Perhaps this year will be my final year focusing on a full triathlon as a green banana. After five years of training for an 11-plus hour race, it has been enough. Unless I qualify for the Ironman World Championships in Nice 2025 …. And if I succeed, then I will train one more year for a full distance race, using a new approach, using Artificial Intelligence (AI).
In early 2018, I used a genuine training plan for the first time in conjunction with my Polar watch and TrainingPeaks. The plan came from 8020-endurance and I had to memorize the workouts, because Polar synced to TrainingPeaks, but did not import the workouts to the watch. In recent weeks, three schedules were sent to my Garmin watch, two of them based on AI.
In 2018, I wondered if I could do a half triathlon, so I looked for a schedule and followed it as best I could. That worked out just fine. The apps I use now tell me that I will finish between 9:45 and 10:39 over the Ironman Cascais. If I follow the schedule correctly. The computer says YES! (One a little louder than the other).

Based on my own experiences and some reading, watching and listening, below a discussion of two training apps based on AI. I compare TriDot® and Humango to the 8020 method combined with TrainingPeaks, as an athlete and as a coach. I will only discuss the main points and obvious differences, as I used the apps. I will leave out some of the apps’ options to keep the text readable. For example I’ll skip the formulas behind the training scores, do I not explain genetic profiling ( Physiogenomix™) and in what way health data (HRV and sleep) are incorporated.

Upfront for transparency:
TriDot® gives me 30% of revenues generated through my links that do not include coaching. I pay 10%-30% remittance for coaching from me through TriDot®.
Humango gives me volume discounts between 20 and 50% on All star subscriptions I pay for. Furthermore, coaching with AI mainly provides me with extra time to coach better, I’ll come back to that later.

Summary, conclusion and advice

For those with little reading time, the very short version.
TrainingPeaks is by now the old-school way to analyze your training data. Great for data nerds. Good to use by and with a coach and/or a static training program. Only the thresholds and thus your zones keep changing. It is the improved version of a training schedule in Excel.
Humango puts fitness and “goal readiness” at the forefront, is fairly lighthearted, nice and funny (also featuring a drill sergeant as coach robot “Hugo”), with plenty of data to look at, and fun workouts. It does, however, do some strange things to your values sometimes (see further). Fine for most athletes.
TriDot® is very solid, comprehensive and more expensive, with the proper execution of workouts at the forefront and is strict about it (you get a unicorn if you succeed). For the demanding athlete who wants a complete package, for example, for training towards a big green banana, such as an IRONMAN, marathon or trail.

Table 1 lists the characteristics of the three platforms, as they themselves promote them, with my comments in (Dutch for now).

In depth

• So, let’s proceed with an in-depth review. Successively I will discuss:
• What is AI and how can you train with it?
• How do the apps underpin their approach?
• This is followed by the use of the apps with
  • the planning and arrangement of training phases;
  • the workouts and classification of training zones;
  • the way they (try to) motivate you;
  • race plans and predictions.

AI and endurance training

Algorithms have been around for some time now, and TrainingPeaks and other platforms use them. The next step in app development, after deploying algorithms, is to automate, or better generate, that which we (coaches, athletes) were still doing ourselves. Briefly from Wikipedia:
“Artificial intelligence (AI) or artificial intelligence (AI) is the mimicking of human skills with a computer system such as learning, planning, reasoning, anticipation and independent decision making without the intervention of a human intelligence.“

In a webinar for TriDot coaches, the CEO of the company behind TriDot explained the evolution of data use in endurance sports using the image below.

He calls the current phase “data intelligence” (DI), which is more than just AI. For example, apps also use “machine learning” and “data science” to generate the plans and predictions.
Training with DI means that the platform can show more than just data. The platforms can also actively monitor your progress, adjust and prescribe your schedule, predict your fitness and finish time, and thereby motivate you.

Underpinning (sla over)

Training and training effect have been studied for years. The well-known training zones also originated from science. The scientific discipline on training is broad and examines psychology (emotion, motivation), injury prevention, performance, physical development, health, nutrition, etc.. As a scientist myself, below I will start with the foundation of the three platforms I tested, taking TrainingPeaks and the 8020-endurance method together as one.

