For the first time, we can reveal that last year, Eddie Dumbar's Walta stage wins were based on a strategy which was a direct result of this tool. I can't exactly remember what stages Eddie won last year, but both of them were solo. Yeah, you a huge pacing component to being solo.
You had a pretty big role if you went through those stages last year. What if a secret algorithm could tell you exactly how to train to win your next race or complete your next target event in a personal best time? We had a few predictions on like in terms of kilogjles kilogjles he would expend for each stage and so being able to model that and model you know the the rate of consumption of those kilogjles over the stages.
This isn't some futuristic dream. It's already here. A powerful new AI system is quietly reshaping how world tour riders prepare and it might just be the edge you've been missing.
And so that's how we kind of measured the impact of temperature elevation on his performance, trying to quantify it and determine the stages that were most optimal for him, but also to try to optimize training before the races. Uh trying to tell him like for examp you have a big drop. On this episode, we sit down with the creator of Vecta, an AI training platform that could completely change how we train, recover, and race.
Stick around. This conversation could rewire the way you ride. Paul, welcome to Rob Podcast.
Thank you. Thank you for having me. Thank you for flying over.
Yeah. Uh, Vector, it's a interesting piece of software. We're going to talk a lot about it today, but from day one, you had buy in from World Tour teams.
That's very rare. Why do you think they trusted a group of students with high stakes data? That's a good question.
Um, so actually we started just out of university. So we were kind of like you know looking where we wanted to go with like building companies with just AI sports and where to go. So we started with like lots of interviews understanding what the coaches want and so we started with like more basic level and tried to increase our time to see like what the coaches are actually using and how they're using the data.
And so we started discussing with water coaches and we were quite like shocked by the fact they only use train pics and that was the only main data stream for them. So we were like maybe we can do something better for you like AI is booming at the moment. They were like they were uh having discussions like in with other industries like we use more AI tools.
So like I think they were yeah they were quite in intricate with like what we can do with AI for them. Yeah, they gave us access to some data and that how we get started actually. Is it a case that AI is moving so fast I've seen projections saying that it's doubling in intelligence every six months at the moment and with them being so busy you see this with a lot of professionals a lot of my friends like especially in uh private equity they love the idea of AI but because it's moving so fast it is an investment in time to figure out its application into your daily life or to tackle your workflow problems.
Yeah, definitely. And I think there is also like a pre chip and after chip with like lots of people after chach like it's amazing to have AI for for them. Uh but yeah, it's time consuming to kind of like learn the process and learn what they can do with AI on a daily bas basis.
I think you I've been shocked by how good chat GPT is for taking data that you currently have like screen capturing your hammerhead training peaks whatever putting it into chat GPT and asking for a recommendation to pace a time trial and giving a GPX file. It'll come back with some pretty good time trial pacing recommendations. Yeah, large language model are just like quite impressive and it's going to continue to like go in this direction.
So like it's uh I think it's good to have like general models like this, but it's also interesting to see like how we can use and customize those models to be even more specific and like for example being more personalized and or precise in uh human performance. That's why I'm excited to dig into your platform. I've played around with AI quite a bit.
I tried to protect at least two hours every day for like going hard on AI. But I can been able to build some pretty useful tools as someone from a non-technical background who can't code, but they're just not like quite there. The no code editors and AI, they're good, but then my diagnostic ability to figure out why to take them from good to great just isn't there.
Yeah. I try also to automate as much as we can in the company just to like we are small team. like it.
We try to automate like for example for sales, for marketing to try to save time basically with all these tools. I got to show you the the Lindy outreacher we're using after the podcast. We spoke about briefly on our call, but we started using it a bit more and it's it's wild.
So I'll tell you about it now. So one of the things that we were looking at is you're bringing in new sponsors on the podcast. Yeah.
So it'll set up a G an AI who's basically the supervisor of this mini little task. Okay. So he'll come along and he'll break it down and go, "Okay, the problem we're trying to solve for is new sponsors on a podcast.
First, let me figure out what the podcast is." So he'll listen to back episodes and transcribe them, try and classify the podcast. Now it'll look for adjacent type content.
So it might say Huberman Labs, Chris Williamson. It'll look who the sponsors are on that. Then the next little guy will go along and he'll go on to LinkedIn Sales Navigator and see who the key decision makers are inside those companies.
and then someone else will formulate emails, send pitch decks and so you just get something booked in your Google calendar sponsorship. It's absolute magic. It's happening so fast as well.
But let's try and ground that what we're talking about because we're going to mention the platform quite a bit. Uh am I pronouncing it right? Vector.
Yep. Perfect. Give me the elevator pitch.
What is it? Yeah. So Vector is a training and coaching platform for athletes and coaches.
Uh so basically the way we try to differentiate ourselves is on three pillars. Um automations so using AI to save time for coaches and get better insights. Um using a better performance model.
Maybe we can talk a little bit about this, but like FTP, TSS are like old school terms that everyone use and user experience because everything now in 2025 is like um not spending too much time in your apps and so trying to be as easy as possible from the time you get back from you're training to like having good insights about what you did and what you should improve or not. Was the motivation to build it born out of frustration with previous? Yeah, because like I used to train a lot.
I have my gin now I have a ring with like sleep recovery data and I'm like after a session I have 10k run for example for this morning an hour and that's all I get and so trying to understand like more what is within the session and also the interaction between like what the data from training the data from recovery from uh sleep from nutrition from lifestyle and kind of understanding all the patterns together and trying to be like yeah have more insights on how I can improve be better and uh and yeah personalize my experience. It's pretty interesting that it's a consumerfacing product because my fear with AI is right now I can log into chat GPT 40 whatever the latest version is right now and I use the exact same chat GPT I put in the same prompts as Bill Gates does. Yeah.
But there is a dystopian split timeline future where I'm stuck on chat GPT4 and he's on chat GPT12 or you know one set of socioeconomic children are tutored by chat GPT4 and Bill Gates's kids are tutored by chat GPT8. And if we take this and apply it into cycling, we could see the same problems where UAE are using large language models and leveraging AI to its full capability where a smaller team archa budget. So the gap just widens and widens and widens.
Yeah, probably. Uh but also it's interesting I think to speak about like yeah applying those large language models to just like having a good data set and so like the the data that you have access to and how the model can interact with the data. Uh so this is I think where it's important to have like you know solutions having access to your the users data so who was the first teams that volunteered some of their data actually no I knew you guys were working with them that's why you set as an example that's true I think because like we are French and so it was near the connection um so yeah I was the first one to trust us and what does that look like they give you total access to all the rider files?
No, it was like just specific riders like under NDAs and kind of like very restricted. Um but like for a mission to kind of like uh yeah build on uh more race analysis um tools and kind of better understand like yeah race performance and kind of link it link it to training data to kind of understand like if there are some optimization to perform. Very interesting.
I'm trying to think at the moment like the different ways it could potentially intersect and impact. How do you see the next 24 to 36 months playing out with it? Are we likely to have directors in a car using real-time AI analysis to predict places where a bunch might split in a crosswinds or based on historic data, weather data, real-time rider data?
What's the bottleneck to this UCI allowing access to that data in real time? Uh that's a good question. I'm not sure I'm the expert to speak about it actually because I'm not involved recycling and working in that.
But I think this is the the direction that uh Visma tried to take last year with their data uh truck they had like for examp so like accessing as much data as they can during the race and so that can inform decision. Uh but the issue with the UCI I think they that they do not have the right to access like for example riders data like power data, heart rate in real time. Uh so like it's going to be limited some point but they can at least analyze you know race dynamics and weather and kind of like have live realtime uh computations.
