POLL: TrainerRoad AI - for better or for worse?

Thanks for your post :slight_smile:

True, but the alternative is to have no comparable feature in the first place. ML-based algorithms can do things you couldn’t practically do at all using algorithms created by hand.

Even if in some narrow circumstances you theoretically could, the code would be incredibly complex and very hard to debug. (I wouldn’t want to be the one programming Plan Builder with all the nested if statements … :sweat_smile:)

My colleagues and I would disagree about the impenetrable part (we are mathematicians and physicists, several of my colleagues actively research the field): you can understand how a ML model functions, but you can not extract cause-and-effect relationships as well as mechanism that is encoded in a trained model (as they are based on statistics). I think you referred to the latter.

How successful an ML model does what it is supposed to hinges on two things: good training data (= high quality and large enough) as well as the right thing to optimize for (the so-called loss function). This is the secret sauce TR brings to the table (in addition to TR’s data set), it takes years of experience and domain expertise just to clean the data, for instance, and to create representative, relevant data sets for training.

At some point you may also need to make judgement calls, e. g. significant improvements to a small group (“outliers”) may make things worse for the vast majority.

You can tell by how often statistics are mentioned in the podcast, heck, entire episodes are built around it. That isn’t a coincidence.

Yes, you have a clearly outlined, systematic process to improve your models. This isn’t traditional programming, yes, it is something new that comes with new challenges, but also new possibilities.

Hey - in term of quality of discourse here, i think that use of language like ‘bitchin about’ when referring to other user’s posts is really unhelpful.

Let’s assume that everyone is posting honestly but maybe from different levels of understanding and perhaps with some genuine frustration from time to time. Calling it bitchin really isn’t adding anything positive to the forum.

I was intending to characterize specifically those people who aren’t interested in solutions but you have a good point. I’ll edit to correct the language.
Thanks :+1:t3:

Overall, I don’t think the AI was worked for me.

Through the magic of Strava MCP server and Claude, I’ve been comparing first 6 months of 2025 and 2026. I did the same events both years, so my calendar is as similar as possible.

My power numbers are down across the board. TR recommended minutes are down across the board.

Granted, I’m a year older and in the range where that matters, but I don’t think it’s all age.

I’m a little wary of overselling the usefulness of aggregated, large datasets for predicting the responses of individual users here. I could get a perfect dataset of 5 billion coin flips, but it doesn’t do me any good for predicting the next. I do AI/ML in healthcare for my day job, and while it’s a fantastic tool at a population level to drive changes at the facility/hospital/network level, it’s frequently just not that helpful for individual patients and can in fact drive bad care when practitioners make assumptions and apply macro principles at the micro level.

And while TR has a massive data with a lot of parameters, there’s still a near-infinite amount of factors that it can’t possibly have insight into: Is someone extra tried because of a low-grade illness they don’t realize they have? How stressful is work that particular day and for the last three months? How are things with the partner and/or kids? Are we getting close the anniversary of beloved family member’s death and motivation starts flagging?

ML can be incredibly accurate, but you also have to understand what you can and can’t measure, and what you can and can’t control for.

I voted somewhat better, but this not an easy choice.

Pros:

  1. Fewer failures
  2. I had lifetime FTP peak this year and I have been on TR for years (also not a spring chicken)
  3. FTP predictor tells you when hard workouts are counterproductive

Cons:

  1. Dynamic progression levels - I don’t understand them and I no longer have a static reference for workout intensity

  2. I felt pre-AI stuff was harder (cognitive dissonance with point #3 under pros)

  3. So many workouts in the inventory, but a few just keep repeating too much. AI loves Warlow (I know I can request alternates and that is appreciated)

Sorry, but I have to ask, when you say that you had a lifetime peak FTP, did you do a test or was that AIFTP, as it’s often said in the forum that this isn’t your ftp and shouldn’t match your ftp, just want to make sure we are comparing apples with apples (not trying to piss on your achievements)

Is anyone else surprised how few people have responded to this survey? There’s less than 300 people as of July 14.

That seems really good no? The forum is a small subset of TR users and other cyclists.

It was AIFTP. I didn’t follow forum discussions on whether this is actual FTP or not, but I guess what matters is performance relative to historical performance, so as long as measurement method is consistent, conclusions about relative performance can be drawn. I appreciate your sensitivity re my ego though :slight_smile:

I don’t think so. The Trainerroad Subreddit, which I would expect to be less active than the forum, gets 3,500 weekly visitors. More people have contributed other forum threads than participated in this poll.