As it is customary in a scholarly review to begin with how you have made the selection of sources, let’s start with that.
I have choosen TriDot and Humango because the two now partner with the two largest triathlon racing organizations: IRONMAN and Challenge, respectively. There are, of course, many other platforms, such as TriQ, AI Endurance, 2Peak and MOTTIV. Big in cycling is also Join Cycling. Some platforms are more advanced than others, sometimes with swimming, sometimes without; sometimes with running on power, sometimes without.
Athletica.ai I will discuss separately, because in the user forum specifically, the scientific foundation is an issue for users.

When I was looking for a schedule for my first half triathlon in 2018, I was very enamoured with 8020-endurance’s approach, especially because of its foundation. That foundation, written down in the subsequent books, came from both large-scale research, convincing case studies from the sport and the authors’ own experiences. So my first question to the TriDot salesperson was what their “model” or endurance sports philosophy was. “None,” was the answer! At least, there was not an explicit philosophy from sports science, but rather from data science. Before I get into that, let me briefly discuss different ways of underpinning.

Succes stories

Development in sports (as well as in other disciplines) is often through the successes of athletes who win and succeed. Sometimes because winners tried something different, sometimes they didn’t. All the platforms I’ve seen use this: “Top athlete A does this and wins. Read her success story with platform Y here.” But scientifically, this is also a well-known way to substantiate the effects of a training approach. Researchers select a group of athletes, determine the degree of success and examine the (differences in) training method. 8020-endurance and proponents of similar approaches, often use these studies. Athletica also refers to this type of research. Humango and TriDot not so much explicitly on the surface, but deeper into the explanations and support you will come across this way of underpinning.

Scientific experiments and reviews

The traditional research methodology to determine the effect of a training approach is experimentation, or writing summaries about many of those experiments (meta-research and review studies). However, it often proves difficult to use large samples for this purpose, so building it into a model for “all” athletes, especially amateurs and recreational athletes, is always difficult. Meta-studies try to overcome this by taking many studies together, but basically have the same problem.
As with the success stories, 8020 and Athletica in particular use these sources.
Interestingly here, Athletica’s approach, High Intensity Interval Training (HIIT), often does not emerge as most effective in this type of research. HIIT does work, but approaches like 8020 (polarized training) tend to be more effective anyway. (I did a quick search and found that outcomes in the articles of the most cited and highest ranked journals, see for example this and this one). Subsequently, I did not try out Athletica.

Examples (case studies)

Individual success stories with an emphasis on the athlete’s process and experience are hardly told on the websites and explanations. Except for the user testimonials as a marketing tool. 8020, however does do so explicitly on their blog and in the books that accompany it. That helps to understand how an approach/workout can work for you. If you look like the example of course. From the “hyper-personal” training plans of AI (as Humango calls it), of course it makes sense not to use on this type of underpinning explicitly.

Data science

Since AI, or better DI, new as substantiation is using data sciences. At first glance this looks like the success stories and also the meta-studies, but it is something different. In fact, rather than groups of athletes or samples, it involves data from as many athletes as possible, of all types. The data sciences analyze that data, remove the irrelevant data and look for connections between the data(points), build models that can predict effects, analyze those again and learn. Basically a completely different branch of science. TriDot is the best example of this. That is why the first reaction to a question on substantiation was also ‘none’, because it was not a classical, sport science substantiation or vision.
Such a denial of using a model behind training is often typical of how AI/DI is approached: “the data determine it”. Like that would be neutral or even objective. Often even the designers of self learning algorithms do not know exactly what they are doing, they are a black box (for that listen to this podcast). Yet in the design, specifically of analyzing the data, choices are constantly made in what are and what are not relevant or reliable data and what is considered successful or effective. Then again, those choices are not made explicit out of commercial interest. In a video interview with someone from TriDot, they later did tell me that the analyses they had done showed that many elements of polarized training are effective and thus reflected in the workouts and training plans.