I would say you guys are interesting because it's not only is it a new software platform you're fundamentally changing the grammar we use to talk about cycling. Since I've started, I actually it's a funny story. My first parameter, I was just out of law school and I'd agreed to sign for a French tame for that year.
You would think I'd have better French after living out there for a year. And I falsified a student loan application to say I needed to buy a car for a master's program. And I took all that money and I bought an SRM at the time.
So it was like €25,000 euro or something for an SRM. But there wasn't, you know, resources. There wasn't YouTube videos on how to use it.
You couldn't get access to like we were talking about Dan Lang on the podcast. Now you can listen to the very best coaches in the world talking on podcasts and they're giving away like you would have paid thousands in consultation fees to chat with these guys and they're giving it away for free. I didn't have access to that.
The only book at the time I could find was Hunter Allen and Andrew Kogan training and racing with a power meter. And a lot of the terms they used in that intensity factor, normalized power, average power, training stress score, chronic training load, acute training load, training stress balance, it's still the vocabulary we use today. So you guys are trying to eradicate that vocabulary.
Yeah. basically because the first reason is like all those terms are quite old now. So like start of the 20s I would say um and also because they're quite linked to train pics as well like under kan like kind of all those terms went into train pics.
Um and also because this is the time to change like for example FTP just like one measure you have a an FTP of 300 watts and this doesn't represent your performance because like if you're a sprinter like you don't really care about your FTP you want to know like your max power or something else. So it's like trying to move into new concepts and the things that has been demonstrated in science I would say. How do you begin to solve this problem?
How do you begin to even understand and frame the problem you're trying to solve? That's a good question. And so that's why we tried from the very beginning to work with the pros and work with the water teams because we wanted to be like you know there are new concepts out there that they use in those teams but we want to have them in our platform and so that everyone can use them.
So that was kind of the first intuition like we want to have those new terms within um vector and so yeah just speaking with the coaches the scientists you kind of learn like yeah the new concepts uh how you can model that and what is interesting I think with our position is like the team is like a team of data scientists mathematicians AI specialist I would say and so it's trying to understand like concepts coming from sport science and trying to kind of like improve them with the concepts and the knowledge that we have in AI and data science and kind of like bring them together. So I think it's yeah trying to apply more mathematics to sports science because sometime it's like in sports was like lots of linear regression kind of like simpler mathematical models and they have some limitations. So it's trying to explain uh a lot more and also because we have way more data sources like in the end we have like uh during train ride we have second by second data with like lots of fields and even more um data.
So it's trying to have models able to take into account those different parameters. It's also a double-sided coin, isn't it? Because you have increasing amounts of data, increasing complexity in the mathematical model, but on the flip side of that, you have increased reluctance to engage with data.
So it has to be a simple user interface for a more complex back end. Yeah, definitely. And that's why for example, we try to have simple terms.
And I think the good example that we try to use is Whoop with their recovery score, sleep score. It's like they they have some scientific metrics behind with HRV resting heart rate but in the end they just simplified in basic scores that everyone can understand. So trying to have quality accuracy and you simplify it but I can't remember which philosopher said it but it's like we can distill a problem with simplest form but not any simpler.
And I wonder that with some of those, have they distilled them beyond the simplest form? I know the terminology we talk about like it it just it did baffle me for quite a long time like CTL like chronic training load. Why did you not just call that fitness?
Exactly. I think they tried. Did they try?
Yeah, I think the CTL they tried. No, fitness is TSB. I think uh TSB I would say is freshness or the balance between the two.
But again, that's the the Yeah. So yeah, I would say CTL. So, if anyone doesn't know what we're talking about, so CTL, it's a running 42-day total of training stress scores.
Yeah. Divided by 42. ATL I would say agreement.
That's fatigue. Yeah. Fatigue is 7day running total.
And then TSB, I don't know how they calculate that. It's some it's the difference between the two actually between the two. Okay.
Yeah. But you take Sarah for instance, who you met at the start. Sarah's training for Badlands this year.
And her background is interesting because she doesn't come from athletic background. and she comes from a drinking background. So, she's taken into cycling for the first time, but she's just like baffled at the complexity of some of these terms where I'm trying to explain to her about ramp rates over the course of a week.
She's like, "What? Why is this so complicated?" Yeah.
And also, yeah, the abbreviations like you need to understand like TSB like training stress balance like Yeah. You could simplify everything. So what's the So even if we take some of these terms one by one and maybe explain the flawed thinking and how you're going to address it start with thresholds is one you touched what's the problem with using threshold so threshold FTP is just like one measure uh that you have and most of the time if you want to test your the first thing if you start with a coach or start cycling we're going to ask you to do like a 20 minute maximal effort and from that we're going to compute your FTP and discount the main parameter.
But the issues with that is 20 max effort is pretty hard to do. You cannot do it like every week to kind of like see improvements and also it's hard with pacing. So like can fluctuate a lot on your the numbers that you have in the end and that represents the most of the time like the one hour max power output that you can sustain.
And so it's trying to shift to something that we call like more critical power. Uh so like the power duration model so that you can have the maximum power for all the durations and kind of like understand if like I'm more a sprinter and you're more a climber for example. Maybe we have the same FTP but I have a way higher max power and you have like something that is way lower.
Um and so it's trying to have something that just represent uh the whole part of the power duration and um being way more precise on the performance metrics that we have. So instead of having just one we have three uh critical power that is most of the time like similar to threshold w prime that is like the anorobic like the available work capacity. So like the amount of kilogjles that you have above your threshold kind of like limited amount and maximum power the maximum power you can do like on 1 second.
So we're using the combination of these two set zones cuz I guess historically I don't know off the top of my head the breakdowns but say we would have taken a 20 minute test as a proxy for what we could do for one hour. Yeah. Then we take the 20-minute test and we multiply it by say 0.
55 to 6 range and that now becomes our zone 2 endurance zone and then we figure out everything else based off a percentage of that 20-minute test. How does the new zone setting model differ from that? It's kind of a similar concept for the moment.
It's kind of like taking percentages but like not only from one parameter but from this threshold parameter but also like the anorobic parameter as well in component so that we have something that is way more precise because yeah the windows can be different depending on your rider type. Interesting because this is something I figured out the limitation in it firstand last year. I wasn't training very much last year but I had a competition with a friend every week.
He set up a indoor time trial league every Tuesday, which was a full gas 20-minute effort every Tuesday. Now, I found out if you do a full gas 20-minute effort every Tuesday, you get very good at full gas 20-minute efforts. You get good at pacing, you get good at delivery.
So, at a point then I'm hitting I think I hit like 405 watts for my 20 minute TT. But then you take that as okay, well, this is my new 20 minute power. And you create zones off this.
And towards the end of the year, I went out to do some zone 5 efforts and I was like, I can't get near these numbers. These are not representative of what I can do at all. They're wildly off what I can do.
Yeah. And also because you were on TT, so maybe the influence can like it's always different on road bike and TT bike uh for the power values I would say. Yeah.
No one's just indoors on the what bike just Okay. indoors. Um yeah.
Yeah. Of course. uh there's this point and also the fact that you know if you use the concept of FTP and in the platform at the moment it's very manual like you need to manually change your FTP change the zones and so what we try to integrate as well is to the concept of adaptive zones and so that we have those metrics that are computed every day adjusted based on your latest training racing data and so that you have the zones that just adjust all time with your progress so that you do not have to do test to update pay them every two months, let's say.
Interesting. So, it's like automatic zone detection. Yeah, definitely.
