The world of AI is changing extraordinarily fast. We spoke with Dr. Paul Laursen about the current challenges and developments of the technology.
Episode Transcript
Chris Case 00:04
Hey everyone, welcome to another episode of fast talk, your source for the science of endurance performance. I’m your host, Chris case here with Trevor Connor. A year and a half ago, we spoke with Dr Paul Larson on fast talk to learn about AI training software. Since then, countless things have changed with the technology, which speaks to the lightning fast way that AI is developing. Generally, we could likely have this conversation again in six months and talk about entirely new things. That’s not only because of the rapid evolution of AI tools, but also because sports science is moving into a new era. We’ve been in what Dr Larson refers to as the era of sports science 2.0 which is all about the collection of data. Of course, data aggregation is still a focus, but Sports Science 3.0 fuses the collection with the interpretation of data, and this is where AI software can be immensely valuable Dr Larson’s company, athletica.ai is working to bring true artificial intelligence to athletes and coaches to help them with training prescription. The challenges are immense, but so is the excitement for what these tools can do. So on today’s episode, we talk with Dr Larson about the current challenges in this field, what AI training software can and can’t do, and what athletes and coaches can do to optimize the AI experience. In our conversation, we’ll revisit what makes something true AI, the meaning of large language models, the challenge of bad inputs and bad outputs, aka, garbage in, garbage out, and how that can be addressed with something called ring fencing. That’s the technical side. We’ll also dive into what you need to know to take full advantage of these tools, including how to best prompt an AI tool, how both athletes and coaches should use it, how the software can learn from athletes, and what the future might bring. So let’s have a real conversation about artificial intelligence, and let’s make you fast
Dr. Paul Laursen 01:58
so we all know that the fastest way to stop progressing is to get injured, and that’s why managing load matters. Athletica keeps an eye on your training stress, helping you adapt before it’s too much. It’s the same principle we talk about at fast talk labs. Train smarter, not harder. Learn more@athletica.ai and use code fast talk to save 15% off a six month plan. We’ll see you there.
Chris Case 02:25
Dr Larson, welcome back to fast talk.
Dr. Paul Laursen 02:27
Hey, Chris, thanks for having me. Man, it’s great to be here as always.
Chris Case 02:31
Yeah, you’ve been on fast talk before, specifically in Episode 292 to talk about AI training software, its impact on training, and that was about a year and a half ago. And you just mentioned before we started recording how things are moving so fast, even in the last month since we scheduled this podcast, that it’s surreal, that it’s crazy. You used a lot of superlatives there. So in a year and a half, it’s must be a different space altogether.
Dr. Paul Laursen 03:01
Yeah, it really is. I’ve had some profound moments with my team who are incredible at Athletica, who teach me these sorts of things, and we can get all into that. But yeah, it is moving quite quickly, and there’s lots to speak about in the next little bit here, about what’s happened in the space since we last
Trevor Connor 03:19
spoke. And I love that you made the comment of, I’m just a college professor, this AI stuff.
Dr. Paul Laursen 03:24
Well, yeah, I was telling you I’m like, really nervous about this, because I so often feel way out of my depth. You guys know my history. I come from the Sports Science Professor background, and I’ve entered this space of AI ultimately, because I knew I could see the writing on the wall, and we wanted a seat at the table from the ultimately in the AI revolution, from the Sports Science profession. I feel an immense responsibility, if I’m quite honest, in terms of where I sit right now with this whole thing. But I’m excited to speak to you and talk to you all about it again.
Trevor Connor 03:58
So I started my career back in the 90s as a computer programmer, website developer, and back then, like I honestly, was taught machine code in school, which I’m not sure anybody knows anymore, but when we were doing website development, you had to write everything yourself. You just sat there and notepad on your computer and wrote all the code. Right? Nowadays, nobody does that anymore. I don’t think anybody could do that anymore. You just have software that will produce the code for you, and you basically tell it what you want it to do. And the impression I’ve gotten with AI is a lot of this AI is self writing, so we are reaching a point where I’m not sure anybody could say they fully understand it and what it’s doing
Dr. Paul Laursen 04:45
that’s good. So we’re in good company today. Then I guess, yeah,
Chris Case 04:48
we’ll try to make some sense of it. We will try to explain it. We’ll try to talk about how best to use it and what it all means. Where do we even begin? Paul with this discussion like it’s too much to talk about. What is. Happened in the past year and a half. But is there a way for you to set the stage? Give us that brief background, that foundation from, where we can launch from?
Dr. Paul Laursen 05:08
Yeah, well, I mean, just to recap from the last episode, we talked about intelligence. We talked about this artificial intelligence, and chatgpt made a cameo and whatnot with we were talking about language models. You know that it was inevitable, and sure enough, it is inevitable. Here we are now, six months down the road. Yeah, and where do we start with it all in that we’ve come a long way since as well. You know, I know the context of Athletica. Of course, there’s lots of other options that are out there for other people, so I can’t help but always refer to the world that I live in with Athletica and whatnot. So I’m going to apologize about that from the get go. But of course, there’s keep in mind, there’s everyone else is doing their own sort of thing too. But one thing I guess, on that, that we’ve tried to start with from the beginning, and we continue to this day, is that we start with models. We start with exercise physiology models, so we avoid any black box just throwing out the word AI. So everything that athletic is built on hit science, the science and application of high intensity interval training.
