Written by Christopher Kelly
Nov. 22, 2023
Chris:
Hello, and welcome to the Nourish Balance Thrive podcast. My name is Christopher Kelly. Paul Arson is a coach and scientist known for his exercise physiology and high performance sports expertise. He holds a PhD in exercise physiology and has a teaching, research, and coaching background. Paul has published over a 140 peer reviewed papers, has more than 14,000 citations and is cofounder of the parent company, Hit Science.
His work often focuses on strategies to optimize performance, including high intensity interval training, heat acclimation, and the role of technology in sports. Paul is also a lifelong athlete having completed 18 Ironman triathlons. Paul is on my podcast today to talk about Athletica, a personalized AI coaching platform that adapts to your current sport, fitness levels, goals, training sessions, and life. We discuss why a coach or athlete might use Athletica, including individualized step by step daily training plans, wearable data input, and comprehensive session analysis. Paul also explains how an AI coach can help combat workout boredom, injury, and scheduling challenges.
Ex well, Paul, tell us about life in Canada.
Paul:
Oh, life in Canada is good right now. We are into November, and those transition seasons between summer and winter. Still getting a few rides in. Just got a a mountain bike ride in the other day, but starting to do a lot more swimming and sauna These days end as a big warm up for the ski season because I got snow all around me. I can kinda just see it up on the hills.
I'm getting pretty excited.
Chris:
Ex Yeah. British Columbia is a nice place to be any time of the year, but I'm particularly jealous during the winter because it is one of the best places in the world for skiing and snowboarding.
Paul:
It is. And and in particular, the little town that I'm in, which is called Revelstoke. So check that out on the map. Although Mhmm. If you a local Heard me telling you that, I'd be shocked.
They like to they like to keep it quiet here, but that that is is pretty sweet.
Chris:
I recently got into surfing, and pill. The culture there is kinda the same way. Like, if I could tell you about this point break, but then I'd have to kill you. So
Paul:
Exactly. It's a bit like that.
Chris:
Do you have lots of triathlete athlete friends in locally that like, what do they do? Like, they have all this incredible snow around them. Do they just hang it up for the season and Discuss skiing, or do they keep doing their triathlon training during the winter?
Paul:
Yeah. It's a real mix. So you'll have the diehard triathletes that They're just not into the winter thing, but they do happen to live here, and they'll just go into their trainers and these sorts of things. And then you have The alternate athletes where they do just move into the cross country skiing or even some downhill, and, let us do some alternate ex type activities while the whilst the snow is here. In my experience, both methods kinda work as well.
So Jan Van Berkel, Andy Buscher, Those are athletes that are top 10 athletes in the world. In in terms of Ironman, they're the same. They'll move into those winter sports. And Even though they're doing these alternate things and just kind of doing the combo of the cross country skiing and the cycling and a little bit of treadmill running, they can really pick back up and ex Still conquer in in the big big events later on
Chris:
this summer. Give us an update on Athletica. You've talked about Athletica on the podcast previously. I just got back from the UK where I was lucky enough to go to Brianna Stubb's wedding in Oxford, and she's been on the podcast. No.
I'm not kidding. Yeah. She's been on the podcast several times before. For people who don't know Brianna, she's a ketogenic diet and exogenous ketone researcher. She's now at the Buck Institute for Research on Aging, And she's been using Athletica.
She's really happy with it.
Paul:
Yep. Yeah. She is. So, yeah, Brianna was one of our original beta users and testers, ex And we've had so many people like this that have sort of paved the way for others. And, I mean, the great story with Brianna is where she kinda bounced around from coach to coach.
Ex And, of course, Brianna's background is as a high performance rower, world champion actually in a couple I think one of the one of the events ex In rowing. And, yes, so she's got a great engine on her, but she just kinda couldn't get the swim bike run together ex And came in on Athletica, long story short, and she couldn't believe it when she that Athletica use was associated with her qualifying ex for Kona, and then she used it to perform very well in Kona as well. And we've had many other stories like that. So for the listener, Athletica is like an AI coach that I've been working on for, ex? Really for 10 years now.
And it's not an easy thing to do, but basically, yeah, just you you plug in your race state, connect your wearable like a Garmin or a Strava, ex. And, it draws a line mathematically between today's date, who you are, and the algorithms optimize your training on a day to day kinda basis. So it doesn't matter what you do. You do more load or less load, well, then the whole system kinda resets and on you kinda go towards your event.
Chris:
Ex what sports do you support now?
Paul:
We do 4 sports, sports of triathlon, running, ex Cycling and duathlon now.
Chris:
Okay. And what can an athlete do now? They have Athletica that they couldn't do before. Well, let's say I had Strava and TrainingPeaks before Yeah. And maybe I'd even worked with a coach, but I'm not super wealthy.
You know? I was never one of those people that could afford to pay a coach. Yeah. Yeah. I mean, I'm gonna date myself now, but back in the day, it was sort of $300 a month to have somebody look at the data after you'd done the workout.
And If you just wanted someone just to create a training plan for you and then send you the calendar, they'd only talk to you maybe once a month via email, then that was a lot cheaper. And so, say, for an athlete like that was just having a coach create a training plan for them, the cheaper one, what can they do now? They've got Athletica that they couldn't do before?
Paul:
Well, yeah, it's quite a few things. So where do we wanna start? So I've just kinda gone through and and said, so we're we're gonna put a an ideal plan together you on a day to day basis from now until your race, whenever your race, you're gonna do. There's a bunch of other tools that are gonna assist you along the way. The first one, let's start really simple, but it's your power or your pace ex profile.
And you will get that if you're in Strava or or on TrainingPeaks. And you'll basically, first of all, have a look at how ex Your sprinting or anaerobic power relative to your aerobic power. Right? It's that curve you're kind of looking at. Now that's really important because that individualizes your training.
Right? So we use that individual power or pace profile To kinda calculate your zones. Right? So your training zones, that's and a and a good coach will do that as well too. So but, of course, you're not paying $500 a month for this, you're you're paying $20 a month or less if you're going on an annual schedule.