How does that work as you transition from a race season into a base period? That's basically the science behind the modeling. It's kind of understanding like, you know, the buildup part, but also like the offseason, trying to understand like the decrease in performance during the offseason part.
So, it's kind of Yeah. the modeling behind it. Yeah.
Because I guess there's a double frustration with that. Sometimes you come back in the off season if you haven't raced at all and you have no race data. There's a maybe a temptation that the model underestimates yeah your threshold power.
But the flip side of that is sometimes I've come back after offse and my threshold has actually dropped significantly and I'm still trying to ride to the old zones because I haven't retested and you don't want to retest because it sucks. So you're kind of thinking I'll do a month's training first and then I'll retest. But if you're going off the old zones, you're getting not the adaptation you think you're getting.
Like you're building extra fatigue that isn't been measured properly. Yeah. So that's why we try to model, you know, like the performance model with the load modeling.
And so like for example, if you start in beginning of November with like only aerobic, right? So like you do not have like max power values. Um but we just use that to kind of like understand and model that there is a progress in CP even though you only try um walked Arabic zone.
You built this program for athletes or for coaches for the moment it's mainly for coaches because we think there are more value in AI for coaches. uh just the fact that it helps them to save time, scale the business and have better insights and because it's quite technical and um we need to have um people who understand the science for the moment before trying to applying to everyone. Yeah.
So you're almost hoping the coaches will become these brands Yeah. evangelicals who teach their athletes about exactly the new science as you did with like the all the terms you mentioned before with TSB and uh TSS and everything. It's like it's easier to just use the coaches to explain the new terms and the new terminology and the concepts.
Okay. And so are we replacing the old performance management chart which was built on training stress score essentially? Yeah.
Because the first issue is uh training stress score. So it comes from TSS and TSS is a measure that is um scaled by FTP. So the first one is like everything relies on FTP.
So if you measure it wrong or like do not update it like everything is wrong. Yeah. Behind it.
Uh so trying to have um a metric that we call volume that is not linked to any threshold value. They just link to kilogjles and weight and so that you can see like for example over time if you if you've been training for five years do you understand like how many kilogjles and how much you progressed over time without any threshold component because otherwise if you divide by threshold it's kind of easier like you cannot see the progress because like over time you increase your threshold as well and since everything is divided by threshold like in the end it's the same values. Can you take a second to talk about kilogjles because I've had a lot of coaches, physiologists on the podcast.
Kilogjles is something that always gets dropped in, but there's a disconnect between the amount that World Tour coaches and athletes use it and talk about it and understanding of your general athlete who's coming from this world of training stress score. Kilogjles isn't something we've ever looked at. Yeah, probably.
Maybe I'm not the the best person to ask this question actually because I'm more mathematician. So like I'm not the sport science and like understanding all those um terms. My point was our point was more to take those concepts they use in more tours and trying to apply to like more mathematical models.
Okay. That is the basis now of the new performance management chart because as you say if you take a bad upstream measurements it can trickle down and have multiple bad downstream measurements like you take a wrong calculation of threshold now it's infected training stress score CTL ATL tsb everything tapering everything yeah so it's like yeah measuring volume in a different um way uh that we think more accurate and also So the coupling um volume and intensity so that we have something that is quite easy for all sessions to have a volumetric intensity metrics that is quite simple everyone can understand it and it's easy for a coach to explain or today this is a high volume session but with a low intensity score. So that means it's going to be like yeah as described but like for to another session kind of explain yeah today the focus is high intensity you're going to have a middle range volume but like the focus is intensity and so it's trying to quantify it uh in two simple scores and also over time to have a performance chart that represent the evolution of load with volume but also the evolution and the distribution of intensity because we talk about a lot about periodization trying to model um in long-term like the evolution distribution of this intensity component.
That's really interesting because I raced the criterium last night 45 minutes full gas legs are absolutely killing me this morning. I could hardly get out of bed, but my training stress score gave me 65 or 70 training stress points. For comparison, for anyone listening, if I do a twohour coffee shop ride this morning, I'll probably get a 6570 training stress score.
They're not the same thing. Exactly. That's kind of the issue because by simplifying they have one term, but you can build up the same performance chart with like very different in the end u stimulus and uh training adaptations.
So it's yeah trying to model that in way better. We had a question in we have a Friday episode of the podcast called writer support where it's like a helpline almost people send in messages if there's a conversation say with Dan Lang and there's a part of that they didn't quite understand they'll send a message in but somebody put a message in saying they had 125 CTL but they were still getting dropped on their local group ride. Why was that?
And I was like I can't answer that because I don't know how you built that 125 CTL. like are you doing 10 hours of you know zone one and you know there five six hours of zone two per week maybe that gets you there. Yeah.
Uh I have this discussion actually with lots of athletes and even pro athletes like talking about TTL and sometimes they are like yeah 120 or like 80 but they feel better at 80 than 120. And it's kind of like the model behind just like doesn't validate your performance. You need to have the highest uh CTL to perform.
Hey everybody, let's take a quick break to talk about the bike I'll be riding this season, Reap. I've been lucky enough to ride all the top brands in the world over the past few years. But these Reap bikes, they're not the same.
And I'll tell you why. Reap is the first company I've seen that isn't chasing sales targets and the mass market. They're chasing something very rare, perfection.
Every bike they make, it's crafted in the UK factory. And it's not about slapping a Made in Britain label on a bike from a Chinese factory. It's about control.
From the first sketch to the final build, they're hands-on, ensuring that every detail is dialed in. That's very rare in an outsourced world of mass production. What sets them apart is innovation.
While others pump out the same old designs, reaps pushing boundaries. They're not following trends, they're setting trends. Think precision and performance like an F1 car for the road.
Absolutely no compromises. And it shows and you can feel it when you ride the bikes. These bikes are built for riders who demand the best.
Whether they're chasing podiums or just want a machine that feels like an extension of your body, a piece of art. It's not hype, it's substance. Ride a reap for yourself.
And you're riding something crafted with intent. So, if you're serious about cycling, check them out. It's reapikes.
com. I absolutely love my one and I couldn't recommend it highly enough. Back to the show.
You mentioned the primary application for the moment is coaches. Yeah, AI is promises this future where everyone has their own assistant and we have this workforce which is working away in the background, sending emails, automating processes. What are the processes that you look to automate that coaches currently perform manually?
The first one was like session analysis. Most of the time coaches just manually zoom on laps, get the intervals data like this way, save it in a note or send a few messages um with like the intervals data. So it was being able to understand what is a training session and what is the structure of a session.
the applications to that or saving time so that as a coach you directly have the data that you want to look at but also comparison and uh evolution because if we are able to automate that it's very easy to say oh yeah today you did like I don't know um a vax session with like 3030s and if we have this model that just detects all the intervals we are able to tell you oh you did this session already three times in the last three weeks and so we can that's interesting com compare the performance and kind of like see the improve improvement over the past 3 weeks. So it's yeah training analysis and kind of like understanding that a training session is not only a total duration normalized power and TSS. It's something where there is a structure and there are some benefits of like uh doing those intervals and so that we can detect them and analyze them.
So, if I just upload a file, I've gone out the door and I've done I don't know, say an old Miki Ferrari session 2040s and I've done three sets of 2040s, three sets of 12 minute 2040s, but I haven't pressed the lap button across any of those sessions, the system's able to figure out the session that I've done based on that. Definitely uh we'll detect them. uh so the beginning and end of all the intervals and then classify the intensity of those intervals so that we can then classify the session and tell you this was like a V2 max or anorobic session let's say and tell you oh yeah we detected five simar sessions since the beginning of the year so here are here is the comparison between all those sessions and here are the improvements that you made how similar does the structure have to be to make that detection all goes back to the data it has been trained H so it's like huge data set that we collected over time where we just um annotated manually all those intervals and so trying to have a system that just detect all the structures.