Trevor Connor 06:11
So this kind of gets at something that you were writing about. You did a series of reviews about this sports science 3.0 which is this idea of taking foundational exercise science knowledge and combining it with cutting edge technology, which would basically be AI software. Can you tell us a little bit more about that? And it might help if you say first, what sports science 1.0 and 2.0
Dr. Paul Laursen 06:37
were so Sports Science 1.0 is what you and I and all the other listeners that did some form of sports science education in their past. These are foundational models. These are foundational principles. We can think of things like AV hills, critical power, critical velocity curve, ultimately, that he showed in 1925 this is banisters, fitness and fatigue modeling, the PMC chart that you’ll be familiar with in the back of training peaks. These are all the models and whatnot. That’s Sports Science. 1.0 like these are all the sort of foundational principles that we can all learn. 2.0 we describe in this series of papers that Trevor refers to are the proliferation, first and foremost, of the sensor technology. All of a sudden, sensors, you’ll notice maybe from 2005 onwards to 2020 all these sensors came to the fray right from the power meters you use, the heart rate monitors, the SAO two monitors, even more and more are coming out, right? We call that sort of sports science 2.0 and these were lots of interesting feedback inputs that are coming and now what we’re calling Sports Science 3.0 is the integration and leverage of all of these tools and AI to actually make sense out of it all. We argue that you can’t just take the sensor data and get insight into that. It’s got to be leveraged on top of the foundational principles that we all grew up with and learning. And that’s what we try to do, ultimately, with Athletica, and that’s what we are encouraging our others, including our competitors, to make sure that they do as well, but with the whole feeling responsibility to the industry. So that’s what we’re sort of all about right now. And we’ll link, hopefully, to the series of papers that we’re publishing in a related journal, sports performance sports reports. Spsr was led by Martin Bucha, my colleague,
Chris Case 08:35
yes, we will what I
Trevor Connor 08:36
really like that you address, and we’ve talked about this before, the sports science, 2.0 phase was great, because we started, as you said, finding all these tools, finding all these things that we could measure. But in some ways, there wasn’t a lot of thought that was attached to it. It was just you got all this data. Great. Your power is going up, your heart rate’s going up, whatever it happens to be. And you would see a lot of athletes, a lot of coaches, not really do a ton of interpretation with it. And what I liked is, when you described Sports Science 3.0 you said we’re going from evidence based to evidence informed, where you’re going back to those foundational principles, looking at the data and then putting a lot more focus on the interpretation of that data, figuring out what it means. Yeah, absolutely,
Dr. Paul Laursen 09:23
that is exactly what we are doing. And again, I don’t want to, I’m racing to the end my last point, I can’t help it, and that’s that these AI agents that we’re now seeing are actually interpreting much of that information and putting forth in a sports science 3.0 form to inform the user, end user, coach or athlete, on the Insight it’s doing work that I as a coach have a difficult time doing. Would you say that phase two is actually over with, or are there things. Things that are still being developed without forethought as to how they will be integrated with AI and how they will be interpreted.
Trevor Connor 10:08
Thank you. Forethought was the word I was looking for when I was saying that and struggling. Yeah, that’s
Dr. Paul Laursen 10:13
a great question. 2.0 is most definitely still happening. More and more sensors are coming out, so I even on my newsreel this morning, was shown that there was another 100 million that was invested into another blood glucose sensor that’s coming out right and they’re getting smaller and smaller these sorts of things. So no, we’re 100% not done with 2.0 the sensor proliferation will continue. So yeah, they’re gonna get better and better, smaller and smaller, there’s so much tech that’s going into all of those but make no mistake about it, we still have to go reach back and use that data in light of in the context of these models that we’ve learned. And of course, the models will improve too, right? Even the 1.0 is probably going to continue, sure the 3.0 phase is going to be what facilitates that, right? And we can talk more about how we’re trying to do that, even from a research standpoint too, where we’re bringing, hopefully, the laboratory into the field. Now. So
Trevor Connor 11:12
everything I’m hearing of the very simple summary is 1.0 was about the principles and foundation. 2.0 was about the data, and 3.0 is saying we’re going to pull those two together and focus on interpreting that data using the principles
Dr. Paul Laursen 11:30
correct. And we’re stealing a page out of Peter Tia’s outlive book too, right? Where, if anyone’s read that, that’s really what he’s describing the field of medicine. We’re just borrowing from Peter.
Suzy Sanchez 11:41
Hey, coaches. This is Susie Sanchez from USA Cycling. We’ve partnered with fast talk laboratories to upgrade our coach education program. By the end of this year, USA Cycling will have over 50 new CEU courses for coaches produced by experts from fast talk labs. Visit USA cycling.org backslash coaches to learn more.
Trevor Connor 12:01
So now let’s dive into the part that you said you were scared to talk about, which is AI platforms in the sports science arena. It does seem like they are growing at an incredibly rapid rate. It does seem like there’s more and more new, sophisticated ways of doing this. You brought up something in one of the papers called rag, and I remember scratching my head going, huh? But let’s talk a little bit about that. And I do want to bring up, I think it’s really important to differentiate. I think AI has become a marketing term, and everybody’s trying to say our stuff is AI, where a lot of it is really just traditional software with a lot of yes, no algorithms in it. Not everything that’s claiming to be AI is true. Ai, would that be accurate? Yeah,
Dr. Paul Laursen 12:48