Chris:
Say it used to be 300, but it's probably $30,000 a month now with inflation.
Paul:
Possibly. Exactly. It would be. So it's a cost effective approach. That's for sure.
So, yeah, that's the first sort of step ex is after you've connected that. And just, again, to take you through the onboarding process, as soon as you onboard, it it downloads, at this point, the last 6 weeks of data if you're on Garmin or ex And, boom, you've got your almost like that blueprint or fingerprint of who you are sits in sort of sits in front. You do some self selecting as well in terms of how ex volume or training you want to do. We have a new feature that's in beta also called our user time constraints. So many of us are you're pigeonholed with your work, and you can only have so much ex time to actually in any given day or you can't train on this day, so you can put those user time constraints in the settings as well.
And then you'll have your plan, and it'll be mapped out. But, of course, it changes one of
Chris:
the a calendar. Right? Like, I've seen it. You've got some really nice videos on YouTube I can link to, and you can sex see that process that you just described, and it seems like the end of that process is you auto populate what looks like a calendar to me.
Paul:
That's right. That's right. So you'll have most of us are familiar with the calendars, whether on any of these platforms. So we have something similar, And your calendar will be populated with the sessions. Of course, things change.
Things get in the way on a daily basis. Ex That's the
Chris:
problem, isn't it? If you just pay for the cookie cutter training plan, you now have to go, oh, I'm gotta go travel for work. Wasn't expecting this the last time we spoke. How do I adjust my plan?
Paul:
Yep. And remember, Chris, the last time you had me on, I talked about science and application of high intensity interval training. So this all kinda went back ex To the book that we first wrote, then we built a course called HIT Science, and now Athletica is the technology arm. Right? So yeah.
So one of the sayings that we have in HIT Science is context rules over content. So we have we've just always known this, and that's ex Basically, like your daily situation, unfortunately, even though you we can have all these cookie cutter perfect plans in place. But at the end of the day, your context is going to rule. Right? Your daughter or your son all of a sudden is homesick and you've got to take care of them, Chris, well, that's gonna potentially ruin some sessions and you've gotta move some things Over and whatnot.
Right? So so we have to have that. So all the sessions are movable and then automatically updated. As well, we have this cool feature called the workout wizard. And this was a very challenging your tech guys, you'd appreciate the technical challenges this one.
But, basically, there's a a lot of programs that are out there. Like, Let's even call them out, right, like trainer road. Right? Thousands of different training sessions that you could potentially have on there. But how do you actually know and how do you actually go and choose 1?
We sort of give you the best options on those. So you press your workout wizard ultimately, and you'll have, like, different options for your changing context. Ex So you want a longer session or a shorter session. We have there's a duration tab. If you are bored of the same old cookie cutter session that ex is giving you.
We go on diversity. You can click on the diversity tab, and you've got a you change it up, but you're still gonna get the same physiological response.
Chris:
I see.
Paul:
And then what else do we have? We have an intensity one as well. So if you want something that's more you're feeling very energized or you're feeling flat, You can switch that up
Chris:
as well. Okay.
Paul:
And then last but not least is an injury tab. These happen as well. So you can have, like might have to switch up the modality ex And moving from, say, for example, a running session to a cycling session or if you're feeling a niggle or something kinda coming on, meditation, yoga, all these sorts ex thing. So Workout Wizard is a brilliant feature. And, again, technologically, it's been very challenging to move those pieces in there that ex To find something that can be catered to you because, again, if you go on one of these alternate platforms with a 1,000 different options, How do you choose?
Right? So we we do the choosing
Chris:
of our courses. It gets worse the more options you give me.
Paul:
Exactly. Exactly. Right? Like, it's that data overload sort of thing. Ex.
So that's another really cool feature that you don't see too many other places to my knowledge.
Chris:
Just before we move on from that, Have you been able to integrate wearables? So for example, I recently started wearing a WHOOP strap just for fun. And it's just ex A WHOOP Strap. A WHOOP. Yeah.
Yeah. Yeah. Yeah. And I'm very impressed by it. Like, the fact that they can take what looks to me like very reasonable heart rate data from my wrist.
Yeah. And the activity detection seems really good. Like, so for example, yesterday, I went for a mountain bike ride, and not only did it get the heart rate data right, It also detected that I'd started a mountain bike ride without me telling it, and it said it was a mountain bike ride, not a row bike ride. That's pretty impressive. And ex.
Just the the quality of the data alone, and, obviously, it measures heart rate variability and sleep diary. I'm just wondering Whether you could integrate any of that into Athletica and then have that be an input where it automatically updates the plan based on those inputs, or is that I I don't know. What do you think about these wearables?
Paul:
Yeah. Well, that's really I mean, that's where it's all going. Right? And this is what we saw 10 years ago, big with the emergence of all these different sensors. And there's I I don't know what the number is.
It's like I think it's it's ridiculous. Like, it's, ex 300,000,000 or something like that in different sensors and stuff. It's it's a crazy number. And so clearly, there's this data overload thing, and that's ex 1 that's we when we saw, that's when we knew we had to make this athletic type program because we need to be able to weed out ex The sensors and the data that that isn't meaningful, we have to and that's so so, yeah, there's all these various different things, body battery, HRV.
Chris:
Ex oh, it's like this thing's packed full of proprietary algorithms I don't really pay too much attention to. But so what do you think is meaningful? Is HRV meaningful for guiding training for Athletica?
Paul:
Yeah. I think it is. And, this is the problem. Like, it's interesting you're mentioning this one. We haven't we don't have anything Yeah.
But this is the fun process that we are working on as we speak. Just prior to this conversation, I was actually, like, ex? I'm going back and forth with the team, talking through all these various different things. So we have our external variables So are our external load markers, and these are, like, our movement speeds. Even things like accelerometry that you just kinda mentioned as well when the WHOOP device Was picking up your your mountain bike.
