinking for quite some time, ever since you first said, about the AI coach that you have, and I was like, oh, that's super cool. I want to think about how I do this for basketball. I have no clue how I put a practice into words at all. It's so visual. I think I would need the visual modality to be completely fleshed out, which it isn't right now to accurately describe. I could tape my basketball practice and try to explain what is happening and things that could be improved, but aren't they?

Speaker 4:

doing this in soccer with those GPX trackers, which is harder indoor, I think, with basketball, but I know in the US you have a few fancy stadiums where they do have GPX trackers inside their location trackers.

Speaker 3:

Yeah, in basketball as well. I know, for example, in the NBA, all the games that you have I don't know what hundreds of cameras on a single basketball court. They just have tracking stats. You can track how much time of possession somebody has. Oh, he has 10%, 20%, 30% of the time this person has the ball in their hands. If this person in an offense, this person has the ball in their hands, they score on average 1.8 points per possession. These type of things they can track. That's crazy, but it only works on an NBA level. I'm not even sure that they have this in the EuroLeague. The EuroLeague is the second best one. I'm not even sure that they have it there because it's super costly.

Speaker 1:

Yeah, but I think maybe there's some types of sports like maybe running it lends more to natural language, maybe basketball lends more to this. I also, for example, comparing basketball to football or soccer, depending where you are. I think it's like these kind of stats, it's not as meaningful in football no, because I have the, for example, american football as well. I think it's a more, in a way, it's a more structured sport in the sense that stops more. You have more shot, like baseball Baseball is a very good example, something that is very structured. And then you have other things, like football, which is a bit less. And I actually have a point on that because I thought of research a thing.

Speaker 3:

There's been research to that, whereas they've asked of the five I think it's five major leagues in the US. You have NBA, nfl, the football league, nhl, the hockey league, you have the MLB major league baseball and then the I'm missing one the MLS, the soccer league. The five leagues and they've asked is data transformative to the way that you do things to the general manager and stuff like that? They saw, for example, in baseball, which is obviously a sport that lends itself extremely well to? It is very discreet, you have discreet intervals. I think the majority of people said it was hugely transformative If you would have the data maturity curve. It really transforms the way they play the sport.

Speaker 2:

That's the money ball. That's the money ball thing.

Speaker 4:

It's a movie, by the way.

Speaker 3:

And then you have NBA, nhl. They sort of were comparable in that some teams really considered it to be already transformative. I think Houston Rockets was an example. There were others that really used data to change the way they play and it's transforming slowly, but there were a couple of parameters. And it might be more exciting, the same article, but there's a couple of parameters. It's how many players do you have on the court, how much stoppage is there? Then you stop at certain points and say, okay, this is an event, and then you have another event. With baseball, it's like somebody pitches you and then you have the event end. With basketball, you can try and wrap it up in 24-second possessions. Same thing goes for a lot of sports Football as well. You have the downs, but when you go to a sport like soccer, it's Football means American football, american football, sorry. When you go to soccer, there is a very fluid kind of sport where you have there's very little stoppage, and then it becomes way more difficult to use analytics to transform the way you do things and I think, running.

Speaker 4:

Why is soccer so much harder than basketball? Like every time, a ball changes position from team.

Speaker 3:

You have more players. You have 22 players compared to 10 in soccer, so more difficult to model. If you would like to try and model things, it's more difficult to do. You have less discrete intervals, so a lot of soccer games end on 0-0. Less things to measure. You can measure passes, but a pass will have a. The end result is your Do I end? 1-0, 2-0? It's very difficult to track individual contributions to a success, hence making it difficult to measure and transform the way you do I have.