that’s totally accurate. We spoke about this in the first podcast that we did on AI, right? And that is really where we started with a lot of this. Keep in mind, we start with sort of template plans in Athletica, which have their own limitations, our own bias, our own flavor, for lack of a better term, we could be called the hit science flavor of coaching. As you guys know, there’s lots of different ways that we can coach, and different philosophies and whatnot that are out there. We started, certainly with that in a binary, if then format, of course, that evolves over time and gets better and better. It’s a good place to start. So start with those template plans. We start with all of the various different models, and just so the user is aware and just how efficient this is becoming when a user comes into athletic so many of us are used to using the sports science 2.0 phase, where we track our training right and you have on Garmin, say, for example, in the Garmin system, you’ve got two years of historic data that’s probably sitting back in there. Well, as soon as you log in and connect your wearable boom, you’ve got two years of historic data is sitting on Athletica when you log in. And then, of course, we see your critical power, critical pace curve, anaerobic speed and power reserve. Basically, we can have a look at what a sort of a sprinty athlete you are versus a diesel engine. And then, of course, if you plug in your race state in the future, we can draw a line based on that history, knowing how much training volume you’ve done in the past, training intensity that you’ve done in the past, and we can draw a line out to that event and optimize for that training. And don’t forget, this is working in both the athlete context and the coach context as well, right? So that’s sort of the first little bit. And then there’s in terms of new changes to Athletica that we’ve made. It was a bit clunky to begin. We have a new UI. We have a mobile app now, and there’s a new feature that we use, also called workout reserve, that you know, this is not too much AI. It’s more looking at maximum mean powers, maximal mean paces. This is across very short durations to very long windows. But it doesn’t matter. It’s agnostic to. Window, and it’s always looking at a ratio of what you are doing in the moment. This is live on your garmin or live on velocity, and it’s actually looking at, you know, how deep into the well you’ve gone relative to your last six weeks of data. So it’s really giving insight, and that can help with letting you know about hitting a personal best to protecting you from injury, right? So it’s a very helpful coaching tool. So that’s embedded within Athletica. The other new one that’s very handy too, is daily availability or user time constraints, whatever you want to call it. That was a very difficult lift for the team. But this is true AI, because you can basically place in as a working person or whatever, how much time you have available, what exercise mode you want to train for, and then you will still get all the things I just described prior, except everything is optimized. The training is still optimized within those windows of time. So all of that happens. And then the big one kind of that you’re referring to, Trevor, was finally the leveraging of the large language models, because those have really advanced, and those are the ones that are blowing my mind. Trevor mentioned the word rag that stands for retrieval, augmented generation. This is actually an AI that’s looking into a library of historic words, and the book actually sits in our reg system, right? And it’s so basically, this is the thing that blew my mind, and we’re just about to publish this on spsr, but it’s basically we proved based on a question that was put to me as an expert, science backed coach, we compared the question response to me completely blinded, not even knowing what I was doing. Andrea signally is the one that did this, and we compared that response with the reg AI model, using leveraging the book science and application of high intensity level training. And the response, we could actually see the 2d response, was almost identical between how I went around in my five minute spiel on how I would fix these athletes. And basically it gave me a question, and I went around in my head saying, Oh, I’ll probably do this first, and then I’d lower the intensity for this individual and lengthen out the exercise duration. And then for this athlete, they’re short on this so I would give them short intervals, blah, blah, blah, all these sorts of things. We basically log the retrieval Right? Like, in your mind, you’re going back and you’re retrieving all this information that’s all stored in there, of course. Now these rag AI systems are doing the same thing. So it’s basically the 2d plot of how I went around was identical from the rag model to me, not identical, but very close in terms of the mapping. So that’s what’s coming next. And that’s mind blowing to me. My mind was blown when I saw this. I was like, Oh my gosh, here we are. That’s
Trevor Connor 18:04
what I want to ask you about, because that’s where you’re getting into true AI. So going back to what we were talking about before, there’s lots of platforms out there where, just give an example, they’ll bring in some coaches, they’ll ask coaches a bunch of questions, and the coaches will write the answers, and they just build a database of if somebody asks this question, here’s the answer, somebody asked this question, there’s the answer, and you just have this huge database of answers. And so when somebody uses the training platform, it’s just pulling this information that’s been put in, but it’s not creating anything unique. It’s not coming up with novel answers. It’s just using the database of answers that already exist. This isn’t that is my understanding. This is where you have fed in a huge amount of information, but now the AI is actually coming up with novel responses, creating its own answers. Is that correct?
Dr. Paul Laursen 18:59
Yeah, I think the big important differentiator in what you described Trevor relative to what I’m talking about, is I’m talking about a ring fenced situation or model. So the one you’ve just described is going out into the internet, and the interwebs, which are vast, right? And what we can now do, and this is experts and other experts that are out there. They have the ability now, with these rag agents, to create a ring fence of their own proprietary IP. And this is what we’ve done, obviously, with hit science and Athletica, where we’ve ring fenced all the info that’s in the book that’s protected with copyright. So the response that I’ve just described, it’s used my brain based on the book, and it’s mapped a very, very similar response. So what we’re looking at here, and others can do this. You guys can do this if you ring, fence your own proprietary info, and then basically the rag, the retrieval, augmented generation. It’s like a librarian. It’s going back in and it’s pulling all the chunks of information and forming them and putting them into the output for the user based on their query, based on what they’re looking for. That’s step one. Now step two, and this is where we’re going next, and this is what we’re working on, is the writing ability, the moving ability. So now we have these agents. The word agency or agent means working independently. This is a spooky thing, right? Is that these agents now are working independently to go and write and form things, and if formed correctly, help you in terms of moving your program around, to keep you safe in accordance with the optimization and objectives that you’re putting into the system. Can
Chris Case 20:44
I ask a follow up question on step one, which is, you’ve effectively constrained the information that you want the rag to pull from? Are you doing that because you’re certain of that information, it’s from your book. Is that part of it? Is that all of it, or is it also like if you had a true AI system where the internet was its oyster, and it could pull from the entire internet, then if you ask that a question like, how do I train for a marathon that’s in six months, there’s so much information on marathon training that you would get some middle ground that wouldn’t be optimized for the individual,
Dr. Paul Laursen 21:25
absolutely. That’s the problem currently, right? Yeah, and you’ll get a great middle ground training plan or training recommendation,
Chris Case 21:35
but it’s basically cookie cutter, so to speak.
