Can AI Actually Make You Faster?

AI is rapidly changing nearly every industry, and endurance coaching is no exception.

In this episode, CTS Head Cycling Coach Adam Pulford explores where AI can genuinely help athletes improve, where it falls short, and why the future of coaching may be more collaborative than competitive.

We discuss training data, prompting, bias, TrainingPeaks, Claude, ChatGPT, Vekta, Kristin Faulkner's use of AI, and the areas where experienced coaches still provide value that algorithms struggle to replicate.

HOST

Adam Pulford has been a CTS Coach for nearly two decades and holds a B.S. in Exercise Physiology. He's participated in and coached hundreds of athletes for endurance events all around the world.

​​Free Cycling Training Assessment: https://trainright.com/cycling-training-assessment-welcome/

Interested in working with a coach? Schedule a free consult: https://trainright.com/coaching/cycling/

Self-coached athlete? Check out our TrainRight Membership: https://trainright.com/membership/

Find more free resources here: https://trainright.com/blog/

Resources:

• Best-ever 20-minute power and 3 Pan-Am golds: Inside the AI tech Olympic Champion Kristen Faulkner built, and is testing on, herself | Cycling Weekly
• Strava Launches MCP Connector, Allowing Athletes to Sync Training History to Claude
• Vekta

Transcript

AI Hype And The Big Question

SPEAKER_00 0:00

How many of you have asked Claude or ChatGPT to build you a training program? Go ahead, raise your hand. Yeah, I have too. More or less out of curiosity to see what I would get back. But also because I've had a ton of friends and even some of my own athletes leveraging AI for some aspect of their training, racing, or recovery insights. And of course, they're not alone. Kristen Faulkner, the reigning Olympic road race champ, U.S. national champ, and multi-time Olympic gold medal winner, just used AI to set a new 20-minute PR for herself. But does that confirm that AI should be used by everyone to optimize their training and performance? Well, my answer to that is similar to whatever is hyped out there right now by the Pro Peloton. Whether it's ketones or broccoli shots, breathing devices, or AI, just because a pro is doing it doesn't mean you should and expect the same result. It's how something is being used and how it works, which is what you want to learn when something is being hyped. And sadly, most things out there are sponsorship endorse bullshit with a splash of applicability these days. However, AI might be a little different because it's rapidly changing the landscape of virtually every industry out there. And I would be silly to say that the world of endurance coaching and training is immune to it. I'm even eating my own words lately by saying AI isn't good enough to make a big difference or replace a decent coach. That was like a year ago, and it's gotten a lot better since, to the point of changing a lot, and I think it could replace a coach of a certain caliber, but not all coaches, in my opinion. So the question is: can AI make you faster? Yeah, I think it can. But there's more to it than just asking Claude or the Faulkner bot coach what is the best workout of the day, or how to plan an annual training program, or something like that. The best question is how coaches and athletes are successfully using forms of AI as tools to enhance their decision making for better training outcomes with the aim of making you faster, or whatever your goals are. And if you think it's weird that me, Adam Pulford, the head coach of cycling at CTS, one of the biggest endurance coaching companies in the world, by the way, might be going rogue by telling everyone how to use AI to coach yourself better. Don't worry, I approved all of this with our CEO, and he's cool with it. So let's get started.

AI As A Coaching Tool

SPEAKER_00 2:44

Couple of main points. And the first one is AI is a tool, just like many other things. And down the road, as well as currently, I think it's being used best to assist a coach or an athlete in making decisions about training. And to do that, you need to feed it information. So if you've done a good job of storing your data in one place like training peaks or intervals.iciu, you can export your data, feed it into Cloud or Chat GPT, have it analyze bigger data sets quickly, then give you feedback on what to do next. I did this myself, and it wasn't wrong. I gave it a prompt and started feeding it files over the past six months of training or so. It showed me where I was strong and where I needed some more work. It didn't tell me something I didn't already know, though. I was good at climbing, good at threshold, had decent enough volume, but I was lacking structured training, especially in the need of VO2 max work. Classic coach AP kind of stuff right there. It gave me advice of what to do next, and on a high level, I would agree with the advice that was given. However, if you don't know enough about training and how to build workouts or how to structure a week or a month of training, this is where it gets a little tricky. It can tell you and guide you through most of this, but it's limiting if you don't know the art of programming and how it'll fit into a bigger overarching plan. I'm sure this will improve, but for right now, taking the information it gives you and putting into a real actionable training program quickly isn't awesome just yet, going from Cloud or Chat GPT into some other platform. But I have good, clean data and a lot of it because I spend a lot of time making sure it's accurate and consistent, writing the same power meters year after year and cleaning up any bad data spikes or wonky calibrations. Which leads me to my next point. Your

Clean Data Or Garbage Advice

SPEAKER_00 4:39

information is only as good as the data that you put in. If you've been, let's just say, writing on vibes because the data stresses you out or something like that, you can't teach AI about yourself, and the tool won't be as useful as somebody who has a ton of data. Additionally, if you have bad data or a mixture of power meter data from the power meter multiverse, like single-sided crank arm units or pedal-based units or spider-based, you won't have as clean data as you possibly could, and you may get some tainted feedback from the AI.