So that's one sort of factor out there. So anything GPS, accelerometry, power meter, These are things that are happening outside of us. So that's 1 division that we have to keep a lookout in. And I'm stealing this off one of the papers that I'm reading actually in sports medicine as a result of ex Just to like, going down this one. So and then the the next one of the internal training load measures, so things like your heart rate, ex Blood lactate, blood glucose, and these are so super sapiens, say for example, they're they definitely have a blood glucose monitor, but they've been talking for a long time now about ex The new blood lactate marker that's coming in as well.
So we'll see this emerging on in again, within the next year. I think we're gonna start to see this, at least first signs of it. It's gonna be fascinating as well. You know your physiology. Like, it is gonna be incredible how this is gonna change the game on so many different levels, Especially if you're low carb athletes.
I know if you remember the MOLOC faster study and the lactate response in that. I mean, lactate's kind of a fuel, but Anyways, we don't have to go there, but it's like the that's another internal load marker, and there's a bunch of other ones. But then also kind of we need to move to subjective markers as ex well. So was our mind or our brain how is that interpreting the systemic situation in your body? We have to trust that one.
And the research suggests that last one is in fact the most important one. So you're Yeah. Maybe shouldn't surprise us too much, but That's probably the one we need to actually weight highest. Now at the moment with Athletica, we actually do already weight that one the highest, and that's why we haven't gone down The path of integrating HIV just yet. We always knew that in our gut.
So at the moment, Athletica, both athlete and coach version, We have, like, a readiness kind of algorithm that looks at rating a perceived exertion. Feel, of course, you can have a session that is hard. You might rate hard, Like an interval session, say, for example, and it feels crap versus feels like it feels really good. Right? And the 2 mean completely different things.
And then we have good we use Google semantic analysis as well for this for your subjective comments
Chris:
Right.
Paul:
Just like a coach would. And now Now also integrating ChatGPT, OpenAI, interpretation of that feedback as well now, which is pretty fascinating. So, ex Yeah. That's a long winded way of talking about the load response variable. So in HIT Science, we classify that.
A load is the wearables that are like your power, your well, internal would be heart rate. But, like, the load response, the load is like, How are you actually responding to that load? And that's what we've just sort of talked about now. It's like so what's the what's my ex Heart rate in response to to power output, say, for example,
Chris:
or what's
Paul:
my HRV in response to that. So
Chris:
So for cyclists, they've definitely got a power meter. Is that required, Or can you go on heart rate data alone?
Paul:
Well, it's not required, but it is definitely improved. For the cyclist, in the context of the cyclist, Yeah. We definitely you're gonna get a way better response if you've got your own power meter that's semi calibrated and fairly reliable. There's some not as good ones out there on the market.
Chris:
But ex okay. Yeah. I got tired of using my status meter because it was just like so many Bluetooth hiccups, like firmware updates dates. So they're not connected and stuff, and at some point, I say, like, screw that. I don't know why don't we're using a power base.
This is just like life is too short. Uh-huh. Tell me about the speaking of life is too short, ex. What is the intervention? So I see the training plan.
It looks great in a web browser, but, obviously, I'm not gonna be looking at my laptop and the web browser when I'm trying to do one of these workouts. So How do you export the workout so that I can see it on my bike or on foot or even in the pool is hardest still. Right?
Paul:
Yep. So the best experience is to use your Garmin device, and, yeah, there's a setting in Athletica where you you can export Automatically the session to Garmin. So that they push to Garmin and then I'll just when I'm ready to do my workout, I'm just move over to my ex cycling workout, and there it is. And I just follow the instructions on my watch. And then even at the very end, it asks me for my rating perceived exertion and ex feel on the session.
So everything's just almost via the watch there. And I can put my comments in later on. But yeah. So it's like the watch almost ex me. I was doing a a short interval 30:30 session the other day on my watch, and it was mint.
I just loved, the experience actually from it.
Chris:
So No. That's great. So the days of me, ex like, scrubbing down things on little pieces of paper and then sellotaping it to my handlebars that are, like, are over. You is what you're telling me.
Paul:
Well, they they don't have to be over, but there's improved methods out there, I think. Yeah. I I quite like it. Again, it just it feels like your coach is right there with you Yeah. A little bit.
Chris:
I have mixed feelings about these devices. I've worn an Apple Watch in the past, and I felt like I didn't really need any more screens in my life, especially not connected to my body. Like, it's just so like, once it's there, it's so tempting to start using it. Whereas, it's one of the things I like about the WHOOP device is it's so noninvasive. He put it on.
There's nothing to look at. You don't think about it again. Like, you forget that you're wearing it. Whereas the Apple Watch is like I don't know. You're just, like, constantly being reminded and pinged and, like, thinking about what it might XP tracking and Yeah.
Notification hell and all that. It's like, I'm not sure which I prefer, the, like, the paper method or the screed method.
Paul:
I I completely relate. Like you, I'm I'm in a tech world. Right? So it's like you almost want your exercise sometimes to be to be just Well, it's just me in the road, right, or me in the forest. And I guess it would be in terms of how I operate is If I'm doing, like, an easy l two, sub l two session, I probably won't I won't want to be guided on that or anything.
I don't need to be. But when I do a HIIT session, which I know is important, that's nice actually because I lose track of my 32nd period sometimes. I'm so sort of focused on the output, and I want someone else to kind of tell me when to turn on and turn off if I wanna set to that. So I think it's I do a little bit of a hybrid version to to what you're but I, like, totally hear what you're saying.
Chris:
What's coming next is mixed reality. Right? Like, I think athletes are gonna be wearing glasses, and all this stuff is gonna be like a heads up display. Like, that's coming this year, I think.
Paul:
Well, that's already happening with the form goggles if we look to the swimming and triathlon. Right? And they've got accelerometers in the goggles themselves now, or so you can actually look expert. Yeah. I was in the high.