Dr. Paul Laursen 21:37
That’s right, exactly. Yeah, exactly. So it’d probably still work well for the majority of people getting started and stuff, but yeah, this is what we do. This is our specialty. This is our expertise. Is what we spent our careers doing. So this gives us a bit of a business case to provide something of value to
Trevor Connor 21:54
take that a step further. One of the issues, I actually gave a presentation on this at conference a couple of years ago, one of the issues they’re starting to run into with these AI platforms like chatgpt is they don’t really have the ability to differentiate high quality information from lower quality information. They just go and consume everything. And you saw when chatgpt came out, a lot of these sources of higher quality information, a lot of the respected newspapers, online magazines, things like that, started suing, saying you don’t have the right to use our information. So what you are actually seeing is platforms like chatgpt essentially would become dumber because they couldn’t consume the really high quality information, and they were getting more and more of their information from user groups and places where people are just ranting and raving and doing conspiracy theories and stuff like that. And the question is, what is chat GPT or these AI is ultimately going to give you at
Chris Case 22:55
bad data in bad data out in a way, if you put garbage into the machine, then it’s going to spit out stuff that’s not so great, either. So
Trevor Connor 23:02
I hadn’t thought about this, but I like your idea of this rag, this retrieval, augmented generation, where you can say, we want to make sure you’re getting high quality answers, so we’re going to restrict what the AI is going to draw from information that we trust. But my question to you is, if somebody asked it a novel question where there isn’t a direct answer in your original hit science content, does your AI have the intelligence to then say I’m going to figure out the answer, or does it say that wasn’t in the content? So I can’t give you an answer? Pause, let
Chris Case 23:38
me ask Dr Larson, that’s what it says.
Dr. Paul Laursen 23:42
No, it’s a great question. And one, I actually will have a answer for it, an answer for it, and that is that we’re very worried about the medical kind of questions, right? We don’t want to be liable for an incorrect medical query that kind of comes in. So we have a focused system that safeguards against any response to a medical related question, because we’re not a medical system, and we don’t want to offer a medical answer and to get in trouble accordingly. So if it’s out of context, then it won’t provide that answer. That’s not what we’re there for. Are
Chris Case 24:16
we getting into a place where you’re also hesitant to talk about it because it’s proprietary system.
Dr. Paul Laursen 24:22
No, we pride ourselves on being not black box and being open and transparent. So I’ll answer any question that you have. I just might not know the answer. Fortunately, I have a lot of smarter people than me that work for me, but Yeah, happy to try.
Trevor Connor 24:36
So I do then have a couple questions. We talked about rag retrieval, augmented generation, you mentioned large language models. What exactly is that? As
Dr. Paul Laursen 24:45
I understand it, rag is sort of an advanced version of the large language models, right? So large language models, as we spoke about on the previous episode on this, are basically, it’s a model, it’s a predictive text model. It’s predicting the next. Chunk of words ultimately that should go to solve that. So when you’re using chatgpt, if you do that, that’s all you’re getting. It’s not truly thinking per se or reasoning, although it appears that way often, but it’s predicting what should be the response, which is why, if you play with it enough, you’ll find that it doesn’t always deliver the appropriate response for you, and that’s because it can only do what it can do in terms of its model. It’s predicting the next words. That’s all it is.
Speaker 1 25:28
Again, a question that might not be relevant, but since chatgpt is maybe the most popular out there, is it completely wide open? Or are there ways to use that that are reminiscent of this rag type system. Good question.
Dr. Paul Laursen 25:44
It probably is using many forms of rag and retrieval, but I think it’s just like Trevor was saying, it’s just so much more vast. So it’s still going in, like the librarian, and pulling all the info that it can, and the better you prompt it. So this is what we have experts on at Athletica is they are prompt engineering experts, and you, the end user, can be an excellent prompter as well. The better your prompt, the more refined you can get that language model to give you the answer that you’re looking for. And a lot of times, it will take multiple attempts right to get that better prompting and response back, and you can certainly use this to your advantage if you are designing your own training programs or your own response to that. So can be a high value tool for coaches and athletes, coaches and self coached athletes. Yeah,
Chris Case 26:35
that prompt is an interesting component, because I feel like we’ve probably all heard about how the system has sort of intrinsic bias in it, and I feel like the prompt, you can also add more bias to it by using certain words or asking for certain things, and as you just put it, getting the answer you’re looking for
Dr. Paul Laursen 26:54
Absolutely. Let’s segue to the carb versus fat debate, because there’s one that we can all relate to, at least we’ve heard of that, and probably even on fast talk, we’ve seen the different arms of those in the carb camp and the fat camp and everything in between. To your point, Chris, if you’re coming in and you’re prompting for the info and you’re looking for the carb information, that’s all you’re going to get. Conversely, if you’re in the fat camp, you know, low carb, high fat, that’s all you’re going to get. So it will confirm your own bias. Now, if I can segue to what we’re trying to do with sports science, 3.0 we have a new study, actually, that we’re trying to provide data for, and we’re going to hopefully leave it open for maybe a year, but we want to get as many people involved as possible. Basically how it’s going to work is you fill out a questionnaire about your fueling habits. The study is called field. So fueling,
Chris Case 27:51
what is it fueling influence on endurance load and Development? Thank you.
Dr. Paul Laursen 27:56
It’s brand new, so I haven’t quite got it fueling influence on endurance load and development. So that’s exactly what we’re looking at. And you come in, give your fueling insight in terms of the questionnaire, and then download your wearable data, and we’ll compare, ultimately, parameters with that are in your that wearable data, critical, power, tramp, durability markers, and ultimately compare that with your fueling camp or beliefs or practices, these sorts of things. That’s, again, we feel that sort of sports science 3.0 in practice.