Soft sciences

What strikes me, as an educational scientist and educator, is that the preference for underpinning is for so-called ‘hard sciences’. I can find little about psychology of (online) coaching and motivating athletes or dealing with setbacks and injuries, goal setting, etc.. It does get talked about in coach communities and certification, but is not used as a rationale for why what graph shows up in the app. Or how the game element of scoring a unicorn in a well-executed workout motivates. Or why Hugo the coaching robot says what he says. There is science about that, too; there are theories about that, too.

In summary, I juxtapose the of each of the three platforms below which rationale they use most.

Using the apps

Now on to how the apps/platforms work in practice. On the two AI/DI platforms, I have had an account for about three months now. On TrainingPeaks, I have been using schedules from 8020-endurance since 2018. I have juxtaposed those below on various aspects, especially on those in which they differ.

Planning and layout

Training for (big) endurance sports challenges means planning well. Not only in your week between work, family, parties and celebrations, but also in the longer term. After all, you need to make sure you are optimally fit on race day. To plan and monitor that, there are a number of values by which you can see progression and fitness. It is too long to explain them all here. The table below lists the terms used by each platform, so that I can then use them to say something about planning and progression.

Let’s start where I would end up, when I follow the training plans, on the day before the race, based on my training history on Sept. 29. The platforms off course assume in forecasting that I follow the schedule. First, the visualization.

In TrainingPeaks, my fitness (CTL) on October 18 is 96 and my form (TSB) is 1. Roughly speaking, for a full triathlon 100 and around zero are recommended, respectively. Fine on target, but beware, it is a rather rough approximation. What you can see here is that TrainingPeaks is more of a monitoring system, because CTL (and TSB) are calculations over the previous six weeks, with the last week counting more heavily. Thus, the values say more about the load (that you can handle), than the performance you can achieve.

Humango’s dashboard shows that my projected readiness is 79% the day before the race. The chronic workload, similar to CTL from TrainingPeaks is 69. The fact that I don’t score 100% is probably because I have hardly done any workouts from Humango. In addition, you can see that Humango generates many parameters, but the correlation is sometimes difficult to understand, that is for example how exactly does CWL relate to readiness?

TriDot calculates my readiness at 7.6 on a scale of 1-10 (on race day 8.5). Pretty much comparable to Humango’s readiness. You can also see that I would have end up at 10 two days after the race. That’s crazy, actually, because if AI/DI optimizes and does so in “real time,” you would expect to end up at 10 or 100% on race day. The fact that that is not the case with both Humango and TriDot is probably because I did not start following their training plans fully and precisely until late.

The plans achieve the aforementioned outcomes by using different phases of training load in their planning. That’s called periodization, for which different models are used. Traditional is the linear model of the season: you start with a lot of volume at low intensity and that changes closer to the main race. The IRONMAN coach certification promotes this model in their standard training plans. The other model, or models, are non-linear. This usually involves dividing the phases into blocks that have a specific training objective and can also be adjusted depending on the number of races and type of races the athlete is going to do.
8020-endurance, Humango and TriDot all follow a variant of a non-linear approach. TriDot opts for ‘reverse periodization’, where intensity (and thus speed) is mainly worked on at the beginning of the season and the ability to maintain it is built up through the phases. This is also becoming increasingly popular in cycling, since successful cyclo-cross racers, who do a lot of short hard efforts in the winter, are also successful on the road.

In TrainingPeaks, I myself used roughly the periodization of 8020-endurance, but taking into account many shorter races during the season. The high volumes I was able to do in the summer, so then the intensity was a bit lower then.
Humango has included in my plan of three months still almost all phases, taking into account the other races prior with the main goal. Noticeable by the multiple taper phases. So from there we can also conclude that Humango does have a view on endurance sports training in addition to data sciences. Which makes sense, but then again, to include both the base and build-up phases is not so logical (although you can adjust that manually). Moreover, they do have my data from the whole year, subsequently Humango could analyze which phase(s) I have already gone through. TriDot seems to do just that. In my plan there is only a race preparation phase of 14 weeks. In other words the analysis of TriDot is that I needed to start working towards the duration and intensity of the race and the building phase has been done.

Workouts and intensity zones

From the long term, now to daily practice. For the week of Sept. 9-15, I juxtaposed the schedules of the three platforms in hours and in training load with the measure that each platform uses itself.

