100% this. Sounds like a really simple way to allow fine tracking and easy comparison (incl’ historically).
Yup, if we think of the bell curve that exists in most things, some people (or use cases) are closer to the middle where the larger population exists, with others more towards the edge of the curve (hence the name). It’s not so much about the person in this instance, but the data that TR is receiving from that user and how it lands in the system (AIFTPD in this case) relative to a larger population of users getting better & desired results.
Those edge cases are legitimate and should be investigated to see if TR can improve to handle those outliers better. But it’s important to recognize that from the many other positive comments about AIFTPD, it seems to be working pretty well overall. The results are likely even better when we note that we tend to hear more about problems than successes because people are more vocal with negative issues and than positive ones.
From what I have seen and what TR tells us, AIFTPD is working well to address a large number of TR users. I replied above to point out what I saw as a broad statement questioning AIFTPD validity on a large scale, based upon the less than ideal results of these edge cases.
4 Likes
Oh definitely more thought experiment!
Just a comment really, and I saw that some of the cases did need looking at and my comment wasn’t really pointed at any one person particularly.
1 Like
I would like to have a continuous aiFTP detection. Intervals require max efforts.
1 Like
Sorry if this has been covered elsewhere in the thread - I haven’t seen it phrased in these terms yet.
I have been guilty (so guilty) of being an over tester, and not pushing through my PLs in a plan. I would like you to take the toys away from me for my own good. I would like to push those PLs up, push my endurance at higher workloads, etc…
So, if - god forbid, I do start actually moving the needle on the FTP dial - how often will TR revise my FTP? Will that 28 day gate apply in the auto model too? What are the constraints it’ll operate within to keep me pushing up those PL’s?
Here is the thing with an FTP value…
I don’t care if you call it TFTP or VFTP or even MosesHasNoSocksFTP as long as the value can be used in real life for rides’ planning and execution.
If AI FTP detection is only good for TR training then what should one do when riding a race or any other non workout challenging ride?
More than that, a few years back on one of TR YT ask a cycling coach podcast the effect and impact of incorrect value of FTP was discussed regarding threshold and VO2 max in which even a 15W higher value shall make them unproductive.
So…whatever the advised method and nomenclature of getting the FTP value you use/ recommend please make sure the value is useful both in the whole training plan as well as in the real world rides.
And…saying all of the above…the effort the TR team is putting into this tool is quite remarkable!
7 Likes
It looks like someone from the team grabbed it now 
I’ve been wondering if maybe they already use the information from every new ramp test (thousands a day?), to compare them to the calculated AI estimates and use the data to keep improving the AI algorithm. So hopefully they already catch new outliers where the difference is large.
One problem occurs to me though, which is that as everyone switches over to using the AI to calculate FTP and no-one uses the ramp test anymore then there will be less data available recently to see if the AI calculations are staying accurate …
TR reviews the following workouts by user’s (the performance outcome and ratings from them), as the main check against AIFTPD.
The whole point of AIFTPD is to avoid testing (Ramp or otherwise) so it’s not required for them to have any testing to evaluate or compare.
What matters to them is how the AIFTPD works in practice for your training.
1 Like
As suspicious/unenthusiastic as a have been of the new-fangled AIFTP and its shortcomings, I’m 100% behind the alternatives and PLs. Only negative was when my PL settings dropped too far after a second assessment after a return to training (a large jump, but one that my total training history should have been able to predict with the right analysis).
The AI/PL and adaptations have really been more predictable since my FTP has returned to TR-historical normal, from winter of liquor, lard, and laziness.
BACK TO THE MINES! 
OUT OF THE FORUM! GET IT DONE! - Me. 
I just don’t want my endurance PLs dropping because most of my extra volume rides are commutes, commutes with detours or recreational unstructured sight-seeing rides. I don’t need to progress in “Endurance” PL from an outdoor ride, but if I’m sitting at 4 .0 and I go ride outside for 2 hours with an average power in my endurance range, no matter how spikey, there’s a pretty good chance I can still complete Pettit +1, for example.