Your head position, which improves your, I guess, the ability of your legs to kinda come up. So They've got and from everything from pace, like, you can actually real time display of your moving pace in the water, your real time display of your head angle, Your heart rate, it's quite impressive, but you're right. The total head display thing. And, again, ex Depends what you want. Right?
Like, personally, I don't think I knee I don't need that during my riding. At least, I don't think I do now or my running, but I can see the benefit for certain sessions intermittently with swimming based on what I've seen with the form.
Chris:
I could imagine it being helpful to have ex intersession feedback on how things are going. Like, if it could say to me, yeah, that interval was great, but the one before it, not so much. Could you do the next one more like the last one and not before did you see what I'm saying? I don't know. Is that too much feedback?
Like, you think that's too the granularity is too fine there.
Paul:
I I mean, it could be Depends on
Chris:
the athlete. Right? Like, so that's the promise of AI is that you can figure out what works best for the athlete.
Paul:
I think so. I think that's that is what it's all about here. And there's just so much tinkering and experimentation going on. Right? And it's like, back to People like Brianna.
Right? Like, the she's not afraid to go and try these sort of new things. And there's the Mhmm. You see this the there's some Some sort of thing with the whole early in early, mid, and late adopter profile and stuff. Right?
So Mhmm. But, yeah, there's a lot of those thank goodness for the early adopters experts because we would be nowhere without them. Right? We're starting to move in now towards, I don't know, between we're between the Early adopter and the mid adopter phase right now with Athletico, which is quite exciting for us because we're just seeing our numbers numbers go up. We're seeing lots of the the slow twitch ex Forum threads, and people are starting to talk about this.
I'm getting invited to conferences and whatnot talking about the future of coaching and these things. So it's Yeah.
Chris:
Well, Let's talk about that. You sent me a link to a podcast I'd not heard of before, Fast Talk Laboratories, and I listened to that episode. What impact will AI have on training software? And it seems to me like it was a podcast for coaches. Would you agree with that?
Paul:
Oh, yeah. Totally. Yeah.
Chris:
And Yeah. I I know I thought it was quite a hostile interview. I'll I'll hopefully make this conversation such that you don't have to listen to this episode to get something out of this conversation, sex. I would recommend people go and listen to it. We could spend the next 2 hours just going through it sentence by sentence.
Like, it was I I didn't agree with a a lot of what they said, but It seemed to me like I'm noticing this shift and interviews Ken Ford on the podcast before. He knows Brie for sure. Ken is the director of the Institute for human and machine cognition, and he talks about AI. He's been doing it for decades. And he said for most of my career, people said, well, you just can't do it.
This can't be done. And then suddenly, they were saying, you you shouldn't do it. Right? Like, oh, that's interesting. You can't really simultaneously hold those 2 ideas in your head at the same time.
Right? And I feel like I'm sensing that shift in that podcast that went from, no, Paul. This is way too complicated. You could never replicate. You could never automate what I do as a coach to, Yeah.
I'm not sure you should do this for some other, like, really fuzzy reason, but is that the type of response you feel like you're getting from the coaching world right now?
Paul:
Well, you don't see the naysayers so much except when you read the forum threads. So, yeah, I guess I get a little bit on both, but I'm more seeing because the the only emails ex messages you get in where I'm sitting are the ones that are saying, oh, I definitely see it going that way. Yeah. I wanna join kinda thing. So our coaching platform is starting to We're getting a lot more coaches onto the coaching platform now as they're sort of in that phase of shifting over from training peaks usually because that's the the gold standard.
Right? So. Or or tell us about
Chris:
the coaching platform. I wonder if the same thing is going to to happen to coaches as has happened already to coders, software engineers like myself. And actually, going back to Ken as well, he prefers the acronym amplified intelligence or augmented intelligence rather than artificial intelligence. And for me as a coder, as I'm sure every coder listening will know, that some really interesting generative AI came along a couple of years ago. I mean, for me, personally, I feel like my productivity is maybe an order of magnitude more than what it was.
And it's not like the AI couldn't really replace me, but I sure as hell spend a lot less time looking up the fine points of things I already know. I used to find spend a lot of time digging around in source code to find out how a library works or I'd be searching for answers on Stack Overflow. Whereas now I could just write a natural language comment and GitHub Copilot pilot will autocomplete that function. And it looks about right, and I tested them like, yeah, I'll tweak this and this, but it's like a super fancy autocomplete. And I wonder if the same thing is now happening for coaches and that you've already just delivered this to them where they can augment their existing skill.
Paul:
Ex Yeah. Exactly. Well, we just leverage all the same features that I've just almost kind of predescribed in the podcast. So, basically and think of the time that we're saving coaches. Expert.
You no longer have to insert your, like, your 8 week plan or your 10 week plan or your 16 week plan. You just plug in the race date and get your athlete to connect their wearable, And you have an infinite plan created for that athlete. And, again, we leverage all these other experiences for the user as well, including the workout wizard. There's one other I actually haven't even spoken about, but it's really it's super cool and helpful. You should actually check it out, Chris.
It's called our workout reserve. Now if you've seen in the sports and AI world, Andrea Zignoli, he works for both us and Supersapiens, and he's just an absolute genius. And he created this this feature. The it's most analogous to a a body battery in terms of and it, basically, it It looks at the power or speed data for your last 6 weeks, and it compares today's session instantly in terms of how close you are to that reserve. And it really helps with the whole principle of of training progression.
Right? So you when that's when your session goes from he we we currently have it, like, on a scale from 100 to 0%. And then if it goes below 0, it means ultimately you've had a breakthrough workout ex or a personal best in any maximal mean power or maximum mean pace relative to that last 6 weeks. So So you can imagine the power of that, right, in terms of it preparing you optimally for the next like, we know that training consistency, for example, expert is one of the most important, if not the most important factor that's going to go to your success and your adaptation as an athlete. So If we are keeping an eye on our workout reserve for any workout, for any sequence of intensity and duration, It's just a wonderful one to kind of look at on a day to day basis.