Better Prompts Beat Confirmation Bias

SPEAKER_00 5:13

Another point here is your results are only as good as your questions. When I got to college, just quite a few years ago now, I started to learn that in order to learn more, you needed better questions. My mind went to the next level when I spent a few years being intellectually beat up by Dean Golich, Tim Cusick, and Andy Coggin when WKO4 came out, attending coding conferences basically in Boulder, Colorado, teaching us not only how to write code to build new fancy charts and graphs, but how to think differently about what we were trying to learn from them. Many of us young coaches were already used to analyzing files in certain ways via Training Peaks Online, SRM Win, Golden Cheetah, or the existing WKO3. But when WKO4 came out, it changed the way that we started to change the way we looked at bigger data sets or training and racing in general because we could go deeper and ask the system what to look for, observe the data for what it was, and see more clearly what the training that we were prescribing was actually doing to the athlete. And sometimes it was different than what we were intending it to be. Anytime I'd come up with what I thought was a good question, Tim or Dean would ask, Well, why do you want to look at that? And what if you took your method out of it and just analyzed the data? They were teaching us to remove bias from our questions, trying to break us from our old ways into new ways of thinking. And that was life-changing for me as a young coach. New, better ways of thinking should be embraced if they provide better solutions, like WKO4 and now WKO5 have done, and what open AI systems seem to have the potential for. My concern is that AI will skew toward whatever bias you already have because without a deep knowledge of training, you're probably not asking great questions. And humans usually have this tendency of just going with what affirms and confirms what we already know to be true. If your AI does this, it will likely only incentivize you to keep doing what you're doing because you're already awesome at it. When in reality, you need external feedback that removes emotion and bias from your strengths and weaknesses and tells you what you need to work on, like a good coach would do. If you're a self-coached athlete, you need to teach yourself to remove bias when analyzing data, even if it hurts your ego. If you're using AI, and you'll likely need to teach or prompt this like I did, unless you want to be told how awesome you are and enjoy this self-fulfilling prophecy.

What Faulkner Got Right

SPEAKER_00 7:54

In the case of Kristen Faulkner, it seems she's built a great tool for her, but she used her data to build the knowledge spine of the AI in order to give her targeted advice. And here's the thing, folks, it hasn't replaced her other coaches. She still leans on them and her teammates for the rest of the stuff that cycling entails. Now I'm not going to sit here and claim that I know everything about AI, her tool or her system. I've only met Kristen a few times at like USA cycling events, one fundraising thing, and she was teammates with one of my athletes at one point. I didn't reach out to her or her director for this particular video. They're a little busy at the Jiro right now, so I didn't want to bother them. But she's super smart and has the professional background off the bike as well as the results and experience on the bike to build an AI that can and should be able to do what she claims it to do. And even though I haven't seen it or used the tool that she has built, all the stuff I've read about, including uh a big article of that uh listed here below, I trust it's pretty solid. And more tools like this will come out soon, if not already existing behind the secret doors of the Pro Peloton. The one thing to remind ourselves about this is it's a very specific use case. Pro athlete, her own data, lots of other resources to leverage, and she has a specific knowledge in this area. I would not use the Faulkner AI tool on myself or on athletes without using it extensively beforehand. But I think down the road, coaches and coaching groups will use tools like this to assist them in analyzing data and helping make decisions about coaching, recovery, and to optimize their process. And I'm fine with that. I personally want this because it's going to speed things up and make me a better coach. And I'm using various AI tools already with good success, even though I'm working on streamlining my process to make it even better. Insider tip here. CTS is working on a little project similar to this. I can't say much more now, but check back soon.