Uh so not only 20 40s or 30s like the basic sessions but like trying to understand like yeah all the different kind of structures that coaches prescribe. Yeah. Because what I find is the difference between what a coach prescribes and what you do in the real world isn't always the same.
The coach might say, "Hey, it's four or five minute V2 efforts today, and I want you resting for eight minutes between each effort or five minutes between each effort, whatever it is." If you're not indoors and you're not in Jerona, the terrain largely dictates when you can go and when you can't. If it's a red light, you're stopped a little bit longer.
If you come to a place of high density traffic, you're like, "My 8 minute break has become a 12-minute break, or my 12-minute break has become a 4-minute break." Can we still detect a V2 session? Of course.
Um in the end the science behind is just like comes from financial uh modeling is kind of like understanding the trends and so like power data is not a power stream is not so different from a financial asset uh price evolution over time. So trying to understand like the breaks and kind of like the big differences they have um to cut them and label them as intervals. Very interesting.
Very interesting. Can we have a look at the platform? Is that possible?
Of course. So what are we looking at here? So we are looking at vector uh the calendar just of one athlete.
Um I think on all the training platform and coaching platform it's kind of like going back to the calendar where you have all the training sessions recovery data sleep data. Uh so we try to make it as simple as possible so that you have at the same time as you can see like training recovery sleep and also psychological data. How do you feel today?
So that it's very important to have this component as well. And do I need to do I need to have recovery data if I'm not using a ring whoop any of these things? Am I limited?
You're not limited. Just you're not you're going to get less insight. Yeah.
And what was the third metric? You said it's subjective feeling. Subjective feeling.
Yeah. How do you feel every morning based on sleep quality, based on stress, soreness? Um and yeah, old school Joe Fre style.
Exactly. I like it. I like it because that's what when I'm chatting on the podcast to world tour coaches, physiologists, everyone's talking about this trifecta between training data, wearable data, and subjective feeling.
Exactly. That seems to be where it's at at the moment. Okay.
Sorry. Go ahead. Um, so if we can start maybe with a profile where it just where you have your performance metric.
So I told you at the beginning like we have uh instead of the FTP we have this power duration model with critical power with like critical power that is your threshold the very prime that is your more available work capacity uh metric um and then the peak power output that you can uh sustain and from that you can get as we call the adaptive training zones that are just computed based on all these parameters and just adjusted every day. So that every morning for example you can come up there and see if there is any change or any improvement uh on your performance and how when you say adaptive I guess you guys need to straddle it between being it's almost like you remember we got power meters at the start and we weren't used to going from heart rate which was quite a stable reading to power which seem to jump all over the place. When you say adaptive can it be too adaptive or how did you skirt around that issue?
Yeah, that's a good point. Uh, actually try to be more stable in the way like you cannot have 10 watts improvement every day or 10 watts like shift every every day. So, it's more something like we study like the long-term progress over time uh so that you can have uh yeah something that is stable because it's not really realistic to have yeah 10 watts improvement uh on a day.
It'd be emotional roller coaster worse and 10's better. It's like when I weigh myself every day. Um, and we were talking just before um about detecting, you know, the intervals.
And so maybe I can go to another athlete I know uh who's done like um 24 is actually an example you mentioned and it was outside. So as you can see it's it's not perfect. Um so this was like the planned structure as you can see like three sets 24 is uh really standard and as you can see what has been detected is like some of them were a bit higher and you can see also the recovery that is a bit different between all the intervals and so directly if you go to the strengths data where you have all the classic uh stuff and kind of zoom in you have this kind of magic button where you have all the intervals that has been that have been detected by the technology itself.
And so what makes it very interesting is like oh yeah for example we know that uh Dominic did this session quite a few times um over the past few weeks and so we're going to try to compare it with uh similar sessions he's done to see if there is any progress or um stagnation over time. Can we or is there plans in the road map for I know your road map's published on your website, but is there plans then to use AI to help a coach create a report that compares performance today compared to historic performance? Definitely.
Um definitely on the plan, lots of things to to do. Um so yeah, so as you can see have the two session is quite easy because we have this kind of like text version of it as well. kind of see like for example, oh yeah, the first set was like um 11 times on the left and on the right as well with like 500 uh yeah 58 watts uh for the new version in April compared to 577 uh 570 watts for Dom's pretty strong is um so yeah so it's having this ability to directly in three clicks having the comparison without spending times you know zooming on all the laps saving that and trying to find when was the last time he's done it.
Um or time. That's very interesting. Even better if just go to the table over there and you have should have clicked this one.
Um have the table here and kind of see like yeah directly the comparison for all the intervals performed. It's nice as well because the graph is very hard to read. I almost to the point that it feels a little pointless when I'm reading these graphs.
you're like, you spend so much time and then you're scribbling down notes. It's nice having the actual text version of it. Yeah.
Uh I know coaches like to have this version as well because when you go back to a session that you've done like three months ago, you just want to have a summary like the key highlights and it reminds you what you thought at that time. Uh so it's good to have the text version as well. And um I think the next big part is if we want to build uh a session for tomorrow uh let's say well do you have a session in mind maybe you can use the 24 days from before uh but like we try to simplify the process of building a structure manually because sometimes it's quite time consuming and this is the reason why coaches kind of build an extensive training library with all the possible sessions they have in mind just drag and drop because it's takes time to build it.
And so that's why we try to simplify and make it as automated as possible and using the text version. And so for example, if you go like indoor platform and just want to work out for this platform, you're going to do something like uh 15 minutes warm up with like um three sets of 10 times uh 20 or um 20 sec at let's say yeah anorobic uh 40s recovery with uh three mean recovery between sets and uh 10 mean cool down and so from that we use uh some large language language models that we've uh mentioned at the beginning so that you we can transform this text text to um a structure that is directly sent to your because I've actually been doing this with chat GPT where I would write a very similar prompt put it into chat GPT and then copy it back into training peak to create the session library. So, you know, once every probably two months if I'll say right, I need to add 10 new sessions here just for variety.
There is a little bit of rearranging chairs around the table of the Titanic. Like there's only so many zones and so many ways you can cut it, but just for variety, you're trying to put some new stuff in. And that would be a very similar prompt that I would use into chat GPT and it would build out the session and then you're copying and pasting it back in to create the session and saving it.
Definitely. And then you have it sent directly to your GI or Wahoo without doing Oh, so it actually builds the structure so you can use that on Ruby's width wherever you're indoors as well. Yeah.
And so we estimate as well a few metrics from the structure so that you you have everything computed and we have some predictions as well depending on those uh metrics. Yeah. And work and calories are pretty prominently to the four in the structure down below.
Yeah, definitely. And something that is coming up as well is the nutrition part having something where we are able to tell you also this kind of session we recommend 90 grams an hour um from the system because you have a hexis integration on your road map I see. Uh yeah of course.
So can you take us to the performance management chart that we're talking about and so we have a look at how this new performance management is structured. Maybe we can start from a session over there. Um so for the moment um the the only metric that has been integrated is the volume component that you have here.
The formula has been publicly announced yet. I'm not sure it's going to be but um you're just like comparing um and having this volume you say hasn't been publicly announced the formula. Yeah.
What's the concern with announce and how it's calculated? Just copycats. Yeah, definitely.
Um and so it's having something that is proprietary I I would say. Okay. So I try to build it with a no comp no code AI coder.