So that's available for athletes and coaches. There's a readiness sort of factor We're actually looking on a day to day basis in terms of what that session is and what that athlete is ranking their feel and RPE. You remember, we We spoke on the subjective markers. So Mhmm. There's a coach and there's a comment section and whatnot that that's seamless and right right through ex Yeah.
I think it's and then there yeah. There's a a library session as well there. So the coaches are gonna have their own sessions that are gonna be it's their own secret sauce or it's close to they might be running a master's program and they wanna have their master swimming sessions in there so they can click and drag those in, whatever it may be. So it's very versatile for the coach.
Chris:
Couldn't you do an experiment here? Like, if the coaches are not sure whether the software will be better than them or better than them with the software, couldn't you just pick a bunch of objective endpoints and maybe the subjective life experience too, and then randomized athletes into maybe 2 or more groups and then Mhmm. Figure out which works best. Right? Like, that that seems doable to me.
Paul:
Ex? You totally could. I will say there's a fair bit of body of research that is suggesting that the most effective Method is the coach that's able to leverage the technology. Like, that's that tends to be the best method. So The
Chris:
Centaur approach.
Paul:
Yeah. Exactly. Because it like the Athletica or Name Your Platform, it should serve as an assistant coach, ex Technology. A service enabler, ultimately. But food coaches are are really important too because they appreciate that context.
Ex. They can and they can do other things that maybe the machine necessarily couldn't do. Right? Like, we spoke about the empathy and all these things, and it's Chat, GPT is getting there, but I think there's still room for a really good human emotional coach to to take on those sorts of things, help with planning, ex All these other various things. I use Aflodica for my coaching, and I'm loving it now.
Chris:
Well, I'm glad you brought this up. I listened to that interview, and I wondered, did they record the interview with you and then record other excerpts with other coaches and athletes and then splice those into ex so you never heard what they said. Right?
Paul:
No. I wasn't aware of any of the other controversies. So I listened to that one back because it was very interesting. So, yeah, yeah, it was quite interesting to hear expert. The negativity, resistance, you were bringing this up.
Right? And and
Chris:
I thought that was Interesting. You you might call that an absentee hatchet job. Right? If you'd have been there to respond to them, that would have been one thing, but splicing it in afterwards where you cannot possibly respond, I I feel is not quite fair. I thought that all of the comments they got from athletes and coaches were either like straw men ex or they're, like, we're basically trying to smuggle in the concept of a soul.
I mean, you just said that, okay, so maybe empathy is important, But are you trying to tell me that those coaches are always empathetic? There's a an assumption there that I'm not sure is true because I've done this for a living, like being empathetic on the phone all day long or on email or any other method of communication is utterly exhausting, and most coaches are not gonna do it. And it's going to depend on the coach's own situation. Right? Like, my kid is sick, and now you're asking me questions about what workout to do whilst you're away on for work.
Honestly, I don't give a crap. Like, I'm just gonna, like, try and get rid of you as quickly as possible. Like and that's just very human, whereas maybe a large language model. Recently, I've been reading again Steven Rolnick's work on motivational interviewing, and motivational interviewing is a communication style that's been studied extensively in the literature for a number of different indications including smoking cessation and drug addiction and all kinds of really hard problems, but also coaching performance as well. And when I understand that communication style, it seems possible to have a large language model detect ex when it's appropriate to use, say, a listening statement versus an affirmation and then generate 1 in response in a way that's likely to get a better response than a human alone would.
And I think we're already seeing this with people using large language models It's like for AI girlfriends. Like, that's a thing now that why is that? Yeah. Absolutely. And it's because I think for this reason, talking to an always on, Always understanding, always empathetic AI assistant may be better for you than talking to your coach or whoever else is around you.
Expert. Mhmm. Plus and minus what your social network looks like. But do you see my point is, like, saying, oh, well, empathy, so you need a real coach. I don't think that's a very good excuse when clearly it could be also automated with a large language model.
But what do you think?
Paul:
Well, I think you're right. And especially, Again, having been in this business for a long time, you made this point. There's a lot of coaches out there that really lack empathy and they lack those great ex Interpersonal skills. And that might not work with your own personality. Right?
So it's easy to to believe that ex Just what you said, the large language models could be a better match for you as an athlete compared to a A coach with a wrong personality.
Chris:
Yeah. I mean, that's another really good point is maybe for some athletes, motivational interviewing works really well, ex but for others, maybe more of a a drill sergeant approach works. Right? Like, you asked me why I didn't do my interval training workout on Saturday. And I say, yeah.
Fuck intervals on Saturday. Saturday is when my buddies are off work, and I just wanna go ride my bike. And one style of you coming back to that is, like, the motivational interviewing style where you're rolling with the resistance and you're using listening statements and affirmations and open questions and all of that good stuff. And then another style might be the drill sergeant, like, What the fuck, Chris? You promised me that if I coached you, you would put in a 100%.
This is not the deal. You can fuck off like there are other people waiting to take you. Like, Maybe that approach works better. Right? And I imagine an an automated software system could do experiments and, like, just figure out which is this style that works best and use that.
Whereas, like, does a coach really have time to do all these different experiments on, like, such an individualized level? Like, I doubt it. Ex
Paul:
Probably not. No. But I don't know. It's crazy where this is all going now or could go. But, yeah, you're you're absolutely right.
So, ex I mean, it's just like, look at where we are now compared to where we were. Right? Like, the speed. Yeah. Think of this year.
Right? When I think, like
Chris:
Oh, it's insane.
Paul:
Chat GPT was was released. Look at where we are after 11 months. Right? It's it's
Chris:
It's absolutely bonkers. It reminds me because I was around in the late nineties when the web came to the fore, and it I it's just insane. I just I mean, it feels similar in the in the Bay Area. The tech companies are laying off tens, hundreds of thousands of employees. At the same time, this new step function change in technology is is happening, and it's almost impossible to keep up with the rate of change.