Again, we see that Humango is the friendliest, or rather most careful. My own planning over reaches a bit because I had put in a 145km bike ride from Roermond to Bergen op Zoom. Anyway, in practice of scheduling volume per week in the month before a race, the platforms do not differ much. The training load differs a bit more, but that too has mainly to do with the methods of calculation.

Before you can actually get started working out, you have to somehow know what the workouts dictate. A peer student of mine over did that 30 years ago by printing out blocks of a workout and sticking it for each athlete in a sequence on a piece of paper. He still coaches and today he sends PDFs created in Excel. That’s a way and does work.
With apps and devices, it works differently: you have to pair your watch (or bike computer) with the app. With all three platforms, you have plenty of options for that. I list them below. Usually you can sync back and forth directly between the device and the platform. Sometimes this does not work and a detour via Strava is needed. Besides through your wearable, all platforms can talk to other training apps, such as Zwift and Rouvy for cycling. TrainingPeaks can further incorporate data from nutrition apps and scales. TriDot can do the same with FuelE.

And next, you go for a nice swim, bike and/or run. This works the same for all apps because it depends mostly on your watch, memorized or printed out schedule. Sometimes the title of a workout almost tells you what to do. All three platforms have a system for naming workouts. 8020-endurance usually chooses to include the goal, physiological, in the name (abbreviation); Cycling Aerobic # (CAE#), where # stands for the number, i.e. the length. It takes a while to figure that out, but worked fine for me in recent years.
Humango alternates in naming between the goal (endurance run), with the designer’s name (Coach Rae’s start slow, finish fast). TriDot only has names that make it clear what you need to do: tresholds 200/300’s.

The biggest difference between the apps in the practice of training is in the way they set the training zones, the intensities. Those zones are used as goals (targets) in the workouts. I found switching between the three systems the hardest. In this you most strongly notice the differences in vision and approach of the systems. I was used to the 8020-endurance model, a variant of polarized training, where 80% of the training takes place in zones 1-3 and 20% in zones 4-5. In doing so, they avoid what they call the “moderate intensity rut”: that area where training just feels tough, from which you then have to recover relatively long, but the training itself has little long-term effect. Those intensities are zones X and Y.
Because I have done so much low-zone training for so long, I was shocked to find that I had to train much more in zones 2,3,4 on Humango and TriDot. Added to which, zones 2 in those models are just a bit more intense than in 8020’s model. See below my zoning for the three platforms (on August 24 in Watts for running).

First thing that stands out is that only TriDot opts for six zones. There’s something to be said for that, because workouts for even higher intensity don’t really need a distinction between very hard and even harder. TriDot furthermore very much agrees with 8020-endurance, but you still notice the difference in workouts. TriDot asks you to do very little in zone 1. Even the warm up goes straight into zone 2, whereas warm ups of 8020 are in zone 1 and long endurance (aerobic) in zone 2.
Second thing that stands out is that Humango estimates my FTP (the functional threshold power, above which you acidify) to be much higher. That is, I suspect, because they continually calculate it based on the last workout. So it’s not a final measurement or test. In fact, if you look up what your 60-minute value is, as that is the definition of FTP, you will find a different (lower) value. That makes it easy to get pretty intense workouts, although they don’t schedule them as often as TriDot.

Feedback and motivation

In addition to the promise of workouts specifically and uniquely for you, which also constantly adapt, the AI/DI platforms also promise to keep you motivated. After all, they are “…designed to improve athletic motivation and performance,” according to Humango. TriDot especially helps you do the “right training right” and “measures and motivates adherence” I’ve already written more about motivation on my website, since I’m particularly interested in how this works. As mentioned, the motivational aspect of the approach is not further substantiated. What I have tried to do is identify how the apps are trying to motivate me based on the feedback they provide

Motives
The attentive reader, especially those in education, will notice that I am trying to get around the word motivation. Motivation can mean so much and so little, which is why I prefer not to use it. The theory I use in my educational research, and thus here as well, speaks of motives. Literally translated, reasons for moving. Since I want to (help) encourage others to get moving, it is an appropriate term. I call them bananas, green and yellow, long-term and short-term motives, respectively.

