Can Chat GPT Write a Good Triathlon Training Plan?

Carrie: Hey everybody, welcome to the show. Today we are taking a hot topic head on, posing the question can Chat GPT write a good training plan? And I can't wait to hear all about it. It's a thing. People are trying this, and we want to hear if it can actually work. This is a collab episode between Try Dot and RunDot. So I'm here, Carrie Tollefson, host of the RunDot Podcast. And Andrew is here. Yay! Host of the TriDot podcast. How's it going, Andrew?

Andrew: Yay. It's going good. Yeah, it's going good. I'm I'm really excited to get into this topic. I have not used Chat GPT myself for anything. Maybe I'm an outlier in that way, or chat as the kids are calling it. but I know I've seen some good chatter about athletes that are that are trying this or talking about this, so I'm excited to learn all about it and see what ChatGPT can or cannot do for training. Our guest today to teach us all about this is Jeff Booher. The founder and CEO of Predictive Fitness, the company behind RunDot Training and TriDot Training, and so much more. Jeff, thanks for joining us again.

Jeff: Absolutely. It's gonna be fun. looking forward to the topic.

Andrew: Well, as always, we'll start off with our warm-up question, move on to the main set learning from Jeff, and then we'll do our cooldown where we ask Jeff an audience question. Lots of good stuff. Let's get to it. For our warm up question today, I've been pondering on this one and I want to know if you went to sleep tonight and woke up tomorrow just forgetting every race you've ever done, just your memory is wiped of those experiences, except for one. Which one race would you most want to remember? Jeff Boer.

Jeff: This is a really good question. I'd say, so it's the triathlon. I'd say the one that it popped into my mind immediately, and I was kind of peculiar that it did. It was, I think, 2011, 2012, Lifetime Triathlon. They did a series back then where they did short course triathlons all over the country. It was a big race. I had just had like a knee surgery really bad about a year before, so I couldn't do anything for a year. I was in a brace for a year, all this stuff. And so it was my first one back after that. And I worked really hard to get back in shape. It was a September, October race, kind of end of the season. And I I thought I was in the lead. I it turned out I was in second place at at the time and I didn't prepare ahead and I didn't check out where the turnaround was. I was just looking for the signs and they hadn't set up the sign yet of here's where the turnaround is on the run. And so I was toward the end of the I thought I was. But and I ran more than a half mile past the turnaround.

Andrew: 'Cause you were just that fast. You were that fast, Jeff. You got there before they could.

Carrie: Ha ha

Jeff: And then I realized like I am off course. There's no one here. And I turned around and went back and ended up from what I was second to 30th or something like that, 32nd. And so that race was it was kind of bittersweet one is I was very proud of myself to get back in shape and I thought I was doing really well. So I knew my potential was there. I saw that. It was very rewarding, but it also taught me a huge lesson to prepare ahead of time. And so I do not want to repeat that. I do not want to learn that again the hard way. And so that was kind of the first thing really that made me want to remember that. Like that is something I I never want to do again. So hard lesson to learn, but I wanna keep it with me.

Andrew: Very interesting that that is the one that she would want to remember out of and that those that lifetime series, those were Olympic distance events, Jeff?

Carrie: Mm.

Jeff: They were they Olympic and Sprint both, yep. There's there, Austin, there's kinda all over the country.

Andrew: Okay. Very, very cool. Carrie, do you wanna remember the Olympics or

Carrie: Yeah, they used to they used to have one here. They had they had the lifetime try here in Minneapolis, which is always a really competitive, yeah. But I'd have to say, Andrew, the one that probably sticks with me is how I made my Olympic team at the Olympic trials. Like the Olympics obviously is like, you know, the the dream, right? But for us to make our Olympic team here in the States is pretty like You get you gotta be top notch, right? And you gotta be able to do it one day every four years. No injury, no illness, nothing. And that one for me, when I made the team in the fifteen hundred in an event, that was probably my second event. that one was wild. I was not supposed to lead. I led from start to finish, basically. My coach was like, You better be happy you won. I got into a little trouble after, but yeah, it was pretty spectacular and so I mean I that life that race changed my life. For sure. Not the Olympic race, but the Olympic trials race. So that's the one I think of. How about you?

Andrew: Yeah, whe where was that race? Where where were where was qualification then?

Carrie: In Sacramento, and it was a hundred and sixteen degrees on the track. We were like prime time NBC, you know, timing for the fifteen hundred, because it was a that's a you know pretty flashy event, right? And so the sprinters, when they were getting into their blocks, they were burning their fingers, so they would pour water on the track. I mean, it was ridiculous. And we made it through. It was it was life-changing, like I said. Yay!

Andrew: Okay. Yeah. You made it through, you qualified. Yeah. Love that for you. My answer here is the Escape from Alcatraz triathlon. All of our triathletes listening on TriDot will be familiar with this. For our RunDot athletes listening, this is a triathlon event. It's very unique. It's a very they call it a modified Olympic distance. It's not a very standard distance, but basically, like they ferry all of the athletes. It's in for San Francisco. And they ferry you out into San Francisco Bay right off the coast of Alcatraz Island. Everybody jumps off the ferry into the water and you swim through San Francisco Bay back to shore. And then once you're where there's sharks! Thank you very much. I was thinking about sharks the entire time. I'm not gonna lie to you. I wanted to get out of that water so bad. and

Carrie: Where they're sharks and it's wavy my gosh. And it's cold, like it's not easy. And I have not done it. I just know this.

Jeff: It is freezing. I got to see it like a month and a half ago. I was there. The first time I've watched the race, actually. It was it was incredible. And the sand ladder. How'd you like that?

Andrew: Yeah, it's very cold. Yeah. The sand ladder. So yeah, and so the the bike course, you're biking all up and down the hills of San Francisco, and it's exactly how you would imagine. the run course, you're on beaches, you're on sidewalk, you're on trail, and yeah, there's the sand ladder where you run about a mile on a beach, so you're running in sand, and then you have to climb up the like back onto the main road. And Jeff, I I forget how how high of a climb it is, but it's literally a ladder made out of sand steps. And there's like ropes that you're have you're like shimming yourself back up. there's a couple of spots

Jeff: It's People are pulling themselves by the ropes, their legs are given out. It's

Carrie: What

Andrew: Yeah, it is so it's kind of like a mix of trail running and beach running and then pathway running. It is it is a wild event. And like you would think the one I would want to remember would be my first Iron Man or my first half Iron Man, which I did in New Zealand, which is awesome. But when it really comes down to it, that Alcatraz event is so unique and so special. I I I tell everybody, it was the wildest three hours of my life. And if I could only remember one race, that's the one I'd want to remember and tell people that I did.