So tell us about your use of Chat GPT in Athletica. That was actually one of my complaints around that podcast interview was making inappropriate comparisons between Athletica and ChatGPT. They are not the same thing. Right? Like, I understand that ex.
You may be using the OpenAI GPT models inside of Athletica, but that doesn't mean that you can go to Chat GPT ex and get the same experience. Right? Like, I get that you can go to Chatt GPT and say, hey. I'm racing cyclocross this autumn. Can you come up with a training plan for me, and it will regurgitate what looks like a coherent plan?
But that is not what Athletica is, and This is almost true of large language models, autoregressive large language models. They're just predicting the next word in the sequence. Right? They're not goal orientated like Athletica is. Right?
You've just given Athletica a goal and then Athletica will work towards that goal. That is not what Chatt GPT does at all. But tell us about how you are using ChatGPT inside of Athletica.
Paul:
Yeah. You described it perfectly. So Athletica is kind of it sits Completely separate away from ChatGPT. And we only use ChatGPT in the communication factor, so in the large language sorta model. We just use it as a a response to your subjective comments.
We also use it in a session analysis ex sort of way. So we found a way to get it to understand numbers and to actually to look at how well your session was executed relative to what was prescribed and what your response was in that. It's still in beta. It's all about the prompt, The massaging and getting the right prompt. Right?
Just like anyone that uses chat g p GPT currently, we're having the same battle in terms of getting that prompt just right. But that's how we do it. And, basically, use some coding. It's not secrets, but the methods and whatnot to get those prompts Just right in there in in accordance with the context of the session.
Chris:
Yeah. Absolutely. I'm very sympathetic to that. I've been doing a bunch of prompt engineering recently, and it it seems like the problem has just moved. Right?
The problem used to be, how do I code the thing? And now the problem is, well, how do I prompt the model to code the thing? But you still got the problem of engineering. Right? Like, it's not driven.
Paul:
Yeah. And, well, I mean, like, internally, a lot of our feedback is that we shouldn't be necessarily using this chat gpt for some of the things that we're looking at. But it's fun to tinker and experiment, And and we are and we certainly are in that experimentation stage, which I guess is what everyone is doing right now, which is why it's evolving so rapidly. This time with recording the podcast, I don't have a clear answer, and I don't exactly know what the final version of the release will be. Yeah.
We're seeing about a 90% 80, 90 ex Good response, but the beta users are, most re rating it 9 out of 10 and finding it quite amazing. We probably will release next week. So it'll be available if likely, one of the per the person the listeners is listening to this. You can check it out and get your own session analyzed and tell us what you think. But, ex Yeah.
Yeah. That's where we're at.
Chris:
I think you've got this problem where you may get a thumbs up or thumbs down, but perhaps the person who's consuming that content, They don't actually know what good content looks like. Right? They don't how how they know really yeah. Exactly. So I've been thinking about this a lot recently.
I started doing jiu jitsu with the kids, and I watch the way that the coaches interact with the kids and then compare that to what I've learned from Rolnick's work. Ex. And it's like one of those things like bad coaching when you see it, but you can't articulate why that it's bad. And One powerful construct for motivational interviewing is the idea of an affirmation, and it's really just another way of saying process over outcome. I learned this from our behavior scientist, doctor Simon Marshall.
It's like Yeah. Concentrate on the process, and you can almost forget about the outcome. Ex and the the jiu jitsu coach will say to the kid, oh, good job. And if if I was a cocky kid at 7 years, I'd say, oh, really? Why is that?
And they wouldn't be able to tell you because they don't actually know what it is that they liked. Right? Like, they're just saying good job because it's easy, but why was it good? Like, it's not useful feedback unless you can tell me why it's good. If they'd said to to the kid, oh, well, I really love the way that you wrap the neck there.
And the way that you posted out your leg like this as that whatever, like specific details about exactly what they did that was good, sex, then that's helpful feedback, whereas just saying, oh, good job, like, not so much, like, that's not feedback I can use. You probably get into a different problem here with Athletica sex is maybe you don't know enough about the process to be able to say what it was about the process that was good. Like, maybe this is where the wearables come in. I don't know. Do you have enough, like, vision to know exactly what the person did well?
Paul:
Ex Yeah. Let's say I'm just actually thinking on the fly, but I I actually love that. And it kinda relates to that. Remember the conversation about the load response ex factors. HRV, all these things.
And I wonder if that's where these things are gonna emerge is where some of those large language models are gonna kinda come into ex Understanding some of those load response markers as we're already implementing and then coming back to how that relates to the process. Ex Again, it's just a fascinating moment in in our evolution of this whole crazy AI world that we're in.
Chris:
Ex It is. Yeah. I mean, the other thing that came up for me in the interview was this idea that, well, only a person could do this.
Paul:
Ex I'm
Chris:
like, oh, really? Like, they use the cats versus dogs example. Now there's, like, extraordinary accuracy in image classification that's just ridiculous. It's like better than humans. And
Paul:
I know it's come a long ways since the example that was given.
Chris:
I know. And so the way that they made progress was creating a model whose architecture kind of resembled how the brain worked. Right? That was a deep neural net.
Paul:
Ex.
Chris:
Yeah. And what was really interesting, Geoff Hinton has been talking about this recently, Geoff Hinton's widely regarded to be the father of deep learning, is that ex in trying to mimic the way that the brain does that type of classification, they came up with an architecture that seems to encode knowledge more efficiently than the brain does. Right? If you look at ChatGPT, what it does right now, do you know another human that can, ex Like, have as many, like, good opinions on such a wide variety of topics, but then you look at the number of, ex like, weights that are stored in that model compared to the number that are in, like, synapses in the human brain. Like, Chat UPT is way more efficient.
Like, what the heck? Ex Yeah. But going back to my original point, like, the criticism is like, oh, well, this is not cats versus dogs, but he never said that it was. Like, Athletica is not a general ex artificial intelligence. Right?