Trevor Connor 28:28
And we’ll put a link to this if anybody wants to participate in the show notes for this episode. It does look really interesting. Yeah,
Chris Case 28:35
the website is pretty easy to remember, too. Athletica.ai/field-study,
Trevor Connor 28:42
we’re back to the 3.0 and that gets me to the, I think the really important question that we touched on a year and a half ago, and want to touch on again, because, like we said, in the world of AI, a year and a half is a huge length of time. Are they at the point where I think of those good old fashioned western movies where Dr Paul Larson goes, my job is done here, and rides off into the sunset, and you got nothing left off for the world, because it’s all in the AI. Can
Chris Case 29:06
you do that again? No, okay, I should have tried one well,
Trevor Connor 29:13
but where can it replace what the coach was doing? And where can’t it? What does the coach and the athletes still need to do. So
Dr. Paul Laursen 29:22
let’s talk to the coach’s job. Let’s remember what the AI still can’t do. It can’t be beside you as a person, yet it doesn’t display the real connection and empathy. You can’t be a partner in what it is that you’re trying to do. There’s degrees of it, but it can’t recognize the context to the same degree we don’t have the sports science 2.0 sensor level, at the same level as the coach that pays attention ultimately and making good observations, what it can do is, and this is where coaches can leverage it as well, is it certainly can speed up time in terms of producing. Creating a reasonably good plan relative to the individual’s history, right? Like basic principles of like, Where does someone start? Right, when you onboard an athlete as a coach and say you’re using training peaks or whatever, where do you actually start with that athlete? You can get a better, granular, more focused and accurate prescription of what the next step is for that athlete in the context of not wanting to injure that individual, because principle of progression is sort of right there. So AI is getting very good at that again. This blew me away the other day when I was because obviously I use Athletica, but we were pinpointing my zone. So there’s another one actually like it can come in right away, and it can pinpoint all of your zones. The other cool one that we’ve put in here as well is the recovery in hit science, we say we’ve got load we’ve got training, but then you’ve also got load response. So that’s things like how you’re feeling about training, your fatigue level, your heart rate variability. So if you’ve done any episodes with Marco altini, dan plues, heart rate variability, all the ways that we can analyze that with insight, and that’s typically looking at 60 day rolling average, which is your normal relative to your seven day normal average. That’s all within Athletica. And again, what’s very handy is that you’ve got an AI coach, or an AI assistant coach that’s actually interpreting that data. When Andrea was wanting my zone data, he says, go out and do three minutes all out. So I went out and did three minutes all out on the on a hill, and it came back. And maybe I shouldn’t have been it shouldn’t been profound, but the first thing that the AI coach told me was, congratulations on your three minute all out. PB, so it all of a sudden knew my right from the history, and that’s just the fact that it knew that, and the AI coach can go right in there immediately and analyze all the data. Now, coaches listening, you’re going to go, Well, that’s pretty handy, right? Because it’s hard to sometimes reach back and find those things and then make mention of it. But like, right off the bat, you’ve got that analyzed sitting in front of you. So I think what it’s doing and can do really well is a speed of the analysis, assuming, to Chris’s point, that we’re not on the GIGO, garbage in, garbage out trail. We have data that we can trust, as long as that is dialed, then you can really leverage these systems.
Trevor Connor 32:25
So back in the 90s, when I was just getting into bike racing, I bought this software package that was pretty cutting edge at the time, and this goes back to what I was talking about, just large databases where some coaches had just put in a ton of workouts, and then in the software, you set the date of your target event and what type of event it is, whether it’s so this was all focused on road cycling. So it was a crit or a time trial or a road race, and then it would map out every single day for you. Just it had a big database of workouts, and it just dumped them in and thought this was really amazing at the time. I’m really glad I didn’t follow it too well, because I’m sure it wasn’t the best plan in the world. Yeah, these AI training platforms are a lot more sophisticated than this. But my question to you is, I said I’m glad I didn’t follow that plan that was built in the 90s. How much now can an athlete trust an AI generated training plan? What can they trust? What shouldn’t they trust? Where do you still need interpretation? Or do you is it good enough now to build a training plan that can adjust to the athlete, that they can just go to a platform like Athletica, say, Build me a training plan, and they’re done so
Dr. Paul Laursen 33:41
at this point, still not perfect. It is a very powerful tool used the right way, but you still have to be your own agent in the whole process. But that being said, it’s getting better and better, but common sense still has to prevail on all of these sorts of things. So if you have an injury, say, for example, we don’t have a mechanism within Athletica where we know about that, and then we know what to do about that. That will be coming. But if you have me back another 18 months, that’s probably going to be present. But right now, as we speak, it is not so you still have to use it as a tool. That being said, it’s a very strong tool. It will give you, if you’re a healthy athlete, and you come in and you’re looking to optimize your fitness for your goal event, it’s very strong at doing that. We have an 85% approval rate where, basically we queried our users and it either met or exceeded expectations. 85% coach would be happy with that score, so it’s certainly doing a very good job at this point in the game, but you still have to do a little bit of learning along the way. You need to learn to take things that the AI coach is saying with a grain of salt and realize that it can’t always be perfect yet. So yeah, that’s where we. We’re certainly a lot better than we were 18 months ago when we last spoke. What if any
Chris Case 35:04
psychology component is there in terms of the delivery of the prescription the post workout news that you receive? Take us inside that, if that exists at all. So
Dr. Paul Laursen 35:15
at this very point, as we record that is not available, I just got to thank and explain how this whole thing has worked in terms of how we’ve built it. It’s been built by the users of Athletica, 1000s of users now that have been on the platform, and they have built it, ultimately via feedback, strong feedback on our forum. We listen to that, and we quickly make changes. So one of the very strong feedbacks is the discussion that the users want in that last week, during the race and in the race following. And we’re working on that as we speak right now with our back end team. We’re ultimately leveraging many of the successful behavior that I’ve tried my best to give to my athletes in the past and whatnot, in terms of mentally preparing them for that race, telling them that the most important muscle in their body, ultimately, during the taper phase shifts from those around the body to the one between their two ears and getting into the game and all those sorts of things. So all of that is coming. And then, of course, a good analysis during the race, and then a recovery post. So that’s being developed as we speak, not present right now at this point. So
Trevor Connor 36:25
that raises an interesting question. A good coach learns from each athlete that they coach, and they bring that knowledge to subsequent athletes that they work with. So what they learn from previous athletes is going to benefit future athletes. Is Is that something that you have in Athletica where it’s learning from all the different athletes that it’s working with, and taking that knowledge and applying it to other athletes, and is there ethical concerns with doing something like that?