Carrie: Mm-hmm.

Jeff: That's cool.

Andrew: we're gonna throw this question out to you, our audience. If you're watching us on YouTube or Spotify, comment right below and let us know of all the races you've done. If you can only remember one, what's the one you want to remember? join the I am try dot Facebook group, join the run dot Facebook group, and you can also answer this question there. I'll throw it out. You'll Andrew Harley asking this question. Let us know what race is this for you. So I admittedly did not realize how many athletes were actually using ChatGPT to generate their training plan. One of the software developers that works for Predictive Fitness on RunDot and TriDot told me a good story where his neighbor raced an IRONMAN, used TriDot for the training, had a good experience, signed up for another IRONMAN, and said, Hey, I'll save a little bit of money. I'll just have ChatGPT write me a training plan. I'm a smooth sailor.

Carrie: Okay. Yeah.

Jeff: huh.

Andrew: Well, went into the event. It was not smooth sailing, had a horrible experience, and got back on the TriDot bandwagon shortly after. And that's the first time I realized like, wow, people are actually doing this for events running and triathlon. so there's a good chance that that if you're listening today, maybe you've thought about this where you're like, hey, I just threw a race on the calendar. I have X number of hours to train. There's no easy way for me to build a plan myself. Let me try ChatGPT to write one for me. and if you do that, what comes back probably looks great. It probably looks legitimate. It looks like it's a plan that's gonna get you ready for that event. And so most athletes walk away feeling like they've solved the problem. I've got a training plan for free. Jeff, you've spent twenty years building training technology and you look at that same moment very differently. Where do you want to start in talking about this one?

Jeff: Well, like you said, you talked about the athlete that it's a normal thing to do. Going to Chat GPT, it's amazing, it's incredible, it solves all these problems, you know, it's just pretty miraculous. So that's a reasonable move for athletes to do. the problem, you come to it with a problem. You have you don't want a training plan, you have limited time, you have no expertise or little expertise to build your plan, or you'd have built it yourself. So Nothing I say is gonna like really knock the athletes or going to it because going to chat GPT is a reasonable thing for sure. but what comes back out of that when you get the plan, it looks professional. Like you mentioned, your friend that you know, you know, he got the even the second plan. It looks great. it looks like something that a coach might have charged you for. And so you feel that relief. You go, cool, I got a plan. So the relief though is really the first part of the problem. It's not the thing that the tool does on purpose, you know, you walk into it and it's not that you're doing it carelessly. but the problem was built in the solution itself. And it's the whole conversation comes down to one distinction. That the plan can look right and the training can look right. I told myself. Alright. So the whole comes the whole conversation comes down to really one distinction. Even when the plan looks right, you know, in training, looking right and being right are completely different things. So I'm not here to, you know, attack Chat GPT, and in fact I I just pick that. I say we you know talk about chat GPT as a representative of the whole category. Claude, Gemini, Grok, you know, all of these they're the same. So we'll just, you know, rather than listing all of them every time. just kind of single out Chat GPT. And they're really brilliant tools. I mean, everybody that's used them, other than you, you ought to check it out. And you use AI though, Andrew, but for specific things, for editing, for different things. You know, you know, all over the place. But and it's basically what it is. It's the wrong tool being handed the wrong job. So and what happens is that you quietly and unknowingly it costs the athlete. It costs them in performance and injury and in so many other ways.

Andrew: Yes, yes, d yeah, to be clear, yeah. Yeah.

Carrie: Hmm. I mean, this is gonna be really interesting, the whole conversation today for me because I am really just learning about all this AI stuff. I'm gonna be listening and I'm gonna be tuning in real close to what you say, Jeff, because I know you study all of this and you know all of them. So when I hear AI powered, sometimes I'm nervous and I don't even want to even go there. But I also think that it means smart. It means it's studied, it knows what's going on in the internet. So what is the one thing that is beneath all of this, underneath all of this? When an athlete types into their computer, build me a training plan, and they use Chat GPT or they use Gemini or these Cloud, whatever they're using, what does that actually mean? What is the answer they're gonna get from AI?

Jeff: Mm-hmm. Yeah, so the first thing, I guess, the place to start is what it is. What is that chat GPT? What is, you know, Jim and I Claude. And it literally is a large language model. So GPT stands for Generative Pre-trained transformer, GPT. And the keyword there is pre-trained, like trained ahead of time. Before you got there, before you asked the question, it was trained. So trained on what? This is a question. Is trained on language. So an enormous pile of text, billions of words and articles and things all across the internet. It's not changed on trained on physiology, not trained on workouts, not one heart rate curve, not one recovery day, nothing has never seen adaptation happen. It did not learn from training. It learned from what people have written about training, which is a very, very different thing than learning from what training actually does to your body. So it was trained on words and not workouts. So the quieter problem beneath that is that the it learned the text. from text it learned but it never was able to learn and check for truth. So you'd get the training that it was training about or talk about training words about training that that were right beside you know accurate stuff from non-accurate stuff but it always comes back you know you have cop confident popular very wrong training advice that sits on the internet right next to the good and the bad and when it's being trained it has no way to know What's good and what's bad. It can't tell apart the good from the bad. So it's absorbing this whole pile. huh.

Andrew: And Jeff and Jeff, just for like this is a oversimplification for sure. There's a number of differences between what tried out and run dot does, what FitLogic powering tried. does versus Chat GPT Claude Gemini. but just right here I want to note as we're talking about how these things are trained on words, they're trained on language, articles, things that are written about training, FitLogic is trained on actual training data, right? That that's one of the big differences between the two.

Jeff: Huge distinction. Yeah. So it's absorbing this whole pile of raw data. and it and it's trained to sound fluent. it's not trained to know what's you know what's true and what's not true. so what it's very good at is predicting the next word. What word comes next in this pattern or this sequence with given the context? So when it writes a plan, it's not reasoning about your body. It's not it's just assembling the words of when a training plan you know exists or when you put it together, what is the next word that should happen? What is the word that would next normally happen when you're comparing training plans or looking for training plans? So for your training plan with Chat GPT, it might as well stand for generating plausible training. Because that's what it looks like. It's just plausible. It's not right or wrong or whatever. It's just it's plausible. It's what you would expect to get back given the context. Yeah. So it's made of the right words, whether or not it's right for you.