Um so yeah the volume component the intensity is going to be shown as well over there uh pretty soon for all the sessions and then it's going to the insight se section where we have uh kind of a date component where you can filter like for example the date range that you are interested and have it organized by different tabs and um uh yeah different tabs and focus. And so the first one is for example the one that is quite similar to performance chart with the TSS where you have instead of TSS the volume that is here and for example each bar is the training stimulus that has been detected of the sessions. advisable if you have there.
Uh as you can see this is like a high volume day but it was like only an Arabic session so probably like very low intensity when we integrate the intensity chart and on top of that it's hasn't been integrated yet but we're going to show two more components. So the vector load and the vector dynamics which are like the long-term evolution of um these volume and intensity metrics. Yeah.
Because we're used to looking at the performance management chart, which feels like a chart that monitors our progression. Yeah, maybe that is flawed to even say that it's measuring our progression because I'm not sure it is measuring our progression, but it at least it feels like it's measuring our progression. Whereas this, it's not a progression chart.
Yeah, this one is the load evolution, but it's quite stable for Dominic and also it misses a few components for the moment. So it's uh the performance model has hasn't been deployed yet fully on the platform. What's load measuring?
Load it's um so you know you were you were talking about CTL uh it's kind of the comparison. The main difference is like it uses volume and not TSS and also the time frame of it. Uh so CTL is 42 days on our side load is computed over 16 weeks window.
Why? because sometimes uh just 16 six weeks uh is pretty small and what you did before kind of like influence what the performance that you have today. So that's why we wanted to have a bigger window uh of analysis.
And sorry just to to clarify terms, load is kles mapped. Yeah. And is it mapped in like whereas training stress score on the CTL it's a running 42-day total.
Is this still an averaged out running average? Yeah, it's a exponentially weighted moving average. Exponentially weighted moving average.
Nice mathematical term. You were waiting to crowbar that one in. four years degree not wasted.
Um but what we like the most is for example if you want to study uh the training data of a rider um and since we have this ability to detect the intervals and classify them is pretty easy to have for example you study this um date window from January to April and you have exactly the number of intervals performed per intensity the average power effort uh per intensity as well and so it's very interesting for example to compare this pair to like what the athlete did the year before see if there is any change any improvement or if for example you organize the training in a very different way for example here we have 10 hours spent at threshold but maybe in the previous series it was like way more I don't know and from this chart we have also this kind of um the whole power duration model with like the peak power output of the athlete during this period but each dot below is an interval performed by the rider so as you can see here it's mostly I think the 2040s uh that he performed. We have like some V2 max over there intervals um in red. And what is interesting as well is if you compare this chart to uh something similar from the previous years, you can of see like for example if you worked on specific durations and where you kind of like um the direction you took for your training plan.
Uh so it's quite also a tool they use to understand if for example uh you did not um train at all specific for example 3 minutes you have no intervals and nothing at this time range or you do not train at a specific intensity. So for example if you have some nothing between 400 and 500 watts on this because that's been a difficult question to answer is somebody and even looking at their data is somebody really poor at two-minute efforts or have they just done very few two-minute efforts? Exactly.
That's takes a lot of time to answer a very simple question if you're to throw all the data. Yeah. And also there's a point is quite important and that's how we started with this uh with the pros is like if you take a pro and you just have the load management chart.
You can have the same values over the last three years let's say. Yeah. But maybe it has been built up very differently.
Like if in a previous team it was like focusing on a specific type of training and a new team very different but in the end maybe the load the final load is the same but just the weight has been added up is very different and so that's how we use this technology and that's how we still use with the pros as well is when they scout someone to kind of like understand how they train in the past and where they are and maybe if there are some improvements to perform um in some specific areas. So if you look at kind of coaching 101, one of the fundamental principles is this idea of specificity. We look at what the target event is.
We try to reverse engineer and deconstruct the demands of the target event and then build that into training plans. Yeah. Do you think this is going to be a more useful tool for understanding if we go the pro analogy of you know a tour of Flanders deconstructing what the demands of that are and then making sure we're ticking all those boxes in training.
Definitely. And for example, if you talk about the race you mentioned, uh the first component that is quite um heavily influential at the moment is the fact uh the well the durability component um in a race. So being able to perform the same watt uh at the end of the race and so that's why we integrated uh when you have a look at your peak power output chart to see what you did in a session but also what you did after 20 30 and even 50 kilogjles per kilogram so that you can see if you are able to sustain the same power data uh after a lot of fatigue and so trying to understand the race dynamics and so the race needs uh so for example if Like you can see that he's not able to sustain the same watts after yeah 50 kilogjles per kilogram.
But you know that that for this race the end is going to be after this threshold that you need to improve in this component for this specific duration for example and trying to better understand uh that way. Really good really good. Durability has become one of those hot topics everyone talking about in the last couple of years.
Another one of them is torque. more coaches are starting to prescribe intervals not on heart rate, not on power, but on torque. Is there an ability to create sessions on torque and how do we look to analyze the results of those sessions?
So, we have when we detect the intervals, we can detect them based on the power in intensity, but also based on uh the cadence. So for example, if you have a high torque uh low cadence interval, you're going to be it's going to be detected over there and you're going to be able to analyze it that way. But it's also pretty easy to say like for example um if we create a session to say like directly in the description, I want to build a high talk session with like four times 8 minutes threshold with high talk.
It's going to build up the session for you and you can have it in just a few clicks and it will specify torque readings for that session as well. Yeah, we have it for example as well um when you because I still don't fully understand what the torque I know Dan Lurang gave a a sample of what a torque session would be like. But so many of us don't have torque as a data field on our computers at the moment that it's like it still feels a little bit foreign.
At least you have it on vector if you want to see it. uh say you have it over there uh to have the talk data so that you have a torque values but also when you compare the session and also um when you analyze it in the chat uh I don't think we have it for this example but if you have some neuromuscular intervals uh we also show the talk value because it's important for us to kind of like explain this term as well. I seen you said chat there.
Is that a chat interface between you and the platform? Like can I ask it questions about hey find me a intervvel session where I done 2x20 threshold in the last not yet. The vector agent is something uh on the plan.
Uh so I'm pretty sure you're all aware about agents in the AI space. Uh but it's something yeah at the moment the chat is something where we have the technology that just summarized the rides. So like with the key data, the intervals data, the sets and a few backs of uh comparison, the coach and the athlete so that the three together can interact and uh the coach can bring way more than just summarizing the intervals for the athlete.
And the next part of it is of course to have an agent and being able to say, "Oh, Vector, can how does this session compare to once two weeks ago or what should I do to improve or do you have some feedback to tell me?" Uh this is the next step. Of course, in AI integration, we have three long form podcasts every week.
Some of them are 90 minutes or two hours long. And I realize you're busy. Sometimes it's hard to cut through the noise and figure out the signal.
Figure out what the important parts are of these conversations. So, I've taken your feedback and once a week I'm delivering to your inbox a newsletter called the Saturday Spin. It's a 4-minute read or less and it's going to break down the very best thing I've learned on the podcast this week.
So, it's insights from the biggest names in the world who are physiologists, sports psychologists, and coaches. It's completely free to sign up to the newsletter. All you have to do is go to rommancycling.
com, stick in your email address, and once a week on a Saturday morning, I will deliver the Saturday spin newsletter direct to your inbox. The link is down below. Go sign up.