Dr. Paul Laursen 36:53
Yeah, you should come and work for Athletica, because you just keep bringing up all the things we’re working on. But just to start with the learning issue. This is what’s being developed right now. Is that historical info. So we’re capturing individual historical info. We all come with our own individual Ness, right? So how can we find that sort of the blueprint of the person when they’re onboarding, and then learn about them from the history, history of their thresholds, history of their training, and reach back. And there’s already degrees of that happening with AI coach. Ai coach is already talking about what happened last week, your frequency of your compliance. How compliant are you with the training? Are you doing way more sessions than you’re being prescribed? Are you doing Are you missing a bunch of sessions? There’s a lot of chatter that’s going to ultimately get into the head of the user to keep them on the right path. So I guess that’s a yes, and it will continue to develop. From the ethical standpoint, I’ve been on a lot of ethical committees in the past in from the university and research settings, so we do our very best to try to make sure that everything that we do in Athletica is being considered and being helpful. Privacy is paramount and the health of the athlete is paramount as well. Consideration is being made there, but I guess misinterpretation is always possible. There always will be the 0.1% thing that we will miss, but all we can do is our best.
Trevor Connor 38:22
It is really, to me, a very interesting question where I’m not sure I have an answer for you. I mean, with a coach, it’s impossible to work with athletes and not learn from them, so the ethics are almost a mute point. You’re going to learn from the athletes you work with, and that knowledge you’re going to bring to future athletes. But with an AI tool, that’s something you can turn on or off, you can say you can’t take knowledge that you gained from other athletes and apply it to future athletes or other athletes on the platform. And I’m not sure what the ethics of that are, personally.
Dr. Paul Laursen 38:56
Yeah, I will just say maybe one important point that I missed is that you can always destroy your data. Any learning, I guess, is going to happen right for the platform, per se, but the individual always has the right because there’s a button that you press and all your data will be completely removed and it’s no longer there, sort of thing no longer in our records. So every user that comes on, has that ability, and that’s the same with the field study as well that we spoke on too. It does have IRB approval, and does have some big wig researchers that are involved, mostly, I will admit, from the low carb, high fat group, Tim Noakes, Philip Prins, Andrew kutnik and Jeff voluk and others.
Trevor Connor 39:40
Listeners, fast talk needs your help. We’re planning to make changes to the fast talk podcast, and we want your feedback. Please take our quick survey at fast talk labs.com/survey this survey is brief, and your responses will help shape the future of fast talk visit fast talk labs.com/survey to get started, you. Yeah, so going back to what we were talking about a little bit before, let’s talk about that athlete who doesn’t have a coach and they’re using a platform like Athletica to help them with building their training plan. What’s your advice to them in terms of here’s where you should really trust the tool, and here’s where you need to bring in your own interpretation, like, how do you differentiate when to say the AI knows better and when to say, I know better? I would say 100%
Dr. Paul Laursen 40:29
you make sure you’re skeptical coming in because it really at first, it doesn’t know a thing about you, and that’s the highest probability place where it could get you wrong. Check your data. This happens all the time. Someone will have a situation, let’s call them a runner, and in their history, they’ll have had these sessions where they’ve gone in their car and they’ve finished a run, they haven’t stopped their watch, and they’ve taken off. And of course, that’s the GIGO problem right there, right you got garbage data, and that skews the thresholds and the zone determinations and your critical power, critical pace. So you’ve got to come in as a skeptic first, make sure the data is clean and you don’t have these types of errors. And then once you’re but once you get comfortable with the process, and maybe you do a test week, and you lay down some test results that you are confident with, you should start to see the value of the platform as you go along there. But certainly it’s clunkiest in the very beginning, the first week. But if you can get over that. And for coaches as well, when they go through this process with their athletes on Athletica, once we get over that, it’s a more smooth sailing operation with less hallucinations and athletes getting in a nice groove. Consistence training, we always know consistency is key, and usually get a really good experience.
Chris Case 41:56
Is hallucination a term that is used in the AI world? Can you define that in this context?
Dr. Paul Laursen 42:02
Yeah, well, it’s, it’s ultimately, it goes back to large language model stuff that we talked about, when the AI hallucinates it. Ultimately, the model goes astray. The model is doing it the best it can, but it just goes off track, ultimately, into something that isn’t what you want. Ultimately, as a isn’t what the user wants, but that’s where the model sort of takes us. And that’s called a I describe that as a hallucination. It’s like, oh, that’s wrong. And you’ve got to call the AI on that. And this is, again, back to the fantastic forum users that we have. They call us a lot, and it’s living in my world of Athletica. You got to have a thick skin, because we get hurt, we get whipped a lot, but that’s what makes us better. So
Trevor Connor 42:42
you say, be skeptical, particularly at the beginning. What about after a year, after the platform has learned you a lot better? How is it best for the athlete to interact?
Dr. Paul Laursen 42:52
I’m reflecting on the users that have been with us for up to four or five years, and yeah, they just can’t believe how that it’s getting we continue to bring new insights and releases to the fray. You get more and more comfortable with the tools that are at your disposal for training, and we spoke about those on the last one, the workout wizard is a fantastic one. So click of a button and you can switch up your training in a new context to something that’s very appropriate, that keeps you on track. Workout reserve is phenomenal as well. And I will say I’m thoroughly enjoying my sessions on velocity. I know you had Robbie and the team on that podcast. I love that. We love our work, our partnership with velocity. Just to give a little bit more plug for that, for only $5 more, you can add on a velocity membership. You can basically do your session live with workout reserve in front of you as a metric, so you can actually see how deep into the well you’re going on any one given session. And then you can do the live sessions or the on demand ones. And again, to the coaching aspect that community is so important, right? So you asked the question before, what’s AI good at? Do we still need the coach? Of course, we still do need the coach to run these sorts of engaging sessions, to have these teaching moments. And this is why the partnership is just so strong with that. Because here we have the science that we’ve spoken about coming in there, but then we also have the pedagogy element, the teaching element that we get on velocity too. So the two together are gold.