Andrew: Yeah. Yeah. This is a training plan. Yeah.

Carrie: Yeah. That I think is key. And I think for someone like me who is very new to AI stuff. and don't get me wrong, I try and use Chat GPT, and especially like when I'm using doing research now as a commentator, I'm trying and I'm finding it is very wrong a lot of times, more times than not right now. But I think that's key, and I want you to kind of just reiterate that for us, Jeff. Like it's not studying the training plans that have been built and that have been published for people. It's studying words out there. And I think that's how it's building it. You know, it's not a famous David Roche you know, training plan necessarily. Maybe it's been on Google, maybe p it can pull from that, but it's the words, it's not the process.

Jeff: Exactly right.

Andrew: Good job, Carrie. Nice, well done. Yeah.

Carrie: Thanks. I kind of wrapped it up.

Jeff: So it's like training it it'll learn the theory, it learns theory, it knows how to repeat theory, it knows how to repeat methods and concepts, but to actually apply it to a person for you, it just is not that's not what it does.

Carrie: Yeah, like I have plugged in give me a good strength workout just to see. And you know, it gives you pretty much kind of what you would think, but definitely not personalized for sure. So yeah. Great.

Andrew: Something. Yeah. So, Jeff, I wanna push on the phrase AI powered for a second. I've worked at this company long enough to have seen different people's reactions to that that phrase. there's a lot of athletes that see training or tech or something is AI powered and they think it's a good thing. It sounds like it's a guarantee of intelligence. This thing's gonna work for me. I've seen some people react negatively because there are so many things out there these days that claim to be AI powered. that are just slapping that label on something that's very inadequate. And so they roll their eyes when they see AI power thinking, this is just another new scam product out there trying to get my attention. In plain terms, what is that label actually telling an athlete about whether a tool can do the job they're handing it?

Jeff: Well, kind of like you mentioned, almost nothing. it just depends on the tool, how it was created, why it was created and what it's used for. So a label tells you, you know, AI that some level of automation is involved. So there's some computerized something, but it doesn't tell you anything about how the thing was built or why it was built or what it can do. So I I mean I think all the time we have it's just an everyday instinct that you would never add a chat bot. for example to get directions to an airport. You know, you don't go to Claude or Chat GPT or anything else. Say, how do I get to the airport? You open up a training, a navigation app, and that's what it was made for. You do that without thinking, and why? You know, because it was built for that job. It knows the live conditions, it's monitoring traffic. It's knows when there's a wreck ahead or when a w road is closed or there's a detour. So it's not guessing at what a plausible route would be or what it might look like. It's actually calculating the best route. It's actually monitoring. It was created to monitor and to ask questions. Hey, here's another route. It's so it's looking for alternative routes as you go along. Chat GPT, they don't know to do that. And so for us, that that's very helpful using a tool that doesn't know the questions that we even know to ask. So you don't even know what the other possibilities are because you're not given that opportunity. It does that for you. It looks for those opportunities and creates the right things. So the tool is created for that job. So a chat bot. Could be descriptive of a plausible route, you know, and you get it back and say, hey, that looks reasonable that you can get there going that way, but it has no idea what the traffic's doing, like right now. And the same thing when you're looking at your training. You know, the Legal that AI doesn't rescue this AI-powered training, it doesn't mean very much. You have a submarine and a canoe, they're both vessels, but they're not the same thing at all. They were not created for the same purpose, and they do very, very different things. So just like the word vessel doesn't tell you anything about. you know, a ship or a boat or anything. it tells you nothing about what the thing is or what it was built to do. And the same thing with training. So when you think of the letters AI, they tell you nothing about whether the tool was built for the job, that you just asked it to do, that you just handed it. So the real question was never, you know, whether or not your app or your tool is AI powered, but it's whether that whatever it is was built to understand training or just talk about training.

Carrie: So how do you trust it? Like say I have my plan and it looks good. Just like I said that that strength plan looked good. You know, it had the right words or it has like for runners, phases, paces, all of that stuff. How do I know that it's right or that it's any good? And what is something that you look for? What's the catch that you can see because you study all this, Jeff, that I am not seeing?

Jeff: That's an excellent question. I mean, that's the heart of the whole issue. Like, start out with a why you went to ChatGPT. You went to Chat GPT to ask for a training plan because you believed that it knew more than you did about putting a plan together. That was the whole point. That belief is the whole point of you asking the question. but it's also the whole problem. So the only way that you could know whether what you got back was actually good or not is would be if you had the expertise to recognize a good one yourself. And where you're able to look at it and say, hey, here's what's missing, here's what's off. Here's what doesn't fit. And like you were, you know, doing research for your commentary. You know certain things and you can look for those things and you can use it for help. But if you don't have that expertise to do that, you're just going with it because it looks right. You're not going to be able to recognize those things. So that expertise needed to validate the plan that comes back from a chat GPT is exactly the same thing that you went there to borrow from it. It has the expertise you don't. And so that creates that problem. So when you get it back, you're not really evaluating the plan. you look at it and you're just thinking, well, this looks good to me. it looks detailed, it looks organized. none of those different qualities though is gonna tell you whether it's right. The one thing that that could catch the mistake is the one thing that you don't have. You know, you asked you asked the question because you couldn't evaluate it, which means that you also can't evaluate the answer that you get back. so a plan, you know, it can name every workout, it could have all the right vocabulary in the right place. And it can still be built on nothing that's specific or nothing that fits you. So the right structure, the wrong plan. And when that happens, you know, no alarm goes off. That plausibility is the problem. It is the trap to this whole scenario. It the better it looks actually, the less you question it, which sets you up for you know problems.

Andrew: Yeah. Yeah, so what makes it so dangerous for an athlete potentially using it is it's not that it's bad at sounding like a coach, it's dangerous because it's actually good at sounding like a coach.

Carrie: Yeah.

Jeff: Exactly. And it was built to sound that way. It was built to sound persuasive and plausible. So that's exactly the thing. If it were obviously bad, then you'd catch it. You'd know. So the danger is exactly right. It's how convincing it is. And plausibility isn't a side effect. That's what is created to do. It w and you can't prompt your way around it. You can't create a better prompt. It's always going to do that. It's in the product. It's how Chat GPT and all the other LLMs, they are built to sound right. That's what they do.