The one that has surprised me the most is that ~10% say it’s the same.

Can we define, “before long”? I’ve given up on TR for the rest of the year and switched to another platform. If it works at the beginning of next year great. I simply don’t want to waste anymore time with my fancy new toy that doesn’t work.

3500 visitors but only 131 contributions per week. Probably a lot of Google search traffic to historic threads rather than active participants. I agree that the TR forums are more active. Interestingly one of the newest subreddit threads is a cross link to this one. :grinning_face:

Anyway, the poll was posted only a couple of days ago, so not really fair to compare engagement versus other posts that have been around for weeks, months or even years in some cases. I’d guess we’ll get 1000+ votes before it’s done. But I also suspect that the results won’t change all that much.

One takeaway is that respondents are quite “polarized” on this topic. :smiling_face_with_sunglasses:

The sample bias is a completely legitimate point. TR themselves could perhaps do a randomized poll across the user base but probably their time is better spent on addressing the underlying concerns.

That’s because coin flips are statistically independent from one another and hence, there are no correlations between them. In datasets like TR’s that’s very different, and ML can pick up on correlations that are not obvious to humans.

This kind of skepticism is good, and it is easy to get carried away by something shiny and new. That’s also why I don’t get worked up by e. g. Apple’s or TR’s lack of a “training readiness score”. If it is a random number generator, it is useless to me.

Skepticism to me also means openness to evidence whether or not the new thing is an improvement. It seems to me TR AI is, on average, a very significant improvement.

TR cannot be omniscient even if you include other sensible data like sleep data. Instead, I’d augment TR’s algorithms with human sensibilities. It doesn’t know that I’m not training, because I have to take care of my wife who has until last weekend be hospitalized, for instance.

Even with a human coach that is still true, if they don’t know relevant things about your personal life, they may prescribe you the wrong workouts.

TR has quoted statistics in its podcast several times at this point where e. g. they pointed out that failure rates have dropped significantly and consistency has improved.

The proof is in the pudding, isn’t it? If TR’s self-reported statistics are anything to go by, TR AI is a significant improvement compared to Adaptive Training. It jives with my personal experience, although that is N = 1, obviously.

Same here (I typically peak at 4.5–4.6 W/kg at age 45, with an all-time peak of 4.7 W/kg at age 41). I have used TR AI since the closed, private beta. It has worked really well for me. That’s just N = 1, of course. And I don’t train outdoors (except for endurance rides), so I have not run into the outdoor FTP bug.

I voted significantly better. My AIFTP has gone from stagnant in the 240-250 range for years to 277 or a 10% increase. I’ve also noticed that I can hang with faster people than I used to get dropped by, even only doing 1-2 of their workouts per week.

I did come close to validating the AIFTP via ramp test where I completed a minute at 360 W giving a FTP of 270 W. During that ramp test I set a bunch of power PRs, 3-12 minutes. The longer duration ones I could likely beat again if I did an interval of that length.

The AIFTP estimation is used to set a level 3 threshold workout as hard and progressively overload you for the next 28 days. Looking at the power curves of workouts, I can see that the blue line (this workout) generally barely peaks above the orange line (last 6 weeks) in places which is what you would want to help you get faster.

The dynamic endurance feature could use some work as far as adding in volume to prepare for long races. Based on the podcast and what I’ve seen, I gather that the ML model tries to increase for shorter length power PRs (<20 minutes, probably) and the data shows that too long of a ride (TSS-wise) you’ll likely skip the next workout so it’ll be “counter productive” for improving those power PRs. To combat that I largely ignored their prescribed 1-2 hour dynamic endurance rides and just joined the group rides that I wanted to do.

For me after years of happily using TR it continually downgrades my FTP when I am continuing to work as hard as ever. Utterly dispiriting. I will not be renewing next year and wish I hadn’t this year

Interesting to me that there’s now 330 votes. Just bumping this up…

Yo noto que a mi si me ha servido mucho, lo prefiero al anterior, trato de seguir el plan lo mejor que puedo ( salvo por algunos sábados que algunas veces salgo en grupo y no puedo hacer mis umbrales correctamente). Los entrenamientos que me programa me funcionan bien.


I’ve found it really helpful—I prefer it to the previous one. I try to stick to the plan as best I can (except for the occasional Saturday when I go out with a group and can’t do my threshold workouts properly). The workouts it schedules for me work well.