But that was super. Really interesting look inside the platform. It's similar to what we're used to.
like it's not a wild change in terms of user interface, but the stuff powering it is quite different and it is going to take a little bit of time to get our head around a lot of the new terminology. I know your John Wakefields, your Dan Langs, they've been talking in this kind of paralons for years now, but most Joe athletes there's an adaptation phase especially I think you know whether you're using Training Peaks or you're using intervals.icu ICU or any of these other platforms, the one that almost everyone is using is Strava.
And so we're taking our files are going from our hammerhead to our training peaks, but also into our Straa and Straa has a totally different set of terms as well. So is that something you're worried about internally that there's going to be a almost an education piece to onboarding? Definitely.
Uh this is kind of the key focus for Vector at the moment education. how we can teach the terms that we use, the new performance metrics that we use and also the all the training analogies that we have. Uh so yeah, education is the key part um and a key focus at the moment.
But also when you mentioned that the user experience it was important for us to not change too many things at the same time because if we want to change the terms change a few things it was important to just use similar calendar calendar that are used to have the training library that they used to have so that at least they kind of find some similarities with what they use at the moment the AI agent especially if you get it on an app on your phone like I'm just trying to think when I was on the Irish track team you would finish like a flying 3k effort as soon as you finish the effort effort you're downloading your file on and then you're looking back and seeing how did that compare to historic like how cool would it be to be able to just pull that out send a voice not be like hey how did that compare to my historic efforts in flying three and it gives you all that data definitely that's uh kind of what we dream of uh but also something like I don't know in the morning you wake up you have a recovery score that is not that good oh big should I adapt something today like and kind of like interact with your data and to have an agent that personalize it to really your uh world. Okay. Yeah.
it becomes almost like we we spoke on the phone a couple of weeks ago and one of the projects I tried super early back in 2014 2015 I think it was it was a coach called pocket coach where it was artificial intelligence machine learning a lot of stuff you've you know managed to successfully roll out now but that was the idea was a coach in your pocket and that's become very real with AI and the way we almost confide in these if you played around and trained up your own my GPT model like the more vulnerable and real you are with it, the better information you get back from it, the more useful information you get back from it. Yeah, definitely. Uh but I think the market is not there yet.
Like I haven't seen any product like at the moment the market is more like in using AI in like all those AI adapted train plans where just like take using it to personalize and adapt it in a little bit but the models doesn't have access to data. So it's kind of like I think the next step in the market you have those AI models having access to your data to being able to personalize and hopefully with your integration with Hexus and other things you we're going to get out with this siloed world where my Whoop recovery data sits over here my cycling data sits here my food data sits here how I feel sits here then we've super sapiens which came along and they've gone bust but that was another data set we've core which is another data set it's What do I do with all this? Like do I go easier or harder today?
Like I have a clue. Yeah, that's the main concern with all these new data streams coming in is like we have lots of apps to use and also lots of data that do not uh we don't understand the interaction between them. So that's why it needs to have like one software that just collects everything kind of do the mixup and understand the insights and the patterns uh between all of these.
you use something called digital twinning with the Simon Yates and some of the other riders. Talk us through what that actually is. So, it's something more for race predictions.
Uh trying to being able to pace more the efforts but also to adapt the strategy based on like what are the the capacities of an athlete today. Uh so it was like more for the taller friends to understand what the athlete can do based on his latest data, recovery data and everything and trying to inform decision makers uh with these um predictions. So yeah being able to predict for example for a climb stage what uh power data is going to be able to perform if the strategy is this or if the strategy is a bit different.
So for example the fatigue at the end of the last climb is going to be a bit higher. should we adapt and kind of like understand and maximize what's the best um strategy for winning a stage. So the data is informant tactics.
Yeah, definitely. And this is I think in war towards the next step. It's kind of like having access to all the training racing data, model it and kind of like um yeah build up different race strategies to optimize it and find the best one.
So, how do you model? Maybe this isn't an answer I'm have the architecture to comprehend, but how do you go about modeling Simon Yates's performance on a hilltop finishing of Wela in the, you know, two days left in a grand tour where he's already raced for two and a half weeks where temperature is different than what he's been training in andor fatigue is different, bunch dynamics. How do we account for all that in a model?
basically with the all the historical data uh that we have available for him because this is something that we really personalize and uh just use his data. So the advantage with like more uh experienced athletes uh more experienced rider is that we have lots of data for them for the past few years. Uh they had they performed and did like a few grand tours.
So we can use this data to understand like how uh they perform after two uh two weeks of racing. Uh so these are all kind of like inputs to the model to predict the output which is uh the power thing. So instead of this that example being the Vela which is the final grand tour of the season let's now say it's the Jurro which is the first Grand Tour of the season.
Is there uh predicted accuracy on the model? Like you can say we have a 90% confidence because it's the VA and we've already seen your performance in Grand Tours this year versus our confidence level that this is the pacing strategies drop to 60% because we actually haven't gotten much data in this calendar year. Yeah, definitely all goes back to the way you measure like the accuracy of the model and the validity validity of a machine learning model.
But um yeah, definitely if you have less data, the accuracy is going to be less of course. And do you think the digital twinning has an application to predict when riders are going to get sick? Um we're not too much in this for the moment.
Um poly moment digital twinning is more like to understand yeah race dynamics and if you pace differently how it's going to impact the performance and how you can basically finish a stage in um the minimum time required. Yeah. So it's like u yeah understanding pacing and also pacing as a group.
So understanding the dynamics if you have u full team of riders uh understand how long they should uh keep like um pulling at uh at the front of the pulit and for how long and what's the base for them. If we take pacing on the last climb to its maybe not as natural conclusion but to one of the possible conclusions is you pace this super wrong and you blow. Now you've got a level of fatigue which you didn't appreciate which rolls into the next day.
Y how do we start to understand the difference between overreaching in a training phase and overtrained? These are complex questions that maybe I'm not the best one to answer again uh more sport science I would say rather than mathematical modeling. Uh but yeah, these are good concepts and hopefully we can help model and understand it a bit uh better.
Yeah, I because I can see definitely applications. You heard of Kitman Labs? You know what they are?
Yeah, of course. So Kitman Labs, for anyone that doesn't know, they model and try and figure out when riders are going to get injured. Actually an Irish company.
Yeah. So they started out with Lster Rugby Team, which is a local rugby team here. and they looked at baselining a bunch of different like your I I don't exactly know what the parameters were, but let's pick some arbitrary ones.
Maybe they measured your speed over a 10 m dash, your vertical jump height, stride length, your gate analysis. They take hundreds of data points and then they look for deviations from those data points to predict your likelihood that you're going to get injured, that you're overtrained. It almost seems like you're building such a powerful data set that we can start to see with like the digital twinning.
Like Simon Yates has gone super well, but the last two times he's gone super well with back to like the 4020 that you can pick out from the data. Hey, you performed this interval now, which also performed it two weeks ago. That pattern recognition, which AI is just so good at pattern recognition.
Hey, you're going super well, but last time you went super well with this sort of power profile, you got sick four days later. So now we suggest pulling back on your training a little bit. Yeah, definitely.
Um, but it's all comes back, I think, to data. And the example you mentioned with kit times is like in team sports they collect lots of you know the injurous data the sickness data and something that we not too much have in cycling actually because everything is power and training data files and we do not have that many uh manual input data to train models on uh but probably it's going to be the future where we have uh this kind of data available and being able to include that and build models on top of that. Um but yeah, I think that's uh one issue with cycling is like we already have so much data available that we kind of like uh do not really care about other data sources that might be useful to build models on.
Yeah, I'm just I'm worried about those AI hallucinations as well of like even using chatp you get them every now and then. You're doing a podcast intro and it's like hey I interviewed Maria last week but you forgot Maria's surname. Can you recall from my Google Docs who that was and it writes an intro for someone that you've never talked to before and you're like I don't think I talked to this person.