Andrew: Yeah. Yeah, so there's probably some athletes listening right now that are thinking I've used Chat GPT for my training, or maybe some run dot and tri dotters listening are thinking I've had training partners use Chat GPT for their training. it gave them clear, confident answers. and they're like, hey, it's been fine. Make the case to that athlete that the confidence in your Chat GPT plan doesn't necessarily mean that it is the right plan for you.

Jeff: Yeah, well like the clear confident, just like you said, that's real. It is. That's what it was built to do. So that's the problem. And but the confidence is there no matter what. You're not gonna get an answer back that's not confident. So and that's what you know, something can be the answer could be, you know, exactly right or completely wrong. You're gonna get the same level. There's no hedging with Chat GPT. It's not gonna say, Well, I'm not really sure about this part, or it depends on this, or you know, I can't see this, I don't know that. It gives you the answer as if it's authoritative. So Confidence and correctness are two completely separate outputs. and they're designed to produce the confident output every single time. So confidently wrong is still wrong. So, you know, you already know who to trust when with confident sounding answers everywhere else in your life. I think when you think of you have a knee injury or something else, you would not let Chat GPT diagnose a knee injury and actually act on it. You may ask questions, learn a little bit, see what the problem could be. But before you're gonna take any action or you know determine what to do, you're gonna ask a doctor, you're gonna go to an MRI get an MRI, get something, get an expert to check that out. and if you were gonna write a will or sign an important contract or something like that, you might get some feedback from ChatGPT. But before you sign anything important, you're gonna have an expert look at that. And in those cases, in the rest of our lives, we know a couple things. One is the stakes are real, they matter, and we know that we can't judge that. So we can't judge and evaluate. what comes back. You know, we can take it as inputs, consider it, inform us a little bit, get a little education, but when it comes down to it, you bring that to someone or something that's created for that that job. So unfortunately though, many runners, tri athletes, when it comes to training plan, a training plan just doesn't trip that instinct, that protective instinct. You know, it shows up, you get it, it's clean, organized, sure of itself, it's confident, but nothing about that experience says slow down. Hey, wait, before you run this, I want to check it out. And they just go with it. So, you know, you double check answers throughout your life, everywhere else, when the stakes are high invisible. But I think with training, it just hides the stakes a lot of time until it's too late. You know, you run the wrong plan for long enough and your knee breaks down and you have these problems. And then when it does, what happens? You take that knee to the doctor for an MRI. but it's funny and kind of ironic that You get the injury and you take it to an expert or a machine that was made for that to diagnose the injury, but you never for a second questioned the plan that caused the injury in the first place.

Carrie: You know, I keep going back and I can't think of what movie it is from back in like the 90s. Is it like the Terminator or like I don't know some reformer where the robot goes wrong and goes starts like you know it's like a horror film? I don't but did the is was the Terminator a robotic thing? I think so. I think it was Schwarzenegger was like

Andrew: Carrie, Carrie, you just described like forty seven different movies from the nineties. Probably.

Carrie: He was the outside, but it w he was like, and it went wrong, right? Like he started to lose his mind. Like that's kind of what makes me uneasy about all of this stuff. Like, you know, you never know. You don't know. It looks good. Everything looks pretty and all of that. But, you know, you've just said that the plan can be, it looked good, like I said, but it can be confidently wrong. Is there a point, Jeff, where it's not just off, but the thing is telling you something that isn't even real?

Jeff: Ha ha. for sure. It's a it's a common fact, I guess, just around it's called a hallucination. So the language models, yes. Yes, it has a name. It it's for that. And it's a fact that's just simply not true, but it presents it as a fact. So they can I've heard so many studies going back so long of just doing research and citing studies just that don't exist. I was talking to my attorney a couple weeks ago. We were talking about a case and some other stuff.

Carrie: And that is crazy that it's actually called something.

Jeff: And he was saying, yeah, he had done research before, and it would come up with precedent, these cases that had never happened, they didn't exist. And so, you know, there's good things about it, but there's these very, very bad things that you have to know. And if you don't recognize it and then research the research, it will correct. It will. It's exactly so the studies that don't exist, cases that don't exist.

Andrew: It's going to give you an answer, right Jeff? Like if you ask it a question, it's going to give you an answer.

Jeff: You know, so it'll describe physiological adaptations incorrectly. It'll and it does all of that with the same fluency it brings to everything else. So it's confident. You know, it reads identical whether the reasoning behind it is sound or whether it's just completely invented. There's no tell, there's no flag, there's no sign, there's no shift in tone, there's no signal at all. and it's not a rare bug or glitch, like I mentioned. It's known, it's a documented property of how LLMs work. They just do that. That's kind of a unfortunate large language model. So that's what all of these are. Chat GPT, Grok, Gemini, they're language models. They were built to understand and interpret language. And they study that. And so they're pro really predicting what word comes next. So given all of the context, all the words before it, here's what word should come next. And that's what it's based on. So they're LLMs. That's what they're trained to do. and if they need to, you know, create another word or create something that seems plausible, then that's what that's what it'll do. And then it even gets more convincing, Carrie, because When you interact with it, you get something back, and now you start to talk to it. You start to describe how you feel, and then it responds, and you ask a follow-up question, and it adjusts, and it comes back and forth. and it feels like there's a mind behind it that's weighing your circumstances, your things, and coming back with a thoughtful, reasoned, calculated answer. It's not, it's just doing the same thing it always does. It's predicting what would be a good, what would a good answer sound like one word at a time, and it brings that good sounding answer back. and so the conversation starts to feel smart and it starts to for some reason, you know, that's why playing can look smart. but both were generated to seem right, but not built to be right. So the danger isn't when it fails to sound right, it's actually when it succeeds and sounds very right, but it's still wrong.

Carrie: Yeah. Yeah. Cause like, I mean, I can ask, what job should I search for? And it will be like, you're great at commentary. You can speak in front of a crowd. Like it knows everything about me. My PRs. If I ask, you know, like what kind of training should I do? Which I have done. I've put it in there. It knows when I should race all of the things. And I do feel a little scared at times that it actually knows me. Sometimes I feel like better than I know myself. But What is actually happening, Jeff, when the plan feels personal like that? What's going on behind this Terminator?