Trevor Connor 44:22
So I guess the final big question that I personally have for you is now, let’s shift focus to the coach and how the coach should best interact with this tool. And we’ve talked in the past. I believe in that previous episode, we brought up the fact that the tool, the AI, can now handle a lot of that data crunching that the coaches used to be very obsessed with when, particularly in that sports science 2.0 phase that we talked about, and now the coach will really be focusing more on the interaction with the athlete and perhaps the interpretation. But can we dive deeper into what that would mean? What does that look like day to day on the platform? If
Dr. Paul Laursen 45:00
you come onto the Athletica coach version, it’s just a training calendar, and you’ve got all your athletes on the left hand side. The advantage is, typically, when you’re a coach, you’re having to plug and play all the various different sessions. I guess the difference from a programming standpoint, is you’re simply entering in the race dates for the athlete, or you get the athlete to do that. So programming itself is taken care of. Of course, you can use the user time daily availability or user time constraints to actually help that athlete. If they can’t train on a Monday, and Monday is going to be a rest day, just plug that in and that’s all gonna that’s all gonna happen. So that’s a time saving feature. Of course, you can have your own sessions in there, your key sessions. So it’s a time saving sort of thing. This has been said by others, but in terms of having you as a coach, your secret sauce to be your training sessions, I think that sort of needs to be probably rethought and to really have a close look at the sessions that are, that are being derived by AI, they’re probably more optimal when you think about how they are leveraging workout reserve, the critical power curve, what the actual athletes actually done, the historic duration. I should be a decent coach. And then if, if you agree, then work towards developing yourself with the community aspect, the relationship aspect, the psychology aspects, you know, what else can you kind of deliver? How that would be? My thinking, the software
Trevor Connor 46:39
can learn the athlete. Can it learn the coach as well? So to give you an example, we joke a lot on the show, and grant holicky is here co hosting with me that he’s all high intensity and I’m all threshold work, and it’s mostly joking, but there is some truth to that. Can Athletica learn that coach’s style? Would it build plans for my athletes that would have a little more on the threshold work type side where, for a grant, it would be much more the high intensity type work that he likes to do. Can you get it to do that sort of learning of the coach, or it’s going to start, as you said, with the base of your high intensity book, and then the coach can adjust it themselves. It’s
Dr. Paul Laursen 47:23
like you work for athletic at Trevor, honestly. So we just now we’re almost at the same phase where we were with the athlete version back four years ago. We’re just starting up now in the context of the coach, because that’s what we’re looking at. We’re looking at doing like a white label version of Athletica with your own sort of flavoring, your own sort of coaching flavor, just like you’ve described, right? You got a high intensity biased coach versus a threshold bias coach versus a polarized bias coach and these sorts of things, and with your own terms that you use and but ultimately, again, the ramble that I gave just before, they are actually they’re very similar. Don’t tell your athletes, but they are very similar at the end of the day, is what we sort of discovered. So yeah, we are working towards that. And there’s basically we’ve just got, like a query thing with a Google form for a white labeled kind of athletic with your own sort of flavoring, your own branding, all that sort of stuff. And we believe this will be the next phase of evolution in the AI training space, and we would like to assist with that. You can fill out the Google Form if you’re a coach, and yet, this is of interest to you, or a coaching platform, and this is of interest to you, and we’ll see if we can work together towards developing that for the future,
Trevor Connor 48:37
because that, I have to admit, would be extraordinarily appealing to me, because I’ve admit, as a coach who has a particular style, if the AI platform built a training plan that I look at and go, This is so different from my style, I don’t know exactly what to do with it, it would be a struggle for me. But if it could build something that’s similar to my style, it would have a huge appeal. And then what would be really interesting to me is there are times where I’m working with an athlete, there’s an event coming up, where I go, where I know my style isn’t right for building this athlete to their best form for this particular event. And to be able to then turn to the AI and say, Okay, come up with what you think is the ideal here that gets out of my style. I mean that that would be the best of both worlds for me. Yeah, that
Dr. Paul Laursen 49:24
feedback has come in loud and clear for us, from coaches. We’ve had over 1000 coaches onto the platform that have poked around and checked it out, and that is very usable, but you are a little bit locked into the hit science style at this point of the game across the various different sports, and it’s still workable, but you have to be inserting your own sessions within that. And if that doesn’t appeal to you, then you know you’re going to churn for us. And yeah, we want to address that audience, because we know the the coaches hold the athletes, so something we want to create for the future. True. So any coaches that are interested, please find that link there on the white label, Athletica, and fill out that Google form, and we’ll see if we can have a chat.
Trevor Connor 50:07
Yeah, and once again, we’ll put that in the show notes. I’m
Chris Case 50:10
just thinking about the mechanics of that. You’re gonna get feedback from a particular coach, and then the people that are programming the system are slot there’s probably not sliders on like this person’s on the polarized side of the scale, or this person’s on this has a sweet spot bias. But you’re not allowing the coach to input or prompt the system with their bias at this point, and maybe you never would. You would do the back end work to inform the system, the AI of their style or bias, and then put it back in their hands and allow them to then use it. Is that correct?
Dr. Paul Laursen 50:53
Yeah, the idea. And again, keep in mind, this is not done. These are live discussions that we’re having right now. But it’s actually more the front end in terms of we started and we let off in the podcast. We started with a template model. And the template model be no different than you working on your training peaks, and have you’ve got your template for your week and these sorts of things. That’s where we started. So that template that you have as a coach now comes in, and maybe it’s very threshold heavy, say, for example, then the idea is that you customized already, that templates already there, and we just make sure that’s aligned with your philosophy and how you work as a coach. And then that’s in there. But then it doesn’t change the Sport Science models working in the back end that can still deliver banisters, trim, critical power, critical pace, all these sorts of things, we can still get appropriate sessions in automated for the athlete, because that’s where I believe the key shift is happening. It’s less in the programming work that you as a coach might be doing where you’re, you know, going blind, trying to program for all your different athletes and stuff. And now much of that is being automated for you, and you’re doing less than that. You’re going less blind, and maybe you’re more blind because you’re doing zoom meetings now, and you’re more more engaged in making those connections with your athletes, but you’re actually like the granularity of the training program or plan design is much more automated and all done, done for you, to your parameters, to your flavor. So here’s
Trevor Connor 52:22
my sports science, 4.0 for you. Chris brought up the sliders. I want to have sliders where you can plug in your personality style as the coach, exactly, and just put mine all the way over to grouchy coach, so that no matter what my athletes ask the AI, it just replies, suck it up, sunshine. That
Chris Case 52:41
seems like it could actually be easy to do if it was just one answer.