Speaker 1:

Like the topic that I brought is very on point with what you're saying. But before I moved I just wanted to touch a bit back on the running thing and also the how do you measure and like the feeling. One thing I think I also have an ordering and one of the things I thought was very nice is that they give you very low level or very raw data, you know, like heart rate, heart rate variability, temperature, all these things. But they also give like high level interpretations of it, right, Because, for example, maybe Bart, you're really into running and you care about all these things, but maybe I don't really care, right. But then they give you a readiness score that is very easy to interpret, right. But it also aggregates these things and I think that's one of the things I really like about the ordering action. You know, I think it kind of caters to a range of people. You know, maybe you don't care what your temperature is or your heart rate, is your lowest heart rate when you're asleep, which you want to know how well you slept or how much activity you've done or, yeah, all these things right. And I think by being able to do both things you can kind of cater to both groups. So what is super easy to interpret, but it's very like subjective, I guess, and I guess you can debate a bit how they come up with this number. And, on the other hand, it's very concrete, it's very that's it, it's factual, but it's also like what does this mean for me? And I also even took that idea a bit more. When I'm presenting insights to stakeholders, right, I can say this is something or like, for example, if we talk about data quality, right, there's so many things we can measure. But like, is this stable, does it have, is it a healthy table? Right, and maybe you can have some high level things like this has an A score and A is green because it's really good. And why are you saying this? And you can drill it down to other things.

Speaker 2:

Yeah, and to bring it back to Tim's, this maybe like with soccer so many days, there is actually so much data being created, like in a play, like and what becomes hard to like. What is the summarized message of that?

Speaker 1:

Right, this is good players, all the data equally important, yeah, right, like if you have a player without the goalkeeper when the ball is all the way in the offense, does it matter so much what he's doing? Maybe probably not right, but like it's still data, right. Maybe another point that I had, like maybe I'm just being a bit of devil's advocate, but I couldn't help but thinking it you have this, you have the program you did and you have these coach apps and all these things. But he also mentioned, like how maybe the runner coach just kind of gives the same workout to everyone, but there's some variability and for me sometimes I'm wondering it's like man, like just do something you know. Like just just go out and run, you know. Like maybe it's not going to be the most efficient, but it's much better than because I also I had like, even with diet and stuff right, it's like, oh yeah, but let me research this diet, let me switch this. It's like man, just do something simple, you know something that is kind of agreed and just stick to that. Like I think sometimes for work, like when I say workout, I'm usually thinking of gym, like lifting weights, right, and then there there's a whole bunch of different types of workouts and maybe you should do leg day rest day, this and this, and then for me it's like it's easy for me and this is very personal to almost like make up an excuse and not do anything. And the approach that I took is almost the opposite. It's like I'm just going to do something very simple, because I'm not the kind of guy that is very interested in this. I do feel better when I work out, but to me it's like let's just go, even if you're doing the same thing, just do something simple, but like have consistency, and I feel like doing that I can get further in my goals, my health goals, but even like other things as well. Then, if I was really trying to optimize for the best workout every time I go, I think it's very personal, right? I think it is. I think it's also because I usually don't hear that argument so much from people. Usually, when people are discussing this, they're discussing what's the most effective workout, what's the most effective diet. But I also feel like maybe people should also think like don't worry about the most efficient, you know, if you're not doing anything depends on yeah, well, yeah, there's a big.

Speaker 2:

There's a big gap between I'm not doing anything and I'm putting in all the time that I can. I want to be as efficient as possible.

Speaker 1:

Yeah. It's a big different like it's, but like people are very different, right, but I feel like, once, people usually don't discuss diet, they usually discuss the higher end right the most. Yeah, sure, that is true. Yeah, that's true. And also I'm thinking like, indeed, like, if people are different and we get the same workout, how much are we optimizing by getting the best workout? Is it the 20% or is it the 80%?