The most visible feedback is the progression that all platforms show (the data visualization in the first image) and the projection of progress in the future, if you are going to follow the schedule: the fitness or readiness you are going to achieve.

Fitness/readiness

Above you can already see how the apps visually monitor your fitness and project readiness. TriDot additionally shows in the same graph the extent to which you deviate from the plan. That can also be found in TraininPeaks and Humango, but is more hidden. So apparently, for them that is not used to motivate. For me, seeing your deviation from the plan does help, especially to see if what I feel is correct: did I have a good week that week (did I do the right workouts right?) and is there an explanation for lower fitness than expected?
The obvious way to know if you are making progress in your workouts is to determine if you are going faster. However, ‘going faster’ (swimming, biking, running) depends quite a bit on conditions, among other things. The common measure to overcome this is the threshold: the power (or pace) that you should be able to sustain for exactly one hour. Usually called FTP (see for more explanation here). Hence, you could see the change in your FTP as evidence for your progress.
In TrainingPeaks, this is a little hard to find. You have to create some tables/graphs in the dashboard yourself for that, or search for your 60-minute value in a workout and compare it to previous workouts. However, TrainingPeaks does automatically signal a change in threshold value and then sends a message (to you and your coach).
In Humango, you can see your FTP immediately on the home page under “My performance” and its development in the graph “Treshold history”. In addition, you can also find a 60-minute value. But as it turns out, those two values are not the same! The explanation is that the treshold history shows the calculated value, so based on most recent training. The 60-minute value shows the actual highest measured value that you sustained for 60 minutes. That’s not the same thing, because based on 45 minutes of workout and a calculation, you could end up with something different than going out 60 minutes full blast and took that value.
TriDot only sets your FTP after tests, which they have you do every four weeks or so. The platform is not going to adjust that threshold based on values in workouts itself. This again shows that TriDot is all about clean and precise data: only values determined according to the testing protocol count (although you can also enter your FTP manually).

Compliance

The second obvious motivator is the way the apps show how well you performed the workouts: compliance. TrainingPeaks shows this by simply giving the executed workout a color on how exactly you performed. Red for a missed workout, orange for done but pretty off target, yellow almost on parr and green for a perfectly executed workout. You can set up exactly how TrainingPeaks determines that, but by default it’s duration and intensity.
Humango immediately shows in the dashboard how well you executed the entire schedule (plan compliance) and indicates with a circle graph, a check mark and a percentage to what extent you have performed the training well. Coach-robot Hugo then also comments, for example: “Well done, another workout done. Was the training too hard, or did you have too little time?” Feedback perfectly according to the principle of ‘first a tap on the shoulder, then a kick in the butt.’

TriDot gives points per workout (session XP) of which you can then see how many of them you have achieved (e.g. 32/42). If you get all the points for the most important training session of the week, you get the full amount: you get a unicorn! Unfortunately, I haven’t received one yet. I did receive a ponicorn because I did a less important training perfectly. In addition, each workout, and each week and month, have a Training Execution score (TrainX). So the most important workouts in the week have the highest TrainX score from 0-100. You also have a personal TrainX score and that score shows your progression.

Gamification

In learning and teaching, especially when connected to apps and the Internet, gamification has been a way to motivate for years now. Meaning that elements of games are used to encourage doing assignments and ensure that participants stay on task. Badges, points, extra lives and rankings etc. are used for this purpose.
TrainingPeaks has medals (bronze, silver gold) in addition to compliance, for thresholds you score (third fastest time at 1km last year, for example). That works to motivate and show your progress to some extent.
In Humango, you won’t find many game elements. Apart from the coaching robot, it is mainly the concrete data that should stimulate you. However, Humango does have a social function, allowing you to work out together in groups or clubs. You can also see what the athletes you follow have done, just like with Strava. That stimulates, of course.
In TriDot the gamification is more obvious. There are the XP points and TrainX scores and the ponicorns and unicorns you can win. You can also compare yourself to others, at least the average of others like you, using the “dots” (hence the name).
Per discipline (swimming, cycling, running) see a score between 0-100 based on the threshold you have. So you can also see from that, for triathletes, what your stronger and weaker parts are, relative to your peers with the same age and gender.