Trevor Connor 52:48
You build that in. I’m sold. All right.
Chris Case 52:52
So, Trevor, you mentioned the 4.0 and joking mix. Do you foresee a 4.0 here? Paul, and what would it
Dr. Paul Laursen 53:02
be? 4.0 would be? Could be spooky. I don’t know, like you might have to go to some of the sci fi movies that we’ve watched in the past, but I don’t want to go there. I want to still be here. Yeah,
Trevor Connor 53:11
let’s save that. I would say my 4.0 is this really frightening paper that I read about all the inserted devices for measuring every fluid and metabolite in your body. Yeah, yeah. That sounds scary to me. That’s just 2.0
Chris Case 53:26
prime. Yeah, exactly
Dr. Paul Laursen 53:28
No. Thank you. Not for me, I’m out.
Trevor Connor 53:33
Well, we have a question for our forum, so please go to forums.fast.labs.com, if you’d like to participate in this conversation, the question is, have you used AI training tools such as Athletica? And if so, what do you like about them, dislike about them, and what would you like to see improve? So if you have some thoughts there, please go to the forum and share your thoughts. And with that, we got our take homes. Dr Larson, would you like to go first? You know how this works? It’s what’s the most important lesson that you’d like our listeners to take away from this episode?
Dr. Paul Laursen 54:08
Yeah, you know, this is similar. I think I’m almost reflecting on the first episode that I did on this with you. If you didn’t hear that last time, it should be more clear this time that AI is here and it’s here to stay and again, becoming more comfortable with what is available can only benefit you if you’re a coach, certainly if you’re a coach, because you should be more in tune with all this stuff than your athletes. Please get involved and learn what’s available there as an advantage for you, in that regard, what can speed your workflow and get better results for you and your athletes? And I’m certainly thinking things like workout reserve, things like HRV analysis, like those are all now embedded within Athletica and available for you. You know that you don’t have to be an expert necessarily at those anymore, where. Used to in the past. So continue to get involved, ultimately, in this. And we are moving into the mid adoption phase. We’re outside of the early adoption phase. So yeah, I think that would be my key take home. Good
Trevor Connor 55:13
take home. Chris, do you have yours?
Chris Case 55:15
Yeah, I would say that at this stage, at least, I see it, AI very much as a tool, and just because you have the tool in your hand doesn’t mean that you know how to use it, necessarily. So there’s that aspect of it, and to me, it goes back to that point I made, that you have to make sure that the data that you’re putting into the system, whatever it might be, whether that’s the data from devices or the prompts you’re giving a system, they have to be well thought out in a way or clean, but you also have to not abandon all of your ability as a human just because you’re dealing with AI. If it tells you to run into a wall, you don’t run into the wall. If it’s the GPS system and you’re driving your car and it tells you to turn into the lake. You don’t turn into the lake like you have to use common sense. Sometimes it’s getting better and it’s getting better and it’s getting better. But like you said, there’s still the opportunity for these hallucinations. So take it with a grain of salt and use it as a tool, but it’s not the be all, end all. At this point, I love
Dr. Paul Laursen 56:19
that. Take home. Awesome, Chris. That was
Trevor Connor 56:21
a great table. I was gonna do something similar. We’re worried about AI replacing everything, but at the end of the day, it’s still a person slapping on running shoes or getting on a bike that has to do the work. So at the end of the day, this will always be a tool. It’s a question of how you use the tools. But I was really fascinated by your description of the sports science, 1.0 2.0 and 3.0 and you really touched on something that I’m interested in, and we’ve talked a lot on the show about, which is that 2.0 phase, which is the huge amount of data. And people get excited all the time. Look at all the data I have. Look at all the charts I have. I remember training platforms that sold themselves by saying, we have this number of charts. We have 400 charts, right? All these sorts of things. But what’s always been a problem is when you ask people, okay, look at all that data you have, what does it mean? You kind of get a no no, yeah. So I like, actually, as somebody who was very concerned and skeptical about AI a couple years ago, I like this idea that we now have a tool that can give that answer to what does it all mean? Because I think the data has gone beyond the point where any human can sit there and process all that data, right? I think that’s going to be the biggest advance that we’re going to see, and that the greatest value that we’re going to see. Makes
Chris Case 57:48
me think that some times people are impressed by just the sheer volume and quantity of information, but if you don’t know what to do with it, then it’s has no quality at all. It’s just information, but if you can interpret it and analyze it and give feedback based off of it, then it turns quantity into quality, and that’s a good thing. It has huge value.
Dr. Paul Laursen 58:11
I can’t believe this is coming up now, and my CEO is listening to this, and he’s saying, Paul, I can’t believe you didn’t bring that up yet, because that’s the key thing that we are trying to do with Athletica, synthesis of the data and interpretation of that data for wherever the user is at. If you are just couch to 5k and you’re just coming out and learning about exercise and the importance to the elite coach, we can do that now with large language models. We can meet the user wherever they are at, and that is what we need to be doing. So thank you for that excellent finish. Trevor, glad I
Trevor Connor 58:46
made that my take home. Very good. Dr. Larson, real pleasure having you on the show as always, and I’m sure we’ll have you on again soon. Thanks,
Dr. Paul Laursen 58:55
Trevor, thanks, Chris, it’s awesome.
Chris Case 58:57
That was another episode of fast talk. The thoughts and opinions expressed on fast talk are those of the individual. Subscribe to fast talk wherever you prefer, to find your favorite podcasts and be sure to leave us a rating and a review. As always, we love your feedback. Tweet us at at fast talk labs, join the conversation@forums.fasttalklabs.com or learn from our experts@fasttalklabs.com for Dr Paul Larson and Trevor Connor, I’m Chris case. Thanks for listening. You.
