I’m watching this forum become dominated by discussion of AI FTP prediction and its many frustrations, and I wondered whether the TR team have any information on or concern about the effects this has on adherence to its training plans and protocols.
I use adherence rather than compliance advisedly, as I’m NOT talking about how well one follows the selected workout–extending rest intervals, pushing too hard on certain work intervals, etc.–but which workouts are done, when they’re done (day of the week, time etc.), whether they’re done at all, or just deleted.
I can’t be alone in having responded to the wild fluctuations in AI FTP prediction by playing with the system, moving things around and back again, changing feedback survey ratings to see what effect it has. So I’m thinking that at some point a loss of trust in the system, attempts to game it, becomes a problem as it would with a human coach. I’m seeing my ‘coach’ promise big gains, then when I do as s/he asks s/he says I’m not getting anywhere. So I’m at risk of entering a circle where I trust them less, so follow their suggestions less, so see fewer gains, so trust them less…
I wonder what others think? What’s the thinking from TR? Are you seeing this? Would you be able to notice it in your data? Have there been more refused AI FTP detections? More changes to the workout calendar?
i’m replying because i didn’t want you to feel like you were screaming into a void and that further reduces your confidence in the product just because you feel like nobody is listening. We are. There is a spectrum of new AI features and there is a spectrum of human behavior. The complexity of trying to get those two to merge effectively is hella hard and i think you’ll get many n=1 folks at most points along the way. There are many struggling and trying to resolve to a satisfactory working system and they appear to be vocal here on the forum. I would also assume there are many perfectly happy that aren’t here on the forum and some that have simply gone elsewhere or disengaged to some degree. Timing is also tricky because many folks head outdoors this time of year and would naturally amend their riding habits and TR usage for that reason alone. It feels very murky at the moment for some of us.
Definitely in that same boat, I’ve seen a steady decline in my belief (and FTP) with some very poor workout choices from the ai leading to more workout failures in the first two/three cycles than I’ve had in most seasons. I’ve generally gone with what it’s suggested so far but experimented with responses, training approaches and workouts to see what effect it has. Trusting it has lead to the point where I’ve gone backwards and am not producing the numbers I was so I’m definitely less motivated to follow the plan now, after all I can get slower without paying for the privilege
I had trouble getting a workable FTP so I had little faith in a prediction that for me should be only a couple of watts change. It just becomes a mind game.
I’m getting good workouts and I’m happy. My ftp prediction drops as the month goes on but I’m not 100% compliant. It was predicting my ftp to remain the same by the last week…so i ramp tested to show it who is the boss. Then it said 10 watt increase. So now I’m worried my workouts will be too tough lol.
Don’t like your ftp? I say do a ramp test and prove it to the ai.
I don’t see how this would prove anything? Trainerroad themselves say that the newest version of aiFTP is not predicting the result of a ramp test… so doing a ramp test to “prove” that this is your ftp (and ignoring any issues with ramp test accuracy ) seems counterintuitive.
The latest iteration of aiFTP, according to TR, aims to place you at around a level thread threshold.
I have taken an opposite track and put a lot of trust in the system to see what happened. I have completed the two hard workouts a week diligently, used the dynamic endurance tools and the yellow days as guidance to dial back and advance. It hasn’t been all smooth sailing as I was also trying to cut weight (doing so too quickly) and had a few other life stresses going on.
A few observations
I have been blown away by how few failed workouts I have had. This happened even after unplanned week of soft pedalling to help manage some of the stress I was under and as a result of not fueling the work. There have been a number of times when I was quite timid about a work out only to complete it within the guidance suggested. A couple of times I rated it less than and a couple of times more than but could see it adjust accordingly.
I have hit personal records for pretty much all durations from 2mins to 60min in the last three weeks and coincidentally have the highest AIFTP I have ever had.
One thing I haven’t done is chased a number. Key focus is one workout at a time, one interval at a time. Lots of effort fueling properly (The EA and Sam Impey stuff has been amazing) and sleep quality, chasing the best sleep for the time I have.
Clearly this isn’t the experience of others and I only share my experience to see if there is something in it that helps others get a better experience because boy when you break through a previously thought barrier it feels pretty bloody good and I hope others can experience that.
I posted a while ago about my initial experience, and I’m seeing pretty much the same thing others are describing here: the predictions start out aggressively high and then gradually come down over time.
In my case, the detection also overshot and increased my FTP by 10 watts above my actual tested value (I retested to confirm). That said, the workout selection has been working well for me. I can see consistent progression, and I’m able to complete the sessions successfully, so I’m not too concerned about the predictions themselves.
For example, my AI Prediction for this month initially showed a 7-watt increase, then dropped to 6, and now it’s at 5. The next detection run is in a couple of days, so we’ll see how it compares.
Looking at my training log, I did my first TR workout in 2017 on a Computrainer. I moved from ErgVideo workouts as a platform for off season training. TR worked really well for that, and still does.
I’ve always wished the system could account for unstructured rides and plan training around them. It seems like we waited years for Workout Levels 2.0. With the introduction of AI, it felt like TR was finally there. After a couple of 28 day cycles of mostly riding outdoors, it has been a disappointment.
I have a detection tomorrow. Although Strava suggests I’ve gotten faster in that time, prediction suggests I’m going to lose about 3% from my FTP. I’ll wait to see if this bug fix helps, but I’m pretty close to going back to ignoring TR from April to November.
I’ve generally gone with what it’s suggested so far but experimented with responses, training approaches and workouts to see what effect it has.
I know that Nate encouraged experimentation early on, but I do think that this type of behavior can create negative outcomes. Endurance training is about steady work, consistency, and progression. As athletes continuously change their training approaches, RPE responses, training disciplines, workout types, etc., most of that consistency and progression goes out the window.
I’m in no way blaming/shaming you for doing those things, as it was encouraged early on, but I do think that sometimes that approach can lead to jagged progression curves rather than the smooth ones that steady training often produces.
Well, I’m also experimenting. Not because Nate encouraged it but because I’m not seeing the outcomes that I want, and so would prefer to use TR, which is well engineered, as a sandpit. I’d rather see the potential negative outcomes that come with my own experimentation, than those that come with blind adherence. Consistency and steady progression were well-drilled training habits well before I started using TR, so I’m reverting to the tried and tested to see what happens.
@eddie are you seeing any interesting data on users departing from the plan, making repeated changes to calendars, etc? Has that changed since the launch of AI FTP? I don’t mean how many fail their workouts, but are people shuffling things around more?
I think a big part of this issue right now is that when you build a specific plan, and you have your workouts set to outside, the prediction is higher than it should be, which makes it really hard to know how the plan you just built is actually going to turn out in the best-case scenario.
Once that bug is fixed, experimentation should be more straightforward, as you should see more accurate predictions from the very beginning.
) seems counterintuitive.