1 Like
I am not here to disagree with you or protect TR. This is more of an interesting observation. I think the problems of putting a PL on an unstructured outdoor ride is much tougher a problem than most people myself included like to believe
If you look at average power it is like look at a tree and not realizing you are in a forest. To make this point go look at Spanish Needle The power average for this is about 70% of FTP which based on your logic means it is a endurance, but I think most of us can agree it is not a endurance ride.
Accurate and fair, but I’m pretty sure Spanish Needle also doesn’t lead to a real-world decrease in “endurance” at the lower end/middle of the endurance PL. Sure, likely it doesn’t advance someone to PL 10 in endurance, but I probably shouldn’t be going from level 4 to level 2 because I haven’t done a structured “endurance” ride with blue targets…
1 Like
If you go to your calendar it looks like you can associate a workout to the outdoor ride this may help reduce the sliding PL you are talking about.
I know this is some added work that you likely don’t want to have to do but for now may be the only solution for your issue?
3 Likes
Will it only be available if your following a training plan ?
The question is can you repeatedly do your intervals in training and get faster.
There’s a “Seven Deadly Sins” of FTP post from a LONG time ago that describes different testing methods and how good they are.
"the seven deadly sins....
...er, ways of determining your functional threshold power (roughly in order of increasing certainty):
1) from inspection of a ride file.
2) from power distribution profile from multiple rides.
3) from blood lactate measurements (better or worse, depending on how it is done).
4) based on normalized power from a hard ~1 h race.
5) using critical power testing and analysis.
6) from the power that you can routinely generate during long intervals done in training.
7) from the average power during a ~1 h TT (the best predictor of performance is performance itself).
Note the key words "hard", "routinely", and "average" in methods 4, 6 and 7..."
AI FTP Detection is essentially doing #6. It’s looking at your routinely generated intervals (but we can now look at long and short) and making a prediction.
Then Adaptive Training doubles down on #6, by constantly looking at what you’re doing in training and adjusting. All the time being able to quantifiably measure that you’re getting faster if your PLs increase.
I would argue that #7 is very very hard for most athletes. I’ve never done a properly paced hour. Hour-long climbs also have elevation and oxygen to take into consideration. And I put out way less power on my TT bike. So I’d need some sort of flat hour-long course I could do on my road bike AND put on steady power (so hard…).
And then I’d have to have the fucking motivation to even do that, which is like something I can generate once or twice a season.
We knew that testing and looking at individual ride files wasn’t the best way. And we knew we can’t expect everyone to do lactate testing routinely at home. We also know people have bad days, extra good days, low motivation, high motivation, high-stress days, etc.
So we built Adaptive Training and AI FTP Detection to constantly look at the athlete and make small changes. For those engineers out there you can think of it like a PID controller for your fitness ;-D.
And we made the system so that you can still do any testing you want or input whatever number you want in that FTP box and train away.
34 Likes
I haven’t actually did an indoors session since May and looking back in my calendar it looks like they were either a just hit or larger miss. I’ve marked most of my outdoors workouts since as ‘did not pass’ but I am getting faster. Motivation probably is the key and I’m guilty of over thinking 
Thanks for the insight Nate. I think the above is a very clear reason why it’s valuable/important for this AI FTP to be ‘close enough’ to what “FTP really is”, and not just some number that can be used to scale training.
2 Likes
Can we please make this a soft gate? So instead of having the hard gate capped at 14 it is set to 12 days? If it was 28 days change that to 26? I sometimes need to move my plan around and that means these hard caps mess up the ability to use AI detection?
100% what [grwoolf] says.
@Nate_Pearson I also think that is important that AIftp is close to a persons actual ftp because people use their ftp to scale for race efforts. I have used my ftp to scale for my power target during triathlon bike legs. If AIftp diverges from that and is just a “training metric”, it loses some value.
2 Likes
You bring up a very valid point.
The counter argument to your logic would be if your most recent FTP test was 4 weeks ago and you hit all your targets during those 4 weeks is that still your true FTP for the event?