Jeff: Yeah, well what's happening is it's taken a shape of something like a training plan. So it can go out there and can find training plans. People have written about training plans, the concepts, all of these different things. And it's just gonna write your numbers on top of it and apply it to you. It's gonna put, you know, like putting your name on a plan doesn't mean that you are in the plan that that's written. So it's trained if it's trains on everybody's word, you know, and nobody's body. So it's just a bunch of words from a bunch of people, but no one's specifically their body. So there's no mechanism in it to do that. There's no mechanism

Carrie: Yeah. Yeah.

Andrew: Yeah.

Jeff: in the large language models to know who you are excuse me, who you are and what you should do. So no nothing from your training history other than the words themselves. So it doesn't know how much work you've done or how much you could build up to, your body composition, your age, your gender. Yeah.

Carrie: Yeah, or what your injuries have been, if you've had a layoff, like that's the stuff it doesn't know. It knows what it can find in the computer, but it doesn't really know you.

Jeff: Right. And even if you tell it those things, you give it those words, it's just words. It doesn't know how your performance ability should be proportional to any kind of intensity or how long you recover. It doesn't have a calculation. There's no mechanism to do any of that. So I mean you could have someone like you, fifteen years older, completely different training history, go in and ask for the same plan. You put in the same thing, you're gonna almost get the same thing back out, regardless of two completely different people. And a plan that would fit anyone is built for no one. So and anyone is not the person that has to run the plan. Now you're the person that has to do that. So it's really important.

Carrie: And not to get off topic, Andrew, like but that's what fit logic does. It is there is AI, there is a component, but it does know our it studies our heart rates, it studies the environment we're in, it studies our sleep patterns and stuff like that. That's the difference, right, Jeff?

Andrew: Yeah.

Jeff: Yeah, so if you think of LLMs, they're trained on words. Ours for twenty years is purpose built, it's AI, but not a language model, it's a fitness model. It's learning from workouts and your physiology and contextualizing that and then count you know, normalizing for weather and age and all of these different things, and it understands training. So it has the actual mechanism to do that and to learn from it and to get better and better over time.

Carrie: Yeah. Yeah. Yeah.

Andrew: And I'll tell you what, like being an athlete on I've trained with TriDot now for gosh, Jeff, when did I start? Twenty eighteen. So I'm coming up on eight years as a TriDot athlete myself. And it it's walking out the door with my workout memorized in my head. I always open up the app and see, okay, here's my interval pattern, here's how long I'm in each zone. these days you can push it to your watch and your watch can beep at you and tell you what zone to go into when, but Knowing that that plan is legitimately personalized for me and is the right workout for me on that day, that there's a certain confidence there in starting that run, starting that bike ride, entering the water for that swim, as opposed to, well, I hope Chat GPT got this right from whatever sources it was sourcing. I don't I don't think that's same confidence would be there.

Carrie: Yeah. Well, and I think what's cool, Andrew, is it's following along with your journey in life. That's what I think is cool about what you all have built, you know, like it's it is watching your day to day. It's not just, okay, plug and play type stuff.

Andrew: Yeah. So Jeff, here's a question I'll ask you. again, I've I'm not a Chat GPT user, but I've seen this reported by them. an example of an AI tool that I use in my day to day workflow. we record these video podcasts in a platform called Riverside. Riverside's a great podcasting platform, state of the art. And Jeff, when we first started doing video episodes, I had to like record everybody's webcam. Get them into Adobe Premiere and by hand cut the show and decide: okay, here's the wide shot. Here's I'm gonna go solo on Jeff Booher for the next 30 seconds, then I'm gonna go back to myself. And I was hand cutting the show on when you're gonna see what camera. If you're watching us right now on Spotify or YouTube, whatever decisions are being made about whose face you're seeing at any given moment, that was done by AI in Riverside. And it's a great tool. It saves me a ton of time. It doesn't always make the exact decisions I would have made, but it's close enough that you're watching this and it's a good polished video podcast. And a lot of video podcasts are using a tool like this. And Jeff, I can have the I can press the AI button, it'll cut the show for me. I start editing some other stuff, and then all of a sudden, if I accidentally hit that button again, it's gonna cut the show again and it's gonna make totally different decisions on when it goes to everybody's camera. And I've seen chat GPT users support a very similar phenomenon where they can ask the same question multiple times to chat and get a different answer every time because I guess it's just finding different sources and saying this is right. Nope, now this is right. Now this is right. Have you tried this with training, Jeff? Would it give you a different response if you asked it for a training plan multiple times?

Okay. So that's exactly right. That's exactly what happens when you know you ask ChatGPT the same thing, even yourself in three different threads. Like start at different times, different days, ask it the same question. Either you're gonna get something very, very generic back, or you're gonna get different plans, not because there's different inputs, but because each time it's trying and going through its process differently. So there's nothing underneath it. There's no understanding of a process or a system or a framework, nothing underneath it, nothing to stay consistent to. so why that happens, it's there's no like underlying model of truth. It's not doing calculations. There's no calculating at all. So it's not ha doesn't have anything to calculate from. And it has no mechanism to calculate. So each time it just generates the new plausible answer at that time. And if you think a real training engine is a real training engine, you know, given the name that the same athlete, the same inputs, it's gonna produce the same prescription every time because it's calculating. It's kind of the whole scientific method. I mean, if you if you have your hypothesis, your test, your protocol, you put in the same inputs, you should get the same output every single time. And that's because there is a standard of truth calculations. things that are that have a weighting and are based on actual truth. It's not composing, it's calculating an answer correctly. Yeah, correct. And so I think if you put those two demos together that we were talking about, you know, you have you can have different athletes nearly the same plan. And you can also do the same athlete and you get very different plan on different days. And so there's no consistency and that's because underneath that there's an absence. There's no framework underneath it to generate accurate

Andrew: It's giving you a right answer and not just an answer.

Jeff: You know, correct training for you.

Carrie: I've heard you say something before and it's a phrase that is used for a plan that looks complete, but there's nothing underneath. We've talked underneath a lot today. We've used that word a lot. I don't use underneath a lot. But you've s talked about a cargo cult. I don't actually know the story, so I want to hear it. And like how does it fit into all of this that you've been talking about?