Speaker 3:

You know, there's this like to add to the point of the of the like the food. You have a lot of people I have a lot of people right now in sort of my inner circle, of people that I know and that I interact with on a daily basis that right now try to go vegetarian and not, and they're like. And then a lot of my friends are like, yeah, but I'm not really a vegetarian because I eat. Like one once a week I eat meat. And I'm like, yeah, but why are you doing this? It's for it's a lot for ecologically ecological reasons, because you're reducing your carbon footprint and stuff like that. If you just eat once a week meat, you've reduced it a lot, like you've done fine, you can be happy. It's not because from time to time you eat meat that you're still a sinner, like you're doing fine, and we reduce a lot of these decisions. And also like this the sportiness reduces to binary, is like you're doing good or you're doing not, and it shouldn't be like that.

Speaker 4:

This week I heard another podcast that we're actually replacing religion by nutrition, health programs, exercise programs and finding the various valid comparison yeah, true.

Speaker 1:

But are we okay to move to a smaller soccer?

Speaker 2:

Some data yes.

Speaker 1:

So this is so. Actually, I've known this for a while. I saw this at a meetup at in Loven here, actually, we actually went to the local team stadium, so they have like a room for doing presentations and whatnot. So this comes from the research department of Q11.

Speaker 3:

Just to be clear, like the local team is a professional soccer club.

Speaker 1:

It's not professional, but it's like. Well, I'm saying this because it's like second division in Belgium, first, first, yeah, it's first, they're quite, they're really stable.

Speaker 2:

It's like it's not the no.

Speaker 3:

Okay, it is a professional team.

Speaker 1:

Yes, but I thought that when I, at least when I first got here, I thought there was a second division. Okay, that's actually true, yeah, right, yeah, okay, so the first division. So they are a professional team. They do have like a nice stadium, they do have like a nice room, the presentation and whatnot. So, basically, this is from a research group at Q11. Actually, I think I even because I'm also interested in this I think the professor that, uh, that leads a lot of this research, he even went to Barcelona, to the football club and they like a conference and they'll present these things, and I really liked it because I think a lot of the problems that you're describing to him, uh, are something that I also thought to myself. So the thing is, um, the name of the, the, the framework, is called Vape, and it's not like vaping like is a V A E. P stands for valuing actions by estimating probabilities. So, and if you're not on the on any live stream, I would recommend having a look at the website later, which will be in the show notes, but they have some nice animations, right. So they start by saying how many actions there are on a football game, but then basically, you can discretize this, so like passes, shots and all these things, right, um, so the first thing. They say that usually. Um, so the first thing, they say that usually less than 1% the actions are like. Actually they say 1% is what counts. So again, there's a lot of actions that don't matter right, traditionally football, you see shots, assist and all these things and that is a very strong statement like do you need all the actions to get to the 1%? yes, they're gonna get that, but so and again, indeed, there is also a little video here. Maybe I'll point to here for you guys. So they're basically saying in football you see shots and assists, but they give a very good example where defender on the on the midfield but he's still like a defender, the guy just passed the goalkeeper. The second one, I guess he makes a pass sideways to the left back. The left back goes since a through ball all the way to the front where I think Sané, I think assist, but basically it's like the guy finds a through ball that puts the ball almost next to the goal and then the guy crosses and finishes. Traditionally there only two statistics that come up the goal scoring, which is Gabriel Jesus, but he basically just pushes the ball in, and the assist of the guy that basically just crosses, but that a very crazy good through ball. It's not considered at all, or it is as considered as the sideways pass on the defense, right, because it's like percentage of passes accepted, right. So which clearly they influence the game differently, right? So then they carry on to kind of say how they model the game of football. So every action has a pre-action game state and a post-action game state and we'll get to the states a bit more. But basically there's the action and that's how the game was before and after, right, and basically what they also says that the game, like there's a probability of scoring a goal and a probability of conceding a goal before the action and then after the action, right, and the value of that action is how much that probability of scoring a goal will increase or decrease, okay. And then they say okay, so the idea, the value is again the probability of scoring a goal minus the probability of conceding a goal.

Speaker 4:

So there are two probabilities there how do you determine the probability of scoring a goal?