Race plans and predictions

For five years I have been trying to estimate my finish time for full triathlons with varying degrees of success. Not so much as to set a performance goal, but mainly to plan the race in terms of nutrition and for possible supporters. My swim is pretty steady and for the four kilometers or so the deviation is not more than 10 minutes. Besides, you can’t eat and drink during the swim, so planning for that is easy. Cycling also worked out reasonably well, with tools like Best Bike split, because using a coursemap and your FTP and margins of no more than 30 minutes, you don’t just run out of gels either. Running is much more difficult, especially because running on power is still new and there are few tools available (Stryd can do that for example). Still, you are not there yet, because all those separate estimates per discipline give a very large margin for the whole (up to two hours).
If there is anything AI/DI has to offer it is the predictions it can make. Hence Humango and TriDot clearly put that feature prominently in their platform. At Humango there is a special tab, seasons, where you can put in the A,B and C races and see what the prediction is for those races. You can upload the courses and indicate how hard you want to go (easy – very hard). Taking into account your readiness and FTP, an expected finishing time is then generated: 9:43 if I go very hard on all parts for my IRONMAN Cascais!. You can play with it a bit by changing the intensity on some parts differently than others and see what effect that has on the finish time.

TriDot has RaceX as a race planner and predictor (with the more expensive subscriptions). That planner has more features, also allowing you to run through what-if scenarios for cycling, similar to Best Bike split. You can adjust your cycling position or wind direction or weight to see what effect that has. Also, the average power you want to pedal, which is similar to what Humango can do with intensity, but more precise. TriDot thinks I will do 10:39 over the race. Still half an hour faster than my fastest triathlon in the lowlands, but already a little less intimidating. And there is “only” a one-hour margin between the two platforms. That will please my father on the finish line for his planning of drinks on the terrace.

So no more need for a coach?

Well. Maybe not, but that would also be the answer for someone who likes to dive into TrainingPeaks data. Besides, most platforms that come with sports watches and cycling computers have training programs. Those also work fine for many. The question is not yes or no, but when and why is it better with AI/DI? And there is no short answer to that that applies to everyone. Good to know is that the CEO of TriDot began the creation of the system as a swim coach, because he needed data to coach better. As such, TriDot is also and primarily built for coaches. Hence, there is no coach robot, as with Humango.

A coach, as far as I’m concerned, helps to keep moving. For some, that means helping with goal setting and race planning, for which a good schedule is then created. For another, it’s more about receiving feedback while following the plan and being encouraged to trust the outcome of the process. For yet another coaching means a combination of those things and more.
How AI/DI can help me as a coach is quite practical: I am less likely to forget to fill a training plan and the customization I build into the schedules is now done for me. So AI/DI allows me to focus entirely on coaching: finding green bananas, encouraging, explaining the process and data, etc.. Does that sounds like a marketing slogan? It is.

Summary

The added value of AI for endurance training is mainly the more precise match of training plan and individual workouts to the athlete and predicting your finishing time for a race. In addition, platforms with AI use your data in a smarter way to visualize and motivate you.
TrainingPeaks, without AI, meanwhile, is the old-school way to analyze your training data. Nice for data nerds. Good to use by and with a coach and/or a static training program, in which only the thresholds and thus your zones keep changing. It is the improved version of a training schedule in Excel, where progress is a number (a CTL of 90, for example).
Humango puts fitness and “goal readiness” at the forefront, is fairly light-hearted, nice and funny (also featuring a drill sergeant as coach robot “Hugo”), with plenty of data to look at, and fun workouts. Does some strange things with your values sometimes, and also encourages with a social feature similar to Strava. Fine for most athletes.
TriDot is very solid, comprehensive and more expensive, with the proper execution of workouts at the forefront and is rigorous about it (you get a unicorn if you succeed). For the demanding athlete who wants a complete package, for example, for training toward a big green banana, such as an IRONMAN, marathon or trail.

Offer

Do you also want to start training with AI? You can now do so at the Banana Shop. Get in touch by emailing to martijn@mvanmartijn.eu.