Yeah, these are kind of issues you need to be careful of when you build them. I would say Eddie Dumbar, big fan of Eddie, Irish man. Had him on the show a bunch of times.
He's always great value. He's hilarious but also super talented bike rider. He talked about his power dropping in the hotter conditions last year in the Vela where he super successful.
I think he won two stages in Valta last year. He used vecta in a run into this. He he said it's a game changer.
How were you able to model that power drop? You're using core sensors I assume. No, not at all.
Um it was actually when we used the power prediction models for climb stages. So we are able to predict like for example oh today it's going to be 400 watts for this climb based on um the modeling that we have for the first part of the race. But with machine learning model most of people people say sometimes that this is a black box and it's kind of impossible to understand why and why there is this prediction in weights higher or lower than in another case.
But there are few techniques that are able to explain why machine learning model is predicting this value. And so with this kind of approach we are able to measure and explain for example in this kind of race conditions you're going to save 10 watts or you're going to be 10 watts lower. And so that's how we kind of measured the impact of temperature elevation on his performance to try to quantify it and determine the stages that were most optimal for him but also to try to optimize training before the races.
Uh so trying to tell him like for example in after this value in temperature you have a big drop. It's not linear. It's really personalized.
Uh we have lots of examples with riders like just temperature doesn't have any impact on them. Yeah. But like for some of them above 28° is like huge drop.
So it's trying to understand that so that they have a number on it and so they can adapt it. It's G like I can't exactly remember what stages 81 last year but both them were solo. Yeah.
You have a huge pacing component to being solo. You had a pretty big role if you went to those stages last year. Yeah.
Yeah. And also in the what is important is also on the nutrition part where we had a few predictions on like in terms of kilogjles kilogjles he would expense for each stage and so being able to model that and model you know the the rate of consumption of those kilogjles over the the stages. So talking to how does that one work?
So it's a individualized kilogjle recommendation. Yeah, based on his data from all the races data he has uh just like build a machine learning model and predict it based on the race um features and this is something that's going to be rolled out as part of the the consumer platform within vector not yet uh because I think the goal with the collaborations that we have with the pros is just like being at the forefront of what we can do with data for performance but not everything is designed to just being applied to everyone. So that's why we try uh and we're going to integrate um a car recommendation tool within Vector, but it's not going to be as precise as what we use for the pros because they don't have the same needs and also it requires like lots of data and that's kind of the issue for the everyday riders.
So what happens now when Eddie comes back for nationals and carves me up? I'm like he's got an unfair advantage here. You're going to have to hook me up.
I need my model. This is the same with equipment. uh maybe you can buy equipment but like data is becoming like unfair advantage and so that's why teams are kind of pushing for more data science within it is interesting because UCI have a rule that all equipment used by the pros has to be commercially available to purchase yeah for consumers but data is doesn't have that rule yet probably uh I'm not an UCI expert uh but they most of the teams now have a data scientist internally so all the teams have kind of start to have more and more data modeling, some prediction tools, uh trying to understand like also the competition, uh how the teams are built in terms of scouting but also in terms of race um squad and um so yeah, data is becoming like crucial.
Yeah, I was talking to Estanis head coach Vasilus and he was saying behind the scenes the clamor has been for data scientists. There's a huge rush to try and get the best data scientists into world tour teams now. Yeah.
Uh we've seen a few shifts between data scientist moving team recently as well. Uh it's kind of like there's the big salaries are going to be coming for data scientists. Yeah, I think so.
In world it's who would have thought data scientists were going to be cool. Good question. But also I think the complexity with data science is like they're very good data scientists but they need to understand um cycling data.
Yeah. they need to understand like sport science and try being able to model in data science models um cycling and also sports psychology and as well I had a friend who a huge advocate of AI I actually coach him and he's doing a time trial last night a 10mi time trial and he sent me back his proposed pacing schedule for the time trial based off analysis on chat GPT where he taken a bunch of his files upload them into chat GPT given the GPX X file the chat GPT and said model me based on today's weather conditions my perfect pacing strategy gave him a very very good pacing strategy and he's like came back he's like would you make any changes to that or is that perfect I was like the only change I'd potentially make is to slightly negative split it and that maybe makes no sense in the data you have to have ridden a time trial and blown and known the psychological cost of dipping below Oh, your target to really understand how painful that can be. It feels like a slow march to the death.
Especially he had he was turned into headwind conditions for the last couple of kilometers. It's like if you're slightly underpace into a headwind. It's just soul destroying and it can become this, you know, momentum that just gathers and now instead of being 10 watts under schedule, you're like, "Oh, this is hopeless.
I'm not even close to the podium." And you're 20 watts under schedule. It's still I think you need that kind of confluence of data sports and a little bit of almost empathy or you know EQ around being an athlete.
Yeah, definitely. And um I think it goes in the same direction as why we want to keep human coach at the center of the platform within vector as well is like psychology and performance so huge and motivation part as well uh that it's important to have this kind of like yeah what human can bring that AI is never going to emotional roller coaster. Exactly.
I'm just I'm picturing me sitting on the side of the road because I've sent these voice messages to coaches over the years where like I can't do it. I just I just can't hit the numbers. But now I'm confiding in a chat GPT model on the side of the road.
I'm telling Vexa. I'm sitting in the shower wearing my full kit sending Vexa a voice message. I just couldn't hit the numbers.
Do you see this as replacing Training Peaks totally or are teams you're working with using Training Peaks along with Vector? At the moment, it depends. I think it's more a contract side um contract issue but at the moment the main goal for us is to yeah replace rank pics because we think we can we do better on all and everything.
Uh so yeah it's replacing this tool that has been there for years but is not in a position where they can improve the platform and something they kind of like uh has been the same over the past few years. Zooming out a little bit, where do you think outside of the Vector platform, where do you think AI has a potential to impact cycling? Huh?
That's um outside of training I think it's um no I think it's seeing like yeah performance as a whole and having like you know as we discussed before all these integrations where training nutrition sleep recovery psychology lifestyle even though it's kind of like combined within one platform and where um AI is going to bring you something from all of these data sources. So it's like yeah seeing like merge and having a combination of all these different u uh yeah parts that have an impact on performance. Have you seen these things that little black box beside you there?
Have you seen them? Gimly I think it's in there open and see is in that is a real time coefficient of frontal drag sense. So you put it on the bike and it give you real time analysis where that starts getting kind of interesting.
And I haven't played around with it enough to fully understand it. But yeah, you can do arrow test and you can figure out this helmet's faster, these shoes are faster. But sitting on this side of the wheel is slightly faster than sitting on this side of the wheel.
You can start to figure out efficiency within the bunch in it. We're gathering so much data every way. And like if you just analyze it that way and with only this data, maybe you're missing a part of the puzzle.
It's kind of like that's why it needs to be everything merged together. What should an AI be doing? What should it stay out of?
What do you mean? Like where should AI not intersect? Like you mentioned that you don't think AI replaces the coach.
It's like almost a coaching assistant if I'm interpreting you correctly. Is there other areas you think that we need to preserve? We shouldn't have AI impacting this part of cycling.
It's example for me would be years ago the great grand tour winner of Venenzo Nebi wanted to see power meters banned. He thought they took the romance out of cycling that Chris Froom would just get to the bottom of a climb, would ignore race situations around them, ride A to B as fast as he could up the climb regardless of what everyone else around them was doing. I can see the argument on both sides for that.