Speaker 1:

that's the cool thing. So they basically do this retroactively. So it's like almost in reinforcement, learning how you have the, the action that gets the reward, and then they give back all the passes and stuff they do something similar with with the goals course. So it's really probabilistic. You know, a pass when the, when the players are laid out like this in this direction, increases the probability by this much because they see statistically that this pass X amount of times increase the probability by this much. So it's probability, it's a mathematical.

Speaker 3:

In the mathematical sense it's like a huge mark of chains of mark of chain of events. Like, every event leads to another event like somebody gives a pass leads to a past and their interception leads to a past, and then when a goal is scored it's sort of back propagates like a neural network back propagation. It's sort of back propagates all the way to the original sort of first pass that was given at the beginning of an offense yes, yeah, and the states are, of course, never the same.

Speaker 1:

You never gonna have the same players in the same location, given the same pass, in the same orientation. But like it's too close enough, right, like it's like, it's just like the state is. Like in a video game, right, like in the game reinforcement learning. The states are not exactly the same, but they are similar, like just because they're not exactly the same doesn't mean they're all equally different, right. And then you can do this again statistically. You do this for a whole bunch of times and you can kind of see what actions, how much they're, how good they are, right, yeah how many like, how do they try to make discrete states?

Speaker 3:

because I'm like you can have an infinite amount of states of like if you were trying to configure soccer players on the court, yeah, and, and you have a ball, you have a, you have a referee, you have a goalkeeper, you have a minute, yeah, there's so many different states that you could possibly, because having a having 22 players configured on the court in the first, in the, in the fifth minute of a soccer game, yeah, between the 85th minute, it's completely different narrative yeah, for sure and how do you make this actionable?

Speaker 1:

so yeah, so I'll get to that. I think, but so the way. So again, I don't have the full answer, if I'm being honest, right, but I think they're showing here the game states and they show like the change of the ball, the change of position in players. I think I believe they may even track all the players here. But then you see, like how body parts before, body part after, all these things, how it influences right and in terms of making it actionable part. So what they do now is that they rate every action and then you can see what actions were good, in terms of what actions increase the probability of square goal or increase the probability of conceding a goal as well, right. And then, once you have all the actions, you can attribute that action to a player so you can see how much each player contributed to the game. And even on the meetup presentation I remember this that they were showing like the percentage of passes completed, and then usually there the defenders, because the defenders make the safest passes, right. But then they were show like, okay, but after you run to this framework, you can see that the most valuable classes was like Kevin de Bruyne, you know, because usually the past that he make leads to a goal, so you can even takes that in consideration, right? So here in the example, like Gabriel Jesus is scored the he's Brazilian, by the way. Just saying he scored the. He scored the goal. But even when you see, the probability of like his score is not as high as other people, you know, because the idea here is that they contribute more to scoring goals. And then you can do this for like. How do you make this actionable? Actionable? They also mentioned there how you can go, I think they went to the Dutch second division and they started rating these players based on the frameworks like this, and then they saw that after two years they were in like the Premier League club. Yeah, so they saw how this was actually taking effect.

Speaker 4:

But this is in scouting, but how can a player improve itself?

Speaker 1:

well, I think that's a different access to. I think maybe, well, again, I'm just kind of thinking a lot here. Right, maybe you can say, well, you're taking, you know, a lot of passes, but usually the one like you're not influencing the game, right, I think that's already a metric. Or maybe you can run simulations and say, well, pass this way or that way, you know yeah, but still like, let's say they play the game.

Speaker 4:

What should this player work on in the next training session?

Speaker 1:

so I think well again, I'm thinking myself and I was practicing right. So, for example, I would basically try to consciously change my style of play. So, for example, if I'm a sent defensive midfielder, I'm gonna try to pay more balls to the tracks attackers straight away right, and then see if, over time, like if that actually influences number of goals and everything, and then, if not, then you can play more conservatively.