Jeff: It's very it's a very visual thing and it's very fascinating. It's goes all the way back to World War II and it's not just a cargo cult, it's not just for training plans or anything. It it's a number of different things. So you'll see how it can be applied here. But after the Second World War, some communities in the islands and in Melanesia they watched the military come in and these cargo planes were landing during the w the war and they were delivering these extraordinary supplies, all this stuff just out of the sky, planes come down, all this stuff comes out. And they're like amazed. And so after the war ends, the planes stopped coming. And so for a few in that in the community, they actually set out to bring the cargo back. So how can we, you know, keep those planes coming back? So they actually built runways out of straw. They built control towers out of wood and bamboo. And they carved headphones out of wood and they sat waving signals. sat set out to trying to wave signals to on this empty airstrip trying to get planes to come down. But the planes never came down. but they just associated this phenomenon, what was happening, with all the visual things they could see. So every visual detail was faithfully replicated. The form was perfect. The function though was completely absent because the form was never what made the thing work. That's not what made the planes come. So in a a language model's plan is kinda like a cargo cult and everything can look good. There's phases, microcycles, will builds and weeks of taper or peak weeks and All the right vocabularies and all the right playing places because it was trained on thousands of descriptions of real plans. So it's describing it. but it learned what a plan looked like very well. but a real plan is not based on those ritual things, it's based on the reasons that those things are put there in the first place. And so that's the you know, in training, you hard block is there for a specific reason. It knows how much you can handle and what you're training for and what your limiters are. In a taper week works for a specific reason, not just that there is a taper, but how much do you pull down, pull back, and what and what is the load of your race coming up? And you know, all of these different things go into the specifics of how the hard blocks, how the taper, how the recovery works, you know, and the phases are just a visible shell. The logic determining though, but what goes into each one and why it goes in each one and that function, that's something that that they can't they can only imitate that, but it's not real. So a plan. can have every phase in the right place and still be built out of nothing. Nothing underneath.

Carrie: Yeah, underneath. You know what got me there? Wooden earphones. They even went that far to carve out wooden earphones to make it to replicate all that. That's crazy. I'm learning so much, Jeff.

Jeff: No. Yep. Yeah, it's crazy. It's so fascinating. But the same thing with so often I think we look at the outer resemblance of something or the form of something, but we don't have that expertise or that depth or the functionality to to create what's underneath it that actually makes it happen.

Carrie: Yeah. Interesting.

Andrew: Yeah, it reminds me, Jeff, of playing with my toddler. And there's certain household tasks that she's seen my wife and I do that she can do verbatim. and she can pretend like she's putting muffins in the oven, but she doesn't know you actually have to like turn the oven on and it needs electricity and it needs to like heat up to actually cook the muffins. She just knows the door opens, the muffins slide in, the door closes, and then a little while later we enjoy muffins. right. She doesn't understand the core in there. And it sounds like, you know, you can dress up a training plan with phases and a taper and some stuff, but without the actual power behind it, underneath it, it's just a costume. It's just empty. it's just an empty ritual like you're describing. Jeff, I I wanna I wanna voice something that I've heard at races. I've heard talking with athletes before. maybe they're g They're getting defensive right now because they've used a plan like this. They've used a Chat GPT plan. Or before Chat GPT was a thing, maybe they use like a couch to 5K style plan, just a just a paper plan. And they'd say, I used a plan like this. I followed it. I got faster. It worked. I finished my race. I had a good time. What's the problem with me doing this? What would you say to an athlete like that?

Jeff: Let's say that's true. Great, you got faster. It worked. You know, what defines work? I think it's the same as self-trained athletes, someone that just goes and does their own thing and they improved. Great. A lot of people have tried that, done that, they followed their plan, they did whatever, and they got faster. and going to chat GPT is understandable. They asked the question and maybe they did improve. But the uncomfortable part is the improvement does not prove that it was a good plan. Almost any structure beats no structure. So an athlete that's training very little and trains more is going to see improvement. So without a real plan, the amount that you're gonna improve is very negligible. and it's not gonna be your potential. But any plan that produces consistency, some periodization, some intentional recovery is gonna produce something. So those gains are real, but they're not evidence that the plan was right. They're evidence that you're understructured to begin with, or that the bar was on the floor is very low, so anything. So improvement, just improvement itself, got faster, is not the standard. So the standard is how close you came to reaching your actual potential based on the time that you invested in the process. And so that's exactly the thing that you can't see. You cannot run a season twice or in parallel. Once with a plausible plan and the other one with the plan that's you know truly right for you. And you can't measure that gap. So you don't have that capacity. So they take the improvement, just some improvement. And they feel validated and they never find out what it costs them, what it costs them in performance or plateau or injury or excessive time spent, you know, that could have been spent elsewhere on more performance or with, you know, other parts of their life. They never see that part. But they're just satisfied with I improved.

Carrie: if we could only run those great seasons back twice. Exactly how it went, right? It never happens that way though. You cannot do it. Jeff, you've said the cost

Jeff: But that's the in in fit logic though, we actually can do that because we have, and that's part of the learning process is we have data from so many different athletes. And so we have, you know, you only have one of you, you know, but we have a bunch of athletes that are just like you across the tens and hundreds of thousands of athletes, and we can see when they do differently. Here's what actually happens, and that's the whole thing with the segmentation and the analysis that we're able to do. So we can see those outcomes. Here's what produces this, and here's what you know produces X, and here's what produces Y.

Carrie: Yeah. Yeah. Yeah. You've said that the cost is invisible. And I don't want to leave it abstract because I think this is the part that actually matters to people. When an athlete pours themselves into a plan that's rougher around the edges than they realize, what does that actually cost them in real life? Like where, you know, what can you how can you explain that?

Jeff: Well, some of what we just chatted about, just you know, the thing, the time that that you could have invested somewhere else. I think it's many, it's several different ways. So it's certainly more than one. It's not just the performance angle, but there's the injury risk that accumulates. And it's not a sudden like boom, boom, I'm injured. I did the wrong plan on day two, I got injured. It's something that quietly accumulates and it stays invisible until the day that it's not. You know, all of a sudden you get injured, but it's from you know, sometimes weeks, months, sometimes years. Of digging a groove that you can't out get out of, and then there's your injury and a chronic problem that you have to deal with for the rest of your life. I think there's the hours themselves. So you're training, you pour hours into the wrong work. It's inefficient work. that work could have gone to sharper, better sessions, delivering more results, more gains. if you've gone into a real recovery to allow your body to recover, or that same time could have been put towards something else, brought a better result. It could be Just the whole fact that you're paying more than you needed to for the same return. So and I think then there's a whole nother area that outlasts all of that. You know, you have seasoned races come and gone, but the part that outlasts that is the time that you spent training on the inefficient training that wasn't for you, that where you're missing dinners with the family, you're missing mornings because you're too torn down to be present for you know, the people are counting on you, your work, your career, your studies, you know, whatever that is. So you're trading parts of your life for training. and that training time was just roughly aimed at something is generic. So I think under all of this sits costs, I think without refunds, you can't get them back. So it's realizing you get to the end of the season or end of a few seasons and you realize, man, I could have spent my time much better. You know, I spent these the best years of my life, this best effort chasing my potential and I was chasing it on a path that was never meant or built or had the capacity to get me there. and you never know how much more was in you. You never know what your potential was. And so there's this constant and forever regret that you can't get back. And even if a plan looks right, sounds right, feels right, all of those things are just the surface. And the surface of that plan is all that a large language model can produce. So if you have you know AI. can define many things, but if you if you strip out the actual intelligence on training, the if you strip out the intelligence of artificial intelligence, all you're left with is the artificial.

