I have been using AI FTP detection since it was offered in Trainerroad. I used the ramp test and before it the 2x 8min and 20 min test. It certainly was easier than suffering for 20 minutes and highly prefered. For most part of the year AI set my FTP around 334 - 339 Watts and I never questioned it. I also have a subscription for Xert and their system “AI” dectected very similar values. However, I looked back into my season and especially looked at 10 mile time trials and the Masters Nats TT and found that my best 20 minute performance in a TT for the year was 322 W. Most other 10 mile TTs were between 305 to 320W. I averaged 285W (307 w NP) at my TT at Nats (20 miles). Best 20 minute there was 307W. I realize that FTP does not necessarily mean that I can hold that power for 1 hour but should I not be able to hold it at least for 20 minutes? What could I be missing?
ok, we need to move beyond the “20min test” as some kind of gold standard. really if you want to determine if your ftp is your ftp is try to see how long you can ride at the threshold you believe to have. if it’s 35mins or so, cool, now you have a basis for which to work on TTE, if it’s longer, also cool, maybe you can move on to vo2 max training.
Personally I’d love to get people in a study to see how long they can ride at their ai ftp, but not sure how many people would volunteer for such a thing
I only recently took a leap of faith by trying to find my FTP by feel (using a Kolie Moore style test), and I’m so happy I did. I have no more doubts what my FTP is (+/- a handful of watts) on a given day, which has given me a lot more confidence to complete long workouts. My head unit also died on the group ride this weekend and I had no trouble pacing the climb by feel (and set a PR). I should have done this sooner.
Do you put out less power in the TT position than in road bike position? There’s a delta for most people so that might well explain it
Are you using your TT bike for your TR FTP Test and workouts? If not, different positions on the bike could be the answer. If you are using the same bike, then you probably test better with the 8 minute ramp test.
Curious, have you done any interval work on TR for 16 minutes or longer above FTP?
Very similar power for my TT bike and road bike.
I am using a Wahoo Kickrbike for my indoor workouts. I have done a few threshold workouts on TR for 16 minutes or longer.
If you power is similar on your trainer bike and TT Bike, then maybe the power reads higher on your Kickrbike…in a way, it doesn’t matter as you are still getting workouts done on the Kickr…it’s just that your outside power doesn’t line up with your TR workout numbers.
This is one issue with having a dedicated indoor bike and running different power meters.
This is the way …
How do you measure power outside, in your 10TTs?
Quarq crank on TT bike.
Yes, as someone that does TTs as described, I would expect that when in a trained state you are capable of holding your threshold power. FTP represents the border between stable and unstable physiology, and if I was in your position, I would manually override AI FTP and not use it based on the other numbers you posted.
I’m suspicious that your Kickr Bike and Quarq are showing different power numbers.
Do you have access to power pedals you could use on both to verify power numbers? Or, do a 300w 20 minute effort on the TT Bike, and then compare to a 300w 20 minute effort on the Kickr and compare RPE and Heart Rate. (Doesn’t have to be 300w, but pick something hard but achievable and use for comparison, and make sure you have a good place for the outdoor test or even borrow some rollers)
Yes if you can’t ride your FTP for at least 20 mins then it’s likely not your threshold.
I have Favero Assioma power pedals and will check how these bikes differ from a power meter perspective. I checked the Kickrbike some time ago and it matched the Favero pedals.
Reason that I am somehow questioning the AI FTP algorithms and not the different bikes with their different power meters is based on the fact that I looked into all my 20 min. efforts in 2023 across all bikes (incl. kickrbike). My best effort for 20min was below AI FTP. I raced Crits, Roadraces, TTs and Gravel races outside and even frequent Zwift races inside.
A 20 minute maximum effort is great by itself. Then, if you do a 3-5 minute maximal effort, you can use CP to estimate threshold. A simple calculation in Excel. CP is one of the 7 deadly sins and has been more reliable for me over ramp tests and many other approximations to estimate threshold.
I use CP for both running and cycling and CP is very close to mlss for me (confirmed with lactate measurements). CP is deterministic.
TTE at CP is usually around 30 minutes for me. I can feel a difference in effort and recovery slightly below and above CP.
Also, what is this AIFTP (not directed at you). Which model did TR copy and slightly modify to call it unique? Where is the white paper? Inquiring minds want to know.
Not sure if you’re just stirring the pot, but There is no whitepaper, it’s their trade secret, secret sauce.
Best description that’s out there is they’re using some type of ML models, and their very large dataset of completed (and rated) workouts to be able to infer their “AIFTP” without testing.
Lots and lots and lots of discussion and many separate threads on it, for many people it matches dead on or very close with other testing (me, for example) and for others it doesn’t. I think there’s an element of “Garbage in, garbage out” where if you’re executing TR created workouts on the trainer with a high degree of compliance, it’s going to match much better.
I’m being a little pedantic, but only because I experience this “trade secret“ stuff more for cycling data analysis than I do for high-end geotechnical engineering. WKO is similar, too. And it’s silly.
You wanna sell your model? Show me how it works. Let me see what the assumptions are. Let me run my own data through and explore the model sensitivity. I could do that in WKO by creating my own FIT files with the desired PDC. I could do the same with TR, but I’m lazy.
Just using the terms ML and AI is marketing. All fluff. No comparison data. At least show me CP (any model) vs AI FTP. And they’re using population data to estimate individual data?
Here’s the thing, they’re not selling their model. All they have to do is show results and create a product that works for people, which for a heck of a lot of people, they do. So they’re under absolutely no obligation to share it.
If I were a betting man, I’d probably be agreeing with you. Probably not an overly complex model, potentially one that is public or very well known, just trained with a pretty good and very large dataset that makes it work. That’s just my guess, but at the same time, it very well could be a custom-developed and more complex underlying model. Who knows.
For me, it matches the output of WKO pretty closely when I feed WKO with enough MMP tests, and it matches the other tests I’ve done, without the max efforts.
But they are. AI FTP is billed as part of their training system. It’s the cornerstone of personalization. No, it’s not required, but has been touted as a new feature.
So that’s why I’m curious. I like models (the numbers type).