Speaker 3:

You know a very simple way also of using the. This type of insight is like imagine that I'm a player and I think at some point that in basketball exists like win shares. It's like what, how much is your contribution to win? And they try to have a metric of, as a player, your contribution to win and it's, I think, is really funny. But imagine that you have a low win share, so you don't contribute as much to win. You can have a look at other players that are hugely influential to win and you could look, just watch film, like just by, just by knowing like, oh yeah, this is somebody that really it's sort of a change, just instead of making it like it's not like you. You, you have this mall and you just push a button and there's a number of action that you can improve, but it's, it gives you insight and sort of you can use these insights to to say like, okay, what is, what is my next level? Alright, I just look in like the next segment of win shares, these type of players. So that's that would be my next step. What are they doing different compared to what I'm doing right now? And you can investigate this and I think you can. There's there's different ways you can look at explainability of the model and you can use explainability techniques to the model and try to find out what you could try to run. What is it that the shop leave and be like oh yeah, this is a huge positive contribution. I have to do this, yeah, sure, but I'm not sure if that will work that way, but I think also one thing that is cool.

Speaker 1:

I remember from the presentation, so it's not on the this they were also saying that once you have that, you can also compare players. Right, you can say how many actions per game do you take and what's the values in average of your action. And from that this was, I think, a little while ago. But there's always the big debate you know Messi or Ronaldo, and then they could, like they came up with a concrete, undisputable way to say the messy is much better than run out, like they were showing like the body and it's like Ronaldo is really good, but Messi was like like you see the cluster of points I think they plotted like literally on the x-axis, the number of actions and the value. So you want to be on the top right, a lot of values were a lot of a lot of actions with a lot of value. And then you see other players kind of like distributed, but then you see Messi like completely isolated. You know, and it was. And again, you, of course you can debate, right, but like you're not gonna debate the outcome, you're gonna debate the process, right, and I think it's still fine, but it's a very indisputable way of saying if you agree with a, b and C, then you have to agree with me which I thought was super cool with something you'd never have in football me coming from Brazil. Big part of our culture is football and a big part of it is just debating and discussing, because it's never concrete, you know yeah, one of my, one of my all-time favorite podcast is it's called thinking basketball and they do these type of like.

Speaker 3:

They constantly do these type of bits like they have an episode and it's what are the ten best passers right now and how do we rank them? And they have just an hour-long debate and I love this. This might sound super weird, but they try to rank. But they use different metrics for like, because these days in basketball, what they used to do was they had sort of the same things. You have points, rebounds, assists. That was the first generation of sort of metrics that they tracked. And then the next thing that you start doing is you try to have an estimate of if you're on the court versus. You have on-off numbers, like if you're on the court, is the team doing good versus if, and then, if you take the difference between these two, you have your impact. And then the next thing was that they sort of had to take into account the fact that, yeah, but you're not alone on the court, so you have to take into account the other four players, and there was a next generation of metrics and then and what they sort of do in all these type of things. And I think it's interesting because if you look at messy versus Ronaldo, who's the better player? It all depends on your, on your, on your view, right? Yeah, what, what is important to you all? Very, very much. But, for example, in basketball, the thing is also like Michael Jordan versus LeBron James. You try to. There's the same debate going on for a long time, except they didn't ever played against each other. But you, they develop these metrics and at some point you want to validate, you want to validate the fact that these metrics make sense. And how do you do that? And then you have the eye test. They always say, like you look at basketball, and if you would ask 25 people like who's the best passer right now, and they all come to the same answer but the guy that they are mentioning is only 25th in your metric that maybe the metric does not make sense. So there is this there, this entire thing of developing descriptive metrics that sort of hold true in a debate, and I love how debate driven it is yeah, yeah, it is.

Speaker 2:

It's very like feeling yeah, and the descriptive metrics are the things that are very limited in the informative volume. Yeah, yeah, that's where natural language shines. Yeah, is there. Do you see in basket or in soccer or like? Do you see places where generative AI will play a big role?

Speaker 3:

I think I haven't given, like I've tried to think about natural language specifically.

