Agreed with points above: TR’s AI has some ability do adapt and “learn” based on what you’ve done in the past, but it’s not a real coach.
For example, it’s very locked in to your weekly schedule, e.g. rest Monday, intervals Tuesday, endurance Wednesday etc. I you miss a workout or do an extra-intense ride on a recovery day, it doesn’t REALLY restructure your week in a meaningful way, it will at most drop the intensity of your next workout. I’m not saying it’s not a useful (probably the most useful currently) training system, but it’s not actually smart.
A dream scenario for me would be the ability to tell TR “hey I have travel/conflicts on Tuesday and Friday next week, but some extra time on Wednesday” or “I need my rest week to fall on X week” and have it re-structure based on that.
This would be really awesome IMO. I find it easy enough to shuffle around days when I have conflicts (although doing it automatically would be cool too), but having Plan Builder automatically adjust things to not always fit the 3:1 structure if you need rest weeks on certain weeks would be nice (e.g. planning around busy or less busy weeks at work).
Was thinking about this and I feel the best was to “Cheat” so AI works for you is to “Override” your FTP and up your number by 5 to 10%. Then the program will force you to “AMP” up your training efforts beyond what TR will suggest with a lower FTP.
Should like a good idea?
Not sure if you’re kidding, but that sounds like a TERRIBLE idea. Your FTP is supposed to be an indicator of your aerobic capacity, not a target metric to reach for. Giving the software an incorrect FTP will throw off all your zones and mismatch your training. There is a much higher probability of overtraining, burnout, suffering (sweet spot becomes threshold), improper progression (you can’t progress TiZ in threshold as fast as in sweet spot), and a ton of other ideas.
If you want harder workouts, use the more aggressive plan settings, or choose some alternates with higher PL’s, or something else. But using a too-high FTP will bite you… HARD.
Way too many of us, including me, have learned this the hard way because the ramp test would overestimate FTP for those of us with strong anaerobic capability. There is NO question that this is a bad idea.
So let’s get real here.
There is no number, no AI FTP calibration and no calculation or gene for determination, perseverance and the human spirit.
We’re all built different, each of us with different capabilities and levels of effort and pain we are capable of sustaining.
So to simply “bet the house” and place all our precious time training towards an end goal based on TR “AI” is foolish.
I can assure you that any serious elite level bike racer that won a major accomplishment never once looked down at their Garmin to check on their Power or Heart numbers while on the way to the podium. Each of them dug fucking deeper beyond anything that some computer generated “suggested effort” could prescribe.
Placing too much emphasis on AI will eventually “dummy down” our free thought, our original thinking and encourage mediocrity…not just training but in most things we do.
The most important muscle in our body is our brain and we need to spend more time listening to that than what some program tells us.
Sure TR is a good starting place, but listen to what your body is telling you and you will achieve much better success.
No, this was a tongue in cheek comment - but let me get in the weeds so
I completely agree that, on race day, a winner (and at every level) is someone who’s able to surpass their own limits, leave it all on the table when it matters.
I however disagree with the idea of “bumping my FTP” for training.
Riders used to train using RPE, then heart rate, then moved to power, for a reason.
Being able to target specific adaptations, from a relatively young age, using power meters, is one of the reasons we have such a crop of amazing riders coming to the front, and why performance, across the board, has improved so much in cycling.
This implies setting up some levels/zones where you can accurately target those adaptations. And while FTP is an imperfect metric (that TR is somewhat addressing by using the progression levels) it remains an efficient way to define those - for the largest number of people (which is what a platform like TR aims to do)
As for the AI comment and following a program - Do you dummy things down by hiring a coach and following their plan for you?
Not everyone has the sufficient sport physiology knowledge to be able to self train, self assess, etc.
I fully concur with the fact that people need to retain a critical view on things, and assess whether or not it’s working for them, but TR is, in my view, not only a good starting place - it can be used in conjunction with that critical spirit to great effect
I had a coach for 3-4 years…did a good job but TR is better. Not perfect but it does the job. Personally I can’t operate (train) unless something is put in front of me to do. I simply don’t want to think…I just want to do what the program tells me to do.
Personally after many years of training and racing my body is telling me that I should be doing greater more intense (vomit-like) sets. I’ve made all the appropriate adjustments to my program, removed the Masters option and still I find when I race I can’t get my “Turbo” going. Or maybe I’m just getting old and my body is saying…he dummy that all that’s in the tank. Be happy and start enjoying what 99% of others never dreamt could ever do.
From my observations using TR off and on over the years (no inside info) here’s my take on where things are at:
Adaptive training plans: not much AI if any, this is just a complex series of “If X do Y” statements that use conventional training load rules, progression levels and workout feedback and workout pass/fails. The logic is probably very complex by now and a big pain to maintain!
Progression levels: No AI, most likely some semi automated admin tools that grade workouts and also check compliance to verify.