It's just really good at answering the question, what kind of workout should I do today? That's a very different problem from All the different problems that Paul Larsen has to solve in his day to day life as a human. Right? Like, it's like Yeah. It again, it's like not a fair comparison, but I didn't really have any thoughts about that.
Paul:
No. And I I guess that just comes from the there's this cloud that sort of sits. So a negative cloud that sits over AI, and I think it's that was clouding the The interview, and that was almost the Yeah. The stance where maybe, you know, Trevor and Rob were kind of coming in a little bit. There's there's some I and the them Speaking to me off air there, yeah, there's a fear.
There's a fear on anytime there's a threat and these guys are really embedded in the whole coaching world. Right? So coaches are ex starting to feel that that fear, right, just like your developers were or are. So, yeah, it's interesting. It's an interesting time.
Chris:
Ex It is. Yeah. I feel like at some point I mean, at the moment, it's like I'm more productive using generative AI as a coder, expert. Probably, at some point, I'm not gonna be needing a door. Right?
Like, somebody will be able to describe the problem. Maybe that's what I'll do then. Right? Like, that's probably the history of technology is people, like being made redundant and then they just go and do something else instead. Like, once upon a time, people used to plow fields with an ox.
Right? And then it'll you know, then tractor came along and then Everybody had to go get different jobs, but no. At this point, like, nobody's saying, oh, man. I wish I could go plow a field with an ox. Right?
Like, you just gotta do something different instead. But I'm sure at the time, it was, like, stressful for all those people that were put out of work by, you know, the steam donkey. Like, that's ex
Paul:
It is exactly that. Right? Yeah. Easy to say, but that is exactly what we're going through right now.
Chris:
But for the coaches right now, the technology is not there yet. So you would predict that coaches that are using Athletica I mean, you're already using ChatGPT inside it. That's gonna give feedback in text form to the athlete, but that's not going to replace a coach in their entirety? Like, that's No. Still going to be absolutely valuing?
Paul:
Ex Yeah. Like, again, the root of or, like, at our heart I'm I'm an educator. Right? I'm a professor. I love to educate.
So Mhmm. What I would like to do like, this is the next step for us in Athletica. On the Chatt GPT integration is educating the coaches as well because I think the coaches have lots ex still to offer their athletes. So I want a version of the chat gpt to analyze help ex analyze the athlete session and explain to the coach also what we're seeing, what Afladeca is seeing, and ex For them to be better off for the process as well. So that's that speaks to just who I am.
That's my desire. I don't wanna That's why we have a coach version and we have an equal amount of coaches and and athletes on the platform. So and this stems from our HIT Science roots. Oh we have a very large a large following from the book and the course and all that sort of stuff. And and yeah we want to help those coaches be better coaches ex And and yes so that's Athletica is a spin off from that and we want to keep that going too.
But I think there's we have we also recognize there's 2 distinct markets ex too. Right? And there needs to be a a language a certain language used for athletes and there's, at least most athletes. And then there's a a certain language that's Potentially used more for coaches, we can actually change the temperature on those languages as well. It's kinda cool too, right, in terms of how technical you want to be ex Or how simple you want to be, say, for example.
Chris:
Yeah. You're just making me think that usually people, persuade them like, the individual athletes, you're probably gonna persuade them with stories. Let let me tell you about this athlete just like you and Yeah. And the training plan they followed and
Paul:
I like it.
Chris:
The come they got. Whereas the coaches are probably going to be I'm guessing, like, there's some variation here as well, but they're about you. Like, they're scientists. They're gonna wanna see evidence. Like, You should look at this 2019 meta analysis of endurance athletes that did this versus that.
So you'd have Some sort of language model that generated citations, some sort of the rag retrieval assisted generation is all the rage now. You have some knowledge base of all the papers that Paul Larson likes, and then you have when you ask a question, then it dips into that knowledge base, then answers the coach's question, but also gives the citation. And this is probably not something that's gonna be in, a large language model by default because it's Paul Larsen's view of the scientific world on endurance training and and and especially high intensity interval training. Right? I'm sure you've read papers you don't agree with.
You'd rather the language model didn't contain. Right?
Paul:
Ex Oh, correct. Excellent. And so kudos to you, Chris. The number of the team members follow your work. And I it was maybe a couple months ago, but you released ex You're releasing your large language models.
I think Huberman was 1 and Malcolm, etcetera.
Chris:
Kendrick.
Paul:
You you had all these various. So of of course, we tend to be
Chris:
more like, try asking ChatGPT what he should eat and it's just gonna give and I've seen this in the Whoop AI coach. I asked that was the first question I asked it. It's like, what should I eat? And it just gave me I I recognize this answer as the vanilla GPT answer. It's just whatever is in the large language model, which has come from reading all of Wikipedia, a whole bunch of web pages, and it says, oh, yeah.
You should eat lean protein and, like, lots of vegetable. I'm like, It's not terrible, but it's not, what I would tell clients, it's not what Ted Nieman talked about the protein leverage hypothesis and satiety per calorie and all of that stuff. I think that's far more and micronutrients, of course, and I find that to be far more valuable than whatever you found on Wikipedia about diet. So So if this were the listener that probably didn't know what the hell I'm talking about, I created a bunch of virtual agents and released them onto our discourse forum. That's what it was.
Yeah. You can tag ex virtual Huberman, which is not my favorite, if I'm honest, virtual Ted, and virtual Malcolm Kendrick. And what happens on the so you ask a question, and then it dips into a knowledge base and then it generates an answer based on and only on what's in the knowledge base. So if you want to know what causes cardiovascular disease, Pretty sure you should ask Malcolm Kendrick about the thrombogenic hypothesis because I find it much more convincing than anything to do with lipids. But, of course, again, that's another really good example something you're not gonna get from Vanilla GPT.
Paul:
Yeah. So, exactly, we did the same thing following
Chris:
the The book museum? Ex The yeah.