Speaker 2:

You can also think multimodal doesn't have to be natural language.

Speaker 3:

We vision can be multimodal right now already, like right now, a lot, of, a lot of vision is used in basketball track, like I said, tracking stats. They're hugely popular up popular right now so it leads to more, you know, more informed decision making. But I think in the net language not multimodal, especially the vision part. Yes, for sure.

Speaker 1:

I think. I think maybe in terms of everything that is like person driven, you know, like maybe the guy that a trainee or he did this or he feel pain, or like these things, I think that's when the NLP and maybe LLMS comes in place. You know it's like oh, how do you feel about this game, how do you feel about this? Are you nervous? Or maybe these things that are closer to the person, like I think this, like even this framework, I think it's more taking a step back and looking at the whole game as a whole. But when you start getting more individual level and I think your question was also very relevant to bowl like how do I as a person can use this to improve myself I think that's when the LLMS part come on. It's harder like to. But yeah, we have time for a little game.

Speaker 2:

I think we need to look at you. You have a hard stop.

Speaker 1:

I do, but I think, if we did.

Speaker 2:

It's an interesting topic, I think, when we shoot. It's about to come back to definitely. We did have a lot of extra topics on the backlog. We did.

Speaker 1:

We did, but I think it's another time, but I think this is a very interesting discussion indeed and I think it merges some passions of mine as well. So be happy to have you guys back in the podcast soon, so for now let's play with quote or not. So do you know what it is about, tim? You played it before.

Speaker 3:

You both you know used to have a different name. Yeah, I know he or Jenny.

Speaker 1:

Kevin, show me down, but I'll explain anyways. You know I always ask but I always explain. So the idea is that a puppet master, he picks two quotes from Jenny, from chat You're your favorite LLM and then he blends that with the real quote and it's our job to identify what's the real quote and what is the fake quote. This time Bart won for the first time.

Speaker 2:

Yes, Emphasis on the winning. I'm not on the first time.

Speaker 1:

So he got to choose the person or the character, the quotes and the fake and real ones. So I see what he got for us.

Speaker 2:

Yes, so I'm going to give two fake quotes, one real, star Trek themed. That's what we decided last time.

Speaker 3:

Yes, we decided like sports. No, actually, that would have been good.

Speaker 2:

We decided last time already, so it's Star Trek themed. It's a quote from data. Does everybody know data? Oh, this is sad. Data is an Android, a humanoid Android, a robot, in other words. That first appeared in Star Trek, the next generation. Star Trek, the next generation, where Patrick Stewart was Jean-Luc Picard.

Speaker 1:

Okay.

Speaker 2:

Arguably the best Star Trek.

Speaker 3:

We're opening up this episode to so much.

Speaker 2:

Arguably, I'd say maybe deep space nine. Anyway, let's go to the quotes of data. So you have the context. It's an Android robot and I have three quotes. Let me get them. One is fake. If you laugh, do I not learn to understand Question mark? That's one. If you question me, do I not ponder? Okay, one second. If you prick me, do I not leak?

Speaker 3:

The fun thing about this is that it's an Android right, so it really feels like. It doesn't feel like the like chatGPD is even like too smooth already in the conversation. It's too human in the conversation to be like Android communication, so it's like the. What is it? The 2016 Facebook bots that they threw online and that sort of started.

Speaker 2:

This is simply data. So the third one. So the third one was if you prick me, do I not leak Fake, fake? You say fake. Clearly, you need to pick real one.

Speaker 1:

I know, I know, but that one is out of the game. I already increased my chances by 50%. The other ones?

Speaker 2:

if you question me, do I not ponder the third one? If you laugh, do I not learn to understand which one do you think is real?

Speaker 1:

I think the first one is true.

Speaker 2:

Okay.

Speaker 4:

I'll follow my order on this one.

Speaker 1:

Good choice, good choice, I must say.

Speaker 3:

I wanted to go for the first. I'm going to go for the second one. Just be controlling.