AI FTP: quite possibly some element of AI, but I doubt the model is improving itself or you’d get constantly shifting numbers (which does get reported but not that often considering).
Red light green light: Seems to be about 90% conventional training load logic with maybe some small element of personalised AI/machine learning stuff. Also gets used in the adaptive training logic (e.g. no hard workout on a yellow or red day).
Only posting this becasue I get slightly annoyed when gullible people assume there is some magic AI black box (aka the TR AI secret sauce) guiding every TR plan. Oh and the “AI sparkle” icons that get used a lot annoy me too. TR is not that smart, but I personally still find value in the software, mainly using AIFTP to set an FTP without having to do maximal efforts and I also find TrainNow useful for workout selection.
When you race, you should know what your body is generally capable of, and you should then dig down deep to try to exceed your estimations.
When you train, this is NOT the recommended way to go. When you train, you should ABSOLUTELY be basing things on a plan based on good training science and methodology, and those plans ABSOLUTELY are based on power, RPE, or other metric intended to gauge your effort and work particular energy systems a particular way.
Neither TR’s AIFTP, nor any other metric, is an “end goal” toward which we are training. They are metrics we use to bracket our effort levels and time-in-zone at a particular time, so we get better results.
You go train all-out and vomit frequently if you like. The science and training history of the world’s best athletes indicates this is NOT the best way to train for most people, but there is always individual variation and preference, so there’s no reason to say it’s a bad idea for you. It may even be the BEST training methodology for you: let’s just not generalize that to everyone else.
Very thoughtful reply thanks, however I’m thinking your opinion could be derived from a background of less real life competition experience. My opinion is solely based on 4 decades of racing all at different stages in my life. I’ve seen and experienced things that I would consider “Intrinsic”…well beyond a calculation.
To add to that, we don’t know how complex the models are. But they are trained on cohorts of people and not individuals. Most individuals wouldn’t have enough data to train a model based on them alone. And even if we could, there’d be no guarantee it would be better.
But there are other factors and questions to consider:
What is the goal of RL/GL, is it to maximize consistency?
Is it to nudge you in the right direction as far as what it thinks you should do long-term or is it mostly concerned with the next few workouts?
The time horizon and the aggressiveness are two important parameters, and you can tweak the latter.
The most difficult thing is not “the model”, but the choices behind it, goals, relevant variables, statistical significance, etc.
With experience you can override TR’s algorithms, but then you’d need experience with how TR’s algorithms react without you overriding them. For me the initial PLs for power ranges I haven’t trained in a while tend to be too low. But RL/GL works well. There are factors like sleep or (very occasionally these days) significant alcohol consumption.
However, you should also be cautious with your experience. At age 44 I no longer recover like my 25-year-old self.
I agree, and I think that’s a deliberate design decision as overcooking an athlete has more serious consequences than leaving a little on the table.
You can adjust how aggressive AT should be in two ways, by changing your previous experience in Plan Builder and by changing the aggressiveness slider.
IMHO that’s unlikely. Not sure what your background is, but let me try to explain why I disagree:
Machine Learning = statistics + fitting a function under constraints. The parameters have various names (e. g. weights in neural networks), but mathematically, these are just implementation details.
The sophisticated algorithms ensure that you e. g. do not get stuck in a local minimum and that you search for what you hope to be a global minimum efficiently. This is a hard computer science problem, but mathematically speaking, these are just implementation details.
Training ML/AI models is not hard these days. Judging by my own experience, I expect it to be much easier than writing algorithms that do these things by hand. Once you have selected the data and decided what to optimize for, doing ML is really, really simple.
The hard parts are data set curation, pre-processing and determining what to optimize for. The latter especially requires a lot of domain expertise beyond ML.
What is and isn’t ML/AI is hard to say. About 10 years ago, the field was known as Big Data and it included e. g. lots of sophisticated methods from statistics. They may be less sexy, but still very useful.
Often the most difficult thing is to pre-process the data, decide which data to include (e. g. including strongly correlated data may result in worse predictions), and, importantly, decide what quantity you want to optimize. I cannot overemphasize this point enough. In my experience, if you start from scratch, data set curation and data pre-processing may take 60–95 % of the time.
Once you have experience with pre-processing (which is the valuable institutional knowledge) and the infrastructure to support it (super important as that takes time!), the focus may shift more towards the training side and judging whether the output does what you want it to.
Some factors that must enter pre-processing are (1) excluding bad heart rate data (dropouts, stuck heart rate data, etc.), (2) how to handle different power meters, and (3) how to deal with “optimistic” responses from workout surveys. I obviously don’t know how the sausage is made inside TR, but I’d be extremely surprised if TR’s algorithms didn’t take those factors into account.
TR is not that smart, but I personally still find value in the software, mainly using AIFTP to set an FTP without having to do maximal efforts and I also find TrainNow useful for workout selection.