Paul:
The book, all the podcasts transcribed, training science podcasts, all the athletic blogs, all the hit science blogs. So it's just, yeah, it's just coming up from our own ring fenced ex language followed by.
Chris:
That's great. That makes me so happy because I listen to a lot of podcasts and quite often, like, either I won't remember what I heard that's ex relevant at the right time, or I just won't remember where I heard it. I listen to so many podcasts. They all and books as well, audiobooks, and they all sort of just melt into one thing. And so I'm like, I know I heard someone talk about this, but goddamn it, which pod I can't even remember which podcast it was, let alone which episode.
And so then when you go back and you're trying to do a Google search, and, of course, that doesn't really help you, especially when you're asking questions about health because all you get is results from bloody the Mayo clinic and health line, whatever. And so having this assistant embedded into my software, I can just ask it a question and it dip into sex. Your book on high intensity interval training and give me the answer, not a list of links that I then have to go look at to see if The answer to my question is on one of those pages. That's, I think, amazing. And that's congratulations on that.
I'm really excited.
Paul:
Well, thank you thank you for the for the prompt as well. So, like, we we're just following you.
Chris:
Well, I mean, what do you think I'm doing? Like, do you think I came up with the idea of RAG? That probably is and it seems like a hack to me as well at some point. I mean, this is you're already seeing this happen with the latest version of GPT 4 has a 128 k context window. So Everything that the prompt and all the chat history and the corpus of text that came for your knowledge base, it all needs to fit inside of this thing they call a context window, and it's quite small.
In the past, they've been as small as 8 or 16,000 tokens, which is largely analogous to words. And so If Paul has written a 700 page tomb on high intensity interval training and you ask a question, I can't put the entire book ex into the context window and have it summarize it. And now the context window is a 128,000 tokens, is it, with the latest version of GPT 4, which is It's getting up there like that's almost a book in terms of the number of words, and
Paul:
so
Chris:
it's getting closer. It's gonna be hella slow for for a while, I think, and so maybe the rag trick will be relevant for a while yet. It'll also be hell of expensive because that's how they bill you. If you're sending these giant prompts, then That's the input tokens is one of the ways they bill you.
Paul:
But we saw the price decrease just the other day as well. Right? So so It's yeah. And that's just gonna keep happening, I think.
Chris:
That's huge. Well, I'm getting a little bit off topic here. Tell people about where
Paul:
expert. I think we're talking about AI and Athletica. Right? So, yeah, it's all I think it's all relevant.
Chris:
Okay. That's good. I'm glad you agree. I hope the listener does too. Yeah.
I really wanted to, like, ex focus on how this software is gonna help athletes. Like that going back to that question, what can I do with Athletica that I couldn't do before? Especially. I'm guessing, like, if you're a self coached athlete, this is definitely gonna be unless you happen to be Joe Freel and you're coaching yourself. Right?
Like, I'm guessing this is gonna be a lot better than you just like coaching yourself.
Paul:
Yeah. Well, I mean, Joe Joe Frills, he's made the call. Right? So and, again, this is on This is with the FastDoc team. Like, he's actually it's his craft for coaching, it's all technology stuff.
And he's basically made the war cry, and he's like, guys, Don't miss the bus on this. Right? Like, this is yeah. The things are moving, and we need to shift. But, yeah, like, just to review all the key features is the automated programming.
Chris:
Right.
Paul:
You basically plug in your race state, and you've got a you've got a plan to that race state that is science based plan. So it's not not not just pulling this out of our ass or anything. So You're gonna basically plug in your wearable to it. So you're gonna see your your overall profile. So it's all individualized to your profile.
The sessions are again, the workout reserve is just as again, you don't see that anywhere else. This is the battery line that I was telling you about. So you, again, remember back expert. The HitScience podcast, no pain no gain, that whole philosophy, that's out. We don't believe in the no pain no gain.
You wanna hit your target, walk away, repeat. Training consistency wins every time over any crazy workout that you might
Chris:
be able to get injured. Right? Like that's I mean, that's saying this a different way of saying the same thing.
Paul:
Ex Exactly. Right? So yeah. So lowering your chance of of getting injured and then the workout wizard as well. It's just because context always rules over the content.
So you've got options on your day in any sort of given day. Yeah. And and again, integration of strength training workouts. We're working with a strength training guru as well. He's got his 500 various different videos.
This is coming in. There's new wearables there that are coming to the I'll leave I might even mention it. So it's One of our partners is Plantiga. I'm not sure if you heard of Plantiga, but, basically, it's a it's a force plate that's in your shoe. Oh, cool.
That's one of our partners that well, they'll be likely ex integrated in the in the future. So imagine it doesn't matter. It could be in your cycling shoe. It could be in a running shoe. Or they could
Chris:
measure power on a bike from something in your shoe.
Paul:
That's right. Exactly. I mean, not just that, but also looking at injury and and imbalances Mhmm. As well. So, like, integrating that as well into the the whole program.
If there's one take home message from the whole podcast, probably, is just the craziness and the possibilities are endless right now because of the 300,000,000 sensors that are out there. You need to figure out who you want to partner with And how are you gonna integrate that new input that's to the model that's kinda coming in there? And is that useful data that's coming in or is it noise? So we need to always just be figuring out the ideal signal from that noise that we want to pay attention to, and AI is gonna help us there do doing that.
Chris:
Awesome. Ex well, athletica dotai is where coaches and athletes can go to find out more about the software. Is there anything else that you'd want people to know about, Paul? Ex
Paul:
No. I I don't think so, Chris. Again, I really appreciate the the opportunity to to speak about it. I really keep up the great work yourself. Our Our whole team follows you, and so keep up the good work there as well.
And, yeah, I I think that's probably done expert now. So, yeah, feel free to reach out to me, paul@athletica.ai, if you wanna have a chat about different things, coaches or athletes. We're here to help
Chris:
Awesome. Thank you so much, Paul. I really appreciate you taking the time. Thank you.
© 2013-2024 nourishbalancethrive