Where AI Breaks Without Fundamentals

SPEAKER_00 10:00

Another main point: if you don't know how something works, you don't know how to fix it when it goes wrong. Here's where I think AI will replace less experienced coaches pretty quickly, but enhance experienced coaches who use it. I think many smart self-coach athletes could be pretty successful using AI to guide them for a while, but when the process doesn't work or something goes wrong, it's going to be even harder to know why or know how to fix it yourself because the fundamentals aren't there. It's like if you have a fancy fucking toilet with the best bidet in the world powered by AI or something. When it breaks, you still need to call the plumber to fix the pipes, and maybe a technician to fix the tech. Fancy things bring fancy problems. And right now, AI is the fancy new toy on the market. Additionally, the human element isn't there. To me, there's something missing about how AI training advice is given. How it will actually feel when it should be done and the broader framework around it. Let's take that hard VO2 workout it gave me, for example. VO2 work is hard, but it brings results. So you need to do it. And apparently I need to do it more. When I'm programming it for my athletes, I'm prescribing it at it at a timing that makes sense for the performance outcome down the road. So when they are racing, I'm gonna do a VO2 block that is six to eight weeks out. So I have plenty of time to stress the system, do the training, then rest, and then come into the race nice and fresh, ready to rip with a bigger VO2 max and more durability. But I also make sure that the VO2 block comes at a point in their life where they can handle it. Volume is lower, kids are in school, work stress is low, sleep has been good, and life is somewhat predictable. If coach Claude gives you a two to three week block of VO2 work of hard stuff two to three times per week, starting today with no other context, you'll probably overcook yourself probably two weeks deep, maybe less. Other things like how to actually drive a bike through corners at speed with 50 other people around you, and the anxiety that comes along with that, it's not felt by AI. Nor do I think it actually ever will. Descending and navigating complicated final circuits at the races when your legs are already tired. You need some real coaching for that. Adaptation to all of this skill and technique, AI will have a really hard time coaching and monitoring this. Experienced coaches do a solid job here. In the end, I do think experienced coaches will become even better if they leverage AI in a way that helps them analyze and organize data to help them make better training decisions faster and give them more time to actually coach. And for self-coached athletes, it will speed things up in the learning process so that you can level up faster than if you were to have started like three or four years ago. So,

Tools Like TrainingPeaks And Strava

SPEAKER_00 13:04

what are some of the existing platforms that may be using AI out there? I know I've been a big promoter of training peaks in the past, content pieces, and I still am. It's not a perfect software for everyone, but I haven't seen anything that is. I've been using Vecta for a few months now, leaning into how it works and how it analyzes data, organizing information, and suggesting training decisions about what comes next. It claims to use AI for this, when in reality, there's a big portion that seems to be like a large language model that generates some of this stuff, but I think it's a cool tool. It's not my preferred tool, but I do see some positives here for people who may not want to go all the way through the steep learnings to leverage WKO5 and may want something more than what Training Peaks has to offer. I should say though, the peeps at TP have launched some new stuff lately on their online app and plan to do quite a bit more, trickling out here in the next couple months. I reached out to Cody uh Stevenson and he said, quote, frequent medium-sized updates would be coming out over the next uh few months, so look for that. For all you premium subscribers to Strava out there, they just announced a few days before I recorded this that they launched an MCP connector allowing users to sync their training history with Claude. An MCP connector simply allows an external tool, in this case Strava, to be connected with an already existing AI, in this case Claude, to kind of do what I showed you earlier, but probably faster. I haven't used it yet, but I'll report back on it soon when I do. Anything is probably uh better than their existing AI at Strava. No dig there, but it wasn't that great. Vecta may provide some new bells and whistles for those who want to try something new that markets itself as AI, but something not as analytical as intervals.icu, slightly more advanced and agile than training peaks, and not as steep of a learning curve like WKO5, like I already mentioned. I wouldn't say I fully endorse it for everyone, because I have my pros and cons with it, but just like anything else, it will require some time spent learning. It's been growing on me, so I think it is better than Chat GPT or Coach Claude alone to analyze training data and what you're supposed to do next.

The Hard Part AI Still Misses

SPEAKER_00 15:22

Now, here's where I'm skeptical. The hardest problem in endurance coaching is not generating workouts or explaining charts. The hard problem is determining the optimal training intervention for a specific athlete at a specific time. Many platforms claim AI coaching, but they're really just doing a combination of physiological calculations, rules-based recommendations, and large language model generation explanations. Now that's useful, but it's not equivalent to a great coach making individualized decisions. Not yet, anyway. Finally, handling life stress. Until AI can feel the stress of human life, I don't think it's reliable to be the only decision maker to deploy workouts to an athlete. And I haven't seen any tool outside of empathy and another human brain come close to measuring the impact of this stress relative to how much training stress an athlete could handle on any given day. And that's a powerful thing not to be overlooked.

Which Kind Of Athlete Are You

SPEAKER_00 16:26

So now there's probably three bins of people listening. First, the ones who are already using AI. Second, the ones who are curious but scared or don't have the time to fiddle with fancy tech stuff yet. And third, the people who have no idea or no interest, I just pay my coach and I or I just write how I feel. If you're in the first bin, drop a comment below and tell me how you're using it and if it's working for you or not. If you're in the second bin, tell me what holds you back from trying AI in your training program. And if you're in the third, tell me what you prefer about working with a coach or just doing your own thing. In

Bottom Line And What Changes Next

SPEAKER_00 17:04

summary, I do think AI will help athletes and coaches reach new limits of human performance both physically and mentally. However, you need to keep your bias out of the equation, understand how human physiology and training works, make sure you have good, clean data that you're feeding it, while also know how the AI works. And that's an ever-changing landscape right now. As of this recording, I still don't think AI is a full replacement for a good coach. The human element is missing. And until they start racing the Tutor France by having just androids biking around, I personally think that we'll still have the need for human coaches. That's it for today. Hope you liked it. I'm sure much of everything that I just talked about will change in some form or fashion in the coming months. And when it does, I'll do another video to help bring some sense to it all, if I can. Until then, I'll see you back here for the next one.