Is there other areas you think need to be ring fenced and AI shouldn't be involved? I think I'm more on the the side of pushing for AI for everything uh because this is what we do and what we like. Um no, but I think it's um going back to psychology where AI is never I don't want to say never but it's going to be always hard to model it in models and always hard to collect enough data to being able to pre to be precise on this side.
Uh so yeah the psychology side or intuition that a coach can have just by seeing someone seeing the face of this person and maybe in the morning you you say that you feel good but you see that the person is not um that good. So yeah this is never going to be replaced and uh the human side is always going to be more powerful. It's really interesting you say that.
I came back from a stage race last weekend. This the first time I've ever been director on a team for a stage race and you could see by the guys at breakfast who was on a good day. Yeah, the eye is just a little bit more sunk in the head.
The glaze that just kind of fixated in one spot and didn't shift as much as it should be. The lack of interaction. The thousand subtle gestures you make speak volumes of the riders freshness, attitude, willingness to put themselves into the heart locker that day.
Yeah. So these are components that's always going to be like you know you need to see them to adapt to it. Are you seeing resistance from any athletes or coaches around integration of AI?
Um I don't think from athletes actually because they want to perform better and just like some tools to perform better are always helpful. Um, and from coaches, I think it depends on the I would say I'm French, so I would say that the French are not we're not pushing for it like we're not data fans. It's like kind of interesting.
Uh, for example, if you compare it to a team like Jula where we've been working with them since quite a long time now, where the Australian culture is way more uh they like data, they love it and want to use it while the French are like more careful and it's like maybe it's not for us for the moment. So, uh, yeah, I think it all goes back to culture. Yeah, it's the almost we have that stereotype of France and Italian as just passionate and are you working with the catalon?
Uh, not yet. So, I have Steven Barrett, the head of performance for the Catathalon coming on the podcast later in the year. So it'll be interesting to talk to him about their resistance if he has resistance to use or if it's just contractual or what's the I don't think he has resistance on No, Stephen's a pretty Yeah.
Steven's a pretty data driven. He's also an academic. Yeah.
So he appreciates that innovation peer-reviewed definitely researchled. Have you seen situations where the AI is making recommendation A and the coach is making recommendation B? No, I don't think so.
Uh because we we're not doing too much recommendations at the moment. So I haven't faced um this part yet. But it has been interesting sometimes to just like show a prediction and what you can see that the coach like doesn't believe in it or like he doesn't understand.
But then you kind of like go in the same route. what I explained before with shop values and kind of explainable AI and explain why the model is predicting this value and directly you have this uh buying on understanding that oh yeah that makes sense now because the model spot something that the the code didn't spot so it's kind of like I think that that is our main challenge and um being able to explain what the model is doing should we be worried about this AI divide it like where it's dumb bar has access to this pacing thing that I don't have access to or is this just naturally the way sports evolve that some teams are innovating faster than others? I think it's the way it evolves and it's also the way we've we wanted to build the company like for example uh big fan of Nike and the way they released the carbon shoes with Kipogi it's like at the beginning it's only for the pros and then when you run a Martin everyone has carbon shoes uh so I think it's the same way with technology it's kind of like developing tools for them for the pros that they use in racing and training and then that we are able to apply to everyone finishing up here recalling your road map from last night.
You have some interesting parts in it. It almost seems like if we talked for another two hours, we'd come up with another 50 things to add to your road map. What's interesting you most or what are you most excited about on the road map?
I think the Victa agent uh because this is like as mathematicians, the AI uh find the the way the model is uh the way the market is going and what is available now. It's kind of like so impressive the way how fast it goes and so like just to have the first like you know agent that is applying to training and performance it's like yeah what excites us the most I would say. Is there anything not on your road map secret stuff that you want to tell us about or anything in the background?
Um, no. I think what we are quite excited about is like, yeah, we do cycling at the moment, but the goal is of course to go like to running, triath, and kind of like being able to to apply the same concepts to different sports. Cycling is always easier because with power data and everything is kind of like um you have lots of data while in other sports sometimes the sensors aren't developed yet.
Uh, so it's like yeah, being able to apply the same thing to everyone. Have you started thinking about strength training? Because kind of my thesis on this at the moment as I'm evolving as an athlete is that world tour guys and rightly so they're optimizing for performance.
Y they don't care really about anything else only performance. For a long time amateurs we just mirror what the world tour guys do. But definitely I'm having a realization that I shouldn't be optimizing for performance.
I should be optimizing for health. Downstream of health might be performance but not to the same level as the world tour guys. And a part of that framework to get me health has to be strength training.
The data is quite clear on that. So strength training load quantification has been historically very difficult in training peaks. Like the day yesterday I raced a crit.
The day before that I had a heavy strength session. I don't know what to put in training stress score for that strength session. Now I'm going into the crit with heavy legs from squatting and deadlifting the day before.
But it's not measured. It's not captured. Yeah.
And this is like the trickiest part when you start to integrate different sports. It's like Yeah. How do you compare and like how's the the load comparing between different sports?
And also the the issue with strength training is like you do not have any power meter or like heart rate don't care about it. Uh so how do you measure it? So I've seen like a few devices that just try to quantify a bit more like strength sessions.
uh but it's not we're not there yet and uh I think it's uh still way to go and I guess another one if when you're moving into that multisport area and again the user case B more me where I'm optimizing for health I'm often thrown in other sports so I'll go for a run maybe once a week a run if I'm to capture it using TSS that's assuming there's a onetoone transfer relationship And there's definitely not like a run makes me fitter on the bike. It fatigues me slightly more than a similar s bike session, but I don't know if it contributes to my bike fitness in the same. So how do we weigh that relationship between running and yeah and um so I think it all goes back to research.
um where having such a big data set with like so many athletes who train differently and so maybe you try to incorporate like a few running sessions in your schedule and trying to understand if it improves a lot. Um the goal also with vector is to have kind of those research projects where we can learn from your past experiences and kind of like being able to then apply to someone else. Uh so you have within those um data data sources that you have understanding the way it impacted your performance and kind of like move forward the science uh research on that.
And so being able to move like from you know the classic research uh way of doing it with like papers you have group A, group B and you measure like oh yeah by doing this plan you improve by X% performance. trying to have it directly with like the data that we collect and trying to being even more personalized and detect patterns that haven't been found yet. Yeah.
Because that's always the problem with these big peer-reviewed studies like I don't care how the sample group of 200 people absorb the run. I care about how I absorb the run. Exactly.
So it's trying to Yeah. personalize this science part and um research part understand you as an as a human. How do you perform and what's make you a better human?
It's really interesting because there's activities that you would weigh up differently. Like if Sarah asks me, "Do I want to go climbing this evening?" Yeah, I want to go climbing, but I've no more data on this.
Only last time I went climbing, I didn't feel brilliant on the bike two days later. Yeah. But if I had a bit more concrete data on that to say this is massively detrimental, then it's like, h can we go for a hike instead of go climbing?
Yeah. And so that's why it's quite useful to have those, you know, um questions that we ask every morning. How do you feel?
How did you sleep? What's your stress? To be able to quantify those kind of like um new things that you try and tell you, for example, oh yeah, the last time you you climbed um you felt bad the next morning, so maybe forget about it and do something else.
I was trying to tell you, yeah, some stuff that you were not like able to detect. It's a fascinating glimpse into the future of sport. Paul, thanks for making this trip over.
Really appreciate it. Thanks a lot. And congratulations on the platform.
Yeah, only the beginning. Thanks for watching Roadman. Take one second.
Click up here to subscribe to the channel and click down here where there's another video I know you'll love. See you next day.