Speaker 1:

But let me explain first my reasoning. Okay, I feel like you also emphasize that it's a robot. I also think that that's what you told chatGPD. So of course the thing is going to say like leak and like bleeding and all the same.

Speaker 2:

Like you, chatGPD will know who data is. I don't need to tell you Fair enough.

Speaker 1:

It's still thing is going to focus a lot on the robot, not bleeding kind of thing, and I think what was the first and second one.

Speaker 2:

If you laugh, do I not learn to understand? If you question me, do I not ponder?

Speaker 1:

Yeah, I feel like ponder is something like chatGPD, but you're in as well. Yeah, but I feel like if you think of a robot, robot doesn't ponder. You know like I don't know, I feel like those but a robot also doesn't yearn.

Speaker 2:

Choices were made.

Speaker 1:

Yes, let's locked in, let's see, and I'm going to all right.

Speaker 2:

The winner Really, the only real one was if you prick me, do I not leak?

Speaker 1:

No, Really Wow, bart, great job, great job.

Speaker 4:

Just a, I must say you got pretty good at this. What's the context in which this quote occurs?

Speaker 2:

It's the one question too much. This specific one, I don't know, by the Google like, give me Of this start. I've actually seen all the episodes, but I've doesn't. Doesn't really have an active memory on this quote. Wow, but it is a recurring team where data Tries to understand human emotion at some point becomes, tries to become.

Speaker 1:

Huh, yeah, it's like the like. What's the name? The? Sheldon Cooper no no, no, no, that was like, sorry, with the lion, the robot and the no. Never mind, it's a person thing. It was to be a real boy. You know, it's like.

Speaker 3:

It's Pinocchio.

Speaker 1:

Yeah, they know that's the one too. Good job, Bart. That means the next week You're also in charge of another one, but maybe that's the team.

Speaker 3:

Did you win though?

Speaker 1:

Yeah, I mean I know I got it right. I kept it. It's still the champion.

Speaker 2:

What is the team for next week?

Speaker 1:

Well, let my guests decide.

Speaker 2:

Yeah, I think it's the ball.

Speaker 1:

Yeah, he was the most critical. No, who is the critical one? I think it's Tim. No critical. Yeah, but like twice. Not a sports figures, yes, but maybe a yeah, go for it. Who do you think Bart should come up with next time?

Speaker 4:

Inspired by Mirillo's outfits. Stay in the Brazilian team, oh okay. For soccer football, whatever you want. If you want to bullsign at all, it's also fun for me.

Speaker 1:

Yeah, maybe let's not go there.

Speaker 3:

Yeah, no, I like it, Ronaldo, like the phenomenon.

Speaker 1:

You know who that is.

Speaker 3:

No, but that's not what that's about. Ronaldo phenomenon and no Ronaldo, but you have there's so many, ronaldo, the one that had knee surgery.

Speaker 1:

Yeah, no, that's too much.

Speaker 3:

The fat one.

Speaker 1:

Okay, the overweight one.

Speaker 3:

The guy with nipples, sorry.

Speaker 2:

Okay, all right, brazilian soccer.

Speaker 1:

Okay, I'll let you choose how many Brazilian football players you know. Ronaldo.

Speaker 3:

He's got a point though.

Speaker 1:

That's okay, we can stay with Ronaldo. I'm not trying to Nice. All right Next week. Ronaldo quote looking forward, I don't think you're going to be able to fool me, just saying let's see Take it as a challenge Cool.

Speaker 2:

Enjoy the weekend, everybody, for our listeners. Thank you for listening, thank you for watching and see you next time.

Speaker 3:

Yes, ciao.

Sports Analytics and Running Coaches
The Future of AI-Based Training Apps
Discussion on Personalized Workout Apps
Sports Training and Data's Value
Discussion on Workout Efficiency and Simplicity
Analyzing Football Actions With Vape Framework
Generative AI in Language and Vision