Andrew: Yeah, so Jeff, a lot of really good information here, especially for folks like Carrie and I who have not used ChatGPT to generate training. you definitely caught us up to speed on the limitations of it, why that this is not the route to go for an athlete that cares about their training or any athlete really. So let's flip this here. If a language model structurally, by just what it's built to do, cannot generate a good viable training plan. What would a tool actually have to look like and actually have to be able to do to help us get our training right? Not sound right, but actually be right.

Jeff: Yep. So being right is not just a finding a smarter chat bot or a better prompt. It's a different kind of machine altogether. So it's built from training data on training data for the purpose of generating training. So it's trained on not trained on words, but trained on workouts. not just words about workouts. So it has to do three things. And we had a whole we had a I think a podcast a month or two ago about the first principles. And that kind goes through it in in a lot more depth than this, but if I narrowed it down, probably two or three things. One is that that this different model, the fit logic, can do that a large language model is just structurally, architecturally not capable of doing. And the first one, you know, basically simply put is just it doesn't know who you are. It to it you know, real individualization means changing the plan itself from the structure, not just the labels on it. Is based on your training history, your performance ability, your body composition, age, gender, your performance relative to others like you, and how fast you adapt, recover, and how those things change over time. So it has to know those things, not just use words that discuss those things or mention those things. Number two, it has to measure the training and what it actually costs you. So training stress is prescribing training stress. It's not just one quantity, it's many different quantities. your you know muscular aerobic threshold neural stress, those different stresses that clear the body differently, those come in different types that affect different people differently. And it has to understand that and to be able to individualize that and measure that stress after it's individualized to you so that that that is specific to what you need. And the can conditions matter. If you're running, you know, at a 45 minute run in 90 degree heat, it's much different than running in a 60 degree heat. And so being able to understand your conditions. and adapting your training, your intensities based on what you actually did so that you're not, you know, under counting a very hard day that was in the hot, but it might have been the same pace and now you're cooking, you know, and then the next prescription the next session that you get is way too much because you worked too hard the first day. So it's understanding how to prescribe training stress, quantifying training stress, our normalized training stress, residual training stress to give that to you specifically relative to you as an individual person, but also your conditions. and prescribing that. And that's the kind of the secret sauce, is that logic of understanding how all of those things work together? And that's what FitLogic is trained on. And I think the last thing that I put there is learning from what happened. So every conversation with a large language model starts over. You know, it was pre-trained before you got it, and then it just learns you start over the conversation. training intelligence is inherently longitudinal, so it's a closed It's a closed loop. We predict, we observe, we compare, we improve. and what comes back is, you know, it's not turning a it's not like a static plan, but it's turning that information into a system where the past weeks actually inform the next weeks. so with other platforms and chat GPT specifically, it's not capable of doing that. there's no mechanism to do that. There's no mechanism to weigh the outcomes. So we have to weigh the outcomes and we have to be able to predict here's what based on All of these things, your environment, your genetics, your you know, all these different things, here's what should come from your training or from your race. Weigh that, how did how did what actually happened compare to what was predicted to happen, and then you can learn from that. And then against that prediction is where you can get better. And so that's what ChatGPT and others that are not made to do this, they don't have the capacity to learn. They're you know, automations of a template, a methodology, or words about training, they don't have a capacity to learn and to improve. And that's what I've spent more than twenty years building FitLogic to do that is that the training logic that powers TriDot and RunDot. So ChatGPT was built to sound correct, FitLogic was built to be correct.

Carrie: Yes! I mean, I think that right there is awesome. That's what we were talking about a number of minutes ago. We were talking about fit logic and why it's so good. And it isn't just Chat GPT building the plan. It's learning us. okay, we're gonna bring it home for all the athletes listening, Jeff. Underneath the plan, I'm using that word now. Underneath the plan and the phases and the paces. Why did these athletes actually go to Chat GPT in the first place? And what do you want them? To understand about whether they got it.

Jeff: Well I'm to think about and remember what they came for. when you go to whatever you go to, Chat GPT or something else, you're not looking for a document. You're not looking for a piece of paper. You're looking for a result to be faster, healthier, to make more of your hours, to stay injury free. that's not a plan. That's not why you opened Chat GPT and went to it. You went to it for the result. And so it handed you back something that looks like a path there, that's organized and confident and detailed. But the one thing that you came for, that result, the training is right for you and it actually works, that's the one thing it can't produce. It can it can describe the result, but it was never built to deliver the result. So again, like as we start, it's not a knock on the tool. LLM's Chat GPT is a brilliant language engine. It was just handed a physiology problem. It's a it's a category error, not You know, that no amount of cleverness fixed. It's just a categorical error. It's a totally different type of machine. So what matters is so why it matters here more than other places is that close is not good enough. a plan that's a little bit wrong, when you hand it some, it's not just gonna give you a little bit of the results, but it's gonna give you something that could lead to injury, to overtraining, to a session a whole season that that's spent going sideways. So all of the small errors. Don't say stay small. They compound. So over trying over time, you know, the training has to be right, not roughly right. You know, roughly right is wrong. So the instinct that you have to tr to trust points the way. Reach for a tool that is built for the actual job that you're handing it. So doing the right training right requires the right tool for the right job.

Andrew: So Jeff, these LLMs, they get updates and all of a sudden one day on social media I'll see a bunch of people excited about like, you know, Chat GPT just updated the version whatever, or this this particular AI image generating tool just got upgraded from this version to that version, and people get excited about these tools all over again because they're more powerful, they're capable of more. If an athlete listening today Maybe they walk away from this conversation understanding, okay, here's the limitations. I'm not gonna go this direction for my training. But say they're tempted in the future to let a chat bot build their training for one reason or another. What's the one thing from this conversation you want sitting in their minds when they're faced with that temptation?

Jeff: run away, run away. No. I think that yeah. well think of it generates training that sounds right. It's not training that is right. And the reason that you can't tell the difference is the same reason you asked the question in the first place. So that's kind of the first thought that we started with. The second, I think that just the AI label. You know, the AI, the fact that it's AI or AI power tells you nothing. The question is whether the tool

Andrew: The Terminator.

Carrie: He's a Terminator.

Jeff: was built to understand training or just talk about it. And no amount of updates, no amount of improvements and the mod the new models coming out, that's not what they were meant to meant to do. So it's not an upgrade away. And then when you think about the stakes, like what's it gonna cost you? It's not the it's not the tool. The stakes are yours. You know, it's your health, your time, your potential. All of those things matter to you. And so give the same amount of care and caution that you would give your knee For a knee injury or an important contract, give it something, you know, to something that's actually built for that job. So if you want to see what you know built for the job looks like, that's fit logic as we described. It powers tri dot for triathletes and run dot for runners. So ChatGPT is generating plausible training. and you know, hopefully after all of this the conversation, the athletes will know what to watch for.

Carrie: Yeah. I loved it. I thought it was great. Thanks, Jeff, for everything. We're on to the cooldown. You want the cooldown question? No, you cannot. So we have one question for you from the audience, and that's from Coach Dennis. I don't know if it's my old coach. We called him D-Dog. But anyway, hi Coach Dennis. Thanks for your question. He asks, Will we see any changes to tri dot and run dot training based

Jeff: Sure. Can I avoid it? No.

Andrew: Yeah. Nope.

Carrie: On the Norwegian training successes. don't you know ya we are Norwegian.

Jeff: Norwegian.

Andrew: Ca Carrie, was that a Minnesota take on Norwegian? What? I don't are you okay. Okay.

Carrie: Yeah, well I'm Norwegian. That's how we talk here. Yes.

Jeff: Ha ha ha. So that's complex answer, but it's kinda simple at the same time. So a lot of thoughts come to mind. so I know you know, Christian, Gustav, there there's some amazing success at the highest level. Just yeah.

Andrew: And that's in triathlon, Jeff. In the running world you have Jacob In Ingebritson and a few others. so it's not just tr in triathlon and running and football at the World Cup, Norway did quite well.

Carrie: Ingebrettson, yeah. And it's Jacob, Andrew.

Jeff: Short much shorter. Yachum. Yeah. So I don't know that the that the World Cup may not be exactly correlating here to the endurance sports, but so I I'd say more address it specifically. You're not gonna see that what we've been doing for twenty years is not gonna change or be untrue suddenly because someone's doing something different. I think the way that we approach it one is always wanna

Andrew: Okay, thank you for that, Carrie. Thank you for that. Yeah, sure.

Jeff: To ask the question, you know, what exactly is Norwegian success? What is the method? What is the repeatable process? How do you actually define what that is? And then when you look at that success and the population, is it one, two, three, five? It's a relatively small number of people doing that. So what's replicable? Are there other factors? I know with on the triathlon side, you see Gustav and Christian with incredibly low sweat rates. And they train together, they push each other, and that's incredible opportunity. There's other factors that that impact. And so how do you isolate exactly what comes from that? So there's some challenges there with a small population of athletes. there are other factors that matter. And I think a big thing, and so we would want to see the numbers get bigger before we apply that, but the big thing here I think is elites are different than age groupers. And so a lot of times everyone is just sees them in the limelight, sees these incredible results. the thing that gets them that last one, two percent as an elite, as a world class, world record, is not the same thing that earned them the first 99% to get there. So the things that that you do at that level are and can be very different. and so a lot of team times you see that that correlation or that see the things that they're doing and want to transfer that. Okay, now all the age groupers, let's go start doing this. Well, I don't have that physique, that body, I'm 30 years older. you know, all of these things are very, very different. and so, but we're always want learn. And so from a fit logic standpoint, we've been looking at this data for 25 years, not just looking at it, but breaking all these things down. How do you isolate different metrics and abstract different factors and judge what is the correlating effect here and what's the causal effect of different things happening? And so we're always curious. So when we see that kind of thing, we want to look, we want to increase the population size. And what can we truly suss out as the true thing that's gonna make the difference? This is the difference maker, or here's a new approach or a new concept that that we can test versus just, hey, it's happening, it's observed, but we don't know why, and we can't replicate it. One of the things I would love to see is kind of more application of that to other non-Norwegians, maybe not that they're just Norwegians, but other elites. So when I see other elites doing the same thing and getting the same results, that would be very interesting. So before we're gonna take something that that different, I guess, to age groupers and to the mass population, I'd love to see others. So let's see what the Germans, the Dutch, the you know, these different people in different parts of the world, the Aussies, you know, if they start adopting some of those training philosophies or techniques and they start getting the same results, then it's the technique. Then it's the thing that they're doing. But until we see that happening, and even at that level, I think that's the first level that has to happen is at the elite level. Let's replicate it in in nearly the same environment. With nearly the same kind of physiology, great body composition, young, recover fast. Let's get all of those other variables the same. Let's replicate it elsewhere. And then we start seeing does that transfer down to normal age groupers and normal other people? But as long as it's just, you know, a relatively small number, even though you named off, you know, four or five on the running, that's relatively a small number overall when you when you think of everybody. So when we see that spread more, that's when we have more stats. And unless you have those numbers and until you have those numbers. it's very hard to isolate the true cause, just because there's so many other factors.

Andrew: Yeah, how would you at this present moment take what Kristian Blumenfeld is doing in his training and apply it to Andrew Harley, thirty eight years old, not nearly the same athlete in Dallas, Texas? It's apples and oranges, right?

Carrie: Yeah. Yeah.

Jeff: Yep. But it's fascinating. We always want to learn from it. We're always watching, we're always doing things. I mean, we're doing a ton of things that no one ever did before that from this very thing. It starts with the concept, the thing, and you start watching it. And then you let that the he data build and grow and you start and develop the mechanism and the means to prove the efficacy of whatever the theory or the concept is. But no doubt they're having great results and kudos, applause. They're incredible.

Carrie: Tough.

Andrew: Yeah.